Untouched open Tcl bug tickets: Reproduced

Triage of 818 open and Pending Bug tickets of Tcl with no comment or check-in from me on 2026-10-16 (in Tcl's workflow Pending means "resolved, waiting for confirmation"). Reproduced on Linux (main 9.1 tip, 9.0 tip, 8.7b1, 8.6 tip, 8.5, 8.4 and old releases), on FreeBSD, NetBSD, OpenBSD and OpenIndiana VMs for platform tickets, and on a Windows 11 VM. Rows marked patch link to a prototype patch (287 in all, in tcl-triage/prototypes). Updated 2026-11-11: the emoji before each ticket id shows its progress.

Reproduced 320

✅409a4e5f282026Abstract-list hooks: lpop and join/lset crash, lindex reads before buffer with a negative indexREPRODUCED (main, 9.0) except F3, which is FIXED on main and 9.0 (lset-abstract-list-uaf, main c7bb2150c9 / 9.0 64dc24bd32, 2026-10-18, unreleased)…

Open ticket 409a4e5f28 · 14. List Object · filed 2026-10-18

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-10-18 check-in 63b421c72c on abstract-list-hook-fixes, in main, 9.0: Fix 409a4e5f28: abstract-list hook misuse. lpop and compiled lset took no reference to the result of setElementProc (use after free); join of a one-element li
  • 🔧 2026-11-04 check-in f6f54cd954 on abstract-list-hook-fixes-90, in main, 9.0: Cherrypick 63b421c72c for abstract list fixes - ticket 409a4e5f28.
  • 🔧 2026-11-05 check-in 92c75c9bbc on core-9-0-branch, in main, 9.0: Fix 409a4e5f28. lpop/lset/join crashes with abstract lists
  • 🔧 2026-11-05 check-in 71ed030c94 on trunk, in main: Fix 409a4e5f28. lpop/lset/join crashes with abstract lists
  • 🔧 2026-11-04 check-in a4e3bf9d77 on abstract-list-hook-fixes, in main: Merge trunk
  • 🔧 2026-10-18 check-in 1198f16e50 on lset-abstract-list-uaf, in main: Fix \[526b01c0b9\]: lset with several indices on an abstract list (lseq) used a freed sublist: it was held only by the cached elements of the list, which were f
  • 💬 2026-10-18 serhiy.storchaka: All four reproduced on trunk. Proposed fix in branch abstract-list-hook-fixes: * F1: [lpop] now takes a reference to the result of setElementProc, like [lset]. The compiled multi-index [lset] (INST_LSET_FLAT) had the same missing reference. * F2: [join] uses getElementsProc for any non-empty list. * F3: the same use-after-free as 526b01c0b9; its fix in branch lset-abstract-list-uaf makes lset l {1 0} X give a X …
  • 💬 2026-10-18 rocketship98: Thanks, this caused a crash in my own extension that claude helped me write and he says there's a documention gap: Addition to F1: doc/ObjectType.3 (SETELEMENT section) is silent on two points, and the core’s callers disagree on the second. lpop calls setElementProc with valueObj == NULL, meaning “remove this element”. A type written from the manual crashes on lpop l 1. A real extension (a bit-array type) segfault…
  • 💬 2026-11-01 apnadkarni: @rocketship98, @serhiy.storchaka, I would prefer that rather than having setElementProc be inconsistent with TclLsetFlat, we document that it should return a Tcl_Obj with the reference count incremented and owned by the caller (as is the case for TclLsetFlat). Do you see any issues with that approach? The documentation is silent but there is a direct mapping between TclLsetFlat and setElementProc and I s…
  • 💬 2026-11-05 apnadkarni: Fixed for 9.0 in 92c75c9bbc and for 9.1 in 71ed030c94. The changes proposed in this branch were applied as is so the assumption moving forward is that setElementProc will return a Tcl_Obj without its reference count incremented (so potentially 0).
Full verdict

REPRODUCED (main, 9.0) except F3, which is FIXED on main and 9.0 (lset-abstract-list-uaf, main c7bb2150c9 / 9.0 64dc24bd32, 2026-10-18, unreleased); not applicable to 8.6 (no abstract lists)

  • Platform generic
  • Tested work-a3/ab62/f1..f5.tcl with tcltest. main c092dae4af: F1 {} x (expected {a c d} x), F2 assertion failure in Tcl_JoinObjCmd (tclCmdIL.c:2194), F3 a X c (OK), F4 1/00, lset l X → "bad index". 9.0 tip build (2026-10-10): F1, F2, F4, F5 same wrong results. Branch tip alone (based on main 2026-10-17, without the F3 fix): F1/F2/F4/F5 fixed, F3 still segfaults. Branch merged into current main in wt-a3 (clean merge): all five correct; abstractlist, lseq, cmdIL, lpop, lset, listObj tests pass.
  • Cause tclCmdIL.c lpop / tclExecute.c INST_LSET_FLAT took no reference to the setElementProc result; Tcl_JoinObjCmd used indexProc for a 1-element list; lstring's indexProc (tclTestABSList.c) didn't check index < 0.
  • Existing work abstract-list-hook-fixes (63b421c72c, 2026-10-18, 42 changed lines), unmerged; needs a manual backport to 9.0 (3 of 5 hunks fail on core-9-0-branch). lset-abstract-list-uaf merged into main and 9.0. Ashok's bug-0c82728f85 (on core-9-0-branch, 3 check-ins 2026-10-22, ticket 0c82728f85 lappend/linsert ignore the Replace hook) touches the same files (tclCmdIL.c, tclExecute.c, tclListObj.c, abstractlist.test): expect merge conflicts in abstractlist.test numbering. rocketship98's doc gap (setElementProc may get valueObj == NULL from lpop; refcount of the returned object) is not addressed: doc/ObjectType.3 has no such text on main or the branch.
  • Prototype none (branch exists)
  • Suggested ticket comment "The doc gap: doc/ObjectType.3 should say that <tt>setElementProc</tt> gets <tt>valueObj</tt> NULL from <nowiki>[lpop]</nowiki> (remove the element) and that the returned object has refcount 0 (the caller takes a reference). Will add it to branch [/timeline?r=abstract-list-hook-fixes|abstract-list-hook-fixes]." (draft; only if Serhiy agrees with the refcount rule)
✅0c82728f852026lappend/linsert ignore the abstract-list Replace hook on some code pathspatchREPRODUCED (main, 9.0); 8.6 has no abstract lists

Open ticket 0c82728f85 · 14. List Object · filed 2026-10-13

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-10-26 check-in d1cb129583 on bug-0c82728f85-91, in main: Create new branch named "bug-0c82728f85-91"
  • 🔧 2026-10-28 check-in 7c3bcf94bb on bug-0c82728f85-91, in main: 9.1 fixes for b24becc2ab and 0c82728f85 abstract list crashes and shimmers
  • 🔧 2026-10-29 check-in 54a03368ce on trunk, in main: Fixes for b24becc2ab and 0c82728f85 abstract list crashes and shimmers
  • 💬 2026-10-16 rocketship98: NOTE: this is a claude fix, I can't be certain this is correct, sorry but he has built and then tested this with and without the patch. This is not something I understand, but the fix is small and hopefully someone will benefit from this and the tests. Where it says I, it means Claude Follow-up with a tested fix: a patch (9 lines added, 2 removed) and six tests, both below. Against trunk of 2026-10-15, commit 102a…
  • 💬 2026-10-20 apnadkarni: Considering other C list functions have been changed to account for abstract list hooks, I can only surmise this was an oversight. Will take a look. It's a bit curious that there is no equivalent abstract list method for "lappend". The lreplace is a superset but I can imagine abstract list implementations where it's overhead is non trivial when compared to a lappend hook.
  • 💬 2026-10-20 rocketship98: Thanks for looking at this, Ashok. Some context on who is hitting it. The bit-array type I mentioned is a real extension that implements all eight abstract-list hooks of TCL_OBJTYPE_V2 (length, index, slice, reverse, getElements, setElement, replace and the in/ni operator). lappend, and linsert in the uncompiled case, are the only commands I have found where the core converts the value to an ordinary list itself i…
  • 💬 2026-10-20 rocketship98: one quick note, having claude do builds (binary search he says) between 9.0.4 and 9.1 to pinpoint when the behavior changed.
  • 💬 2026-10-21 rocketship98: While looking at why linsert behaves differently from lappend for us, I believe I found the cause of the linsert case (claude is more certain, and he's the I below in this comment). In Tcl_LinsertObjCmd (tclCmdIL.c) there is a special case for inserting exactly one element at the end of the list: if ((objc == 4) && (index == len)) { result = Tcl_ListObjAppendElement(NULL, listPtr, objv[3]); Tcl_ListObjAppendElemen…
  • 💬 2026-10-22 apnadkarni: It appears as though the public list related functions (Tcl_Xxx) have afaict been all changed to check for abstract list methods. However, some internal functions / macros (TclXxx) have been modified similarly but others have not. The unmodified ones depend on the caller to make the check. Not sure whether that is by intent or oversight. Seems to me it would make sense to make it consistent throughout. Started…
  • 💬 2026-10-23 apnadkarni: Just a note that makes me a bit wary of using lstring for testing after inspecting the source. Its implementation breaks EIAS. Consider `` % set l1 [lstring abc] a b c % set l2 [list a b c] a b c % lappend l1 xyz a b c x % lappend l2 xyz a b c xyz `` That's a no-no. Additionally, it's implementation of replaceProc is incomplete in that it does not handle multiple indices. It may suffice for very specific test ca…
  • 💬 2026-10-23 oehhar: Thanks Ashok! What is lstring? Is is an abstract list, which presents a string as a list where each character is a list element? Then, you can not add "xyz", as it violates this abstract list. Ok, I understnd the example. There are two possible solutions: * error out * transform to tradional list Take care, Harald
  • 💬 2026-10-23 apnadkarni: Harald, lstring is only part of the test suite used to test abstract lists. Erroring out is not an option IMO as a command must treat two string representations ("a b c" in both cases) exactly the same. It cannot accept one and not the other. For something like lstring, only alternative is the second solution you proposed. I did not mean to imply lstring needs to be fixed, it's only used in tests. Rather, my point…
  • 💬 2026-10-23 oehhar: Thanks for the explanation, I appreciate! Take care, Harald
  • 💬 2026-10-27 apnadkarni: Branch bug-0c82728f85 has been merged into 9.0 to pick up committed and verified fixes. Please don't close branch bug-0c82728f85. The abstract list test object lstring implementation test suite is missing tests which result in crashes within lstring itself preventing the missing tests from being added. These are still being worked on. 9.1/trunk still pending because of significant differences in both the bytecode …
  • 💬 2026-10-29 apnadkarni: Fixed for 9.1 in 54a03368ce and 9.0 in d27dbcd1b3.
  • 💬 2026-10-29 apnadkarni: Just a note: not all shimmers have been eliminated. In particular, nested indices and the lset implementation not tackled. Partly because not worth fixing in practice imo and partly because of complexity. lset for instance works on an unshared value while the corresponding abstract list method returns a new object. (Shimmer will occur with the C API, not with the script level command.)
  • 💬 2026-10-29 oehhar: Great work, thank you! Perhaps the documentation of lseq may talk a bit about conversion to normal list. I suppose, that is always the case if a list element is written (lset). Thanks for all, Harald
  • 💬 2026-10-30 apnadkarni: > Perhaps the documentation of lseq may talk a bit about conversion to normal list. I don't think we ever document internal shimmering details and imo nor should we. As far as scripts are concerned, a list is a list is a list. The internal implementation should not matter. Note the same is also true for string which can internally shimmer into various forms.
  • 💬 2026-11-04 oehhar: Or a document on the tcl core wiki? I see the next command (concat) being handled... Thanks for all, Harald
🧪 Prototype check (2026-10-24)

0c82728f85.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before ./tcltest r0c8.tcl (the ticket's lstring script) -> lappend 1 elem compiled in a proc : list, lappend 2 elems compiled in a proc : list, linsert end 1 elem NOT compiled : list (same as the report)
  • Patch applied cleanly
  • After same script -> all seven lines lstring; tests abstractlist (124/0 failed), lseq, linsert, listObj, list, lreplace, lrange, lindex, append, appendComp: 0 failed
  • Test existing abstractlist-bug-0c82728f85 fails before / passes after
Full verdict

REPRODUCED (main, 9.0); 8.6 has no abstract lists

  • Platform generic
  • Tested reporter's script with own main tcltest (9.1, wt-a1) and tcl9.0 tip tcltest; identical output to the report on both.
  • Cause generic/tclListObj.c:TclListObjAppendElements (used by Tcl_ListObjAppendElement, compiled lappend in procs, and the single-element-at-end shortcut of Tcl_LinsertObjCmd) goes straight to TclListObjGetRep, which converts the abstract list; only Tcl_ListObjReplace consults the type's replaceProc.
  • Prototype prototypes/0c82728f85.patch (56 lines, test abstractlist-bug-0c82728f85). In TclListObjAppendElements, if the object has replaceProc and lengthProc, delegate to TclObjTypeReplace(interp, toObj, length, 0, elemCount, elemObjv). Fixes all three cases; abstractlist/list*/lappend/linsert/lseq tests pass (11700 tests, 0 failed). Design question (whether append should honour the hook) still for the maintainers; applies to 9.0 as well.
  • Suggested ticket comment TclListObjAppendElements() does not check for a replaceProc, unlike Tcl_ListObjReplace(), so lappend in a proc and single-element "linsert ... end" convert the abstract list. Proposed fix: delegate to the replaceProc there (patch with test abstractlist-bug-0c82728f85).
⬜b989a6a0e02026regexp -nocase: quantified k misses U+212A KELVIN SIGN that bare k matchespatchREPRODUCED (main, 9.0, 8.6, 8.5)

Open ticket b989a6a0e0 · 43. Regexp · filed 2026-10-12

🧪 Prototype check (2026-10-24)

b989a6a0e0.patch: VERIFIED. It is over the size guideline, the table should be produced by tools/uniParse.tcl, and the glob path and backrefs stay ToLower-only. Also note that i now matches U+0131/U+0130 under -nocase (the same as Python's re.IGNORECASE, but string equal -nocase does not do this).

  • Base main (tip d6dd6d0f71)
  • Before work-a3/vb989.tcl against "K" -> ^k$=1 ^k{1}$=0 ^k+$=0 ^[k]$=0 ^[k]+$=0 ^[a-z]$=0; regexp -nocase {^K+$} K = 0; ſ (U+017F) vs ^[s]$/^s+$ = 0
  • Patch applied cleanly (80 changed lines, about 15 of them the generated table). I regenerated the 53-entry table from the tip's case tables (work-a3/vb989gen.tcl: c with fold=lower(upper(c)) not in {fold, upper, title}), and it matches the patch exactly.
  • After all the ticket forms and [a-z] -> 1; {^K+$} vs K -> 1; ſ vs [s]/s+ -> 1. As the triage said, bare ^s$ vs ſ is still 0 (glob path, ToLower only). Tests regexp (275/0 failed), regexpComp, reg (1072/0), string, utf: 0 failed
  • Test existing regexp-33.1 and 33.2 (from the patch) fail on the unpatched tip / pass after
Full verdict

REPRODUCED (main, 9.0, 8.6, 8.5)

  • Platform generic
  • Tested own main build, 9.0.5, 8.6.18, 8.5.19 — same results on all. Also asymmetric: regexp -nocase {^K$} K is 0 while it matches k.
  • Cause generic/regc_locale.c:allcases()/range() build the -nocase set from only ToLower/ToUpper/ToTitle of the pattern char, so {k,K} never gets U+212A (whose lower is k, but nothing maps back to it). Bare ^k$ "works" only because the compiler turns it into a string match -nocase (TclReToGlob), which compares ToLower of both sides. A scan of the Tcl case tables finds 53 such one-way mappings (U+212A, U+212B, U+2126, U+017F, U+0130/0131, U+03C2, U+1C80–1C88, U+2C6x, U+A7xx, ...).
  • Prototype prototypes/b989a6a0e0.patch (175 lines incl. a generated 53-entry table and tests regexp-33.1/33.2; bigger than the "small" limit). Adds a static {fold, other} table (fold = ToLower(ToUpper(c))) to regc_locale.c; allcases() and range() add the other members of the class. Fixes all forms in the report and ranges like [a-z]; reg*/string* tests pass (3273, 0 failed). A proper fix should generate the table in tools/uniParse.tcl along with tclUniData.c. Remaining inconsistency: the glob path still uses ToLower only, so bare ^s$ does not match U+017F while [s] now does; backrefs (casecmp) are also ToLower only.
  • Suggested ticket comment The regexp -nocase sets come from ToLower/ToUpper/ToTitle of the pattern char only, so chars like U+212A whose lowercase is k are not reached from k; bare ^k$ matches only because it is compiled to string match. Proposed fix: add the missing case counterparts (53 in the current tables) in regc_locale.c allcases()/range(); prototype with tests regexp-33.*.
🟢1c722347fe2026Tool generate wrong links for targets containing spaces/newlinespatchREPRODUCED (main)

Open ticket 1c722347fe · 53. Configuration and Build Tools · filed 2026-10-06

🟢 Since triage (updated 2026-11-11)

Status now: Open

  • 🔧 2026-10-20 check-in 53e5bd4306 on man2html-keyword-anchors, not merged: Fix 1c722347fe: the HTML documentation had about 300 dead links to the Keywords pages. The links in the KEYWORDS section used the raw keyword, while the ancho
  • 🔧 2026-10-20 check-in d61292cc5f on man2html-keyword-anchors, not merged: Use nospace-text for all anchor links, as suggested by Thomas (TWu).
  • 🔧 2026-10-20 check-in 66f0f3eaa0 on man2html-meta-viewport, in main: Fix 7f6005a867: add a viewport meta tag to the generated HTML pages, so they render properly on mobile browsers.
  • 💬 2026-10-20 serhiy.storchaka: Thank you, Thomas. The attached ZIP is not available to me, so I made the fix from your description. Confirmed on main: a link check of the HTML generated for Tcl and Tk finds 302 dead anchors and one dead page. The links in the KEYWORDS sections did not go through nospace-text, which also replaced only single spaces (some keywords contain a newline), and there was no Keywords page for "3D". Proposed fix in branch…
  • 💬 2026-10-20 anonymous: Thank You very much, Serhiy! Unfortunately this is not the complete request. The anchors are left. Probably our URI-check found only the connection to the main commands or does not check the name=/id= values at all as local targets from other pages. I use "CSS HTML Validator 2026" to found all problems in the doc of 9.0.4 version. I save my solution to https://blindenfreizeiten.de/TCL/tools_tcltk-man2html.zip so Y…
  • 💬 2026-10-20 serhiy.storchaka: Thank you, Thomas, the ZIP is available now. Most of it was already in the branch (nospace-text maps any whitespace, including newlines, and there are Keywords pages for names starting with a digit). The remaining links now also go through nospace-text in d61292cc5f: the three #$here links of the tables of contents, the M$name standard option link and the #$manual(...-id-...) cross-reference. Their targets alrea…
  • 💬 2026-10-21 anonymous: I accept the both fix-points for version 9.1.0 - thank You, Serhiy! @Harald: Can they merged too? I will look on the remaining dl, dd, dt and ol fails, open a new ticket on finding a solution...
  • 💬 2026-10-21 serhiy.storchaka: The meta viewport is merged now, into trunk (785fa0c704), core-9-0-branch and core-8-6-branch.
Full verdict

REPRODUCED (main)

  • Platform docs
  • Tested main tools/tcltk-man2html.tcl run with own main build (Tcl only and Tcl+Tk docs), link checker over the generated HTML. The attached ZIP is not in the local repo (blob missing), so the reporter's files were not compared.
  • Cause tools/tcltk-man2html-utils.tcl: the KEYWORDS section writes href="../Keywords/X.html#$key" with the raw keyword, while the Keywords pages use [nospace-text $k] for the anchor, so every multi-word keyword link is dead (192 dead anchors in Tcl docs, 302 with Tk). nospace-text only maps single spaces. tcltk-man2html.tcl: the letter list is A-Z only, so Tk's "3D" keyword links to a non-existent Keywords/3.html.
  • Prototype prototypes/1c722347fe.patch (12 changed lines, no test: build tool; verified with a link checker: 302 -> 0 dead anchors, 3.html now generated)
  • Suggested ticket comment Confirmed: keyword links in the KEYWORDS section did not go through nospace-text, and keywords starting with a digit had no index page. Proposed fix: apply nospace-text to the href (and make it map any whitespace run), and add index pages for leading digits. The meta-viewport and the ol/dd/dl nesting are separate issues.
🟢05b158ed562026http::geturl -keepalive intermittent socketWrState errorspatch 3764f4e81fpatchREPRODUCED (main, 9.0); not in 8.6 (http 2.9.9)

Open ticket 05b158ed56 · 29. http Package · filed 2026-09-15

🟢 Since triage (updated 2026-11-11)

Status now: Open

  • 🔧 2026-10-16 check-in 00a815eb5e on http-keepalive-reuse-closed, in main: Fix 05b158ed56 and 3764f4e81f: http::OpenSocket opens a new connection if the persistent socket it was going to reuse has been closed in the meantime (e.g.
  • 🔧 2026-11-10 check-in 98a1b78011 on trunk, in main, 9.0: Fix 05b158ed56 and 3764f4e81f: http::OpenSocket opens a new connection if the persistent socket it was going to reuse has been closed in the meantime
  • 🔧 2026-11-10 check-in 655b9df308 on core-9-0-branch, in 9.0: Fix 05b158ed56 and 3764f4e81f: http::OpenSocket opens a new connection if the persistent socket it was going to reuse has been closed in the meantime
  • 💬 2026-10-16 serhiy.storchaka: Reproduced on 9.0 and main with a local server that closes idle keep-alive connections: http::CheckEof closes the idle socket after geturl has chosen to reuse it but before http::OpenSocket runs, so the reuse path reads the removed socketWrState element. Proposed fix in branch http-keepalive-reuse-closed: OpenSocket opens a new connection in that case. Test http-4.19. It also fixes 3764f4e81f.
Full verdict

REPRODUCED (main, 9.0); not in 8.6 (http 2.9.9)

  • Platform generic
  • Tested main (own build, http 2.10.3), 9.0.5 (http 2.10.2), 8.6.18 (http 2.9.9), with a localhost server that keeps the connection alive but closes it while the client is idle (work-a4/s/05b_server.tcl, 05b_client.tcl). Every reuse after a server close fails on 9.x.
  • Cause library/http/http.tcl: geturl (PreparePersistentConnection path) decides to reuse the persistent socket and schedules OpenSocket with after idle; in http::wait's vwait the pending readable event runs CheckEof -> CloseSocket -> CloseQueuedQueries -> Unset, which removes socketMapping/socketWrState for the connection. OpenSocket then still takes the reusing branch, and ScheduleRequest reads the unset socketWrState($connId).
  • Prototype prototypes/05b158ed56.patch (79 lines, test http-4.19.*): in OpenSocket, if $state(reusing) but socketMapping($connId) no longer exists, switch to a fresh placeholder, start the timeout timer and call PreparePersistentConnection, i.e. open a new connection. The repro passes; http*.test passes.
  • Suggested ticket comment Reproduced on 9.0 and main with a local server that closes idle keep-alive connections. The idle socket is closed by http::CheckEof after geturl has chosen to reuse it but before http::OpenSocket runs, so the reuse path reads the removed socketWrState element. Prototype: open a new connection in OpenSocket in that case; test http-4.19.
⬜7bad415e3e2026Free-form clock scan reads "+02" as two minutes in 8.6 (backport of 51aa536160 without semantic predicates); the 9.x fix misreads a zone followed by a numberREPRODUCED (8.6: +02/-02 trailing zone read as minutes; main, 9.0: zone followed by a number, e.g. +0200 15 Nov 2023 00:13:20 -> 1699935200, …

Open ticket 7bad415e3e · 06. Time Measurement · filed 2026-09-03

Full verdict

REPRODUCED (8.6: +02/-02 trailing zone read as minutes; main, 9.0: zone followed by a number, e.g. +0200 15 Nov 2023 00:13:20 -> 1699935200, +02 2023-11-15 00:13:20 -> 1700007080 on all three)

  • Platform generic
  • Tested main c3bf6dd052 (own build), 9.0.5, 8.6.18; work-a6/t1.tcl
  • Cause generic/tclGetDate.y zone: sign tUNUMBER (8.6 always hhmm); in 9.x nmzone uses yyDigitCount, which already belongs to the lookahead number when one follows.
  • Prototype none (reporter's branch bug-7bad415e3e-8.6 138b037189, 2026-09-08, exists and is unmerged; sebres is reviewing and prefers not to change 8.6)
  • Suggested ticket comment Still reproducible: 8.6.18 reads a trailing +02 as minutes, main and 9.0 misread a zone followed by a number (+0200 15 Nov 2023 00:13:20). Branch bug-7bad415e3e-8.6 is not merged; the lookahead part would also apply to 9.x nmzone.
⬜9d79d6b2f22026winFCmd tests use the wrong drive.patchREPRODUCED by code reading (main: test disabled with nonPortable; 9.0/8.6 still hardcode the drive); on this VM C: and D: both exist, so the hard…

Open ticket 9d79d6b2f2 · - New Builtin Commands · filed 2026-08-18

🧪 Prototype check (2026-10-24)

9d79d6b2f2.patch: VERIFIED

  • Base main (tip c092dae4af)
  • Before winFCmd-16.10 SKIPPED: nonPortable. The VM has C: and D: (the VBox Guest Additions CD, CDFS, cd works), so the old hard-coded D: would happen to be right here; a machine whose second drive is not D: was not available (no subst, as no system changes are allowed)
  • Patch applied cleanly, 28 changed lines, not changed
  • After winFCmd-16.10 PASSED (the probe picks D:) when run via cmd.exe and also when tcltest91 is started directly from PowerShell (no =C: variable); winFCmd.test admin 234 total, 3 failed (6.1, 6.13, 9.3 as baseline), low 0 failed
  • Test the patch is a test fix; the test now runs and passes
Full verdict

REPRODUCED by code reading (main: test disabled with nonPortable; 9.0/8.6 still hardcode the drive); on this VM C: and D: both exist, so the hardcoded choice happens to work

  • Platform Windows (test suite)
  • Tested code reading; main's winFCmd-16.10 was turned off with nonPortable by 08133f796d (2026-09-14) but still hardcodes D:/C:. In 9.0/8.6 the test is different (file norm ${dd}:foo with no cd), still uses the hardcoded name, and only checks syntax. Repro: win/9d79d6b2f2.tcl
  • Cause tests/winFCmd.test winFCmd-16.10 picks dd as D: (or C:) instead of finding a volume that exists. The second failure (run from PowerShell) is outside Tcl: the per-drive current directory is kept in cmd.exe's hidden =X: environment variables, which PowerShell does not set.
  • Prototype prototypes/9d79d6b2f2.patch (28 changed lines, test winFCmd-16.10 given a new constraint otherDrive, which looks for a native volume other than the current one that can be cd'd to; nonPortable removed). The PowerShell case still fails; notInCIenv is kept.
  • Suggested ticket comment Proposed fix: winFCmd-16.10 now finds a second native drive with &#91;file volumes&#93; and no longer assumes D:. The failure when the suite is started from PowerShell remains: the per-drive current directory comes from cmd.exe's hidden =X: environment variables.
Windows VM, run 3
  • Platform Windows (test suite)
  • Verdict REPRODUCED by code reading (main: test disabled with nonPortable; 9.0/8.6 still hardcode the drive); on this VM C: and D: both exist, so the hardcoded choice happens to work
  • Tested VM win/9d79d6b2f2.tcl: volumes C:/ D:/; cd D:; file norm C:foo → cwd/foo on 9.1b1 even when started from PowerShell (ssh), but C:/foo on 8.6.18 from PowerShell (no =C: variable). Prototype on tcl-wt91: winFCmd.test with tcltest91 under cmd.exe: 16.10 PASSED (unpatched: SKIPPED nonPortable), 172 passed / 3 failed — the same 3 failures (winFCmd-6.1, 6.13, 9.3, ACL/permission tests) occur without the patch. 8.6.18 winFCmd-16.* all pass.
  • Cause tests/winFCmd.test winFCmd-16.10 picks D: (or C:) instead of a volume that exists
  • Prototype prototypes/9d79d6b2f2.patch (existing, 28 changed lines, constraint otherDrive) tested on Windows: works when a second drive exists; the no-second-drive case could not be tested (no subst drives allowed here). Tested on Windows.
  • Suggested ticket comment Proposed fix: winFCmd-16.10 finds a second native drive with &#91;file volumes&#93; instead of assuming D: (constraint otherDrive), tested on Windows 11; with it 16.10 runs and passes on main again. Note 9.1 normalizes C:foo correctly even when started from PowerShell, while 8.6.18 needs cmd.exe's =C: variable.
✅0e3f777ed52026Segfault upon sequence of execution tracesREPRODUCED (main, 9.0, 8.6)

Open ticket 0e3f777ed5 · 46. Traces · filed 2026-07-25

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-11-10 check-in 48d092ceb0 on interp-trace-delete-self, in main: 0e3f777e merge main
  • 💬 2026-11-10 oehhar: Fix and test are effective. Will merge tomorrow if CI is clean. Thanks Serhiy, Harald
  • 💬 2026-11-11 oehhar: Merged to: * main 245d2b1c * core-9-0-branch ff4839e9 * core-8-6-branch a583dec2 Thanks for all. Bug closed. Harald
Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform generic
  • Tested reporter's exercise.tcl on main (2026-10-22), 9.0.5, 8.6.18: SIGSEGV after the first "p invoked"; branch tip 711f947fb2: both calls run, trace.test 291 total, 0 failed
  • Cause generic/tclBasic.c TclCheckInterpTraces(): when an enterstep trace deletes itself while another one is set, the reverse scan continues from the deleted trace and runs off the list
  • Existing work branch interp-trace-delete-self (tip 711f947fb2, 2026-10-14), unmerged (not in trunk/9.0/8.6); merges into current trunk without conflicts (dry run)
  • Prototype none (branch exists)
  • Suggested ticket comment none needed (Serhiy's 2026-10-14 comment is current).
⬜e27c3246c32026Tcl_UtfToChar16 should replace invalid chars with U+FFFDpatchREPRODUCED (main, 9.0; 8.6 not applicable — Tcl 8 keeps cp1252/Latin-1 fallback by design)

Open ticket e27c3246c3 · - New Builtin Commands · filed 2026-06-05

🧪 Prototype check (2026-10-24)

e27c3246c3.patch: VERIFIED (affects every Win32 API call on Windows, not run there)

  • Base main (tip d6dd6d0f71)
  • Before build with only the test command (tclTest.c) from the patch: utf-27.2 FAILED: {0041 0080 0042} {0178 00A0 00BF} {00C3 0042 00E4 00B8 0042 00F0 009F 0042} {00C1 0081 00F8 00FF} (a leading \x80 stays U+0080, \x9F becomes U+0178 via cp1252, broken leads stand for themselves)
  • Patch applied cleanly (tclTest.c hunks offset up to 142 lines; 90 changed lines, over the size limit as the triage said: ~15 in tclUtf.c)
  • After {0041 FFFD 0042} {FFFD FFFD FFFD} {FFFD 0042 FFFD FFFD 0042 FFFD FFFD 0042} {FFFD FFFD FFFD FFFD}; tests utf 401/0 failed, encoding 230/0, string 1411/0, stringObj 85/0, icu 58/0, cmdIL 164/0, format 271/0, scan 184/0, regexp 273/0; doc/Utf.3 paragraph renders correctly. TclNumUtfChars also calls Tcl_UtfToChar16 but only uses the byte count, which is unchanged
  • Test existing utf-27.2 fails before (with the test command) / passes after; utf-27.1 (valid input) passes on both
Full verdict

REPRODUCED (main, 9.0; 8.6 not applicable — Tcl 8 keeps cp1252/Latin-1 fallback by design)

  • Platform generic (mainly matters on Windows, where Tcl_UtfToWCharDString = Tcl_UtfToChar16DString feeds every Win32 API call)
  • Tested main build with a new testutftochar16 test command. Before the patch, A\x80B gives 0041 0080 0042; in the middle of a string the same \x80 gives U+20AC via cp1252, so the result changes with the position. Naked 0xA0–0xBF and broken lead bytes come out as themselves. 9.0 has the same code (by reading).
  • Cause generic/tclUtf.c Tcl_UtfToChar16 maps 0x80–0x9F through the cp1252 table and every other invalid byte to itself. The tail loop of Tcl_UtfToChar16DString also stores incomplete sequences as UCHAR(*p++). tclIcu.c already avoids Tcl_UtfToChar16DString for this reason (comment there).
  • Prototype prototypes/e27c3246c3.patch (90 changed lines, over the size limit: about 15 in tclUtf.c, the rest is the test command testutftochar16 in tclTest.c, a doc/Utf.3 paragraph for Tcl_UtfToChar16, and tests utf-27.1/27.2). Invalid bytes become U+FFFD. The surrogate-pair path (high surrogate from the 4-byte lead, low surrogate from the trail bytes) and the C0 80 NUL are kept. utf/encoding/string/stringObj/icu/cmdIL tests pass. Tcl_UtfToUniChar(DString) keeps cp1252 (not touched; changing it would affect all string handling).
  • Suggested ticket comment Confirmed on main and 9.0. Proposed patch makes Tcl_UtfToChar16 and the tail of Tcl_UtfToChar16DString store U+FFFD for invalid or incomplete sequences instead of cp1252 / the raw byte, with a testutftochar16 test command and tests utf-27.*. This is a visible behaviour change for Windows API strings, so it probably needs a TIP or at least a changes.md entry.
🔧76952ef5922026libtclzip unavailable for install on Windows when building with noembedpatchREPRODUCED (main): nmake core shell dlls OPTS=noembed then nmake install-libraries OPTS=noembed fails with File not found - libtcl9.1b1.zip /…

Open ticket 76952ef592 · 53. Configuration and Build Tools · filed 2026-06-04

🔧 Since triage (updated 2026-11-11)

Status now: Open

  • 🔧 2026-10-20 check-in 53bf37e425 on win-install-libraries-zip, not merged: Fix 76952ef592: in makefile.vc, install-libraries depends on libtclzip, so the library zip is built for OPTS=noembed instead of the install failing.
  • 💬 2026-10-20 serhiy.storchaka: Confirmed on main with MSVC 2022: after nmake core shell dlls OPTS=noembed, nmake install-libraries OPTS=noembed fails with "File not found - libtcl9.1b1.zip". Keeping the current design (the zip is installed for noembed builds too), the install could build the zip instead of failing. Proposed fix in branch 53bf37e425: install-libraries depends on libtclzip. Tested with nmake: the zip is built and installed.
Full verdict

REPRODUCED (main): nmake core shell dlls OPTS=noembed then nmake install-libraries OPTS=noembed fails with File not found - libtcl9.1b1.zip / xcopy return code 4

  • Platform Windows (nmake)
  • Tested code reading of win/makefile.vc on main (unchanged on 9.0): core depends on libtclzip only when TCL_EMBED_SCRIPTS, but install-libraries unconditionally copies $(TCLSCRIPTZIP) without depending on it. Unix not affected (different build). Repro script: win/76952ef592.tcl (run in an MSVC env against a scratch checkout).
  • Cause win/makefile.vc: install-libraries lacks a libtclzip prerequisite.
  • Prototype prototypes/76952ef592.patch (1 line: install-libraries: libtclzip ..., untested on Windows). Ashok's last comment says the current behaviour is by design; the patch is the low-cost alternative of building the zip on demand instead of failing.
  • Suggested ticket comment If installing the zip for noembed builds is intended, install-libraries could simply depend on libtclzip, so nmake install-libraries OPTS=noembed builds it instead of failing. One-line change, needs a Windows check.
Windows VM, run 1
  • Platform Windows (nmake)
  • Verdict REPRODUCED (main): nmake core shell dlls OPTS=noembed then nmake install-libraries OPTS=noembed fails with File not found - libtcl9.1b1.zip / xcopy return code 4
  • Tested VM tcl-wt91 (main of 2026-09-21), MSVC 2022, separate BUILDDIRTOP and INSTALLDIR under %TEMP%
  • Cause win/makefile.vc: install-libraries copies $(TCLSCRIPTZIP) but does not depend on libtclzip (only core does, and only when embedding)
  • Prototype prototypes/76952ef592.patch (1 line, unchanged): with it install-libraries builds and installs lib\libtcl9.1b1.zip. Tested on Windows.
  • Suggested ticket comment Confirmed on main: with OPTS=noembed, install-libraries fails because the zip was never built. If installing it is intended, a one-line fix is to make install-libraries depend on libtclzip (tested with nmake).
⬜477e96c2502026classvariable inside classmethod creates unusable variablepatchREPRODUCED (main, 9.0, 8.7b1); 8.6 has no classmethod/initialize

Open ticket 477e96c250 · 35. TclOO Package · filed 2026-06-04

🧪 Prototype check (2026-10-24)

477e96c250.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before tcltest 477e.tcl (initialize sets m 1; classmethod setit {} {classvariable m; incr m}) -> setit: 1 showit: 1 (the classmethod gets a fresh variable in the delegate's namespace); the same inside namespace eval ::app -> app: 1 10
  • Patch applied cleanly (34 changed lines). I checked that the name test fits the tip: TclOOGetClassDelegateName builds <class object namespace>:: oo ::delegate, so the parent namespace is the class's object namespace (::oo::ObjNN), also for classes created in other namespaces
  • After -> setit: 2 showit: 2, app: 11 11; tests oo.test 373 passed/16 skipped/0 failed, ooUtil.test 34/0/0, ooNext2.test 57/5/0, ooProp.test 59/0/0
  • Test existing ooUtil-6.4 fails before / passes after
Full verdict

REPRODUCED (main, 9.0, 8.7b1); 8.6 has no classmethod/initialize

  • Platform generic
  • Tested wt-a3 main, 9.0.5, 8.7b1
  • Cause generic/tclOOBasic.c:TclOOClassVariableObjCmd uses the method's declaringClassPtr; for a [classmethod] that is the class delegate "<ns>:: oo ::delegate", so the variable is created in the delegate's own namespace instead of the class's namespace.
  • Prototype prototypes/477e96c250.patch (69 lines incl. test; C change 14 lines, test ooUtil-6.4). Detects the delegate by its command name (namespace " oo " / command "delegate") and uses the parent namespace, i.e. the owning class's namespace; works for subclasses too. oo*.test all pass.
  • Suggested ticket comment The delegate class that holds classmethods is taken as the "declaring class", so &#91;classvariable&#93; links into the delegate's namespace. Proposed fix (map the delegate back to its class) with test ooUtil-6.4.
⬜0be26484a02026mutex in TclpFinalizeNotifier in the epoll and kqueue notifierspatchREPRODUCED (main, 9.0; 8.6 has no epoll/kqueue notifier). The mutex is dead code, so this is a cleanup, not a functional bug.

Open ticket 0be26484a0 · 02. Event Loops · filed 2026-05-25

🧪 Prototype check (2026-10-24)

0be26484a0.patch: VERIFIED (kqueue half by reading only)

  • Base main (tip d6dd6d0f71)
  • Before code reading on the tip: tsdPtr->notifierMutex in tclEpollNotfy.c/tclKqueueNotfy.c is only initialised in PlatformEventsInit and locked/unlocked/destroyed in TclpFinalizeNotifier (grep: lines 311/337/338/375 epoll, 284/301/302/343 kqueue), never used by TclpAlertNotifier or anything else; Tcl_ThreadAlert/TclFinalizeNotifier serialise on listLock (generic/tclNotify.c). Cleanup, no functional bug to reproduce.
  • Patch applied cleanly (offsets 25 and 4 lines; 44 changed lines). After it no tsdPtr->notifierMutex reference remains (the remaining notifierMutex hits are the separate static one in tclSelectNotfy.c/tclUnixNotfy.c).
  • After the default Linux build uses epoll (NOTIFIER_EPOLL=1): tclEpollNotfy.o rebuilt without warnings; make test TESTFLAGS="-file 'event.test io.test socket.test thread.test timer.test notify.test interp.test'" -> 1746 total, 1644 passed, 102 skipped, 0 failed. kqueue file not compilable here (no sys/event.h); its diff only removes the same field/init/lock/unlock/destroy lines and comments.
  • Test none possible (dead-code removal, no behaviour change)
Full verdict

REPRODUCED (main, 9.0; 8.6 has no epoll/kqueue notifier). The mutex is dead code, so this is a cleanup, not a functional bug.

  • Platform unix (epoll: Linux; kqueue: BSD/macOS)
  • Tested code reading; tsdPtr->notifierMutex is used only in PlatformEventsInit (init) and TclpFinalizeNotifier (lock/unlock/destroy) in both tclEpollNotfy.c and tclKqueueNotfy.c, and never by TclpAlertNotifier. Tcl_ThreadAlert and TclFinalizeNotifier already serialize on listLock in generic/tclNotify.c, so alert vs finalize is protected anyway.
  • Cause leftover from the select notifier, which uses the global notifierMutex in TclpAlertNotifier.
  • Prototype prototypes/0be26484a0.patch (44 changed lines, no test since behaviour is unchanged): removes the field, init, lock/unlock/destroy and fixes the comments in both files. Built with --enable-epoll in a separate dir; event/io/socket/thread/timer/notify/interp tests pass (1745 total, 0 failed). kqueue part compiled only by inspection (CANNOT-TEST here, needs a BSD VM).
  • Suggested ticket comment Agreed: notifierMutex is only taken in TclpFinalizeNotifier, and alert vs finalize is already serialized by listLock in tclNotify.c (Tcl_ThreadAlert, TclFinalizeNotifier). Proposed cleanup removes it from both the epoll and kqueue notifiers.
🔧67769bcad32026Tcl_RegisterConfig idempotency issuepatchREPRODUCED (main, 9.0, 8.6)

Open ticket 67769bcad3 · 54. Configuration Reporting · filed 2026-05-07

🔧 Since triage (updated 2026-11-11)

Status now: Open

  • 🔧 2026-10-20 check-in d9ca7f22fe on registerconfig-twice, not merged: Fix 67769bcad3: a second Tcl_RegisterConfig call for the same package merges its keys instead of losing the package.
  • 💬 2026-10-20 serhiy.storchaka: Confirmed on main, 9.0 and 8.6; your analysis is right: creating the new pkgconfig command deletes the old one, and its QueryConfigDelete removes the just merged package dict. Proposed fix in branch d9ca7f22fe: keep a reference to the merged dict and put it back after the command is created. RegConfig.3 now says that a second call adds its keys (and replaces the values of existing ones). New test pkgconfig-3.1, …
Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform generic
  • Tested small stubs extension calling Tcl_RegisterConfig twice for the same package (work-a2/rc): after the 2nd call demopkg::pkgconfig list fails with "package not known" on main (own build), 9.0 tip, 8.6 tip.
  • Cause generic/tclConfig.c:Tcl_RegisterConfig: the merged package dict is put into the database first, then Tcl_CreateObjCommand replaces the old ::pkg::pkgconfig, whose QueryConfigDelete removes that package entry (and frees the merged dict).
  • Prototype prototypes/67769bcad3.patch (58 lines incl. new testregisterconfig command in tclTest.c, test pkgconfig-3.1; the fix itself is 5 lines: hold a reference to the dict and re-put it after the command is created). Test fails without the fix, config.test passes with it; valgrind clean. Note: values are stored as bytes and decoded with the encoding of the latest call, so merging calls with different encodings would decode old values with the new encoding.
  • Suggested ticket comment Confirmed in 8.6, 9.0 and main: replacing the query command runs QueryConfigDelete, which removes the just-merged package dict. Proposed fix: keep a reference and put the dict back after Tcl_CreateObjCommand, with a test (pkgconfig-3.1 via a new testregisterconfig command).
⬜8014353dcb2026append env(varname) overwrites existing valueREPRODUCED (main, 9.0, 8.6): append env(username) XYZ turns USERNAME=serhiy into USERNAME=XYZ

Open ticket 8014353dcb · 08. Environment Variables · filed 2026-04-26

Full verdict

REPRODUCED (main, 9.0, 8.6): append env(username) XYZ turns USERNAME=serhiy into USERNAME=XYZ

  • Platform Windows (generic root cause)
  • Tested Unix main/9.0/8.6. On Unix env names are case-sensitive, so the env symptom cannot occur; but a generic test with an array read trace that creates the element shows append never fires read traces (result "new", not "OLDnew"); top-level lappend does ("OLD new") while compiled lappend in a proc does not.
  • Cause append (TclAppendObjCmd and INST_APPEND_*) calls TclPtrSetVarIdx without TCL_TRACE_READS, so the env read trace (EnvTraceProc -> TclGetEnv, case-insensitive on Windows) is not run for env(username); the element is created fresh and the write trace sets USERNAME (case-insensitive) to just the appended text. Same family as 3057639fff.
  • Prototype none (making append fire read traces is a behaviour change across all traced variables; needs a decision. A narrower env-only fix would be to canonicalise env element names on Windows in the array).
  • Suggested ticket comment Cause is generic: &#91;append&#93; does not invoke read traces, so the env read trace never fetches "username" from the case-insensitive Windows environment and the write trace replaces USERNAME. Same root as &#91;3057639fff&#93;.
Windows VM, run 1
  • Platform Windows (generic root cause)
  • Verdict REPRODUCED (main, 9.0, 8.6): append env(username) XYZ turns USERNAME=serhiy into USERNAME=XYZ
  • Tested VM 9.1b1, 9.0.5, 8.6.18
  • Cause as in the Linux analysis: append does not fire read traces, so the case-insensitive env lookup never runs and the write trace replaces USERNAME
  • Prototype none (needs a decision, see Linux analysis)
  • Suggested ticket comment Confirmed on Windows with 8.6, 9.0 and 9.1. Cause: &#91;append&#93; does not fire read traces, so env(username) is not fetched from the case-insensitive environment before the write. Same root as &#91;3057639fff&#93;.
⬜b108e00ac82026Defining const with a different value should raise an errorpatchREPRODUCED (main, 9.0; 8.6 has no const). The behaviour is what TIP 677 and const.n specify, so this is really a decision to change the spec.

Open ticket b108e00ac8 · - New Builtin Commands · filed 2026-01-05

🧪 Prototype check (2026-10-24)

b108e00ac8.patch: VERIFIED (doc part was missing from the saved patch, now added; still a TIP 677 behaviour change needing a decision)

  • Base main (tip d6dd6d0f71)
  • Before work-a4/v4const.tcl: namespace eval ::c1 {const v 2; const v 3; set v} -> 0 2; compiled (apply) the same, no error
  • Patch applied cleanly, then completed: (1) the triage section says doc/const.n is updated but the saved patch did not contain it; added: "it is an error unless that variable is marked as a constant and value is equal (as a string) to its current value, in which case const is a no-op" (the loop example in const.n uses a fixed value and stays valid); (2) TclConstValueDiffers had been inserted between Tcl_ConstObjCmd's header comment and its body; moved above that header. 65 changed lines now.
  • After -> can't make constant "v": constant has a different value for both uncompiled and compiled paths, errorCode TCL LOOKUP CONST; same value (const v [expr {1+1}]) -> 0 2; make test TESTFLAGS="-file 'var.test info.test oo.test compile.test namespace.test upvar.test proc.test'" -> 1493 total, 1456 passed, 37 skipped, 0 failed; const.n renders with 0 groff warnings. Library const uses (idna.tcl, autoexecok.tcl) are per proc call with fixed or per-call values.
  • Test existing var-26.10.1 / var-27.10.1 (rewritten) fail before (var.test unpatched: 217 passed, 2 failed) / pass after; var-26.10.3 / 27.10.3 (same value) pass both
Full verdict

REPRODUCED (main, 9.0; 8.6 has no const). The behaviour is what TIP 677 and const.n specify, so this is really a decision to change the spec.

  • Platform generic
  • Tested main, 9.0.5. const v 2; const v 3 returns quietly and v stays 2. A loop with const c $x returns 1 1. The current behaviour is pinned by tests var-26.10.1/27.10.1 ("unmodifiable by const but not an error"). Four more tests (var-26.16, 27.16, 30.1, 30.2) redefine a constant with another value along the way.
  • Cause generic/tclVar.c Tcl_ConstObjCmd and the INST_CONST_IMM/STK code in generic/tclExecute.c return OK on an existing constant without looking at the value.
  • Prototype prototypes/b108e00ac8.patch (62 changed lines). New TclConstValueDiffers() (string compare with TclStringCmp); both paths raise can't make constant "X": constant has a different value (errorcode TCL LOOKUP CONST). const.n is updated. var-26.10.1/27.10.1 now expect the error, new var-26.10.3/27.10.3 check that the same value is still a no-op, and the 4 incidental tests use the same value. var/info/oo/compile/namespace/upvar/proc tests pass. The const uses in the library (idna.tcl, autoexecok.tcl) are per call and not affected. Changes TIP 677 behaviour, so it needs dkf's or a TCT decision (the loop question).
  • Suggested ticket comment Prototype: redefining an existing constant with a different string value raises "constant has a different value" (both compiled and uncompiled const); the same value stays a no-op. This changes TIP 677 behaviour and tests var-26.10.1/27.10.1, so it needs a decision first.
⬜bc7925fc502025Inconsistencies in long path handling in Tcl 8.6 and 9REPRODUCED (main, 9.0, 8.6; LongPathsEnabled=0, the Windows default). The LongPathsEnabled=1 case was not tested (HKLM system setting, not allowed …

Open ticket bc7925fc50 · - New Builtin Commands · filed 2025-12-18

Full verdict

REPRODUCED (main, 9.0, 8.6; LongPathsEnabled=0, the Windows default). The LongPathsEnabled=1 case was not tested (HKLM system setting, not allowed on this VM).

  • Platform Windows
  • Tested code reading. Main has TIP 744 long-path support from apn-win-longpaths (merged; 9.1 only). With the LongPathsEnabled policy on, plain Win32 calls handle long paths. With it off, win/tclWinFile.c TclNativeCreateNativeRep still adds \\?\ (or \\?\UNC) to paths longer than MAX_PATH (!TclpLongPathSupported()), so the ../glob/cd differences from the ticket remain, as the reporter says in his own comment. Unix not affected. Repro: win/bc7925fc50.tcl (run with LongPathsEnabled 0 and 1, on 9.0 and main).
  • Cause win/tclWinFile.c TclNativeCreateNativeRep: the \\?\ prefix turns off Win32 path processing (.., /), and code paths that do not go through it (cd via SetCurrentDirectory, glob) do not get it.
  • Prototype none (design question: drop the \\?\ fix-up and fail with "file name too long", or apply it everywhere. Not small.)
  • Suggested ticket comment Main (TIP 744) is consistent when LongPathsEnabled is on. Without the policy, TclNativeCreateNativeRep still adds the \\?\ prefix, so the inconsistencies remain there and in 9.0/8.6. Needs a decision whether to drop the prefixing and report "file name too long".
Windows VM, run 3
  • Platform Windows
  • Verdict REPRODUCED (main, 9.0, 8.6; LongPathsEnabled=0, the Windows default). The LongPathsEnabled=1 case was not tested (HKLM system setting, not allowed on this VM).
  • Tested VM win/bc7925fc50.tcl, path of 1054 chars: file mkdir and writing a file inside work, but file isdirectory/exists $long/.. = 0, glob $long/* finds nothing (8.6: "no such file or directory"), cd $long → "file name too long", on 9.1b1 (manifest is longPathAware, TIP 744 merged), 9.0.5 and 8.6.18. Extra finding (work-win3/lpdel2.tcl): file delete -force $base on that tree returns success but deletes nothing (all 5 levels and the leaf remain) on all three versions; deleting the leaf file directly works.
  • Cause win/tclWinFile.c TclNativeCreateNativeRep adds \\?\ to paths over MAX_PATH, which disables .. processing; cd (SetCurrentDirectory) and glob/recursive delete (win/tclWinFCmd.c TraverseWinTree builds its own native paths without the prefix) fail. The silent delete: the traversal fails with ENOENT and generic/tclFCmd.c FileDeleteCmd (the "race condition" bypass after TclpObjRemoveDirectory) treats ENOENT as "already gone".
  • Prototype none (design decision on the \\?\ fix-up, as in the Linux analysis; the silent delete could be fixed separately by checking existence again before bypassing ENOENT)
  • Suggested ticket comment Confirmed on Windows 11 (LongPathsEnabled=0) with 8.6.18, 9.0.5 and 9.1: mkdir works, but .., glob and cd fail on the same long path. Worse, &#91;file delete -force&#93; of such a tree returns OK and removes nothing: the traversal's ENOENT is taken as "already deleted" in FileDeleteCmd.
⬜844f778a6c2025(How) does TimerExitProc remove idle event sourcespatchREPRODUCED (main; same code in 9.0, 8.6): idle handlers still pending at thread exit/finalization are leaked

Open ticket 844f778a6c · 03. Timer Events · filed 2025-11-15

🧪 Prototype check (2026-10-24)

844f778a6c.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before -DPURIFY build (work-a6/purify), extension il/idleleak.c (Tcl_DoWhenIdle 100 times, then exit), TCL_FINALIZE_ON_EXIT=1 valgrind --leak-check=full tclsh il.tcl -> 3,200 (32 direct, 3,168 indirect) bytes in 1 blocks are definitely lost ... by Tcl_DoWhenIdle (tclTimer.c:871); total definitely 128 B/3 blocks, indirectly 3,541 B/106 blocks
  • Patch applied cleanly
  • After the Tcl_DoWhenIdle record is gone; definitely 96 B/2 blocks, indirectly 373 B/7 blocks (the two remaining records are present before too, unrelated); tests timer (136/0 failed; one timing failure of timer-15.8 in the first run did not recur in 3 patched reruns of timer-15.*; flaky timing test, unrelated to the change), event, thread, unload: 0 failed
  • Test none possible (leak; needs a C extension and valgrind on a PURIFY build)
Full verdict

REPRODUCED (main; same code in 9.0, 8.6): idle handlers still pending at thread exit/finalization are leaked

  • Platform generic
  • Tested main built with -DPURIFY under valgrind, TCL_FINALIZE_ON_EXIT=1, with a small extension (work-a4/s/idleleak.c) that calls Tcl_DoWhenIdle 100 times and exits: 3,200 bytes in 100 blocks definitely/indirectly lost from Tcl_DoWhenIdle; none with the prototype. Pure Tcl scripts do not leak, because after idle and bgerror handlers are cancelled when the interp is deleted. Only C code that leaves idle handlers pending is affected.
  • Cause generic/tclTimer.c:TimerExitProc frees both timer queues but never touches tsdPtr->idleList; the header comment claims it removes idle "event sources" (there is no idle event source; idle handlers are run directly by TclServiceIdle).
  • Prototype prototypes/844f778a6c.patch (51 lines, no test: a leak test needs a mem-debug build). It frees idleList, resets lastIdlePtr and corrects the comment. timer/event/thread tests pass.
  • Suggested ticket comment Harald is right: TimerExitProc frees only the timer handlers, so idle handlers that C code left pending at thread exit leak (valgrind with a -DPURIFY build). Proposed fix: free tsdPtr->idleList in TimerExitProc and correct the comment.
⬜aa9f3ef8202025Inconsistent behavior of "file exists" commandpatchREPRODUCED (main, 9.0, 8.6)

Open ticket aa9f3ef820 · 37. File System · filed 2025-11-12

🧪 Prototype check (2026-10-24)

aa9f3ef820.patch: VERIFIED (real or busy COM ports still untested)

  • Base main (tip c092dae4af)
  • Before tclsh91 aa9b.tcl -> aux/prn/lpt1/com1 exists=1 although file type fails; \\.\nul //./nul \\.\con \\.\CONIN$ exists=0 although type=characterSpecial
  • Patch applied cleanly, 86 changed lines, not changed
  • After aux/prn/lpt1/com1 exists=0, the \\.\ devices exists=1 readable=1, \\.\pipe\nosuch 0; aa9f3ef820.tcl VERDICT: NOT REPRODUCED (no COM ports on the VM). File tests admin 8 failed, low 6 failed, as baseline
  • Test existing winFile-5.1 (0 0 0 0) and 5.2 (com1 com9 lpt1 lpt9) FAIL with win/tclWinFile.c reverted, pass with the patch
Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform Windows
  • Tested code inspection; Unix has no equivalent (no reserved device names)
  • Cause win/tclWinFile.c NativeAccess: file exists (F_OK) only calls GetFileAttributesW. For bare reserved names comN (N=1..9, recognised by WinIsReserved) Windows reports attributes whether or not the device exists; for \\.\comN device paths GetFileAttributesW fails, so exists/readable/writable return 0. file type goes through NativeStat, which opens the device with CreateFileW and therefore reflects reality.
  • Prototype none (needs Windows; a fix would be to fall back to a CreateFileW(OPEN_EXISTING, 0 access) probe in NativeAccess for device paths / reserved names, like NativeStat does)
  • Suggested ticket comment NativeAccess() in win/tclWinFile.c relies on GetFileAttributesW(), which is meaningless for devices: it succeeds for any reserved name com1..com9 and fails for \\.\comN. NativeStat() opens the device instead, hence the difference with &#91;file type&#93;.
Windows VM, run 4
  • Platform Windows
  • Verdict REPRODUCED (main, 9.0, 8.6)
  • Tested VM (no serial or parallel ports) win/aa9f3ef820.tcl and work-win4/aa9b.tcl on 9.1b1, 9.0.5, 8.6.18: file exists is 1 for com1–com9, lpt1, aux, prn although file type and open fail; 0 for \\.\nul, //./nul, \\.\con, \\.\CONIN$ although file type gives characterSpecial. A real COM port (busy or ≥ COM10) could not be tested.
  • Cause win/tclWinFile.c NativeAccess uses GetFileAttributesW, which returns FILE_ATTRIBUTE_ARCHIVE for any reserved device name and fails for \\.\ paths; NativeStat opens the device with CreateFileW instead
  • Prototype prototypes/aa9f3ef820.patch (main, 86 changed lines incl. tests; 69 in win/tclWinFile.c with comments; tests winFile-5.1, winFile-5.2; applies to 9.0 and 8.6 with offsets): for \\.\ paths (except \\.\pipe\, to avoid connecting to a pipe server) and bare reserved names (WinIsReserved), NativeAccess opens the device with CreateFileW (GENERIC_READ for F_OK, like NativeStat), counting ACCESS_DENIED/SHARING_VIOLATION as existing for F_OK, and reports devices as not executable. Built on the VM: exists now agrees with type for all names above; winFile.test 6 passed + 5 failed (winFile-4.0–4.4 cacls tests, fail identically unpatched), fCmd.test, cmdAH.test, fileSystem.test 0 failed, winFCmd.test 3 failed as unpatched (6.1, 6.13, 9.3). Tested on Windows.
  • Suggested ticket comment NativeAccess() in win/tclWinFile.c relies on GetFileAttributesW(), which "finds" every reserved name com1..com9/lpt1/aux/prn and fails for \\.\ device paths, while &#91;file type&#93; opens the device. Proposed fix: open devices in NativeAccess() as NativeStat() does, tests winFile-5.1/5.2 (tested on Windows 11 without real COM ports; a check with an existing and a busy COM port would be welcome).
⬜5d5b1052282025execution trace on nested switch - inconsistent line depending on optionspatchREPRODUCED (main, 9.0, 8.6)

Open ticket 5d5b105228 · 18. Commands M-Z · filed 2025-11-10

🧪 Prototype check (2026-10-24)

5d5b105228.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before attached switch_test.tcl (fossil artifact 9c39f3a13b) on the tip -> 10 ok, several !! ... type eval line 4 cmd {return aa} (expected line 2) for the 3/4/5-arg cases
  • Patch applied with offset +2 (re-saved, 46 changed lines, content unchanged)
  • After attached script -> all lines ok (line 2); tests switch, info, trace, compile: 894 total, 876 passed, 18 skipped, 0 failed
  • Test existing switch-20.1 fails before (C fix stashed: 3 3 3 -7 -4) / passes after (2 2 2 -7 -4)
Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform generic
  • Tested the attached switch_test.tcl on main/9.0.5/8.6.18 (the 3/4/5-arg cases give wrong lines). Smaller repro without traces: a proc whose body is not a literal, with both switches called as $sw so neither is compiled inline. info frame 0 in the inner arms gives 3 3 3 2 instead of 2 2 2 2.
  • Cause generic/tclCmdMZ.c Tcl_SwitchObjCmd. When the pattern/body list is split and no source lines are known, it fills ctxPtr->line with -1 for the list elements, but the frame keeps type TCL_LOCATION_BC (the outer arm body was compiled with relative counting). TclInitCompileEnv then calls TclGetSrcInfoForPc again on that BC invoker, which puts back the word lines of the switch command itself. The list index j of the body is then looked up in the switch's argument words. Bodies with j < number of switch words get a wrong line, so the number of bad arms grows with the number of options, as reported.
  • Prototype prototypes/5d5b105228.patch (46 changed lines, test switch-20.1). In that branch it changes the frame type from BC to TCL_LOCATION_EVAL, so the compiler counts lines relatively. The test also checks that absolute lines in a sourced literal body do not change. switch/info tests pass; the attached script then prints all "ok". Applies to 9.0 and 8.6 with fuzz.
  • Suggested ticket comment Cause: when switch splits its body list in a BC frame without source info, TclInitCompileEnv fetches the switch command's own word lines again (TclGetSrcInfoForPc) and indexes them with the list position of the arm. Proposed fix: mark the frame as EVAL in that case; test switch-20.1.
⬜1d061800062025channel reads prematurely from stdin on WindowsREPRODUCED (main, 8.6; console stdin only)

Open ticket 1d06180006 · 01. Notifier · filed 2025-11-08

Full verdict

REPRODUCED (main, 8.6; console stdin only)

  • Platform Windows
  • Tested win/1d06180006.tcl on Linux main, 9.0.5, 8.6.18: chan pending input stdin is 0 inside the readable handler (pipe stdin). The script also has an interactive console mode for the Windows console case the ticket is about.
  • Cause win/tclWinConsole.c reader thread reads console input into the channel buffer to make the channel readable (Ashok's analysis in the ticket). Ashok's experimental branch apn-winconsole-noreadahead (last check-in 0341d9eef3, 2025-12-29, merge from trunk) is not merged and has no descendants. The reporter concluded it cannot work reliably in line mode, and Ashok would merge only the Alt-key filtering part.
  • Prototype none (Windows console code; branch already exists)
  • Suggested ticket comment Reporter and Ashok agree the read-ahead cannot be avoided in console line mode, and the experimental branch apn-winconsole-noreadahead is unmerged. Propose closing as Won't Fix, unless the Alt/Esc wake-up filtering from e70736d496 should be merged on its own.
Windows VM, run 2
  • Platform Windows (console stdin)
  • Verdict REPRODUCED (main, 8.6; console stdin only)
  • Tested VM, the ticket's C program (rd.c in work-win2/rd: Tcl_CreateChannelHandler on stdin, raw _read(0) in the callback) built with MSVC against 9.1b1 and 8.6.18, run in a real console (ConPTY via ssh -tt) with typed lines hello, world, q: the first callback's _read blocks and returns world — hello was consumed by Tcl's console reader thread and is lost to the raw read. Same on 9.1b1 and 8.6.18. With stdin as a pipe there is no read-ahead (win/1d06180006.tcl pipe mode OK on 9.1b1, 9.0.5, 8.6.18); chan pending input stdin is 0 in the handler even for the console, because the data sits in the console driver's own buffer, not the channel buffer.
  • Cause win/tclWinConsole.c reader thread reads console input ahead to make the channel readable (as Ashok described); branch apn-winconsole-noreadahead (unmerged) not tested here
  • Prototype none (branch exists)
  • Suggested ticket comment Confirmed on Windows 11 with 8.6.18 and 9.1: with console stdin the first line is read by Tcl's console thread before the handler runs, so a raw read() in the handler misses it; pipes are not affected. Ashok's branch apn-winconsole-noreadahead is the candidate fix (unmerged).
✅2ef3430fdc2025Don't assume that host=*mingw32* means cross-compilingpatchREPRODUCED (main): native MSYS2 UCRT64 configure --host=x86_64-w64-mingw32 --build=x86_64-w64-mingw32 gives TCL_EXE = tclsh; with no tclsh on P…

Open ticket 2ef3430fdc · 53. Configuration and Build Tools · filed 2025-10-26

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-10-20 check-in e4fd3a27b4 on win-native-mingw-tclsh, in main, 9.0: Fix 2ef3430fdc: win/configure decides whether to use an installed tclsh from $cross_compiling, not from host=*mingw32*, so native MSYS2/Cygwin builds use the
  • 🔧 2026-10-20 check-in c87b92e2ee on trunk, in main: Merge win-distclean-tcl-pc and win-native-mingw-tclsh: fix 59a9b957c6 and 2ef3430fdc.
  • 🔧 2026-10-20 check-in 063874bae3 on core-9-0-branch, in main, 9.0: Fix 59a9b957c6 and 2ef3430fdc: Windows "make distclean" removes tcl.pc; win/configure uses an installed tclsh only when cross-compiling, not for any *mingw3
  • 🔧 2026-10-20 check-in 161779a5c4 on core-8-6-branch, in main, 9.0, 8.6: Fix 59a9b957c6 and 2ef3430fdc: Windows "make distclean" removes tcl.pc; win/configure uses an installed tclsh only when cross-compiling, not for any *mingw3
  • 💬 2026-10-20 serhiy.storchaka: Confirmed with a native MSYS2 UCRT64 build: with no tclsh on PATH, make install fails at install-tzdata with "tclsh: command not found". win/configure.ac treated any *mingw32* host as a cross build. Proposed fix in branch e4fd3a27b4: use autoconf's $cross_compiling instead, which is what the anonymous comment asked for. Tested natively in MSYS2 (configure says "cross compiling... no", install works) and with a L…
  • 💬 2026-10-20 jan.nijtmans: I like this solution! +1 to merge
Full verdict

REPRODUCED (main): native MSYS2 UCRT64 configure --host=x86_64-w64-mingw32 --build=x86_64-w64-mingw32 gives TCL_EXE = tclsh; with no tclsh on PATH, make succeeds but make install fails at install-tzdata (tclsh: command not found, Error 127)

  • Platform build (Windows MSYS2/Cygwin)
  • Tested win/configure.ac on main/9.0 and win/configure.in on 8.6 have the case ${host_alias} in *mingw32*) TCL_EXE=tclsh block. Prototype checked on Linux with mingw-w64 cross gcc: win/configure --host=x86_64-w64-mingw32 still gives TCL_EXE = tclsh (autoconf reports cross compiling = yes). Native MSYS2 case not testable here; repro script win/2ef3430fdc.tcl (needs MSYS2 on the VM).
  • Cause win/configure.ac decides "cross build" from host_alias instead of autoconf's $cross_compiling.
  • Prototype prototypes/2ef3430fdc.patch (configure.ac 8 changed lines + regenerated win/configure with autoconf 2.73): use if test "$cross_compiling" = yes. Side effect: a Linux cross build whose binfmt runs .exe via Wine gets cross_compiling=no and would run the built tclsh through Wine.
  • Suggested ticket comment Proposed fix: base TCL_EXE on autoconf's $cross_compiling rather than on host_alias, so native MSYS2/Cygwin builds use the freshly built tclsh while real cross builds keep using an installed tclsh. Checked for the Linux cross build; needs confirmation in MSYS2.
Windows VM, run 1
  • Platform build (Windows MSYS2)
  • Verdict REPRODUCED (main): native MSYS2 UCRT64 configure --host=x86_64-w64-mingw32 --build=x86_64-w64-mingw32 gives TCL_EXE = tclsh; with no tclsh on PATH, make succeeds but make install fails at install-tzdata (tclsh: command not found, Error 127)
  • Tested VM tcl-wt91, MSYS2 UCRT64 gcc (no mingw32/mingw64 toolchains installed), full build + install to a temp prefix, unpatched and patched
  • Cause win/configure.ac decides cross builds from host_alias instead of $cross_compiling
  • Prototype prototypes/2ef3430fdc.patch (unchanged; applies to the VM's main): configure reports "cross compiling... no", TCL_EXE uses the built tclsh, make install succeeds (tzdata installed). Tested on Windows.
  • Suggested ticket comment Confirmed with MSYS2 UCRT64: make install fails with "tclsh: command not found" when no tclsh is installed. Proposed fix: base TCL_EXE on $cross_compiling; tested natively in MSYS2 (install works) and for a Linux cross build.
✅ac3ea873052025Don't assume that Windows has no usable zlib/tommathREPRODUCED (main, 9.0; 8.6 for zlib) - effectively an RFE

Open ticket ac3ea87305 · 53. Configuration and Build Tools · filed 2025-10-26

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-10-20 check-in 61fe2461e3 on win-system-zlib-tommath, in main: Fix ac3ea87305: win/configure gets --with-system-zlib and --with-system-libtommath to link the zlib and libtommath of the build system (e.g. MSYS2) instead of
  • 🔧 2026-10-20 check-in be527d2e2d on trunk, in main, 9.0: Fix ac3ea87305: Don't assume that Windows has no usable zlib/tommath
  • 🔧 2026-10-20 check-in 809ce55fbc on core-9-0-branch, in main, 9.0, 8.6: Fix ac3ea87305: Don't assume that Windows has no usable zlib/tommath
  • 🔧 2026-10-27 check-in 0890c290ed on core-8-6-branch, in main, 9.0, 8.6: Fix ac3ea87305 and 17ffc49790. New option --with-system-zlib and fix out-of-tree build with "configure --srcdir=..."
  • 💬 2026-10-20 serhiy.storchaka: Proposed in branch 61fe2461e3: win/configure gets two options, off by default, so the standard build is unchanged: [code] With them the bundled zlib1.dll / libtommath.dll and their import libraries are neither copied nor installed (for static builds the bundled objects are not compiled in), and configure fails if the library is not found. Tested natively with MSYS2 UCRT64 and its zlib and libtommath packages: tc…
  • 💬 2026-10-20 jan.nijtmans: I don't see anything wrong with this, so go ahead. +1
  • 💬 2026-10-20 serhiy.storchaka: Merged to be527d2e2d and 809ce55fbc.
  • 💬 2026-10-20 jan.nijtmans: Please don't forget the changes.md/changes entries
Full verdict

REPRODUCED (main, 9.0; 8.6 for zlib) - effectively an RFE

  • Platform build (Windows MinGW/MSYS2)
  • Tested code reading. win/configure.ac always links the bundled zlib/libtommath import libs (compat/zlib/win*/, libtommath/win*/) and win/Makefile.in always copies the bundled zlib1.dll / libtommath.dll into the build and install; there is no equivalent of unix/configure.ac's zlib.h/deflateSetHeader probe or --with-system-libtommath.
  • Cause win/configure.ac lines 110-161 hard-code the bundled libraries.
  • Prototype none (needs a new configure option plus Makefile.in changes for DLL copy/install, likely > 60 lines; design choice for the maintainers). The reporter now uses a packaging workaround (drops the bundled DLLs), which Jan approved as drop-in compatible.
  • Suggested ticket comment Still as described: win/configure has no switch for system zlib/libtommath. Porting the unix probes (--with-system-libtommath, zlib.h check) to win/configure.ac and making the DLL copy rules conditional would resolve it; meanwhile dropping the bundled DLLs is the endorsed workaround.
⬜ce32712f832025Inconsistent locking during zipfs initializationpatchREPRODUCED (main, 9.0) by code inspection — main: dead code only; 9.0: theoretical race

Open ticket ce32712f83 · - New Builtin Commands · filed 2025-10-03

🧪 Prototype check (2026-10-24)

ce32712f83.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before code reading on the tip: TclZipfs_Mount (line 2447) and TclZipfs_MountBuffer (2562, with the TODO - how come a *read* lock suffices for initialzing ?) call ZipfsSetup() under ReadLock(); the other callers (TclZipfsMountExe 4323, TclZipfsMountShlib 4376, TclZipfsInit 6522) use WriteLock(). On main TclInitSubsystems already runs TclZipfsInit() first, so the branch is dead; a runtime race is not reproducible.
  • Patch applied cleanly (offset 6; 9 changed lines): both functions call TclZipfsInit() (WriteLock + check) before ReadLock(). (The triage line says 38 lines; the patch has 9.)
  • After builds without warnings; make test TESTFLAGS="-file zipfs.test" -> 535 total, 498 passed, 37 skipped, 0 failed
  • Test none possible (initialisation race; on main the path is never reached after TclInitSubsystems)
Full verdict

REPRODUCED (main, 9.0) by code inspection — main: dead code only; 9.0: theoretical race

  • Platform generic
  • Tested code inspection of main and 9.0 tclZipfs.c; zipfs.test on main with prototype (533 tests, 0 failures)
  • Cause generic/tclZipfs.c TclZipfs_Mount and TclZipfs_MountBuffer call ZipfsSetup() under ReadLock(), so two threads could both run it (double Tcl_FSRegister / hash init). In practice: on main, TIP 732 (bc851af66a, apnadkarni, 2025-12-22) made TclInitSubsystems call TclZipfsInit() under WriteLock, so the read-lock branch is never taken; on 9.0 every Tcl_CreateInterp runs TclZipfs_Init (WriteLock) first, so a race needs C-level mounts from two threads before any interp exists.
  • Prototype prototypes/ce32712f83.patch (38 lines, no test — race not testable): in both functions call TclZipfsInit() (WriteLock + double check) before taking ReadLock(), removing the TODO. Applies to 9.0 as well.
  • Suggested ticket comment On main TclZipfsInit() is now called from TclInitSubsystems (TIP 732), so the ReadLock()+ZipfsSetup() branches in TclZipfs_Mount/MountBuffer are dead code. Proposed cleanup: call TclZipfsInit() before ReadLock() there (patch available).
✅17ffc497902025Building Tcl 9.0.2 release fails when using "configure --srcdir=<path> ..."patchREPRODUCED (main; same code in 9.0 and 8.6)

Open ticket 17ffc49790 · 53. Configuration and Build Tools · filed 2025-09-25

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-10-20 check-in 0e472d58ee on configure-pkgs-no-srcdir, in main: Fix 17ffc49790: an out-of-tree build with "configure --srcdir=..." failed to configure the bundled packages, since Tcl passed its own --srcdir on to their con
  • 🔧 2026-10-27 check-in 9ac77da0c6 on trunk, in main, 9.0: Fix 17ffc49790: Make packages work in an out-of-tree build with "configure --srcdir=..."
  • 🔧 2026-10-27 check-in 2e8ab95738 on core-9-0-branch, in main, 9.0, 8.6: Fix 17ffc49790: Make packages work in an out-of-tree build with "configure --srcdir=..."
  • 🔧 2026-10-27 check-in 0890c290ed on core-8-6-branch, in main, 9.0, 8.6: Fix ac3ea87305 and 17ffc49790. New option --with-system-zlib and fix out-of-tree build with "configure --srcdir=..."
  • 💬 2026-10-20 serhiy.storchaka: Reproduced with an out-of-tree build of the current main with Thread and Itcl in pkgs/: .../unix/configure --srcdir=.../unix, then make fails in configure-packages with "cannot find sources () in .../unix". The relative --with-tcl=../.. is not the problem: Tcl passes its own --srcdir on to each package's configure (in PKG_CFG_ARGS), so the package looks for its sources in Tcl's unix directory. The packages do not …
  • 💬 2026-10-27 jan.nijtmans: Fixed in 9ac77da0c6, core-9-0-branch and core-8-6-branch Many thanks!
Full verdict

REPRODUCED (main; same code in 9.0 and 8.6)

  • Platform build (unix)
  • Tested own main checkout with Thread trunk unpacked into pkgs/ (from thread.fossil), out-of-tree build. ../tcl9.1/unix/configure --srcdir=../tcl9.1/unix: make fails in configure-packages with "configure: error: cannot find sources () in ../tcl9.1/unix". Without --srcdir it builds. (The reporter's --srcdir=../tcl9.0.2, i.e. the top dir rather than unix/, is a user error: config.status cannot find ../unix/Makefile.in.)
  • Cause unix/configure.ac: PKG_CFG_ARGS="$ac_configure_args ..." forwards Tcl's own --srcdir to each package's configure, so it looks for its sources in Tcl's unix dir. The --with-tcl=../.. relative path is not the problem; the package configure is run via the absolute $(TCL_SRC_DIR)/pkgs/<pkg>/configure.
  • Prototype prototypes/17ffc49790.patch (configure.ac 4 lines + regenerated configure): strip --srcdir=X / --srcdir X from PKG_CFG_ARGS. Verified: both forms then build pkgs/thread. (make install of my Thread trunk tarball then fails for an unrelated reason: the fossil tarball lacks tclconfig/install-sh.)
  • Suggested ticket comment The failure comes from Tcl's own --srcdir being forwarded to the bundled packages' configure via PKG_CFG_ARGS. Proposed fix: drop --srcdir from PKG_CFG_ARGS; the packages find their sources from the absolute path of their configure script. Note --srcdir must point to the unix directory, not the top of the source tree.
⬜475db438432025float cast overflows in lseqpatchREPRODUCED (main, 9.0 by reading the same code; 8.6 has no lseq)

Open ticket 475db43843 · 17. Commands I-L · filed 2025-09-23

🧪 Prototype check (2026-10-24)

475db43843.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before gcc -fsanitize=float-cast-overflow build (work-a6/ubsan), tcltest r475.tcl (lseq 1e50 count 1 1; compiled lseq 0 count -Inf) -> tclArithSeries.c:653:6 and :785:2: runtime error: 1e+50 is outside the range ... long long int, tclExecute.c:9407:6: runtime error: -inf is outside the range ...
  • Patch applied cleanly
  • After same UBSan build: no runtime errors, same results (1e+50, expected integer but got "-Inf"); lseq.test 147/0 failed with no UBSan reports; normal build lseq, lrange, abstractlist: 0 failed
  • Test existing lseq-bug-475db43843-0/1 trigger all three UBSan reports before and none after; on a normal build they pass both ways (UB has no visible effect on x86-64), so no non-sanitizer failure is possible
Full verdict

REPRODUCED (main, 9.0 by reading the same code; 8.6 has no lseq)

  • Platform generic
  • Tested out-of-tree main build with -fsanitize=float-cast-overflow (work-a6/ubsan). lseq 1e50 count 1 1 reports UB at tclArithSeries.c ExtractNumber ((Tcl_WideInt)value) and TclNewArithSeriesObj (end = dend). The compiled lseq 0 count -Inf adds a third one, tclExecute.c GenerateArithSeries ((Tcl_WideInt) dCount before any range check). Results are right on x86-64, but this is undefined behaviour.
  • Cause casting doubles outside the Tcl_WideInt range in generic/tclArithSeries.c (ExtractNumber, 2 places; TclNewArithSeriesObj end = dend) and generic/tclExecute.c GenerateArithSeries. ArithSeriesLenDbl's guard end <= (double)WIDE_MAX also lets 2^63 through.
  • Prototype prototypes/475db43843.patch (42 changed lines, tests lseq-bug-475db43843-0/1). New saturating DoubleToWide() helper in tclArithSeries.c; GenerateArithSeries checks the range before the cast (as the uncompiled lseq in tclCmdIL.c already does); the guard becomes <. lseq.test passes with no UBSan reports; without the patch the new tests trigger all three reports.
  • Suggested ticket comment Confirmed with -fsanitize=float-cast-overflow: the doubles are cast to Tcl_WideInt without a range check in ExtractNumber, TclNewArithSeriesObj and the compiled GenerateArithSeries (count -Inf). Proposed fix: saturating conversion plus a range check first; tests lseq-bug-475db43843-*.
🟢3764f4e81f2025http11 test suite is unreliable under loadpatchREPRODUCED (main, 9.0; 8.6 partly, see below)

Open ticket 3764f4e81f · 69. Other · filed 2025-09-05

🟢 Since triage (updated 2026-11-11)

Status now: Open

  • 🔧 2026-10-16 check-in 00a815eb5e on http-keepalive-reuse-closed, in main: Fix 05b158ed56 and 3764f4e81f: http::OpenSocket opens a new connection if the persistent socket it was going to reuse has been closed in the meantime (e.g.
  • 🔧 2026-11-10 check-in 98a1b78011 on trunk, in main, 9.0: Fix 05b158ed56 and 3764f4e81f: http::OpenSocket opens a new connection if the persistent socket it was going to reuse has been closed in the meantime
  • 🔧 2026-11-10 check-in 655b9df308 on core-9-0-branch, in 9.0: Fix 05b158ed56 and 3764f4e81f: http::OpenSocket opens a new connection if the persistent socket it was going to reuse has been closed in the meantime
  • 💬 2026-10-16 serhiy.storchaka: The failures happen when a new test server gets the same port as an earlier one, so the client reuses a dead persistent socket, and http.tcl does not replay the request: an empty line left after a chunked -channel response is taken as the status line (8.6 too), and in 9.x http::CheckEof can close a socket that a queued request is about to use ("can't read socketWrState(...)", 05b158ed56). Proposed fix in branch …
Full verdict

REPRODUCED (main, 9.0; 8.6 partly, see below)

  • Platform generic
  • Tested main build (c3bf6dd052) running 48 parallel http11.test copies x3: 1/144 runs failed (http11-2.12.x, same "eof {} crc32 mismatch" as the ticket); a deterministic repro (work-a2/h11/repro.tcl: keepalive request, restart httpd11.tcl on the same port, request again with no event loop in between) fails on main and 9.0.5 every time, on 8.6.18 only after a -channel request.
  • Cause each test starts a new httpd on port 0; sometimes the kernel hands out a port the http package still holds a persistent socket for (instrumented runs: the failure matched such a reuse). Reusing that dead socket hits two http.tcl bugs: (1) library/http/http.tcl http::Event: the CRLF left over after a chunked -channel response is read as an empty status line, so state goes to "header", and the eof that follows ends the request as eof without the TestForReplay retry (8.6 too); (2) 9.x only: http::CheckEof can close the socket after geturl has queued the idle-time OpenSocket that reuses it, and ScheduleRequest then throws can't read "socketWrState(localhost:N)": no such element in array.
  • Prototype prototypes/3764f4e81f.patch (67 changed lines, 22 of them code; test http11-5.1.*). With it the repro passes and 144/144 stress runs pass; http.test and http11.test pass.
  • Suggested ticket comment The failures happen when a new test server gets the same port as an earlier one, so the client reuses a dead persistent socket. http.tcl then doesn't replay the request: an empty line left over from a chunked -channel response is taken as the status line, and in 9.x http::CheckEof can close a socket that a queued request is about to use (this throws "can't read socketWrState(...)"). Proposed fix with test http11-5.1 attached for review (no branch yet).
⬜404eaa51e12025tcl::pkgconfig not updated for zipfs installsREPRODUCED (main 9.1b1, 9.0.5)

Open ticket 404eaa51e1 · 54. Configuration Reporting · filed 2025-09-02

Full verdict

REPRODUCED (main 9.1b1, 9.0.5)

  • Platform generic (unix zipfs builds)
  • Tested installed main and 9.0.5 (zipfs builds): scriptdir,runtime / scriptdir,install = <prefix>/lib/tcl9.x, which does not exist (no init.tcl there), while info library is //zipfs:/lib/tcl/tcl_library. 8.6.18 reports the real installed dir.
  • Cause unix/Makefile.in compiles tclPkgConfig.c with CFG_INSTALL_SCRDIR=$(SCRIPT_INSTALL_DIR) / CFG_RUNTIME_SCRDIR=$(TCL_LIBRARY) regardless of whether make install installs the script library (INSTALL_LIBRARIES is empty for zipfs builds).
  • Prototype none (design decision: Jan argues to keep the value because extensions use its dirname to find the module dir; dgp argues not to report locations nothing is installed to; a separate moduledir entry was also proposed).
  • Suggested ticket comment Still the case on 9.0.5 and main: scriptdir,* point to <prefix>/lib/tcl9.x, which a zipfs install does not create. Needs a decision between dropping/changing these keys and adding a moduledir entry first.
⬜1ecea011962025string is lower, upper not consistent with Unicode database and other languagesREPRODUCED (main, 9.0, 8.6) -- behaviour change, not a small bug fix

Open ticket 1ecea01196 · 18. Commands M-Z · filed 2025-07-23

Full verdict

REPRODUCED (main, 9.0, 8.6) -- behaviour change, not a small bug fix

  • Platform generic
  • Tested string is lower ª 0, :lower: on ª 0, string is lower ⅰ 0 while string toupper ⅰ gives Ⅰ, string is lower ⓐ 0, string is upper Ⅰ 0 -- identical on main, 9.0, 8.6.
  • Cause tools/uniParse.tcl / generic/tclUniData.c classify case only by General_Category (Lu/Ll), ignoring Other_Lowercase/Other_Uppercase (derived Lowercase/Uppercase properties).
  • Prototype none (needs regenerated Unicode tables from PropList.txt and a decision; apnadkarni targets 9.1+/possibly a TIP; affects regexp :lower:/:upper: and -nocase too)
  • Existing work none
  • Suggested ticket comment Still the same on main. Needs a decision whether 9.1.x may change it or it waits for 9.2/TIP; no branch yet.
⬜b69e00ecf62025Tcl does not raise an error even when it detects an invalid internal UTF-8 sequenceREPRODUCED (main, 9.0) — design/reminder ticket; 8.6 has no profiles

Open ticket b69e00ecf6 · 11. Conversions from String · filed 2025-07-16

Full verdict

REPRODUCED (main, 9.0) — design/reminder ticket; 8.6 has no profiles

  • Platform generic
  • Tested own main tcltest and 9.0 tip tcltest: encoding convertto utf-8 [testbytestring \xEF\xBF\x7F] returns c3afc2bf7f even with -profile strict; to iso8859-1 it gives efbf7f (bytes passed through as Latin-1); convertfrom errors as expected.
  • Cause invalid bytes in Tcl_Obj.bytes are decoded by TclUtfToUCS4/Tcl_UtfToUniChar as single Latin-1/cp1252 chars before the encoder sees them, so the profile never applies.
  • Prototype none (needs a policy decision on how internal invalid UTF-8 is treated across the core, not a local fix).
  • Suggested ticket comment Still the same in 9.1 and 9.0: convertto with -profile strict accepts invalid internal UTF-8 (turned into Latin-1 chars). Needs a decision on the policy before any fix.
✅b4637c43122025upvar segfault for destroyed objectREPRODUCED (main, 9.0, 8.6)

Open ticket b4637c4312 · 07. Variables · filed 2025-05-08

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-11-01 check-in eed1c88593 on oo-destructor-deleted-namespace-90, in main, 9.0: Backport of Fix b4637c4312: [upvar] to an object variable after the object was destroyed in its method crashed.
  • 🔧 2026-11-01 check-in 298d1d97d5 on core-9-0-branch, in main, 9.0: Fix 28de60cef5 and b4637c4312
  • 🔧 2026-11-01 check-in f3f08274f7 on trunk, in main: Fix 28de60cef5 and b4637c4312
  • 🔧 2026-10-13 check-in de6527681f on oo-destructor-deleted-namespace, in main, 9.0: Fix b4637c4312: [upvar] to an object variable after the object was destroyed in its method crashed. Same cause as 28de60cef5; add a test.
  • 💬 2026-11-01 dkf: Confirming the crash with 9.1.0, and it looks like something that wants a backport fix to 9.0 too. No plans to backport to 8.6, given the support status of that release line.
  • 💬 2026-11-01 dkf: Fix applied to 9.0 onwards.
  • 💬 2026-11-01 dkf: BTW, thanks for hunting this down, Serhiy.
Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform generic
  • Tested reporter's script on main (2026-10-22), 9.0.5, 8.6.18: SIGSEGV on all; branch tip de6527681f: script returns normally, oo.test 390 total, 0 failed
  • Cause resolving the class variable V for [upvar] in a method after [my destroy] looks it up in the deleted object namespace (same cause as 28de60cef5, per Serhiy)
  • Existing work branch oo-destructor-deleted-namespace (tip de6527681f, 2026-10-13), unmerged (not in trunk/9.0/8.6); merges into current trunk without conflicts (dry run)
  • Prototype none (branch exists)
  • Suggested ticket comment none needed (Serhiy's 2026-10-13 comment is current).
⬜03cff7a7772025-Wconversion warnings, like TclAdvanceContinuations()REPRODUCED (main, partly addressed). The example in the ticket is fixed: TclAdvanceContinuations now takes Tcl_Size loc. Three rounds of fixes ([…

Open ticket 03cff7a777 · - New Builtin Commands · filed 2025-04-16

Full verdict

REPRODUCED (main, partly addressed). The example in the ticket is fixed: TclAdvanceContinuations now takes Tcl_Size loc. Three rounds of fixes (ca9b2d5796, d92999333a, 9f799357d8 on the conversion-warning branch, merged into trunk as 932cc0af74 on 2025-04-20) went into 9.1.

  • Platform build
  • Tested main libtcl built with CFLAGS_WARNING="-Wconversion -Wno-sign-conversion" (gcc 15): 361 warnings left. Most are in tclEncoding.c (73), tclDate.c (70, generated parser), tclZipfs.c (59), tclClockFmt.c (29), tclIO.c (23), tclInt.h (20, so they repeat in every file), tclZlib.c (19). The core parse/compile/execute files have 1–5 each. Adding -Wsign-conversion gives far more (e.g. tclCompile.c 77).
  • Cause n/a (ongoing cleanup; the ticket says the goal is not to fix them all)
  • Prototype none (open-ended cleanup, no single bug)
  • Suggested ticket comment The TclAdvanceContinuations case is fixed on main. Main still has about 360 -Wconversion warnings (not counting sign conversion), mostly in tclEncoding.c, tclDate.c, tclZipfs.c and tclClockFmt.c. Close this as done for the cases that matter, or list the remaining ones that are real bugs.
⬜fb9acd64a52025effect of close affects parent interp in specific casepatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

Open ticket fb9acd64a5 · 20. [interp] · filed 2025-04-02

🧪 Prototype check (2026-10-24)

fb9acd64a5.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before [interp create] eval {close stdout}; puts "hello" -> can not find channel named "stdout", exit 1
  • Patch applied with offset +20 in interp.test; test renumbered interp-41.1 -> interp-42.1 because 2114165fff.patch also adds interp-41.1/41.2 (re-saved, 39 changed lines)
  • After same script -> hello, exit 0; tests interp, io, ioCmd, chan, chanio, safe, iogt, ioTrans: 2754 total, 2570 passed, 184 skipped, 0 failed
  • Test existing (renumbered) interp-42.1 fails before (tclInterp.c fix stashed: 1:can not find channel named "stdout") / passes after. Note: its interp.test hunk has the same insertion point as 2114165fff's, so when both are applied the second one needs a trivial manual merge (C hunks apply fine).
Full verdict

REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

  • Platform generic
  • Tested reporter's script on wt-a3 main, 9.0.5, 8.6.18, 8.5.19, 8.4.20 — all fail with can not find channel named "stdout"; with any prior channel op in the parent all work.
  • Cause generic/tclIO.c:GetChannelTable registers stdin/stdout/stderr lazily on an interp's first channel operation. If the child does it first, it holds the only reference, so close stdout there really closes the thread's stdout; the parent later finds no std channel.
  • Prototype prototypes/fb9acd64a5.patch (65 lines, test interp-41.1): new internal TclInitChannelTable(), called by ChildCreate (tclInterp.c) for the parent before the child is created, so the parent always holds its own reference (same as the "puts start" case). interp/io*/chan*/safe* tests pass. Probably not enough for the related 5338d8e82.
  • Suggested ticket comment The parent registers the std channels only on its first channel operation, so a child that closes stdout first drops the last reference. Proposed fix: initialise the parent's channel table when creating a child; test interp-41.1.
⬜0b3595a8192025Invocation of imported nested command is affected by recursionlimit and has O(n) instead of O(1)REPRODUCED (main, 9.0, 8.6)

Open ticket 0b3595a819 · 21. [namespace] · filed 2025-03-29

Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform generic
  • Tested reporter's PoC (time instead of timerate) on wt-a3 main, 9.0.5, 8.6.18: chain of 100 imports works (2–4 us/call), chains of 1000 and 5000 fail with "too many nested evaluations (infinite loop?)".
  • Cause an imported command is a forwarding command (tclNamesp.c InvokeImportedNRCmd) that calls the next ImportedCmdData->realCmdPtr, which can itself be an import, so each call walks the whole chain with one NRE level per link (counted against the recursion limit). TclGetOriginalCommand already resolves the chain, but invocation does not use it.
  • Prototype none (sebres has a rewrite in his fork, to be done together with 62dafd9c53; changing import resolution/ownership is beyond a small fix).
  • Suggested ticket comment Still present in main, 9.0 and 8.6 (chain of 1000 imports gives "too many nested evaluations"). Waiting for the backport of the fork's rewrite.
⬜62dafd9c532025Namespace import can cause quadratic big-O and affects all siblings of same original commandREPRODUCED (main, 9.0, 8.6)

Open ticket 62dafd9c53 · 21. [namespace] · filed 2025-03-27

Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform generic
  • Tested PoC with 1000/10000 namespaces importing ::A::*: deleting a namespace takes 5–7 us at n=1000 and 68–114 us at n=10000 (main, 9.0.5, 8.6.18), so total deletion grows quadratically; import cost stays flat.
  • Cause Command.importRefPtr is a singly linked list of all imports of the real command; deleting an import (tclNamesp.c DeleteImportedCmd) walks the list to unlink itself, O(n) per deletion.
  • Prototype none (sebres' fix replaces importRefPtr by refcounting; a data-structure change, not small. A minimal alternative would be a doubly linked list, but it changes the Command/ImportRef layout in tclInt.h).
  • Suggested ticket comment Still present in main, 9.0, 8.6 (per-deletion cost grows ~linearly with the number of imports). Waiting for the backport of the fork's rewrite.
⬜48150cdad22025Command creation with overwrite (prevention against endless loop by its recreation via trace) is affected by drastic mem-leakREPRODUCED (main, 9.0, 8.6)

Open ticket 48150cdad2 · 46. Traces · filed 2025-03-27

Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform generic
  • Tested PoC loop, RSS growth ~125 bytes/iteration (100000 iterations: +13 MB main, +12.8 MB 9.0.5, +12.7 MB 8.6.18). ASan+PURIFY build of main (work-a3/asan): reports the TraceCommandInfo from TraceCommandObjCmd (tclTrace.c:599) leaked once per iteration; no use-after-free found, also after namespace delete ::A.
  • Cause generic/tclBasic.c:TclCreateObjCommandInNs (same code in Tcl_CreateCommand): if the delete trace re-created the command, the new Command is just Tcl_Free()d, without running/freeing its traces, deleteProc or client data.
  • Prototype none. A proper fix has to decide the semantics (silently delete the re-created command without calling its traces, which still needs a way to free trace client data with TCL_TRACE_DESTROYED and no script; or refuse the creation, which the void-ish Tcl_CreateObjCommand API cannot report). Not a small change.
  • Suggested ticket comment Still leaks in main, 9.0 and 8.6: the command re-created by the delete trace is freed with Tcl_Free only, so its trace data (and client data) leak. Needs a decision on the intended semantics.
⬜ebc39f14d72025OO-command implicit tracingREPRODUCED (main, 9.0, 8.6: object commands still get an internal rename/delete trace, tclOO.c ObjectRenamedTrace); proposed fix branch not merged

Open ticket ebc39f14d7 · 35. TclOO Package · filed 2025-03-25

Full verdict

REPRODUCED (main, 9.0, 8.6: object commands still get an internal rename/delete trace, tclOO.c ObjectRenamedTrace); proposed fix branch not merged

  • Platform generic
  • Tested source check of main generic/tclOO.c (trace installed at ~line 1002); branch br-ebc39f14d7-oo-no-trace has one check-in 495ffc0324 (2025-03-25, sebres, 68 changed lines in tclOO.c), forked from core-8-6-branch da7ca485a4, no descendants, never merged to 8.6/9.0/main. No user-visible failure is described; it is a refactoring/performance RFE plus a theoretical use-after-free.
  • Cause design (TclOO uses a command trace for object-command delete/rename cleanup).
  • Prototype none (branch exists; refactoring, needs dkf review and forward-port to 9.x where tclOO.c differs).
  • Suggested ticket comment Proposed fix in branch br-ebc39f14d7-oo-no-trace (8.6 only, not merged). Needs forward-port to 9.0/main and review by the TclOO maintainer.
⬜38a31dffdf2024file delete -force reports permission deniedpatchREPRODUCED (main, 9.0, 8.6, 8.4)

Open ticket 38a31dffdf · 37. File System · filed 2024-11-30

🧪 Prototype check (2026-10-24)

38a31dffdf.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before fd.tcl (d/models/svg/yoke.svg, models and svg mode 0555; file delete -force d) -> 1 {error deleting "d/models/svg/yoke.svg": permission denied} 1
  • Patch applied with offset -9 in fCmd.test (re-saved, 40 changed lines, content unchanged; the triage line's "fCmd-5.13" is actually fCmd-20.3 in the patch)
  • After same script -> 0 {} 0; tests fCmd, unixFCmd, fileSystem, cmdAH: 17518 total, 17403 passed, 115 skipped, 0 failed (fCmd-20.1, unreadable subdirectory, still passes)
  • Test existing fCmd-20.3 fails before (C fix reverted: error deleting "tfa/a/b/f": permission denied) / passes after
Full verdict

REPRODUCED (main, 9.0, 8.6, 8.4)

  • Platform unix (reported on macOS; same on Linux)
  • Tested Linux: tree d/models/svg/yoke.svg with models and svg mode 0555 → file delete -force d fails with "error deleting .../yoke.svg: permission denied" on main, 9.0.5, 8.6.18, 8.4.20. (rm -rf fails the same way, but file.n promises that with -force "Tcl will attempt both to change permissions ... if that is necessary to allow the deletion".)
  • Cause unix/tclUnixFCmd.c: DoRemoveDirectory() does chmod u+rwx only on the top directory; TraversalDelete() does nothing on DOTREE_PRED, so read-only subdirectories block unlinking their entries.
  • Prototype prototypes/38a31dffdf.patch (63 lines, test fCmd-5.13; fails on 9.0, passes on main with fix; fCmd.test and unixFCmd.test pass): on DOTREE_PRED add u+wx to readable directories. Unreadable (000) subdirectories are deliberately left alone, because fCmd-20.1 expects that case to fail; the reviewer may prefer full u+rwx and changing fCmd-20.1 (the top-level directory already gets u+rwx, so file delete -force tfa/a works while file delete -force tfa does not).
  • Suggested ticket comment Reproduced on Linux too: -force only makes the top-level directory writable (DoRemoveDirectory in unix/tclUnixFCmd.c), not read-only subdirectories, contrary to file.n. Proposed fix: make read-only subdirectories writable during the traversal; test fCmd-5.13.
⬜bc4ac0dd5f2024lsearch -sorted -inline -subindices incorrect resultpatchREPRODUCED (main, 9.0, 8.7b1; 8.6 differs: -subindices is ignored with -inline there and the whole element is returned)

Open ticket bc4ac0dd5f · 17. Commands I-L · filed 2024-11-18

🧪 Prototype check (2026-10-24)

bc4ac0dd5f.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before work-a3/vbc4.tcl -> a (expected b); the -real list gives 1.0 2.0 2.0 4.00 4.00 6.000; also lsearch -sorted -subindices -inline -stride 2 -index 0 {a 1 b 2 c 3 d 4 e 5} c -> b (expected c)
  • Patch applied with an offset (C hunk +8 lines). I moved test lsearch-25.7 above the "# 26* mimics 19* but with -stride added" comment, where the original had put it below. 12 changed lines.
  • After b, 1.0 2.0 3.000 4.00 5.0 6.000, stride case c. Tests lsearch (166/0 failed), cmdIL: 0 failed
  • Test existing lsearch-25.7 (from the patch) fails on the unpatched tip / passes after
Full verdict

REPRODUCED (main, 9.0, 8.7b1; 8.6 differs: -subindices is ignored with -inline there and the whole element is returned)

  • Platform generic
  • Tested main, 9.0.5, 8.7b1, 8.6.18 with the ticket's examples. lsearch -sorted -subindices -inline -index 0 {{a 1} {a 2} {b 3} {c 4} {c 5}} b → a; the -real list gives 1.0 2.0 2.0 4.00 4.00 6.000.
  • Cause generic/tclCmdIL.c Tcl_LsearchObjCmd, single inline result with -subindices. It uses listv[i+groupOffset], where i is the last element the binary search probed, instead of the found index. For linear search the two are the same, so only -sorted is affected. Annotate shows the line coming in with the merge 9d080b5730 (2018-03-05, merge 8.7).
  • Prototype prototypes/bc4ac0dd5f.patch (12 changed lines, test lsearch-25.7). i → index. lsearch.test passes; the new test fails on 9.0.5.
  • Suggested ticket comment Cause: the -inline -subindices result is taken from the last element probed by the binary search (i) instead of the found index. Proposed one-line fix plus test lsearch-25.7.
⬜3d1c409f1a2024Bad performance: tclThreadAlloc vs. CPU-cache (fragmentation/granularity/etc)...REPRODUCED (main, 8.6)

Open ticket 3d1c409f1a · 41. Memory Allocation · filed 2024-11-18

Full verdict

REPRODUCED (main, 8.6)

  • Platform generic
  • Tested reporter's script (single thread, 40 rounds of 1M array set/get) on main -O2, main -O2 -DPURIFY, 8.6.18
  • Cause generic/tclThreadAlloc.c: the per-thread bucket/object caches are plain LIFO lists, so after many alloc/free cycles consecutive allocations are scattered in memory. With zippy: set 0.09 -> 0.31 us, get 0.04 -> 0.22 us after 40 rounds (8.6 same pattern). With -DPURIFY (system malloc), the same build stays flat at 0.048/0.022 us.
  • Prototype none (changing the allocator design, not a small fix; sebres has an implementation in his fork)
  • Suggested ticket comment Still reproducible on main and 8.6: after 40 rounds, set/get timings grow 3-6x with the threaded allocator, but stay flat in a -DPURIFY (system malloc) build of the same check-in. So the slowdown comes from the allocator's free-list ordering, not from the hash tables.
⬜66aa81c90a2024MS-Windows+TCL9: if a load failes due to a missing secondary DLL, the error message says "No error".patch 86patchREPRODUCED (main, 9.0, 8.6; Release builds with the threaded allocator, long DLL paths only)

Open ticket 66aa81c90a · 17. Commands I-L · filed 2024-11-11

🧪 Prototype check (2026-10-24)

66aa81c90a-86.patch: VERIFIED+TEST-ADDED

  • Base core-8-6-branch (tip 6b2fff63e6; fossil update core-8-6-branch: already up to date)
  • Before tclsh86t.exe 66aa.tcl (win/66aa81c90a.tcl with its temp dir moved into the scratch dir; win/66aa81c90a/needsdep.dll) -> 51-char path: this library or a dependent library could not be found in library path (WIN_LOAD MOD_NOT_FOUND); 171-char path: couldn't load library "...": No error (POSIX {unknown error} {No error}), VERDICT: BUG
  • Patch applied cleanly (Git's patch.exe -p0); 11 changed lines in win/tclWinLoad.c, plus the new test (35 changed lines in all)
  • After same script -> both paths ... could not be found in library path (WIN_LOAD MOD_NOT_FOUND), VERDICT: OK; tests load.test: 31 total, 14 passed, 17 skipped, 0 failed (with load-12.1)
  • Test added load-12.1 (constraint ddeDll = win + a *dde*.dll next to the test executable): it copies the dde DLL into a 120-char directory with its import of USER32.dll changed to USERXX.dll, so loading it fails on the missing dependency. Unpatched (only win/tclWinLoad.c reverted): load-12.1 FAILED (13 passed / 1 failed); patched: PASSED. The same test also fails unpatched / passes with 66aa81c90a.patch on core-9-0-branch (No error, POSIX {unknown error} vs WIN_LOAD MOD_NOT_FOUND)

66aa81c90a.patch: VERIFIED+TEST-ADDED

  • Base main (tip c092dae4af)
  • Before tclsh91 66aa.tcl -> 171-char path: No error (POSIX {unknown error} {No error}), VERDICT: BUG
  • Patch applied cleanly; 11 changed lines in win/tclWinLoad.c, plus the new test (34 changed lines in all)
  • After both paths this library or a dependent library could not be found in library path (WIN_LOAD MOD_NOT_FOUND), VERDICT: OK; nmake test TESTFLAGS=-file load.test 31 total, 14 passed, 17 skipped, 0 failed
  • Test added load-12.1 (the same test as in 66aa81c90a-86.patch: a copy of the dde DLL with its USER32.dll import renamed, in a 120-char directory); FAILED with win/tclWinLoad.c reverted (No error), PASSED with the patch. The same test also fails unpatched / passes patched on core-9-0-branch with this patch
Full verdict

REPRODUCED (main, 9.0, 8.6; Release builds with the threaded allocator, long DLL paths only)

  • Platform Windows
  • Tested code reading only; Unix uses dlopen/dlerror and is not affected. Repro script: win/66aa81c90a.tcl with a prebuilt x86_64 mingw test DLL win/66aa81c90a/needsdep.dll (it imports a missing missingdep66aa.dll). The script loads it from a short and from a long (>100 chars) path.
  • Cause win/tclWinLoad.c TclpDlopen reads GetLastError() for the second LoadLibraryExW only after Tcl_DStringFree(&ds). When the native path is longer than the DString static buffer, the free goes through Tcl's allocator. The threaded allocator (default in 9.0; not in 8.6 MSVC builds) calls TlsGetValue, which sets the last error to 0. ERROR_SUCCESS then hits the default: case and gives "No error". This explains why Jan's short ./tcl9memorymoduletest.dll got the correct message and Harald's long C:/... path did not. The code is the same in 8.6, so 8.6 builds with the threaded allocator would show it too.
  • Prototype prototypes/66aa81c90a.patch (11 changed lines; saves the second LoadLibraryExW error before Tcl_DStringFree). Not compiled or tested on Windows.
  • Suggested ticket comment Probable cause: TclpDlopen calls GetLastError() after Tcl_DStringFree; for paths that do not fit the DString static buffer the free goes through the threaded allocator (TlsGetValue), which clears the last error, so it reads 0 = "No error". Proposed patch saves the error right after LoadLibraryExW; it still needs to be checked on Windows.
Windows VM, run 3
  • Platform Windows
  • Verdict REPRODUCED (main, 9.0, 8.6; Release builds with the threaded allocator, long DLL paths only)
  • Tested VM win/66aa81c90a.tcl with win/66aa81c90a/needsdep.dll (imports a missing DLL): 42–53 char path → this library or a dependent library could not be found in library path (WIN_LOAD MOD_NOT_FOUND); 162–173 char path → couldn't load library "...": No error (errorCode POSIX {unknown error} {No error}) on 9.1b1, 9.0.5 and 8.6.18. The 9.1 memdebug Debug build gives the right message for both, as expected if the allocator clears the last error.
  • Cause confirmed: win/tclWinLoad.c TclpDlopen reads GetLastError() after Tcl_DStringFree, which (for a native name too long for the static buffer) goes through the threaded allocator and resets the last error to 0
  • Prototype prototypes/66aa81c90a.patch (main, existing, 11 changed lines) and new prototypes/66aa81c90a-86.patch (core-8-6-branch, 11 changed lines; the 9.0 file has the same code as main): save the error right after the second LoadLibraryExW. Built on Windows: the long path now gives the correct "dependent library could not be found" message on main and 8.6; load.test passes (13 passed, 17 skipped, 0 failed) on both. No test added (needs a DLL with a missing dependency). Tested on Windows.
  • Suggested ticket comment Confirmed on Windows 11 with 8.6.18, 9.0.5 and 9.1: with a long path, GetLastError() is read after Tcl_DStringFree, which resets it, hence "No error". Proposed fix (tested on Windows, main and 8.6): save the error right after LoadLibraryExW.
⬜02977e00042024test oo-1.23 failsREPRODUCED (main, 9.0, 8.6) — stack-depth dependent; oo-1.23/1.24 still marked knownBug

Open ticket 02977e0004 · 35. TclOO Package · filed 2024-10-22

Full verdict

REPRODUCED (main, 9.0, 8.6) — stack-depth dependent; oo-1.23/1.24 still marked knownBug

  • Platform generic
  • Tested -constraints knownBug -match oo-1.2[34]: wt-a3 main (-g, -O0) passes with 8 MB stack, segfaults with 4 MB; 9.0 and 8.6 tip tcltest (-O2) pass with 8 MB and 4 MB, segfault with 2 MB and 1 MB. Control: oo-1.20–1.22 pass at 2 MB.
  • Cause recursive object/class teardown (Tcl_DeleteNamespace -> ObjectNamespaceDeleted -> TclOODeleteDescendants -> ...) uses C stack proportional to ownership depth. Partial fix b569739692 (2024-11-01) was backed out (53462a5b29 main, 2024-11-12) because of macOS mem-debug crashes; branches bug-02977e0004, bug-02977e0004-sebres, bug-02977e0004-jn are unmerged.
  • Prototype none (destruction-order redesign, see dkf's 2024-12-03 comment).
  • Suggested ticket comment Still reproducible on main, 9.0 and 8.6 with a small C stack (ulimit -s 2048 for -O2 builds, 4096 for -O0). The fix branches are not merged.
⬜209fd9adce2024package files documented behaviourpatchREPRODUCED (main, 9.0, 8.7b1; 8.6 has no package files)

Open ticket 209fd9adce · - New Builtin Commands · filed 2024-09-29

🧪 Prototype check (2026-10-24)

209fd9adce.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before package-bug-209fd9adce scenario (tcllib-style top-level pkgIndex.tcl that sources subdir indexes; package aa does a failing nested package require): package files aa -> a/pkgIndex.tcl a/aa.tcl a/pkgIndex.tcl b/pkgIndex.tcl, i.e. the indexes rescanned by package unknown are attributed to aa
  • Patch applied cleanly (48 changed lines, 15 in C)
  • After -> a/pkgIndex.tcl a/aa.tcl; tests package.test 403 passed/1 skipped/0 failed, pkgMkIndex.test 27/0/0, tm.test 21/0/0, safe.test 153/2/0, load.test 30/0/0
  • Test existing package-bug-209fd9adce fails before / passes after
Full verdict

REPRODUCED (main, 9.0, 8.7b1; 8.6 has no package files)

  • Platform generic
  • Tested tcllib-style layout (top-level pkgIndex.tcl that sources subdir pkgIndex.tcl files, as tcllib does) plus a package whose load does a failing nested package require; main build, 9.0.5, 8.7b1
  • Cause generic/tclPkg.c PkgRequireCoreStep1 runs the package unknown script while the requiring package's name is on top of the tclPkgFiles stack. tclPkgUnknown itself uses source -nopkg, but plain source calls inside pkgIndex.tcl (tcllib's top-level index) are recorded via TclPkgFileSeen, so every nested failed require adds all those pkgIndex.tcl files. Jan's idea (remember the pkgIndex.tcl in the ifneeded data) is already implemented (30f95df8b0); this is the remaining leak.
  • Prototype prototypes/209fd9adce.patch (48 changed lines, 15 in C, test package-bug-209fd9adce): hide the names stack (set to NULL) while the unknown script runs, restore in PkgRequireCoreStep2, as source -nopkg does. package.test, pkgMkIndex.test pass.
  • Suggested ticket comment Still present: tclPkgUnknown uses <tt>source -nopkg</tt>, but plain <tt>source</tt> calls inside pkgIndex.tcl (e.g. tcllib's top-level index) run during a nested <tt>package unknown</tt> are recorded for the package being loaded. Proposed fix: hide the package-name stack while the <tt>package unknown</tt> script runs (patch + test package-bug-209fd9adce available).
⬜f14b33ec9f2024changed behaviour wrt command names, namespaces and resolutionREPRODUCED (main, 9.0, 8.7b1, 8.6.18) — not 9.0-specific: the restriction disappeared in 8.6.9

Open ticket f14b33ec9f · 21. [namespace] · filed 2024-09-05

Full verdict

REPRODUCED (main, 9.0, 8.7b1, 8.6.18) — not 9.0-specific: the restriction disappeared in 8.6.9

  • Platform generic
  • Tested namespace eval foo {proc :bar {} ...} succeeds, namespace which :bar gives ::foo:::bar, and ::foo:::bar is "invalid command name" on main, 9.0.5, 8.7b1, 8.6.9–8.6.18; 8.5.19 and 8.6.0–8.6.8 refuse with "can't create procedure ":bar" in non-global namespace with name starting with ":"". namespace eval foo {namespace eval :baz {}} also creates an unreachable ::foo:::baz.
  • Cause the check in generic/tclProc.c:Tcl_ProcObjCmd was removed by 4039b2b895 (2018-02-15, pooryorick, "merge pyk-backport-to-8-6"), shipped in 8.6.9.
  • Prototype none (policy question: pooryorick argues such names should be allowed, others want the 8.6.8 error back; also concerns commands created by other means, not only proc).
  • Suggested ticket comment Not new in 9.0: the check was dropped in 8.6.9 by 4039b2b895 (merge pyk-backport-to-8-6), so 8.6.9+ behaves the same. Needs a decision whether to restore the error or make namespace which return a reachable name.
⬜d8f12171562024windows dos device paths inconsistencies and missing functionalitypatchREPRODUCED (main, 8.6) for //?/UNC/...; //./C:/... is part of the RFE (not supported)

Open ticket d8f1217156 · 37. File System · filed 2024-08-11

🧪 Prototype check (2026-10-24)

d8f1217156.patch: VERIFIED

  • Base main (tip c092dae4af)
  • Before tclsh91 d8f.tcl (win/d8f1217156.tcl with the temp dir in scratch) -> //?/UNC/localhost/C$/... exists=0 pathtype=relative (also with the illegal. tail); //./C:/... exists=0; VERDICT: REPRODUCED (3 of 7)
  • Patch applied cleanly, 28 changed lines, not changed
  • After both //?/UNC/... rows exist=1, pathtype=absolute, so native access through \\?\UNC\localhost\C$\... works, including the illegal. name. Only //./C:/... is still exists=0; it is the RFE part and not in the patch. File tests admin 8 failed, low 6 failed, as baseline
  • Test existing filename-6.31 and 6.32 FAIL on Windows with generic/tclFileName.c reverted and pass with the patch
Full verdict

REPRODUCED (main, 8.6) for //?/UNC/...; //./C:/... is part of the RFE (not supported)

  • Platform Windows (path parsing part is generic code, testable with testsetplatform win)
  • Tested Linux tcltest with testsetplatform windows: file pathtype //?/UNC/host/share/x → relative, file split → {} host share x on main and 8.6 tip; //?/UNC//host/share/x (extra slash) → volumerelative. //?/c:/x is fine.
  • Cause generic/tclFileName.c ExtractWinRoot: after skipping //?/UNC/ the remainder host/share/... no longer starts with //, so the UNC branch is never taken and the path falls through to "relative"; with an extra slash (//?/UNC//host) it becomes volume-relative. //./c:/x is parsed as UNC host "." share "c:" (win/tclWinFile.c then has no special handling either).
  • Prototype prototypes/d8f1217156.patch (93 lines, tests filename-6.31, filename-6.32; fail on 9.0, pass on main with fix; fileName.test and fileSystem.test pass): parse the host directly after //?/UNC/ and don't prepend another //. Native access of the resulting \\?\UNC\host\share still needs checking on Windows (win/d8f1217156.tcl).
  • Suggested ticket comment //?/UNC/host/share/... is parsed as a relative path: ExtractWinRoot() in generic/tclFileName.c strips the //?/UNC/ prefix and then expects the host to start with // again. Proposed fix with tests filename-6.31/6.32 (reproducible on any platform with testsetplatform win). //./ paths are a separate RFE.
Windows VM, run 4
  • Platform Windows (path parsing in generic code)
  • Verdict REPRODUCED (main, 8.6) for //?/UNC/...; //./C:/... is part of the RFE (not supported)
  • Tested VM win/d8f1217156.tcl on 9.1b1 and 8.6.18: //?/UNC/localhost/C$/... exists=0, pathtype relative (also with an illegal. tail); //./C:/... exists=0; //?/C:/..., //localhost/C$/... and //./UNC/localhost/C$/... work.
  • Cause generic/tclFileName.c ExtractWinRoot expects // again after the //?/UNC/ prefix
  • Prototype prototypes/d8f1217156.patch (existing, main, 28 changed lines by the BRIEF count, tests filename-6.31/6.32) applied to tcl-wt91 and built: both //?/UNC/... rows now exist=1 with pathtype absolute (native access through \\?\UNC\host\share works); fileName.test 285 passed / 0 failed (incl. the new tests), fileSystem, fCmd, cmdAH 0 failed; winFCmd 3 and winFile 5 failures as unpatched. //./C:/... still fails (RFE). Tested on Windows.
  • Suggested ticket comment //?/UNC/host/share/... is parsed as a relative path by ExtractWinRoot(). Proposed fix with tests filename-6.31/6.32; on Windows 11 such paths then work natively (&#91;file exists&#93; on \\?\UNC\localhost\C$\...). //./C:/... support remains a separate RFE.
✅3ff2d724d02024build failure on Alpine Linux (MUSL) - sys/queue.h no such file or directorypatchREPRODUCED by code reading (main, 9.0; 8.6 too: tclEpollNotfy.c includes <sys/queue.h> unconditionally) - CANNOT-TEST on real musl (no musl toolcha…

Open ticket 3ff2d724d0 · 53. Configuration and Build Tools · filed 2024-08-07

✅ Since triage (updated 2026-11-11)

Status now: Closed (Fixed)

  • 🔧 2026-10-20 check-in f7ece61aa8 on epoll-no-sys-queue, in main: Fix 3ff2d724d0: Tcl did not build with musl (e.g. Alpine Linux without bsd-compat-headers), which has no <sys/queue.h>. Define the six list macros that tclEpo
  • 🔧 2026-10-20 check-in 5e89bab5e2 on trunk, in main, 9.0: Fix 3ff2d724d0: build failure on Alpine Linux (MUSL) - sys/queue.h no such file or directory
  • 🔧 2026-10-20 check-in 617c86df64 on core-9-0-branch, in main, 9.0: Fix 3ff2d724d0: build failure on Alpine Linux (MUSL) - sys/queue.h no such file or directory
  • 💬 2026-10-20 serhiy.storchaka: As Harald asked, Tcl itself can avoid this: tclEpollNotfy.c only needs six list macros from , which musl does not provide. Proposed fix in branch epoll-no-sys-queue: define them in tclEpollNotfy.c instead of including that header, so no configure check and no bsd-compat-headers are needed. Tested on Alpine 3.24 without bsd-compat-headers: the current main fails as reported, with the branch it builds and the event,…
  • 💬 2026-10-20 oehhar: Great. The preprocessor gymnastics feel good. It would be ok for me to merge. Thanks, Harald
  • 💬 2026-10-20 jan.nijtmans: Please don't forget the changes.md/changes entries
Full verdict

REPRODUCED by code reading (main, 9.0; 8.6 too: tclEpollNotfy.c includes <sys/queue.h> unconditionally) - CANNOT-TEST on real musl (no musl toolchain/Alpine VM here)

  • Platform unix (Linux/musl)
  • Tested simulated a libc without the header (ac_cv_header_sys_queue_h=no) with the prototype: builds, and event/socket/io/unixNotfy tests pass (1112 passed, 0 failed) with the fallback macros on the epoll notifier.
  • Cause unix/tclEpollNotfy.c needs only LIST_HEAD/ENTRY/INIT/INSERT_HEAD/REMOVE/FOREACH from the BSD <sys/queue.h>, which musl does not ship (Alpine needs the bsd-compat-headers package).
  • Prototype prototypes/3ff2d724d0.patch (~30 hand-written lines: AC_CHECK_HEADERS([sys/queue.h]) in the Linux notifier branch, HAVE_SYS_QUEUE_H in tclConfig.h.in, fallback definitions of the six LIST_* macros in tclEpollNotfy.c; plus regenerated configure). To verify on Alpine: build without bsd-compat-headers.
  • Suggested ticket comment Proposed fix: check for sys/queue.h in configure and define the six LIST_* macros tclEpollNotfy.c uses when it is missing, so musl builds work without bsd-compat-headers. Tested by forcing the header check off on glibc; needs a run on Alpine.
⬜ba3a158b522024Additional commands to be disabled in safe interpspatchREPRODUCED (main, 9.0, 8.6) — info loaded in a safe interp lists full paths of shared libraries loaded anywhere in the process; info hostname a…

Open ticket ba3a158b52 · - New Builtin Commands · filed 2024-08-04

🧪 Prototype check (2026-10-24)

ba3a158b52.patch: VERIFIED (hostname still exposed; policy decision pending)

  • Base main (tip d6dd6d0f71)
  • Before safe-6.4 FAILED (info loaded works in a safe interp); patched interp.test hidden-list tests interp-21.5, 21.8, 23.2 fail on the unpatched build (expected, they list tcl:info:loaded)
  • Patch applied cleanly (7 changed lines)
  • After make dltest.marker; make test -file 'interp.test safe.test safe-stock.test safe-zipfs.test info.test load.test package.test' -> 1288 passed / 0 failed
  • Test existing safe-6.4 fails before / passes after
Full verdict

REPRODUCED (main, 9.0, 8.6) — info loaded in a safe interp lists full paths of shared libraries loaded anywhere in the process; info hostname and tcl_platform also exposed (design question)

  • Platform generic
  • Tested main build (parent loads dltest/tcl9pkga.so, interp create -safe child's info loaded shows the full path), 9.0.5, 8.6.18 (8.6 even exposes info nameofexecutable); info script returns {} in a plain safe interp and Safe Base tokenizes source paths
  • Cause generic/tclCmdIL.c defaultInfoMap: the "loaded" entry has unsafe flag 0 (nameofexecutable and cmdtype have 1)
  • Prototype prototypes/ba3a158b52.patch (7 changed lines, test safe-6.4; interp.test hidden list updated): mark info loaded unsafe. interp, safe*, info, load, package tests pass. Library uses of info loaded (tcl_findLibrary inside catch, safe::interpCreate -load in the parent) are unaffected. hostname left alone; needs a decision.
  • Suggested ticket comment Confirmed for <tt>info loaded</tt> (full paths of all loaded shared libraries are visible in a safe interp). Proposed fix: mark it unsafe like <tt>info nameofexecutable</tt> (one-line change in tclCmdIL.c, test safe-6.4). <tt>info hostname</tt> and <tt>tcl_platform</tt> need a separate decision; for 9.1 or later, since this changes behaviour.
⬜b5a02c53902024Guard http::Connect against cleared socket infopatchREPRODUCED (main, 9.0, 8.6)

Open ticket b5a02c5390 · 29. http Package · filed 2024-07-16

🧪 Prototype check (2026-10-24)

b5a02c5390.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before the patch's test used as repro with the library change stashed: TCL_LIBRARY=wt-a2/library ./tcltest ../tests/http.test -match 'http-4.20*' -> both thread levels FAILED with {{can't read "state(socketinfo)": no such element in array}} {} 1 sock... (bgerror after cleanup, token still exists, socket leaked)
  • Patch applied with offset +7 in http.test (re-saved, 39 changed lines, content unchanged)
  • After http.test: 156 total, 150 passed, 6 skipped, 0 failed; http11.test: 98 total, 86 passed, 12 skipped, 0 failed (httpProxy/httpPipeline not run, per rules)
  • Test existing http-4.20.* fails before / passes after
Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform generic
  • Tested localhost, no TLS needed (work-a4/s/b5a.tcl): geturl -command ... -keepalive 0, update idletasks (socket opened, connect pending), http::cleanup, then enter the event loop. The writable handler still calls http::Connect on the cleaned-up token, which raises a bgerror (8.6.18/http 2.9.9: can't read "state(sock)", as in the ticket; 9.0.5 and main: can't read "state(socketinfo)" from Finish, because Connect's eof check is now caught). The socket also leaks. With -keepalive 1 the same bgerror happens on 9.0.
  • Cause library/http/http.tcl:http::cleanup only cancels timers/coroutines and unsets the state array. For a transaction still in progress it leaves the socket open and its fileevent handlers pointing at the dead token. Harald's proposed guard in Connect would hide the error, but the socket and its handlers would still leak.
  • Prototype prototypes/b5a02c5390.patch (72 lines, test http-4.20.*): in http::cleanup, if state(status) is still "" and a socket exists, unset state(-command) (no user callback from cleanup) and catch {reset $token}, which closes or releases the socket before the state is unset. The repro is clean (no bgerror, no leaked channel); the http*.test files pass; http-4.20 fails on unpatched 9.0. Side note, not addressed: with -keepalive 1 a reset in the middle of a request leaves the socket in socketMapping, so it may be reused with a half-finished exchange.
  • Suggested ticket comment Reproduced without TLS: cleanup while the socket is still connecting leaves the writable handler (http::Connect) and the socket behind, giving this bgerror on 8.6 and a similar one (state(socketinfo)) on 9.0/main. Proposed fix: http::cleanup resets an unfinished transaction (without calling -command) instead of guarding Connect only; test http-4.20.
⬜929a918ec12024docs should say oo::configurable can only be mixed in with oo::configurablepatchREPRODUCED (main, 9.0); DOCS

Open ticket 929a918ec1 · 69. Other · filed 2024-07-09

🧪 Prototype check (2026-10-24)

929a918ec1.patch: VERIFIED

  • Base main (tip d6dd6d0f71); docs only
  • Before tcltest 929a.tcl (oo::configurable class with property p plus a mixin of a plain class V) -> with mixin V: unknown method "configure": must be destroy or hello; with mixin -append V: 0 3 hi. The doc does not mention it
  • Patch applied cleanly (6 changed lines)
  • After man -l configurable.n renders the new sentences in ADVANCED USAGE; they match the behaviour above
  • Test none possible (docs only)
Full verdict

REPRODUCED (main, 9.0); DOCS

  • Platform docs (generic)
  • Tested main build, 9.0.5. In a configurable class, mixin V gives unknown method "configure"; mixin -append V works (work-a2/oo/mix.tcl). The obj5-bad.tcl attachment is missing from the repository. 8.6 has no oo::configurable.
  • Cause as dmcooke said: oo::configurable works by mixing in oo::configuresupport::configurable, and a plain mixin in the class definition replaces the mixin list. The title's diagnosis ("only mixed in with oo::configurable") is wrong.
  • Prototype prototypes/929a918ec1.patch (6 changed lines, docs only; doc/configurable.n ADVANCED USAGE: use mixin -append)
  • Suggested ticket comment The cause is that mixin replaces the mixin list, which removes oo::configuresupport::configurable. mixin -append V works. Proposed doc fix (patch attached) says this in configurable.n.
⬜e9b7cb3b452024Missing ticket category: Doc or DocumentationREPRODUCED (site): still missing

Open ticket e9b7cb3b45 · 69. Other · filed 2024-07-09

Full verdict

REPRODUCED (site): still missing

  • Platform site (Fossil ticket config)
  • Tested ticket-common config in the local tcl.fossil clone. subsystem_choices has no documentation entry (it has "75./91. Tutorial", "90. FAQs", "51. Release Notes"), and type_choices is "Bug Support Patch RFE".
  • Cause ticket-common subsystem_choices
  • Prototype none (site setup, admin only). Suggestion: add e.g. {58. Documentation} to subsystem_choices.
  • Suggested ticket comment Still true: the ticket form has no documentation category. An admin could add a "Documentation" entry to subsystem_choices in the ticket setup; until then "69. Other" is used.
⬜2b7d3233ae2024minor doc errorpatchREPRODUCED (main, 9.0); DOCS

Open ticket 2b7d3233ae · 69. Other · filed 2024-07-09

🧪 Prototype check (2026-10-24)

2b7d3233ae.patch: VERIFIED

  • Base main (tip d6dd6d0f71)
  • Before running the define.n example (extracted to work-a4/v4define2.tcl) prints DB: create row in ::Users matching username=abc (doc says "from"), and the doc's Group find -> invalid command name "Group".
  • Patch applied cleanly (8 changed lines); with Groups find the example runs and prints DB: locate row from ::Groups matching groupname=webadmins, as the patched comment says (object numbers in the doc are illustrative)
  • After groff 25 warnings (same 25 before)
  • Test none possible (docs only)
Full verdict

REPRODUCED (main, 9.0); DOCS

  • Platform docs
  • Tested ran the doc/define.n DBTable example on the main build. It prints "DB: create row in ::Users ...", while the page shows "create row from". The same example also calls a non-existent Group find (the class is Groups), and its shown output "locate row ::Group with ..." does not match the code ("locate row from ::Groups matching ...").
  • Cause doc/define.n example output out of sync with its code
  • Prototype prototypes/2b7d3233ae.patch (8 changed lines, docs only: fixes both "create row" lines, plus Groups find and its output)
  • Suggested ticket comment Confirmed. The example output in define.n does not match the code: "create row in", and the Group/Groups mismatch in the find example. Proposed doc fix attached.
⬜708b0056952024info frame via invokehidden does not return full frame information.REPRODUCED (main, 9.0, 8.6)

Open ticket 708b005695 · - New Builtin Commands · filed 2024-07-02

Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform generic
  • Tested main build, 9.0.5, 8.6.18: child invokehidden info frame 0 → type precompiled, info frame → 1; for comparison child eval {info frame 0} reports the parent's source frame (type source, file, line)
  • Cause invokehidden (generic/tclInterp.c ChildInvokeHidden → TclObjInvoke) runs the command without pushing any CmdFrame in the child, so iPtr->cmdFramePtr is NULL; InfoFrameCmd (tclCmdIL.c) falls back to the "precompiled" placeholder added for the crash 0de6c1d79c
  • Prototype none (needs a design decision on what frame invokehidden should report and CmdFrame plumbing across interps/NRE; not a small fix)
  • Suggested ticket comment Still the same in 9.1 and 8.6: invokehidden pushes no CmdFrame, so <tt>info frame 0</tt> gives <tt>type precompiled</tt> and <tt>info frame</tt> gives 1. Behaviour is consistent and harmless; it needs a decision on what frame info invokehidden should expose (e.g. the parent's invoking frame, as <tt>interp eval</tt> does).
⬜d0d47f81962024Some httpProxy tests require internet accesspatchREPRODUCED (main, 9.0; 8.6 has no httpProxy.test). The no-proxy tests are still unconstrained.

Open ticket d0d47f8196 · 29. http Package · filed 2024-06-14

🧪 Prototype check (2026-10-24)

d0d47f8196.patch: FIXED-MEANWHILE (2b7247bc38). Different approach from the prototype: the file probes socket -async www.google.com 80 (5 s timeout) whenever it is sourced, so loading httpProxy.test still opens an outside connection and the tests run whenever the network is up. The opt-in needsInternet of the prototype would avoid that; it may still be worth suggesting. The rule "never run httpProxy.test" therefore still applies

  • Base main (tip d6dd6d0f71)
  • Before does not reproduce as reported: on the tip httpProxy-2.1..5.2 carry a new constraint googleReachable, added by 2b7247bc38 (max, 2026-10-20, "Add test constraint googleReachable to skip tests that access google.com in build environments without network access"). Checked statically only (httpProxy.test must not be run here)
  • Patch no longer applies (8 of 9 hunks fail: the constraint lists changed); not ported, per the brief
  • After n/a
  • Test n/a
Full verdict

REPRODUCED (main, 9.0; 8.6 has no httpProxy.test). The no-proxy tests are still unconstrained.

  • Platform generic (tests)
  • Tested tests/httpProxy.test on main and 9.0: httpProxy-2.1, 3.1, 4.1, 5.1 have -constraints {} and fetch http://www.google.com/; the https variants 2.2/3.2/4.2/5.2 need only needsTls, which is opt-in. Note: an earlier make test TESTFLAGS="-file http*.test" in this triage run did include httpProxy.test, so these 4×2 tests really did contact www.google.com from this machine (they passed). Nobody meant to use the network.
  • Cause the tests were written for a Squid test setup where internet access is assumed. The proxy tests are opt-in (needsSquid*), but the direct ones are not.
  • Prototype prototypes/d0d47f8196.patch (123 lines of context, 10 changed lines, no new test): adds an opt-in constraint needsInternet (documented in the header next to needsSquid*) to the 8 no-proxy tests. With it, httpProxy.test is fully skipped by default (100 skipped, 0 failed); with -constraints needsInternet they run as before.
  • Suggested ticket comment Still present: httpProxy-2.1/3.1/4.1/5.1 have no constraint and access www.google.com directly. Proposed fix: an opt-in needsInternet constraint on the no-proxy tests (2.x–5.x .1 and .2), like the existing needsSquid* constraints. A local server can't stand in here, because these tests compare the direct and proxied paths to the same site.
⬜7ec0aa14002024http -progress callback invoked multiple times with no progresspatchREPRODUCED (main, 9.0, 8.6)

Open ticket 7ec0aa1400 · 29. http Package · filed 2024-06-01

🧪 Prototype check (2026-10-24)

7ec0aa1400.patch: VERIFIED. The TLS empty-read case from the report goes through the same code path but was not tested (no TLS here).

  • Base main (tip d6dd6d0f71)
  • Before the patch's http-4.21 (a local chunked server, parts 20 ms apart) with the unpatched tip library -> ok abcdefghij {5 10 10 10} (two callbacks without new bytes) at ThreadLevel 0 and 1
  • Patch applied cleanly (the http.test hunk is offset by 7 lines; 41 changed lines)
  • After {5 10}; http-4.21.* passed 20/20 runs (timing-based test, no flakiness seen). With Thread (levels 0-2): http.test Total 234 Passed 225 Skipped 9 Failed 0; http11.test 129 passed/0 failed. httpPipeline/httpProxy were not run (network rule).
  • Test existing http-4.21.$ThreadLevel (from the patch) fails before / passes after
Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform generic
  • Tested localhost server only (no TLS available): a chunked response sent in small separate packets (work-a4/s/prog_server.tcl, prog_client.tcl) gets 2 callbacks with no new bytes on main, 9.0.5 and 8.6.18. New test http-4.21: unpatched 9.0 gives 5 10 10 10, where 5 10 is expected. With TLS (as in the report), every readable event that carries no application data gives a 0-byte read and another identical callback. That is the same code path, but untested here.
  • Cause library/http/http.tcl:http::Event calls the -progress command after every readable event in the body state, even when n == 0 (chunk-size line read alone, final 0 chunk, trailer, or an empty read from a TLS channel).
  • Prototype prototypes/7ec0aa1400.patch (74 lines, test http-4.21.*): call -progress only if $n > 0. http/http11/httpcookie/httpPipeline tests pass. (The -channel/CopyContent path calls -progress per copied block and was not changed.)
  • Suggested ticket comment Reproduced locally with a chunked response: http::Event calls -progress after every readable event, also when no body bytes were read (chunk-size lines, the final chunk, and empty reads on TLS sockets). Proposed fix: call it only when data was read; test http-4.21.
⬜0f3e1921c52024TIP 426 C API not implementedREPRODUCED (main, 9.0) — still not public; effectively an RFE (needs a stubs-table/API decision)

Open ticket 0f3e1921c5 · - New Builtin Commands · filed 2024-05-19

Full verdict

REPRODUCED (main, 9.0) — still not public; effectively an RFE (needs a stubs-table/API decision)

  • Platform generic
  • Tested code reading: main has only MODULE_SCOPE TclRegisterCommandTypeName/TclGetCommandTypeName (generic/tclInt.h:3730), nothing in tcl.decls; used internally by tclBasic.c, tclZlib.c (and TclOO)
  • Cause TIP 426 public functions were implemented as internal-only
  • Prototype none (adding public stubs entries is an API change for the next minor release, not a bug fix)
  • Suggested ticket comment Still internal-only in 9.1 (<tt>TclRegisterCommandTypeName</tt>/<tt>TclGetCommandTypeName</tt> are MODULE_SCOPE). Since there is now an interested user (sqlite3 command types), exporting them via tcl.decls could be considered for 9.2; otherwise the TIP text should be amended and the ticket closed.
⬜31fd8427062024Interim increment of iPtr->compileEpoch causes reset of line-numberREPRODUCED (main, 9.0, 8.6)

Open ticket 31fd842706 · 47. Bytecode Compiler · filed 2024-05-13

Full verdict

REPRODUCED (main, 9.0, 8.6)

  • Platform generic
  • Tested reporter's script with testbumpinterpepoch on wt-a3/unix/tcltest (main), tcl9.0 and tcl8.6 tip tcltest: first run after the bump prints "line 2" instead of "line 4" for info frame 0; second run is correct.
  • Cause generic/tclExecute.c TEBCresume, label instStartCmdFailed: after an epoch change the rest of the current command is re-evaluated from a fresh string object (GetSrcInfoForPc -> Tcl_NewStringObj -> instEvalStk), losing the line/context information of the enclosing bytecode, so line numbers restart at 1 of that substring. Test-only branch bug-31fd84270644f67d (81c5544b70, 6a05960d39, 2024-05-13, sebres) has no fix and is not merged.
  • Prototype none (needs the eval to carry the original command's line and continuation-line info, i.e. a CmdFrame/ContLineLoc for the re-evaluated substring; not a small change in TEBC).
  • Suggested ticket comment Still reproducible in main, 9.0, 8.6. Cause: instStartCmdFailed re-evaluates the command text as a new string without location info. Branch bug-31fd84270644f67d has the test only.
⏳910d67a2292024NS-qualified invocations of command evade [namespace unknown]REPRODUCED (main, 9.0, 8.6) -- design question, waiting for TIP 689 (State: Draft, Vote: Pending, Tcl-Version 9.1)

Open ticket 910d67a229 · 21. [namespace] · filed 2024-03-05 · Pending (resolved, waiting for confirmation)

⏳ Since triage (updated 2026-11-11)

Status now: Pending (Fixed)

    Full verdict

    REPRODUCED (main, 9.0, 8.6) -- design question, waiting for TIP 689 (State: Draft, Vote: Pending, Tcl-Version 9.1)

    • Platform generic
    • Tested reporter's script on main build (c3bf6dd052), 9.0.5, 8.6.18: ::tcl::xxx::test and namespace inscope ::tcl {xxx::test} both give "invalid command name"
    • Cause generic/tclBasic.c TEOV_NotFound uses the unknown handler of the caller's namespace only (TIP 181 design). Fix on branch tip-689 (last merge from 9.0 by sebres 2025-07-26, f464fbae80); the 2024-03 merge to trunk/core-8-branch was backed out on 2024-03-20, not merged to main.
    • Prototype none (implementation already exists on tip-689; needs TIP vote, not a bug fix)
    • Suggested ticket comment Ticket is Pending/Fixed but the change is not in main, 9.0 or 8.6: the merge was backed out and TIP 689 is still a Draft. Implementation is on branch tip-689. Suggest status Open (or Deferred) until the TIP is voted.
    🔧f91ab723f32024exec+wish fails to handle windows batch file with arguments and spacespatchREPRODUCED (main 9.1.1, 9.0.5, 8.6 tip, installed 8.6.18; wish only). Same bug as 6dacb3b6f0

    Open ticket f91ab723f3 · 16. Commands A-H · filed 2024-02-22

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-25 check-in e8d1e56815 on bug-f91ab723f3-quoting-main, not merged: Rebase of bug-f91ab723f3-quoting to main. Fix 6dacb3b6f0 and f91ab723f3: put the command after "cmd.exe /c" in double quotes, so cmd.exe does not strip the
    • 🔧 2026-10-14 check-in 30bc26a91e on bug-f91ab723f3-quoting, not merged: Fix 6dacb3b6f0 and f91ab723f3: append the closing quote for "cmd.exe /c" in BuildCommandLine(), before the conversion to WCHAR. Appending it to the converte
    • 💬 2026-10-25 serhiy.storchaka: Rebased the fix onto main as bug-f91ab723f3-quoting-main: the command after cmd.exe /c is enclosed in double quotes, and the closing quote is added in BuildCommandLine() before the conversion to WCHAR. Tested on Windows 11 with wish and tclsh.
    Full verdict

    REPRODUCED (main 9.1.1, 9.0.5, 8.6 tip, installed 8.6.18; wish only). Same bug as 6dacb3b6f0

    • Platform Windows
    • Tested Linux not affected (Windows-only cmd.exe path). Repro win/f91ab723f3.tcl: creates "bat dir with spaces\test.bat" echoing [%*], runs the reporter's cases in tclsh (console) and in wish (no console → "cmd.exe /c"), prints VERDICT; pass the wish path if it is not next to tclsh.
    • Cause win/tclWinPipe.c TclpCreateProcess prefixes cmd.exe /c for APPL_DOS (.bat/.cmd) when HasConsole() is false, without the outer quotes cmd.exe needs when the command has more than two quotes.
    • Existing work bug-f91ab723f3-quoting (0530fbcf2d oehhar 2024-02, 30bc26a91e Serhiy 2026-10-14), unmerged, against an old core-8-6-branch; applies to current 8.6 with fuzz, needs a small rebase for 9.0/main (BuildCommandLine now uses Tcl_UtfToWCharDString, dead-code removal changed indentation). Alternative bug-f91ab723f3-no-cmd rejected by Serhiy (breaks console output of batch files in wish); bug-f91ab723-dead-code-removal merged into core-8-6-branch (ce2bd7a315, 2024-02-27); bug-f91ab723f3-no-cmd unmerged.
    • Prototype none (branch exists)
    • Suggested ticket comment none needed (Serhiy's 2026-10-14 comment covers it); run win/f91ab723f3.tcl with wish on 8.6/9.0/9.1 builds of the rebased branch before merging.
    Windows VM, run 6
    • Platform Windows
    • Verdict REPRODUCED (main 9.1.1, 9.0.5, 8.6 tip, installed 8.6.18; wish only). Same bug as 6dacb3b6f0
    • Tested win/f91ab723f3.tcl, same matrix and same results as 6dacb3b6f0 (2 of 3 cases fail in wish, 0 in tclsh); branch bug-f91ab723f3-quoting build: FIXED; main + rebased patch: FIXED, winPipe.test/exec.test pass.
    • Cause win/tclWinPipe.c TclpCreateProcess / BuildCommandLine, see 6dacb3b6f0.
    • Prototype prototypes/f91ab723f3.patch (14 changed lines; rebase of bug-f91ab723f3-quoting onto main, applies to 9.0 with offsets; no test, see above)
    • Suggested ticket comment Proposed fix in branch bug-f91ab723f3-quoting verified on Windows 11 (8.6); the same change rebased onto main also fixes it there. It still needs a merge into 9.0/main.
    🔧b2e3cbdc8f2023make html failspatchREPRODUCED (main; same findversion code in 9.0 and 8.6)

    Open ticket b2e3cbdc8f · 53. Configuration and Build Tools · filed 2023-12-15

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-20 check-in 09a2f4899f on man2html-find-tk-generic, not merged: Fix b2e3cbdc8f: "make html" failed with "not table of contents" when the Tk source directory had a release name like tk9.1b1: findversion paired each search r
    • 🔧 2026-10-18 check-in cddebff8c5 on man2html-tk-beside, not merged: tcltk-man2html.tcl: a Tk tree found next to the source directory replaced the common prefix also used for Tcl, so all Tcl paths were wrong and the tool failed w
    • 💬 2026-10-20 serhiy.storchaka: Reproduced on main with source directories named tcl9.1b1 and tk9.1b1, exactly as reported. The cause is in findversion in tools/tcltk-man2html.tcl: foreach top1 {$top $top/..} sub {{} generic} pairs the two lists, so it only tries $top//tk.h and $top/..//generic/tk.h. After Tcl is found the search for Tk starts from the parent directory, so tk9.1b1/generic/tk.h is never checked, and Tk is only found if its direct…
    Full verdict

    REPRODUCED (main; same findversion code in 9.0 and 8.6)

    • Platform build/docs
    • Tested sibling source dirs named tcl9.1b1 / tk9.1b1 (real dirs, docs + generic copied), tcltk-man2html.tcl --useversion=9.1 --srcdir=.../tcl9.1b1 (what make html runs): "using Tk source directory" empty, then ": make-manpage-section: not table of contents." Works when the dirs are named tcl9.1/tk9.1.
    • Cause tools/tcltk-man2html.tcl:findversion iterates foreach top1 {$top $top/..} sub {{} generic} in parallel, so it only tries <top>/<dir>/tk.h and <top>/../<dir>/generic/tk.h. After Tcl is found, tcltkdir becomes the parent dir, so Tk's <parent>/tk9.1b1/generic/tk.h is never checked; the empty Tk dir then breaks the section scan.
    • Prototype prototypes/b2e3cbdc8f.patch (30 changed lines, almost all re-indentation: nested loops instead of the parallel foreach). Verified: both --srcdir=<parent> and --srcdir=<parent>/tcl9.1b1 now find tk9.1b1 and generate TkCmd/TkLib; --tcl from the checkout unchanged.
    • Suggested ticket comment Cause: findversion pairs $top with no subdir and $top/.. with generic/, so a Tk tree beside Tcl is only found if its directory matches tk$useversion. Proposed fix: try both subdirs for each search root.
    ⬜52b83aa2792023tcl9.0 finds wide character but doesn't report line numberpatchREPRODUCED (main, 9.0) -- 8.6 not affected (no strict encoding profile, source reads the bad byte silently)

    Open ticket 52b83aa279 · 45. Parsing and Eval · filed 2023-11-23

    🧪 Prototype check (2026-10-24)

    52b83aa279.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before tclsh bad52.tcl (byte 0xB0 on line 4 inside if 0 {...}) -> couldn't read file "bad52.tcl": invalid or incomplete multibyte or wide character (no position)
    • Patch applied cleanly
    • After ... wide character at line 4; also right for a bad byte on line 1 and on line 50001 of a 50001-line file (chunked reads), and with source -encoding utf-8; tests source (24/0 failed), io (884/0), ioCmd, main, cmdAH, fCmd: 0 failed
    • Test existing source-7.7 fails before / passes after
    Full verdict

    REPRODUCED (main, 9.0) -- 8.6 not affected (no strict encoding profile, source reads the bad byte silently)

    • Platform generic
    • Tested file with byte 0xB0 inside if 0 {...} on main build, 9.0.5: couldn't read file "...": invalid or incomplete multibyte or wide character, no position
    • Cause generic/tclIOUtil.c Tcl_FSEvalFileEx / TclNREvalFile: on Tcl_ReadChars failure only Tcl_PosixError is reported, although objPtr already holds the data decoded before the bad byte.
    • Prototype prototypes/52b83aa279.patch (70 changed lines, most of it a new helper SetReadFileError with header comment, replacing 4 identical error blocks; test source-7.7). For EILSEQ it appends " at line N" (N counted from the data read so far). source/io/ioCmd/main/cmdAH tests pass.
    • Suggested ticket comment Still the case in 9.0/9.1: the read error is reported without position, although the data before the bad byte is available. Proposed fix: report "... at line N" for encoding errors in &#91;source&#93; (patch with test source-7.7).
    ⬜453d66ef702023http ThreadLevel 2 test errorspatchREPRODUCED (main, 9.0; 8.6's http 2.9 has no -threadlevel) — deterministic with a reset during the socket-thread wait; the unmodified http.test pas…

    Open ticket 453d66ef70 · - New Builtin Commands · filed 2023-10-04

    🧪 Prototype check (2026-10-24)

    453d66ef70.patch: VERIFIED. The triage asked for an http patchlevel bump, but it is no longer needed on main: 9.1.0 and 9.0.4 shipped http 2.10.2, and main already says 2.10.3 (unreleased). A 9.0 backport would still need its own bump.

    • Base main (tip d6dd6d0f71); Thread 3.0.7 (thread.fossil trunk) built in work-a3/thread
    • Before TCLLIBPATH=work-a3/thread tcltest v453.tcl (-threadlevel 1, geturl to a local server, after idle {http::reset $tok}) -> BGERROR: invalid command name "::http::1--SocketCoroutine" (the message from the ticket), then status: reset
    • Patch applied cleanly (the http.test hunk is offset by 7 lines; 24 changed lines)
    • After same -> only status: reset, no bgerror. With Thread: http.test Total 234 Passed 225 Skipped 9 Failed 0; http11.test 0 failed (httpcookie.test is all skipped here)
    • Test existing http-4.11.1.$ThreadLevel (from the patch): with the unpatched library http-4.11.1.1 and 4.11.1.2 FAIL; with the patch they pass
    Full verdict

    REPRODUCED (main, 9.0; 8.6's http 2.9 has no -threadlevel) — deterministic with a reset during the socket-thread wait; the unmodified http.test passes 40/40 on main and 40/40 on 9.0 tip with Thread 3.0.7 (6 parallel runs) on this machine, so the suite only hits it under timing pressure

    • Platform generic
    • Tested main build and 9.0 tip tcltest with Thread 3.0.7; repro: -threadlevel 1, http::geturl ... -command, after idle {http::reset $tok} → bgerror invalid command name "::http::1--SocketCoroutine" (exact message from the ticket)
    • Cause library/http/http.tcl http::cwaiter::CwaitHelper does after 0 $coroName when the socket thread writes its result variable; if the coroutine was deleted while yielded (http::Finish renames ${token}--SocketCoroutine on reset/timeout, cf. tests http-4.11/4.12 "-timeout 1" + reset), the after script fails
    • Prototype prototypes/453d66ef70.patch (24 changed lines, test http-4.11.1.$ThreadLevel): resume through a helper CwaitResume that checks the coroutine still exists. Full http.test passes on main with Thread; the new test fails on the 9.0 tip library (ThreadLevel 1 and 2). An http patchlevel bump (2.10.3 → 2.10.4) still needs to be added, since http is a released package.
    • Suggested ticket comment Reproduced deterministically: an <tt>http::reset</tt> (or timeout) while the socket is being opened in a helper thread deletes ${token}--SocketCoroutine, and CwaitHelper's <tt>after 0 $coroName</tt> then raises a bgerror. Proposed fix: resume through a helper that checks the coroutine still exists (patch + test http-4.11.1.* available).
    ⬜5030db5deb2023zipfs find does not take links into accountpatchREPRODUCED (main, 9.0, 8.7b1)

    Open ticket 5030db5deb · - New Builtin Commands · filed 2023-09-24

    🧪 Prototype check (2026-10-24)

    5030db5deb.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before zipfs-find-native-links FAILED: a a/f a/l a/l/l a/l/l/f a/l/l/l ... (a/l -> ../b, b/l -> ../a, recursion until the path is too long)
    • Patch applied cleanly (hunk offset 52 lines; 16 changed lines)
    • After a a/f a/l b b/l; make test TESTFLAGS="-file zipfs.test" 499 passed / 0 failed
    • Test existing zipfs-find-native-links (unix) fails before / passes after
    Full verdict

    REPRODUCED (main, 9.0, 8.7b1)

    • Platform generic (native file systems with symlinks)
    • Tested Linux, xx/a/l → ../b, xx/b/l → ../a: zipfs find returns 166 entries on main and 9.0.5, 3982 on 8.7b1 (recursion only stops when the path gets too long, so the count depends on the base path length); expected 6
    • Cause generic/tclZipfs.c, script proc ::tcl::zipfs::Find recurses into every entry returned by glob, including symlinks to directories.
    • Prototype prototypes/5030db5deb.patch (49 lines, test zipfs-find-native-links; fails on 9.0, passes on main with fix; zipfs.test passes): list links but do not descend into them (file type ne "link").
    • Suggested ticket comment ::tcl::zipfs::Find descends into symbolic links, so link cycles are followed until the path gets too long. Proposed fix: don't recurse into links (like find without -L); test zipfs-find-native-links.
    ⬜4ae42446ab2023overlaid zipfs mounts co-mingle directory contentsREPRODUCED (main, 9.0)

    Open ticket 4ae42446ab · - New Builtin Commands · filed 2023-09-09

    Full verdict

    REPRODUCED (main, 9.0)

    • Platform generic
    • Tested Linux main and 9.0.5: mounting teststored.zip on //zipfs:/test/testdir (a directory inside mounted test.zip) succeeds and glob shows abac-repeat.txt test2; after unmount test2 again
    • Cause generic/tclZipfs.c keeps all mounts in one flat ZipFS.fileHash; ZipFSCatalogFilesystem silently skips names already present ("should not happen but skip it anyway"), and unmount removes entries without restoring shadowed ones.
    • Prototype none (needs a decision: reject mounts on existing directories/inside archives vs. defined overlay semantics; juliannoble2 and griffin argued for allowing it)
    • Suggested ticket comment Still the same on main and 9.0. Needs a decision whether to refuse a mount point that is an existing directory in another archive, or to support overlays properly (restoring shadowed entries on unmount).
    ⬜89fbf1f86e2023Unclean memory handlingREPRODUCED (main; 8.6 per ticket comments, not rebuilt with PURIFY here)

    Open ticket 89fbf1f86e · 41. Memory Allocation · filed 2023-08-31

    Full verdict

    REPRODUCED (main; 8.6 per ticket comments, not rebuilt with PURIFY here)

    • Platform generic
    • Tested main -O2 -DPURIFY build under valgrind; 2-line script set a 0; exit $a already gives 144 bytes definitely lost, if 1 {exit $a} adds CFWord blocks from TclArgumentEnter
    • Cause [exit] calls Tcl_Exit -> Tcl_Finalize -> FreeMainInterp and never returns to the C frames that hold references: TclEvalEx's substituted word objects (TclSubstTokens) and CFWord records (TclArgumentEnter/TclArgumentRelease), and Tcl_FSEvalFileEx's normalized path. Once the main interp's tables are freed, these become unreachable and valgrind reports them as definitely lost. 8.5 showed no such report because it did not delete the main interp at exit, so the same blocks stayed "still reachable". The leak is bounded (once per process) and only exists in exit paths.
    • Prototype none (fixing it needs [exit] to unwind the C stack to Tcl_Main before finalizing, which changes how exit behaves; alternatively add these frames to tools/valgrind_suppress)
    • Suggested ticket comment Still on main: set a 0; exit $a in a script file is enough with a PURIFY build. The lost blocks are owned by C frames (TclEvalEx, Tcl_FSEvalFileEx) that Tcl_Exit never returns to, so they leak once per process, at exit. A real fix means making &#91;exit&#93; unwind to Tcl_Main before Tcl_Finalize; a cheap option is adding these frames to tools/valgrind_suppress.
    ✅13d3c0d7642023lset on empty list is allowed when index argument is largepatchREPRODUCED (main, 9.0; 8.6 rejects such indices as "bad index", so not affected)

    Open ticket 13d3c0d764 · - New Builtin Commands · filed 2023-08-13

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-17 check-in 7325d81876 on lset-empty-huge-index, in main: Fix 13d3c0d764: lset on an empty list accepted a huge index as end+1. In the special case for empty lists, accept the index only if it is end+1 when evaluated
    • 🔧 2026-10-29 check-in 62a9a4d3c2 on lset-empty-huge-index, in main: Rebase lset-empty-huge-index to main. Fix 13d3c0d764: lset on an empty list accepted a huge index as end+1. In the special case for empty lists, accept the in
    • 🔧 2026-10-29 check-in 1b4052be16 on trunk, in main, 9.0: Fix 13d3c0d764: lset on empty list is allowed when index argument is large
    • 🔧 2026-10-29 check-in 8a9e91396f on core-9-0-branch, in main, 9.0: Fix 13d3c0d764: lset on empty list is allowed when index argument is large
    • 💬 2026-10-17 serhiy.storchaka: The ambiguity can be resolved locally in TclLsetFlat(), without changing the index encoding (so Tk is not affected): when the list is empty and the index comes back as TCL_SIZE_MAX, evaluate it again as if the list had one element, and accept it only if this gives 1 (end+1 and similar forms). Proposed fix in branch lset-empty-huge-index, tests lset-18.1 to lset-18.5. 8.6 is not affected.
    Full verdict

    REPRODUCED (main, 9.0; 8.6 rejects such indices as "bad index", so not affected)

    • Platform generic
    • Tested main build, 9.0.5, 8.6.18: on an empty list lset L 0x7fffffffffffffff X (and 0x8000000000000000, and nested lset L 0x7fff... 0 X) succeeds; 0x7ffffffffffffffe is rejected
    • Cause generic/tclListObj.c TclLsetFlat: with endValue -1 the index code returns TCL_SIZE_MAX both for "end+1" (GetEndOffsetFromObj's WIDE_MAX encoding) and for huge indices, and the special case from 0b05b5c750 maps every TCL_SIZE_MAX to 0. On 32-bit builds end+2 and 0x7ffffffffffffffe would probably also be accepted (WIDE_MAX-1 clamps to TCL_SIZE_MAX); not tested.
    • Prototype prototypes/13d3c0d764.patch (34 changed lines, 10 in C, tests lset-17.1–17.5): in the special case, re-evaluate the index with endValue 0 and accept only if it gives 1 (i.e. end+1 forms). No change to the shared index encoding, so Tk is unaffected. lset, lsetComp, list, lindex, lreplace, linsert, listObj tests pass; the new tests fail on the 9.0 tip. bigdata tests not run.
    • Suggested ticket comment The ambiguity can be resolved locally in TclLsetFlat: when the list is empty and the index comes back as TCL_SIZE_MAX, re-evaluate it with end=0 and accept only if it gives 1 (end+1). Proposed fix with tests lset-17.*; it does not touch the index encoding, so Tk is unaffected.
    ⬜6f22c7a1fc2023Cannot load sampleextension when built with mingw64patchREPRODUCED (tclconfig trunk 2026-08-20) — not MinGW-specific: any extension configured with a --prefix (or a config.site setting prefix) that conta…

    Open ticket 6f22c7a1fc · - New Builtin Commands · filed 2023-08-01

    🧪 Prototype check (2026-10-24)

    6f22c7a1fc.patch: VERIFIED

    • Base tclconfig trunk (tip bed05771cd), tested through Thread trunk (work-a1/thread, nested tclconfig checkout, configure regenerated with autoconf 2.73 -f)
    • Before ../configure --with-tcl=~/tcltk/9.0/lib --prefix=~/tcltk/8.6 -> checking for Tcl public headers... ~/tcltk/8.6/include (8.6 headers for a 9.0 build); without --prefix -> 9.0/include
    • Patch the original was a --- a/tcl.m4 git-style diff (needs -p1); applied cleanly and re-saved as a fossil diff (-p0). 9 changed lines
    • After with --prefix=~/tcltk/8.6 -> ~/tcltk/9.0/include, and Thread builds (libtcl9thread3.0.7.so); without --prefix still 9.0/include
    • Test none possible (no TEA test suite); the configure runs above (work-a1/v1/6f22.sh) stand in for one
    Full verdict

    REPRODUCED (tclconfig trunk 2026-08-20) — not MinGW-specific: any extension configured with a --prefix (or a config.site setting prefix) that contains another Tcl's tcl.h picks that header over the one of --with-tcl

    • Platform build (TEA / tclconfig, any Unix-like platform)
    • Tested Linux, Thread trunk out-of-tree: configure --with-tcl=~/tcltk/9.0/lib --prefix=~/tcltk/8.6 → "Tcl public headers... ~/tcltk/8.6/include" (8.6 headers with 9.0 stubs); without --prefix → 9.0/include
    • Cause tclconfig/tcl.m4 TEA_PUBLIC_TCL_HEADERS searches the package's ${includedir} before ${TCL_PREFIX}/include and ${TCL_BIN_DIR}/../include. MSYS2's /etc/config.site sets prefix=/mingw64, where the system Tcl 8.6 tcl.h lives.
    • Prototype prototypes/6f22c7a1fc.patch (against tclconfig trunk tcl.m4, not tcl.fossil; 9 changed lines, no test): search Tcl's own prefix and the dir next to tclConfig.sh before the package prefix. Verified with Thread's configure regenerated with autoconf 2.73: the same command now finds ~/tcltk/9.0/include. Since 3317aec2c5 a mismatched tcl.h already fails at compile time.
    • Suggested ticket comment Not MinGW-specific: TEA_PUBLIC_TCL_HEADERS in tclconfig's tcl.m4 prefers the extension's <tt>$includedir</tt> over Tcl's own <tt>TCL_PREFIX/include</tt>, so any --prefix (or config.site) with another Tcl's tcl.h wins. Proposed fix: move the package prefix last in the search list (tclconfig patch available).
    💬52fd7694562023cmdMZ-6.6 fails on low performance systempatchREPRODUCED (main, 9.0, 8.6: the absolute thresholds are unchanged; not reproducible on this fast machine, judged from the test code and the reporte…

    Open ticket 52fd769456 · 06. Time Measurement · filed 2023-06-26

    💬 Since triage (updated 2026-11-11)

    Status now: Open

    • 💬 2026-10-20 serhiy.storchaka: The thresholds are unchanged on main, 9.0 and 8.6: cmdMZ-6.6 requires the 10 µs case to cost less than 100 µs per call and to run more than 500 times at more than 10000/s, a fixed budget for the interpreter overhead. A possible fix is to compare it with the 1 ms case instead (ideally 100 times faster; 5 times leaves room for slow systems): [code] I could not reproduce the failure (my machine is too fast, also unde…
    🧪 Prototype check (2026-10-24)

    52fd769456.patch: VERIFIED

    • Base main (tip d6dd6d0f71); tests only
    • Before this machine is too fast (passes, also under valgrind), so I simulated a slow one: taskset -c 5 tcltest cmdMZ.test -match cmdMZ-6.6 with 12 own busy loops pinned to the same core -> FAILED 3/3, {1 0 1 0 1 0 1 1 1} {184.8 µs/# 412 # 5411 #/sec ...} {12858 µs/# 3 # 77.8 #/sec ...}, the same three checks that fail in the ticket
    • Patch applied cleanly (6 changed lines)
    • After same contention -> cmdMZ-6.6 passes 5/5; without contention cmdMZ.test 96 passed/1 skipped/0 failed, 3/3 runs
    • Test the patch is the test fix; cmdMZ-6.6 under CPU contention fails before / passes after
    Full verdict

    REPRODUCED (main, 9.0, 8.6: the absolute thresholds are unchanged; not reproducible on this fast machine, judged from the test code and the reporter's numbers)

    • Platform generic
    • Tested main test passes here; t4b.tcl shows m1 at ~82 us/# under valgrind, near the 100 us limit
    • Cause tests/cmdMZ.test cmdMZ-6.6 requires the 10 us busy wait to cost < 100 us per call and run > 500 times / > 10000/s, i.e. a fixed interpreter overhead budget.
    • Prototype prototypes/52fd769456.patch (6 changed lines, test cmdMZ-6.6): compare m1 with m2 (5x faster per call, 5x more iterations and rate) instead of absolute numbers; the reporter's values pass (108 vs 1093 us, 463 vs 46, 9257 vs 914), cmdMZ.test passes.
    • Suggested ticket comment cmdMZ-6.6 uses absolute thresholds for the 10 us case. Proposed fix: compare it relative to the 1 ms case (5x faster, 5x more iterations), which keeps the test's point and passes with your numbers.
    ✅176de86b3b2023TEBCResume(): pointer overflow for INST_LREPLACE4patchREPRODUCED (main; 9.0 has the same size_t numNewElems code, by reading; 8.6 has no INST_LREPLACE4)

    Open ticket 176de86b3b · 17. Commands I-L · filed 2023-05-08

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Duplicate)

    • 💬 2026-11-08 chrstphrchvz: Closing as duplicate in favor of 94c6511e.
    🧪 Prototype check (2026-10-24)

    176de86b3b.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before clang 21 build -fsanitize=pointer-overflow (work-a6/ubcl), tcltest r176.tcl (proc p {} {linsert {} 0 a b}; p) -> tclExecute.c:5529:21: runtime error: addition of unsigned offset to 0x... overflowed to 0x...; linsert.test shows the same report
    • Patch applied cleanly
    • After same UBSan build, r176.tcl, linsert.test (28/0 failed), lreplace.test (3579/0), execute.test (157/0): no runtime errors; the same files pass on the normal build
    • Test none possible (only visible under clang UBSan; behaviour is unchanged otherwise)
    Full verdict

    REPRODUCED (main; 9.0 has the same size_t numNewElems code, by reading; 8.6 has no INST_LREPLACE4)

    • Platform generic
    • Tested main built with clang 21 -fsanitize=pointer-overflow: linsert.test/lreplace.test report tclExecute.c:5529:21: runtime error: addition of unsigned offset ... overflowed. gcc's pointer-overflow check does not flag it
    • Cause generic/tclExecute.c INST_LREPLACE: size_t numNewElems, so OBJ_AT_DEPTH(numNewElems - 1) = tosPtr[-(size_t)k] adds a huge unsigned offset when numNewElems ≥ 2 (and wraps when it is 0)
    • Prototype prototypes/176de86b3b.patch (2 changed lines, no test — only visible under UBSan): declare numNewElems as Tcl_Size. With it the whole test suite under clang pointer-overflow UBSan (excluding http*) shows no runtime errors; the only failures are zipfs ones caused by the out-of-tree build.
    • Suggested ticket comment Still present on main and 9.0: <tt>numNewElems</tt> is <tt>size_t</tt>, so <tt>OBJ_AT_DEPTH(numNewElems - 1)</tt> adds a wrapped unsigned offset. Proposed fix: make it <tt>Tcl_Size</tt> (one line); with it the full test suite is clean under clang <tt>-fsanitize=pointer-overflow</tt>.
    ⬜5338d8e82f2023Closing the std* channels and opening replacements sometimes doesn't register the std* channel name in interpsREPRODUCED (main, 9.0, 8.7b1, 8.6, 8.5). The behaviour follows from the documented design (NOT-A-BUG, doc clarification at most).

    Open ticket 5338d8e82f · 25. Channel System · filed 2023-05-04

    Full verdict

    REPRODUCED (main, 9.0, 8.7b1, 8.6, 8.5). The behaviour follows from the documented design (NOT-A-BUG, doc clarification at most).

    • Platform generic
    • Tested reporter's script with a file instead of the tty (work-a4/s/std.tcl) on main, 9.0.5, 8.7b1, 8.6.18, 8.5.19: all give can not find channel named "stdout".
    • Cause the standard channels are per thread (tclIO.c tsdPtr->stdoutChannel), and each interp that uses them registers its own reference. The parent's puts registers stdout in the parent, so close stdout in the child only drops the child's reference. The thread-level channel stays open, the stdout slot is not emptied, and Tcl_CreateChannel makes the new file an ordinary fileN. Without the parent's puts the child holds the last reference, the slot is emptied, and the new channel becomes stdout (StdChannels.3, "Re-initialization"), but then for every interp in the thread, the parent included.
    • Prototype none (by design; per-interp std channels would be a redesign)
    • Suggested ticket comment Standard channels are per thread, not per interp. close stdout in a child only unregisters it there while the parent still holds it, so the slot is not empty and the next opened channel does not become stdout (see StdChannels.3, "Re-initialization of Tcl standard channels"). To redirect a child's output, give the child its own channel (e.g. open it in the child and use interp alias/a puts wrapper), or close stdout in all interps of the thread first. Propose closing as Works as designed, or turning this into a doc clarification.
    ⬜3f90d66af92023Inconsistent behaviour of trace add execute leave on a command already on the stackpatchREPRODUCED (main, 9.0, 8.6)

    Open ticket 3f90d66af9 · 46. Traces · filed 2023-05-04

    🧪 Prototype check (2026-10-24)

    3f90d66af9.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before reporter's script (work-a2/v2/tr.tcl) -> "would fire fini break" only after the second foo; with a dummy enter trace added first (tr.tcl pre) it fires after the first; behaviour unchanged on the tip, trace.n silent about it
    • Patch applied cleanly; last sentence reworded to state both cases exactly ("invoked when that execution completes only if the command already had execution traces when the execution started; otherwise it is first invoked at the end of a later execution") instead of "it is only invoked at the end of later executions", which the pre case contradicts (8 changed lines now)
    • After groff -man -Tutf8 doc/trace.n renders the paragraph correctly
    • Test none possible (docs only; no code fix proposed)
    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform generic
    • Tested reporter's script on main build, 9.0.5, 8.6.18: without a pre-existing execution trace the leave trace added by foo on itself fires only at the end of the second call; with a dummy enter trace registered first it fires at the end of the first call
    • Cause generic/tclBasic.c EvalObjvCore schedules the TEOV_RunLeaveTraces callback only if iPtr->tracePtr || (cmdPtr->flags & CMD_HAS_EXEC_TRACES) at the time the command starts; TEOV_RunLeaveTraces then re-checks the current trace list. Scheduling the callback for every command would cost every command dispatch.
    • Prototype prototypes/3f90d66af9.patch (6 changed lines, doc only: doc/trace.n states that whether leave traces run is decided when an execution starts). No code fix: a cheap one would need finding active executions of the command when the first exec trace is added, which NRE does not make easy.
    • Suggested ticket comment Confirmed in 8.6, 9.0 and 9.1: leave traces are scheduled only if the command already had execution traces when it started (EvalObjvCore). Making this dynamic would slow down every command, so the proposed patch documents the behaviour in trace.n instead.
    ⬜147b28d0432023Win32 WaitForInputIdle and potentially unwanted 5 seconds delaypatchREPRODUCED (main, 8.6)

    Open ticket 147b28d043 · 49. Threading · filed 2023-05-01

    🧪 Prototype check (2026-10-24)

    147b28d043.patch: VERIFIED

    • Base main (tip c092dae4af)
    • Before tclsh91 wfi.tcl with busy.exe (work-win5/gui/busy.c rebuilt in scratch: a /subsystem:windows child that is busy for 3 s) -> exec busy.exe & returned after 3177 ms, open |busy.exe after 3159 ms
    • Patch applied cleanly, 10 changed lines, not changed
    • After 22 ms / 8 ms; exec.test, winPipe.test, process.test, pid.test, ioCmd.test: 655 total, 514 passed, 141 skipped, 0 failed
    • Test none possible (needs a GUI executable that is busy before its message loop; the test suite cannot build one)
    Full verdict

    REPRODUCED (main, 8.6)

    • Platform Windows
    • Tested source inspection only; Unix has no equivalent (fork/exec does not wait)
    • Cause win/tclWinPipe.c:TclpCreateProcess calls WaitForInputIdle(procInfo.hProcess, 5000) after CreateProcessW, a workaround for an NT 3.5 leak (Q124121). When the child is a GUI program that is busy before it enters its message loop, [exec ... &] and [open |...] block for up to 5 s. (win/nmakehlp.c has the same call, but that only affects the build.)
    • Prototype prototypes/147b28d043.patch (28 lines of diff, removes the call and rewrites the comment; no test, it is a timing behaviour; Windows build not checked). Repro: win/147b28d043.tcl (starts wish with after 3000 before the event loop, times exec &).
    • Suggested ticket comment The WaitForInputIdle call is still in TclpCreateProcess on all branches. Proposed fix: remove it (the NT 3.5 leak it worked around is long gone). Repro script: exec of a GUI child that is busy for 3 s takes ~3 s to return.
    Windows VM, run 5
    • Platform Windows
    • Verdict REPRODUCED (main, 8.6)
    • Tested VM work-win5/wfi.tcl with work-win5/gui/busy.c (a /subsystem:windows child that creates a message queue, stays busy 3 s, never shows a window; used instead of win/147b28d043.tcl to avoid wish): exec busy.exe & returns after 3065 ms (9.1b1) / 2958 ms (8.6.18), open |busy.exe after 2676 / 2954 ms.
    • Cause win/tclWinPipe.c TclpCreateProcess calls WaitForInputIdle(procInfo.hProcess, 5000) after CreateProcessW (NT 3.5 workaround, Q124121).
    • Prototype prototypes/147b28d043.patch (existing, main, 10 changed lines, no test) built on the VM: both forms return after 4 ms; exec.test, winPipe.test, process.test, pid.test, ioCmd.test 0 failed. Applies to 9.0 with an offset; for 8.6 the hunk needs a trivial manual port (same code). Tested on Windows.
    • Suggested ticket comment Confirmed on Windows 11 (8.6.18, 9.1): &#91;exec gui.exe &&#93; and &#91;open |gui.exe&#93; block until the GUI child is idle, up to 5 s. Proposed fix: drop the WaitForInputIdle() call (the NT 3.5 leak it worked around is long gone); with it the exec returns in a few ms and the exec/pipe tests pass.
    🔧4d3f2130292023VFS support causes infinite recursion during 'file exists' if working directory has been removedpatch addonREPRODUCED (main, 9.0; 8.6 only with tclvfs, not tested) — the branch fixes the reporter-like case but NOT a second crash

    Open ticket 4d3f213029 · 37. File System · filed 2023-04-25

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-22 check-in e2e5c2656c on eval-cwd-deleted, not merged: Fix 4d3f213029 for the case where [pwd] already failed after the current directory was deleted: TclFSCwdPointerEquals() compared the path with the NULL cached
    Full verdict

    REPRODUCED (main, 9.0; 8.6 only with tclvfs, not tested) — the branch fixes the reporter-like case but NOT a second crash

    • Platform generic (crash on 9.x without any extension; 8.6 needs tclvfs)
    • Tested Linux, script run with a relative path, cd into a dir, delete it, then file exists .: main 2026-10-22 and 9.0.5 crash (SIGSEGV / "alloc: invalid block" panic), 8.6.18 OK. Branch eval-cwd-deleted tip 58b479ca8a: the variant without an earlier [pwd] is fixed (source.test 24/0 failed), but this still crashes:

    set sub [file join [pwd] sub]; file mkdir $sub; cd $sub; set x [pwd]; file delete $sub; set r [catch {pwd}]; puts [file exists .]

    (SIGSEGV in TclFSCwdPointerEquals tclIOUtil.c:534 ← Tcl_FSGetNormalizedPath tclPathObj.c:1844 ← TclEvalEx). The ticket's exact script (uncaught [pwd]) just errors on main/branch.

    • Cause (1) generic/tclBasic.c TclEvalEx: goto error before parsing when the script path cannot be normalized (fixed in the branch); (2) generic/tclIOUtil.c TclFSCwdPointerEquals: after a failed [pwd] the cached cwd is NULL, and it is passed to TclGetStringFromObj() when the script path still caches the old cwd
    • Existing work branch eval-cwd-deleted (tip 58b479ca8a, 2026-10-13), unmerged (not in trunk/9.0/8.6)
    • Prototype prototypes/4d3f213029-addon.patch, an addition to be applied on top of eval-cwd-deleted (27 changed lines incl. the test; code: 7 lines in TclFSCwdPointerEquals returning 0 when tsdPtr->cwdPathPtr is NULL; test source-9.2 fails without it, passes with it; source.test, cmdAH.test, fCmd.test, fileSystem.test pass)
    • Suggested ticket comment "The branch [/timeline?r=eval-cwd-deleted|eval-cwd-deleted] still crashes when the script called <code>&#91;pwd&#93;</code> in the deleted directory and then <code>&#91;pwd&#93;</code> failed: <code>TclFSCwdPointerEquals()</code> then compares with a NULL cwd. Proposed addition to the branch: return 0 in that case, test source-9.2."
    🟢9f4feaaaa12023tcltest test contraints with a dash (-) in the name doesn't workpatchREPRODUCED (main, 9.0, 8.6) — partly fixed: the stray output and the missing skip table were fixed by 98f03d1dc3 (bug 0aaeb9c380, 2023-09-22, so …

    Open ticket 9f4feaaaa1 · 34. tcltest Package · filed 2023-04-24

    🟢 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-20 check-in 98752639ed on tcltest-dash-constraints, in main: Fix 9f4feaaaa1: a "-" between name characters is part of a constraint name, not a subtraction, so -constraints bug-1234 works.
    • 🔧 2026-10-22 check-in 891b7e7921 on trunk, in main, 9.0: Many tcltest improvements, 773924ffff, 9f4feaaaa1, 811659ffff, 572400ffff, 8413e79232 and 1528500fff
    • 🔧 2026-10-22 check-in e5335d304d on trunk, in main, 9.0: Revert change for 9f4feaaaa1, so this one can be decided upon separately
    • 🔧 2026-10-27 check-in 1d88292331 on tcltest-dash-constraints, not merged: Fix 9f4feaaaa1: a "-" between name characters is part of a constraint name, not a subtraction
    • 💬 2026-10-20 serhiy.storchaka: The stray output and the missing skip table were fixed by 98f03d1dc3 (0aaeb9c380), but on main, 9.0 and 8.6 such names are still parsed as bug - 0306a5563, so -constraints bug-0306a5563 on the command line has no effect. Proposed fix in branch 98752639ed (stacked on e055230591, for 773924ffff): a "-" between name characters is part of the name, so bug-1, a-b || c and !a-b work; a "-" next to an operator …
    • 💬 2026-10-20 dgp: This is not a new feature, but a compatibility breaking change, so best introduced in a new major release tcltest 3. Given the freedom to break compat, I would be happier breaking it harder and eliminating constraint expressions completely. We can sort out the enthusiasts for the different approaches in a TIP.
    • 💬 2026-10-22 jan.nijtmans: I already merged the proposed fix to trunk, but - thinking about it - better discuss about this first seperately.
    Full verdict

    REPRODUCED (main, 9.0, 8.6) — partly fixed: the stray output and the missing skip table were fixed by 98f03d1dc3 (bug 0aaeb9c380, 2023-09-22, so since 8.6.14/9.0.0). Still broken: -constraints bug-0306a5563 on the command line does not enable the test, because the name is still taken as the expression bug - 0306a5563.

    • Platform generic
    • Tested main (own build), 9.0.5, 8.6.18 with the reporter's rde.test via all.tcl
    • Cause library/tcltest/tcltest.tcl:Skipped — any character outside [.:_a-zA-Z0-9 ws] turns -constraints into an expression, and the regsub [.\w]+ splits names at "-".
    • Prototype prototypes/9f4feaaaa1.patch (29 changed lines, test tcltest-28.1) — a "-" between name characters now counts as part of the name, both for the "is it an expression" check and in the regsub, so bug-1, a-b || c and !a-b work. Side effect: a-b no longer means subtraction (which makes no sense for booleans). No tests in tests/*.test use "-" in constraint expressions; tcltest.test and cmdAH.test pass. It is a small behaviour change; dgp suggested a tcltest version bump for it.
    • Suggested ticket comment The stray output was fixed by 98f03d1dc3 (0aaeb9c380), but such names are still parsed as a - b, so -constraints bug-0306a5563 on the command line has no effect. Proposed fix: a "-" between name characters is taken as part of the name; test tcltest-28.1.
    ⬜d6d85099332023Tcl_InitBignumFromDouble() does not check for NaNpatchREPRODUCED (main, 9.0, 8.6) by code reading; not reachable from script

    Open ticket d6d8509933 · 48. Number Handling · filed 2023-04-08

    🧪 Prototype check (2026-10-24)

    d6d8509933.patch: VERIFIED+TEST-ADDED

    • Base main (tip d6dd6d0f71)
    • Before tclStrToD.c Tcl_InitBignumFromDouble on the tip still has only the isinf() check; NaN is not reachable from core commands (all callers filter it), so a test hook was needed: with the new testbignumobj fromdouble (below) and no fix, testbignumobj fromdouble 1 $nan succeeds without error (frexp/ldexp on NaN)
    • Patch applied cleanly; added test infrastructure + test (53 changed lines now: 11 fix, 33 tclTestObj.c, 9 obj.test)
    • After testbignumobj fromdouble 1 $nan -> error floating point value is Not a Number, errorcode TCL VALUE DOUBLE NAN; Inf -> integer value too large to represent; 1e20 -> 100000000000000000000; tests obj, expr, expr-old, lseq: 2859 total, 2815 passed, 44 skipped, 0 failed
    • Test added testbignumobj fromdouble varIndex double (generic/tclTestObj.c, uses Tcl_GetNumberFromObj so NaN passes through) and obj-34.3, fails before (no error for NaN) / passes after
    Full verdict

    REPRODUCED (main, 9.0, 8.6) by code reading; not reachable from script

    • Platform generic (C API)
    • Tested generic/tclStrToD.c on main/9.0/8.6 has only the isinf() check. All core callers (entier/int/wide/round in tclBasic.c, number compare in tclExecute.c, lseq in tclCmdIL.c) filter NaN first: expr {entier($nan)} etc. give "floating point value is Not a Number" on main and 8.6.18. Only extensions calling the public API can pass NaN.
    • Cause tclStrToD.c:Tcl_InitBignumFromDouble — no isnan() test before frexp()/ldexp() (unspecified results).
    • Prototype prototypes/d6d8509933.patch (36 lines, no test: no test command reaches the function with NaN). Adds an isnan() branch returning TCL_ERROR with the same message/errorcode as Tcl_GetDoubleFromObj (TCL VALUE DOUBLE NAN); expr*.test pass.
    • Suggested ticket comment Still no NaN check in main/9.0/8.6; core callers filter NaN, so only the C API is affected. Proposed fix: return TCL_ERROR with "floating point value is Not a Number" for NaN.
    🔧cd609219f52023directory creation temporarily changes the umaskpatchREPRODUCED (main, 9.0, 8.6) — and worse than "temporary": 8 threads running file mkdir in a loop leave the process umask permanently at 0000 (3/3…

    Open ticket cd609219f5 · 16. Commands A-H · filed 2023-04-05

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-17 check-in 5d7e74108f on mkdir-no-umask-call, not merged: Fix cd609219f5 and 623585ffff: [file mkdir] read the umask with umask(0) + umask(mode), which threads could leave at 0 for the whole process, and masked the
    • 💬 2026-10-17 serhiy.storchaka: Confirmed, and worse than temporary: with several threads running [file mkdir], the process umask ends up permanently at 0, so later files are created world-writable (main, 9.0, 8.6). Proposed fix in branch mkdir-no-umask-call, with 623585ffff: call mkdir(path, 0777) and let the kernel apply the umask; umask() is not called any more. The permissions of new directories do not change. Test unixFCmd-5.2. The old co…
    Full verdict

    REPRODUCED (main, 9.0, 8.6) — and worse than "temporary": 8 threads running file mkdir in a loop leave the process umask permanently at 0000 (3/3 runs on main, 9.0 and 8.6; /proc/self/status shows the umask 0 thousands of times while the threads run). Also, with umask 0777, file mkdir makes a mode 000 directory, although the docs promise that u+rwx is always added.

    • Platform unix
    • Tested main (own build), 9.0 and 8.6 tip tcltest (testthread), Linux
    • Cause unix/tclUnixFCmd.c:DoCreateDirectory reads the umask with umask(0); umask(mode), which is process-wide and races between threads. It then passes (0777 & ~umask) | u+rwx to mkdir(), which applies the umask again and so drops the u+rwx it was meant to force.
    • Prototype prototypes/cd609219f5.patch (58 changed lines, ~20 in C; tests unixFCmd-5.2, unixFCmd-5.3) — mkdir(path, 0777) lets the kernel apply the umask, and then a chmod adds u+rwx only if the umask removed any of these bits. The umask is never touched. 5.3 (threads, then check that a new file is 0644 under umask 022) fails without the fix (666) and 5.2 fails too (umask 0777 → 0); unixFCmd.test (5×) and fCmd.test pass.
    • Suggested ticket comment Confirmed: with several threads running file mkdir, the process umask ends up permanently at 0 (main, 9.0, 8.6). Proposed fix: let mkdir() apply the umask and add u+rwx with chmod() afterwards, never calling umask(); tests unixFCmd-5.2/5.3.
    ⬜90058f0f3a2023macOS: outdated comment in Tcl_InitNotifier()patchREPRODUCED (main, 9.0, 8.6). The comment is still there and still wrong; checked by reading the code (CANNOT-TEST at runtime: macOS, but the change…

    Open ticket 90058f0f3a · 01. Notifier · filed 2023-03-28

    🧪 Prototype check (2026-10-24)

    90058f0f3a.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before macosx/tclMacOSXNotify.c:494-500 still has the comment about re-entering via Tcl_ServiceAll() in kCFRunLoopBeforeWaiting, while both observers are created with only kCFRunLoopEntry|kCFRunLoopExit and UpdateWaitingListAndServiceEvents (line 1355) only does the waitingList bookkeeping for those two activities (checked by reading)
    • Patch applied cleanly (comment only)
    • After new comment matches the code; not compiled (macOS-only file, comment-only change)
    • Test none possible (comment)
    Full verdict

    REPRODUCED (main, 9.0, 8.6). The comment is still there and still wrong; checked by reading the code (CANNOT-TEST at runtime: macOS, but the change is comment-only).

    • Platform macOS
    • Tested macosx/tclMacOSXNotify.c on main (l. 495), 9.0 (l. 498), 8.6 (l. 549)
    • Cause since 7495bd2a4429 ("Address macOS hangs in Tcl_WaitForEvent") and ca8b8183b62d (cleanup), UpdateWaitingListAndServiceEvents handles only kCFRunLoopEntry/Exit (waitingList bookkeeping) and no longer calls Tcl_ServiceAll in kCFRunLoopBeforeWaiting. The comment's re-entry rationale no longer applies; the second observer exists because tclEventsOnlyRunLoopMode is not one of the common modes.
    • Prototype prototypes/90058f0f3a.patch (23 lines, comment only, no test)
    • Suggested ticket comment Still present on main, 9.0 and 8.6. Proposed new comment: the second observer covers tclEventsOnlyRunLoopMode, which is not a common mode, so that waitingList is also updated on run loop entry/exit when Tcl_WaitForEvent runs the run loop in that mode.
    ⬜c71eea294d2023namespace delete is unbearably slow -- variable editionpatchREPRODUCED (main, 9.0, 8.6)

    Open ticket c71eea294d · 07. Variables · filed 2023-01-13

    🧪 Prototype check (2026-10-24)

    c71eea294d.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before tclsh rc71.tcl 10000 50000 100000 (reporter's script, debug build) -> delete 43 ms / 1.67 s / 3.58 s (quadratic)
    • Patch applied cleanly
    • After 5.3 ms / 23 ms / 51 ms (linear); tests trace (291/0 failed), namespace, namespace-old, var, upvar, interp, oo, proc: 0 failed; valgrind on the -DPURIFY build: trace-18.* and rc71b.tcl (3000 vars whose unset traces each create a new variable in the dying namespace) 0 errors, terminates
    • Test existing trace-18.5 (correctness of the snapshot with unset traces that unset and create variables) passes before and after; a test failing before would have to be a timing test (flaky), so none added
    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform generic
    • Tested reporter's script, deletion of 10k/50k/100k vars: main build 34 ms / 1.9 s / 3.6 s (debug), 9.0.5 6.5 ms / 0.47 s / 0.93 s, 8.6.18 9.8 ms / 0.72 s / 1.4 s -- superlinear
    • Cause generic/tclVar.c TclDeleteNamespaceVars restarts the hash search with VarHashFirstVar after every deleted variable (to survive unset traces). Each restart scans the emptied buckets from 0 again, hence quadratic. The command table was fixed the same way earlier (TclTeardownNamespace takes a snapshot of commands).
    • Prototype prototypes/c71eea294d.patch (73 changed lines incl. a 25-line test trace-18.5; new helper NextVarToDelete takes a snapshot of the table with raised refCounts and takes a new one when it is used up and the table is not empty, e.g. traces created new variables). Debug build: 100k vars 3.6 s -> 24 ms, linear. trace/namespace/var/upvar/interp/oo/proc tests pass, also under ASan with -DPURIFY.
    • Suggested ticket comment Still quadratic in 8.6/9.0/9.1: TclDeleteNamespaceVars restarts the hash search after every variable. Proposed fix: snapshot the variables first, like TclTeardownNamespace does for commands (patch with test trace-18.5).
    ⬜d02728896c2022“Recently changed” report empty when logged outREPRODUCED (core.tcl-lang.org/tcl report rn=7, from the local clone's reportfmt)

    Open ticket d02728896c · None · filed 2022-09-29

    Full verdict

    REPRODUCED (core.tcl-lang.org/tcl report rn=7, from the local clone's reportfmt)

    • Platform other (Fossil ticket report configuration, not Tcl code)
    • Tested the report SQL in ~/tcltk/tcl.fossil reportfmt rn=7; evaluated its WHERE clause with fossil sql: NULL login -> 0 of 8965 tickets, a non-empty login -> 8965 (also showing the 149 Deleted ones, which it means to hide).
    • Cause WHERE ($login <> '' || status <> 'Deleted') uses ||, which is string concatenation in SQL, not OR, and binds tighter than <>. It parses as ($login <> status) <> 'Deleted', which is NULL (no rows) when $login is NULL (logged out) and always true when logged in. Tk's report rn=7 uses plain WHERE status <> 'Deleted'.
    • Prototype none (repository setting, needs an admin to edit report 7). Fix: WHERE (coalesce($login,'') <> '' OR status <> 'Deleted'), or just WHERE status <> 'Deleted' as in Tk.
    • Suggested ticket comment The report's WHERE clause uses || (string concatenation in SQL) instead of OR, so it is NULL for anonymous users and lists nothing. A ticket admin needs to change it to WHERE status <> 'Deleted' (as in Tk) or (coalesce($login,'') <> '' OR status <> 'Deleted').
    ⬜a63a35d5ae2022reunite old comment with relevant code in unix/configure.acpatchREPRODUCED (main, 9.0, 8.6) — comment-only issue; the orphaned "AIX remembers this path" note sits above the Darwin block (main unix/configure.ac:7…

    Open ticket a63a35d5ae · 53. Configuration and Build Tools · filed 2022-04-16

    🧪 Prototype check (2026-10-24)

    a63a35d5ae.patch: VERIFIED

    • Base main (tip d6dd6d0f71); build system, comment only
    • Before tip unix/configure.ac:719-724 still has the orphaned "AIX remembers this path" note directly above if test "uname -s" = "Darwin", about 60 lines away from the TCL_BUILD_LIB_SPEC it explains
    • Patch applied with offsets (-82 lines in configure, -16 in configure.ac); 14 changed lines (8 in configure.ac). ~/tcltk/autoconf-2.73/bin/autoconf -f on the patched configure.ac gives a byte-identical configure to the patched one (and on the unpatched tip it reproduces the checked-in configure, 0 diff lines)
    • After out-of-tree configure with the patched and the unpatched script: tclConfig.sh differs only in the build-directory name
    • Test none possible (comment-only build change); the autoconf regeneration and the tclConfig.sh comparison stand in for one
    Full verdict

    REPRODUCED (main, 9.0, 8.6) — comment-only issue; the orphaned "AIX remembers this path" note sits above the Darwin block (main unix/configure.ac:736, 8.6 configure.in:786); Tk's unix/configure.ac:1338 has the same orphan

    • Platform build
    • Tested code reading main/8.6, history per reporter (Tcl 7.5 placement above TCL_BUILD_LIB_SPEC)
    • Cause the note explains the -L\pwd\`` in TCL_BUILD_LIB_SPEC; code inserted between them over time (Darwin/framework blocks)
    • Prototype prototypes/a63a35d5ae.patch (main, 8 changed lines in configure.ac, 14 with configure regenerated by autoconf 2.73; no test)
    • Suggested ticket comment The note belongs to the -L\pwd\`` in TCL_BUILD_LIB_SPEC. Proposed fix: move it next to that assignment in the non-framework branch (comment only; Tk's unix/configure.ac has the same orphan).
    🔧69b4f7eb372022FilterInputBytes(): signed integer overflow, poor worst-case performance lengthening objPtrpatchREPRODUCED (main, 9.0, 8.6)

    Open ticket 69b4f7eb37 · 24. Channel Commands · filed 2022-03-06

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-18 check-in 34cbbe86a5 on gets-long-lines, not merged: Fix 69b4f7eb37: [gets] of a line longer than 2 GB panicked in Tcl_SetObjLength, returned a negative length with -translation binary, or crashed when a \r ende
    • 💬 2026-10-18 serhiy.storchaka: Still present in 9.1 (9.0 has the same code): Tcl_GetsObj(), TclGetsObjBinary() and FilterInputBytes() keep offsets and counts in int, so [gets] of a line over 2 GB panics in Tcl_SetObjLength ("length requested is negative"), returns a negative length with -translation binary, or crashes when a \r ends a buffer. Proposed fix in branch gets-long-lines: Tcl_Size there, and the doubling of the buffer no longer overfl…
    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform generic
    • Tested main (own build, 2026-10-16 tip c3bf6dd052), 8.6.18; dd bs=4096 count=N if=/dev/zero | tclsh x.tcl with gets stdin. 1 GB line: fine on both (8 s / 5 s, no slow linear growth on this 62 GB box). 2.3 GB line: main panics "Tcl_SetObjLength: length requested is negative: -2147483647" (SIGILL), 8.6 panics similarly (expected 2 GB limit there).
    • Cause generic/tclIO.c:FilterInputBytes still uses int offset, length, spaceLeft, GetsState.totalChars is int and copiedTotal in Tcl_GetsObj/TclGetsObjBinary is int, so on Tcl 9 offset + offset overflows (UB; Tcl_AttemptSetObjLength rejects the negative value, then Tcl_SetObjLength panics). So 9.x cannot gets a line > 2 GB although Tcl_Size would allow it. The slow linear growth only appears when the doubling allocation fails (low memory); not addressed.
    • Prototype prototypes/69b4f7eb37.patch (17 changed lines, main; no test: needs > 4 GB RAM). Widens these variables to Tcl_Size and clamps the doubling at TCL_SIZE_MAX. 2.3 GB line then reads correctly (16 s, 9 GB RSS); io/chanio/ioCmd/encoding tests pass.
    • Suggested ticket comment Still present in 9.x: FilterInputBytes() and Tcl_GetsObj() keep int offsets/counts, so &#91;gets&#93; of a line over 2 GB panics in Tcl_SetObjLength. Proposed fix: use Tcl_Size there and clamp the doubling (no test, needs > 4 GB). The growth-strategy part is unchanged.
    ⬜a5a99f0aef2021lsearch returns different results depending on order of -bisect and -exact argumentspatchREPRODUCED (main, 9.0, 8.6; 8.6 has no -stride, but lsearch -bisect -exact {a c e} d gives -1 while -exact -bisect gives 1)

    Open ticket a5a99f0aef · 17. Commands I-L · filed 2021-12-30

    🧪 Prototype check (2026-10-24)

    a5a99f0aef.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before tclsh ra5a.tcl -> -bisect -exact ... gives -1, -exact -bisect ... gives 0
    • Patch applied cleanly
    • After both 0; tests lsearch (165+1, 0 failed), cmdIL: 0 failed
    • Test existing lsearch-22.7 fails before / passes after
    Full verdict

    REPRODUCED (main, 9.0, 8.6; 8.6 has no -stride, but lsearch -bisect -exact {a c e} d gives -1 while -exact -bisect gives 1)

    • Platform generic
    • Tested main (own build), 9.0.5, 8.6.18 (8.5 has no -bisect)
    • Cause generic/tclCmdIL.c:Tcl_LsearchObjCmd — -bisect sets mode=SORTED plus bisect=1, and a later -exact resets mode to EXACT (a linear search), so the bisect flag is silently ignored. Per lsearch.n -bisect implies -sorted, and -sorted compares exactly anyway, so -exact should not cancel it.
    • Prototype prototypes/a5a99f0aef.patch (12 changed lines, test lsearch-22.7) — after option parsing, if (bisect && mode == EXACT) mode = SORTED;. -glob/-regexp after -bisect still win by the "last matching style" rule (unchanged). The test fails without the fix, and lsearch.test passes.
    • Suggested ticket comment -exact after -bisect switched back to a linear exact search and silently dropped the bisection. Since -bisect implies -sorted, which already compares exactly, the proposed fix keeps the sorted mode in that case; test lsearch-22.7.
    ⬜057b7dbd3d2021Const-folding by Expression Compiler borks Ensemble rewrite machinerypatchREPRODUCED (main, 9.0, 8.6)

    Open ticket 057b7dbd3d · 47. Bytecode Compiler · filed 2021-12-01

    🧪 Prototype check (2026-10-24)

    057b7dbd3d.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before namespace eval ::y {proc p {x} {expr {$x <= -1}}; namespace export p; namespace ensemble create}; catch {y p} m; puts $m -> wrong # args: should be "::y::p x" (same on 8.6 tip 326e1bf923)
    • Patch applied cleanly (23 changed lines, unchanged)
    • After same repro -> wrong # args: should be "y p x"; tests compExpr, compExpr-old, expr, expr-old, namespace, oo, proc: 3637 total, 3586 passed, 51 skipped, 0 failed
    • Test existing compExpr-9.1 fails before (C fix reverted, test kept) / passes after
    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform generic
    • Tested attached weird-good.tcl / weird-bad.tcl on main build, 9.0.5, 8.6.18: bad case gives should be "::x::point validate x"; reduced repro without TclOO: namespace eval ::y {proc p {x} {expr {$x <= -1}}; namespace export p; namespace ensemble create}; y p gives should be "::y::p x" instead of "y p x" (only on the first call, when the body gets compiled)
    • Cause generic/tclCompExpr.c ExecConstantExprTree executes the folded constant subexpression with TclNRExecuteByteCode while the outer command (whose body is being compiled before arg checking) has the ensemble rewrite set up; that execution resets iPtr->ensembleRewrite, as the reporter analysed.
    • Prototype prototypes/057b7dbd3d.patch (23 changed lines, test compExpr-9.1): save the three ensembleRewrite fields before executing the constant bytecode and restore them afterwards. Test fails without the fix; compExpr, compExpr-old, expr, expr-old, namespace, oo, proc tests pass.
    • Suggested ticket comment Confirmed in 8.6, 9.0, 9.1; a plain proc in an ensemble with expr {$x <= -1} shows it on the first call. Proposed fix: ExecConstantExprTree saves/restores iPtr->ensembleRewrite around executing the folded constant (patch with test compExpr-9.1).
    ⬜e68d7234df2021[time] is not yieldablepatchREPRODUCED (main, 9.0, 8.6) -- RFE-like (time was never NRE-enabled)

    Open ticket e68d7234df · 60. NRE and coroutines · filed 2021-11-16

    🧪 Prototype check (2026-10-24)

    e68d7234df.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before reporter's script (tcltest e68d.tcl) -> cannot yield: C stack busy
    • Patch applied with offsets (tclCmdMZ.c +2, cmdMZ.test +3 lines); 86 changed lines, slightly over the size limit as the triage says
    • After -> 0ok. A semantics script (e68d2.tcl: result format, errorInfo of error inside time and of a proc using return -code error -errorcode X, break/continue codes 3/4, return from a proc inside time, count 0 and -1, wrong # args) gives byte-identical output before and after. Tests cmdMZ.test 98 passed/1 skipped/0 failed, coroutine.test 81/1/0, nre.test 29/0/0, basic.test 118/29/0, tailcall.test 39/4/0, interp.test 384/7/0, safe.test 153/2/0
    • Test existing cmdMZ-5.9 and 5.10 fail before / pass after
    Full verdict

    REPRODUCED (main, 9.0, 8.6) -- RFE-like (time was never NRE-enabled)

    • Platform generic
    • Tested reporter's script on main build, 9.0.5, 8.6.18: "cannot yield: C stack busy"
    • Cause generic/tclCmdMZ.c Tcl_TimeObjCmd loops over TclEvalObjEx; no nreProc registered in tclBasic.c builtInCmds. (Same for timerate.)
    • Prototype prototypes/e68d7234df.patch (86 changed lines, ~65 code + tests cmdMZ-5.9, cmdMZ-5.10): new TclNRTimeObjCmd plus TimeLoopCallback that re-schedules itself for each iteration and uses TclNREvalObjEx; Tcl_TimeObjCmd becomes a Tcl_NRCallObjProc2 wrapper. Results, errorInfo, break/continue/return behaviour unchanged; cmdMZ/coroutine/nre/basic/tailcall tests pass; no ASan leak when a coroutine suspended inside time is deleted. Note the measured time then includes time spent suspended. Slightly over the size limit.
    • Suggested ticket comment Confirmed. &#91;time&#93; has no NRE implementation. Proposed patch makes it NRE-enabled (callback per iteration), so &#91;yield&#93; works inside it; tests cmdMZ-5.9/5.10. &#91;timerate&#93; would need the same.
    ⬜ba7491332a2021Tcl_Panic from Itcl_SafeInitpatchREPRODUCED (Itcl trunk 4.3.8 652875dd3d, built against main): namespace eval itcl {set some variable}; package require itcl aborts with "Itcl: …

    Open ticket ba7491332a · 81. Bundled Packages · filed 2021-09-21

    🧪 Prototype check (2026-10-24)

    ba7491332a.patch: VERIFIED (the ticket's second point, locating itcl.tcl, is an enhancement and not covered)

    • Base itcl trunk (tip 652875dd3d, 4.3.8), own checkout work-a1/v1/itcl with nested tclconfig, built against the main tip wt-a1
    • Before tcltest ba74.tcl (safe child creates ::itcl, then load libtcl9itcl4.3.8.so Itcl $child) -> Itcl: cannot create namespace: "::itcl" and the process aborts (exit 132)
    • Patch applied cleanly (15 changed lines)
    • After safe child -> 1 {can't create namespace "::itcl": already exists}, same in the main interp; after namespace delete itcl a second load works and itcl::class is usable (so the failed init leaves nothing behind). The itcl suite make test: 728 passed/66 skipped/0 failed (794)
    • Test existing interp-1.6 (itcl tests/interp.test) panics before (Test file error: Itcl: cannot create namespace) / passes after
    Full verdict

    REPRODUCED (Itcl trunk 4.3.8 652875dd3d, built against main): namespace eval itcl {set some variable}; package require itcl aborts with "Itcl: cannot create namespace: "::itcl"". The same happens through Itcl_SafeInit (load lib Itcl $safeInterp after the safe interp created ::itcl). The code is unchanged since the 4.2.1 report, so 8.6 + Itcl 4.2/4.3 is affected the same way.

    • Platform generic (Itcl, bundled package)
    • Tested Itcl trunk built in work-a6/itcl/build against the wt-a6 main build
    • Cause itcl generic/itclBase.c:Initialize calls Tcl_Panic when Tcl_CreateNamespace(::itcl) fails, which it does when the namespace already exists.
    • Prototype prototypes/ba7491332a.patch (diff against itcl.fossil trunk; 15 changed lines, test interp-1.6) — free infoPtr and return TCL_ERROR with Tcl's message can't create namespace "::itcl": already exists. The full itcl test suite passes (794 tests). Not changed: the second panic (::itcl::internal::dicts), which cannot happen right after ::itcl was freshly created. Also not addressed: the ticket's second point (locating itcl.tcl via pkgIndex/ITCL_LIBRARY), which is an enhancement.
    • Suggested ticket comment Still reproducible with Itcl trunk, also from a safe interpreter. Proposed fix: return TCL_ERROR (Tcl_CreateNamespace already left the "already exists" message) instead of Tcl_Panic in Initialize; test interp-1.6.
    ✅3b00b2198d2021async test suite calls are missing cleanuppatchREPRODUCED (main, 8.6) for interp-36.7: a race in the test. after 0 error foo and after 10 set done both fire in the same timer pass if the pro…

    Open ticket 3b00b2198d · 34. tcltest Package · filed 2021-07-20

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-10-20 check-in a81652d24d on interp-36.7-race, in main: Fix 3b00b2198d: interp-36.7 waits for the bgerror handler instead of a fixed 10 ms timer, which lost the race on a loaded host.
    • 🔧 2026-10-21 check-in 0f5fdaa945 on trunk, in main, 9.0, 8.6: Fix 3b00b2198d: async test suite calls are missing cleanup
    • 🔧 2026-10-21 check-in ff7935f994 on core-9-0-branch, in main, 9.0: Fix 3b00b2198d: async test suite calls are missing cleanup
    • 🔧 2026-10-21 check-in 6b2fff63e6 on core-8-6-branch, in main, 9.0, 8.6: Fix 3b00b2198d: async test suite calls are missing cleanup
    • 💬 2026-10-20 serhiy.storchaka: interp-36.7 has a race: if the process is held up for more than 10 ms, after 0 error foo and after 10 set done fire in the same timer pass and vwait returns before the bgerror handler has run. Reproduced on main and 8.6 by adding a 20 ms sleep before the vwait (result "untouched"). Proposed fix in branch a81652d24d: wait for the variable the handler sets, with a 10 s timeout. It passes with the added sleep and s…
    • 💬 2026-10-21 jan.nijtmans: +1 Test-case only fixes/improvements don't need an entry in the changelog.
    • 💬 2026-10-21 serhiy.storchaka: Merged into trunk 0f5fdaa945, backported to core-9-0-branch ff7935f994 and core-8-6-branch 6b2fff63e6.
    Full verdict

    REPRODUCED (main, 8.6) for interp-36.7: a race in the test. after 0 error foo and after 10 set done both fire in the same timer pass if the process is held up for more than 10 ms, and then vwait returns before the bgerror idle handler has run. Simulated with a 20 ms sleep before the vwait: result "untouched" on main and 8.6. socket-14.12: NOT-REPRODUCED (10 runs of socket-14.*; randport already uses a fresh port from socket -server {} 0, which has been the case since 2017). The yocto patch URL in the ticket can't be checked offline.

    • Platform generic (test suite)
    • Tested main (own build), 8.6.18 (simulated delay)
    • Cause tests/interp.test interp-36.7 uses a fixed 10 ms timer instead of waiting for the handler
    • Prototype prototypes/3b00b2198d.patch (15 changed lines, interp-36.7 itself) — vwait on result, which the bgerror handler sets, with a 10 s timeout. interp.test passes.
    • Suggested ticket comment interp-36.7 has a race: under load both timers fire before the bgerror idle handler runs. Proposed fix: wait for the handler's variable with a timeout instead of a fixed 10 ms. socket-14.12 doesn't fail here, so a separate ticket with details would help if it still happens.
    ⬜6235d23a762021Inconsistent handling of environment variables with [interp create]patchREPRODUCED (main, 9.0, 8.6): after unset env(HOME), interp create sets HOME again in parent and child

    Open ticket 6235d23a76 · 08. Environment Variables · filed 2021-06-19

    🧪 Prototype check (2026-10-24)

    6235d23a76.patch: VERIFIED

    • Base main (tip c092dae4af on the VM)
    • Before tclsh91 6235d23a76.tcl (win/) -> after unset+interp create: parent=1 child=1, VERDICT: REPRODUCED
    • Patch applied with offsets (+11 lines in win/tclWinInit.c); saved again from fossil diff; 27 changed lines
    • After same -> parent=0 child=0, NOT-REPRODUCED; env.test 33 total, 32 passed, 1 skipped, 0 failed
    • Test existing env-12.0 FAILED with win/tclWinInit.c reverted (tcltest rebuilt), passes with the patch
    Full verdict

    REPRODUCED (main, 9.0, 8.6): after unset env(HOME), interp create sets HOME again in parent and child

    • Platform Windows
    • Tested Linux 9.0.5 with win/6235d23a76.tcl; code reading of win/tclWinInit.c on main. Attachment exercise.tcl (d25b6f3b3c) is missing from the repository; the ticket body has the full repro.
    • Cause win/tclWinInit.c TclpSetVariables runs for every new interpreter and, if env(HOME) does not exist, sets it from HOMEDRIVE+HOMEPATH / USERPROFILE / c:\. The environment is process-wide, so creating a child re-creates HOME in the parent too.
    • Prototype prototypes/6235d23a76.patch (27 changed lines, test env-12.0): do the HOME setup only once per process (static flag under a mutex). Compile-checked with mingw -fsyntax-only only; env.test passes on Linux (the new test also runs there). Repro: win/6235d23a76.tcl.
    • Suggested ticket comment Cause: TclpSetVariables (win/tclWinInit.c) sets env(HOME) for each new interpreter if it is missing, and the environment is process-wide. Proposed fix: do this only once per process (patch with test env-12.0).
    Windows VM, run 1
    • Platform Windows
    • Verdict REPRODUCED (main, 9.0, 8.6): after unset env(HOME), interp create sets HOME again in parent and child
    • Tested VM 9.1b1, 9.0.5, 8.6.18; prototype in tcl-wt91
    • Cause win/tclWinInit.c TclpSetVariables sets env(HOME) for every new interpreter
    • Prototype prototypes/6235d23a76.patch (27 changed lines, test env-12.0, unchanged): tested on Windows: repro gives parent=0 child=0; env.test passes (33 tests, env-12.0 included)
    • Suggested ticket comment Confirmed on Windows with 8.6, 9.0 and 9.1. Proposed fix: set up HOME only once per process in TclpSetVariables (test env-12.0, tested on Windows).
    ⬜84cdaceffb2021Make Install fails on Windows looking for tcldde14sx.libduplicate of 2b874804b5: REPRODUCED (8.6 only, run 1); NOT-REPRODUCED (main run 1, 9.0 now)

    Open ticket 84cdaceffb · - New Builtin Commands · filed 2021-06-06

    Full verdict

    duplicate of 2b874804b5: REPRODUCED (8.6 only, run 1); NOT-REPRODUCED (main run 1, 9.0 now)

    • Platform build (Windows, nmake win/buildall.vc.bat)
    • Tested code reading main/9.0/8.6 win/buildall.vc.bat + makefile.vc
    • Cause the static pass of buildall.vc.bat runs nmake ... shell OPTS=static[,msvcrt] %1. With install, install-binaries copies the static dde/registry lib (tcldde14sx.lib in 8.6), but target shell never builds it; only dlls/release do. The reporter's workaround (build everything with OPTS=static,nomsvcrt, then install) avoids this.
    • Prototype none new — see prototypes/2b874804b5.patch (static pass builds shell dlls) and repro win/2b874804b5.cmd
    • Suggested ticket comment Duplicate of 2b874804b5: the static pass of buildall.vc.bat builds only shell, so install cannot find the static dde library. Proposed fix (there): build shell dlls in that pass.
    Windows VM, run 3
    • Platform build (Windows, nmake win/buildall.vc.bat)
    • Verdict duplicate of 2b874804b5: REPRODUCED (8.6 only, run 1); NOT-REPRODUCED (main run 1, 9.0 now)
    • Tested VM, MSVC 2022, tcl-wt90 (core-9-0-branch 16cd1f3e89): nmake shell OPTS=static then nmake install OPTS=static (work-win3/r2b-90.cmd) succeeds and installs tcl9dde14s.lib and tcl9registry13s.lib, because 9.x static builds always link the static packages. Run 1 (result-win-1.md): main also succeeds; 8.6 fails (now first on zlib1.dll), fixed by building shell dlls. (The static artifacts were removed from tcl-wt90 afterwards and its shared tclConfig.sh/tcl.nmake regenerated.)
    • Cause 8.6 win/buildall.vc.bat static pass builds only shell, while install copies the static dde/registry libs and zlib1.dll that only dlls builds
    • Prototype prototypes/2b874804b5-86.patch (core-8-6-branch, 5 changed lines, from run 1); nothing needed for 9.0/main
    • Suggested ticket comment Duplicate of 2b874804b5. Only 8.6 is affected: 9.0 and 9.1 static builds include the static packages, so install after shell works (tested on 9.0 and main). Proposed fix for 8.6 there: build shell dlls in the static pass.
    ⬜7d207629522021extra characters after close quote or brace missing contextREPRODUCED (main, 9.0, 8.6). The behaviour is unchanged, but this is an RFE, not a bug.

    Open ticket 7d20762952 · 69. Other · filed 2021-04-16

    Full verdict

    REPRODUCED (main, 9.0, 8.6). The behaviour is unchanged, but this is an RFE, not a bug.

    • Platform generic
    • Tested main build and 8.6.18: set a "x"y gives "extra characters after close-quote". The context is already in errorInfo ("while executing \"puts \"abc\"d\""), which stops at the offending character.
    • Cause generic/tclParse.c ParseTokens/Tcl_ParseCommand fixed messages (lines 283/289 on main)
    • Prototype none. The message text is checked by 39 test results in 9 test files and may be matched by user code. That makes it an RFE (new wording), not a small fix.
    • Suggested ticket comment The location is already shown in errorInfo, which ends at the offending character. Changing the message itself would affect many tests and possibly scripts that match it. Re-classify as RFE (or close)?
    🔧f2ff05fc842021Usage of static systemEncoding is not thread safe (vulnerability)patchREPRODUCED (main, 9.0) — crash; 8.6 has the same code (not run: no Thread 2.8 build)

    Open ticket f2ff05fc84 · 11. Conversions from String · filed 2021-04-06

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-17 check-in d518a8ab5d on system-encoding-keep-retired, not merged: Fix f2ff05fc84: a conversion with the system encoding could crash if another thread changed the system encoding. Keep former system encodings until finalizati
    • 🔧 2024-04-26 check-in 246fae69b1 on sebres-encoding-perf-branch, not merged: increase cache size (4 now, can be configurable later), free cache in lock, more performance test-cases
    • 💬 2026-10-17 serhiy.storchaka: Reproduced on 9.0 and main: one thread doing [file exists] on a long path while another switches [encoding system] crashes every run. Also Tcl_GetEncoding(NULL, NULL) returned systemEncoding read again after unlocking the mutex. Branch system-encoding-keep-retired is a simple alternative to sebres-encoding-perf-branch: former system encodings are kept until finalization, so a conversion can keep using the one it r…
    Full verdict

    REPRODUCED (main, 9.0) — crash; 8.6 has the same code (not run: no Thread 2.8 build)

    • Platform generic (threads)
    • Tested Thread 3.0.7 built in work-a1. Script: one thread loops file exists <600 KB path> (Tcl_UtfToExternalDStringEx with NULL encoding), main thread loops encoding system over koi8-r/koi8-u/cp1251/cp1252/iso8859-5/macCyrillic for 4 s. SEGV in 40/40 runs on 9.0.5, 10/10 on unpatched main, 3/3 under ASan (TableFromUtfProc via Tcl_UtfToExternalDStringEx/CheckAccess). With the prototype: 40/40 clean, 5/5 clean under ASan.
    • Cause generic/tclEncoding.c: the conversion functions use systemEncoding without taking a reference when encoding is NULL, so Tcl_SetSystemEncoding in another thread can free it mid-conversion. Second bug found on the way: Tcl_GetEncoding(NULL, NULL) increments the refCount of the system encoding under the mutex but then returns systemEncoding re-read *after* unlocking, so it can return a different, unreferenced encoding (same in 9.0 and 8.6).
    • Prototype prototypes/f2ff05fc84.patch (204 lines with context, 74 +/- lines, test thread-9.1, constraint thread): Tcl_GetEncoding returns the pointer it referenced; Tcl_ExternalToUtfDStringEx, Tcl_UtfToExternalDStringEx, Tcl_ExternalToUtfEx, Tcl_UtfToExternalEx and Tcl_ExternalToUtf take a reference with Tcl_GetEncoding(NULL, NULL) and re-enter themselves when encoding is NULL. That is one mutex lock/unlock pair per NULL-encoding conversion (the cost sebres objected to); encoding/io/chan/fCmd/cmdAH/utf tests pass, thread.test passes. Tcl_GetEncodingName(NULL) / Tcl_GetEncodingNulLength(NULL) still unprotected (the API cannot be fixed). sebres-encoding-perf-branch (8.5-based) is not merged.
    • Suggested ticket comment Reproduced on 9.0.5 and main: one thread doing file exists on a long path while another loops encoding system crashes every run. Also Tcl_GetEncoding(NULL, NULL) returns systemEncoding re-read after unlocking the mutex. Proposed fix: take a reference to the system encoding for NULL-encoding conversions and return the referenced pointer, test thread-9.1.
    ⬜c2a8c0a5f62021Tcl should try to read env(VAR) before unsetting itREPRODUCED (main, 9.0, 8.6)

    Open ticket c2a8c0a5f6 · 08. Environment Variables · filed 2021-03-29

    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform generic
    • Tested test extension calling setenv() in Init, on main build (9.1.1), 9.0.5, 8.6.18. unset env(MYVAR) fails with "no such element in array"; unset -nocomplain env(MYVAR) silently leaves the variable in the environment (a later info exists env(MYVAR) is 1).
    • Cause generic/tclVar.c:TclObjUnsetVar2 (and the INST_UNSET_ARRAY* bytecodes) look up the element with createPart2=0, so for an element that exists only in environ the lookup fails before any trace runs and EnvTraceProc never calls TclUnsetEnv. Read traces (info exists, $env(X)) and array traces (array unset, array names) do resync, which is why array unset env MYVAR works.
    • Prototype none (needs a decision on generic trace semantics: e.g. fire read/array traces on a missing element before unset, or have the env array resync itself on unset; not env-local in the current design)
    • Suggested ticket comment Still present in 9.1/9.0/8.6: unset looks up env(X) without creating it, so no trace fires and TclUnsetEnv() is never called for a variable set behind Tcl's back. Workaround: info exists env(X) (or array unset env X) before unset.
    ⬜84284637b62021make html/nmake htmlhelp require tk folder beside tclREPRODUCED (main, 9.0, 8.6 — same tool code)

    Open ticket 84284637b6 · - New Builtin Commands · filed 2021-02-08

    Full verdict

    REPRODUCED (main, 9.0, 8.6 — same tool code)

    • Platform build
    • Tested main tools/tcltk-man2html.tcl with --srcdir=<wt-a6> (no tk beside it): it silently printed "using Tk source directory ../tkmain", an unrelated Tk tree two levels up.
    • Cause tools/tcltk-man2html.tcl assumes Tcl and Tk share one parent (tcltkdir, --srcdir); findversion globs $top and $top/.. for any tk.h matching the version and takes the first hit. Tk's win/makefile.vc htmlhelp passes no --srcdir at all; unix make html passes --srcdir=$(TOP_DIR)/... There is no option to name the Tk tree.
    • Prototype none (needs a --tkdir=/--tcldir= option plus reworking all $tcltkdir/$tkdir path joins, and a matching change in Tk's makefile.vc/Makefile.in — two repos, more than a small fix)
    • Suggested ticket comment Still the case on main: tcltk-man2html.tcl only knows one --srcdir and takes the first Tk below it or its parent, without telling the user. Fix needs a separate --tkdir option, passed by Tk's makefile.vc/Makefile.in.
    ⬜77527e2f992021Tcltest disregards output from child interpreterspatchREPRODUCED (main, 9.0, 8.6) — by design; limitation not documented explicitly

    Open ticket 77527e2f99 · 34. tcltest Package · filed 2021-01-13

    🧪 Prototype check (2026-10-24)

    77527e2f99.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before work-a4/v4tt.test: test x-1 -body {interp create c; c eval {puts hello-child}; interp delete c} -output hello-child -> FAILED, "Output was: (empty)", and hello-child goes straight to stdout.
    • Patch applied cleanly at offset 7 (3 changed lines)
    • After groff 0 warnings (0 before)
    • Test none possible (docs only)
    Full verdict

    REPRODUCED (main, 9.0, 8.6) — by design; limitation not documented explicitly

    • Platform generic
    • Tested own repro (attachment exercise.tcl is not in the repository) — child interp puts / puts stderr inside a test with -output/-errorOutput fails on main build, 9.0.5, 8.6.18.
    • Cause library/tcltest/tcltest.tcl intercepts output by importing tcltest::Replace::puts into :: of the test interp only; child interps keep the real puts. tcltest.n already says capture "depends on usage of puts" in the test interp. Capturing everything would need a channel transform on stdout/stderr (a behaviour redesign).
    • Prototype prototypes/77527e2f99.patch (3 changed lines, docs only: tcltest.n states that puts in other/child interpreters is not captured)
    • Suggested ticket comment tcltest captures output by replacing puts in the interpreter running the test only, so child interpreters are not covered; this is the documented "pure Tcl puts" limitation. Proposed doc clarification: say so explicitly in tcltest.n. A workaround is to alias puts in the child to the parent's: interp alias $child puts {} puts.
    ⬜45ddcfc3b12021Tcl_WideInt might be > long long (64-bit)REPRODUCED (main) — effectively an RFE/portability cleanup

    Open ticket 45ddcfc3b1 · 48. Number Handling · filed 2021-01-08

    Full verdict

    REPRODUCED (main) — effectively an RFE/portability cleanup

    • Platform build
    • Tested main configured with CFLAGS=-DTCL_WIDE_INT_TYPE=__int128 (gcc, Linux x86-64): the build fails already in generic/tclCmdMZ.c:4121 (Tcl_TimeObjCmd passes long long * to TclGetWideIntFromObj, -Wincompatible-pointer-types is an error in gcc 15); configure also still defines TCL_WIDE_INT_IS_LONG=1 regardless. The test suite was never reached.
    • Cause the core assumes Tcl_WideInt is exactly 64 bits in many places (long long mixed with Tcl_WideInt, WIDE_MAX vs LLONG_MAX, libtommath i64 conversions, binary/format code).
    • Prototype none (large audit, not a small fix). Arguably the tcl.h comment should simply say Tcl_WideInt must be 64-bit.
    • Suggested ticket comment Still the case on main: with TCL_WIDE_INT_TYPE=__int128 the build fails already in tclCmdMZ.c (long long vs Tcl_WideInt). Suggest documenting in tcl.h that Tcl_WideInt must be exactly 64 bits, or treating this as an RFE.
    🔧1c4ba6fa402020--enable-64bit and Linux aarch64patchREPRODUCED (main, 9.0, 8.6 by code; emulated on x86_64 with a CC wrapper that rejects -m64, as aarch64 gcc does)

    Open ticket 1c4ba6fa40 · 53. Configuration and Build Tools · filed 2020-11-12

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-20 check-in c3120d629c on configure-64bit-default, not merged: Fix 1c4ba6fa40: on platforms whose compiler has no -m64 flag (e.g. Linux aarch64), --enable-64bit gave "64bit support being disabled" and TCL_CFG_DO64BIT was
    • 💬 2026-10-20 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1. On aarch64 the compiler is already 64-bit by default, so nothing is broken: the only effects are the warning and [tcl::pkgconfig get 64bit] being 0. Proposed fix in branch configure-64bit-default, and the same in Tk branch configure-64bit-default: when the compiler does not accept -m64, unix/tcl.m4 checks whether it is 64-bit by default (sizeof(void *) == 8) and accepts that. Che…
    Full verdict

    REPRODUCED (main, 9.0, 8.6 by code; emulated on x86_64 with a CC wrapper that rejects -m64, as aarch64 gcc does)

    • Platform build (Linux aarch64 and other 64-bit Linux targets whose gcc has no -m64)
    • Tested main configure --enable-64bit with the wrapper: "64bit support being disabled" warning, TCL_CFG_DO64BIT not defined
    • Cause unix/tcl.m4 SC_CONFIG_CFLAGS Linux case sets do64bit_ok only if the compiler accepts -m64. Harmless for the build (the default ABI is already LP64); the only effect is the warning and tcl::pkgconfig get 64bit = 0. Note x86_64 without --enable-64bit also reports 0, so that key means "--enable-64bit was given", not "64-bit build".
    • Prototype prototypes/1c4ba6fa40.patch (main, 9 hand-edited lines in tcl.m4, 48 with regenerated configure; no test — verified with the wrapper: "checking if compiler is 64-bit by default... yes", TCL_CFG_DO64BIT set). Falls back to a sizeof(void*)==8 compile check when -m64 is rejected. Tk's tcl.m4 and tclconfig need the same change.
    • Suggested ticket comment On aarch64 the default mode is already 64-bit, so the warning is cosmetic; only tcl::pkgconfig get 64bit is affected. Proposed fix: when -m64 is not accepted, accept the default mode if sizeof(void*)==8 (unix/tcl.m4).
    ✅9cdb0d58582020unix/Makefile.in problem when building pkgs for static buildspatchREPRODUCED (main, 9.0, 8.6 — by code and a dummy pkgs/ package that prints its configure args)

    Open ticket 9cdb0d5858 · 53. Configuration and Build Tools · filed 2020-11-02

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Wont Fix)

    • 🔧 2026-10-20 check-in 9d87df83e9 on pkgs-respect-disable-shared, not merged: Fix 9cdb0d5858 and 8228750a0a: with --disable-shared the bundled packages were still built as shared libraries, since --enable-shared came after the argumen
    • 🔧 2026-10-20 check-in 732583df8c on pkgs-shared-make-var, not merged: Fix 9cdb0d5858 and 8228750a0a: the bundled packages were always built as shared libraries, even with --disable-shared. Keep that default, so that [package r
    • 💬 2026-10-20 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1, on Unix and Windows: the fixed --enable-shared comes after the arguments passed on to the packages' configure, so --disable-shared never reaches them. This may be intended: shared, stub-enabled packages can be loaded with [package require] also from a static Tcl, while static ones only help when linked into the application. So there are two proposed fixes, please choose: * pkgs-r…
    • 💬 2026-10-21 dgp: The purpose of the make *packages targets in Tcl's unix Makefile is to make installable builds of the bundled collection of packages that can then successfully be [package require]d or [load]ed into the Tcl built by the same Makefile. Forcing the --enable-shared build is intentional to achieve this aim. Back when it made a difference, forcing --enable-threads had similar justification. I believe in Tcl 9 we can …
    • 💬 2026-10-30 jan.nijtmans: I agree with Don. If you build a package with --disable-shared, the only thing built is a static library (*.a or *.lib). This library is not useful in itself, it still has to be linked in the application, neither Tcl nor Tk have the machinery to do that. So, when being part of the Tcl 'batteries included', --disable-shared makes no sense for the included packages. So, closing as 'Won't fix' (in fact 'works as desi…
    Full verdict

    REPRODUCED (main, 9.0, 8.6 — by code and a dummy pkgs/ package that prints its configure args)

    • Platform build (unix)
    • Tested main configure --disable-shared; make configure-packages passes ... --disable-shared ... --enable-shared to the package configure, so packages are always shared (8.6 also appends --enable-threads)
    • Cause unix/Makefile.in configure-packages puts the fixed --enable-shared after $(PKG_CFG_ARGS) (which carries the top-level ac_configure_args). Possibly intentional: shared stub-enabled packages also work with a static tclsh and are loadable via pkgIndex, while static package libs are not loadable by [package require] unless linked into the shell.
    • Prototype prototypes/9cdb0d5858.patch (main, 4 changed lines, no test): --enable-shared $(PKG_CFG_ARGS) --libdir=$(PACKAGE_DIR) — keeps --libdir last (the reporter's version would let a user --libdir override PACKAGE_DIR). For 8.6 the same reorder with --enable-threads.
    • Suggested ticket comment Confirmed: the fixed --enable-shared comes after PKG_CFG_ARGS, so --disable-shared never reaches bundled packages. Proposed fix: put the defaults before $(PKG_CFG_ARGS) and keep --libdir last; needs a decision whether static package libraries are wanted, since [package require] can't load them from pkgs/.
    ⬜b1add8f61f2020coroutine "::http::dechunksockd72420" is already runningpatchREPRODUCED (main, 9.0, 8.6)

    Open ticket b1add8f61f · 29. http Package · filed 2020-08-29

    🧪 Prototype check (2026-10-24)

    b1add8f61f.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before TCL_LIBRARY=<wt>/library ./tcltest ../tests/http11.test -constraints knownBug -match 'http11-2.4.2*' -> coroutine "::http::dechunksock6085a6cdbd90" is already running, result timeout {HTTP/1.1 200 OK} {crc32 mismatch: ...} ..., FAILED for ThreadLevel 0 and 1
    • Patch applied cleanly
    • After http11.test (98 total, 0 failed, 2.4.2 now enabled), http.test (154, 0 failed); localhost only, no proxy/pipeline tests run
    • Test existing http11-2.4.2 (knownBug dropped): with the new test file but the unpatched http.tcl it fails again (2 failures, "already running"); passes with the fix
    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform generic
    • Tested http11-2.4.2 with -constraints knownBug on main (own build), 9.0 tip and 8.6 tip tcltest: all print coroutine "::http::dechunksock..." is already running and the test fails with timeout/crc mismatch.
    • Cause library/http/http.tcl:http::ReceiveChunked (used for -channel without -handler and chunked transfer) keeps its chan event readable binding while it calls the CopyChunk callback; a -progress callback that enters the event loop resumes the running coroutine. Not really a stacked-channel problem: the same fix as for c2dc1da315 works (unbind during the callback, rebind afterwards).
    • Prototype prototypes/b1add8f61f.patch (20 changed lines, test http11-2.4.2 enabled by dropping knownBug; it now uses a new progressPauseShort (1 ms pause), because the 100 ms pause times 193 chunks of 671 bytes goes over the 5 s -timeout once callbacks no longer run in parallel). http11.test 88/0 failures, http.test passes; without the http.tcl change the enabled test fails again.
    • Suggested ticket comment Still present in 8.6/9.0/main. Cause: http::ReceiveChunked leaves the readable fileevent bound while the CopyChunk/-progress callback runs. Proposed fix: unbind it during the callback, as in EventGateway for c2dc1da315; enables test http11-2.4.2.
    🔧be38540b322020Deadlock during fork in TclpAllocREPRODUCED (main); fixed by Serhiy's branch

    Open ticket be38540b32 · 49. Threading · filed 2020-08-18

    🔧 Since triage (updated 2026-11-11)

    Status now: Open

    • 🔧 2026-10-13 check-in ecb454f543 on alloc-fork-deadlock, not merged: Fix be38540b32: a child process created by fork() could deadlock on its first allocation, if another thread of the parent held a lock of the threaded allocato
    Full verdict

    REPRODUCED (main); fixed by Serhiy's branch

    • Platform unix (threaded builds with the threaded allocator)
    • Tested Linux, forkhang.c (ticket attachment) linked against my builds, 6 parallel copies x 20000 forks, twice per build (240000 forks each): main c092dae4af: 39 hung children (26 + 13); branch alloc-fork-deadlock (ecb454f543): 0 hung (0 + 0). 9.0/8.6 not run (same allocator code; Serhiy's comment).
    • Cause generic/tclThreadAlloc.c: the allocator's list/bucket mutexes can be held by another thread at fork(); the child inherits them locked and its first allocation that needs a new cache (GetCache via Tcl_GetThreadData in AtForkChild) blocks forever.
    • Existing work alloc-fork-deadlock (Serhiy, 1 check-in ecb454f543 2026-10-13, branched from trunk 9a3e24fe8a, unmerged; 40 changed lines in tclThreadAlloc.c, pthread_atfork holding all allocator locks; no test, as the race needs ~1/6000 forks).
    • Prototype none (Serhiy's branch exists)
    • Suggested ticket comment none needed; ready for merge, then backport to 9.0 and 8.6 (8.6 is where the report came from).
    ⬜66c2e8c9742020Interpreter name resolution has problems with namespacespatchREPRODUCED (main, 9.0, 8.6)

    Open ticket 66c2e8c974 · 20. [interp] · filed 2020-07-21

    🧪 Prototype check (2026-10-24)

    66c2e8c974.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before interp create foo; foo eval {set a abc}; interp create ::foo -> succeeds (::foo), then ::foo eval {set a} -> can't read "a": no such variable, interp children -> ::foo (the first interp was silently replaced); interp-2.14/2.15 fail
    • Patch applied cleanly; ported cosmetically: the new ChildNameKey() and its header were inserted between GetInterp's header comment and its body, moved it above that header (no code change; 45 changed lines)
    • After interp-2.14/2.15 pass; tests interp 386/0 failed, safe 153/0, safe-stock 11/0, safe-zipfs 22/0, basic 118/0, info 285/0
    • Test existing interp-2.14, interp-2.15 fail before / pass after
    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform generic
    • Tested main build, 9.0.5, 8.6.18: interp create foo; interp create ::foo succeeds, the new interp's command replaces ::foo and silently deletes the first interp.
    • Cause generic/tclInterp.c: ChildCreate/GetInterp key childTable by the literal name, while the interp command is created relative to the global namespace, so "foo" and "::foo" are two keys for one command.
    • Prototype prototypes/66c2e8c974.patch (45 changed lines, tests interp-2.14, interp-2.15). Adds ChildNameKey() that strips leading "::" (all leading colons, as namespace resolution does) for the hash key in both creation and lookup; interp create ::foo now reports "already exists", and interp exists {::foo ::bar} works. Side effect: interp children lists "foo" for an interp created as "::foo". interp.test, safe*.test, basic.test pass. Does not canonicalise inner "ns::::x" forms.
    • Suggested ticket comment Still present in 9.1/9.0/8.6. Proposed fix: key the child table by the name without leading "::", so "foo" and "::foo" name the same interpreter (tests interp-2.14/2.15).
    ⬜3d768414de2020Use of the thread package prevents interp from being fully deletedpatchREPRODUCED (main + Thread 3.0.7, 9.0.5 + Thread 3.0.7, 8.6.18 + Thread 2.8.13)

    Open ticket 3d768414de · 49. Threading · filed 2020-07-14

    🧪 Prototype check (2026-10-24)

    3d768414de.patch: VERIFIED

    • Base Thread trunk (tip 39d9f08aa0), built in work-a1/thread/buildv1 against the main tip
    • Before thread-bug-3d768414de (child interp writes a buffered x to a file, then thread::send -async $t {set x 1} ::done resp. -command, vwait, interp delete; file size shows whether the interp was really deleted and its channel flushed) with threadCmd.c unpatched -> 0 0
    • Patch applied cleanly (43 changed lines)
    • After -> 1 1; full Thread suite 128 passed/16 skipped/0 failed (144)
    • Test existing thread-bug-3d768414de fails before / passes after
    Full verdict

    REPRODUCED (main + Thread 3.0.7, 9.0.5 + Thread 3.0.7, 8.6.18 + Thread 2.8.13)

    • Platform generic (Thread package)
    • Tested child interp with assoc-data delProc (work-a6/assoc) and with a buffered file channel. Plain package require Thread, thread::create/send/release, tsv, mutex, transfer and tpool all let the interp be deleted. thread::send -async $t script ::var and thread::send -command $t script cmd leave it alive: the delProc never runs and the channel is never flushed.
    • Cause thread generic/threadCmd.c: ThreadSend calls Tcl_Preserve(clbkData->interp) for the callback, but when the result is routed back to the originating thread (ThreadEventProc, with execProc ThreadClbkSetVar/ThreadClbkCommand) nothing calls the matching Tcl_Release. The Release in the other branches only covers the same-thread case.
    • Prototype prototypes/3d768414de.patch (diff against the Thread repo trunk, 97 lines, mostly the test; C change ~15 lines in ThreadEventProc; test thread-bug-3d768414de in tests/thread.test covers the var and -command forms). It fails without the fix and passes with it; thread.test passes 126/126.
    • Suggested ticket comment Reproduced with Thread 2.8.13 and 3.0.7: thread::send -async with a result variable (or -command) Tcl_Preserve()s the caller's interp, and the routed-back callback never releases it, so the interp is never deleted. Proposed fix: release it in ThreadEventProc after running ThreadClbkSetVar/ThreadClbkCommand; test thread-bug-3d768414de. (Belongs to the Thread package.)
    ⬜ba6ab5eb7b2020[return -code] behaviour depends on how coroutine was startedpatchREPRODUCED (main, 9.0, 8.6)

    Open ticket ba6ab5eb7b · 60. NRE and coroutines · filed 2020-07-05

    🧪 Prototype check (2026-10-24)

    ba6ab5eb7b.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before tclsh rba6.tcl (proc gen {} {yield; return -code break}; coroutine t2 gen; catch {while 1 t2}), as a file and on stdin -> 1:invoked "break" outside of a loop
    • Patch applied cleanly
    • After 0: both ways; tests coroutine (85/0 failed), nre, basic, proc, apply, interp, oo, tailcall: 0 failed. Extra check rba6e.tcl (after-scripts with break run from vwait inside a coroutine, uplevel #0 {break}/{return -code error}/{eval continue} inside a coroutine and at top): identical output patched and unpatched
    • Test existing coroutine-7.20/7.21 fail before / pass after; 7.22 (plain resume at level 0 still errors) passes both ways
    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform generic
    • Tested main build, 9.0.5, 8.6.18, both as a script file and on stdin: coroutine created at top level + return -code break gives "invoked "break" outside of a loop"; created inside apply it works. The tailcall break variant from the comment works in all three now.
    • Cause generic/tclBasic.c:TEOV_PushExceptionHandlers adds the top-level TEOV_Exception handler when numLevels == 1. TclNRCoroutineObjCmd evaluates the body after iPtr->numLevels--, so the body of a coroutine created at level 0 runs at level 1 and its break/continue is turned into an error inside the coroutine instead of being returned to the resumer.
    • Prototype prototypes/ba6ab5eb7b.patch (44 changed lines, tests coroutine-7.20..7.22). Skips the top-level handler when iPtr->execEnvPtr->corPtr != NULL (the resumer still gets it, so plain t at top level still errors, test 7.22). Tests run in a subprocess (exec [interpreter]) because interp eval allows exceptions; added the exec constraint to coroutine.test. 7.20/7.21 fail on unpatched 9.0, pass patched; coroutine, nre, basic, proc, apply, interp, oo tests pass. Side effect: errorInfo for 7.22 no longer has the "gen" invoked from within frame.
    • Suggested ticket comment Still present. The coroutine body of a coroutine created at level 0 runs at numLevels 1 and gets the top-level "no break/continue" handler. Proposed fix: skip that handler inside a coroutine's execEnv, so the code reaches the resumer (tests coroutine-7.20..7.22).
    ⬜9ee9f4d7be2020zlib push/pop corrupts underlying channelREPRODUCED (main, 9.0, 8.6)

    Open ticket 9ee9f4d7be · 57. zlib · filed 2020-06-16

    Full verdict

    REPRODUCED (main, 9.0, 8.6)

    • Platform generic
    • Tested reporter's script on main (30d8ccce48), 9.0.5, 8.6.18: source size grows 5 7 9 11 (2 bytes per push/pop). Already captured as knownBug test zlib-14.1.
    • Cause generic/tclZlib.c:ZlibTransformClose always runs deflate(Z_FINISH) for a deflate-mode transform and writes the result (empty raw-deflate block, 2 bytes; more for zlib/gzip) to the parent if the channel is writable, even when nothing was written through the transform. The script reads through a write-side compressor on a read-write channel.
    • Prototype none (behaviour decision: skipping the trailer when nothing was written would change the output for "push, write nothing, close" which today yields a valid empty stream; dkf left it undecided)
    • Suggested ticket comment Still the same on 9.1/9.0/8.6 (knownBug zlib-14.1). Options: only finalize when data was written through the transform, or when the transform was pushed on a channel opened for writing only; either needs a decision about the empty-stream case.
    ⬜f45d0dc1a72020Tcl_UniCharAtIndex is not used in the Tcl core, and has no test-casespatchREPRODUCED (main, 9.0, 8.6) — still no caller in generic/ and no test calls it: the utf-8.* tests labelled "Tcl_UniCharAtIndex" use `[string index]…

    Open ticket f45d0dc1a7 · 44. UTF-8 Strings · filed 2020-04-26

    🧪 Prototype check (2026-10-24)

    f45d0dc1a7.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before still no caller of Tcl_UniCharAtIndex in generic/ and no test command for it (the utf-8.* tests use [string index]); nothing to reproduce, it is a coverage gap
    • Patch applied cleanly (tclTest.c hunks offset up to 142 lines; 48 changed lines)
    • After utf.test 404 passed / 0 failed (399 before + the 5 new utf-8.13..8.17)
    • Test existing utf-8.13..8.17 (constraint testunicharatindex); coverage tests, no failing behaviour to show before
    Full verdict

    REPRODUCED (main, 9.0, 8.6) — still no caller in generic/ and no test calls it: the utf-8.* tests labelled "Tcl_UniCharAtIndex" use [string index], which never reaches it

    • Platform generic
    • Tested grep of main/9.0/8.6 sources; prototype tests on main build
    • Cause tests/utf.test utf-8.1–8.12 test Tcl_GetUniChar indirectly; tclTest.c has no wrapper for Tcl_UniCharAtIndex.
    • Prototype prototypes/f45d0dc1a7.patch (48 changed lines, new test command testunicharatindex in tclTest.c, tests utf-8.13–utf-8.17: index 0, index > 0, negative index, emoji, end of string); utf.test passes
    • Suggested ticket comment Still true on main: utf-8.* only call &#91;string index&#93;. Proposed test command <code>testunicharatindex</code> with tests utf-8.13–17 (patch available); deprecating the function is a separate TIP question.
    ✅28de60cef52020Segfault in destructor using tsv::lockREPRODUCED (main, 9.0, 8.6); not in 8.6.0; fixed by Serhiy's branch

    Open ticket 28de60cef5 · 10. Objects · filed 2020-02-16

    ✅ Since triage (updated 2026-11-11)

    Status now: Closed (Fixed)

    • 🔧 2026-11-01 check-in db7f34af59 on oo-destructor-deleted-namespace-90, in main, 9.0: Backport of Fix 28de60cef5: reading an object variable from code compiled at run time (e.g. [eval]) in a destructor...
    • 🔧 2026-11-01 check-in 298d1d97d5 on core-9-0-branch, in main, 9.0: Fix 28de60cef5 and b4637c4312
    • 🔧 2026-11-01 check-in f3f08274f7 on trunk, in main: Fix 28de60cef5 and b4637c4312
    • 🔧 2026-10-13 check-in de6527681f on oo-destructor-deleted-namespace, in main, 9.0: Fix b4637c4312: [upvar] to an object variable after the object was destroyed in its method crashed. Same cause as 28de60cef5; add a test.
    • 💬 2026-11-01 dkf: It was a missing NULL check. Ought to be fixed now (9.0 onwards).
    Full verdict

    REPRODUCED (main, 9.0, 8.6); not in 8.6.0; fixed by Serhiy's branch

    • Platform generic
    • Tested Linux. Reporter's script (tsv::lock) on main build + Thread 3.0.7: SIGSEGV 3/3. oo-3.13 body (rename [self] {}; eval {set thing} in the destructor): SIGSEGV on main, 9.0.5, 8.6.18; 8.6.0 prints the value. Branch oo-destructor-deleted-namespace (de6527681f): reporter's script runs to the end (prints thing: something, since the destructor frame's compiled local is found by name), oo-3.13 gives the "no such variable" error; oo.test, ooNext2.test, ooProp.test pass.
    • Cause generic/tclOOMethod.c ProcedureMethodCompiledVarConnect/GetObjectVar: after rename [self] {} the object's namespace is deleted (oPtr->namespacePtr == NULL), and the object variable resolver for code compiled at run time (eval/tsv::lock/uplevel bodies) dereferences it.
    • Existing work oo-destructor-deleted-namespace (Serhiy, f38a74a978 + de6527681f 2026-10-13, branched from trunk 88c6ad4c75, unmerged; 45 changed lines; also fixes b4637c4312, tests oo-3.13, oo-3.14).
    • Prototype none (Serhiy's branch exists)
    • Suggested ticket comment none needed (already announced); needs merge + backport to 9.0 and 8.6.
    ⏳554117edde2020Tcl_SaveInterpState/Tcl_RestoreInterpState are not usable to save interpreter state (error info/code)REPRODUCED (main, 9.0, 8.6) — Pending/Fixed is wrong: the fix was never merged

    Open ticket 554117edde · 45. Parsing and Eval · filed 2020-02-10 · Pending (resolved, waiting for confirmation)

    ⏳ Since triage (updated 2026-11-11)

    Status now: Pending (Fixed)

      Full verdict

      REPRODUCED (main, 9.0, 8.6) — Pending/Fixed is wrong: the fix was never merged

      • Platform generic
      • Tested small extension doing Save/Eval(catch {error ...})/Restore (the branch's testevalobjv -safe test result-1.9.1) on main build, 9.0.5, 8.6.18: all print {UNEXPERR INVISIBLE} {ERROR in safe block} instead of the saved errorCode/errorInfo.
      • Cause generic/tclResult.c: Tcl_RestoreInterpState restores iPtr->errorInfo/errorCode but the ::errorInfo/::errorCode variables were overwritten in between and are only re-synced when the iPtr fields are non-NULL. Branch bug-554117edde (9cace6cc78, 2020-02-10; rebased to 8.6 by Jan in 45213b1bc0, 2023-09-12) is not merged into core-8-6-branch, core-9-0-branch or main; Jan's 2023 comment reports it breaks init-0.1, iocmd-21.8, iortrans-2.7.
      • Prototype none (fix exists on branch but needs rework for the 3 regressions; not small)
      • Suggested ticket comment Still reproducible on 9.1/9.0/8.6; branch bug-554117edde was never merged and its 8.6 rebase breaks init-0.1, iocmd-21.8, iortrans-2.7. Setting back to Open.
      ⏳75d525d37c2020Memory corruption transferring server (listening) sockets between threadsREPRODUCED (main, 9.0, 8.6 on Windows: crash)

      Open ticket 75d525d37c · 25. Channel System · filed 2020-01-31 · Pending (resolved, waiting for confirmation)

      ⏳ Since triage (updated 2026-11-11)

      Status now: Pending (Fixed)

        Full verdict

        REPRODUCED (main, 9.0, 8.6 on Windows: crash)

        • Platform generic (crash reported for Windows)
        • Tested main (own build + -DPURIFY build under valgrind) with Thread 3.0.7 built from thread.fossil trunk: apn's attachment (adapted to port 0), plus variants: transfer to a worker; transfer from a child interp that is then deleted; transfer from a worker thread that then exits. No crash, valgrind clean. On Unix the accept callback keeps running in the ORIGINAL thread and interp (the listening fd's file handler stays in the creating thread's notifier); after that interp is deleted the accept callback is silently dropped and the new owner never sees connections.
        • Cause unix/tclUnixSock.c / win/tclWinSock.c keep acceptProc/interp in the AcceptCallback for the creating interp; nothing re-targets it on thread::transfer. sebres' fix 455c680c3a on branch bug-75d525d37c-sf-chan-transfer (2020-05-29, adds statePtr->interp) was never merged into any release branch, so the "Pending/Fixed" status is not backed by a merged change.
        • Prototype none (needs an API/semantics decision: re-bind accept to the receiving interp or refuse to transfer server sockets)
        • Windows repro win/75d525d37c.tcl (needs Thread; runs a child, reports crash vs. which interp runs accept).
        • Suggested ticket comment Pending, but branch bug-75d525d37c-sf-chan-transfer was never merged. On Linux (main) it no longer crashes: the accept callback keeps running in the original thread/interp and stops once that interp is deleted, so transferring a server socket still does not work. Should not be closed as fixed; needs a decision whether to re-bind or to refuse transferring listening sockets.
        Windows VM, run 2
        • Platform generic (crashes on Windows; on Unix the accept callback silently stays in the old thread)
        • Verdict REPRODUCED (main, 9.0, 8.6 on Windows: crash)
        • Tested VM win/75d525d37c.tcl with Thread 3.0.7 (9.x) / 2.8.13 (8.6): after thread::transfer of a listening socket, the first connection runs the ORIGINAL interp's accept proc in the receiving thread. 9.0.5 and 8.6.18: child dies with SIGSEGV (0xC0000005); 9.1b1 Release: "can not find channel named stdout" from the main interp's accept proc, then trace trap; 9.1 memdebug Debug build: Tcl_DbIsShared: Tcl_Obj allocated in another thread panic. 1 run each, deterministic.
        • Cause generic/tclIOCmd.c AcceptCallback keeps the creating interp; on Windows the accept notification is delivered to the thread that now owns the channel (win/tclWinSock.c), which then evaluates the script in an interp belonging to another thread. sebres' fix on branch bug-75d525d37c-sf-chan-transfer (455c680c3a, 2020-05-29, 188 changed lines against 2020 8.6) was never merged.
        • Prototype none (the branch exists but needs a rebase onto current 8.6/9.x, and a decision: re-bind accept to the receiving interp or refuse to transfer listening sockets)
        • Suggested ticket comment Not fixed: branch bug-75d525d37c-sf-chan-transfer was never merged. On Windows 8.6.18, 9.0.5 and 9.1 still crash when a listening socket is transferred and a client connects (the old interp's accept script runs in the new thread; the memdebug build panics "Tcl_Obj allocated in another thread"). On Linux it does not crash, but accept keeps running in the old thread.
        ✅8413e792322020tcltest should not exit an applicationREPRODUCED (main, 9.0, 8.6) — intentional, undocumented

        Open ticket 8413e79232 · 34. tcltest Package · filed 2020-01-29

        ✅ Since triage (updated 2026-11-11)

        Status now: Closed (Fixed)

        • 🔧 2026-10-20 check-in 2052c30735 on tcltest-no-exit-when-sourced, in main: Fix 8413e79232: cleanupTests exits under Tk only when the test file is the main script, not in an application that sources it.
        • 🔧 2026-10-22 check-in 891b7e7921 on trunk, in main, 9.0: Many tcltest improvements, 773924ffff, 9f4feaaaa1, 811659ffff, 572400ffff, 8413e79232 and 1528500fff
        • 🔧 2026-10-22 check-in e44f1d0e87 on core-9-0-branch, in main, 9.0: Many tcltest improvements, 773924ffff, 811659ffff, 572400ffff, 8413e79232 and 1528500fff
        • 💬 2026-10-20 serhiy.storchaka: Still the case on main, 9.0 and 8.6. The exit is there so that wish foo.test does not stay in the event loop after the tests, so simply removing it would change how Tk test files run. Proposed fix in branch 2052c30735: cleanupTests exits only when, in addition, the test file is the main script of the application ([file normalize [info script]] eq [file normalize $::argv0]). An application that sources a test fil…
        • 💬 2026-10-22 jan.nijtmans: Fixed in 891b7e7921, not backported yet to 9.0 and 8.6
        • 💬 2026-10-28 jan.nijtmans: Now also backported to core-9-0-branch and core-8-6-branch Closing. Many thanks!
        Full verdict

        REPRODUCED (main, 9.0, 8.6) — intentional, undocumented

        • Platform generic
        • Tested script that sets ::tk_version (what Tk does) then calls tcltest::cleanupTests: process exits before the next line on main build, 9.0.5, 8.6.18.
        • Cause library/tcltest/tcltest.tcl cleanupTests (non-all-file branch) does exit when ::tk_version exists and the interactive constraint is false, so wish foo.test terminates. The code comment itself calls it wrong. Removing it changes how Tk test files run standalone under wish.
        • Prototype none (behaviour change; options: exit only when [info script] is $::argv0, or make it a configurable option, and document it in tcltest.n)
        • Suggested ticket comment Still there: cleanupTests calls exit when Tk is loaded and tcl_interactive is 0, so that wish foo.test terminates. Removing it would break running Tk test files standalone; one option is to exit only when the test file is the main script, and to document the behaviour.
        ⬜934e6eb9cf2020ensemble unknown handler confused by tailcallpatchREPRODUCED (main, 9.0, 8.6, 8.6.0)

        Open ticket 934e6eb9cf · 21. [namespace] · filed 2020-01-03

        🧪 Prototype check (2026-10-24)

        934e6eb9cf.patch: VERIFIED

        • Base main (tip d6dd6d0f71)
        • Before work-a3/v934.tcl (the ticket script, plus a call from a proc and from a coroutine) -> unknown subcommand "x": namespace ::ns1 does not export any commands in all three cases
        • Patch applied cleanly (15 changed lines)
        • After one two three; from a proc caller continued: a b (the handler's tailcall does not replace the calling proc); in a coroutine it works too. Tests namespace (313/0 failed), tailcall, nre, oo, coroutine, apply: 0 failed
        • Test existing namespace-53.12 (from the patch) fails on the unpatched tip / passes after
        Full verdict

        REPRODUCED (main, 9.0, 8.6, 8.6.0)

        • Platform generic
        • Tested reporter's script on main build, 9.0.5, 8.6.18, 8.6.0: "unknown subcommand "x": namespace ::ns1 does not export any commands".
        • Cause generic/tclEnsemble.c:EnsembleUnknownCallback calls TclSkipTailcall() before a synchronous Tcl_EvalObjv() of the handler. That marks an NRCommand below the eval's callback root, so the handler's tailcall is deferred until after the whole ensemble command; the handler's result is empty, which means "reparse", and the reparse fails.
        • Prototype prototypes/934e6eb9cf.patch (15 changed lines, test namespace-53.12). Removes the TclSkipTailcall() call, so Tcl_EvalObjv pushes its own NRCommand and the tailcall runs inside the handler eval (deferredCallbacks is consumed by the ensemble command itself, so no outer mark leaks in). namespace, tailcall, nre, oo tests pass.
        • Suggested ticket comment Still present. EnsembleUnknownCallback() calls TclSkipTailcall() before a non-NRE Tcl_EvalObjv(), so the handler's tailcall runs only after the ensemble returns and the handler looks like it returned an empty list. Proposed fix: drop that call (test namespace-53.12).
        ⬜7cc22390872020continuation lines \<newline> produce wrong line numbers in tracebackspatchREPRODUCED (main, 9.0, 8.6, 8.5)

        Open ticket 7cc2239087 · 45. Parsing and Eval · filed 2020-01-01

        🧪 Prototype check (2026-10-24)

        7cc2239087.patch: VERIFIED (the [info body] part of the ticket is language behaviour, rule [9], and is not changed)

        • Base main (tip d6dd6d0f71)
        • Before reporter's proc (tcltest 7cc2.tcl) -> errorInfo (procedure "foo" line 3) for expr 1/0 on body line 5
        • Patch applied cleanly (56 changed lines)
        • After -> (procedure "foo" line 5); tests error.test 321 passed/8 skipped/0 failed, info.test 285/3/0, proc.test 37/1/0, basic.test 118/29/0, execute.test 157/0/0, compile.test 160/11/0, apply.test 41/3/0, trace.test 286/4/0, namespace.test 312/2/0, oo.test 373/16/0, coroutine.test 81/1/0, nre.test 29/0/0, tailcall.test 39/4/0
        • Test existing error-23.1 and 23.2 fail before / pass after; error-23.3 (dynamic body) passes on both, as intended (unchanged case)
        Full verdict

        REPRODUCED (main, 9.0, 8.6, 8.5)

        • Platform generic
        • Tested reporter's proc on main build, 9.0.5, 8.6.18: errorInfo says (procedure "foo" line 3) for an error on body line 5, while info frame already reports the right line. info body returning the body with backslash-newlines replaced is language behaviour (rule [9] applies inside braces too), not a bug.
        • Cause generic/tclNamesp.c:TclLogCommandInfo counts '\n' in the bytecode source, which no longer has the substituted continuation lines. The TIP #280 ExtCmdLoc data (iPtr->lineBCPtr) has the real line of each command.
        • Prototype prototypes/7cc2239087.patch (56 changed lines, ~25 code; tests error-23.1..23.3). In TEBC's checkForCatch path (generic/tclExecute.c), after TclLogCommandInfo, look up the command's ECL entry and set iPtr->errorLine = line - eclPtr->start + 1 when the info is there. Checked file-sourced procs, procs typed on stdin, dynamic bodies, eval/uplevel/apply strings, and nested [cmd] on a continuation line: all correct, dynamic cases unchanged. The full suite (all files except http*) passes: 522081 total, 0 failed. Non-bytecode paths (Tcl_EvalEx direct) still count newlines.
        • Suggested ticket comment Still present. errorInfo's "line N" is computed by counting newlines in the body after backslash-newline substitution. Proposed fix: use the TIP 280 command line info when compiled code has it (tests error-23.*). [info body] is unchanged by design.
        ⬜f7afd3dbef2019tailcall does not trigger an [namespace unknown] handlerpatchREPRODUCED (main, 9.0, 8.6)

        Open ticket f7afd3dbef · 60. NRE and coroutines · filed 2019-12-24

        🧪 Prototype check (2026-10-24)

        f7afd3dbef.patch: VERIFIED (the side effect on other lookupNsPtr users, i.e. ensemble -map, oo forward and coroutine, noted in the triage, stays a behaviour change for a maintainer to accept)

        • Base main (tip d6dd6d0f71)
        • Before reporter's script (tcltest f7af.tcl) -> namespace eval ns1 nosuchroutine -> hello, but apply {{} {tailcall nosuchroutine} ns1} -> invalid command name "nosuchroutine"
        • Patch applied cleanly (20 changed lines)
        • After both -> hello; tests tailcall.test 40 passed/4 skipped/0 failed, namespace.test 312/2/0, namespace-old.test 126/0/0, nre.test 29/0/0, basic.test 118/29/0, unknown.test 7/0/0, oo.test 373/16/0, ooNext2.test 57/5/0, coroutine.test 81/1/0, proc.test 37/1/0, apply.test 41/3/0, interp.test 384/7/0, safe.test 153/2/0
        • Test existing tailcall-bug-f7afd3dbef fails before / passes after
        Full verdict

        REPRODUCED (main, 9.0, 8.6)

        • Platform generic
        • Tested reporter's script on main build, 9.0.5, 8.6.18: apply {{} {tailcall nosuchroutine} ns1} gives "invalid command name" while the same call without tailcall uses ns1's handler.
        • Cause generic/tclBasic.c:TEOV_NotFound picks the unknown handler from varFramePtr->nsPtr. A tailcall runs in the caller's frame (global) with only iPtr->lookupNsPtr set to ns1 (TclNRTailcallEval), so ns1's handler is ignored.
        • Prototype prototypes/f7afd3dbef.patch (20 changed lines, test tailcall-bug-f7afd3dbef). TEOV_NotFound prefers lookupNsPtr's handler when lookupNsPtr is given and has one, then falls back as before. Side effect: the same applies to the other lookupNsPtr users (ensemble -map targets resolved in the ensemble's namespace, oo forwards in the object's namespace, coroutine creation) — arguably also right, but a behaviour change for those. tailcall, namespace(-old), nre, basic, interp, safe, oo, ooNext2, unknown, coroutine, proc, apply, execute, compile tests pass.
        • Suggested ticket comment Still present. TEOV_NotFound() chooses the handler from the current call frame's namespace, but a tailcall only sets the lookup namespace. Proposed fix: prefer the handler of the explicit lookup namespace (test tailcall-bug-f7afd3dbef); note that it also affects ensemble maps and oo forwards to missing commands.
        🔧fc88534d382019duplicate trace with namespace exported procpatchREPRODUCED (main, 9.0, 8.6); works in 8.5.19 and 8.4.20 (regression from the NRE rewrite in 8.6)

        Open ticket fc88534d38 · 46. Traces · filed 2019-11-29

        🔧 Since triage (updated 2026-11-11)

        Status now: Open

        • 🔧 2026-10-17 check-in c8c14f3e58 on imported-cmd-step-traces, not merged: Fix fc88534d38: step and interpreter traces saw an imported command twice, because InvokeImportedNRCmd() re-dispatched to the real command through TclNREvalOb
        • 💬 2026-10-17 serhiy.storchaka: Regression since 8.6 (8.5 traces the call once): InvokeImportedNRCmd() re-dispatches to the real command through TclNREvalObjv(), which runs the step and interpreter traces a second time. Proposed fix in branch imported-cmd-step-traces: an internal flag skips the interpreter traces for that re-dispatch; execution traces on the real command still fire. Tests trace-41.1 and trace-41.2.
        Full verdict

        REPRODUCED (main, 9.0, 8.6); works in 8.5.19 and 8.4.20 (regression from the NRE rewrite in 8.6)

        • Platform generic
        • Tested reporter's script on main build, 9.0.5, 8.6.18 (enterstep and leavestep both doubled for the imported name, single for ::Counter::test); 8.5.19, 8.4.20 single.
        • Cause generic/tclNamesp.c:InvokeImportedNRCmd re-dispatches to the real command through TclNREvalObjv, which runs the interpreter-wide (step) traces again; 8.5 called the real objProc directly.
        • Prototype prototypes/fc88534d38.patch (85 changed lines, ~43 of them code; tests trace-41.1, trace-41.2). New internal eval flag TCL_EVAL_NO_INTERP_TRACES (tclInt.h InterpEvalFlags, 0x80) passed by InvokeImportedNRCmd; TclNREvalObjv/TEOV_RunEnterTraces skip TclCheckInterpTraces with it, and a TEOV_RunLeaveCmdTraces variant skips them on leave. Execution traces set on the original command still fire (trace-41.2; 8.6+ behaviour kept, 8.5 did not fire them). trace, basic, namespace(-old), tailcall, nre, oo, interp, proc, cmdInfo, unknown tests pass. Slightly over the size guideline because of the leave-trace plumbing.
        • Suggested ticket comment Still present since 8.6 (8.5 is fine): an imported command re-dispatches to its target via TclNREvalObjv(), which runs the step/interp traces a second time. Proposed fix: an internal flag that skips the interp traces for that re-dispatch (tests trace-41.1/41.2).
        ⬜c6897e6e6a2019select notifier crashes/misbehaves on fd >= FD_SETSIZEREPRODUCED (main with select notifier, 8.6, 8.5); not on Linux main/9.0 default builds (epoll)

        Open ticket c6897e6e6a · 01. Notifier · filed 2019-11-05

        Full verdict

        REPRODUCED (main with select notifier, 8.6, 8.5); not on Linux main/9.0 default builds (epoll)

        • Platform unix (select notifier: 8.6 everywhere; 9.x on platforms without epoll/kqueue, and Linux builds configured without epoll)
        • Tested reporter's script (600 async client sockets, ulimit -n 4096): 8.6.18 and 8.5.19 abort with "* bit out of range 0 - FD_SETSIZE on fd_set *"; main built with ac_cv_header_sys_epoll_h=no (select notifier, -O2) aborts the same way; main and 9.0.5 default (epoll) builds handle 1201 channels fine.
        • Cause unix/tclSelectNotfy.c (Tcl_CreateFileHandler/DeleteFileHandler and the notifier thread loop) uses FD_SET/FD_ISSET on any fd without checking fd < FD_SETSIZE. sebres' branch bug-c6897e6e6a (2019, last merge-mark 2024-05-22) only rejects such sockets in the socket layer and was never merged.
        • Prototype none (a real fix is replacing select() with poll() in tclSelectNotfy.c, or a policy decision on rejecting fds; both beyond a small patch)
        • Suggested ticket comment Still reproducible with the select notifier: 8.6 and 9.x builds without epoll/kqueue abort with "bit out of range 0 - FD_SETSIZE". Linux 9.x uses epoll and is not affected. Branch bug-c6897e6e6a was never merged; the clean fix would be a poll()-based select notifier.
        ⬜0c7ddf85362019array variable traces don't disable additional traces on the arraypatchREPRODUCED (main, 9.0, 8.6, 8.4)

        Open ticket 0c7ddf8536 · 46. Traces · filed 2019-11-02

        🧪 Prototype check (2026-10-24)

        0c7ddf8536.patch: VERIFIED (behaviour change still needs a core decision, as the triage said)

        • Base main (tip d6dd6d0f71)
        • Before tcltest tests/trace.test -match 'trace-bug-0c7ddf8536*' -> -1 FAILED: {::foo {} read} 1 two {::foo {} read} 1 two {::foo baz read} 1 two {::foo bar read} 1 two (array trace re-fires recursively, the reporter's case); -2 FAILED: {::foo bar read} {::foo baz read} {::foo baz read}
        • Patch applied cleanly (68 changed lines)
        • After all 3 new tests pass; tests trace.test 289/0 failed, var 217/0, upvar 70/0, info 285/0, env 26/0, set 64/0, incr 69/0, append 52/0
        • Test existing trace-bug-0c7ddf8536-1/-2 fail before / pass after (-3, unset traces, passes on both: guards the kept behaviour)
        Full verdict

        REPRODUCED (main, 9.0, 8.6, 8.4)

        • Platform generic
        • Tested reporter's script on main build, 9.0.5, 8.6.18, 8.4.20: the array's read trace fires again for info exists ::foo and for other elements read inside the trace. Unchanged since at least 8.4.
        • Cause generic/tclTrace.c:TclCallVarTraces marks only the element (varPtr) VAR_TRACE_ACTIVE; the existing !TclIsVarTraceActive(arrayPtr) test never matches for element accesses.
        • Prototype prototypes/0c7ddf8536.patch (68 changed lines, 21 code; tests trace-bug-0c7ddf8536-1..3). Marks the containing array active too (not for unset traces, which stay enabled as trace.n says) while the traces run; element-specific traces on other elements still fire. trace, var, upvar, info, env tests pass. Long-standing behaviour, so it may break code that relies on recursive array traces; needs a decision.
        • Suggested ticket comment Confirmed back to 8.4: only the element is marked trace-active, so the array's own traces run again from inside the trace. Proposed fix: mark the array active as well for non-unset traces (tests trace-bug-0c7ddf8536-*); it changes long-standing behaviour, so opinions welcome.
        ⬜ab4307e01b2019changing [encoding system] can corrupt [encoding dirs]patchREPRODUCED (8.6); NOT-REPRODUCED on main, 9.0 and 8.7b1 (fixed as a side effect: zipfs initialisation sets the encoding search path through Tcl_Set…

        Open ticket ab4307e01b · 10. Objects · filed 2019-10-21

        🧪 Prototype check (2026-10-24)

        ab4307e01b.patch: VERIFIED

        • Base core-8-6-branch (tip 326e1bf923); main is not affected (vab43.tcl gives same: 1 on the main tip)
        • Before work-a3/vab43.tcl (encoding system gb12345, read encoding dirs, switch back) on the 8.6 tip -> same: 0 now=#?home#?serhiy#?tcltk/tcl-triage#?wt#?a3#..., and later encoding convertto koi8-r -> unknown encoding "koi8-r"
        • Patch applied cleanly (34 changed lines). Checked the logic against TclSetProcessGlobalValue/TclGetProcessGlobalValue: pgv->value holds native bytes and the per-thread copy is decoded with the system encoding, so bumping the epoch alone is correct.
        • After same: 1 now=wt-a3-86/li..., koi8-r loads. Tests encoding, unixInit, cmdAH, env, fileName, main, basic: 0 failed
        • Test existing encoding-27.3 (from the patch) fails on the unpatched 8.6 tip / passes after
        Full verdict

        REPRODUCED (8.6); NOT-REPRODUCED on main, 9.0 and 8.7b1 (fixed as a side effect: zipfs initialisation sets the encoding search path through Tcl_SetEncodingSearchPath, which stores it with no encoding, so it is never re-converted)

        • Platform generic
        • Tested encoding system gb12345; encoding system utf-8; encoding dirs on main build, 9.0.5 (installed and the tip build with a filesystem TCL_LIBRARY), 8.7b1: unchanged. 8.6.18: garbage with embedded NULs (#?home#?serhiy#?tcltk/8.6/lib/tc^@...), and later encoding loads fail ("unknown encoding").
        • Cause generic/tclUtil.c:TclGetProcessGlobalValue. On a system encoding change it runs Tcl_UtfToExternal(old) and then Tcl_ExternalToUtf(new) on pgvPtr->value, which already holds native bytes, so the conversions go the wrong way. With a non-ASCII-compatible encoding such as gb12345 (double-byte only) this is lossy, and switching back cannot restore the value. The same code is still on main/9.0, but there it is only reached for PGVs whose initializer supplies an encoding (libraryPath, the initial encodingSearchPath without zipfs).
        • Prototype prototypes/ab4307e01b.patch, against core-8-6-branch (34 changed lines, test encoding-27.3, run in a subprocess so it sees the initial value). Keeps the native bytes and only starts a new epoch, so per-thread copies are decoded again with the new system encoding. While gb12345 is the system encoding the dirs read as gb12345 text, as native bytes should; after switching back they are exact again. The test fails without the fix; encoding, unixInit, cmdAH, env, fileName tests pass. The same change applies to main/9.0 as a latent fix.
        • Suggested ticket comment Reproducible in 8.6 only; in 8.7/9.x the search path is set via Tcl_SetEncodingSearchPath() at init and is not re-converted. Cause: TclGetProcessGlobalValue() converts the stored native bytes as if they were UTF-8. Proposed fix for 8.6: keep the native bytes and only bump the epoch (test encoding-27.3).
        ⬜f7ac7469622019minor improvement to Tcl man page.REPRODUCED (main HTML generator)

        Open ticket f7ac746962 · 69. Other · filed 2019-10-03

        Full verdict

        REPRODUCED (main HTML generator)

        • Platform docs (tools/tcltk-man2html-utils.tcl)
        • Tested ran main tools/tcltk-man2html.tcl on main doc/. In TclCmd/Tcl.html rule [8], the last two paragraphs ("There may be any number..." and "Note that variables may contain...") come out after </dl> wrapped in a new <dl><dd>, one level deeper than the rule's own text. groff renders them at the rule's body indent (doc/Tcl.n is correct: .RS / .TP list / .PP / .PP / .RE).
        • Cause tools/tcltk-man2html-utils.tcl output-IP-list/output-RS-list: when a .TP list inside .RS ends at .PP, the remaining .RS content is put in a fresh <dl><dd> instead of continuing at the enclosing .IP body level.
        • Prototype none (the .RS/.IP handling in man2html is heuristic, and changing it affects every page; it needs checking across the whole doc set)
        • Suggested ticket comment Still true in the HTML from tcltk-man2html (groff output is correct). The paragraphs that follow a .TP list inside .RS ... .RE get an extra <dl><dd> level.
        ⬜7bdd1d62e220198.6.10 testinghttp-4.14 REPRODUCED (main on NetBSD 10.1, every run; 8.6 passes), proposed fix in branch http-test-bad-port (de7efe2cc2, ticket 5350f32153, no…

        Open ticket 7bdd1d62e2 · - New Builtin Commands · filed 2019-10-01

        Full verdict

        http-4.14 REPRODUCED (main on NetBSD 10.1, every run; 8.6 passes), proposed fix in branch http-test-bad-port (de7efe2cc2, ticket 5350f32153, not merged). cmdAH-24.20: NOT-REPRODUCED (main, 8.6 on NetBSD). Other parts (macOS, AIX) CANNOT-TEST.

        • Platform generic (test-suite report: Linux, macOS, AIX, NetBSD)
        • Tested http-4.14* passes on main build and 8.6 tip tcltest (Linux). http.test now uses HOST localhost (comment mentions http-4.14 name-resolution problems). cmdAH-24.20.1/2 (NetBSD) are now guarded by the filetime64bit constraint (file system 2038 check, fd91b0ca09). socket_inet-5.1/5.3 (Mojave) and AIX socket-14.* not checked; attachments are logs only.
        • Cause http-4.14 depended on host name resolution; cmdAH-24.20 on file system timestamp range.
        • Prototype none
        • Suggested ticket comment This is an 8.6.10 release-testing report. http-4.14 now uses localhost and the cmdAH-24.20 tests are skipped on file systems without 64-bit timestamps. Any failures still seen with 8.6.18/9.0 should go in their own tickets. Suggest closing.
        Unix VMs, run 1
        • Platform generic (test-suite report)
        • Verdict http-4.14 REPRODUCED (main on NetBSD 10.1, every run; 8.6 passes), proposed fix in branch http-test-bad-port (de7efe2cc2, ticket 5350f32153, not merged). cmdAH-24.20: NOT-REPRODUCED (main, 8.6 on NetBSD). Other parts (macOS, AIX) CANNOT-TEST.
        • Tested NetBSD 10.1 amd64 (FFS root). main: cmdAH.test 17001/0 failed, socket.test 191/0, http.test 1 failed: http-4.14.0 (ThreadLevel 0) returns "" instead of "connect failed: connection refused", 4 of 4 runs. 8.6: cmdAH 323/0, http 75/0, socket 175/0. Debug run: NetBSD gives out ephemeral ports in descending order (65195, 65194, ...), so $port+1 is the listener of the httpd started by the outer (non-ThreadLevel) run of http.test, still open, and the "bad" URL connects. With the de7efe2cc2 diff applied, http.test on main passes (154/0 failed). FreeBSD 15.1 (random ports) passes http.test on main/9.0/8.6.
        • Cause tests/http.test: set badurl //${::HOST}:[expr {$port+1}] assumes port+1 is free.
        • Prototype none new (branch http-test-bad-port, verified on NetBSD)
        • Suggested ticket comment The NetBSD http-4.14 failure still happens on main: ports are allocated in sequence there, so $port+1 belongs to the outer run's httpd. Proposed fix in branch http-test-bad-port (de7efe2cc2, see 5350f32153), verified on NetBSD 10.1. cmdAH-24.20 passes on NetBSD now. Other failures should get their own tickets; suggest closing this one after the merge.
        ✅8228750a0a2019Makefile issue when building packagesREPRODUCED (main, 9.0, 8.6) — code unchanged, but the order looks intentional

        Open ticket 8228750a0a · 69. Other · filed 2019-08-13

        ✅ Since triage (updated 2026-11-11)

        Status now: Closed (Duplicate)

        • 🔧 2026-10-20 check-in 9d87df83e9 on pkgs-respect-disable-shared, not merged: Fix 9cdb0d5858 and 8228750a0a: with --disable-shared the bundled packages were still built as shared libraries, since --enable-shared came after the argumen
        • 🔧 2026-10-20 check-in 732583df8c on pkgs-shared-make-var, not merged: Fix 9cdb0d5858 and 8228750a0a: the bundled packages were always built as shared libraries, even with --disable-shared. Keep that default, so that [package r
        • 💬 2026-10-20 serhiy.storchaka: Same as 9cdb0d5858, where two proposed fixes are described: pkgs-respect-disable-shared and pkgs-shared-make-var.
        Full verdict

        REPRODUCED (main, 9.0, 8.6) — code unchanged, but the order looks intentional

        • Platform build
        • Tested read unix/Makefile.in configure-packages on main, 9.0, 8.6 (no build needed)
        • Cause unix/Makefile.in configure-packages passes $(PKG_CFG_ARGS) --libdir=... --enable-shared (8.6 also --enable-threads), so the forced --enable-shared always wins over anything in PKG_CFG_ARGS. PKG_CFG_ARGS is Tcl's own $ac_configure_args plus the PKG_CFG_ARGS environment variable (unix/configure.ac:39), so simply moving it last (as requested) would also make a Tcl configured with --disable-shared build static bundled packages, which package require from pkgs/ cannot load (pkgIndex.tcl uses load on a shared lib). 9.0 has the same in both the pkgs8 and pkgs loops.
        • Prototype none (needs a decision: e.g. append only the user's environment PKG_CFG_ARGS after --enable-shared, keeping Tcl's own args before it; or tell users to build the static package outside the Tcl tree)
        • Suggested ticket comment Still the same in main/9.0/8.6. --enable-shared is forced on purpose because bundled packages are loaded from shared libraries; a static itcl is better built separately with its own configure. If wanted, only the user-supplied PKG_CFG_ARGS environment value could be placed after --enable-shared.
        🔧7b51dbf6ca2019Code using namespace exported commands not regenerated correctlyREPRODUCED (main, 9.0, 8.6, also 8.6.0)

        Open ticket 7b51dbf6ca · 47. Bytecode Compiler · filed 2019-05-23

        🔧 Since triage (updated 2026-11-11)

        Status now: Open

        • 🔧 2026-10-14 check-in 866a927c0c on compile-imported-cmd-redefined, not merged: Fix 7b51dbf6ca: compiled code calling an imported command kept the old compiled form after the original command was redefined or replaced, since the imported
        Full verdict

        REPRODUCED (main, 9.0, 8.6, also 8.6.0)

        • Platform generic
        • Tested reporter's script (disassembly still push1 "" after redefinition) and sebres' tcl::dict::get variant on main (2026-10-22), 9.0.5, 8.6.18, 8.6.0; branch tip 866a927c0c: both variants call the new definition, namespace.test 319/0 failed, compile.test 171/0 failed
        • Cause generic/tclNamesp.c: an imported command copies the compileProc of the original at import time and keeps it after the original is redefined (Serhiy's comment)
        • Existing work branch compile-imported-cmd-redefined (tip 866a927c0c, 2026-10-14), unmerged (not in trunk/9.0/8.6); dgp asked for care about import vs alias semantics, Serhiy answered 2026-10-14 (tcllib control::assert checked)
        • Prototype none (branch exists)
        • Suggested ticket comment none needed now (discussion is current; waiting for dgp's/Jan's review).
        ✅67fd4f973a2019incorrect results of 'info exists' when unset env var in one interp and check for existence from another interppatchREPRODUCED (main, 9.0, 8.6)

        Open ticket 67fd4f973a · 17. Commands I-L · filed 2019-05-20

        ✅ Since triage (updated 2026-11-11)

        Status now: Closed (Duplicate)

        • 🔧 2026-10-19 check-in f1438aeb70 on env-stale-element, in main: Fix 3602289fff, 67fd4f973a and 845241ffff: after another interp unset an env variable, a read failed but left the old element, so [info exists ::env(X)] s
        • 💬 2026-10-19 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1. The env read trace returns an error for a name no longer in the environment, but keeps the stale element, so [info exists] still sees it. Proposed fix in branch env-stale-element, together with 3602289fff and 845241ffff: the read trace removes the stale element. Test env-11.3.
        • 💬 2026-10-20 jan.nijtmans: Dup of 845241ffff
        Full verdict

        REPRODUCED (main, 9.0, 8.6)

        • Platform generic
        • Tested reporter's script on main build, 9.0.5, 8.6.18: after unset ::env(X) in one interp, the other interp's info exists ::env(X) returns 1 while reading fails.
        • Cause generic/tclEnv.c EnvTraceProc: on a read of a name no longer in environ it returns "no such variable" but leaves the stale element of this interp's env array defined; TclVarTraceExists (tclTrace.c) and the bytecode info exists ignore the trace error and see a defined element.
        • Prototype prototypes/67fd4f973a.patch (29 changed lines, test env-12.0): in the read trace, when getenv fails and the element still has a value, unset it (elements without a value are kept, since they may hold traces, see env-7.3). env/var/trace/info/interp.test pass; env-12.0 fails without the fix.
        • Side issue (not fixed): unset ::env(X) in a child for a name set by the parent but never read in the child gives "no such element in array" (the child's array is only synced on array traces).
        • Suggested ticket comment The env read trace in tclEnv.c returns an error for a name no longer in the environment but keeps the stale element, so info exists still sees it. Proposed fix: unset the stale element in the read trace; test env-12.0.
        ⬜c6f1ffa7a92019clock scan/freescan deviation (wrong conversion) if TZ specified in the inputREPRODUCED (main, 9.0, 8.6) in part — the format/free-scan mismatch is gone in 9.0+; the rest is a design question

        Open ticket c6f1ffa7a9 · 16. Commands A-H · filed 2019-04-29

        Full verdict

        REPRODUCED (main, 9.0, 8.6) in part — the format/free-scan mismatch is gone in 9.0+; the rest is a design question

        • Platform generic
        • Tested sebres' snippets on main build, 9.0.5, 8.6.18.
          • clock scan {10:00 GMT} -format {%H:%M %Z} -timezone :America/New_York -base 1556240400: 8.6 gives Apr 26 (date of base taken in the input zone), 9.0/main give Apr 25 like free scan (date of base taken in the -timezone zone). Format scan and free scan now agree (9.0 clock rewrite).
          • Free scan clock scan CET -base 100000000 -gmt 1 still keeps the base time of day in GMT and reads it as CET (99996400 = base − 1h) on all versions, while format scan gives midnight. This is the documented "no time given" difference between free and format scan, plus the open question which zone wins for the base parts. The fix proposed in the ticket (input zone wins) is not merged.
          • -timezone GMT+05 is POSIX TZ syntax (5 h west of UTC), so its opposite sign from +05 is correct, not a bug.
        • Cause library/clock.tcl / generic/tclClock.c take the missing fields from the base in the -timezone (or current) zone and apply the zone from the input only to the parsed fields.
        • Prototype none (precedence change, behaviour redesign)
        • Suggested ticket comment Since 9.0 format scan and free scan agree (date from base is taken in the -timezone zone). Whether a zone given in the input should win for the parts taken from -base is a design decision still open; GMT+05 vs +05 is POSIX TZ sign convention and works as documented.
        ⬜a8e4f76ce42019load library (dll) from zipfs-library causes a leak in temporary folderREPRODUCED (main, 9.0)

        Open ticket a8e4f76ce4 · 38. Init - Library - Autoload · filed 2019-04-11

        Full verdict

        REPRODUCED (main, 9.0)

        • Platform Windows
        • Tested Linux main build: loading dltest/tcl9pkga.so from a mounted zip leaves nothing in TMPDIR, because generic/tclIOUtil.c TclLoadFile deletes the temp copy right after loading (Unix allows unlinking a loaded file). On Windows the DLL is locked, so the TCL* temp dir stays.
        • Cause generic/tclIOUtil.c TclLoadFile copies the DLL to a temp file; Windows can't delete a loaded DLL. The registry/dde case (6ce3c0de9d, closed) has a workaround (364bd985d6). For other DLLs nothing changed: Jan's MemoryModule branch (b74bd939ad/4e88af7df7) and Ashok's apn-bug-a8e4f76ce4 (delete on exit, one check-in 4317ebe71d, 2024-11-26) are both unmerged; the comments say it needs a TIP.
        • Prototype none (design disputed, needs a TIP)
        • Repro win/a8e4f76ce4.tcl ?ext.dll? (zips a DLL, loads it from zipfs in a child tclsh, checks %TEMP% for a new TCL* dir); verified on Linux (NOT-REPRODUCED there, as expected)
        • Suggested ticket comment Still open on Windows for DLLs other than registry/dde (those are worked around by 364bd985d6). Unix is not affected because the temp copy is unlinked right after loading. Both proposals (MemoryModule, apn-bug-a8e4f76ce4) are unmerged and need a TIP.
        Windows VM, run 3
        • Platform Windows
        • Verdict REPRODUCED (main, 9.0)
        • Tested VM win/a8e4f76ce4.tcl: a child tclsh mounts a zip holding an extension DLL and loads it; after the child exits a new %TEMP%\TCLxxxxxxxx directory with the DLL copy remains. Thread 3.0.7 DLL on 9.1b1 and 9.0.5, tcl9dde14.dll on 9.0.5: left behind every time (the script deletes them afterwards). The registry/dde workaround 364bd985d6 covers only Tcl's own packages, not a DLL loaded from an application zip.
        • Cause generic/tclIOUtil.c TclLoadFile copies the DLL to a temp dir and cannot delete it on Windows while it is mapped; nothing removes it at exit
        • Prototype none (MemoryModule and apn-bug-a8e4f76ce4 are unmerged; needs a TIP, per the comments)
        • Suggested ticket comment Still happens on Windows 11 with 9.0.5 and 9.1: every load of a DLL from a mounted zip leaves a TCLxxxxxxxx directory in %TEMP%. The open proposals (MemoryModule, apn-bug-a8e4f76ce4) are unmerged.
        ⬜c03621bf9b2019[clock scan] this modifier, againREPRODUCED (main, 9.0, 8.6): this day/this week/this month are syntax errors everywhere. 8.6: thiss day = now day = start of tomorrow, `l…

        Open ticket c03621bf9b · 06. Time Measurement · filed 2019-03-28

        Full verdict

        REPRODUCED (main, 9.0, 8.6): this day/this week/this month are syntax errors everywhere. 8.6: thiss day = now day = start of tomorrow, last January = Dec 30 of the previous year. main/9.0: thiss day and now day are now syntax errors, last January fails with "invalid day".

        • Platform generic
        • Tested main c3bf6dd052, 9.0.5, 8.6.18 (work-a6/t5.tcl, base 1554818068)
        • Cause generic/tclGetDate.y OtherTable has { "this", tSEC_UNIT, 0 }, but LookupWord strips the plural "s" first, so "this" never matches; next/last are tNEXT / tUNUMBER(-1) with different month semantics (sebres' comment).
        • Prototype none (needs a decision on what "this"/"last <month>" should mean; sebres planned a broader rework)
        • Suggested ticket comment Still present in 9.x: "this" is unreachable (the lexer strips the "s"), and last January now fails with "invalid day" instead of returning Dec 30. Either drop "this" from OtherTable and the docs, or define it.
        ⬜b43efdfd612019file delete -force does not work on read-only subdirectoriespatchREPRODUCED (main, 9.0, 8.6, 8.4)

        Open ticket b43efdfd61 · 37. File System · filed 2019-03-16

        🧪 Prototype check (2026-10-24)

        b43efdfd61.patch: VERIFIED

        • Base main (tip d6dd6d0f71)
        • Before tcltest b43e.tcl (reporter's script without fileutil: subdir chmod a-w, file delete -force $dir) -> error deleting ".../subdir/testfile": permission denied, dir still exists
        • Patch applied (fCmd.test offset -9 lines); 40 changed lines
        • After -> 0, exists: 0; tests fCmd.test 261 passed/45 skipped/0 failed, unixFCmd.test 30/19/0, fileSystem.test 141/21/0, cmdAH.test 16971/30/0. The default build uses FTS (HAVE_FTS=1), so I also built the opendir path (out-of-tree, tcl_cv_api_fts=no): the repro is fixed and fCmd.test 261/45/0 there too
        • Test existing fCmd-20.3 fails before / passes after
        Full verdict

        REPRODUCED (main, 9.0, 8.6, 8.4)

        • Platform unix
        • Tested main (30d8ccce48), 9.0.5, 8.6.18, 8.4.20: error deleting ".../subdir/testfile": permission denied. A read-only top-level directory is deleted fine.
        • Cause unix/tclUnixFCmd.c: DoRemoveDirectory adds u+rwx only to the directory passed to file delete; TraversalDelete does nothing for DOTREE_PRED, so read-only subdirectories keep their mode and unlink() inside them fails. file.n says -force "will attempt ... to change permissions".
        • Prototype prototypes/b43efdfd61.patch (27 changed lines, test fCmd-20.3). TraversalDelete adds u+wx on pre-order directories. It deliberately does not add u+r, so an unreadable subdirectory still fails (fCmd-20.1 unchanged, and the FTS and opendir code paths stay consistent, since the opendir path opens the directory before the pre-order callback). fCmd.test passes; fCmd-20.3 fails without the fix.
        • Suggested ticket comment Still in 9.1/9.0/8.6: only the top directory gets its permissions changed, subdirectories don't. Proposed fix: TraversalDelete makes each subdirectory owner-writable/searchable before deleting its contents (test fCmd-20.3).
        🔧055240025a2019Calling Tcl_DeleteEventSource from inside Tcl_DoOneEvent segfaultspatchREPRODUCED (main; 9.0, 8.6 by code reading, identical loops)

        Open ticket 055240025a · 04. Async Events · filed 2019-02-06

        🔧 Since triage (updated 2026-11-11)

        Status now: Open

        • 🔧 2026-10-18 check-in 97243eeba2 on notifier-delete-source-in-loop, not merged: Fix 055240025a: Tcl_DeleteEventSource called from a setup or check proc freed the event source while Tcl_DoOneEvent or Tcl_ServiceAll was walking the list. Du
        • 💬 2026-10-18 serhiy.storchaka: Still present in 9.1 (9.0 and 8.6 have the same loops): Tcl_DeleteEventSource() frees the source at once, while Tcl_DoOneEvent() or Tcl_ServiceAll() continue the traversal through its nextPtr, and a deleted source may still have its procs called. Proposed fix in branch notifier-delete-source-in-loop: during a traversal the source is only marked deleted, and the outermost traversal frees it. Tests notify-3.1 to 3.4…
        Full verdict

        REPRODUCED (main; 9.0, 8.6 by code reading, identical loops)

        • Platform generic
        • Tested small extension (work-a4/b22/evsrc.c + evsrc.tcl) whose checkProc (and, second case, setupProc) deletes its own event source; main -DPURIFY build under valgrind: "Invalid read of size 8 at Tcl_DoOneEvent (tclNotify.c:1054)", reading sourcePtr->nextPtr of the freed entry.
        • Cause generic/tclNotify.c:Tcl_DeleteEventSource frees the EventSource immediately, while Tcl_DoOneEvent and Tcl_ServiceAll are walking the list and continue via sourcePtr->nextPtr.
        • Prototype prototypes/055240025a.patch (47 changed lines, no test: tcltest has no command creating event sources). Adds a traversal depth counter; during a traversal Tcl_DeleteEventSource only NULLs both procs, and the outermost traversal unlinks and frees such entries. valgrind clean on the repro; event/notify/timer/unixNotfy/socket/io/thread tests pass.
        • Suggested ticket comment Confirmed on 9.1: Tcl_DeleteEventSource() called from a setup/check proc frees the entry Tcl_DoOneEvent() is iterating over (valgrind: invalid read in Tcl_DoOneEvent). Proposed fix: mark the source deleted during the traversal and free it afterwards.
        🔧82049f96f82018file copy attempts to copy directory into itselfpatchREPRODUCED (main, 9.0, 8.6)

        Open ticket 82049f96f8 · 37. File System · filed 2018-11-21

        🔧 Since triage (updated 2026-11-11)

        Status now: Open

        • 🔧 2026-10-19 check-in c51f96d014 on copy-into-itself, not merged: Fix 82049f96f8 and 875545270d: [file copy] of a directory into itself recursed until the path was too long; now it is an error. tcl::CopyDirectory, used for
        • 💬 2026-10-17 serhiy.storchaka: The deep tree that [file copy 1 1/2] leaves behind cannot be removed with [file delete -force] either: daba77ae1d.
        • 💬 2026-10-19 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1: [file copy 1 1/2] (and [file copy -force d d]) copies until the path is too long, and leaves a tree that [file delete -force] cannot remove (daba77ae1d). Proposed fix in branch copy-into-itself, together with 875545270d: copying a directory into itself is now an error, "trying to copy a directory into itself". The normalized paths are compared, so links are handled, and so is…
        • 💬 2026-10-21 dgp: Just want to add a caution that Tcl's virtual filesystems and ability to mount anywhere undermines a lot of "common sense" reasoning about the relationship between directories. Be sure to consider all the possibilities and not limit your analysis to the assumption that all paths are inside a single filesystem.
        Full verdict

        REPRODUCED (main, 9.0, 8.6)

        • Platform generic
        • Tested main (30d8ccce48), 9.0.5, 8.6.18: file copy 1 1/2 recurses until "file name too long" and leaves a deep 1/2/1/2/... tree behind. file copy -force d d (target becomes d/d) does the same. Side finding: file delete -force then cannot remove that tree either ("file name too long", path > PATH_MAX); only rm -rf could.
        • Cause generic/tclFCmd.c:CopyRenameOneFile has no check that the target lies inside the source directory (rename gets EINVAL from rename(2), copy has nothing similar).
        • Prototype prototypes/82049f96f8.patch (79 changed lines incl. 20 test lines; tests fCmd-21.13, fCmd-21.14). New static IsSubdirectory() compares normalized paths (prefix + '/'). It reports only when the target's parent exists, so the "no such file or directory" error is kept (fCmd-8.3 copies [file home] to a missing path from a cwd inside home). Error: error copying "1" to "1/2/1": trying to copy a directory into itself. fCmd, cmdAH, fileSystem tests pass.
        • Suggested ticket comment Still in 9.1/9.0/8.6, and the runaway copy also can't be removed by file delete -force. Proposed fix: CopyRenameOneFile refuses to copy a directory into itself, comparing normalized paths (tests fCmd-21.13/21.14).
        ⬜469c5ea5072018VFS breaks when [pwd] is non-existent native directorypatchREPRODUCED (main, 9.0, 8.6, 8.5)

        Open ticket 469c5ea507 · 37. File System · filed 2018-11-08

        🧪 Prototype check (2026-10-24)

        469c5ea507.patch: VERIFIED

        • Base main (tip d6dd6d0f71)
        • Before work-a4/v4gone.tcl run by tcltest started in a directory removed before exec: file delete spong -> error deleting "spong": bad address; file mkdir spong -> ... bad address; file rename a b, file copy a b, file lstat spong s -> bad address
        • Patch applied cleanly (fCmd.test hunk at offset -9; 65 changed lines)
        • After file delete spong -> 0 (ok, like a missing file), mkdir/rename/copy/lstat -> no such file or directory; make test TESTFLAGS="-file 'fCmd.test cmdAH.test fileSystem.test fileName.test unixFCmd.test'" -> 17824 total, 17662 passed, 162 skipped, 0 failed
        • Test existing fCmd-33.1 fails before (Result was error deleting "spong": bad address) / passes after
        Full verdict

        REPRODUCED (main, 9.0, 8.6, 8.5)

        • Platform unix
        • Tested tclsh started in a removed directory: main (30d8ccce48), 9.0.5, 8.6.18, 8.5.19: file delete spong -> "error deleting "spong": bad address", file mkdir x -> "bad address". (If the dir is removed after a Tcl cd, the cached cwd hides the delete case but mkdir still fails.)
        • Cause relative paths are normalized against getcwd(); when it fails, Tcl_FSGetNativePath() returns NULL and several unix wrappers pass that NULL straight to the syscall (EFAULT): unix/tclUnixFile.c:TclpObjLstat, TclpUtime; unix/tclUnixFCmd.c:TclpObjDeleteFile, TclpObjCreateDirectory, TclpObjRenameFile, TclpObjCopyFile, and the chown/chmod attribute setters. TclpObjStat/TclpObjAccess/TclpObjChdir already check for NULL.
        • Prototype prototypes/469c5ea507.patch (65 changed lines incl. 22 test lines, test fCmd-33.1). Adds the NULL checks, so errno from getcwd (ENOENT) is reported, and file delete of a missing relative file succeeds as usual. fCmd/cmdAH/fileSystem pass; fCmd-33.1 fails without the fix. Making relative paths work without a cwd at all would be a bigger VFS change; not attempted.
        • Suggested ticket comment Still in 9.1/9.0/8.6: several native FS wrappers pass a NULL native path to lstat/unlink/mkdir/... when the cwd can't be determined. Proposed fix: check for NULL like TclpObjStat already does (test fCmd-33.1); errors now say "no such file or directory".
        ⏳723a2f4ac32018return -code break/tailcall break from coroutine goes wrong at the top-levelpatchREPRODUCED (main, 9.0, 8.6) for the coroutine cases. The ticket is Pending/Fixed, but neither fix branch was ever merged. The tailcall cases (`proc…

        Open ticket 723a2f4ac3 · 60. NRE and coroutines · filed 2018-07-31 · Pending (resolved, waiting for confirmation)

        ⏳ Since triage (updated 2026-11-11)

        Status now: Pending (Fixed)

          Full verdict

          REPRODUCED (main, 9.0, 8.6) for the coroutine cases. The ticket is Pending/Fixed, but neither fix branch was ever merged. The tailcall cases (proc p {} {tailcall break}; while 1 p, and the tailcall inside a coroutine) now work on all three.

          • Platform generic
          • Tested main, 9.0, 8.6 as top-level scripts: proc p0 {} {yield; return -code break}; coroutine c0 p0; while 1 c0 and the coroutine c0 try {yield; break} from itop.tcl both fail with invoked "break" outside of a loop.
          • Cause generic/tclBasic.c:TclNRCoroutineObjCmd evaluates the body without Tcl_AllowExceptions(), so at numLevels 0 the break is turned into an error when it comes back through a resume. Branches bug-723a2f4ac3 (dgp, e81fd56959, the same one-liner) and sebres-bug-723a2f4ac3 (34526b0071) have no merge children, so they were never merged.
          • Prototype prototypes/723a2f4ac3.patch (15 changed lines: dgp's one-line Tcl_AllowExceptions() before Tcl_NREvalObj in TclNRCoroutineObjCmd, plus test coroutine-bug-723a2f4ac3, run in a subprocess because it must be at top level). The test fails without the fix. coroutine, nre, tailcall, basic and proc .test pass. interp.test fails only interp-32.1, which also fails without the patch (cwd-related).
          • Suggested ticket comment Still reproduces on 9.1/9.0/8.6 for the coroutine cases (the tailcall ones are fixed). Neither bug-723a2f4ac3 nor sebres-bug-723a2f4ac3 was ever merged. dgp's one-line Tcl_AllowExceptions() in TclNRCoroutineObjCmd still fixes it on main; proposed fix with test coroutine-bug-723a2f4ac3. Status should go back to Open.
          ⬜29aa1c7b132018Mixed-in destructors aren't calledpatchREPRODUCED (main, 9.0, 8.6)

          Open ticket 29aa1c7b13 · 10. Objects · filed 2018-07-17

          🧪 Prototype check (2026-10-24)

          29aa1c7b13.patch: VERIFIED

          • Base main (tip d6dd6d0f71)
          • Before the ticket script (work-a3/v29aa.tcl) -> a: a b: a b c: c (A's destructor, mixed into object $c, is not called)
          • Patch applied cleanly (28 changed lines)
          • After same -> a: a b: a b c: a c; tests oo (374/0 failed), ooNext2, ooProp, ooUtil: 0 failed
          • Test existing oo-3.13 (from the patch) fails on the unpatched tip / passes after
          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform generic
          • Tested main, 9.0, 8.6. Destructors of a class mixed into a class are called; destructors of a class mixed into an object (oo::objdefine obj mixin A) are never called, also for oo::object instances.
          • Cause generic/tclOOCall.c:AddSimpleChainToCallContext skips oPtr->mixins under if (!(flags & SPECIAL)), which covers DESTRUCTOR as well as CONSTRUCTOR. TclOOGetCallContext already avoids caching the destructor chain when oPtr->mixins.num != 0 (since 2008), so object mixins were meant to take part.
          • Prototype prototypes/29aa1c7b13.patch (28 changed lines, test oo-3.13): walk object mixins for DESTRUCTOR too, skipping the private/instance-method lookups that need a method name. The reporter's script now prints c: a c. oo, ooNext2, ooProp and ooUtil .test pass. As with class mixins, a mixin destructor that calls next with nothing after it raises "no next destructor implementation".
          • Suggested ticket comment Object-level mixins are skipped when building the destructor chain (AddSimpleChainToCallContext skips them for all SPECIAL chains), although the destructor-chain cache already allows for them. Proposed fix with test oo-3.13. Constructors stay as they are, since an object mixin is normally added after construction.
          ✅815e2468062018tcl crash on channel thread transfer and fileevent writable in placeREPRODUCED (main); fixed by Serhiy's branch

          Open ticket 815e246806 · 49. Threading · filed 2018-06-08

          ✅ Since triage (updated 2026-11-11)

          Status now: Closed (Fixed)

          • 🔧 2026-10-13 check-in bb3d0c3db9 on channel-refcount-thread-safe, in main: Fix 815e246806: transferring a channel to another thread from its own file event handler could panic with "Channel released more than preserved": both threads
          • 💬 2026-10-30 jan.nijtmans: Fixed in f65229e7a7, core-9-0-branch and core-8-6-branch Many thanks!
          Full verdict

          REPRODUCED (main); fixed by Serhiy's branch

          • Platform generic
          • Tested Linux, my main build (c092dae4af) + Thread 3.0.7 trunk: the description script, 40 runs x 10 s, 4 in parallel: 19/40 abort with "Channel released more than preserved", 21/40 survive. Branch channel-refcount-thread-safe (bb3d0c3db9), same conditions: 40/40 survive. thread.test (incl. thread-9.1), io.test, ioCmd.test, chan.test, socket.test pass on the branch. (9.0/8.6 not re-run; Serhiy's comment covers them.)
          • Cause generic/tclIO.c TclChannelPreserve/TclChannelRelease: channel refCount is changed by both threads (sender's NotifyChannel/event script vs. receiver after the transfer) without synchronisation.
          • Existing work channel-refcount-thread-safe (Serhiy, 1 check-in bb3d0c3db9 2026-10-13, branched from trunk 88c6ad4c75, unmerged; 51 changed lines in tclIO.c + thread.test). Older fix-815e246806-8-6 / -8-5 (sebres 2018, atomics) never merged.
          • Prototype none (Serhiy's branch exists)
          • Suggested ticket comment none needed (Serhiy already posted "Proposed fix in branch"); branch is ready for review/merge; needs backports to core-9-0-branch and core-8-6-branch.
          ✅75ecc0e6cb2018TclInitSubsystems issues: the sporadic lack of encodingspatchREPRODUCED (8.6; main only with --disable-zipfs, a new regression); 9.0 OK (fixed by 17c53b7edd, 2024-08-11, for fccb9f322f, in 9.0.0)

          Open ticket 75ecc0e6cb · 38. Init - Library - Autoload · filed 2018-05-30

          ✅ Since triage (updated 2026-11-11)

          Status now: Closed (Fixed)

          • 🔧 2026-10-16 check-in cbc900f06b on init-encoding-dirs, not merged: Fix 75ecc0e6cb: add $tcl_library/encoding to the encoding search path again, lost when TIP 732 moved the auto_path code from init.tcl to C
          • 🔧 2026-11-06 check-in 87da0a5db8 on apn-bug-75ecc0e6cb, in main: Create new branch named "apn-bug-75ecc0e6cb"
          • 🔧 2026-11-06 check-in e743ee1843 on apn-bug-75ecc0e6cb, in main: Ticket 75ecc0e6cb - non-zipfs builds with relocated library misses encodings
          • 💬 2026-10-16 serhiy.storchaka: The system-encoding part was fixed in 9.0 by 17c53b7edd; 8.6 still shows it for an uninstalled or relocated tclsh. On main, TIP 732 dropped the [encoding dirs] addition from init.tcl, so a --disable-zipfs tclsh with a relocated library has no encoding directories at all (9.1.0). Proposed fix in branch init-encoding-dirs, test encoding-26.1.
          • 💬 2026-10-16 dgp: I thought one of the intents of TIP 732 was to simplify and reduce the search paths. Before growing them again, someone involved in TIP 732 should comment.
          • 💬 2026-10-17 serhiy.storchaka: It does not grow the search path, it restores what 9.0 does: init.tcl adds [file join $::tcl_library encoding] to [encoding dirs] if it is not there. TIP 732 (83e88416ac) moved that part of init.tcl to InitAutoPathObjCmd(), whose comment still says "The function also adds the encoding subdirectory of tcl_library to the encodings search path if not already present", but the code was not moved. So in 9.1 a tclsh w…
          • 💬 2026-10-17 apnadkarni: I have to retrace but will need some time before I can look at it. I vaguely recall it was supposed to be done elsewhere (which would mean the comment was not fixed) but I could be wrong.
          • 💬 2026-11-06 apnadkarni: Proposed fix in branch [apn-bug-75ecc0e6cb](75ecc0e6cb">https://core.tcl-lang.org/tcl/timeline?r=apn-bug-75ecc0e6cb). Implementation differs from the init-encoding-dirs branch. Encoding directory should be added irrespective of whether InitAutoPath is called by a (custom) init.tcl or not so it's called from LocatePreInitScript which sets tcl_library. Implementation is shared with the zipfs enabled case. Question, in cas…
          • 💬 2026-11-07 apnadkarni: Merged into 9.1 in 1bc3a58eac. Not applicable to 9.0. Please re-open if you still see an issue.
          Full verdict

          REPRODUCED (8.6; main only with --disable-zipfs, a new regression); 9.0 OK (fixed by 17c53b7edd, 2024-08-11, for fccb9f322f, in 9.0.0)

          • Platform generic (seen on unix)
          • Tested LANG=japanese / ru_RU.KOI8-R, no TCL_LIBRARY, with a tclsh whose compiled-in library dir does not exist:
            • 8.6 tip (out-of-tree build with a dummy prefix, run as X/unix/tclsh with X/library): encoding system is iso8859-1 instead of euc-jp/koi8-r; encoding system euc-jp later works. Installed 8.4–8.6.18, 9.0.5, 9.1 and uninstalled zipfs builds of main/9.0 are fine.
            • main (own build) with --disable-zipfs, run as X/bin/tclsh with X/lib/tcl9.1: encoding dirs is EMPTY, the system encoding is iso8859-1 and only the 13 built-in encodings exist (encoding system euc-jp fails: unknown encoding).
          • Cause (1) 8.6: TclpSetInitialEncodings runs in TclInitSubsystems before init.tcl has found tcl_library, so the encoding search path holds only the compiled-in or TCL_LIBRARY dirs; init.tcl adds $tcl_library/encoding later, but the system encoding is not set again. 9.0 calls TclpSetInitialEncodings() again at the end of Tcl_Init (17c53b7edd). Branch fix-initsubsystem-75ecc0e6cb9899bf (rebased to 8.6 in 2022, last check-in c88e1bbfbf, 2022-12-01) was never merged. (2) main: TIP 732 (83e88416ac on apn-init-refactor, merged e89c38915e, 2026-01-09, so in 9.1.0) moved the auto_path part of init.tcl to C (InitAutoPathObjCmd in generic/tclInterp.c) but dropped the encoding dirs addition of $tcl_library/encoding, though the function comment still describes it.
          • Prototype prototypes/75ecc0e6cb.patch (28 changed lines, test encoding-26.1) for main: InitAutoPathObjCmd appends $tcl_library/encoding to the encoding search path again. Fixes the --disable-zipfs repro (euc-jp, koi8-r, 95 encodings); encoding, interp, safe, init, unixInit tests pass; the new test fails without the fix. For 8.6 the fix would be the one line from 17c53b7edd (TclpSetInitialEncodings() at the end of Tcl_Init), but see the side effect below; no 8.6 patch made.
          • Side effect found since 17c53b7edd every Tcl_Init resets the system encoding chosen by the script: encoding system koi8-r; interp create; encoding system gives utf-8 on main, 9.1.0 and 9.0.5 (koi8-r on 8.6 and 8.7b1; a safe interp does not reset it). Threads and Tk interps are probably affected the same way (not tested). I found no ticket for this.
          • Suggested ticket comment The system-encoding part was fixed in 9.0 by 17c53b7edd (Tcl_Init sets the system encoding again); 8.6 still shows it for an uninstalled/relocated tclsh. On main TIP 732 dropped the encoding dirs addition from init.tcl, so a --disable-zipfs tclsh with a relocated library has no encoding dirs at all; proposed fix for that and test encoding-26.1 in branch (name tbd). Note that 17c53b7edd has a side effect: &#91;interp create&#93; now resets an explicit &#91;encoding system&#93; setting back to the locale default.
          ⬜8b3626f9582018Mismatching section reference and title in READMEpatchREPRODUCED (main, 9.0, 8.6) — README.md section 5 still says see the "Support and bug fixes" section below, the section is "8. Support and Training"

          Open ticket 8b3626f958 · 51. Release Notes · filed 2018-05-13

          🧪 Prototype check (2026-10-24)

          8b3626f958.patch: VERIFIED

          • Base main (tip d6dd6d0f71)
          • Before README.md:101 see the "Support and bug fixes" section below.; the section is ## <a id="support">8.</a> Support and Training (l.119)
          • Patch applied cleanly (2 changed lines)
          • After see the "[Support and Training](#support)" section below., anchor support exists (also used by the table of contents, l.26)
          • Test none possible (documentation only)
          Full verdict

          REPRODUCED (main, 9.0, 8.6) — README.md section 5 still says see the "Support and bug fixes" section below, the section is "8. Support and Training"

          • Platform docs
          • Tested README.md in main, 9.0, 8.6 tips
          • Cause section renamed, reference not updated
          • Prototype prototypes/8b3626f958.patch (main, 2 changed lines): reference becomes a link [Support and Training](#support)
          • Suggested ticket comment Still present in README.md on all branches. Proposed fix: refer to "Support and Training" and link it to #support.
          ⬜4bbb308ab22018trace causes generation of string representation of "commandPrefix" argumentREPRODUCED (main, 9.0, 8.6)

          Open ticket 4bbb308ab2 · 18. Commands M-Z · filed 2018-04-28

          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform generic
          • Tested main, 9.0 tip, 8.6 tip; variable read trace and execution enter trace: trace add gives the pure list a string rep, and inside the callback the argument is a pure string (list intrep lost)
          • Cause generic/tclTrace.c stores the command prefix as a C string (char command[TCLFLEXARRAY] in TraceVarInfo/TraceCommandInfo) and runs it with Tcl_EvalEx after appending the trace arguments to a DString, so the words are re-parsed from strings. Branch bug-3485833 (related, last check-in a2fcb3e560, 2018-04-30) was never merged.
          • Prototype none (needs storing the prefix as a Tcl_Obj and calling Tcl_EvalObjv in all variable/command/execution trace paths and in trace info; well over 60 lines; it is a performance improvement, not a functional bug, as dgp noted)
          • Suggested ticket comment Still the case in 8.6, 9.0 and main: tclTrace.c keeps the command prefix as a string and evaluates it with Tcl_EvalEx, so the list intrep is lost. Fixing it means keeping a Tcl_Obj prefix and using Tcl_EvalObjv; this is an optimisation (RFE), not a correctness bug.
          🟢cc592ec70d2018vwait not NRE-enabled: C stack busypatchREPRODUCED (main, 9.0, 8.6) for "cannot yield: C stack busy" (vwait/update/time/etc. are not NRE-enabled); the panic from the comments ("tailcall c…

          Open ticket cc592ec70d · - New Builtin Commands · filed 2018-04-14

          🟢 Since triage (updated 2026-11-11)

          Status now: Open

          • 🔧 2026-10-17 check-in d10cadf3ca on core-8-6-branch, in main, 9.0, 8.6: Fix cc592ec70d: a second yieldto after one that failed with "C stack busy" panicked with "tailcall cannot find the right splicing spot". Backport of [eb50e83e
          • 💬 2026-10-17 serhiy.storchaka: "Cannot yield: C stack busy" from a coroutine in an event handler under [vwait] is the same in all versions: vwait runs the event loop by nested C calls, and making it NRE-aware would be a larger design change. The panic "tailcall cannot find the right splicing spot" after a failed yieldto (sebres' scripts in the comments) was fixed for 9.0 in eb50e83ee1 (f9800d52bd) and is now backported to core-8-6-branch in…
          Full verdict

          REPRODUCED (main, 9.0, 8.6) for "cannot yield: C stack busy" (vwait/update/time/etc. are not NRE-enabled); the panic from the comments ("tailcall cannot find the right splicing spot") is FIXED since 9.0.0 (8.7 branch, eb50e83ee1, 2021-06-20, ticket f9800d52bd) but still REPRODUCED on 8.6 tip and 8.6.18

          • Platform generic
          • Tested main, 9.0 tip, 8.6 tip, 8.6.18, 8.7b1: the reporter's script gives "cannot yield: C stack busy" everywhere. sebres's 2nd and 3rd scripts (a second yieldto after a failed one, from an after handler under vwait or from inside time) abort with Tcl_Panic on 8.6; on 9.0/main they only give the error again
          • Cause Tcl_VwaitObjCmd runs the event loop by recursive C calls (Tcl_DoOneEvent), so a yield from an event handler finds corPtr->stackLevel != stackLevel in TclNRCoroutineActivateCallback (generic/tclBasic.c). In 8.6 the failed yieldto leaves its tailcall list in the NRCommand callback (data[1]) of the caller's env, so the next TclSetTailcall finds no free splice spot and panics; eb50e83ee1 records corPtr->yieldPtr and removes the stale tailcall when the yield fails
          • Prototype prototypes/cc592ec70d.patch (core-8-6-branch; 66 changed lines, 25 in code, tests coroutine-7.13 (from 9.0) and coroutine-7.14): a backport of eb50e83ee1 to 8.6. Fixes both panic repros; coroutine, tailcall, nre and basic tests pass; both tests panic/fail on unpatched 8.6. Making vwait itself NRE-enabled (the ticket title) is a design change (event loop shared by interps, see sebres' comment): no prototype.
          • Suggested ticket comment "C stack busy" from yield inside vwait is still the case on all versions: vwait runs the event loop by recursive C calls, and making it NRE-aware is a larger design change. The panic after a failed yieldto (from the comments) was fixed for 9.0 in eb50e83ee1 (f9800d52bd) but never backported; proposed 8.6 backport with tests coroutine-7.13/7.14 in branch (name tbd).
          ⬜e86e178aff2018namespace deletion: trace that calls [apply], specifying the namespace, doesn't fire.REPRODUCED (main, 9.0, 8.6)

          Open ticket e86e178aff · 21. [namespace] · filed 2018-01-19

          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform generic
          • Tested main, 9.0, 8.6: the delete trace with apply {args ... ::ns1} prints nothing, both with namespace delete from inside and from outside the namespace. A plain command trace does fire. Calling apply explicitly from such a trace fails with namespace "::ns1" not found, and delete-trace errors are silently ignored.
          • Cause generic/tclProc.c:TclNRApplyObjCmd/InitLambdaCall resolves the lambda's namespace by name (TclGetNamespaceFromObj), and generic/tclNamesp.c:GetNamespaceFromObj rejects NS_DYING namespaces; with activations the namespace is also already unlinked from its parent. The proposed fix on bug-e593adf103-core-8 (pooryorick, 2018, about 23 check-ins on namespace lifecycle and refcounting) was never merged (no merge children).
          • Prototype none (needs a decision on whether a dying namespace may be resolved by name; the existing branch is a large refactor)
          • Suggested ticket comment Still reproduces on 9.1/9.0/8.6: apply cannot resolve a namespace that is being deleted ("namespace ... not found"), and delete-trace errors are ignored. Branch bug-e593adf103-core-8 was never merged.
          ⬜ca6a9e56592017Can't flush one line through a socket when using zlib compression/decompressionpatchREPRODUCED (main, 9.0, 8.6, 8.6.6)

          Open ticket ca6a9e5659 · 57. zlib · filed 2017-09-29

          🧪 Prototype check (2026-10-24)

          ca6a9e5659.patch: VERIFIED

          • Base main (tip d6dd6d0f71)
          • Before zlib-17.1 (the reporter's case: server pushes deflate, puts a line, -flush sync; a separate tcltest process pushes inflate on a blocking localhost socket and does gets) with tclZlib.c unpatched -> timeout after 5 s
          • Patch applied cleanly (39 changed lines). I checked the loop on the tip: the new avail_in == 0 break is reached only after a decBytes > 0 round, so it never returns 0 (EOF) by mistake
          • After -> qwertyuiop (3/3 runs); tests zlib.test 73 passed/2 skipped/0 failed, http11.test 86/12/0, http.test 148/6/0
          • Test existing zlib-17.1 fails before / passes after
          Full verdict

          REPRODUCED (main, 9.0, 8.6, 8.6.6)

          • Platform generic
          • Tested reporter's server/client as two processes on 127.0.0.1 with main (30d8ccce48), 9.0.5, 8.6.18, 8.6.6. The client's gets returns only when the server closes (3 s later). A raw client shows the server does send the full sync-flushed block (18 bytes, ending 0000ffff), so the sending side is fine.
          • Cause generic/tclZlib.c:ZlibTransformInput, inflate side: after decompressing the bytes it got, it loops back to Tcl_ReadRaw on the parent whenever toRead is not yet satisfied. On a blocking channel that read blocks even though decompressed data is ready. Non-blocking channels break out on EAGAIN, so they are not affected.
          • Prototype prototypes/ca6a9e5659.patch (39 changed lines incl. 30 test lines, test zlib-17.1). Return what was decompressed once all buffered input is consumed (inStream.avail_in == 0) instead of reading more. zlib.test and http11.test pass; zlib-17.1 times out without the fix.
          • Suggested ticket comment Still in 9.1/9.0/8.6. The server sends the sync-flushed data; the inflate transform's input proc keeps reading the blocking socket instead of returning the bytes it already decompressed. Proposed fix: stop once buffered input is consumed (test zlib-17.1).
          🔧875545270d2017tcl::CopyDirectory copying /a/a /x/a/a errorpatchREPRODUCED (main, 9.0, 8.6, 8.4)

          Open ticket 875545270d · 38. Init - Library - Autoload · filed 2017-09-14

          🔧 Since triage (updated 2026-11-11)

          Status now: Open

          • 🔧 2026-10-19 check-in c51f96d014 on copy-into-itself, not merged: Fix 82049f96f8 and 875545270d: [file copy] of a directory into itself recursed until the path was too long; now it is an error. tcl::CopyDirectory, used for
          • 💬 2026-10-19 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1. The check uses [string first], so a target that only contains the source path somewhere is rejected. It matters for [file copy] between file systems (from zipfs or a VFS), which uses tcl::CopyDirectory. Proposed fix in branch copy-into-itself, together with 82049f96f8: compare the [file split] prefix. Tests init-6.0 and 6.1.
          Full verdict

          REPRODUCED (main, 9.0, 8.6, 8.4)

          • Platform generic
          • Tested main build (c3bf6dd052), 9.0.5, 8.6.18, 8.4.20; work-a2/875545270d.tcl (src $w/cd875/a, dest $w/cd876/a$src)
          • Cause library/init.tcl:tcl::CopyDirectory checks "dest inside src" with string first $nsrc $ndest (substring anywhere) plus a comparison of a single path element, so a dest that contains the src path further down, with a matching element at the same depth, is rejected as "move a directory into itself". Reachable from file copy only through the cross-filesystem path (zipfs/vfs to native).
          • Prototype prototypes/875545270d.patch (34 changed lines, tests init-6.0, init-6.1): compares the file split prefix instead. init.test and fCmd.test pass.
          • Suggested ticket comment The into-itself check in tcl::CopyDirectory uses &#91;string first&#93; instead of a path-prefix test. Proposed fix: compare the &#91;file split&#93; prefix (test init-6.0).
          ⬜941d52a3e52017[dict with] discards string rep even when it doesn't modify the dictionarypatchREPRODUCED (main, 9.0, 8.6, 8.5)

          Open ticket 941d52a3e5 · 16. Commands A-H · filed 2017-08-24

          🧪 Prototype check (2026-10-24)

          941d52a3e5.patch: VERIFIED

          • Base main (tip d6dd6d0f71)
          • Before work-a4/v4dictwith.tcl: set d "\n\tone 1\n\ttwo 2\n"; dict with d {} -> {one 1 two 2} (string rep regenerated); compiled dict with d {} in apply -> {a {\n x 1 } b 2}; dict with d a {} -> {a {x 1} b 2}
          • Patch applied cleanly (77 changed lines)
          • After all three keep the original strings ("\n\tone 1\n\ttwo 2\n", "a {\n x 1 } b 2"); make test TESTFLAGS="-file 'dict.test var.test upvar.test trace.test'" -> 958 total, 948 passed, 10 skipped, 0 failed; dict-20*..22* under valgrind with the -DPURIFY build (work-a4/build-purify): 69 passed, no valgrind errors
          • Test existing dict-22.26 and dict-22.27 fail before (dict.test unpatched: 373 passed, 2 failed) / pass after
          Full verdict

          REPRODUCED (main, 9.0, 8.6, 8.5)

          • Platform generic
          • Tested main build, 9.0.5, 8.6.18, 8.5.19 (work-a2/941d52a3e5.tcl; compiled and uncompiled dict with behave the same). dict update with an empty body has the same effect (not changed by the prototype).
          • Cause generic/tclDictObj.c:TclDictWithFinish always unshares the dict, re-puts every key (Tcl_DictObjPut invalidates the string rep even for the same value object) and writes the variable back.
          • Prototype prototypes/941d52a3e5.patch (77 changed lines, most of them moved code; tests dict-22.26, dict-22.27): looks up the leaf read-only first and unshares/updates only at the first key whose variable no longer holds the very same Tcl_Obj; if none changed it returns without writing the variable. Side effect: a write trace on the dict variable no longer fires for a no-op dict with. dict, var, upvar, trace tests pass; the new tests fail without the fix.
          • Suggested ticket comment TclDictWithFinish re-puts every key and writes the variable back even when no value changed. Proposed fix: skip the write-back when every variable still holds the same value object (tests dict-22.26/27); a no-op &#91;dict with&#93; then no longer fires write traces.
          ⬜77f0dbc0f82017Wrong errorstack when using slave interppatchREPRODUCED (main, 9.0, 8.6)

          Open ticket 77f0dbc0f8 · 20. [interp] · filed 2017-07-26

          🧪 Prototype check (2026-10-24)

          77f0dbc0f8.patch: VERIFIED

          • Base main (tip d6dd6d0f71)
          • Before interp-31.2 FAILED: -errorstack INNER {returnImm {caught error} {}} UP 1 UP 1 (stale INNER from an earlier caught error)
          • Patch applied cleanly (test hunk offset 20 lines; 26 changed lines)
          • After interp-31.2 passes (INNER {invokeStk foo} ...); tests interp 385/0 failed, error 318/0, info 285/0, cmdMZ 96/0, execute 157/0, ioTrans 72/0, basic 118/0, proc 37/0, namespace 312/0, coroutine 81/0, nre 29/0, oo 373/0
          • Test existing interp-31.2 fails before / passes after
          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform generic
          • Tested main (own build), 9.0 tip, 8.6 tip; stale INNER {returnImm {catched error} {}} from an already-caught error shows up in -errorstack after a child->parent alias error
          • Cause generic/tclResult.c:Tcl_GetReturnOptions puts iPtr->errorStack into the options even when iPtr->resetErrorStack is set (the stack is logically empty but still holds the previous error). The alias target in the parent fails without any Tcl_LogCommandInfo, so the reset never happens, and Tcl_TransferResult copies the stale list into the child; Tcl_SetReturnOptions then clears resetErrorStack in the child, so the child never adds its own INNER either.
          • Prototype prototypes/77f0dbc0f8.patch (26 changed lines, test interp-31.2). Tcl_GetReturnOptions reports an empty -errorstack when a reset is pending; Tcl_SetReturnOptions leaves the reset pending when it gets an empty -errorstack. Result is now INNER {invokeStk foo} CALL {run_in_slave foo}, which is what the reporter expected. interp, error, result, info, cmdMZ, execute, ioTrans, basic, proc, namespace, coroutine and nre tests all pass; full suite: only the pre-existing zipfs.test environment failures (they also fail on a clean build). Side effect: an explicit return -code error -errorstack {} now lets the next level record its INNER context.
          • Suggested ticket comment Tcl_GetReturnOptions() returned the stale error stack while a reset was pending, and Tcl_TransferResult() carried it into the child interp. Proposed fix with test interp-31.2 (prototype, not yet on a branch).
          ✅59a9b957c62017windows: make distclean doesn't clean uppatchREPRODUCED (main, 9.0, 8.6) — partly: Makefile, config.status, tclConfig.sh are removed, so configure must be rerun, but tcl.pc (made by configur…

          Open ticket 59a9b957c6 · 53. Configuration and Build Tools · filed 2017-06-21

          ✅ Since triage (updated 2026-11-11)

          Status now: Closed (Fixed)

          • 🔧 2026-10-20 check-in 2025c9111f on win-distclean-tcl-pc, in main, 9.0: Fix 59a9b957c6: Windows "make distclean" now removes the generated tcl.pc.
          • 🔧 2026-10-20 check-in c87b92e2ee on trunk, in main: Merge win-distclean-tcl-pc and win-native-mingw-tclsh: fix 59a9b957c6 and 2ef3430fdc.
          • 🔧 2026-10-20 check-in 063874bae3 on core-9-0-branch, in main, 9.0: Fix 59a9b957c6 and 2ef3430fdc: Windows "make distclean" removes tcl.pc; win/configure uses an installed tclsh only when cross-compiling, not for any *mingw3
          • 🔧 2026-10-20 check-in 161779a5c4 on core-8-6-branch, in main, 9.0, 8.6: Fix 59a9b957c6 and 2ef3430fdc: Windows "make distclean" removes tcl.pc; win/configure uses an installed tclsh only when cross-compiling, not for any *mingw3
          • 💬 2026-10-20 serhiy.storchaka: tcl.pc is generated by win/configure (AC_CONFIG_FILES) but was missing from the distclean rule in win/Makefile.in, so it survived and kept the old prefix. The other generated files were already removed; unix/Makefile.in removes tcl.pc too. Proposed fix in branch 2025c9111f. Checked with a MinGW build: make distclean now leaves the build directory empty.
          • 💬 2026-10-20 jan.nijtmans: Just go ahead! +1 Thanks!
          Full verdict

          REPRODUCED (main, 9.0, 8.6) — partly: Makefile, config.status, tclConfig.sh are removed, so configure must be rerun, but tcl.pc (made by configure, contains the old prefix) survives distclean

          • Platform build (Windows configure/Makefile.in: MSYS2/mingw)
          • Tested main win/configure --host=x86_64-w64-mingw32 --prefix=... in a separate build dir, make, make distclean: only tcl.pc left (contains the prefix). Code reading for 9.0/8.6 (same rule; unix/Makefile.in distclean does remove tcl.pc)
          • Cause win/Makefile.in distclean does not list tcl.pc (added to AC_CONFIG_FILES later)
          • Prototype prototypes/59a9b957c6.patch (main, 2 changed lines, no test; verified: build dir empty after distclean). Same one-word change for 9.0/8.6.
          • Suggested ticket comment distclean already removes Makefile/config.status, so configure has to be rerun, but the configured tcl.pc (with the old prefix) was left behind. Proposed fix: add tcl.pc to the win/Makefile.in distclean rule, as unix/Makefile.in does.
          ⬜da7c0799782017FS strips root '/' from FS root volume "zvfs:/"patchREPRODUCED (main, 9.0, 8.6)

          Open ticket da7c079978 · 37. File System · filed 2017-06-08

          🧪 Prototype check (2026-10-24)

          da7c079978.patch: VERIFIED

          • Base main (tip d6dd6d0f71)
          • Before tcltest with testsimplefilesystem 1: file normalize of simplefs:/simpledir/.., simplefs:/a/b/../.., simplefs:/. -> simplefs: (separator lost); simplefs:/a/../b -> simplefs:/b
          • Patch applied cleanly (test hunk at offset -16; 23 changed lines)
          • After -> simplefs:/ simplefs:/ simplefs:/ simplefs:/b; make test TESTFLAGS="-file 'fileSystem.test fileName.test cmdAH.test zipfs.test fCmd.test winFCmd.test winFile.test'" -> 18555 total, 18130 passed, 425 skipped, 0 failed (fileName.test includes the testsetplatform windows drive cases)
          • Test existing filesystem-7.9 fails before (Result was simplefs: simplefs: simplefs: simplefs:/b) / passes after
          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform generic (VFS)
          • Tested tcltest with testsimplefilesystem (volume "simplefs:/") on main (30d8ccce48), 9.0 tip, 8.6 tip: file normalize simplefs:/dir/.. -> simplefs: (should be simplefs:/). zipfs volumes (//zipfs:/) are fine because of their special case. The reporter's other points (cd / from a volume-relative cwd on Windows going through TclWinVolumeRelativeNormalize, rename not routed to the VFS) were not reproduced: they need a Windows build and his chiselapp test fs, which we don't have offline.
          • Cause generic/tclPathObj.c:TclFSNormalizeAbsolutePath strips ".." back to the last separator and re-adds a trailing "/" only for zipfs volumes and Windows "X:" drives.
          • Prototype prototypes/da7c079978.patch (23 changed lines incl. 10 test lines, test filesystem-7.9). Also re-adds the "/" when the result ends in ':' and contains no separator, i.e. a bare "name:" volume (covers "C:" too). fileSystem, fileName, cmdAH, zipfs, fCmd tests pass; filesystem-7.9 fails without the fix.
          • Suggested ticket comment The normalize part still reproduces with tcltest's simplefs (file normalize simplefs:/dir/.. -> simplefs:). Proposed fix: keep the trailing separator of any "name:/" volume, not only C:/ and zipfs (test filesystem-7.9). The Windows volume-relative issues from the comments need separate checking.
          ⬜46450b464e2017A namespace loses track of itselfREPRODUCED (main, 9.0, 8.6, 8.5), but it follows from the documented semantics

          Open ticket 46450b464e · 21. [namespace] · filed 2017-06-02

          Full verdict

          REPRODUCED (main, 9.0, 8.6, 8.5), but it follows from the documented semantics

          • Platform generic
          • Tested main, 9.0 tip, 8.6 tip, 8.5.19: both reporter scripts behave as described on all of them
          • Cause after namespace delete [namespace current] the running namespace is NS_DYING. It stays current for the running code, but by design (namespace.n: "marked to prevent other code from looking it up by name") the name ::ns1 no longer resolves to it. namespace eval [namespace current] {} therefore creates a second ::ns1. Relative doc creates/looks up a child of the dying namespace, while apply ... ::ns1 runs in the new one. So two namespaces share the name ::ns1.
          • Prototype none (needs a design decision, e.g. refuse to create children in, or namespace eval into, a dying namespace, or have namespace current reflect the deleted state)
          • Suggested ticket comment Still the same on 8.5 through 9.1. Code that is still running in a deleted namespace keeps using it, but the name ::ns1 no longer finds it (as namespace.n documents), so namespace eval [namespace current] creates a new namespace under the same name. Changing this needs a design decision about what code running in a dying namespace may do.
          🟢a4a7f2abd32017Tcl [string match] and [glob] are exponential complexity for bad patterns, could be linearpatchREPRODUCED (main, 9.0, 8.6)

          Open ticket a4a7f2abd3 · 16. Commands A-H · filed 2017-04-24

          🟢 Since triage (updated 2026-11-11)

          Status now: Open

          • 🔧 2026-10-19 check-in 2e0da1e14e on glob-match-linear, in main: Fix a4a7f2abd3: [string match], [glob] and [lsearch -glob] took exponential time for patterns like a*a*a*...b. When the string runs out before the pattern, an
          • 🔧 2017-05-22 check-in c84d53f778 on aspect-string-match, not merged: fix advancement of snext: could not cook up a test for this, but the error is clear
          • 💬 2026-10-19 serhiy.storchaka: Still exponential in 8.6, 9.0 and 9.1: with string match [string repeat a* 30]b [string repeat a 30] each failed * retries every suffix, although nothing after an exhausted string can match. Proposed fix in branch glob-match-linear: a smaller change than the rewrite in aspect-string-match, in all four matchers (string, byte array and both Unicode ones). As in wildmat, a match that fails because the string ran out …
          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform generic
          • Tested work-a2/a4a7f2abd3.tcl (string match [string repeat a* n]b [string repeat a n]): time grows about 10x per +4 of n on main and 8.6.18 (n=22: 12 ms on 8.6, 27 ms on the debug main build; n=30 takes seconds). Same for -nocase, lsearch -glob, the Unicode and bytearray paths. Branch aspect-string-match (2017, Tcl_StringCaseMatch only) was never merged.
          • Cause the four matchers Tcl_StringCaseMatch / TclByteArrayMatch (generic/tclUtil.c) and TclUniCharCaseMatch / TclUniCharMatch (generic/tclUtf.c) recurse at every * and retry every suffix even after an inner * already failed for all suffixes.
          • Prototype prototypes/a4a7f2abd3.patch (127 changed lines, above the size guideline but mechanical; test string-11.56.*): wildmat-style ABORT. Each matcher becomes a static helper returning 1/0/-1 (-1 = string ran out), the public function returns helper(...) == 1, and an outer * stops retrying on -1. That makes matching O(len·pattern); n=30 now takes about 1 us. 20000 random pattern/string pairs (ASCII, Unicode, nocase, bytearray, lsearch) give identical results to 9.0.5 and unpatched main. string, lsearch, switch, fileName, cmdIL tests pass; the new test fails on the 9.0 tip. The [a- / trailing-backslash oddities from aspect's comment are untouched.
          • Suggested ticket comment Still exponential on 8.6/9.0/9.1. Proposed fix: wildmat-style "abort" return in the four glob matchers so an outer * stops retrying once the string is exhausted. Same results as before on 20000 random cases (test string-11.56).
          ⬜b7b707a3102017[win32] NRT-only - NativeGetTime backwards time-drifts bug (fixed)REPRODUCED (main, 9.0, 8.6) on the VirtualBox VM — backward steps of up to ~150 ms; whether native hardware shows it was not tested

          Open ticket b7b707a310 · 03. Timer Events · filed 2017-03-31

          Full verdict

          REPRODUCED (main, 9.0, 8.6) on the VirtualBox VM — backward steps of up to ~150 ms; whether native hardware shows it was not tested

          • Platform Windows
          • Tested Linux main/8.6 sanity run of the repro (no backward steps, microseconds vs monotonic drift 0..1 us); code check: sebres' fix f6637d3dd8 (branch sebres-8-5-timerate, 2017-05-23, calibrationInterv, no backward ticks on calibration) is present in win/tclWinTime.c on main, 9.0 and 8.6 (merged 2019-03, per sebres).
          • Cause win/tclWinTime.c calibration thread (NativeGetTime/UpdateTimeEachSecond) re-bases the virtual clock; the 2017 backward jumps were fixed, but sebres reports (2026-09-14) that [clock microseconds] still drifts 1-2 ms against [clock monotonic] around calibrations and plans to backport the inline calibration from fdfbd5e10.
          • Prototype none (Windows-only, owner is working on it)
          • Windows repro win/b7b707a310.tcl (default 30 s; counts backward steps of [clock microseconds] and, on 9.1, its drift range against [clock monotonic]).
          • Suggested ticket comment none for now: the original backwards-drift fix f6637d3dd8 is in 8.6/9.0/main; leave open for sebres' pending calibration work (2026-09-14 comment) and close when that lands.
          Windows VM, run 2
          • Platform Windows
          • Verdict REPRODUCED (main, 9.0, 8.6) on the VirtualBox VM — backward steps of up to ~150 ms; whether native hardware shows it was not tested
          • Tested VM win/b7b707a310.tcl for 60 s each: clock microseconds went backwards 11 times on 9.1b1 (max 147005 us), 15 on 9.0.5 (max 117523 us), 12 on 8.6.18 (max 147041 us); on 9.1 microseconds vs monotonic drifted by up to 3.17 s. Control (work-win2/ck/ck.c, same 60 s): GetSystemTimePreciseAsFileTime never goes backwards (0 of 1.9e9 samples), but it drifts −1.3 .. +2.0 s against QueryPerformanceCounter, because the VM's virtual TSC and VBoxService time sync disagree. So on this VM the system clock itself is monotonic, and the backward steps come from Tcl's calibration.
          • Cause win/tclWinTime.c NativeGetTime/UpdateTimeEachSecond interpolates the system time from QPC and re-bases it at each calibration; when QPC and the system clock diverge (here by up to ms per second) the re-based value can jump back. The 2017 fix f6637d3dd8 (present in 8.6/9.0/main) limits but does not prevent this. Tcl never calls GetSystemTimePreciseAsFileTime (Win8+), which would make the whole calibration thread unnecessary.
          • Prototype none (owner sebres is working on the calibration; a switch to GetSystemTimePreciseAsFileTime would be a larger change)
          • Related on 9.0 the same clock disagreement triggers a ~72 min hang in Tcl_Sleep (separate bug, see the last section).
          • Suggested ticket comment On a VirtualBox Windows 11 guest, where QPC and the system clock drift apart by up to 2 s per minute, &#91;clock microseconds&#93; still goes backwards by up to ~150 ms (8.6.18, 9.0.5, 9.1), while GetSystemTimePreciseAsFileTime stays monotonic. Maybe use GetSystemTimePreciseAsFileTime directly instead of the calibrated QPC?
          🔧96c3f3b47d2017No processing of after eventspatchREPRODUCED (8.6 on Win64); FIXED since 9.0 (64-bit Tcl_Time.sec)

          Open ticket 96c3f3b47d · 03. Timer Events · filed 2017-02-12

          🔧 Since triage (updated 2026-11-11)

          Status now: Open

          • 🔧 2026-10-19 check-in d7e3050ca6 on after-overflow-8-6, not merged: Fix 96c3f3b47d and 1802895fff: [after] with a delay over about 12 years overflowed the seconds where long is 32 bits (Windows, 32-bit systems), so later tim
          • 💬 2026-10-19 serhiy.storchaka: Proposed fix for 8.6 in branch after-overflow-8-6, in the spirit of chw's patch: both forms of [after] add the delay with a new TimeAddMs(), which clamps the wakeup time (to 2038-01-19 where long is 32 bits) instead of letting it wrap. Tested with a MinGW Win64 build (32-bit long) under Wine: with after [expr {2**40}] ... pending, an after 200 never fired before and fires after 201 ms now, and the blocking form no…
          Full verdict

          REPRODUCED (8.6 on Win64); FIXED since 9.0 (64-bit Tcl_Time.sec)

          • Platform generic (32-bit long only: 8.6 on Win64 and on any 32-bit build)
          • Tested main, 9.0.5, 8.6.18 on Linux x86_64 (not reproducible there: 64-bit long); no 32-bit toolchain here. 8.6 on 64-bit Linux still accepts after 9223372036854775808 (bignum wraps to negative, fires at once; harmless).
          • Cause generic/tclTimer.c:Tcl_AfterObjCmd (8.6) wakeup.sec += (long)(ms / 1000) overflows a 32-bit long; the wrapped negative sec sorts first in TclCreateAbsoluteTimerHandler while TimerSetupProc computes a huge positive wait, so every later timer waits behind it.
          • Prototype prototypes/96c3f3b47d.patch (against core-8-6-branch; 30 changed lines, test timer-11.3) — clamps the wakeup to LONG_MAX seconds instead of wrapping (simplified chw's patch, without the signed-overflow TCL_TIME_BEFORE change). timer.test passes on Linux; the test only bites on a 32-bit-long build.
          • Suggested ticket comment Not an issue in Tcl 9 (64-bit Tcl_Time.sec). For 8.6 (32-bit long, e.g. Win64) proposed fix: clamp the wakeup time in Tcl_AfterObjCmd instead of letting it wrap, test timer-11.3.
          Windows VM, run 2
          • Platform generic (32-bit long: 8.6 on Win64 and any 32-bit build)
          • Verdict REPRODUCED (8.6 on Win64); FIXED since 9.0 (64-bit Tcl_Time.sec)
          • Tested VM win/96c3f3b47d.tcl: 9.1b1 and 9.0.5 after 0 fires; 8.6.18 hangs for good — even the after 3000 exit timer never fires behind the after 665544332211 timer (killed after a few minutes).
          • Cause generic/tclTimer.c Tcl_AfterObjCmd (8.6): wakeup.sec += (long)(ms / 1000) overflows the 32-bit long, as in the Linux analysis
          • Prototype prototypes/96c3f3b47d.patch (core-8-6-branch, 30 changed lines, test timer-11.3), unchanged: with it the repro passes and timer.test passes on Windows (55/55). Without the fix timer-11.3 hangs the test run (verified: unpatched tcltest86 still running after 15 s), so maybe add a constraint or run it in a child process. Tested on Windows.
          • Suggested ticket comment Confirmed on 8.6.18 Win64: a far-future &#91;after&#93; blocks every later timer; 9.0/9.1 are fine. Proposed fix for 8.6: clamp the wakeup time instead of wrapping (test timer-11.3, tested on Windows).
          ⬜322720f99d2017Poor bytecode (optimizations?) for || and &&REPRODUCED (main; 9.0/8.6 compile the same way) — performance/RFE, not a correctness bug

          Open ticket 322720f99d · 47. Bytecode Compiler · filed 2017-02-06

          Full verdict

          REPRODUCED (main; 9.0/8.6 compile the same way) — performance/RFE, not a correctness bug

          • Platform generic
          • Tested main disassembly of a 3-term || and &&: each inner pair still yields jumpTrue L1; push 0; jump L2; L1: push 1; L2: jumpTrue/jumpFalse ..., the same pattern dgp and dkf showed in 2017
          • Cause generic/tclCompExpr.c compiles each ||/&& to a boolean value, and the peephole optimizer (generic/tclOptimize.c) does not fold a "push constant boolean then conditional jump" back into a direct jump to the final target
          • Prototype none (optimizer work: jump threading over push-0/1 sequences, with checks for other jumps into the sequence; more than a small fix)
          • Suggested ticket comment Still the same on 9.1: chained ||/&& materialise a 0/1 at each step and then test it again. This needs jump threading in tclOptimize.c (or in how the expression compiler emits ||/&&). It is an optimisation, not a correctness bug.
          ⬜120922d1d42017[clock format] returns timezone [clock scan] doesn't understandREPRODUCED (main, 9.0, 8.6) in the numeric-abbreviation form. "MHT" is gone (tzdata now writes "+12"), but `clock scan [clock format $t -timezone +…

          Open ticket 120922d1d4 · 06. Time Measurement · filed 2017-01-30

          Full verdict

          REPRODUCED (main, 9.0, 8.6) in the numeric-abbreviation form. "MHT" is gone (tzdata now writes "+12"), but clock scan [clock format $t -timezone +12] ("... 14:18:26 +12 2017") is 42480 s off on main, 9.0 and 8.6. On 8.6 also :Kwajalein/:Pacific/Kiritimati/:America/Sao_Paulo (they format as +12/+14/-02 there).

          • Platform generic
          • Tested main c3bf6dd052, 9.0.5, 8.6.18 (work-a6/t5.tcl)
          • Cause the same as 7bad415e3e: in 9.x nmzone takes the digit count of the following year (2017), in 8.6 any signed number is hhmm. Side finding: ist is -HOUR(11/2) in generic/tclGetDate.y, and HOUR() is integer, so IST scans as +05:00 (1800 s off for :Asia/Kolkata round trip); nst (13/2), jt (15/2), cast/cadt (19/2) lose their half hour the same way (main, 9.0, 8.6).
          • Prototype none (covered by the fix for 7bad415e3e; the half-hour table entries are a separate 10-line fix in tclGetDate.y + tclDate.c)
          • Suggested ticket comment "MHT" is no longer produced, but the default format still does not round-trip for zones abbreviated as +12: the scanner misreads a numeric zone followed by the year, see 7bad415e3e. Also IST/JT/CAST scan without their half hour (HOUR(11/2) is integer division).
          ⬜0520d172842017Exploding runtime with growing list of after idle eventspatchREPRODUCED (main, 9.0, 8.6)

          Open ticket 0520d17284 · 02. Event Loops · filed 2017-01-27

          🧪 Prototype check (2026-10-24)

          0520d17284.patch: VERIFIED+TEST-ADDED

          • Base main (tip d6dd6d0f71)
          • Before tcltest 0520.tcl (reporter's script, 10k/20k/40k after idle then update) -> 10000: 89 ms / 20000: 262 ms / 40000: 1380 ms (quadratic)
          • Patch applied cleanly
          • After same script -> 3 ms / 5 ms / 9 ms; tests timer.test 138/138 passed, event.test 56 passed 10 skipped 0 failed, interp.test 384 passed 7 skipped 0 failed
          • Test existing test (renamed timer-bug-0520d17284.1) is a list-consistency test and passes before too; added timer-bug-0520d17284.2 (relative timing: update for 32000 vs 2000 idle events must be < 64x; ~16x linear, ~220x before), fails before (3940 us vs 873770 us, 3/3 runs) / passes after (5/5 runs)
          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform generic
          • Tested reporter's script plus an after 0 variant on main (own build), 9.0.5, 8.6.18: [update] for 10k/20k/40k events takes ~60/250/1450 ms on all three (quadratic). With the prototype: 2/5/9 ms.
          • Cause generic/tclTimer.c: AfterInfo list per interp is singly linked with new entries at the head, but events fire oldest-first, so AfterProc (and FreeAfterPtr for [after cancel]) walk the whole list to unlink → O(n²). Separately, creating n after <ms> timers is also O(n²) (CreateTimerHandler sorted insert walks from the head; 40k after 0 take ~1.3 s to create) — not addressed by the prototype. sebres' large rework (branch sebres-8-6-event-perf-branch, last touched 2022-05-31) was never merged.
          • Prototype prototypes/0520d17284.patch (main; 92 changed lines, of which ~40 are removed loops and a comment header; test timer-bug-0520d17284) — adds prevPtr to AfterInfo with LinkAfterPtr/UnlinkAfterPtr helpers (as in chw's androwish change), O(1) unlink. timer/event/interp tests pass.
          • Suggested ticket comment Still quadratic in 9.0/main: AfterProc unlinks from the head of a singly linked list while events fire oldest-first. Proposed fix: doubly linked AfterInfo list (prevPtr), test timer-bug-0520d17284. Timer creation with many pending after <ms> is still O(n) per insert.
          ⬜931e4956b92017Interpreter command count not always incrementedpatchREPRODUCED (main, 9.0, 8.6; 8.5 counts correctly)

          Open ticket 931e4956b9 · 47. Bytecode Compiler · filed 2017-01-12

          🧪 Prototype check (2026-10-24)

          931e4956b9.patch: VERIFIED (behaviour change: command limits trip one command earlier, as the triage says)

          • Base main (tip d6dd6d0f71)
          • Before work-a2/v2/cnt.tcl in a child interp: 5x ip eval {incr i} -> 5x incr: 0 counted; ip eval {set x 1; set y 2} -> set;set: 1
          • Patch applied with offset +20 in interp.test (re-saved, 36 changed lines, content unchanged)
          • After same -> all incr/set commands counted (6 and 3 with the script's arithmetic, i.e. 5 and 2 plus the double-counted info cmdcount the triage already noted); full suite ./tcltest ../tests/all.tcl -notfile 'http*.test zipfs.test': 521570 total, 519479 passed, 2091 skipped, 0 failed
          • Test existing info-bug-931e4956b9 fails before (C fix reverted: 1 instead of 5) / passes after; the updated interp-34.4..34.7 also fail before (4 failures) since they encode the new limit timing. Note: the test's expected 5 includes info cmdcount itself counted twice (pre-existing double count), so it would need updating if that is fixed later.
          Full verdict

          REPRODUCED (main, 9.0, 8.6; 8.5 counts correctly)

          • Platform generic
          • Tested main, 9.0 tip, 8.6 tip, 8.5.19 in a child interp. 5x ip eval {incr i} adds 0 to info cmdcount on 8.6+ (8.5: 5). A 3-command proc body counts 2. set x 1; set y 2 counts 1.
          • Cause generic/tclCompile.c:CompileCmdCompileProc (the atCmdStart switch). TclInitCompileEnv sets atCmdStart = 1, and for codeNext == codeStart no INST_START_CMD is emitted and no count is recorded. So the first compiled command of every ByteCode is never added to iPtr->cmdCount; non-compiled first commands are still counted by Dispatch. This also affects interp limit -commands.
          • Prototype prototypes/931e4956b9.patch (36 changed lines, test info-bug-931e4956b9, placed at the end of info.test because info frame tests hard-code line numbers). At the start of a script it emits INST_START_CMD instead of skipping it. Compact (main-interp, no-limit) code is unaffected; other ByteCodes grow by 9 bytes. Behaviour change: command limits now trip one command earlier, so interp-34.4..34.6 expected {6 4 b} changes to {5 5 b} and interp-34.7 4 to 5 (updated in the patch). Full suite: only the pre-existing zipfs.test environment failures remain (they also fail on a clean build).
          • Note compiled ensemble subcommands without their own compiler (e.g. info cmdcount) are counted twice, by INST_START_CMD and by Dispatch. This is pre-existing and not fixed.
          • Suggested ticket comment The first compiled command of each bytecode is never counted, because the compiler omits INST_START_CMD at offset 0. Proposed fix with test info-bug-931e4956b9 (prototype). It makes command limits trip one command earlier, so the interp-34.4..34.7 expectations change.
          ⬜e7c21ed6782016entier() fails to canonicalize resultREPRODUCED (main, 9.0, 8.6, 8.5). On 9.0+ int() also returns its argument unchanged, since it became the same as entier() (TIP 514, 10bc1a9af5 20…

          Open ticket e7c21ed678 · 45. Parsing and Eval · filed 2016-12-16

          Full verdict

          REPRODUCED (main, 9.0, 8.6, 8.5). On 9.0+ int() also returns its argument unchanged, since it became the same as entier() (TIP 514, 10bc1a9af5 2020-01-23 refactor).

          • Platform generic
          • Tested main, 9.0 tip, 8.6 tip, 8.5.19. tcl::mathfunc::entier/round/abs/max 0x2 return 0x2, and expr {abs($x) eq "2"} is 0 while expr {+$x eq "2"} is 1. A top-level expr result is canonical (2) everywhere.
          • Cause generic/tclBasic.c ExprIntFunc (entier/int), ExprRoundFunc, ExprAbsFunc, ExprCeilFunc/ExprFloorFunc (integer args), ExprDoubleFunc, ExprMaxMinFunc and ExprSqrtFunc do Tcl_SetObjResult(interp, objv[1]) when the value is already of the result type. The argument's original string rep (0x2, 00.02) is kept.
          • Prototype none. This is a behaviour/performance trade-off (identity return avoids an allocation) across ~9 call sites. A fix would be a small helper that duplicates the object and calls Tcl_InvalidateStringRep when objv[1]->bytes is set. It needs a maintainer decision first, since aspect argued either way is defensible.
          • Suggested ticket comment Still the same on 9.1, and int() now behaves like entier() here. Math functions return their argument unchanged when it is already of the right numeric type, so the non-canonical string rep survives. Should they canonicalise, like unary + does? If yes, a small helper in tclBasic.c covers all of them.
          ⬜391bc0fd2c2016glob, -nocomplain, -types, and insufficient privilegespatchREPRODUCED (main, 9.0, 8.6) for the -types hidden part. The EACCES-should-raise part is a design question that aspect argued against; not changed.

          Open ticket 391bc0fd2c · 16. Commands A-H · filed 2016-11-16

          🧪 Prototype check (2026-10-24)

          391bc0fd2c.patch: VERIFIED (the MAC_OSX_TCL branch was checked by reading only)

          • Base main (tip d6dd6d0f71)
          • Before work-a3/v391.tcl -> in a mode-0600 dir glob -types hidden -dir d * = "" (it has .h); with dangling links -types l = bad but -types {hidden l} = "" (it has .bad)
          • Patch applied cleanly (48 changed lines)
          • After hidden in 0600 dir: .. . .h, hidden l: .bad. Returning ./.. matches what -types hidden already gives in a searchable dir (checked on 9.0.5 and main). Tests fileName (261/0 failed), fCmd, fileSystem, cmdAH: 0 failed
          • Test existing filename-11.52 and 11.53 (from the patch) fail on the unpatched tip / pass after
          Full verdict

          REPRODUCED (main, 9.0, 8.6) for the -types hidden part. The EACCES-should-raise part is a design question that aspect argued against; not changed.

          • Platform unix
          • Tested work-a2/391bc0fd2c.tcl on main, 9.0.5, 8.6.18, 8.6.6: in a mode-0600 dir glob -types hidden finds nothing, and glob -types {l hidden} .foo misses a dangling hidden symlink that -types l foo finds. On 9.x these silently return {} (no-match is not an error there); 8.6 raises "no files matched". aspect's fix e83297c2af (branch aspect-bug-391bc0fd2c, off core-8-5-branch, 2017) was never merged.
          • Cause unix/tclUnixFile.c:NativeMatchType calls stat() whenever any perm bit is set, including TCL_GLOB_PERM_HIDDEN, which on Unix needs only the name; a failing stat (EACCES, dangling link) rejects the entry.
          • Prototype prototypes/391bc0fd2c.patch (48 changed lines, tests filename-11.52, filename-11.53): a port of aspect's idea. The hidden check is done by name before any stat, stat only runs for non-hidden perm bits, and a statDone flag keeps the type and macOS branches correct. fileName and fCmd tests pass; the new tests fail on the 9.0 tip.
          • Suggested ticket comment -types hidden still stat()s every entry, so it fails in non-searchable dirs and on dangling hidden links. Proposed fix: check hidden by name before stat (port of e83297c2af, tests filename-11.52/53). Raising EACCES for -types d is left as is.
          ⬜bd2656d8e32016unset fails for declared-only namespace vars, but removes declarationpatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4). Long-standing behaviour; the choice is between DOCS and a behaviour change.

          Open ticket bd2656d8e3 · 07. Variables · filed 2016-10-28

          🧪 Prototype check (2026-10-24)

          bd2656d8e3.patch: VERIFIED+TEST-ADDED (doc + behaviour-pinning test)

          • Base main (tip d6dd6d0f71); also applies to core-8-6-branch (fuzz 1 in unset.n, offset in var.test)
          • Before unset.n says only that an error occurs "when the named variable does not exist"; nothing about the side effect. Behaviour (work-a1/bd26.tcl, main and 8.6 the same):

          namespace eval a {variable A}; info vars a::* -> ::a::A; unset a::A -> error can't unset "a::A": no such variable; info vars a::* -> ` (declaration gone; namespace which -variable a::A -> ). unset -nocomplain and unset of a variable`-linked local in a proc do the same.

          • Patch 43 changed lines. doc/unset.n: new paragraph (declared-only variable doesn't exist, so unset errors unless -nocomplain, but the declaration is removed; no longer listed by info vars / found by namespace which -variable; the only way to undo a declaration); doc/variable.n: one sentence pointing to it; SEE ALSO cross references. tests/var.test: var-7.18, var-7.19.
          • After man pages render, troff warnings 0/0 for both. tests var.test on main: 223 total, 219 passed, 4 skipped (memory), 0 failed.
          • Test added var-7.18 (unset of declared-only ns var: error message, then info vars and namespace which empty) and var-7.19 (same via variable link in a proc, both plain unset and unset -nocomplain). Pins current behaviour: pass on main and on 8.6 tip (main's var.test run with the 8.6 tcltest, -match var-7.1[89]: 2 passed).
          Full verdict

          REPRODUCED (main, 9.0, 8.6, 8.5, 8.4). Long-standing behaviour; the choice is between DOCS and a behaviour change.

          • Platform generic
          • Tested main, 9.0 tip, 8.6 tip, 8.5.19, 8.4.20. unset a::A on a variable A without a value errors "no such variable" and also removes ::a::A from info vars; unset -nocomplain removes it silently. Same inside a proc through variable C.
          • Cause generic/tclVar.c:TclPtrUnsetVarIdx computes the error from TclIsVarUndefined() but still calls UnsetVarStruct(), which ends with TclClearVarNamespaceVar(varPtr). CleanupVar() then frees the declared-only variable.
          • Prototype none. Keeping the declaration would leave no command that undoes variable A (unset -nocomplain is currently the only way), so this needs a decision. The cheaper option is to document the side effect in unset.n (and variable.n).
          • Suggested ticket comment Same on 8.4 through 9.1. unset clears the namespace-variable declaration even though it reports "no such variable", because the error is computed before UnsetVarStruct() runs. This is currently the only way to undo variable A, so I'd rather document it in unset.n than change it.
          🔧0ef291c1c22016memory corruption with trace + transchanpatchREPRODUCED (main, 8.6; 9.0 by code — same InvokeTclMethod)

          Open ticket 0ef291c1c2 · 26. Channel Transforms · filed 2016-09-20

          🔧 Since triage (updated 2026-11-11)

          Status now: Open

          • 🔧 2026-10-17 check-in 9c18d04ed6 on iortrans-reentrant-argv, not merged: Fix 0ef291c1c2: memory corruption with trace + transchan. InvokeTclMethod() used the command words preallocated per transform, so a handler which re-entered t
          • 🔧 2026-10-17 check-in 2ae9cc6780 on iortrans-reentrant-argv, not merged: Renumber test iortrans-bug-0ef291c1c2 to iortrans-12.1.
          • 💬 2026-10-17 serhiy.storchaka: Reproduced on main and 8.6 (heap-use-after-free under ASan; 9.0 has the same code): InvokeTclMethod() in tclIORTrans.c fills the command words preallocated for the transform, so when the handler re-enters the same transform (here the trace's [puts]), the nested call overwrites and frees the words of the outer call that is still running. Proposed fix in branch iortrans-reentrant-argv: copy the words for each call. …
          Full verdict

          REPRODUCED (main, 8.6; 9.0 by code — same InvokeTclMethod)

          • Platform generic
          • Tested reporter's script; release builds run silently, ASan+PURIFY builds of main and 8.6 report heap-use-after-free in InitArgsAndLocals ← InvokeTclMethod (tclIORTrans.c).
          • Cause generic/tclIORTrans.c:InvokeTclMethod writes the method name and arguments into the shared preallocated rtPtr->argv and evaluates it. When the handler re-enters the same transform (here an enterstep trace does [puts] to the transformed channel), the nested call overwrites the slots and frees its methObj while the outer Tcl_EvalObjv still uses that array. tclIORChan.c already builds a fresh command list per call and is not affected.
          • Prototype prototypes/0ef291c1c2.patch (main; 39 changed lines, test iortrans-bug-0ef291c1c2) — copies rtPtr->argv to a per-call array. The test fails under ASan without the fix and passes with it; ioTrans.test and ioCmd.test pass. Should apply to 9.0/8.6 with offsets.
          • Suggested ticket comment Cause: InvokeTclMethod in tclIORTrans.c reuses the shared rtPtr->argv, so a handler that re-enters the transform (via the trace's &#91;puts&#93;) overwrites the words of the outer call that is still running. Proposed fix: copy argv for each call, test iortrans-bug-0ef291c1c2.
          ⬜477c24cb392016rename ::set and lose custom return optionsREPRODUCED (main, 9.0, 8.6) as dgp restated it: bytecode leaks the return options of an earlier TCL_OK command into the final -options (the reverse…

          Open ticket 477c24cb39 · 18. Commands M-Z · filed 2016-08-31

          Full verdict

          REPRODUCED (main, 9.0, 8.6) as dgp restated it: bytecode leaks the return options of an earlier TCL_OK command into the final -options (the reverse of the original report)

          • Platform generic
          • Tested work-a2/477c24cb39.tcl: catch {set status [p1]} m o with p1 doing return -custom1 1 1 gives -custom1 1 -code 0 -level 0 when compiled and -code 0 -level 0 when set is called through a variable. catch {p1; list x} also keeps -custom1.
          • Cause TclExecuteByteCode does not reset iPtr->returnOpts after a TCL_OK result from an invoked command, so the options survive into the enclosing catch; non-compiled evaluation resets them per command. Same root cause as open ticket 2433843fff "return options can leak to next bytecode".
          • Prototype none (the obvious fix, clearing returnOpts after every INST_INVOKE, costs performance and needs a decision on how long options of a TCL_OK return should live; dkf suggested not fixing)
          • Suggested ticket comment Still present on 8.6/9.0/9.1: compiled code leaks the options of an earlier TCL_OK return into &#91;catch&#93; -options. This duplicates 2433843fff; suggest closing this one as a duplicate and keeping the discussion there.
          ⬜67a5eabbd32016refchan, coroutine, and postevent from the "watch" procREPRODUCED (main, 9.0, 8.6). The current behaviour is a TCT decision: the fix e066e24f36 was backed out in 2021, and again in 2024 with TCT appro…

          Open ticket 67a5eabbd3 · 60. NRE and coroutines · filed 2016-08-15

          Full verdict

          REPRODUCED (main, 9.0, 8.6). The current behaviour is a TCT decision: the fix e066e24f36 was backed out in 2021, and again in 2024 with TCT approval.

          • Platform generic
          • Tested main, 9.0 tip, 8.6 tip: the reporter's script (with a finalize proc added) gives bgerror coroutine "::c1" is already running
          • Cause chan postevent calls Tcl_NotifyChannel() synchronously (generic/tclIORChan.c). When it is called from the refchan's watch handler, which runs inside chan event inside the coroutine, the handler fires before yield.
          • Prototype none. The asynchronous fix exists and was deliberately rolled back (see the wiki page "Rationale for rollback of refchan event generation in core"); oehhar asked for a TIP, and a separate ticket if this is a documentation issue.
          • Suggested ticket comment Still reproduces on 9.1. Since the TCT rolled back the asynchronous chan postevent, the remaining options are a TIP or documenting in refchan.n/chan.n that chan postevent fires handlers synchronously, so watch must not call it directly (wrap it in after 0).
          ⬜cdd157d4842016file copy -force does an invalid call to openpatchREPRODUCED (main, 9.0, 8.6)

          Open ticket cdd157d484 · 37. File System · filed 2016-07-18

          🧪 Prototype check (2026-10-24)

          cdd157d484.patch: VERIFIED

          • Base main (tip d6dd6d0f71)
          • Before strace -e trace=open,openat tcltest vcdd.tcl (file copy -force of a 0640 file) -> openat(AT_FDCWD, "cdd/out.txt", O_WRONLY|O_CREAT|O_TRUNC, 0100640)
          • Patch applied cleanly (5 changed lines)
          • After openat(..., O_WRONLY|O_CREAT|O_TRUNC, 0640); the copy keeps 0o640, and a setuid 4750 source still gives a 0o4750 copy. Tests fCmd, fileSystem, cmdAH, unixFCmd: 0 failed
          • Test none possible (the mode bits are visible only via strace or a picky filesystem, with no hook in tcltest; the strace output above is the proof)
          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform unix
          • Tested strace of file copy -force in.txt out.txt with main (30d8ccce48): openat(..., "out.txt", O_WRONLY|O_CREAT|O_TRUNC, 0100664). Same code in 9.0 and 8.6 sources (unix/tclUnixFCmd.c TclUnixCopyFile).
          • Cause unix/tclUnixFCmd.c:TclUnixCopyFile passes the source's full st_mode, including the S_IFREG type bits, as the open() mode.
          • Prototype prototypes/cdd157d484.patch (5 changed lines, no test: only visible with strace or a picky filesystem). Masks the mode with 07777 as agreed in the comments; CopyFileAtts sets the final permissions afterwards anyway. With the fix strace shows mode 0640 for a 0640 source, the copy keeps 0o640, fCmd.test passes.
          • Suggested ticket comment Still present. Proposed fix: pass st_mode & 07777 to open() in TclUnixCopyFile (the permissions are still copied by CopyFileAtts afterwards).
          ⬜2923d528322016INST_STR_MATCH != TclStringMatchObj()patchREPRODUCED (main; same duplication in 9.0/8.6). Code duplication with diverging conditions, not a wrong result.

          Open ticket 2923d52832 · 10. Objects · filed 2016-03-30

          🧪 Prototype check (2026-10-24)

          2923d52832.patch: VERIFIED+TEST-ADDED

          • Base main (tip d6dd6d0f71)
          • Before the two copies still differ on the tip. Observable effect: set x [format %s abc]; set p [format %s a*]; set cmd string; $cmd match $p $x (direct, non-compiled call -> TclStringMatchObj) converts both untyped values to the Unicode string type: representation string/string (the compiled INST_STR_MATCH leaves them pure/pure)
          • Patch applied cleanly; test added (36 changed lines now)
          • After same script -> 1, representation pure/pure; tests string, stringComp, regexp, lsearch, switch: 2004 total, 1995 passed, 9 skipped, 0 failed
          • Test added string-11.56.$noComp (string match keeps plain strings): string-11.56.1 (testevalex, direct eval) fails before / passes after; string-11.56.0 (compiled) passes on both, as expected
          Full verdict

          REPRODUCED (main; same duplication in 9.0/8.6). Code duplication with diverging conditions, not a wrong result.

          • Platform generic
          • Tested code review of main. INST_STR_MATCH (tclExecute.c) promotes to Unicode if *either* argument has tclStringType. TclStringMatchObj (tclUtil.c, used by direct-eval string match and by regexp when the RE reduces to a glob) promotes if the *string* has tclStringType or no internal rep at all, so it makes a needless Unicode conversion of plain strings. Both then use the same TclByteArrayMatch / Tcl_StringCaseMatch fallbacks, so results agree; only speed and memory differ.
          • Cause two copies of the same dispatch logic that drifted apart
          • Prototype prototypes/2923d52832.patch (28 changed lines, no new test: the change is not visible from scripts). TclStringMatchObj takes the INST_STR_MATCH condition and INST_STR_MATCH calls TclStringMatchObj. string, lsearch, regexp and switch tests pass; full suite: only the pre-existing zipfs.test environment failures.
          • Suggested ticket comment The two still differ: TclStringMatchObj converts untyped strings to Unicode, INST_STR_MATCH does not. Results are the same, only speed differs. Proposed cleanup (prototype): INST_STR_MATCH calls TclStringMatchObj, which uses the "either side is already Unicode" rule.
          ⬜8527d8becf2016Commands cd, glob and normalize may fail on Windows if used with network locations containing NTFS junctionspatchREPRODUCED (main, 9.0, 8.6) for mapped network drives; UNC paths are not affected

          Open ticket 8527d8becf · 36. Pathname Management · filed 2016-03-02

          🧪 Prototype check (2026-10-24)

          8527d8becf.patch: VERIFIED

          • Base main (tip c092dae4af)
          • Before temporary net use M: \\localhost\C$ /persistent:no (deleted in the same ssh call; net use empty afterwards), tclsh91 8527m.tcl -> normalize M:/.../Foo/subdir -> C:/.../Bar/subdir, cd; pwd -> C:/..., VERDICT: REPRODUCED
          • Patch applied cleanly, 33 changed lines, not changed
          • After normalize M:/.../Foo/subdir -> M:/.../Foo/subdir, pwd -> M:/..., glob works, VERDICT: NOT-REPRODUCED; local C: junctions are still resolved. File tests admin 8 failed, low 6 failed, as baseline
          • Test none possible (needs a mapped network drive)
          Full verdict

          REPRODUCED (main, 9.0, 8.6) for mapped network drives; UNC paths are not affected

          • Platform Windows
          • Tested code reading of main win/tclWinFile.c; repro script win/8527d8becf.tcl (needs an elevated tclsh: creates a temporary share //localhost/tkt8527 with junctions Foo -> local Bar and Baz -> Q:\nosuch\Bar)
          • Cause win/tclWinFile.c:TclpObjNormalizePath follows every intermediate component with FILE_ATTRIBUTE_REPARSE_POINT via WinReadLinkDirectory and restarts normalization from the junction target. A junction's target is an absolute path on the machine that holds the volume (P:\Share\Bar on the server), so on a share it is interpreted on the client; cd/glob go through the normalized path and fail. Unix is not affected (symlinks are resolved by the client there by design).
          • Prototype none (needs Windows to verify; likely fix: skip reparse-point resolution when the path is UNC or GetDriveTypeW(root) == DRIVE_REMOTE, about 10 lines)
          • Suggested ticket comment Still present on main: TclpObjNormalizePath() resolves junctions on network paths using the server-local target. Not following reparse points on UNC/DRIVE_REMOTE paths should fix it.
          Windows VM, run 4
          • Platform Windows
          • Verdict REPRODUCED (main, 9.0, 8.6) for mapped network drives; UNC paths are not affected
          • Tested VM, junction Foo -> C:\...\Bar, accessed through the default admin share (no new share created). Via a temporary non-persistent net use M: \\localhost\C$ (created and deleted in the same ssh session, work-win4/8527m.tcl): file normalize M:/.../Foo/subdir → C:/.../Bar/subdir and cd + pwd lands on C: on 9.1b1, 9.0.5, 8.6.18 (it works here only because the "server" is localhost; on a real server the client-side C: path is wrong, as reported). Via UNC //localhost/C$/.../Foo/subdir (work-win4/8527c.tcl, adapted from win/8527d8becf.tcl, which needs net share): paths stay under the share, cd/glob work, all three versions.
          • Cause win/tclWinFile.c TclpObjNormalizePath follows every intermediate reparse point via WinReadLinkDirectory, also on a network drive, where the junction target is a path on the server
          • Prototype prototypes/8527d8becf.patch (main, 33 changed lines, no test: needs a network drive; applies to 9.0 with offsets, one hunk fails on 8.6): don't follow links when WinIsRemotePath() (UNC, or GetDriveTypeW == DRIVE_REMOTE). Built on the VM: M:/.../Foo/subdir stays M:/.../Foo/subdir, cd/pwd/glob work; fileSystem, fileName, fCmd, cmdAH 0 failed; winFCmd 3 and winFile 5 failures as unpatched. Tested on Windows.
          • Suggested ticket comment Still happens with 8.6.18, 9.0.5 and 9.1 on Windows 11 for mapped network drives (UNC paths are fine): TclpObjNormalizePath() resolves the junction to the server-side target path. Proposed fix: don't follow reparse points on remote drives (GetDriveTypeW() == DRIVE_REMOTE) or UNC paths; tested with a mapped drive.
          🔧6141c151862015Transchan loses input data when written topatchREPRODUCED (main, 9.0, 8.6)

          Open ticket 6141c15186 · 26. Channel Transforms · filed 2015-12-27

          🔧 Since triage (updated 2026-11-11)

          Status now: Open

          • 🔧 2026-10-17 check-in b31ec18fbe on transform-write-keep-input, not merged: Fix 6141c15186: a transform on a non-seekable channel lost the buffered input when written to. WillWrite() discards the input only if seeking back succeeded,
          • 🔧 2026-10-17 check-in 102b356c60 on transform-write-keep-input, not merged: Renumber test iortrans-bug-6141c15186 to iortrans-13.1.
          • 🔧 2016-03-08 check-in 8b47978cb6 on aku-tkt-6141c15186, not merged: Update to trunk work
          • 💬 2026-10-17 serhiy.storchaka: Still in 8.6, 9.0 and 9.1: WillWrite() (from 5af0d249de, for [2901998]) drops the buffered input and then seeks back, and a transform always has a seek proc, also over a socket, so the seek fails after the input is gone. Proposed fix in branch transform-write-keep-input, along the lines dgp suggested: discard the input only if the seek back succeeded, and make the transform seek fail early when the channel below…
          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform generic
          • Tested attachment transchan-bug.tcl is missing from the repository; own repro work-a4/t614.tcl (localhost socket, pass-through [chan push] transform, non-blocking): server sends 5 lines, client reads 1 and writes "Hi!" → only 1 line received on main/9.0.5/8.6.18; 5 without the push or without the write. With the prototype: 5.
          • Cause generic/tclIO.c:WillWrite discards buffered input whenever the top channel type has a wideSeekProc, then seeks back; tclIORTrans.c's type always has one even when the base channel (socket) cannot seek, so the seek fails after the input is already gone. aku's fix on branch aku-tkt-6141c15186 (128ad6929b, 2016-01-28: per-transform channel type without seek) was never merged.
          • Prototype prototypes/6141c15186.patch (main; 65 changed lines, of which 31 are the test iortrans-bug-6141c15186, uses a |tclsh script r+ pipe) — WillWrite discards the input only if the seek back succeeded, and ReflectSeekWide fails up front (before clear/flush) when the parent cannot seek. Test fails without the fix ("a d"), passes with it; ioTrans/io/ioCmd/chanio tests pass.
          • Side finding a *blocking* [gets] on a transformed socket hangs even with 5 lines available (main, 9.0, 8.6): ReflectInput loops on Tcl_ReadRaw until the full 4096-byte request is satisfied. Separate issue; worth a search for an existing ticket.
          • Suggested ticket comment Still in 9.0/main: WillWrite() in tclIO.c drops the input buffer before trying to seek back, and a transform always looks seekable. Proposed fix: discard only if the seek succeeds, and make ReflectSeekWide fail early for a non-seekable base; test iortrans-bug-6141c15186. Branch aku-tkt-6141c15186 was never merged.
          ⬜972c35d41d2015file copy "permission denied" for Windows symlinkspatchREPRODUCED (main, 9.0, 8.6) for file symlinks (copied as a regular file); directory symlinks are copied as links when the user may create symlinks …

          Open ticket 972c35d41d · 37. File System · filed 2015-11-23

          🧪 Prototype check (2026-10-24)

          972c35d41d.patch: VERIFIED

          • Base main (tip c092dae4af)
          • Before tclsh91 972c35d41d.tcl (admin) -> file copy bar.txt baz.txt -> ok, baz.txt ... (file): the file symlink is copied as a regular file
          • Patch applied cleanly, 48 changed lines, not changed
          • After same -> baz.txt ... (link). The script still prints "is link: 0" because file readlink fails on symlinks, a separate known issue. Unprivileged fallback, now tested: a symlink made as admin and copied as basic user (972low.tcl) gives copy: 0, baz.txt type=file content=Hello (same as unpatched; no error). winFCmd, fCmd, cmdAH: admin 3 failed (as baseline), low 0 failed (winFCmd-2.18 skipped by winSymLinks)
          • Test existing winFCmd-2.18 FAILED with win/tclWinFCmd.c reverted (file tf1), passes with the patch
          Full verdict

          REPRODUCED (main, 9.0, 8.6) for file symlinks (copied as a regular file); directory symlinks are copied as links when the user may create symlinks (the "permission denied" for non-admins not tested: the ssh account is an elevated admin)

          • Platform Windows
          • Tested Linux main: file/dir symlinks are copied as links (OK); Windows repro script win/972c35d41d.tcl (needs elevated tclsh or Developer Mode; also checks [file delete] of a dir symlink, bug 3c13c4ed31)
          • Cause win/tclWinFCmd.c:DoCopyFile calls CopyFileW without COPY_FILE_COPY_SYMLINK (CopyFileExW), so a file symlink is followed. Directory reparse points go to TclWinSymLinkCopyDirectory -> NativeWriteReparse (win/tclWinFile.c), which was written for junctions; IO_REPARSE_TAG_SYMBOLIC_LINK is even defined as 0 there (the real tag is 0xA000000C), and setting a symlink reparse point needs SeCreateSymbolicLinkPrivilege, giving "permission denied" for non-admins.
          • Prototype none (needs Windows to verify; file part is a one-line CopyFileExW(..., COPY_FILE_COPY_SYMLINK) change, directory part should use CreateSymbolicLinkW with the read target)
          • Suggested ticket comment Still in main: DoCopyFile() uses CopyFileW() which follows file symlinks, and directory symlinks are copied with the junction code (NativeWriteReparse). CopyFileExW() with COPY_FILE_COPY_SYMLINK should handle file symlinks (and probably directory symlinks too; to be verified).
          Windows VM, run 4
          • Platform Windows
          • Verdict REPRODUCED (main, 9.0, 8.6) for file symlinks (copied as a regular file); directory symlinks are copied as links when the user may create symlinks (the "permission denied" for non-admins not tested: the ssh account is an elevated admin)
          • Tested VM win/972c35d41d.tcl, work-win4/sl2.tcl, sl3.tcl on 9.1b1, 9.0.5, 8.6.18: file copy bar.txt baz.txt (bar.txt a symlink) gives a regular file; file copy src/myres dest1 (dir symlink ..\res) gives an identical SYMLINKD reparse point (fsutil). Further findings, all versions: file readlink fails for symlinks ("invalid argument" for dir symlinks, "not a directory" for file symlinks; only junctions work), so the repro's link check reports 0 even for correct copies; file delete src/myres (dir symlink) fails with "is a directory" (bug 3c13c4ed31); file delete -force of a directory containing a file symlink fails with "invalid argument".
          • Cause win/tclWinFCmd.c DoCopyFile uses CopyFileW, which follows symlinks; directory reparse points go to TclWinSymLinkCopyDirectory (needs SeCreateSymbolicLinkPrivilege for symlink tags). Readlink: win/tclWinFile.c WinReadLinkDirectory handles only mount points.
          • Prototype prototypes/972c35d41d.patch (main only, 48 changed lines, test winFCmd-2.18 with a winSymLinks probe constraint): DoCopyFile copies via CopyFileExW(COPY_FILE_COPY_SYMLINK), falling back to CopyFileW on ERROR_PRIVILEGE_NOT_HELD so unprivileged users keep the old behaviour. Built on the VM: the file symlink copy is now a link with the target's content; winFCmd-2.18 passes (fails unpatched: file tf1); winFCmd.test 172 passed / 3 failed as unpatched, fCmd.test and cmdAH.test 0 failed. The unprivileged fallback is untested. Directory symlinks for non-admins, readlink and recursive delete not addressed. Tested on Windows.
          • Suggested ticket comment On Windows 11 (8.6.18, 9.0.5, 9.1) &#91;file copy&#93; of a file symlink still copies the target, while a directory symlink is copied as a link (if the user may create symlinks). Proposed fix for the file case: CopyFileExW() with COPY_FILE_COPY_SYMLINK and a fallback to the old copy without the privilege, test winFCmd-2.18. Also noticed: &#91;file readlink&#93; fails on symlinks (only junctions work) and &#91;file delete -force&#93; fails on a directory holding a file symlink.
          ⬜3e30dc62e02015compiled [lappend] in proc fails to fire read tracesREPRODUCED (main, 9.0, 8.6, 8.5)

          Open ticket 3e30dc62e0 · 47. Bytecode Compiler · filed 2015-11-16

          Full verdict

          REPRODUCED (main, 9.0, 8.6, 8.5)

          • Platform generic
          • Tested main c3bf6dd052 build, 9.0.5, 8.6.18, 8.5.19. Global lappend x a/append x a fire the read trace; in a proc compiled lappend x a and append x a (and append x a b) do not; [subst lappend] x a and lappend x a b (INST_LAPPEND_LIST) do.
          • Cause tclExecute.c INST_APPEND_*/INST_LAPPEND_* call TclPtrSetVarIdx without TCL_TRACE_READS (tclVar.c comment: removed on purpose for [3057639], env lappend), while the direct-eval commands and INST_LAPPEND_LIST read the variable through Tcl_ObjGetVar2 and fire traces. Tests append-7.2..7.4/trace-3.x want traces; append-9.*/appendComp-9.* want none for nonexistent env/array elements. [3057639] is still Open.
          • Prototype none (needs a decision on the intended semantics, compiled/direct and append/lappend all need to change together; not a small fix)
          • Suggested ticket comment Still the same in 9.1, 9.0 and 8.6: compiled single-value append/lappend skip read traces (no TCL_TRACE_READS in INST_(L)APPEND_*), INST_LAPPEND_LIST and direct eval fire them. Needs a decision together with &#91;3057639&#93;.
          ⬜39bb9b3ebb2015deleting a coroutine bypasses try..finally handlersREPRODUCED (main, 9.0, 8.6)

          Open ticket 39bb9b3ebb · 60. NRE and coroutines · filed 2015-11-11

          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform generic
          • Tested main build, 9.0.5, 8.6.18: neither rename foo {} nor namespace delete runs the finally clause; the unset trace workaround does fire.
          • Cause tclBasic.c DeleteCoroutine -> RewindCoroutine sets eePtr->rewind=1 and resumes the coroutine; tclExecute.c TEBC then jumps straight to abnormalReturn (checkForCatch / after INST_INVOKE) skipping every catch range, so the bytecoded finally body never runs (same mechanism as interp cancel -unwind). Only NRE callbacks and var/command traces run.
          • Prototype none (semantic decision: running script code during rewind needs a design for errors/yield inside the handler; Miguel considered it not a bug)
          • Suggested ticket comment Still so in 9.1/9.0/8.6: deletion rewinds the coroutine (execEnv rewind flag) and TEBC skips all catch ranges, like &#91;interp cancel -unwind&#93;. Running finally clauses here would be a semantic change (what if the handler yields or errors?); needs a decision, maybe a TIP.
          🔧5ea71fdcd32015Regexp fails to enforce LACON restrictions inside nested parenthesesREPRODUCED (main, 9.0, 8.6); Proposed fix in branch regexp-lacon-nested-parens (unmerged) fixes it

          Open ticket 5ea71fdcd3 · 43. Regexp · filed 2015-11-07

          🔧 Since triage (updated 2026-11-11)

          Status now: Open

          • 🔧 2026-10-16 check-in 055b73e260 on regexp-lacon-nested-parens, not merged: Fix 5ea71fdcd3: restrictions of lookahead constraints were not applied nested. Rebase to trunk
          Full verdict

          REPRODUCED (main, 9.0, 8.6); Proposed fix in branch regexp-lacon-nested-parens (unmerged) fixes it

          • Platform generic
          • Tested main/9.0/8.6 give xy y and x {}. Branch tip 055b73e260 built in wt-a1: first example "invalid backreference number", second x; x(?=(a)(b))(\w) on xabc -> xa a (renumbered as documented); (?:...) inside LACON unaffected. reg/regexp/regexpComp.test pass on the branch (1648, 0 failed). Dry-run merge into current main: code merges cleanly, only conflict is changes.md.
          • Cause generic/regcomp.c parse() recursion for nested parens did not pass the LACON context type.
          • Prototype none (Serhiy's branch exists)
          • Existing work regexp-lacon-nested-parens (Serhiy bac9120832, rebased to trunk by Jan 055b73e260, 2026-10-16) unmerged
          • Suggested ticket comment none needed; breaking change already flagged by Serhiy. Branch still merges cleanly into main apart from changes.md.
          ⬜880855cb092015Tcl C API OO symbols are not exported to libtcl.sopatchREPRODUCED (main, 9.0, 8.6) — intentional since [3010352]

          Open ticket 880855cb09 · 35. TclOO Package · filed 2015-08-26

          🧪 Prototype check (2026-10-24)

          880855cb09.patch: VERIFIED. The fix works, but it also exports the 18 internal TclOO* functions, which may not be wanted; the policy question from [3010352] remains open.

          • Base main (tip d6dd6d0f71)
          • Before nm libtcl9.1.so -> t Tcl_NewObjectInstance (local); linking work-a3/oolink.c (calls Tcl_GetObjectFromObj, Tcl_GetObjectAsClass, Tcl_NewObjectInstance) with -ltcl9.1 -> "undefined reference" for all three
          • Patch applied cleanly (7 changed lines); full rebuild (rm *.o, since tcl.h changed)
          • After the dynamic symbol table gains 53 symbols and loses none: 35 public Tcl_* (tclOODecls.h, tclPlatDecls.h) and 18 internal TclOO* from tclOOIntDecls.h (TclOOMakeProcMethod, TclOOObjectCmdCore, TclOOIsReachable, ...). The triage mentioned only the 35. oolink links and runs (o2 m -> hi). Tests oo, ooNext2, ooProp, ooUtil, load: 0 failed
          • Test none possible in the Tcl test suite (linkage only; nm plus the direct-link program is the proof)
          Full verdict

          REPRODUCED (main, 9.0, 8.6) — intentional since [3010352]

          • Platform unix (shared builds; on Windows MODULE_SCOPE has no effect)
          • Tested nm -D of main build, 9.0.5 and 8.6.18 libtcl: Tcl_NewObjectInstance etc. are local (t); with the prototype 35 more Tcl_* symbols are exported and a C program linking -ltcl9.1 calls Tcl_GetObjectFromObj/Tcl_NewObjectInstance directly.
          • Cause generic/tcl.h redefines TCLAPI as MODULE_SCOPE when BUILD_tcl before including tclPlatDecls.h, so tclOODecls.h (included later via tclOO.h) keeps that definition. Stubs users (all extensions using Tcl_InitStubs/TclOOInitializeStubs) are unaffected.
          • Prototype prototypes/880855cb09.patch (7 changed lines, no test; removes the TCLAPI override, verified with nm and a direct-link C program, oo*.test pass). Note: the unix Makefile does not rebuild objects after a tcl.h change; remove tclOO*.o etc. by hand.
          • Suggested ticket comment Still MODULE_SCOPE by design of &#91;3010352&#93; (tcl.h redefines TCLAPI before tclOODecls.h). dkf says they should have normal linkage; removing the override in tcl.h exports them (7 lines). Needs a decision between the two positions.
          ⬜ef28eb1f152015TCL_FILE_EVENTS cannot drive async I/O aloneREPRODUCED (main, 9.0, 8.6) — design limitation

          Open ticket ef28eb1f15 · 25. Channel System · filed 2015-08-10

          Full verdict

          REPRODUCED (main, 9.0, 8.6) — design limitation

          • Platform generic
          • Tested work-a4/tef.tcl with tcltest (testfilehandler oneevent = Tcl_DoOneEvent(TCL_FILE_EVENTS|TCL_DONT_WAIT)): pipe with "one\ntwo\n", readable handler doing one [gets]; 50 file-only event passes deliver only "one", the second line waits until a timer event ([update]) — identical on main, 9.0 tip, 8.6 tip.
          • Cause generic/tclIO.c: readable events for data already in the channel buffer are re-fired by a SYNTHETIC_EVENT_TIME timer (ChannelTimerProc, ~line 8797/8843), [chan copy] starts with a 0 ms timer (ZeroTransferTimerProc), and [chan postevent] since e066e24f36 also goes through timers; so TCL_TIMER_EVENTS are required as well. dgp's 2021 comment that e066e24f36 did not address it still holds.
          • Prototype none (needs a redesign: channel-ready events queued as file events instead of timers; touches all of tclIO.c's notification)
          • Suggested ticket comment Still true on 9.0/main: buffered input is re-signalled via a timer in tclIO.c (ChannelTimerProc), so Tcl_DoOneEvent(TCL_FILE_EVENTS) alone stalls with data in the buffer (verified with testfilehandler oneevent). Fixing needs channel events queued as file events; at least Notifier.3 could say that TCL_TIMER_EVENTS are needed too.
          ⬜3eace1b21f2015'glob -dir $path *' doesn't work for deeply nested pathspatchREPRODUCED (main, 9.0, 8.6) on Windows 11 with the default LongPathsEnabled=0: TIP 744 does not fix it (its long-path tests are skipped without the…

          Open ticket 3eace1b21f · 36. Pathname Management · filed 2015-07-07

          🧪 Prototype check (2026-10-24)

          3eace1b21f.patch: VERIFIED

          • Base main (tip c092dae4af)
          • Before tclsh91 3eace1b21f.tcl (temp dir moved to scratch) -> glob -dir path *: empty, glob subdir/*: 0; g3.tcl (432-char dir, 2 files) -> {} 0 0
          • Patch applied cleanly, 49 changed lines, not changed
          • After glob -dir path *: path, glob subdir/*: 1; g3 -> {f1 f2} 2 2. Still failing, as the triage said (not addressed by the patch): cd into it, recursive file delete -force (deleted: 0), and glob -dir //?/$path * (empty before and after). File tests (winFCmd fCmd fileSystem fileName cmdAH winFile): admin 8 failed, low 6 failed, the same lists as baseline
          • Test existing winFCmd-20.3 FAILED with win/tclWinFile.c reverted ({} 0 0), passes with the patch
          Full verdict

          REPRODUCED (main, 9.0, 8.6) on Windows 11 with the default LongPathsEnabled=0: TIP 744 does not fix it (its long-path tests are skipped without the policy; with the policy set the path would work, not tested since HKLM must not be changed)

          • Platform Windows
          • Tested Linux main and 8.6 (no issue, as expected); code reading: main win/tclWinFile.c TclNativeCreateNativeRep adds the \\?\ prefix for paths > MAX_PATH when long paths are not enabled, tclsh manifest is longPathAware, and tests/winFCmd.test has dir-glob-longpath-0/1/2 (glob, glob -directory, glob -path). Repro script win/3eace1b21f.tcl (reporter's script, non-interactive) for main, 9.0 and 8.6.
          • Cause 9.0/8.6 pass the plain path (> MAX_PATH) to FindFirstFileW in TclpMatchInDirectory.
          • Prototype none (fixed on main by TIP 744; backport is a feature-sized change)
          • Suggested ticket comment Fixed in 9.1 by TIP 744 (Windows long path support, 5e98a389b8); tests dir-glob-longpath-* cover glob -directory. Not planned for 9.0/8.6.
          Windows VM, run 4
          • Platform Windows
          • Verdict REPRODUCED (main, 9.0, 8.6) on Windows 11 with the default LongPathsEnabled=0: TIP 744 does not fix it (its long-path tests are skipped without the policy; with the policy set the path would work, not tested since HKLM must not be changed)
          • Tested VM win/3eace1b21f.tcl and work-win4/g3.tcl (432-char dir with two files): glob -directory $dir * and glob $dir/* return nothing (no error) on 9.1b1 and 9.0.5; 8.6.18 raises "couldn't read directory .../*: no such file or directory". Also seen: file delete -force of such a tree returns success but deletes nothing (9.1b1, 8.6.18, work-win4/g5.tcl); cd into it fails; file mkdir fails for an intermediate dir of 248–260 chars (CreateDirectoryW limit is MAX_PATH-12, the native rep adds \\?\ only above MAX_PATH).
          • Cause win/tclWinFile.c TclpMatchInDirectory builds the FindFirstFileExW search string with Tcl_UtfToWCharDString from the UTF-8 dir + pattern, bypassing TclNativeCreateNativeRep, which would add the \\?\ prefix for paths > MAX_PATH; the failure is ERROR_PATH_NOT_FOUND, which 9.x IsNoSuchFileError treats as "no match". Recursive delete: win/tclWinFCmd.c TraverseWinTree appends components to the root's native path the same way.
          • Prototype prototypes/3eace1b21f.patch (main only, 49 changed lines, test winFCmd-20.3): when the search string is ≥ MAX_PATH and has no \\?\ prefix, add \\?\ (drive) or \\?\UNC (UNC) and convert / to \. Built on the VM: the repro's glob -dir and glob subdir/* now work, winFCmd-20.3 passes (fails unpatched), winFCmd.test 172 passed / 3 failed (6.1, 6.13, 9.3, as unpatched), fCmd.test, fileSystem.test, fileName.test 0 failed. Does not apply to 9.0/8.6 (different code there, needs a backport). Recursive delete and cd not addressed. Tested on Windows.
          • Suggested ticket comment Still fails on Windows 11 with 8.6.18, 9.0.5 and 9.1 unless LongPathsEnabled is set: TclpMatchInDirectory() builds the FindFirstFileExW() pattern without the \\?\ prefix, so 9.x silently returns no matches (8.6 reports an error). Proposed fix: add the prefix for long search paths, test winFCmd-20.3. Note &#91;file delete -force&#93; of such a tree also silently does nothing (TraverseWinTree, same cause).
          ⬜57e44de76e2015Array Read Trace fires only if Array has been written topatchREPRODUCED (main, 9.0, 8.6, 8.4)

          Open ticket 57e44de76e · 07. Variables · filed 2015-07-05

          🧪 Prototype check (2026-10-24)

          57e44de76e.patch: VERIFIED+TEST-ADDED (doc + behaviour-pinning tests)

          • Base main (tip d6dd6d0f71); also applies to core-8-6-branch (offsets only)
          • Before trace.n says a trace on a whole array "is invoked whenever any element of the array is manipulated", and that a traced undefined variable is created without a value; nothing about element reads before it becomes an array. Behaviour (work-a1/57e4.tcl, identical on main and 8.6 tip): trace add variable a read tr; set a(x) -> can't read "a(x)": no such variable, trace not called (also compiled, from a proc); after array set a {} the same read calls the trace (read x) and returns the value it sets; a scalar read set b of a traced undefined variable does call the trace; trace add variable c(e) read tr makes c an array and its read trace fires.
          • Patch 59 changed lines. doc/trace.n (trace add variable): such a variable is not an array yet; reading an element fails with "no such variable" without invoking a read trace on name; whole-array traces fire for elements only once the variable is an array (after an element write or array set name {}); an element trace makes the variable an array. tests/trace.test: trace-41.1..41.3.
          • After man page renders, troff warnings 0/0. trace.test on main: 293 total, 289 passed, 4 skipped, 0 failed.
          • Test added trace-41.1 (uncompiled read: error, no trace call, array exists 0; after array set tvar {} the trace fires and supplies the value), trace-41.2 (same through a compiled read in a proc, array made by an element write), trace-41.3 (element trace makes the variable an array and fires). Pins current behaviour: pass on main and on 8.6 tip (main's trace.test run with the 8.6 tcltest, -match trace-41.*: 3 passed).
          Full verdict

          REPRODUCED (main, 9.0, 8.6, 8.4)

          • Platform generic
          • Tested main build, 9.0.5, 8.6.18, 8.4.20; global and in a proc (compiled INST_LOAD_ARRAY): reading a(x) from a traced but undefined variable gives "no such variable" without firing the trace; after array set a {} the trace fires and can supply the value.
          • Cause tclVar.c Tcl_ObjGetVar2 (and the compiled load path) look up with createPart1=0, so TclLookupArrayElement returns NOSUCHVAR for an undefined non-array variable before any trace is called; a write uses createPart1=1 and turns it into an array, so later reads fire. trace.n only promises element traces for "a whole array", and an undefined traced variable is not yet an array.
          • Prototype none (calling read traces on an undefined non-array variable would change semantics on every read path, compiled and not; better a doc note)
          • Suggested ticket comment Still the same in 9.1/9.0/8.6. A read with createPart1=0 fails before traces run because the traced variable is not an array yet; array set a {} before adding the trace is the workaround. Suggest documenting this in trace.n rather than changing the lookup.
          ⬜58b96f67442015Difference in [info commands] behaviour in global namespacepatchREPRODUCED (main, 9.0, 8.6)

          Open ticket 58b96f6744 · 17. Commands I-L · filed 2015-06-11

          🧪 Prototype check (2026-10-24)

          58b96f6744.patch: VERIFIED

          • Base main (tip d6dd6d0f71)
          • Before tclsh r58b.tcl (namespace path ::tcl::mathfunc in ::) -> exact: <> glob: <signbit sin sinh>
          • Patch applied cleanly
          • After exact: <sin> glob: <signbit sin sinh>; tests namespace (315/0 failed), namespace-old, info, cmdIL: 0 failed
          • Test existing namespace-51.14.1 fails before / passes after
          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform generic
          • Tested main (own build), 9.0.5, 8.6.18. After namespace path ::tcl::mathfunc in ::, info commands sin returns "" while info commands si* returns sin sinh (the pattern case already follows the path); inside a namespace both work.
          • Cause generic/tclCmdIL.c:InfoCommandsCmd, exact-name fast path: the namespace path is searched only if ((nsPtr != globalNsPtr) && !specificNsInPattern), so the global namespace's own path is skipped. Present since TIP 229 (f1050beab4). aspect's fix on branch [bug-58b96f6744] (a756018479, 2017-02-12) is the same one-line change; never merged (the branch has only that check-in, marked closed-color by dkf).
          • Prototype prototypes/58b96f6744.patch (11 changed lines, test namespace-51.14.1; namespace/info/cmdIL tests pass; the test fails without the fix)
          • Suggested ticket comment Still present on main/9.0/8.6: the exact-name fast path in InfoCommandsCmd skips the namespace path when the current namespace is ::, while the glob path does not. Proposed fix (same as a756018479 on branch bug-58b96f6744, rebased onto main) with test namespace-51.14.1.
          ⬜0e7c09eb492015tclsh in a command pipeline: ::tcl_interactive is 0, but interpreter is in interactive mode (halfway).REPRODUCED (main, 9.0, 8.6). Behaviour as described; changing it is a behaviour change (TIP material)

          Open ticket 0e7c09eb49 · 69. Other · filed 2015-04-14

          Full verdict

          REPRODUCED (main, 9.0, 8.6). Behaviour as described; changing it is a behaviour change (TIP material)

          • Platform generic
          • Tested tclsh < script with puts too many arguments followed by another command: it prints the error, runs the next command and exits 0. tclsh script exits 1. Same on main, 9.0.5, 8.6.18.
          • Cause generic/tclMain.c Tcl_MainEx: when there is no script argument it always runs the REPL loop, and is.tty only switches prompts and result printing. Errors are printed and ignored, and the exit code stays 0.
          • Prototype none. The reporter's 4-line patch (stop with exit code 1 on the first error when stdin is not a tty) is a visible behaviour change for anything that feeds scripts to tclsh on stdin and expects it to carry on (test suites, echo cmd | tclsh idioms). It needs a TIP or at least core agreement, and the attached diff (dd3bf62e82) is not in the repository.
          • Suggested ticket comment Still the case in 8.6.18, 9.0 and 9.1: tclsh < file ignores errors and exits 0, tclsh file exits 1. Making non-tty stdin stop on the first error is a compatibility change and needs a TIP. Until then, tclsh /dev/stdin gives the script semantics in a pipeline.
          ⬜50f67a3ace2014INST_START_CMD bug?REPRODUCED (main, 9.0, 8.6; for source not in 8.5)

          Open ticket 50f67a3ace · 47. Bytecode Compiler · filed 2014-11-27

          Full verdict

          REPRODUCED (main, 9.0, 8.6; for source not in 8.5)

          • Platform generic
          • Tested main build, 9.0.5, 8.6.18, 8.5.19 with set x [proc set args {puts FOO}]42: tclsh TT (direct eval) prints FOO and x stays unset everywhere; source TT sets x on 8.6+ (FOO and unset on 8.5); stdin (whole-command compile) sets x on all versions.
          • Cause the compiled set is inlined as INST_STORE_*; INST_START_CMD checks the compile epoch only before the command's words are substituted, so a redefinition from inside an argument is not seen until the next command. source is compiled since 42090a4d6 (NRE source in 8.6).
          • Prototype none (needs a second epoch check after word substitution for every inlined command, or the Tcl.n change dkf proposed: resolve the command after reading the first word; both are design decisions)
          • Suggested ticket comment Unchanged in 9.1/9.0/8.6: INST_START_CMD checks the epoch before the arguments are substituted, so an inlined command redefined from its own arguments still runs the old compiled version. Fixing it means another epoch check after the words, or defining that the command is resolved after its first word (dkf's proposal, a TIP).
          ✅fdcc860df52014Test suite tests not running independentpatchREPRODUCED (main, 9.0, 8.6) for cmdMZ-return-3.* only; the other ~190 listed tests now pass standalone

          Open ticket fdcc860df5 · 69. Other · filed 2014-11-13

          ✅ Since triage (updated 2026-11-11)

          Status now: Closed (Fixed)

          • 🔧 2026-10-20 check-in acd6e13723 on cmdMZ-return-standalone, in main: Fix fdcc860df5: cmdMZ-return-3.* set ::errorCode and ::errorInfo first, so they pass when run alone.
          • 🔧 2026-10-21 check-in 03c91ff6ac on trunk, in main, 9.0, 8.6: Fix fdcc860df5: Test suite tests not running independent
          • 🔧 2026-10-21 check-in 8f4fc33974 on core-9-0-branch, in main, 9.0: Fix fdcc860df5: Test suite tests not running independent
          • 🔧 2026-10-21 check-in eeeec49664 on core-8-6-branch, in main, 9.0, 8.6: Fix fdcc860df5: Test suite tests not running independent
          • 💬 2026-10-20 serhiy.storchaka: Rechecked on main by running every test from the 2015 list alone. Only cmdMZ-return-3.* still depend on earlier tests: they read ::errorCode/::errorInfo, which don't exist until some earlier test has raised an error. Proposed fix in branch acd6e13723: set both before the loop. The tcltest-21.x/24.x/25.3 entries are not dependencies: these tests define nested tests named e.g. tcltest-21.4.0, which an exact -match…
          • 💬 2026-10-21 jan.nijtmans: +1 Many thanks!
          • 💬 2026-10-21 serhiy.storchaka: Merged into trunk 03c91ff6ac, backported to core-9-0-branch 8f4fc33974 and core-8-6-branch eeeec49664.
          Full verdict

          REPRODUCED (main, 9.0, 8.6) for cmdMZ-return-3.* only; the other ~190 listed tests now pass standalone

          • Platform generic (test suite)
          • Tested main build: every test named in the 2015 list run alone (all.tcl -file X.test -match NAME); 24 fail. 14 are tcltest-21.x/24.x/25.3, which are not dependencies: they define nested tests tcltest-21.4.0 etc. that the exact -match pattern filters out (with -match 'tcltest-21.4*' all pass). The 10 cmdMZ-return-3.x failures are real; with -match 'cmdMZ-return-3.*' 3.0 still fails on main, 9.0 and 8.6 tips.
          • Cause tests/cmdMZ.test cmdMZ-return-3.* read $::errorCode/$::errorInfo, which do not exist until some earlier test has raised an error.
          • Prototype prototypes/fdcc860df5.patch (3 changed lines, sets ::errorCode/::errorInfo before the loop; cmdMZ-return-3.0, 3.9, 3.* alone and the whole cmdMZ.test pass)
          • Suggested ticket comment Rechecked on main: of the listed tests only cmdMZ-return-3.* still depend on earlier tests (they read ::errorCode/::errorInfo before any error has set them); proposed fix: initialize both before the loop. The tcltest-2x.y entries are not dependencies, they need -match 'tcltest-21.4*' because the nested test names have a .0 suffix. Everything else in the list passes standalone now, so this can be closed after the cmdMZ fix.
          ⬜027ba061a82014Standard channels do not exist on new threads on WindowsREPRODUCED (main, 9.0, 8.6)

          Open ticket 027ba061a8 · 69. Other · filed 2014-09-25

          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform Windows (wish only)
          • Tested not on Linux — there threads get the real fd 0/1/2 so the issue cannot occur; code reading only.
          • Cause (by code): std channels are thread-specific (tclIO.c tsdPtr->stdinChannel…). In wish on Windows there are no std handles (GUI subsystem); Tk's console (tk/generic/tkConsole.c Tk_InitConsoleChannels → Tcl_SetStdChannel) installs console channels only in the thread that created the console, and TclpGetDefaultStdChannel finds no OS handles in a new thread. So this is really a Tk/console design point, as dgp noted in 2014.
          • Prototype none (needs a decision: e.g. Thread/Tk passing the console channels to new threads, or null channels)
          • Suggested ticket comment Cause: std channels are per-thread; in wish on Windows they are Tk console channels created only in the main thread, and a new thread has no OS std handles. Windows repro script prepared; probably belongs to Tk.
          Windows VM, run 2
          • Platform Windows (wish)
          • Verdict REPRODUCED (main, 9.0, 8.6)
          • Tested VM interactive session (scheduled task, wish started by wscript without std handles, main window withdrawn; script work-win2/027c.tcl) with Thread 3.0.7 / 2.8.13: the main thread has stderr/stdin/stdout (Tk console channels), a thread::create thread has none — puts stdout there fails with can not find channel named "stdout". Same on Tk/Tcl 9.1b1, 9.0.5, 8.6.18.
          • Cause std channels are per-thread (tclIO.c); Tk's console channels are installed only in the main thread, and a new thread finds no OS std handles in a GUI-subsystem process
          • Prototype none (design decision, probably for Tk/Thread)
          • Suggested ticket comment Confirmed with wish 8.6.18, 9.0.5 and 9.1 on Windows 11: threads created with thread::create have no stdout/stderr. The main thread's std channels are Tk console channels that exist only in that thread; a fix would need Tk (or Thread) to provide channels for new threads.
          ✅2b874804b52014Make Install fails on Windows looking for tcldde14sx.libpatch 86patchREPRODUCED (8.6 only); NOT-REPRODUCED on main (9.0 has the same makefile logic)

          Open ticket 2b874804b5 · 53. Configuration and Build Tools · filed 2014-08-15

          ✅ Since triage (updated 2026-11-11)

          Status now: Closed (Fixed)

          • 🔧 2026-10-20 check-in 0e4b74d5ed on buildall-static-dlls-8-6, in main, 9.0, 8.6: Fix 2b874804b5: the static pass of buildall.vc.bat also builds "dlls", so "install" finds the static dde/registry libraries and zlib1.dll.
          • 🔧 2026-10-20 check-in a4ac7741a1 on core-8-6-branch, in main, 9.0, 8.6: Merge buildall-static-dlls-8-6: fix 2b874804b5, the static pass of buildall.vc.bat also builds "dlls".
          • 💬 2026-10-20 serhiy.storchaka: Still fails with 8.6 (MSVC 2022), now first on zlib1.dll: the static pass of buildall.vc.bat builds only shell, but install also copies the static dde/registry libraries and zlib1.dll, which only the dlls target builds. 9.0 and later are fine, since static builds there always include the static packages. Proposed fix for 8.6 in branch 0e4b74d5ed: build shell dlls in the static pass. Tested with nmake: install su…
          • 💬 2026-10-20 oehhar: Serhiy, just merge this fix. Thanks, Harald
          Full verdict

          REPRODUCED (8.6 only); NOT-REPRODUCED on main (9.0 has the same makefile logic)

          • Platform build (Windows, nmake win/buildall.vc.bat)
          • Tested code reading main + 8.6 win/buildall.vc.bat, makefile.vc
          • Cause buildall.vc.bat's second pass runs nmake -f makefile.vc shell OPTS=static %1 (8.6: static,msvcrt → suffix sx). With %1=install, install-binaries copies $(TCLDDELIB) (tcldde14sx.lib in 8.6, tcl9dde15s.lib on main), but target shell (setup core $(TCLSH)) never builds it; only dlls/release do. (oehhar's VC6 run passed, probably because the static lib was left over from an earlier build.)
          • Prototype prototypes/2b874804b5.patch (main, 5 changed lines, CRLF kept): the static pass builds shell dlls; the same change applies to 8.6. Windows repro win/2b874804b5.cmd (static shell then install into %TEMP%).
          • Suggested ticket comment The static pass in buildall.vc.bat builds only shell, which does not make the static dde library that install copies. Proposed fix: build shell dlls there.
          Windows VM, run 1
          • Platform build (Windows, buildall.vc.bat)
          • Verdict REPRODUCED (8.6 only); NOT-REPRODUCED on main (9.0 has the same makefile logic)
          • Tested VM, MSVC 2022. main (tcl-wt91): nmake shell OPTS=static then nmake install OPTS=static succeeds: in 9.x static builds always link the static packages, so shell builds tcl9dde15s.lib. 8.6 (tcl-wt86): nmake shell OPTS=static,msvcrt then install fails, now even earlier at File not found - zlib1.dll; with shell dlls the install succeeds (installs tcldde14sx.lib, tclreg13sx.lib, zlib1.dll)
          • Cause 8.6 win/buildall.vc.bat static pass builds only shell; without staticpkg, install copies the static dde/registry libs and zlib1.dll that only dlls builds
          • Prototype prototypes/2b874804b5-86.patch (core-8-6-branch, 5 changed lines, CRLF kept): static pass builds shell dlls; tested by running the two nmake steps. The main patch prototypes/2b874804b5.patch is not needed (main already works).
          • Suggested ticket comment Still fails with 8.6 (now on zlib1.dll first); 9.0 and later are fine because static builds always include the static packages. Proposed fix for 8.6: build shell dlls in the static pass of buildall.vc.bat.
          ⬜537f9f98f42014windows handling "entry point not found" exceptionREPRODUCED (main, 9.0, 8.6)

          Open ticket 537f9f98f4 · 27. Channel Types · filed 2014-06-23

          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform Windows
          • Tested code reading only (Windows-specific).
          • Cause win/tclWinPipe.c:Tcl_WaitPid maps a fixed list of EXCEPTION_* codes to 0xC0000000|SIG*, but the default case stores any other exit code as is; win/tclWinPort.h WIFSIGNALED tests & 0xC0000000 and WTERMSIG takes & 0x7F, so any other NTSTATUS error (0xC0000139 → "signal" 57, "unknown signal"; 0xC0000409 fast-fail → 9 = "kill signal") is reported as CHILDKILLED with a meaningless signal.
          • Prototype none (needs a decision on how to report unlisted NTSTATUS codes: e.g. as CHILDSTATUS with the full code, or new cases for the DLL-load statuses; the fix itself would be ~10 lines in Tcl_WaitPid's default case)
          • Suggested ticket comment Still in main: Tcl_WaitPid's default case passes NTSTATUS codes like 0xC0000139 through, and WIFSIGNALED/WTERMSIG in tclWinPort.h turn them into a bogus signal ("unknown signal"; 0xC0000409 even becomes "kill signal"). Windows test script prepared; fix needs a decision on how to report such codes.
          Windows VM, run 2
          • Platform Windows
          • Verdict REPRODUCED (main, 9.0, 8.6)
          • Tested VM win/537f9f98f4.tcl (child tclsh exits with the given code): 0xC0000139, 0xC0000135, 0xC0000142 and 0xC0000409 all give CHILDKILLED <pid> {unknown signal} {unknown signal} / "child killed: unknown signal" on 9.1b1, 9.0.5 and 8.6.18; 0xC0000005 → SIGSEGV, exit 3 → CHILDSTATUS 3 (both fine). (0xC0000409 shows "unknown signal" here, not "kill signal" as the code reading suggested.)
          • Cause win/tclWinPipe.c Tcl_WaitPid's default case passes unlisted NTSTATUS codes through; WIFSIGNALED/WTERMSIG in win/tclWinPort.h then take them as a signal
          • Prototype none (needs a decision on how to report unlisted NTSTATUS codes)
          • Suggested ticket comment Confirmed on 8.6.18, 9.0.5 and 9.1: a child exiting with STATUS_ENTRYPOINT_NOT_FOUND, DLL_NOT_FOUND, DLL_INIT_FAILED or a fast-fail (0xC0000409) is reported as "child killed: unknown signal". Should these become CHILDSTATUS with the full code, or get their own messages?
          ⬜d5ddbc7f492014Inconsistent terminology in error messages: should be vs. must be, unknown/ambiguous vs. badREPRODUCED (main, 8.6) — cosmetic, RFE-like

          Open ticket d5ddbc7f49 · 45. Parsing and Eval · filed 2014-06-17

          Full verdict

          REPRODUCED (main, 8.6) — cosmetic, RFE-like

          • Platform generic
          • Tested main build and 8.6.18: wrong-# messages say "should be", value lists say "must be"; ensembles say unknown or ambiguous subcommand, Tcl_GetIndexFromObj says bad/ambiguous.
          • Cause tclEnsemble.c (line ~2024) builds "unknown[ or ambiguous] subcommand" by hand instead of using Tcl_GetIndexFromObj wording; "must be" comes from Tcl_GetIndexFromObj, "should be" from Tcl_WrongNumArgs.
          • Prototype none (message changes touch ~1400 lines in 20+ test files and break scripts matching error strings; an incompatible cosmetic change for a major release, if at all)
          • Suggested ticket comment Still as described in 9.1. Unifying the wording would change hundreds of expected results in the test suite and could break scripts that match error messages, so it only makes sense for a major release; suggest converting this to a Feature Request.
          ⬜35a839cba82014Can not reinitialize Tcl after Tcl_Finalize()patchREPRODUCED (main, 9.0, 8.6; 8.5.19 segfaults on the second round instead)

          Open ticket 35a839cba8 · 50. Embedding Support · filed 2014-04-14

          🧪 Prototype check (2026-10-24)

          35a839cba8.patch: VERIFIED

          • Base main (tip d6dd6d0f71)
          • Before embedding program r35a.c (3 rounds of Tcl_FindExecutable/CreateInterp/Init/Eval/DeleteInterp/Tcl_Finalize, linked against the tip libtcl9.1.so) -> round 0: rc=0 ... then Tcl_InitSubsystems called while exiting, SIGILL (panic), rc=132
          • Patch applied cleanly
          • After rounds 0, 1, 2 all rc=0 42 9.1.1 1970 a b ..., REINIT OK, rc=0; tests event (66/0 failed), init, main, thread: 0 failed
          • Test none possible (needs an embedding program that calls Tcl_Finalize and starts again; no test command does that)
          Full verdict

          REPRODUCED (main, 9.0, 8.6; 8.5.19 segfaults on the second round instead)

          • Platform generic
          • Tested C demo work-a2/35a.c (3 rounds of Tcl_FindExecutable/CreateInterp/Init/eval/DeleteInterp/Tcl_Finalize) linked against main, 9.0.5, 8.6.18: panic "Tcl_InitSubsystems called while exiting" in round 1; 8.5.19: SIGSEGV in round 1 (so it did not work in 8.5 either, as dgp found). With the prototype all 3 rounds succeed on main, valgrind-clean on a -DPURIFY build.
          • Cause tclEvent.c InvokeExitHandlers sets the static inExit and nothing resets it, so Tcl_InitSubsystems panics after any Tcl_Finalize.
          • Prototype prototypes/35a839cba8.patch (5 changed lines; ferrieux's one-liner inExit = false; at the end of Tcl_Finalize; no test, needs an embedding program; event/init/main/thread.test pass)
          • Suggested ticket comment Still panics in 9.1/9.0/8.6 (8.5 segfaults), since inExit is never reset. A demo program doing 3 init/finalize rounds works with inExit = false at the end of Tcl_Finalize and is valgrind-clean; proposed patch available.
          ⬜4af4d24baa2014Timezone setting fails for non-DST TZs under WindowspatchREPRODUCED (main, mocked registry on Windows; 9.0/8.6 have the same code, their registry DLL is not built in place on the VM); with the real regist…

          Open ticket 4af4d24baa · 06. Time Measurement · filed 2014-03-05

          🧪 Prototype check (2026-10-24)

          4af4d24baa.patch: VERIFIED

          • Base main (tip d6dd6d0f71)
          • Before with a mocked registry (Bias 420, empty StandardStart/DaylightStart; the clock.test _setupRegistry mock), ::tcl::clock::GuessWindowsTimeZone -> :localtime
          • Patch applied cleanly (the clock.test hunk is offset by 5 lines; 24 changed lines)
          • After -> :America/Phoenix; clock.test (8757 passed/0 failed) and clock-ivm.test (8769/0) with TCL_LIBRARY=wt/library: 0 failed
          • Test existing clock-49.3 (from the patch) fails on the unpatched tip library (result :localtime) / passes after. Whether real Windows registries still store empty start values remains NEEDS-WINDOWS (win/4af4d24baa.tcl).
          Full verdict

          REPRODUCED (main, mocked registry on Windows; 9.0/8.6 have the same code, their registry DLL is not built in place on the VM); with the real registry of Windows 11 it does not trigger

          • Platform Windows
          • Tested work-a6/t8.tcl overrides registry/_hasRegistry: empty StandardStart/DaylightStart give :localtime, 16 zero bytes give :America/Phoenix. Windows script: win/4af4d24baa.tcl (prints the real registry values, then the mocked verdict).
          • Cause library/clock.tcl GuessWindowsTimeZone: binary scan $stdtzi @0s val on an empty value leaves val unset, the lappend errors, and the catch falls back to :localtime.
          • Prototype prototypes/4af4d24baa.patch (24 changed lines, 3 in clock.tcl; test clock-49.3): pad both start values with 16 zero bytes. The test fails without the fix; clock.test and clock-ivm.test pass.
          • Suggested ticket comment GuessWindowsTimeZone fails when StandardStart/DaylightStart are empty: binary scan leaves the variable unset and the catch falls back to :localtime. Proposed fix: treat empty values as zeros (pad to 16 bytes), test clock-49.3.
          Windows VM, run 3
          • Platform Windows
          • Verdict REPRODUCED (main, mocked registry on Windows; 9.0/8.6 have the same code, their registry DLL is not built in place on the VM); with the real registry of Windows 11 it does not trigger
          • Tested VM win/4af4d24baa.tcl with 9.1b1: the live TimeZoneInformation values are 16-byte SYSTEMTIMEs (zone FLE, real guess :Europe/Helsinki); with the registry mocked to empty StandardStart/DaylightStart the guess falls back to :localtime. In the Time Zones database all 101 of 144 no-DST zones (e.g. US Mountain Standard Time) carry 16 zero bytes, not empty values, in TZI, so empty values come from older Windows or third-party tools.
          • Cause library/clock.tcl GuessWindowsTimeZone: binary scan on an empty value leaves val unset and the catch falls back to :localtime
          • Prototype prototypes/4af4d24baa.patch (existing, 24 changed lines, test clock-49.3) tested on Windows: the mocked repro gives :America/Phoenix, clock-49.1–49.3 pass, the full clock.test passes (8759/8759) with tcltest91 and the patched library. Tested on Windows.
          • Suggested ticket comment Still in the code: GuessWindowsTimeZone falls back to :localtime when StandardStart/DaylightStart are empty (current Windows 11 writes 16 zero bytes, so it only shows with such registries). Proposed fix: treat empty values as zeros, test clock-49.3 (tested on Windows and Linux).
          ⬜d18d06b5fc2013Wrong lang messages for "uk" locale in clock packagepatchREPRODUCED (main, 9.0, 8.6) for DATE_FORMAT/DATE_TIME_FORMAT only: uk.msg uses "%e/%m/%Y" (%x -> "16/10/2026"); Ukrainian uses "16.10.2026" (as r…

          Open ticket d18d06b5fc · 06. Time Measurement · filed 2013-12-16

          🧪 Prototype check (2026-10-24)

          d18d06b5fc.patch: VERIFIED+TEST-ADDED (only the date formats; the lower-case day/month names the ticket also asks to change are correct Ukrainian, as the triage says)

          • Base main (tip d6dd6d0f71)
          • Before clock format 1760572800 -gmt 1 -locale uk -format "%x | %c" -> 16/10/2025 | 16/10/2025 0:00:00 +0000
          • Patch applied cleanly (library/msgs/uk.msg, 4 changed lines) + test added
          • After 16.10.2025 | 16.10.2025 0:00:00 +0000; clock scan 16.10.2025 -locale uk -format %x round-trips; clock.test (TCL_LIBRARY=<wt>/library) 8764 total, 0 failed
          • Test added clock-72.1 (uk %x, %c and scan %x), fails with the unpatched uk.msg / passes after
          Full verdict

          REPRODUCED (main, 9.0, 8.6) for DATE_FORMAT/DATE_TIME_FORMAT only: uk.msg uses "%e/%m/%Y" (%x -> "16/10/2026"); Ukrainian uses "16.10.2026" (as ru.msg already does). NOT-A-BUG for the rest: Ukrainian writes day and month names in lower case, and the genitive month names ("жовтня") are right for "%d %B %Y" (Tcl has no separate nominative form, so "%B %Y" reads "жовтня 2026").

          • Platform generic
          • Tested main c3bf6dd052 (work-a6/t7.tcl); 9.0 file identical, 8.6 the same values written as \u escapes
          • Cause library/msgs/uk.msg generated from an old ICU.
          • Prototype prototypes/d18d06b5fc.patch (4 changed lines): DATE_FORMAT "%d.%m.%Y", DATE_TIME_FORMAT "%d.%m.%Y %k:%M:%S %z"; no test (no test uses uk); clock.test unaffected.
          • Suggested ticket comment The date formats are wrong: Ukrainian uses dd.mm.yyyy. Proposed fix: "%d.%m.%Y" in DATE_FORMAT and DATE_TIME_FORMAT. Lower-case day and month names are correct Ukrainian, and the genitive month names fit "%d %B %Y", so those stay.
          ⬜c31f72813b2013Tcl crashes when calculating the string representation of deep nested listsREPRODUCED (main, 9.0, 8.6) — SIGSEGV (stack overflow) at 100000 levels with the default 8 MB stack; 30000 levels still work

          Open ticket c31f72813b · 14. List Object · filed 2013-11-08

          Full verdict

          REPRODUCED (main, 9.0, 8.6) — SIGSEGV (stack overflow) at 100000 levels with the default 8 MB stack; 30000 levels still work

          • Platform generic
          • Tested work-a1/c31f.tcl with N=10000/30000/100000 on main build, 9.0.5, 8.6.18
          • Cause generic/tclListObj.c:UpdateStringOfList calls TclGetStringFromObj on each element, which recurses into UpdateStringOfList for nested lists: about 3 C frames per nesting level, with no depth limit. Freeing such lists was made iterative earlier, but string generation cannot easily be.
          • Prototype none (an iterative string generator or a depth/stack guard is a redesign well over 60 lines; Tcl has no stack-space check left, since TclpCheckStackSpace was removed)
          • Suggested ticket comment Still crashes on 8.6, 9.0 and main at 100000 levels: UpdateStringOfList recurses through Tcl_GetString once per level. A fix needs an explicit-stack string generator or a nesting limit; small patches won't do it.
          ✅d80f8cf33b2013tcl/macosx Makefile has fragile make distcleanpatchREPRODUCED (main; the same code is in 9.0 and 8.6)

          Open ticket d80f8cf33b · 53. Configuration and Build Tools · filed 2013-08-03

          ✅ Since triage (updated 2026-11-11)

          Status now: Closed (Fixed)

          • 🔧 2026-10-20 check-in dd3e012b6a on macosx-distclean-partial, in main: Fix d80f8cf33b: "make distclean" in macosx/ stopped with an error unless both the Development and the Deployment build had been made, since the clean rules al
          • 🔧 2026-10-20 check-in 3d871b1598 on trunk, in main, 9.0: Fix d80f8cf33b: tcl/macosx Makefile has fragile
          • 🔧 2026-10-20 check-in e6f4b1445d on core-9-0-branch, in main, 9.0, 8.6: Fix d80f8cf33b: tcl/macosx Makefile has fragile make distclean
          • 🔧 2026-10-20 check-in ec8a966c79 on core-8-6-branch, in main, 9.0, 8.6: Fix d80f8cf33b: tcl/macosx Makefile has fragile make distclean
          • 💬 2026-10-20 serhiy.storchaka: Still present on main (and the same code is in 9.0 and 8.6): clean-tcl and distclean-tcl in macosx/GNUmakefile always run make -C "${OBJ_DIR}", and distclean does that for both the Development and the Deployment build. So it stops with "No such file or directory" as soon as one of them was never configured. Proposed fix in branch macosx-distclean-partial: the clean rules skip the sub-make when the build directory …
          • 💬 2026-10-20 jan.nijtmans: Merged to 3d871b1598, core-9-0-branch and core-8-6-branch Many Thanks!
          Full verdict

          REPRODUCED (main; the same code is in 9.0 and 8.6)

          • Platform build (macosx/GNUmakefile; the logic is generic GNU make, so I tested it on Linux)
          • Tested main make -f macosx/GNUmakefile distclean BUILD_DIR=... with nothing built → make -C .../tcl/Development: No such file or directory, stops at distclean-develop. The same thing happens for the Deployment style after only make develop / make test-tcl.
          • Cause macosx/GNUmakefile clean-tcl/distclean-tcl always run ${MAKE} -C "${OBJ_DIR}" clean|distclean, and that fails when the build style's OBJ_DIR was never configured. distclean iterates over both styles (develop, deploy), so a partial build always breaks it.
          • Prototype prototypes/d80f8cf33b.patch (7 changed lines, no test: build target). Adds DO_MAKE_CLEAN = test ! -f "${OBJ_DIR}/Makefile" || ${MAKE} -C ... for the clean/distclean rules. I verified on Linux that distclean with nothing built, and after a develop-only build, now exits 0. Not tried on macOS.
          • Suggested ticket comment clean-tcl/distclean-tcl in macosx/GNUmakefile run make -C $OBJ_DIR even for a build style that was never configured. Proposed fix: skip the sub-make when $OBJ_DIR/Makefile does not exist.
          🔧6dacb3b6f02013exec and open | fail for .bat files if 8dot3name is disabled and path + command contain spacesREPRODUCED (main 9.1.1, 9.0.5, 8.6 tip, installed 8.6.18; only without a console, i.e. wish). Same bug as f91ab723f3

          Open ticket 6dacb3b6f0 · 52. Portability Support · filed 2013-07-22

          🔧 Since triage (updated 2026-11-11)

          Status now: Open

          • 🔧 2026-10-25 check-in e8d1e56815 on bug-f91ab723f3-quoting-main, not merged: Rebase of bug-f91ab723f3-quoting to main. Fix 6dacb3b6f0 and f91ab723f3: put the command after "cmd.exe /c" in double quotes, so cmd.exe does not strip the
          • 🔧 2026-10-14 check-in 30bc26a91e on bug-f91ab723f3-quoting, not merged: Fix 6dacb3b6f0 and f91ab723f3: append the closing quote for "cmd.exe /c" in BuildCommandLine(), before the conversion to WCHAR. Appending it to the converte
          • 💬 2026-10-25 serhiy.storchaka: The fix is rebased onto main as bug-f91ab723f3-quoting-main, see f91ab723f3.
          Full verdict

          REPRODUCED (main 9.1.1, 9.0.5, 8.6 tip, installed 8.6.18; only without a console, i.e. wish). Same bug as f91ab723f3

          • Platform Windows
          • Tested Linux not affected (the "cmd.exe /c" prefix exists only in win/tclWinPipe.c TclpCreateProcess, used for APPL_DOS i.e. .bat/.cmd when HasConsole() is false). Repro win/6dacb3b6f0.tcl (runs the cases in wish, prints VERDICT).
          • Cause win/tclWinPipe.c TclpCreateProcess: cmd.exe /c "dir with spaces\x.bat" "arg" has more than two quotes, so cmd.exe strips the first and last one. Still present unchanged in main, 9.0 and 8.6.
          • Existing work bug-f91ab723f3-quoting (oehhar 0530fbcf2d + Serhiy 30bc26a91e, 2026-10-14), unmerged, based on core-8-6-branch of 2024-02. Applies to current 8.6 (hunk 1 with fuzz); on main/9.0 hunk 2 fails (BuildCommandLine now ends with Tcl_UtfToWCharDString) and hunk 1 needs fuzz after the dead-code removal: a trivial rebase is needed for 9.x. Not buildable/testable here (Windows only).
          • Prototype none (branch exists)
          • Suggested ticket comment none needed (Serhiy's 2026-10-14 comment already points to the branch); after a rebase onto 9.0/main, run win/6dacb3b6f0.tcl with wish on Windows before merging.
          Windows VM, run 6
          • Platform Windows
          • Verdict REPRODUCED (main 9.1.1, 9.0.5, 8.6 tip, installed 8.6.18; only without a console, i.e. wish). Same bug as f91ab723f3
          • Tested win/6dacb3b6f0.tcl run from tclsh, with the matching wish (no console, so "cmd.exe /c" path; wish windows withdrawn, in the ssh session, not on the user's desktop). In wish exec "...\bat dir with spaces\test.bat" "" and ... {arg with spaces} 8 fail with '%USERPROFILE%\tmp\triage6\bat' is not recognized as an internal or external command; noSpaces 8 works; all 3 cases work in tclsh (console). Note: 8dot3 name creation is ENABLED on C: (fsutil 8dot3name query C:: volume state 0, registry state 2), so 8dot3name being disabled is not needed to reproduce; the disabled case itself was not tested (no other volume, and the setting was not to be changed).
          • Branch bug-f91ab723f3-quoting (30bc26a91e, 8.6.13-based) built in a new checkout: all 3 cases pass in wish (with tk86 wish86t.exe loading the branch's tcl86t.dll), VERDICT FIXED.
          • Rebased onto main (c092dae4af): prototypes/f91ab723f3.patch (14 changed lines, the same two hunks; also applies to 9.0 with offsets), built on the VM: both repros pass in wish91; winPipe.test + exec.test on the patched build: Total 255 Passed 220 Skipped 35 Failed 0 (run from ssh, i.e. with a console, so the cmd.exe path itself is not covered by the test suite).
          • Cause win/tclWinPipe.c TclpCreateProcess prefixes cmd.exe /c for APPL_DOS without a console; with more than two quotes cmd.exe strips the first and last quote.
          • Prototype prototypes/f91ab723f3.patch (main rebase of the branch; no automated test: the bug needs a process without a console, e.g. a test that runs a .bat via a detached tclsh/wish)
          • Suggested ticket comment Reproduced on Windows 11 with current main, 9.0 and 8.6 (wish only), also with 8dot3 names enabled. Branch bug-f91ab723f3-quoting fixes it on 8.6; the same change rebased onto main fixes it there too.
          ⬜b198d6560a20133-argument switch doesn't get bytecompiledpatchREPRODUCED (main, 9.0, 8.6)

          Open ticket b198d6560a · 47. Bytecode Compiler · filed 2013-07-18

          🧪 Prototype check (2026-10-24)

          b198d6560a.patch: VERIFIED+TEST-ADDED (with the -nocase -integer fix added)

          • Base main (tip d6dd6d0f71)
          • Before switch-20.2 FAILED (switch -exact $a {...} disassembly has no jumpTable: not compiled)
          • Patch applied cleanly; extended. The original patch turned the existing compiled/runtime mismatch into a new regression: switch -nocase -integer $a {...} raised "-nocase option cannot be used with -integer option" before the patch and was silently accepted once compiled. (-nocase -integer -- $a was already silently accepted on the tip, while set c switch; $c -nocase -integer -- ... raises the error.) Added a punt to the interpreted version when noCase && mode == Switch_Integer, plus test switch-20.4. Now 39 changed lines
          • After compiled and runtime both raise the error for -nocase -integer with and without --; switch.test 148/0 failed, compile 160/0, execute 157/0; the original patch also passed info 285/0, trace 286/0, basic 118/0
          • Test existing switch-20.2 fails before / passes after; added switch-20.4, fails before (... 0 x for the -- form) / passes after
          Full verdict

          REPRODUCED (main, 9.0, 8.6)

          • Platform generic
          • Tested disassemble on main build, 9.0.5, 8.6.18: switch $a {...} and switch -exact -- $a {...} compile to a jump table; switch -exact $a {...}, switch -glob $a {...}, switch -nocase $a {...} (options without --) are invoked at runtime.
          • Cause tclCompCmdsSZ.c TclCompileSwitchCmd: the option loop runs while numWords>=3; when the options are followed by just the value and body list it exits with numWords==2 and the following if (numWords < 3) return TCL_ERROR (meant for the "--" break case) punts. TclNRSwitchObjCmd never parses the last two words as options, so compiling this form is safe.
          • Prototype prototypes/b198d6560a.patch (28 changed lines incl. tests switch-20.1..20.3; switch, compile, execute, info, trace, basic.test pass; switch-20.2 fails on 9.0 as expected). Side note: compiled switch -nocase -integer -- ... already accepts what the interpreted command rejects ("-nocase option cannot be used with -integer"); the patch extends that to the no--- form.
          • Suggested ticket comment Still so in 9.1/9.0/8.6: the option loop in TclCompileSwitchCmd only succeeds when it stops at --. With options followed by exactly value + body list the interpreted command also never reads options, so it can be compiled; proposed patch with tests switch-20.*.
          ⏳3612422fff2013tclvars(n) man page went missingpatchREPRODUCED (main, 9.0, 8.6) as designed: no tclvars.n is installed. The Pending "Fixed" resolution (references removed) holds only partly.

          Open ticket 3612422fff · 53. Configuration and Build Tools · filed 2013-05-01 · Pending (resolved, waiting for confirmation)

          ⏳ Since triage (updated 2026-11-11)

          Status now: Pending (Fixed)

            Full verdict

            REPRODUCED (main, 9.0, 8.6) as designed: no tclvars.n is installed. The Pending "Fixed" resolution (references removed) holds only partly.

            • Platform docs
            • Tested main unix/installManPage doc/tclvars.n dir installs argc.n … tcl_version.n (16 names), no tclvars.n. Grepped doc/ for references on main, 9.0 and 8.6.
            • Cause the NAME line deliberately lists the variables, not "tclvars" (dkf 2013). dkf's 2013 reference cleanup is done, but new references came back: main doc/filename.n:151 ("documented in the manpages for \fBtclvars\fR and \fBtm\fR") and doc/library.n:169 ("see the tclvars manual page"). 9.0 has only the filename.n one, 8.6 has neither.
            • Prototype prototypes/3612422fff.patch (4 changed lines, docs only). Points both references to auto_path instead.
            • Suggested ticket comment Installing under the NAME names is by design. Two newer references to "tclvars" (filename.n in 9.0+, library.n in 9.1) should point to auto_path instead. Once that is done the ticket can be closed.
            ⬜3610400fff2013tailcall clashes with execution tracesREPRODUCED (main, 9.0, 8.6)

            Open ticket 3610400fff · 60. NRE and coroutines · filed 2013-04-09

            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested reporter's script on main build, 9.0.5, 8.6.18: the enterstep trace for the tailcalled target runs at level 1 in namespace ::, where namespace which target gives ''.
            • Cause tclBasic.c TclNRTailcallObjCmd stores the proc's namespace with the words and TclNRTailcallEval resolves the command through iPtr->lookupNsPtr after the proc frame is popped; the traced command string keeps the unqualified word, and no frame with that namespace exists when the trace fires (same class of issue as ensembles/oo forwards, [3610393]).
            • Prototype none (would need tailcall to pass a qualified/resolved command name to traces, or a frame for the lookup namespace; a design question, dkf: "no idea what to do with this")
            • Suggested ticket comment Unchanged in 9.1/9.0/8.6: the tailcalled command is resolved in the proc's namespace via lookupNsPtr after its frame is gone, so traces see an unqualified name in the caller's context. A fix would be for &#91;tailcall&#93; to hand the trace a fully qualified command word; needs a decision on what traces should see.
            ✅3602289fff2013EnvTraceProc errors when env var doesn't existpatchREPRODUCED (main, 9.0, 8.6, 8.5) as a stale element: when another interp unsets an env var, a read in this interp fails with "no such variable" but…

            Open ticket 3602289fff · 08. Environment Variables · filed 2013-01-27

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Duplicate)

            • 🔧 2026-10-19 check-in f1438aeb70 on env-stale-element, in main: Fix 3602289fff, 67fd4f973a and 845241ffff: after another interp unset an env variable, a read failed but left the old element, so [info exists ::env(X)] s
            • 💬 2026-10-19 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1 in a narrower form: after another interp unsets an env variable, reading it here fails, but the stale element stays, so [info exists ::env(X)] still returns 1. The lappend case no longer errors. Proposed fix in branch env-stale-element, also for 67fd4f973a and 845241ffff: the read trace removes the stale element and keeps the error. Unlike the attached patch, [set env(missing)…
            • 💬 2026-10-20 jan.nijtmans: Dup of 845241ffff
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5) as a stale element: when another interp unsets an env var, a read in this interp fails with "no such variable" but leaves the element in place, so [info exists ::env(X)] still returns 1. The original lappend/append symptom (bug 3057639) does not reproduce: [lappend env(NEW) a] / [append] work both compiled and uncompiled since 8.5.

            • Platform generic
            • Tested main (own build), 9.0.5, 8.6.18, 8.5.19 (work-a3/360228.tcl, 360228b.tcl)
            • Cause generic/tclEnv.c:EnvTraceProc, the read branch: when TclGetEnv returns NULL it returns an error but never removes the stale array element. The attached patch (unset plus a silent NULL return) would also change [set env(missing)] semantics and panics on failure. An unconditional unset breaks env-7.3 (bug [9b4702], traces on not-yet-existing elements).
            • Prototype prototypes/3602289fff.patch (27 changed lines, test env-bug-3602289). It unsets the element only when it is defined and keeps the error. env/var/trace/info/upvar tests pass.
            • Suggested ticket comment Still present in a narrower form: after another interp unsets an env var, a read in this interp errors but leaves a stale element, so &#91;info exists&#93; returns 1. Proposed fix: drop a defined stale element in EnvTraceProc and keep the error (an unconditional unset breaks env-7.3, bug [9b4702]). Patch with test env-bug-3602289 is available. The lappend case from 3057639 no longer errors.
            ⬜3601284fff2013[testinterpresolver] breaks epoch rulesREPRODUCED (main, 9.0, 8.6)

            Open ticket 3601284fff · 21. [namespace] · filed 2013-01-17

            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic (test command only)
            • Tested tcltest from my main build, the 9.0 tip and the 8.6 tip, with dgp's extended resolver-1.1 scenario (work-a3/360128.tcl). It prints r0=Y r1=Y r2=Z r3=Z, where a context-correct resolver would give Y, "unknown command", Z, Y.
            • Cause generic/tclTest.c:InterpCmdResolver decides by context (the calling proc's name and namespace, the caller's namespace) that the "cmdName" intrep validation (namespace + cmdEpoch/refEpoch) never sees, so cached resolutions leak into other contexts. resolver-1.x only pass because TclInvalidateCmdLiteral clears literals on [namespace import] etc. Same root issue as 3418547 (dgp's 2016 branch bug-3418547 was never merged).
            • Prototype none (needs a decision on what custom resolvers may depend on, plus a redesign of the test resolver and resolver.test)
            • Suggested ticket comment Still true on main/9.0/8.6: InterpCmdResolver depends on the calling proc's name, which the cmdName epoch check cannot see. Extending resolver-1.1 with a final &#91;x&#93; returns Z instead of Y. Tied to [3418547]; fixing it needs a rule for what resolvers may depend on.
            ⬜3600058fff2013Doctools nroff/groff output not supported by tcltk-man2htmlREPRODUCED (main; by code reading): branch bug-3600058-td (last check-in cdb671dd50, 2014-04-28) was never merged.

            Open ticket 3600058fff · 55. Other Tools · filed 2013-01-09

            Full verdict

            REPRODUCED (main; by code reading): branch bug-3600058-td (last check-in cdb671dd50, 2014-04-28) was never merged.

            • Platform docs (tools/tcltk-man2html)
            • Tested read the main tools/tcltk-man2html-utils.tcl. output-name still needs ^([^-]+) - (.*)$ (issue #1: an empty description makes $head undefined), and the .SH/.SS handler still does [llength $rest] on a string (issue #5). No .SH CATEGORY support.
            • Cause the branch is a large rework (font-state rewrite, categories, --pkgdir, .SH CATEGORY added to every doc file). It is far over the size limit and conflicts with today's trunk.
            • Prototype none (a redesign; the small parts, #1 regex and #5 string trim $rest eq "", could be cherry-picked if wanted)
            • Suggested ticket comment Branch bug-3600058-td was never merged and no longer applies to trunk. Small robustness fixes (#1 empty NAME description, #5 llength on a non-list) could be split out. The rest is an RFE, and doctools output has improved since then.
            ⬜3592042fff2012pkg_mkIndex fails on package presentREPRODUCED (main, 8.6), but resolution is already "Wont Fix" (dgp: stubbing package present would break indexing scripts that use `package vsatis…

            Open ticket 3592042fff · 39. Package Manager · filed 2012-12-03

            Full verdict

            REPRODUCED (main, 8.6), but resolution is already "Wont Fix" (dgp: stubbing package present would break indexing scripts that use package vsatisfies [package present ...] with -load)

            • Platform generic
            • Tested reporter's foo/bar/test example on main (own build) and 8.6.18: "warning: error while sourcing init.tcl: package bar is not present"
            • Cause library/package.tcl: pkg_mkIndex stubs package require (returns "") in its child interp but leaves package present alone, so the asserted package really is absent.
            • Prototype none (Won't Fix; workaround: use package require instead of package present, or create pkgIndex.tcl by hand)
            • Suggested ticket comment Unchanged in 9.x. As discussed, stubbing package present would break other indexing uses; use package require in such scripts. Closing as Won't Fix.
            ⬜3587928fff2012info.test cannot be added toREPRODUCED (main, 9.0, 8.6) — still true by design of the tests

            Open ticket 3587928fff · None · filed 2012-11-16

            Full verdict

            REPRODUCED (main, 9.0, 8.6) — still true by design of the tests

            • Platform generic (test suite)
            • Tested read tests/info.test on main (2648 lines, 144 type source line N expectations; comments at lines 1014/1958 say to add tests only at the end)
            • Cause the info frame tests compare absolute line numbers of the test file itself.
            • Prototype none (would mean moving ~1000 lines of info frame tests into their own file, or rewriting expectations relative to a marker line; larger than a quick fix)
            • Suggested ticket comment Still the case on main; the current workaround is to add new tests only at the end of info.test. Splitting the &#91;info frame&#93; tests into their own file (e.g. infoFrame.test) would solve it.
            🔧3577801fff2012Issue with [file normalize] on Linux with symlink wtih ".."patchREPRODUCED (main, 9.0, 8.6, 8.5)

            Open ticket 3577801fff · 37. File System · filed 2012-10-17

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-19 check-in 854be12c0a on normalize-dotdot-after-link, not merged: Fix 3577801fff, 3570244fff, 3183016fff and 3099612fff: on Unix, [file normalize] and [cd ..] used the target of a link literally when ".." followed it,
            • 💬 2026-10-19 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1: when ".." follows a link, the link target is used literally, so ".." or "." in the target, a trailing "/" or a chain of links give a wrong result (d/l/.. with l -> ../x gives d/..). Proposed fix in branch normalize-dotdot-after-link, together with 3570244fff, 3183016fff and 3099612fff: on Unix the link target is normalized and further links are followed before backing up, s…
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5)

            • Platform generic (unix symlinks)
            • Tested main (own build), 9.0.5, 8.6.18, 8.5.19, 8.4.20 on Linux. With d/lx -> ../x: [file normalize d/lx/..] gives "<t>/d/.." (".." left in the result), d/lx/../.. gives "<t>/d" (should be the parent of <t>); with a trailing slash in the target (d/lx2 -> ../x/) d/lx2/.. gives "<t>/x" (should be <t>); the reporter's l -> ../f (a file) case gives t/d/f. Link chains (ll -> l -> ../x) back up to the wrong directory too.
            • Cause generic/tclPathObj.c:TclFSNormalizeAbsolutePath, ".." handling: the link target is pasted in literally (relative ones after the link's directory) and then the last element is cut off textually, so ".." / "." components or a trailing "/" in the target, or a target that is itself a link, give a wrong parent.
            • Prototype prototypes/3577801fff.patch (57 changed lines: 24 in tclPathObj.c, 33 tests filesystem-1.55, filesystem-1.56). Normalizes the substituted target with TclFSNormalizeToUniquePath, follows further links (max 32, so self-links terminate), strips trailing "/", then cuts the last element. All results now match realpath(1) / Python os.path.realpath; full test suite (minus http*) passes.
            • Suggested ticket comment Still present in 9.1/9.0/8.6: when ".." follows a symlink, the link target is used literally, so ".." or a trailing "/" in the target or link chains give a wrong result. Proposed fix: normalize the link target before backing up over it (patch with tests filesystem-1.55/1.56 available, not committed).
            🔧3570244fff2012"cd .." behaves wrong if last element in pwd is a symlinkREPRODUCED (main, 9.0, 8.6, 8.5); 8.4.20 instead resolves the link at [cd] time ([pwd] shows the target)

            Open ticket 3570244fff · 37. File System · filed 2012-09-20

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-19 check-in 854be12c0a on normalize-dotdot-after-link, not merged: Fix 3577801fff, 3570244fff, 3183016fff and 3099612fff: on Unix, [file normalize] and [cd ..] used the target of a link literally when ".." followed it,
            • 💬 2026-10-19 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1. As dkf suspected, the cause is in normalization: when ".." follows a link, the link target is used literally (3577801fff). Proposed fix in branch normalize-dotdot-after-link: on Unix [cd ..] now goes to the parent of the link target in all five cases of the report, as cd -P .. does. Test filesystem-1.56.
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5); 8.4.20 instead resolves the link at [cd] time ([pwd] shows the target)

            • Platform generic (unix symlinks)
            • Tested the five cases of the report on main, 9.0.5, 8.6.18, 8.5.19, 8.4.20: link_abs ok; link_abs/ (trailing slash) and ../dir/ stay in the target dir; ../dir gives "B/test/.."; ./link_abs gives "B/test/."
            • Cause same as 3577801fff: [cd ..] normalizes "<pwd>/.." through TclFSNormalizeAbsolutePath, which uses the link target literally.
            • Prototype prototypes/3577801fff.patch (shared; test filesystem-1.56). With it all five cases give B (the parent of the physical target, as POSIX cd -P ..).
            • Suggested ticket comment Same cause as 3577801fff: the link target is used literally when ".." follows a link. The proposed fix there also fixes all cases here (test filesystem-1.56).
            ⬜3560322fff2012[exec <<] cannot set binary translationREPRODUCED (main, 9.0, 8.6)

            Open ticket 3560322fff · 27. Channel Types · filed 2012-08-21

            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested work-a4/t356.tcl: exec od -An -tx1 << "\xff\x80\x00\r\n" gives c3 bf c2 80 00 0d 0a (UTF-8-encoded) on main, 9.0.5 and 8.6.18 with a UTF-8 system encoding; only with LANG=C (iso8859-1) are the bytes intact. The newer exec -encoding option (9.0/main) only decodes the *output*; the << input is still system-encoded (and on Windows also newline-translated, as exec.n says).
            • Cause generic/tclPipe.c:TclCreatePipeline passes the << literal to TclpCreateTempFile, which converts it with the system encoding (and translates newlines on Windows); there is no option to choose encoding/translation for it, and Tcl_OpenCommandChannel has no parameter for it.
            • Prototype none (needs a design/TIP — e.g. make exec -encoding apply to << input too, accept -encoding binary, or the -binarystdout ideas from wiki 39989; plumbing goes through the public TclCreatePipeline/Tcl_OpenCommandChannel path)
            • Suggested ticket comment Still true on 9.0/main: the &#91;exec -encoding&#93; option only decodes output, << input is always converted with the system encoding. A fix needs a small design decision (e.g. let -encoding also apply to << input, with binary meaning raw bytes).
            🔧3556218fff2012string match & glob: close-bracket in character setpatchREPRODUCED (main, 8.6) (no way to put ] in a [...] set; []a] never matches anything, and backslash is not special inside a set)

            Open ticket 3556218fff · 18. Commands M-Z · filed 2012-08-11

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-19 check-in 7ca44862f9 on glob-bracket-posix, not merged: Fix 219233ffff and 3556218fff: in glob patterns ([string match], [glob], [lsearch -glob], [switch -glob]), a "-" right before the closing bracket of a set n
            • 💬 2026-10-19 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1: there is no way to include a closing bracket in a set. Backslash is not special inside a set, and {[]a]} never matches anything. The only workaround was a range ending in the closing bracket, which always includes other characters: the narrowest one, {[\-]]}, also matches a backslash. And such ranges are what 219233ffff removes. Proposed fix for 9.1 in branch glob-bracket-posix…
            Full verdict

            REPRODUCED (main, 8.6) (no way to put ] in a [...] set; []a] never matches anything, and backslash is not special inside a set)

            • Platform generic
            • Tested main (own build), 8.6.18: string match {[]a]} ]/a = 0, glob {[]a]} finds nothing. Workaround today: a range ending in ], e.g. [\-]].
            • Cause generic/tclUtil.c:Tcl_StringCaseMatch/TclByteArrayMatch and generic/tclUtf.c:Tcl_UniCharCaseMatch/TclUniCharMatch: the set loop returns no-match as soon as it sees ], even right after [.
            • Prototype prototypes/3556218fff.patch (39 changed lines: about 24 code lines in the 4 matchers, POSIX-style "a ] right after [ is literal", plus doc/string.n text and tests string-11.56/11.57; string, util, utf, lsearch, switch, fileName, parse and cmdMZ tests pass; the new tests fail without the fix). Compatibility: only patterns containing [], which today can never match, change meaning. It still is a semantic extension and may want a TIP or at least a changes.md note; glob.n could get the same sentence.
            • Suggested ticket comment Still the case: a ] can't appear in a glob bracket set, and []...] never matches. Proposed fix: treat a ] right after [ as a literal (as POSIX fnmatch and [scan] do) in all four matchers. This only affects patterns that currently can never match. Test string-11.56/11.57.
            ⬜3532972fff2012Tcl Scripts results in "called Tcl_FindHashEntry on deleted"REPRODUCED (main, 9.0, 8.6) only with TCL_FINALIZE_ON_EXIT=1; NOT-REPRODUCED with the default quick exit

            Open ticket 3532972fff · 80. Thread Package · filed 2012-06-07

            Full verdict

            REPRODUCED (main, 9.0, 8.6) only with TCL_FINALIZE_ON_EXIT=1; NOT-REPRODUCED with the default quick exit

            • Platform generic (Tcl finalization + Thread package)
            • Tested reporter's script with the delay cut to 300 ms (work-a6/t13.tcl), 40 runs per version and setting. Default: 0/40 failures on main, 9.0.5 and 8.6.18. TCL_FINALIZE_ON_EXIT=1: main 19/40, 9.0.5 37/40, 8.6.18 29/40, panicking with "called Tcl_FindHashEntry on deleted table" or "Tcl_Release couldn't find reference", or segfaulting.
            • Cause after thread::release, the worker is still tearing down its interp and its thread state (Tcl_FinalizeThread) while the main thread's exit runs a full Tcl_Finalize, which frees process-wide tables (preserve list, encodings, etc.) under it. Tcl_Finalize does not wait for other Tcl threads; this is the reason 8.6 made quick exit the default ([2001201]).
            • Prototype none (would need Tcl_Finalize to synchronize with or reap other Tcl threads, which is a design change; scripts can avoid it with -joinable + thread::join before exit: 0/40 failures on main and 8.6 with TCL_FINALIZE_ON_EXIT=1, work-a6/t13j.tcl)
            • Suggested ticket comment Still reproducible with TCL_FINALIZE_ON_EXIT=1 on 8.6/9.0/9.1: a full Tcl_Finalize races with a worker that was just released and is still finalizing. With the default quick exit it does not happen (0/40). Joining the worker before exit avoids it.
            ⬜3485833fff2012trace - inexplicable large overhead by ensemble commandsREPRODUCED (main, 9.0, 8.6, 8.5) for ensemble subcommands that are not bytecompiled. Compiled ones such as [string range] stopped showing it in 8.6…

            Open ticket 3485833fff · 46. Traces · filed 2012-02-08

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5) for ensemble subcommands that are not bytecompiled. Compiled ones such as [string range] stopped showing it in 8.6 because they compile inline.

            • Platform generic
            • Tested a tiny stubs extension calling Tcl_CreateObjTrace(..., TCL_ALLOW_INLINE_COMPILATION, proc that returns TCL_OK) (work-a3/tr/, work-a3/348583b.tcl), with a 50k-char $s. [string first A $s 2000]: untraced 0.2-0.4 us, traced 34 us (9.0.5), 40 us (8.6.18), 59 us (main debug build). [::tcl::string::first ...] traced stays at 0.2-0.5 us. [string range] on 8.5.11/8.5.19: 44-50 us traced vs 0.15 us. testcmdtrace resulttest no longer shows it for string range/wordend/totitle, because those compile or are O(N) anyway.
            • Cause generic/tclBasic.c:TEOV_RunEnterTraces (and the leave counterpart) calls TclGetStringFromObj on the command, which for an ensemble rewrite is a freshly built list, so the trace costs O(total argument length). The Tcl_CmdObjTraceProc API passes that command string, so avoiding it needs a lazy/optional string (an API change). Branch bug-3485833 (dgp/sebres kludge, 2012, merged with 8.5 in 2018) was never merged to a release branch. dgp-scan-element (only a TclScanElement speedup) was merged, which explains Resolution=Fixed.
            • Prototype none (needs an API/design change to the obj-trace interface, or reuse of the original source string as in bug-3485833)
            • Suggested ticket comment Still present on 8.6.18, 9.0 and main for non-compiled ensemble subcommands: with an obj-trace installed, &#91;string first A $s 2000&#93; on a 50k string takes ~35-60 us instead of 0.2 us. Building the list string rep for the trace callback's command argument is O(N). Branch bug-3485833 was never merged; the resolution "Fixed" refers only to the TclScanElement speedup.
            ⬜3463315fff2011file normalize breaks C:/Windows/system32/drivers/etc/hostspatchREPRODUCED (8.6); FIXED in 9.1.0 and core-9-0-branch (unreleased 9.0.5) by 0ac314cb96 (2026-07-16, fix for d40d8db3fb)

            Open ticket 3463315fff · 37. File System · filed 2011-12-21

            🧪 Prototype check (2026-10-24)

            3463315fff.patch: VERIFIED+TEST-ADDED (the original WOW64 case was not rerun: it needs a 32-bit build; it was verified earlier in result-win-4.md)

            • Base core-8-6-branch (tip 6b2fff63e6)
            • Before no 32-bit build is possible without a new build tree, so I used the same failure branch reached through an ACL (as in d40d8db3fb): the n3.tcl script, run with runas /trustlevel:0x20000 (a basic-user token: no SeBackupPrivilege; the ssh session has backup and restore privileges enabled, so it can list any directory and does not reproduce), creates n3/middledir/tail and denies the user (RD) on middledir with icacls. file normalize .../n3/middledir/tail -> %USERPROFILE%/tmp/verifyW1/n3/middledir/ (the component is dropped) on 8.6.18. 9.0.5 and 9.1.1 give .../middledir/tail (fixed there by 0ac314cb96). The x64 file normalize C:/WINDOWS/System32/drivers/etc/hosts is fine on every version (no WOW64)
            • Patch applied cleanly; 16 changed lines in win/tclWinFile.c (not changed), plus the test (36 changed lines in all)
            • After same command -> .../n3/middledir/tail; tests run as basic user: winFCmd.test 179 total, 165 passed, 14 skipped, 0 failed (with -constraints allowfilesecuritychanges). As the ssh admin: 3 failed (winFCmd-6.1, 6.13, 9.3, as unpatched; these are caused by the backup privilege and pass as basic user). fileName/fileSystem/cmdAH/winFile/fCmd: basic user 11 failed, admin 8 failed. These are identical lists with and without the patch: cmdAH-16.2, 17.2, 17.3 and winFile-4.0 to 4.4 for both; fCmd-28.12, 28.16 and filename-11.17.2 only for the basic user, who has no symlink privilege
            • Test added winFCmd-16.15 (8.6 counterpart of 9.0's winFCmd-16.25.1, but using exec icacls instead of the 9.0-only testfilesddl command; constraints win + allowfilesecuritychanges, which is set in CI or with -constraints). As basic user: FAILED unpatched (.../middledir/ instead of .../middledir/tail), PASSED patched. NB: as an elevated user with SeBackupPrivilege enabled it passes even unpatched

            Note on the VM session: from 21:37 to 00:10 a Tk verification session (~/tcltk/triage, tmp\v1024) was running tktest91 against tcl91.dll and using the 'triage' task. While it ran I did not touch tcl91 or the task: I only ran the unpatched tcl91 build for the "before" repros and did the 8.6/9.0 work. The tcl91 work below ran from 00:10 to 00:45, after the coordinator said the VM was free. Runs without privileges: the ssh session has SeBackupPrivilege, SeRestorePrivilege and SeCreateSymbolicLinkPrivilege enabled, which bypasses directory ACLs. "low" runs below use runas /trustlevel:0x20000 (a basic-user token with only SeChangeNotifyPrivilege). That is not a settings change, and the 'triage' task was not used at all.

            Baseline (unpatched tcl91, tcltest91.exe tests/all.tcl -file "winFCmd winFile fCmd cmdAH fileName fileSystem winPipe exec env winDde io ioCmd chanio load process pid"): admin 20451 total, 8 failed (winFCmd-6.1 6.13 9.3, winFile-4.0 to 4.4); low 6 failed (filename-11.17.2, winFile-4.0 to 4.4). These are the reference lists below.

            Full verdict

            REPRODUCED (8.6); FIXED in 9.1.0 and core-9-0-branch (unreleased 9.0.5) by 0ac314cb96 (2026-07-16, fix for d40d8db3fb)

            • Platform Windows
            • Tested code reading (main/9.0/8.6 win/tclWinFile.c); 8.6 prototype syntax-checked with mingw; repro script win/3463315fff.tcl (needs a 32-bit tclsh on 64-bit Windows)
            • Cause win/tclWinFile.c:TclpObjNormalizePath: GetFileAttributesExW sees drivers\etc (exempt from WOW64 redirection), but FindFirstFileW of the component fails because listing System32\drivers is redirected to SysWOW64\drivers; the failure branch appended only "/" ("This is usually the '/' in 'c:/'"), dropping the component and giving drivers//hosts. Same mechanism as d40d8db3fb (ACL-denied listing); the fix there copies the original component instead.
            • Prototype prototypes/3463315fff.patch (16 changed lines, 8.6 backport of the d40d8db3fb code fix against core-8-6-branch, no test: the 9.0 test needs the ACL test infrastructure in tclWinTest.c)
            • Suggested ticket comment Same cause as d40d8db3fb: FindFirstFileW() fails for the component (WOW64 redirects the listing of System32\drivers) and the component was dropped. Fixed in 9.1.0 and core-9-0-branch by 0ac314cb96; proposed 8.6 backport available.
            Windows VM, run 4
            • Platform Windows (32-bit tclsh on 64-bit Windows)
            • Verdict REPRODUCED (8.6); FIXED in 9.1.0 and core-9-0-branch (unreleased 9.0.5) by 0ac314cb96 (2026-07-16, fix for d40d8db3fb)
            • Tested VM win/3463315fff.tcl with 32-bit (MACHINE=IX86) tclsh built from tcl-wt86, tcl-wt90, tcl-wt91 (removed afterwards): 8.6.18 tip file normalize C:/WINDOWS/System32/drivers/etc/hosts → C:/Windows/System32/Drivers//hosts (exists 0); 9.0.5 and 9.1b1 → C:/Windows/System32/Drivers/etc/hosts (exists 1).
            • Cause win/tclWinFile.c TclpObjNormalizePath: FindFirstFileW of etc fails because WOW64 redirects the listing of System32\drivers, and the failure branch appended only "/"
            • Prototype prototypes/3463315fff.patch (existing, core-8-6-branch, 16 changed lines, no test) applied to tcl-wt86: 32-bit repro gives .../Drivers/etc/hosts (exists 1); x64 fileName.test, fileSystem.test 0 failed; winFCmd.test 3 failed, winFile.test 5 failed, cmdAH.test 3 failed (16.2, 17.2, 17.3), all identical to unpatched 8.6.18. Tested on Windows.
            • Suggested ticket comment Same cause as d40d8db3fb: in a 32-bit tclsh on 64-bit Windows FindFirstFileW() of "etc" fails (WOW64 redirects the listing of System32\drivers) and the component was dropped. Fixed in 9.1.0 and core-9-0-branch by 0ac314cb96 (verified with 32-bit builds on Windows 11); proposed 8.6 backport tested with a 32-bit 8.6 build.
            ⬜3447273fff2011tpool::* commands respect interp limitspatchREPRODUCED (main + Thread 3.0.7, 8.6.18 + Thread 2.8.13)

            Open ticket 3447273fff · 80. Thread Package · filed 2011-12-01

            🧪 Prototype check (2026-10-24)

            3447273fff.patch: VERIFIED

            • Base Thread trunk (tip 39d9f08aa0, Thread 3.0.7), built in work-a1/thread/buildv1 against the main tip wt-a1 build
            • Before new tests/tkt-3447273fff.test with threadPoolCmd.c unpatched: -1 (interp cancel -unwind after 200 ms while tpool::wait waits on a 2 s job) -> 1 {eval unwound} 0 1 (unwinds only after the job is done, not under 1 s); -2 (time limit 200 ms) -> 0 1 0 (no limit error at all)
            • Patch applied cleanly (118 changed lines with the new test file; saved file is the original, unchanged)
            • After both tests pass (3/3 runs); full Thread suite make test 129 passed/16 skipped/0 failed (145)
            • Test existing tkt-3447273fff-1/-2 fail before / pass after
            Full verdict

            REPRODUCED (main + Thread 3.0.7, 8.6.18 + Thread 2.8.13)

            • Platform generic (Thread package)
            • Tested child interp that waits with tpool::wait on a 2 s job, after 300 {interp cancel -unwind $c} from the parent. The cancel only takes effect after 2000 ms, when the job is done. thread::send + vwait, for comparison, unwinds at 300 ms. A time limit behaves the same way.
            • Cause thread generic/threadPoolCmd.c: TpoolPostObjCmd (2 sites) and TpoolWaitObjCmd loop on Tcl_DoOneEvent until a worker signals them and never check Tcl_Canceled/Tcl_LimitExceeded, unlike ThreadWait in threadCmd.c. (TpoolRelease has the same loop, but it has no interp and must finish tearing down the pool, so it is left unchanged.)
            • Prototype prototypes/3447273fff.patch (diff -N against the Thread repo trunk. C part: 73 changed lines, about half of them the helper and its header comment, a new WaitForSignal() helper that checks cancel/limits, unlinks the thread's waiter from the pool's waiter list on bail-out and sets "limit exceeded" with errorCode {TCL EVENT LIMIT} the way vwait does. New file tests/tkt-3447273fff.test, following the existing tkt-84be1b5a73.test, with cancel and time-limit cases.) Both new tests fail without the fix and pass with it; the full Thread suite passes (145 tests, 16 skipped).
            • Suggested ticket comment Still true in Thread 2.8.13/3.0.7: tpool::wait (and tpool::post waiting for a worker) ignores &#91;interp cancel&#93; and &#91;interp limit&#93; until the job finishes. Proposed fix: check Tcl_Canceled/Tcl_LimitExceeded after each Tcl_DoOneEvent, as thread::wait does, and take the waiter off the pool's list on bail-out; test tkt-3447273fff.
            ⬜3431943fff2011clock scan with %y is not handled correctlypatchREPRODUCED (main, 9.0, 8.6): clock scan 01221 -format %m%d%y gives 2001-01-22, though clock(n) says %y "accepts two digits" on input (%m and %d s…

            Open ticket 3431943fff · 06. Time Measurement · filed 2011-11-01

            🧪 Prototype check (2026-10-24)

            3431943fff.patch: VERIFIED+TEST-ADDED (doc + behaviour-pinning tests)

            • Base main (tip d6dd6d0f71). The doc hunk does not apply to core-8-6-branch as is (same text, different context lines); the test applies with fuzz.
            • Before clock.n %y: "On input, accepts two digits". Behaviour (work-a1/3431.tcl, -gmt 1): clock scan 01221 -format %m%d%y -> 2001-01-22 (main and 8.6); 1/5/7 %m/%d/%y -> 2007-01-05; { 7} and { 7} accepted (leading whitespace); 1/5/123/123 %y -> "input string does not match supplied format". (Side note, not part of this ticket: a lone %y without %m/%d sets the year on main but is ignored on 8.6, which keeps today's date.)
            • Patch 19 changed lines. doc/clock.n %y: "accepts one or two digits, possibly with leading whitespace", plus the 01221 example and the hint to check with regexp when exactly two digits are required (the wording now matches %m/%d/%e, which already say "one or two digits"). tests/clock.test: clock-72.1, 72.2.
            • After man page renders; troff warnings 1/1 (the pre-existing "cannot adjust line" on p. 12, shifted). clock.test on main: 8765 total, 8756 passed, 7 skipped, 2 failed: clock-71.1/71.2 (he locale) failed only because the wt's embedded libtcl zip was stale (Oct 16, before the he.msg fix); they pass with TCL_LIBRARY=<wt>/library and after rebuilding the zip, so unrelated.
            • Test added clock-72.1 (one/two digits at the end, in the middle, with leading space: 01221, 012207, 1/5/7, 1/5/26, " 7-01-05", 7-01-05, 70105) and clock-72.2 (three digits -> error). Pins current behaviour: pass on main; the same tests extracted into work-a1/p3431.test pass on the 8.6 tip (main's clock.test does not load in 8.6).
            Full verdict

            REPRODUCED (main, 9.0, 8.6): clock scan 01221 -format %m%d%y gives 2001-01-22, though clock(n) says %y "accepts two digits" on input (%m and %d say the same, and also take one digit).

            • Platform generic
            • Tested main c3bf6dd052, 9.0.5, 8.6.18 (work-a6/t9.tcl)
            • Cause generic/tclClockFmt.c (9.x) / library/clock.tcl (8.6): numeric fields are scanned as 1 to N digits, so the last field may be short.
            • Prototype none (requiring exactly two digits would break common inputs like 1/5/26 with %m/%d/%y; this is better fixed in the docs, "one or two digits")
            • Suggested ticket comment On input, %y, %m, %d etc. accept one or two digits, which inputs like 1/5/26 rely on, so making them strict would be incompatible. The manual should say "one or two digits"; use &#91;regexp&#93; first if exactly two digits are required.
            ⬜3429618fff2011side effect comment for ::tcl::clock::GetSystemTimeZonepatchREPRODUCED (8.6 only; main/9.0 comment is accurate)

            Open ticket 3429618fff · 06. Time Measurement · filed 2011-10-28

            🧪 Prototype check (2026-10-24)

            3429618fff.patch: VERIFIED

            • Base core-8-6-branch (tip 326e1bf923)
            • Before library/clock.tcl:2982-2984 says it stores the zone in CachedSystemTimeZone unconditionally; with installed 8.6.18 and TCL_LIBRARY=wt-a6-86/library: TZ=Europe/Kyiv -> info exists ::tcl::clock::CachedSystemTimeZone = 0, without TZ -> 1
            • Patch applied cleanly on core-8-6-branch (comment only; does not apply to main, where the proc was rewritten and its comment is accurate)
            • After new comment ("If neither TCL_TZ nor TZ is set ... stores") matches; clock.tcl still loads, same results
            • Test none possible (comment)
            Full verdict

            REPRODUCED (8.6 only; main/9.0 comment is accurate)

            • Platform docs
            • Tested read library/clock.tcl on main, 9.0, 8.6 tips
            • Cause 8.6 library/clock.tcl GetSystemTimeZone header says it always stores the zone in CachedSystemTimeZone; it does so only when neither TCL_TZ nor TZ is set. In 9.0/main the proc was rewritten for the C engine and the header ("stores ... in engine configuration") matches the code, which always calls configure -system-tz.
            • Prototype prototypes/3429618fff.patch (5 changed lines, comment only, against core-8-6-branch; no test)
            • Suggested ticket comment In 9.0 the comment was rewritten together with the C clock engine and is accurate. For 8.6, proposed comment fix: the zone is cached only when neither TCL_TZ nor TZ is set.
            ⬜3418547fff2011cmd literal sharing vs. cmd resolversREPRODUCED (main, 9.0, 8.6) in the form dgp described in 2016: after resolver-1.1's [namespace import], a call to [x] gives "Z" instead of the reso…

            Open ticket 3418547fff · 45. Parsing and Eval · filed 2011-10-04

            Full verdict

            REPRODUCED (main, 9.0, 8.6) in the form dgp described in 2016: after resolver-1.1's [namespace import], a call to [x] gives "Z" instead of the resolver's "Y", and step 2 ([z] outside x) gives "Y" instead of an unknown-command error. The original symptom (no invalidation on [namespace import]) was fixed by TclInvalidateCmdLiteral (trunk, 2011-10-20; resolver-1.x tests).

            • Platform generic
            • Tested tcltest from my main build, the 9.0 tip and the 8.6 tip (work-a3/360128.tcl)
            • Cause cmdName intreps (generic/tclObj.c ResolvedCmdName) cache resolver results and validate them only by namespace and epochs; a context-dependent resolver like tclTest.c:InterpCmdResolver (or nsf's colon resolver) is invisible to that. TclInvalidateCmdLiteral patches only literals. dgp's branch bug-3418547 (2016, removes the cmdName cache as a baseline) was never merged.
            • Prototype none (design issue: needs rules for custom resolvers or uncached resolution; see also d4e7780ca1)
            • Suggested ticket comment Still reproducible on main/9.0/8.6 in dgp's extended form of resolver-1.1 (a final &#91;x&#93; returns Z instead of Y). Branch bug-3418547 was never merged. This needs a decision on what custom command resolvers may depend on; duplicate in spirit of [3601284].
            ⬜3408408fff2011Erratic sharing/duplication of values in [expr]patchREPRODUCED (main, 9.0, 8.6)

            Open ticket 3408408fff · 48. Number Handling · filed 2011-09-13

            🧪 Prototype check (2026-10-24)

            3408408fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before expr-bug-3408408 FAILED: 1 0 0 (proc p b {expr {$b ? 1 : 0}} returns a fresh int with no string rep instead of the shared literal "1")
            • Patch applied cleanly (26 changed lines)
            • After 1 0 1; tests expr 2154/0 failed, expr-old 448/0, execute 157/0, compile 160/0, mathop 385/0, compExpr 80/0, compExpr-old 183/0, incr 69/0, string 1411/0
            • Test existing expr-bug-3408408 fails before / passes after (expr-bug-3408408-1 guards normalization and passes on both)
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested main build (c3bf6dd052), 9.0.5, 8.6.18 with tcl::unsupported::representation: expr {$b?1:0} returns a fresh unshared int with no string rep; expr {$b==1} and expr {$b?$L1:$L2} return shared values.
            • Cause generic/tclExecute.c INST_TRY_CVT_TO_NUMERIC: a shared numeric value with a string rep is always duplicated (string rep dropped) to force canonical form, even when its string rep already is canonical (the literal "1").
            • Prototype prototypes/3408408fff.patch (26 changed lines, tests expr-bug-3408408, expr-bug-3408408-1). For ints only, formats the value with TclFormatInt and keeps the same Tcl_Obj if the string rep already matches; 007, "+1", 0x10, " 2 " are still normalized. Full suite passes. Doubles left alone (Tcl_PrintDouble is costlier).
            • Suggested ticket comment Still present in 9.1: INST_TRY_CVT_TO_NUMERIC duplicates a shared literal even when its string rep is already canonical. Proposed fix for integers (keep the value if TclFormatInt gives the same string), test expr-bug-3408408.
            ⬜3400944fff2011[exit] creates leaksREPRODUCED (main) — by design: [exit] does not unwind, so what unwinding would free is reported by valgrind; changing it needs a TIP (Serhiy's comm…

            Open ticket 3400944fff · 53. Configuration and Build Tools · filed 2011-08-30

            Full verdict

            REPRODUCED (main) — by design: [exit] does not unwind, so what unwinding would free is reported by valgrind; changing it needs a TIP (Serhiy's comment). Not a real leak (the process ends)

            • Platform generic
            • Tested main (2026-10-22) -DPURIFY build, valgrind --leak-check=full: namespace eval foo exit → 168 bytes definitely lost in 5 blocks (+215 indirect); the same script without [exit] → 0
            • Cause generic/tclEvent.c Tcl_Exit() exits without unwinding the C stack of TclEvalEx/TclNRRunCallbacks
            • Existing work none (test suite side already handled: notValgrind constraint, no [testthread exit] in tests)
            • Prototype none (needs a TIP: [exit] as an unwinding exception)
            • Suggested ticket comment none needed (Serhiy's 2026-10-12 comment covers it). Could be retyped as RFE / closed "Won't fix" unless someone writes the TIP.
            ⬜3393714fff2011overflow in toupper deltaREPRODUCED (main, 9.0, 8.6, 8.5) — the remaining part (the delta overflow itself was fixed in 2011)

            Open ticket 3393714fff · 44. UTF-8 Strings · filed 2011-08-18

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5) — the remaining part (the delta overflow itself was fixed in 2011)

            • Platform generic
            • Tested main build, 9.0.5, 8.6.18, 8.5.19: [string toupper]/[string totitle] of U+023F U+0240 U+0250 U+0251 U+0252 U+0265 U+0266 U+026B U+0271 U+027D return the char unchanged, [string tolower] of U+023A U+023E unchanged; the reverse direction (U+2C65 -> U+023A, U+A78D -> U+0265) works, so case mapping is asymmetric. Tcl_UniCharToUpper itself is correct.
            • Cause generic/tclUtf.c Tcl_UtfToUpper/ToLower/ToTitle convert in place and skip any char whose UTF-8 form would grow (2 -> 3 bytes); [string toupper/tolower/totitle] (tclCmdMZ.c, 6 paths) and INST_STR_UPPER/LOWER/TITLE (tclExecute.c) all go through these in-place functions.
            • Prototype none (needs a non-in-place conversion helper wired into the 3 commands x 2 paths plus the bytecode path; well over the size limit, and the public in-place API cannot change).
            • Suggested ticket comment Still present in 9.1: Tcl_UtfToUpper/Lower/Title work in place and skip the 12 chars whose case partner has a longer UTF-8 form, and [string toupper] etc. use them. A fix needs a non-in-place conversion for the Tcl commands; the C API can only be documented.
            ⬜3380073fff2011declining [expr] performanceREPRODUCED (main, 9.0, 8.6): the gap is still there and has not grown since 8.6

            Open ticket 3380073fff · 47. Bytecode Compiler · filed 2011-07-27

            Full verdict

            REPRODUCED (main, 9.0, 8.6): the gap is still there and has not grown since 8.6

            • Platform generic
            • Tested the reporter's distance proc called in a bytecompiled for-loop, best of 5 x 200000 calls (work-a3/338007.tcl), on optimized installed builds: 8.4.20 0.43 us, 8.5.19 0.53 us, 8.6.18 0.80 us, 9.0.5 0.79 us, 9.1b1 0.79 us. My main wt is a debug build and is not used for timing.
            • Cause not a single bug. The usual 8.5/8.6 costs apply: math functions dispatched as ::tcl::mathfunc commands (8.5) instead of dedicated bytecodes, and NRE proc call overhead (8.6). The bytecode of the expressions is unchanged (as ferrieux noted).
            • Prototype none (performance work, no small fix)
            • Suggested ticket comment Still measurable: 0.43 us (8.4.20), 0.53 (8.5.19), 0.80 (8.6.18 / 9.0 / 9.1) per call in a compiled loop. Stable since 8.6, so this is a general performance RFE (mathfunc dispatch, NRE call overhead) rather than a bug. Suggest moving it to the RFE tracker or closing.
            🔧3328635fff2011Virtual Time broken on threaded unixpatchREPRODUCED (8.6; 9.0 by code reading in the select notifier and Tcl_ConditionWait); NOT-A-BUG on main (Tcl_SetTimeProc is unsupported since 9.1, TI…

            Open ticket 3328635fff · 03. Timer Events · filed 2011-06-24

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-21 check-in bc65a512d2 on unix-monotonic-condwait, not merged: Fix 3328635fff: Use CLOCK_MONOTONIC for timed condition waits on Unix where pthread_condattr_setclock() is available, so changes of the system time do not aff
            • 💬 2026-10-21 serhiy.storchaka: On main the timers are monotonic since TIP 723, but timed condition waits on Unix still used an absolute wall-clock deadline (default CLOCK_REALTIME condition variables, deadline from Tcl_GetDayTime()). A step of the system time during a timed wait made it longer or shorter by the size of the step. This affects every timed wait, including the 0-timeout one in zipfs init. Proposed fix in branch unix-monotonic-condw…
            • 💬 2026-10-21 oehhar: Would be great, if a Linux expert could verify. On MS-Win all is ok. Harald
            • 💬 2026-10-21 serhiy.storchaka: Harald, I tested it on Linux: with the wall clock shifted 1 hour forward by an LD_PRELOAD shim, after 2000 hangs with current trunk and takes 2 seconds with the branch. The full test suite passes on Linux, and the notifier, timer, thread and I/O tests on FreeBSD, NetBSD, OpenBSD, DragonFly and OpenIndiana. Windows needs no change, since its waits use relative timeouts. macOS has no pthread_condattr_setclock(), so …
            • 💬 2026-10-21 oehhar: Thanks. Lets give other Linux folks one week of time. If there is no 2nd opinion, we will merge in one week. Thanks for all, Harald
            Full verdict

            REPRODUCED (8.6; 9.0 by code reading in the select notifier and Tcl_ConditionWait); NOT-A-BUG on main (Tcl_SetTimeProc is unsupported since 9.1, TIP 723)

            • Platform unix
            • Tested small extension calling Tcl_SetTimeProc with time = real + 3600 s, then after 200 + vwait: 8.6.18 hangs (killed after 20 s), 9.0.5 on Linux works (epoll notifier, relative timeout). Main: Tcl_SetTimeProc panics "not supported in TCL 9.1"; notifier/timers use Tcl_GetDayTime/Tcl_GetMonotonicTime (TIP 723 merged in acb0a91c2c, 2026-01-08). Repro in work-a4/vt/.
            • Cause unix/tclUnixNotfy.c Tcl_WaitForEvent (8.6; 9.0: tclSelectNotfy.c, used where neither epoll nor kqueue exists) and unix/tclUnixThrd.c Tcl_ConditionWait add the timeout to the virtual Tcl_GetTime() to get an absolute time for pthread_cond_timedwait().
            • Prototype prototypes/3328635fff.patch (against core-8-6-branch; 16 changed lines, no test: there is no script-level access to Tcl_SetTimeProc). Uses gettimeofday() for the absolute deadline. Fixes the repro; timer.test, event.test pass. 9.0 needs the same change in tclSelectNotfy.c and tclUnixThrd.c.
            • Suggested ticket comment Still reproducible with 8.6 (and 9.0 with the select notifier): Tcl_WaitForEvent and Tcl_ConditionWait add the timeout to the virtual Tcl_GetTime() when computing the pthread_cond_timedwait() deadline. Proposed fix for 8.6/9.0: use the real time there. In 9.1 Tcl_SetTimeProc is no longer supported (TIP 723).
            ⬜3307281fff2011Tcl_SetStdChannel not respected by 'puts'patchREPRODUCED (main, 9.0, 8.6)

            Open ticket 3307281fff · 25. Channel System · filed 2011-05-25

            🧪 Prototype check (2026-10-24)

            3307281fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before io-bug-3307281 scenario (child tcltest: puts -nonewline "" and puts -nonewline stdout "" first, then testchannel setstdchannel $f stdout, puts a, puts stdout b) with only tclIO.c unpatched -> stdout got a, the file only b: bare puts still writes to the old stdout
            • Patch applied cleanly (tclTest.c hunk offset 142 lines); 46 changed lines
            • After -> {} {a b }; tests io.test 864 passed/21 skipped/0 failed, chanio.test 766/15/0, ioCmd.test 276/101/0, iogt.test 13/4/0, ioTrans.test 72/34/0
            • Test existing io-bug-3307281 fails before / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested reporter's tout5.c (attachment) built against main (own build), 9.0.5 and 8.6.18: puts "..." without a channel still writes to the old stdout after Tcl_SetStdChannel; puts stdout is also stuck once the literal "stdout" was resolved before the switch (the reporter's second case).
            • Cause generic/tclIO.c TclGetChannelFromObj keeps a cached lookup (chanObjType intrep) valid while the interp and statePtr->epoch are unchanged; Tcl_SetStdChannel changes the channel that "stdin"/"stdout"/"stderr" mean but does not bump the epoch, so cached "stdout" objects (including the TSD stdoutObjPtr used by bare puts) keep the old channel. The 2011 fix.patch (reset the puts TSD) covers only the bare puts case.
            • Prototype prototypes/3307281fff.patch (46 changed lines: 11 in tclIO.c — bump the old std channel's epoch in Tcl_SetStdChannel; the rest is a new testchannel setstdchannel test subcommand and test io-bug-3307281). Fixes tout5.c; io.test, chanio.test, ioCmd.test pass; the test fails without the fix; valgrind clean.
            • Suggested ticket comment Still present in 8.6–9.1: cached lookups of "stdout" are only invalidated by a channel epoch change, which Tcl_SetStdChannel does not do. Proposed fix: bump the epoch of the replaced channel in Tcl_SetStdChannel (test io-bug-3307281).
            ⬜3302370fff2011Performance gap between [global] and [::]REPRODUCED (main, 9.0, 8.6, 8.5, 8.4): the gap is still about 4-5x

            Open ticket 3302370fff · 07. Variables · filed 2011-05-15

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4): the gap is still about 4-5x

            • Platform generic
            • Tested 2M [incr] loop (work-a3/330237.tcl), optimized installed builds. Ratio of :: time to global time: 8.4.20 3.0x, 8.5.19 4.1x, 8.6.18 4.9x, 9.0.5 4.6x, 9.1b1 4.2x (about 40 ms vs 187 ms). Main still compiles [incr ::zzz] to push "::zzz"; incrStkImm (disassembled with my build).
            • Cause generic/tclCompCmds.c / tclCompile.c: qualified names do not get an LVT slot, so every access goes through TclObjLookupVarEx, which has no fast cached path for FQNs in TEBC. msofer's 2011 suggestion (auto-upvar'd qualified compiled locals) was never implemented.
            • Prototype none (compiler optimization work; not a small fix)
            • Suggested ticket comment Still a 4-5x gap on 8.6.18/9.0/9.1 (&#91;incr ::zzz&#93; still compiles to incrStkImm). This is an optimization RFE (msofer's qualified compiled-local idea) rather than a bug.
            ⬜3260718fff2011[while 1 body] costly shimmer of 1patchREPRODUCED (main, 9.0, 8.6), but only for the uncompiled [while]/[for] (non-literal body, e.g. while 1 $body). A literal body is compiled and has…

            Open ticket 3260718fff · 47. Bytecode Compiler · filed 2011-03-30

            🧪 Prototype check (2026-10-24)

            3260718fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before tclsh r326.tcl: after while $cond $body with cond 1 (non-literal body, so uncompiled [while]) -> cond rep: exprcode (the constant is compiled into an expression intrep on every loop entry)
            • Patch applied cleanly (offsets 2 lines; tclInt.h change used only by tclCmdAH.c/tclCmdMZ.c, both rebuilt)
            • After cond rep: int; 1e6-iteration while 1 $body 8.39 s before vs 8.08 s after (debug build, one run each); tests while (48/0 failed), for, while-old, foreach, basic, cmdAH (17001/0), cmdMZ, info: 0 failed
            • Test existing while-8.1 fails before / passes after (while-8.2 passes both, edge case)
            Full verdict

            REPRODUCED (main, 9.0, 8.6), but only for the uncompiled [while]/[for] (non-literal body, e.g. while 1 $body). A literal body is compiled and has no condition test at all.

            • Platform generic
            • Tested main, 9.0.5, 8.6.18 (work-a3/326071*.tcl). After while 1 $body the shared literal "1" has the exprcode intrep, so it shimmers back to int wherever "1" is used as a number. That costs about 5-8% per iteration in a small body (8.5.19: 1.54 vs 1.31 us).
            • Cause generic/tclCmdAH.c:TclNRForIterCallback calls Tcl_NRExprObj(cond) on every iteration, and that compiles the condition into the shared literal's intrep.
            • Prototype prototypes/3260718fff.patch (44 changed lines, tests while-8.1/8.2). It adds a ForIterData.constCond field, set once with Tcl_GetBooleanFromObj in [for]/[while], and skips the expr evaluation for constant conditions. Results match 9.0 over an edge-value table (nan, "", 0x10, " 1 ", 1e400, yes/off, $x, !0) for both while and for. while/for/while-old/foreach/basic/cmdAH/cmdMZ/info tests pass, and while-8.1 fails without the fix.
            • Suggested ticket comment Still present for the uncompiled while/for: the condition literal "1" gets an exprcode intrep on every iteration. Proposed fix: detect a constant boolean condition once and skip the expr evaluation; patch with tests while-8.1/8.2 is available.
            ⬜3185623fff2011Tcl treatment of ".." as path separator differs from WindowsREPRODUCED (main, 9.0, 8.6, 8.5)

            Open ticket 3185623fff · 36. Pathname Management · filed 2011-02-18

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5)

            • Platform Windows
            • Tested main tcltest with testsetplatform windows: [file split C:/foo/bar/...] gives "C:/ foo bar ...", [file join] / [file dirname] treat "..." as a name; repro script win/3185623fff.tcl compares [file normalize $SystemRoot/System32/...] with [file normalize $SystemRoot/System32]
            • Cause win/tclWinFile.c:TclpObjNormalizePath, "Path is just dots" branch, deliberately appends an all-dots component unchanged ("to be nice ... we don't mangle the path"), while Win32 strips trailing dots so "..." opens the same directory as ".".
            • Prototype none (deciding whether to collapse "..." on Windows is a behaviour change; small to do in that branch, but it affects [file normalize] results that scripts may compare)
            • Suggested ticket comment Still the same on 9.1: [file normalize] keeps "..." on Windows although the OS treats it as ".". Collapsing all-dots components in TclpObjNormalizePath() would fix it; needs a decision whether that behaviour change is wanted.
            Windows VM, run 5
            • Platform Windows
            • Verdict REPRODUCED (main, 9.0, 8.6, 8.5)
            • Tested VM win/3185623fff.tcl on 9.1b1, 9.0.5, 8.6.18, 8.5.19 (Windows 11): file exists C:/WINDOWS/System32/... = 1 and isdirectory 1 (Windows opens it as System32), but file normalize gives C:/Windows/System32/... (and .../System32/.../drivers/etc/hosts), not C:/Windows/System32.
            • Cause win/tclWinFile.c TclpObjNormalizePath, "path is just dots" branch, keeps an all-dots component unchanged, while Win32 strips trailing dots.
            • Prototype none (behaviour change of [file normalize] needs a decision, see result-43.md)
            • Suggested ticket comment Still so on Windows 11 with 8.6.18, 9.0.5 and 9.1: C:/Windows/System32/... exists as a directory (Windows treats ... as .), but &#91;file normalize&#93; keeps the ... component. Collapsing all-dots components in TclpObjNormalizePath() would fix it; needs a decision whether that change is wanted.
            🔧3183016fff2011cd .. from link in Windows confuses TclpatchREPRODUCED (main, 9.0, 8.6, 8.5) for junctions (what &#91;file link -symbolic&#93; creates for directories); a mklink /D symlink is consistent (b…

            Open ticket 3183016fff · 36. Pathname Management · filed 2011-02-16

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-19 check-in 854be12c0a on normalize-dotdot-after-link, not merged: Fix 3577801fff, 3570244fff, 3183016fff and 3099612fff: on Unix, [file normalize] and [cd ..] used the target of a link literally when ".." followed it,
            • 💬 2026-10-19 serhiy.storchaka: Still present on Windows 11 with 8.5.19 to 9.1: after cd test1/test2/link; cd .. with a junction to test1, [pwd] and [glob] show the parent of test1, while [file exists], [open] and cmd /c cd use test2. Tcl resolves ".." physically after the link, but Windows itself resolves it lexically. Proposed fix in branch normalize-dotdot-after-link: on Windows ".." is resolved lexically, as the system does (and as Python's …
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5) for junctions (what &#91;file link -symbolic&#93; creates for directories); a mklink /D symlink is consistent (both sides lexical)

            • Platform Windows (Unix has a related cosmetic bug)
            • Tested Linux main, 9.0.5, 8.6.18, 8.5.0, 8.5.19, 8.4.20 with the reporter's error.tcl. On Linux pwd, glob, file exists and open all agree (physical parent of the link target), so the Windows inconsistency is not reproducible here. But with a *relative* link, file normalize link/.. (and pwd after cd ..) returns <target>/.. unresolved since 8.5 (8.4 was lexical).
            • Cause generic/tclPathObj.c:TclFSNormalizeAbsolutePath expands a relative link as dir/../../test1 and strips the last element, leaving dir/../..; unix/tclUnixFCmd.c:TclpObjNormalizePath runs realpath() on everything except the leaf, so a trailing .. survives.
            • Prototype prototypes/3183016fff.patch (21 changed lines, test filesystem-1.55); Unix only: a trailing ./.. leaf goes through realpath() too. fileSystem.test and the full suite pass. Windows repro: win/3183016fff.tcl.
            • Suggested ticket comment On Unix, a relative link followed by a trailing .. was not fully normalized (pwd gave .../test1/..), because TclpObjNormalizePath leaves the leaf to realpath(). Proposed fix with test filesystem-1.55; the Windows part still needs checking on a current release.
            Windows VM, run 5
            • Platform Windows (Unix has a related cosmetic bug, see result-44.md)
            • Verdict REPRODUCED (main, 9.0, 8.6, 8.5) for junctions (what &#91;file link -symbolic&#93; creates for directories); a mklink /D symlink is consistent (both sides lexical)
            • Tested VM win/3183016fff.tcl and work-win5/lk.tcl on 9.1b1, 9.0.5, 8.6.18, 8.6.0, 8.5.19: after cd test1/test2/link; cd .. (link a junction to test1) pwd and glob * show the physical parent of test1, but cmd /c cd, file exists, open use test1/test2: exists(test1)=0, exists(in_test2)=1, a new file lands in test2.
            • Cause win/tclWinFile.c TclpObjChdir passes the translated path (...\link\.., or a relative ..) to SetCurrentDirectoryW, which Windows resolves lexically, while Tcl_FSChdir records the normalized (junction-resolved) path as the Tcl cwd; relative native paths then resolve against a different directory.
            • Prototype prototypes/3183016fff.patch (now Unix part from result-44.md + Windows part, 51 changed lines in total, 30 for Windows; tests filesystem-1.55 and winFile-5.1): TclpObjChdir calls SetCurrentDirectoryW with the native rep of the normalized path. Built on the VM (main): repro consistent (pwd, glob, file exists, open all in the parent of test1); winFile-5.1 passes (fails unpatched); winFile.test 5 failures (4.0–4.4, as unpatched), cmdAH, cmdMZ, fCmd, fileSystem, fileName, exec, winPipe 0 failed, winFCmd 3 failures as unpatched. Same code in 9.0 and 8.6. Tested on Windows.
            • Suggested ticket comment Still happens on Windows 11 (8.5.19 to 9.1) with junctions: TclpObjChdir() gives SetCurrentDirectoryW() the unnormalized path, which Windows resolves lexically (link\.. = the link's parent), while &#91;pwd&#93; and &#91;glob&#93; use the junction-resolved one. Proposed fix: chdir to the normalized path, test winFile-5.1 (plus a Unix fix for a trailing .. after a relative link, test filesystem-1.55).
            ⬜3160902fff2011socket_inet6 tests failingpatchREPRODUCED (main, 8.6 on OpenIndiana with the IPv6 loopback interface down); NOT-REPRODUCED with a normal IPv6 loopback

            Open ticket 3160902fff · 24. Channel Commands · filed 2011-01-18

            🧪 Prototype check (2026-10-24)

            3160902fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before no IPv6-less loopback on this Linux host, so simulated with a process-local LD_PRELOAD shim (work-a4/v4shim/noconn6.c: connect() on AF_INET6 fails with EHOSTUNREACH; bind still works, as on the reporter's host): LD_PRELOAD=.../noconn6.so t.sh main tests/socket.test -match 'socket_inet6-2.*' -> socket_inet6-2.1, 2.2, 2.3 ... FAILED; the full socket.test under the shim hangs (killed by timeout 300)
            • Patch applied cleanly (offset 13; 24 changed lines, test-only)
            • After full socket.test under the shim -> 115 total, 93 passed, 22 skipped, 0 failed, no hang (IPv6 tests skipped); without the shim -> 191 total, 179 passed, 12 skipped, 0 failed (IPv6 tests run as before)
            • Test the patch is the test-suite fix itself (constraint probe); shown failing/hanging before and skipping cleanly after with the shim
            Full verdict

            REPRODUCED (main, 8.6 on OpenIndiana with the IPv6 loopback interface down); NOT-REPRODUCED with a normal IPv6 loopback

            • Platform unix (Solaris 9 / SPARC)
            • Tested code reading; on Linux all socket_inet6 tests run and pass.
            • Cause The host supported AF_INET6 but had no IPv6 on loopback; the old constraint only checked that the family existed. Shortly after this report rmax reworked the constraints (86730a63b2, 2011-06-28, before 8.6.0): supported_inet6 now requires binding a server to the IPv6 loopback address, and localhost_v6 requires that localhost resolves to it. Tcl tests now also discover the loopback addresses (4337c99176, tcltests.tcl). So such a host should skip the IPv6 tests instead of failing.
            • Prototype none
            • Suggested ticket comment The address family constraints in socket.test were reworked in 86730a63b2 (they now need a working IPv6 loopback), and the loopback addresses are now discovered at run time. Solaris 9 is long obsolete; please reopen if a current Solaris/illumos still fails; proposing to close.
            • To run on the OpenIndiana VM (optional): ./tcltest ../tests/socket.test -match 'socket_inet6-*', once with IPv6 on lo0 disabled.
            Unix VMs, run 1
            • Platform unix (Solaris 9 SPARC; tested on illumos)
            • Verdict REPRODUCED (main, 8.6 on OpenIndiana with the IPv6 loopback interface down); NOT-REPRODUCED with a normal IPv6 loopback
            • Tested OpenIndiana. With lo0 inet6 up, socket_inet6-* passes on main and 8.6. With pfexec ifconfig lo0 inet6 down (restored afterwards), binding a server to ::1 still works, so supported_inet6 and localhost_v6 stay true. Then socket_inet6-2.1 "tcp connection" FAILS on main and 8.6, and the next test hangs until killed (blocking socket ::1 $port never returns). This matches the report (2.1 failing on a host where AF_INET6 exists but IPv6 loopback does not work; the error there was EHOSTUNREACH). The 86730a63b2 rework only checks that a bind works.
            • Cause tests/socket.test: the supported_$af constraint only does socket -server -myaddr $localhost 0; an address on a down interface can be bound but not connected to.
            • Prototype prototypes/3160902fff.patch (24 changed lines, test-only): new helper canConnect binds a server and makes a socket -async connection to it with a 2 s limit; supported_inet/supported_inet6 use it. OpenIndiana with lo0 inet6 down: all IPv6 tests skipped, no hang, socket_inet-* 71 passed/0 failed. With lo0 up: socket_inet6-* 71/0 failed, full socket.test unchanged (the same 11 socket-14.x failures as unpatched). Linux (9.0 tcltest): socket.test 191/0 failed.
            • Suggested ticket comment Reproduced on OpenIndiana with the IPv6 loopback down: ::1 can still be bound, so supported_inet6 is true, but connections fail or hang (socket_inet6-2.1, as reported). Proposed fix: the constraint also checks that an async connect to the probe server succeeds.
            • Side finding (not this ticket): on OpenIndiana socket.test has 11 failures on main and 9 on 8.6 (socket-14.2, 14.7.2, 14.8.2, 14.11.0, 14.12, 14.14–14.19; [socket -async] to a port with no listener) even with a normal loopback, plus scan-15.1 fails. These may deserve their own tickets.
            ⬜3107983fff2010Glitch in the constant expression optimization.patchREPRODUCED (main, 9.0, 8.6, 8.5) inside procs: expr {foo(1?" 0x42 ":" 0x41")} passes 66 to the function, while $x?... passes " 0x42 ". At top l…

            Open ticket 3107983fff · 47. Bytecode Compiler · filed 2010-11-12

            🧪 Prototype check (2026-10-24)

            3107983fff.patch: VERIFIED (the semantic choice against TIP 232's wording still needs a maintainer decision, as the triage says)

            • Base main (tip d6dd6d0f71)
            • Before tcltest 3107.tcl (inside a proc, mathfunc foo returns <$x>) -> foo($x?" 0x42 ":" 0x41") = < 0x42 > but foo(1?" 0x42 ":" 0x41") = <66>; also foo(1 ? 0x10 : 2) = <16>, while foo(0x10) and foo($x ? 0x10 : 2) give <0x10>
            • Patch applied cleanly (28 changed lines)
            • After foo(1?" 0x42 ":" 0x41") = < 0x42 >, foo(1 ? 0x10 : 2) = <0x10> (now the same as the non-constant forms); whole-expression results unchanged: {1?" 0x42 ":"a"} = 66, {(1?" 0x42 ":"a")+1} = 67, {" 0x42 "} = 66. Tests compExpr.test 81 passed/2 skipped/0 failed, expr.test 2152/16/0, expr-old.test 448/13/0, compile.test 160/11/0, mathop.test 385/0/0, assemble.test 277/4/0, execute.test 157/0/0
            • Test existing compExpr-9.1 fails before / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5) inside procs: expr {foo(1?" 0x42 ":" 0x41")} passes 66 to the function, while $x?... passes " 0x42 ". At top level (one-off, no folding) both pass " 0x42 ".

            • Platform generic
            • Tested main, 9.0.5, 8.6.18, 8.5.19 (work-a3/310798.tcl, 310798b.tcl: a 17-expression table)
            • Cause generic/tclCompExpr.c:CompileExprTree. When a constant subtree is folded, ExecConstantExprTree compiles it with the QUESTION node as root, and the case START: case QUESTION: if (convert && nodePtr == rootPtr) TRY_CVT_TO_NUMERIC rule then converts it. The non-constant path converts only at the real root.
            • Prototype prototypes/3107983fff.patch (28 changed lines, test compExpr-9.1). It converts only at START and sets convert=true after folding, so the outer root still converts the folded value. Only the constant-in-function-argument rows change; whole-expression results ({1?" 0x42 ":"a"} -> 66, + 1 -> 67) stay the same. compExpr/expr/expr-old/compile/mathop/assemble tests pass. dgp noted that TIP 232 literally prescribes the converted form, so this picks the long-standing non-constant behaviour; that needs a maintainer decision.
            • Suggested ticket comment Still present on 8.5-main inside procs: folding a constant ?: subexpression applies the end-of-expression TRY_CVT_TO_NUMERIC to it, so mathfunc arguments differ between constant and variable conditions. Proposed fix (consistency with the non-constant case) with test compExpr-9.1 is available.
            ⬜3059758fff2010tclTraceExec==3 : wrong pc+INST on uncaught error returnpatchREPRODUCED (main, only with -DTCL_COMPILE_DEBUG and tcl_traceExec 3; code identical in 9.0/8.6)

            Open ticket 3059758fff · 47. Bytecode Compiler · filed 2010-09-05

            🧪 Prototype check (2026-10-24)

            3059758fff.patch: VERIFIED. Aside (pre-existing, not from this patch): for a caught TCL_BREAK the trace prints TCL_OK, range at 40, new pc 56 because result = TCL_OK is set before the TRACE_APPEND in processExceptionReturn.

            • Base main (tip d6dd6d0f71), out-of-tree build work-a4/build-cdebug (CFLAGS=-DTCL_COMPILE_DEBUG configure --enable-symbols)
            • Before work-a4/v4trace.tcl (proc B {} {error boom}; proc A {} {B; return ok}; set tcl_traceExec 3; catch {A}) -> 2: 1 (9) push1 => TCL_ERROR, result="boom" and 1: 1 (9) push1 => ...: pc 9 is the last operand byte of (5) invokeStk 1, decoded as an opcode
            • Patch applied cleanly (35 changed lines)
            • After -> 2: 1 (10) pop 0 => ... after "B": TCL_ERROR, result="boom" / 1: 1 (10) done 0 => ... after "A": TCL_ERROR, ..., same format as the TCL_OK trace ((10) pop 0 => ... after "B": TCL_OK, result="x"); a caught error traces ... after "B": TCL_ERROR, result="boom" + found catch at 5. execute.test on the debug build: 157 passed / 0 failed; normal build -file 'execute.test compile.test nre.test basic.test': 504 total, 464 passed, 40 skipped, 0 failed
            • Test none possible (output of a -DTCL_COMPILE_DEBUG-only stdout trace)
            Full verdict

            REPRODUCED (main, only with -DTCL_COMPILE_DEBUG and tcl_traceExec 3; code identical in 9.0/8.6)

            • Platform generic (TCL_COMPILE_DEBUG builds only)
            • Tested main built with -DTCL_COMPILE_DEBUG, attached traceexec.tcl: prints 1: 3 (19) pop => TCL_ERROR where 19 is the last operand byte of (15) invokeStk 3.
            • Cause tclExecute.c TEBCresume: the NR callback does pc-- before goto processExceptionReturn, whose TCL_COMPILE_DEBUG switch (*pc) then decodes an operand byte as opcode; its INVOKE/EVAL cases are therefore dead (all those paths come through the callback).
            • Prototype prototypes/3059758fff.patch (35 changed lines, no test: debug-only stdout trace). Traces in the callback before pc-- (same format as the TCL_OK trace, e.g. (20) done 0 => ... after "B": TCL_ERROR) and jumps past the trace in processExceptionReturn; removes the dead switch. execute/compile/nre/basic.test pass (normal build; execute.test also with the debug build).
            • Suggested ticket comment Still present: after pc-- in the NR callback the trace in processExceptionReturn decodes an operand byte. Proposed fix: trace the failed invocation before backing up pc, like the TCL_OK case does, and drop the now dead switch (*pc).
            ⬜3057639fff2010lappend inconsistency with respect to read tracesREPRODUCED (main, 9.0, 8.6, 8.5) (the inconsistency msofer reopened with in 2012; the ::env part was fixed in 2010)

            Open ticket 3057639fff · 17. Commands I-L · filed 2010-09-01

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5) (the inconsistency msofer reopened with in 2012; the ::env part was fixed in 2010)

            • Platform generic
            • Tested main (own build), 9.0.5, 8.6.18, 8.5.19. With a read trace that sets the variable to 1, compiled lappend foo 2 gives 2 (no read trace) and eval lappend foo 2 gives 1 2 (trace fired). lappend ::env(NOSUCH) x works in both paths on all versions. appendComp-7.2..7.8 still carry the bug_3057639 constraint (never run).
            • Cause e6f97d1e4a (2010) dropped TCL_TRACE_READS from the INST_LAPPEND_* store flags (tclExecute.c, doStore*). Direct Tcl_LappendObjCmd (tclVar.c) still reads through TclPtrGetVarIdx: it fires read traces and silently ignores trace errors (treats them as "variable does not exist"), which is why the ::env/nonull case passes there. I tried adding TCL_TRACE_READS back to the 6 LAPPEND flag sites: compiled lappend then fires the trace (1 2), but appendComp-7.3/9.0/9.1 fail, because TclPtrSetVarIdx turns a trace error into a command error where the direct path ignores it. A consistent fix needs a decision: either "lappend fires read traces and ignores their errors" in both paths (new TclPtrSetVarIdx flag), or no read traces in either, like append. msofer and dgp considered this TIP-worthy.
            • Prototype none (needs a behaviour decision; experiment reverted)
            • Suggested ticket comment Still inconsistent on 8.6/9.0/main: compiled lappend skips read traces, direct lappend fires them but ignores their errors (the reason the ::env case works). Restoring TCL_TRACE_READS in INST_LAPPEND_* alone breaks appendComp-9.x, because TEBC would propagate trace errors. We need to decide which semantics we want (probably a TIP) before fixing.
            🔧3048170fff2010bgerror called when limit exceed proc calls [vwait]patchREPRODUCED (main, 9.0, 8.6; 8.5.19 shows the related #3027822 symptom instead)

            Open ticket 3048170fff · 20. [interp] · filed 2010-08-18

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-17 check-in b8ee0282eb on time-limit-handler-reentry, not merged: Fix \[3027822fff\] and \[3048170fff\]: the time limit timer callback could re-run the limit check while the time limit handlers were already running (e.g. one e
            • 💬 2026-10-17 serhiy.storchaka: Still present in 8.6–9.1. Same cause as 3027822fff: the child's time-limit timer fires inside the limit handler's [vwait] and raises a background error in the busy child, whose bgerror handler cannot run because the limit is exceeded. Proposed fix in branch time-limit-handler-reentry (shared with 3027822fff): TimeLimitCallback does nothing while the time-limit handlers are running. Test interp-34.15.
            Full verdict

            REPRODUCED (main, 9.0, 8.6; 8.5.19 shows the related #3027822 symptom instead)

            • Platform generic
            • Tested attached testInterpLimits2.tcl on main (9.1.1 build), 9.0.5, 8.6.18, 8.5.19 (40 s timeout). 9.x/8.6: "error in background error handler: time limit exceeded" each time the limit is renewed; 8.5: sin() returns {} ("Break now").
            • Cause same as #3027822: tclInterp.c TimeLimitCallback. The child exceeds its time limit at a command boundary, the limit handler (in the parent) runs [vwait], the child's still-pending time-limit timer fires inside it, Tcl_LimitCheck fails again (the active handler is skipped) and TimeLimitCallback calls Tcl_BackgroundException on the busy child. The child's bgerror handler cannot run because the limit is exceeded.
            • Prototype prototypes/3048170fff.patch (= 3027822fff.patch, 36 changed lines, test interp-34.15). TimeLimitCallback returns early if TCL_LIMIT_TIME is already marked exceeded (its handlers are running); a new limit re-arms the timer. Both attached scripts then run cleanly; interp/event/timer.test pass.
            • Suggested ticket comment Still present in 8.6–9.1. The child's time-limit timer fires inside the limit handler's &#91;vwait&#93; and raises a background error in the busy child. Proposed fix (shared with &#91;3027822&#93;): skip TimeLimitCallback while the time-limit handlers are running; test interp-34.15.
            ⬜3048057fff2010race in Tcl_MutexLock ?patchREPRODUCED (main, 9.0, 8.6 by code; ThreadSanitizer report on main)

            Open ticket 3048057fff · 49. Threading · filed 2010-08-18

            🧪 Prototype check (2026-10-24)

            3048057fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before out-of-tree TSan build of the tip (work-a2/v2/tsan, -fsanitize=thread -g -O1 -DPURIFY), 4 threads locking one static Tcl_Mutex (work-a2/v2/mtx.c, copied from work-a6): runmtx.sh 40 -> data race in 40/40 runs, at tclUnixThrd.c:647/662 (unlocked *mutexPtr read vs. store under globalLock in Tcl_MutexLock) and :109/:114 (PMutexLock owner field)
            • Patch applied with offset +8 (re-saved, 52 changed lines, content unchanged)
            • After same, 40 runs -> 0 runs with a race report; tests event, io, thread, notify, async, timer: 1176 total, 1090 passed, 86 skipped, 0 failed; with Thread 3.0.7 (TCLLIBPATH=work-a2/thread) thread, async, ioTrans, socket: 364 total, 360 passed, 4 skipped, 0 failed
            • Test none possible in tcltest (needs a TSan build); the TSan program above is the repro
            Full verdict

            REPRODUCED (main, 9.0, 8.6 by code; ThreadSanitizer report on main)

            • Platform unix (generic double-checked locking; win/tclWinThrd.c has the same pattern)
            • Tested TSan build of main (work-a6/tsan, -fsanitize=thread -DPURIFY) with a 4-thread C program locking one static Tcl_Mutex (work-a6/3048057-mtx.c): data race reported 5/5 runs at tclUnixThrd.c:639 (unlocked *mutexPtr == NULL read) vs :648 (store under globalLock), exactly the ticket's report; plus a second race on the reentrant-mutex owner field (PMutexLock reads pmutexPtr->thread unlocked; 9.0/main only, 8.6 has no PMutex). With the prototype: 0 reports in 40/40 runs. 8.6 and 9.0 have the same unlocked read in Tcl_MutexLock/Tcl_ConditionWait.
            • Cause unix/tclUnixThrd.c Tcl_MutexLock/Tcl_ConditionWait/Tcl_ConditionNotify use double-checked locking with plain loads/stores of the lazily created pointer (no acquire/release), and PMutexLock/Unlock/PCond*Wait access the owner thread field without atomics. Benign on x86 in practice, formally a data race (could matter on weakly ordered CPUs).
            • Prototype prototypes/3048057fff.patch (52 changed lines incl. comment; no tcltest test, needs TSan; thread.test with Thread, event/socket/io.test and Thread's own suite pass) — GCC/clang __atomic acquire/release for the published pointers, relaxed atomics for the owner field; plain accesses on other compilers.
            • Suggested ticket comment Still present: ThreadSanitizer reports the unlocked *mutexPtr read in Tcl_MutexLock on 9.1 (and a similar one on the reentrant owner field in PMutexLock). Proposed fix: use __atomic acquire/release loads/stores for the lazily created mutex/condition pointers and relaxed atomics for the owner field (patch available, unix only).
            ⬜3047802fff2010inconsistent ns treatment of aliasespatchREPRODUCED (main, 9.0, 8.6)

            Open ticket 3047802fff · 20. [interp] · filed 2010-08-18

            🧪 Prototype check (2026-10-24)

            3047802fff.patch: VERIFIED+TEST-ADDED (doc + behaviour-pinning tests)

            • Base main (tip d6dd6d0f71)
            • Before interp.n says only that srcCmd "gives the name of a new command, which will be created in the source interpreter", and that targetCmd "is used to locate a command procedure in the target interpreter". Behaviour (work-a1/3047.tcl, 3047b.tcl; identical on main and 8.6 tip): in namespace eval ns, interp alias {} myputs {} puts creates ::myputs; interp alias {} ns::myputs2 {} puts creates ::ns::ns::myputs2 (token ns::myputs2); ::ns::myputs3 is used as given. child alias x puts from inside a namespace creates ::x in the child. An unqualified targetCmd tgt given inside ns (where ns::tgt exists) calls the global ::tgt when invoked, also when invoked from ns.
            • Patch 57 changed lines. doc/interp.n: srcCmd is interpreted as by Tcl_CreateObjCommand, not as by proc (unqualified -> global namespace of the source interp, even when called from another namespace; relative qualified -> relative to the current namespace; recommend fully-qualified names); child alias refers to that rule; ALIAS INVOCATION: a not fully qualified targetCmd is looked up in the target interp's global namespace at each invocation. tests/interp.test: interp-27.9..27.12.
            • After man page renders, troff warnings 2/2 (the same two pre-existing "expected numeric expression" warnings, now at shifted lines). interp.test on main: 395 total, 388 passed, 7 skipped, 0 failed.
            • Test added interp-27.9 (unqualified srcCmd from namespace eval foo -> ::aliasBug3047), interp-27.10 (foo::x from inside foo -> ::foo::foo::x, token as given; ::foo::y as given), interp-27.11 (unqualified targetCmd resolves to the global proc, not foo::), interp-27.12 ($child alias from a namespace -> child's ::). Pins current behaviour: pass on main; the same 4 tests extracted into work-a1/p3047.test pass on the 8.6 tip (main's interp.test itself does not load in 8.6: it needs tcl::build-info).
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested main, 8.6.18: in namespace eval ns, interp alias {} myputs {} puts creates ::myputs, interp alias {} ns::myputs2 {} puts creates ::ns::ns::myputs2 (and interp aliases lists the token ns::myputs2, not the real name).
            • Cause tclInterp.c TclAliasCreate passes the name to Tcl_CreateObjCommand2/Tcl_NRCreateCommand2, whose documented rule (CrtObjCmd.3) is: unqualified → global namespace, qualified → relative to the current namespace. [proc] resolves both relative to the current namespace.
            • Prototype none (behaviour change: either rule would break existing scripts; needs a decision, and interp.n should document the rule either way)
            • Suggested ticket comment Still so in 8.6–9.1. Aliases follow the Tcl_CreateObjCommand rule (unqualified names go to the global namespace, qualified names are relative to the current one), unlike &#91;proc&#93;. Changing that is incompatible; at least interp.n should document it.
            ⬜3046359fff2010Tcl_FSNewNativePath() can destroy intreppatchREPRODUCED (main, 9.0, 8.6; by code reading, unchanged)

            Open ticket 3046359fff · 37. File System · filed 2010-08-16

            🧪 Prototype check (2026-10-24)

            3046359fff.patch: VERIFIED+TEST-ADDED (after rework; the original prototype leaked the clientData)

            • Base main (tip d6dd6d0f71)
            • Before new test command testnewnativepath /foo/bar (tclTest.c: a minimal registered "nativepath" filesystem whose internalToNormalizedProc caches its own native rep "cached:<path>" through Tcl_FSGetInternalRep; the command passes clientData "native:<path>" to Tcl_FSNewNativePath and reports the native rep kept and how many reps were freed after release) -> native:/foo/bar 2: the driver's cached rep is thrown away, as the ticket says
            • Patch applied cleanly (offset 6), but the ORIGINAL patch is BROKEN in one respect: it returns early for any path-typed object and never stores or frees clientData, so the clientData passed in leaks; the same test with the original patch gives cached:/foo/bar 1 (only one of the two reps freed). It also skipped the conversion when the intrep belonged to another filesystem (e.g. a driver that normalized through the native fs), so the new object was no longer tied to fromFilesystem. Reworked: return early only when the intrep's fsPtr is fromFilesystem; then store clientData if the native rep is still NULL, free it with freeInternalRepProc if the native rep is a different one. Otherwise the old code path. 17 changed lines in tclPathObj.c; with the new test filesystem and command in tclTest.c (~105 lines) and the test, 127 changed lines in all
            • After testnewnativepath /foo/bar -> cached:/foo/bar 2; tests fileSystem.test 142 passed/21 skipped/0 failed, cmdAH.test 16971/30/0, fCmd.test 260/45/0, fileName.test 259/47/0, unixFCmd.test 30/19/0, unixFile.test 7/0/0, zipfs.test (make test) 498/37/0
            • Test added filesystem-14.1 (constraint testnewnativepath) plus the testnewnativepath command; fails before (native:/foo/bar 2) / passes after; it also catches the leak in the original patch (cached:/foo/bar 1)
            Full verdict

            REPRODUCED (main, 9.0, 8.6; by code reading, unchanged)

            • Platform generic
            • Tested code reading of main/9.0/8.6 generic/tclPathObj.c. No core filesystem has an internalToNormalizedProc that sets a path intrep (zipfs has NULL), and nothing in the core calls Tcl_FSNewNativePath, so only extension filesystems are affected.
            • Cause generic/tclPathObj.c:Tcl_FSNewNativePath always frees the returned object's intrep and builds a new FsPath. If the driver kept clientData in its own FsPath, freeing that rep calls freeInternalRepProc on the same clientData that is then stored again (use-after-free), not just wasted work.
            • Prototype prototypes/3046359fff.patch (8 changed lines, no test: needs a test filesystem with internalToNormalizedProc); returns the object unchanged if it already has fsPathType. The fs/path tests pass (zipfs.test's 25 failures happen on an unpatched build too when TCL_LIBRARY is set).
            • Suggested ticket comment Still as described. Proposed fix: Tcl_FSNewNativePath returns the driver's object as is when it is already of the path type. That also avoids freeing clientData if the driver stored it in its own intrep.
            ⬜3038316fff2010Encoding freeProc deadlock riskpatch 86REPRODUCED (8.6 only; not on 8.7b1, 9.0, main)

            Open ticket 3038316fff · 10. Objects · filed 2010-08-02

            🧪 Prototype check (2026-10-24)

            3038316fff-86.patch: VERIFIED (doc only, no test possible)

            • Base core-8-6-branch (tip 326e1bf923). 8.6 only: since TIP 509 (8.7/9.0) Tcl mutexes are reentrant and the same-thread case no longer deadlocks (main survives the repro).
            • Before Encoding.3 says only "This freeProc function is called when the encoding is deleted"; no word about locking. Behaviour: work-a2/enc/encfree.c (copied to work-a1/enc; freeProc calls Tcl_GetEncoding/Tcl_FreeEncoding, then the last reference is dropped): timeout 5 wt-a1-86/unix/tcltest work-a1/encfree.tcl encfree86.so -> hangs, rc=124; main ... encfree91.so -> survived. Cause: tclEncoding.c FreeEncoding() calls freeProc with encodingMutex held (also from Tcl_SetSystemEncoding and TclFinalizeEncodingSubsystem); the functions taking that mutex in 8.6 are Tcl_GetEncoding, Tcl_FreeEncoding, Tcl_GetEncodingNames, Tcl_SetSystemEncoding, Tcl_CreateEncoding (and Init/Finalize).
            • Patch 10 changed lines, doc/Encoding.3 only: freeProc runs with Tcl's internal, non-reentrant encoding lock held; it must not call Tcl_GetEncoding, Tcl_GetEncodingFromObj, Tcl_FreeEncoding, Tcl_GetEncodingNames, Tcl_CreateEncoding, Tcl_SetSystemEncoding (directly or via e.g. Tcl_EvalObjEx), or wait for a thread that may call them.
            • After man page renders; troff warnings 1/1 (pre-existing).
            • Test none possible (a C-API locking note; a test would have to deadlock the test suite, and the documented rule is a restriction on callers, not behaviour Tcl can check).
            Full verdict

            REPRODUCED (8.6 only; not on 8.7b1, 9.0, main)

            • Platform generic (deadlock seen on Unix 8.6)
            • Tested small extension (work-a2/enc/encfree.c) that creates an encoding whose freeProc calls Tcl_GetEncoding/Tcl_FreeEncoding, then drops the last reference: 8.6.18 hangs (timeout), 8.7b1, 9.0.5 and main print "survived".
            • Cause tclEncoding.c FreeEncoding calls freeProc with encodingMutex held (unchanged in all versions). Since TIP 509 (8.7/9.0) Tcl mutexes are reentrant, so a freeProc using encoding functions in the same thread no longer deadlocks; in 8.6 (pthread mutex) it does. A cross-thread lock-order deadlock is still possible in theory if the freeProc waits for a lock held by a thread that is calling encoding functions.
            • Prototype none (in 9.x the remaining risk is theoretical; for 8.6 a doc warning in Encoding.3 is the only cheap remedy, restructuring the locking is not small)
            • Suggested ticket comment Since TIP 509 (reentrant mutexes, 8.7/9.0) a freeProc may call encoding functions; in 8.6 this still deadlocks (freeProc runs with encodingMutex held). Suggest a one-line warning in Encoding.3 on core-8-6-branch and closing for 9.x.
            🔧3027822fff2010Failure of [expr] with [interp limit] and [vwait]patchREPRODUCED (main, 9.0, 8.6, 8.5)

            Open ticket 3027822fff · 20. [interp] · filed 2010-07-10

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-17 check-in b8ee0282eb on time-limit-handler-reentry, not merged: Fix \[3027822fff\] and \[3048170fff\]: the time limit timer callback could re-run the limit check while the time limit handlers were already running (e.g. one e
            • 💬 2026-10-17 serhiy.storchaka: Still present in 8.6–9.1. The child hits its time limit, the limit handler enters [vwait], and the child's pending time-limit timer fires there. It runs the limit check again and raises a background error in the busy child, which clobbers the child's interp result. The math function then returns garbage to [expr]. Proposed fix in branch time-limit-handler-reentry: TimeLimitCallback does nothing while the time limi…
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5)

            • Platform generic
            • Tested attached testInterpLimits.tcl on main (9.1.1 build), 9.0.5, 8.6.18, 8.5.19: expr {cos($nloop) * sin($nloop)} fails with "cannot use a list as left operand" (9.x), "non-numeric string" (8.6), "empty string" (8.5); 9.x/8.6 also print "error in background error handler: time limit exceeded".
            • Cause tclInterp.c TimeLimitCallback. The child hits its time limit inside a nested command (the math function), the handler enters [vwait], the child's pending time-limit timer fires there, and Tcl_LimitCheck plus Tcl_BackgroundException overwrite the busy child's interp result. When the handler extends the limit, the math function "returns" that garbage to [expr].
            • Prototype prototypes/3027822fff.patch (36 changed lines, test interp-34.15). TimeLimitCallback returns early when TCL_LIMIT_TIME is already marked exceeded (the handlers are running in an outer Tcl_LimitCheck, which decides; a new limit re-arms the timer). Both attached scripts run cleanly; interp-34.15 fails 2/3 runs without the fix, 10/10 passes with it; interp/event/timer.test pass.
            • Suggested ticket comment Still present in 8.5–9.1: the child's time-limit timer fires inside the limit handler's &#91;vwait&#93; and clobbers the child's interp result. Proposed fix: TimeLimitCallback does nothing while the time-limit handlers are running (also fixes &#91;3048170&#93;); test interp-34.15.
            ⬜3021922fff2010Error loading statically built bundled packages ITcl & tdbcREPRODUCED (9.0, 8.6; main has the same pkgs rule) for Itcl; the lib\lib\itcl4.x install dir part is FIXED (installs to lib\itcl4.3.4); tdbc not …

            Open ticket 3021922fff · 81. Bundled Packages · filed 2010-06-27

            Full verdict

            REPRODUCED (9.0, 8.6; main has the same pkgs rule) for Itcl; the lib\lib\itcl4.x install dir part is FIXED (installs to lib\itcl4.3.4); tdbc not tested (no tdbc sources offline)

            • Platform Windows (MSVC makefile.vc)
            • Tested code reading only (win/makefile.vc, win/rules.vc on main/9.0/8.6). The pkgs/ target calls each package's makefile.vc with only TCLDIR=$(ROOT) and $(MAKEFLAGS); whether OPTS=static and INSTALLDIR leak through (giving static .lib packages and the reported lib\lib\itcl4.x install dir) depends on nmake macro inheritance and the TEA rules-ext.vc in the bundled itcl/tdbc. On 9.x "staticpkg" is ignored (makefile.vc:150-159).
            • Cause unknown without a Windows build; candidates are win/makefile.vc pkgs/install-pkgs targets and the package's rules.vc _INSTALLDIR=$(_INSTALLDIR)\lib adjustment (rules.vc:420) when INSTALLDIR already ends in \lib.
            • Prototype none (needs Windows). Repro: win/3021922fff.cmd <tcl-src-with-itcl-and-tdbc-in-pkgs> (static build+install into %TEMP%, prints VERDICT).
            • Suggested ticket comment Needs re-checking with a current 9.0/8.6 static MSVC build (nmake OPTS=static) with itcl/tdbc in pkgs/; the 2010 report predates the TEA rules-ext.vc rewrite.
            Windows VM, run 4
            • Platform build (Windows, nmake makefile.vc)
            • Verdict REPRODUCED (9.0, 8.6; main has the same pkgs rule) for Itcl; the lib\lib\itcl4.x install dir part is FIXED (installs to lib\itcl4.3.4); tdbc not tested (no tdbc sources offline)
            • Tested VM, MSVC 2022 x64: fresh checkouts of core-8-6-branch and core-9-0-branch with Itcl 4.3.4 (itcl.fossil tag itcl-4-3-4) in pkgs\, win/3021922fff.cmd (nmake OPTS=static,staticpkg INSTALLDIR=... release + install, work-win4/b3021.cmd). Both builds succeed, but Itcl is built as a static library (itcl434ts.lib / tcl9itcl434s.lib) and installed with a pkgIndex.tcl that loads itcl434ts.dll / tcl9itcl434.dll, so package require itcl in the installed tclsh86ts/tclsh90s fails with "couldn't load library ...". Build dirs removed afterwards.
            • Cause win/makefile.vc pkgs/install-pkgs run the package's makefile.vc with $(MAKE) -$(MAKEFLAGS) ... TCLDIR=$(ROOT); nmake passes the command-line OPTS=static on to the sub-make, so packages become static libraries, which neither the static tclsh links in (only dde/registry are linked statically on 9.x) nor load can use.
            • Prototype none (design choice: build bundled packages shared in static builds, e.g. override OPTS for the sub-make, or link them into tclsh; needs a decision and a test of both)
            • Suggested ticket comment Still happens with 8.6 and 9.0 (Itcl 4.3.4, MSVC 2022): with OPTS=static the pkgs\ packages inherit the option, are built as .lib and installed with a pkgIndex that tries to load a DLL, so &#91;package require itcl&#93; fails. The lib\lib install path is fixed. Either the pkgs targets should build the packages shared, or static builds must link them in.
            ✅3019348fff2010Nested lset can modify list on errorpatchREPRODUCED (main, 9.0, 8.6; not 8.5, which has no append-by-lset)

            Open ticket 3019348fff · 14. List Object · filed 2010-06-22

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Fixed)

            • 🔧 2026-10-17 check-in c4114a5e1f on lset-undo-append-on-error, in main: Fix 3019348fff: a nested [lset] which failed after appending an empty sublist for an end+1 index left that sublist in the list. Remove it on error.
            • 🔧 2026-10-19 check-in 6a23978bf0 on trunk, in main, 9.0: Fix [3019348] - Nested lset can modify list on error
            • 🔧 2026-10-19 check-in 27d3ea715e on core-9-0-branch, in main, 9.0: Fix [3019348] - Nested lset can modify list on error
            • 🔧 2026-10-19 check-in 479483454f on apn-lset-undo-append-on-error, in main: Merge trunk: keep the 526b01c0b9 reference hold in the replace branch of TclLsetFlat.
            • 💬 2026-10-17 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1: since TIP 331 (41ad4c4b67) TclLsetFlat() appends an empty sublist for an index equal to the list length before checking the remaining indices, and does not remove it when one of them fails. Inside a procedure the literal list is shared and copied, which hides the problem. Proposed fix in branch lset-undo-append-on-error: remove the appended sublist on error. Tests lset-17.1 to …
            • 💬 2026-10-19 apnadkarni: @serhiy.storchaka Looks fine but I've made some modifications in branch [apn-lset-undo-append-on-error](https://core.tcl-lang.org/tcl/timeline?r=apn-lset-undo-append-on-error) to eliminate some redundant/unnecessary checks. If you are ok with those changes, you can commit that branch to trunk and 9.0. Or I can, if you prefer.
            • 💬 2026-10-19 serhiy.storchaka: Your changes look fine to me. The branch conflicted with trunk, where the fix for 526b01c0b9 holds a reference to the sublist around TclListObjSetElement(); I merged trunk into apn-lset-undo-append-on-error and kept that hold in your replace branch (lset-19.1 crashes without it). Please merge it to trunk and 9.0.
            • 💬 2026-10-19 apnadkarni: Fixed for 9.1 in 6a23978bf0. Pending 9.0 merge.
            • 💬 2026-10-19 apnadkarni: Fixed for 9.0 in 27d3ea715e. Closing. Thanks!
            Full verdict

            REPRODUCED (main, 9.0, 8.6; not 8.5, which has no append-by-lset)

            • Platform generic
            • Tested main build, 9.0.5, 8.6.0, 8.6.18, 8.5.0/8.5.9/8.5.19. set foo {{x x} {x x}}; lset foo 2 1 x errors but leaves foo = {x x} {x x} {} at global level / with lset foo {2 1} x / lset foo 2 bogus x. Inside a proc the literal is shared and copied, so the bug is hidden there.
            • Cause generic/tclListObj.c TclLsetFlat: when an intermediate index equals the list length (TIP 331 append, 41ad4c4b67, 2008-10-05), an empty sublist is appended to the (unshared) parent before the remaining indices are validated, and is not removed when a later index fails.
            • Prototype prototypes/3019348fff.patch (26 changed lines, tests lset-17.1..17.3). Remembers the outermost list an empty sublist was appended to and removes it with Tcl_ListObjReplace on error. lset.test fails 3 tests on 9.0 tip, passes with the patch; full suite passes (only cmdAH-20.5 atime failed, unrelated filesystem timing).
            • Suggested ticket comment Still present in 9.1: TclLsetFlat appends an empty sublist for an end+1 intermediate index before checking the remaining indices (since TIP 331, 41ad4c4b67). Proposed fix: undo the append on error, tests lset-17.*.
            ⬜3017879fff2010proc emptyness is cached too aggressivelypatchREPRODUCED (main, 9.0, 8.6, 8.5)

            Open ticket 3017879fff · 47. Bytecode Compiler · filed 2010-06-18

            🧪 Prototype check (2026-10-24)

            3017879fff.patch: VERIFIED (as noted in triage: imports of a target redefined as an empty proc lose the no-op inlining, because the new proc's compileProc is set only after relinking; harmless)

            • Base main (tip d6dd6d0f71)
            • Before tclsh r301.tcl (import an empty args proc, call it from a compiled proc, redefine the target with a puts body, call again) -> only done, the new body is not run
            • Patch applied cleanly
            • After You have to do _something_.. then done; tests namespace (315/0 failed), namespace-old, proc, basic, compile, rename, oo, ooNext2, ooProp: 0 failed
            • Test existing namespace-9.10 fails before / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5)

            • Platform generic
            • Tested main, 9.0.5, 8.6.18, 8.5.19: after namespace import foo::do-nothing of an empty args proc, redefining foo::do-nothing with a body is ignored by compiled callers (nothing printed). Point 3 of the report (enter traces on empty procs) works on all versions.
            • Cause tclNamesp.c Tcl_Import copies the target's compileProc (TclCompileNoOp for an empty proc) into the import command. On [proc] redefinition tclBasic.c Tcl_CreateObjCommand2/TclCreateObjCommandInNs relink the imports to the new command but keep the stale compileProc, so callers are recompiled (compile epoch bumped) with the no-op again.
            • Prototype prototypes/3017879fff.patch (49 changed lines, test namespace-9.10). New helper ResetImportCompileProcs, called in both relink loops, copies the new command's compileProc into the import and recursively into imports of the import. namespace/proc/basic/compile/rename/oo*.test pass. Imports of a proc redefined as empty are then no longer inlined (only lost optimisation).
            • Suggested ticket comment Still present in 8.5–9.1: when a proc is redefined, its imports are relinked but keep the old compileProc (TclCompileNoOp of the empty proc). Proposed fix: refresh the compileProc of the imports (recursively) when relinking; test namespace-9.10.
            ✅3016135fff2010traceback using clock format with locale of he_ILpatchREPRODUCED (main, 9.0; regression since 8.7a1 from a6c8a2df48, 2017-04-13; 8.6/8.5 fine)

            Open ticket 3016135fff · 06. Time Measurement · filed 2010-06-14

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Fixed)

            • 🔧 2026-10-16 check-in e359a46a3b on clock-he-msg-quote, in main, 9.0: Fix 3016135fff again: escape the quote in the he.msg era names, which a6c8a2df48 turned into a bare quote, so that [clock format -locale he] failed with "ex
            • 🔧 2026-10-17 check-in f3d8c6e911 on trunk, in main: Fix 3016135fff: traceback using clock format with locale of he_IL
            • 🔧 2026-10-17 check-in 313539c373 on core-9-0-branch, in main, 9.0: Fix 3016135fff: traceback using clock format with locale of he_IL
            • 💬 2026-10-16 serhiy.storchaka: Regression in 9.0: a6c8a2df48 turned the escaped quote in the he.msg era names into a bare quote, so [clock format ... -locale he] fails again with "extra characters after close-quote". 8.6 is not affected. Proposed fix in branch clock-he-msg-quote, test clock-71.1.
            • 💬 2026-10-17 jan.nijtmans: +1 Please go ahead
            • 💬 2026-10-17 serhiy.storchaka: Merged into main in f3d8c6e911 and into 9.0 in 313539c373.
            Full verdict

            REPRODUCED (main, 9.0; regression since 8.7a1 from a6c8a2df48, 2017-04-13; 8.6/8.5 fine)

            • Platform generic
            • Tested main build, 9.0.5 ("extra characters after close-quote"); 8.6.18, 8.5.19 OK. A loop over all library/msgs/*.msg on main fails only for he.
            • Cause a6c8a2df48 "Translate all \u???? sequences to their UTF-8 counterpart in *.msg files" turned " in he.msg lines 47–48 (BCE/CE) back into a bare " inside a quoted word, which brought back the 2010 bug. The tags core-8-7-a1/a3/a5 and core-9-0-0 all have the broken line.
            • Prototype prototypes/3016135fff.patch (8 changed lines, test clock-71.2; against main, applies to 9.0)
            • Suggested ticket comment Regression in 8.7/9.0: a6c8a2df48 converted the " in he.msg BCE/CE to a bare quote. Proposed fix: escape it as \", test clock-71.2. 8.6 is not affected.
            ⬜3011585fff2010duplicate switch (non-bug)patchREPRODUCED (8.6, 8.5 only; 9.0/main format in C, the proc is gone)

            Open ticket 3011585fff · 06. Time Measurement · filed 2010-06-04

            🧪 Prototype check (2026-10-24)

            3011585fff.patch: VERIFIED

            • Base core-8-6-branch (tip 326e1bf923)
            • Before library/clock.tcl ParseClockFormatFormat2 still has the unreachable second % arm (line 1025, after the one at line 767 in the same switch)
            • Patch applied cleanly (3 changed lines)
            • After clock format 0 -format {%% %Y%%} -gmt 1 -> % 1970%; clock.test 8574/0 failed (8.6 reads library/ directly, no zipfs)
            • Test none possible (dead code removal, no behaviour change)
            Full verdict

            REPRODUCED (8.6, 8.5 only; 9.0/main format in C, the proc is gone)

            • Platform generic
            • Tested read library/clock.tcl in the 8.6 and 8.5 tips; main/9.0 have no ParseClockFormatFormat2
            • Cause 8.6 library/clock.tcl ParseClockFormatFormat2 has a second % case (line 1025) after the first one (line 767), and it can never be reached.
            • Prototype prototypes/3011585fff.patch (3 changed lines, against core-8-6-branch; clock.test passes, no new test since this is dead code)
            • Suggested ticket comment Only 8.6 is affected (9.0 formats in C). Proposed cleanup: remove the second, unreachable % case.
            ⬜2997967fff2010interp-36.7 timing sensitive?patchREPRODUCED (main, 8.6; test unchanged in all branches)

            Open ticket 2997967fff · 20. [interp] · filed 2010-05-07

            🧪 Prototype check (2026-10-24)

            2997967fff.patch: FIXED-MEANWHILE (0f5fdaa945, 2026-10-21; 9.0 ff7935f994, 8.6 6b2fff63e6). Ticket can be closed.

            • Base main (tip d6dd6d0f71)
            • Before does not reproduce: interp-36.7 was rewritten on the tip to vwait for the bgerror handler itself (with a 10 s safety timeout) instead of a fixed 10 ms done timer, in 0f5fdaa945 (main, 2026-10-21, "Fix 3b00b2198d: async test suite calls are missing cleanup"), also ff7935f994 (core-9-0-branch) and 6b2fff63e6 (core-8-6-branch)
            • Patch no longer applies (hunk FAILED, the lines it changes are gone); left unchanged
            • After n/a
            • Test n/a (the patch is itself a test fix)
            Full verdict

            REPRODUCED (main, 8.6; test unchanged in all branches)

            • Platform generic (tests)
            • Tested main, 8.6.18 with a script that blocks 30 ms after scheduling both timers (busy machine): result "untouched"; with the proposed change: "foo". interp.test passes with the change.
            • Cause tests/interp.test interp-36.7: if both timers are due in the same pass, the done timer fires before the idle callback that runs the background error handler, so [vwait] returns too early.
            • Prototype prototypes/2997967fff.patch (2 changed lines): after 10 [list after idle [list ::set ... {}]], so done is set only after the bgerror idle callback, which was queued first.
            • Suggested ticket comment Still timing sensitive: when both timers are due at once, done is set before the bgerror idle handler runs. Proposed fix: set done from an &#91;after idle&#93; scheduled by the 10 ms timer.
            ⬜2994388fff2010vwait's "would wait forever" lost in threaded buildsREPRODUCED (main, 9.0, 8.6; only threaded builds — all three are threaded, 8.5.19 here is unthreaded and reports the error)

            Open ticket 2994388fff · 02. Event Loops · filed 2010-04-29

            Full verdict

            REPRODUCED (main, 9.0, 8.6; only threaded builds — all three are threaded, 8.5.19 here is unthreaded and reports the error)

            • Platform unix
            • Tested after 10 {error boom}; vwait forever: 8.5.19 (unthreaded) returns "can't wait for variable "forever": would wait forever"; main, 9.0.5 and 8.6.18 hang (killed after 5 s).
            • Cause in threaded builds the notifier's Tcl_WaitForEvent blocks on a condition variable with no timeout (another thread could Tcl_ThreadAlert it), so Tcl_DoOneEvent never returns 0 and Tcl_VwaitObjCmd never reaches the "would wait forever" error.
            • Prototype none (needs a design: counting event sources/threads, as the ticket says; ferrieux 2011: "No hope to go back, but leaving it open for reference")
            • Suggested ticket comment Still the case in 8.6–9.1 threaded builds. Fixing it needs a way to know that no other thread can wake the notifier, so this is a design question, not a small fix.
            ⬜2990516fff2010TM default paths do not search custom --libdirpatchREPRODUCED (main, 9.0 by code reading; not 8.6 in a standard layout)

            Open ticket 2990516fff · 39. Package Manager · filed 2010-04-21

            🧪 Prototype check (2026-10-24)

            2990516fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before out-of-tree build of the tip with --prefix=$V/inst2990 --libdir=$V/inst2990/lib64, make install, foo2990-1.0.tm in lib64/tcl9/site-tcl: tcl::pkgconfig get libdir,runtime -> .../lib64, but tcl::tm::path list has only .../lib/tcl9/{site-tcl,9.0,9.1} and //zipfs:/lib/tcl/tcl9/...; package require foo2990 -> can't find package foo2990
            • Patch applied cleanly (23 changed lines)
            • After rebuilt zip and reinstalled: path list now also has .../lib64/tcl9/{site-tcl,9.0,9.1}; package require foo2990 -> 1.0. Tests (TCL_LIBRARY=wt/library): tm.test 22/0/0, package.test 402 passed/1 skipped/0 failed, safe.test 153/2/0, safe-zipfs.test 22/0/0, safe-stock.test 11/0/0, interp.test 384/7/0
            • Test existing tm-3.14 fails before (with tm.tcl reverted) / passes after
            Full verdict

            REPRODUCED (main, 9.0 by code reading; not 8.6 in a standard layout)

            • Platform generic (unix install layout)
            • Tested main built with --libdir=<prefix>/lib64 and installed: [tcl::tm::path list] contains only <prefix>/lib/tcl9/... and //zipfs:/lib/tcl/tcl9/..., while make install puts modules into $(libdir)/tcl9 (MODULE_INSTALL_DIR) and [tcl::pkgconfig get libdir,runtime] correctly says lib64; a foo-1.0.tm in lib64/tcl9/site-tcl is not found. 8.6: [info library] is $libdir/tcl8.6 so the first root already equals libdir; in 9.x [info library] is in zipfs, so only the nameofexecutable guess remains (hard-coded "lib").
            • Cause library/tm.tcl ::tcl::tm::Defaults roots only [file dirname [info library]] and <exe>/../lib.
            • Prototype prototypes/2990516fff.patch (23 changed lines, test tm-3.14): also root at [tcl::pkgconfig get libdir,runtime]. tm.test, package.test, safe*.test pass; the lib64 module is then found.
            • Suggested ticket comment Still present in 9.x: with the library in zipfs, [info library] no longer points into $libdir, so a --libdir=.../lib64 install never searches $libdir/tcl9. Proposed fix: add [tcl::pkgconfig get libdir,runtime] to the default TM roots (test tm-3.14).
            🔧2982713fff2010I cannot run exec command on Linux using Tcl/TkpatchREPRODUCED (main, 9.0, 8.6, 8.4) when tclsh is started with SIGCHLD set to SIG_IGN by its parent

            Open ticket 2982713fff · 69. Other · filed 2010-04-06

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-17 check-in b3c28a91a0 on tclsh-inherited-sigchld, not merged: Fix 2982713fff: if tclsh was started with SIGCHLD ignored (e.g. from system() in a program which ignores it), [exec] failed with "child process lost". Tcl_Mai
            • 💬 2026-10-17 serhiy.storchaka: Reproduced when tclsh is started with SIGCHLD ignored, as by system() from a program which ignores it, or env --ignore-signal=CHLD tclsh: the kernel reaps the children itself and waitpid() fails with ECHILD. Proposed fix in branch tclsh-inherited-sigchld: Tcl_MainEx() restores the default SIGCHLD handling if it is ignored. Applications which embed Tcl keep their own setting, as Tk now keeps the signal handlers of …
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4) when tclsh is started with SIGCHLD set to SIG_IGN by its parent

            • Platform unix
            • Tested env --ignore-signal=CHLD tclsh x.tcl and perl -e '$SIG{CHLD}="IGNORE"; exec @ARGV' tclsh x.tcl. exec echo yo fails with "child process lost (is SIGCHLD ignored or trapped?)", and exit codes are lost. Note: dash's trap '' CHLD (as in ferrieux's example) does not actually set SIG_IGN, but bash's does.
            • Cause with SIGCHLD set to SIG_IGN, the kernel reaps children itself and waitpid() in Tcl_WaitPid fails with ECHILD. Tcl already resets SIGCHLD to the default in its own children (RestoreSignals in unix/tclUnixPipe.c) and already sets SIGPIPE itself in TclpInitPlatform, but it never fixes an inherited SIG_IGN on SIGCHLD.
            • Prototype prototypes/2982713fff.patch (28 changed lines, test exec-23.1): TclpInitPlatform (unix/tclUnixInit.c) sets SIGCHLD back to SIG_DFL only if it is SIG_IGN at startup. Test exec-23.1 runs tclsh under bash -c "trap '' CHLD; exec ..." and checks that CHILDSTATUS ... 3 comes back. It fails without the C change. exec, process, pid, unixInit, event and io tests pass (1180 total, 0 failed). Trade-off: an embedding application that ignores SIGCHLD on purpose before initializing Tcl gets it reset. It can ignore it again afterwards, and [exec] cannot work in such a process anyway.
            • Suggested ticket comment Reproducible when the parent starts tclsh with SIGCHLD ignored (e.g. env --ignore-signal=CHLD tclsh): waitpid() then fails with ECHILD. Proposed fix: TclpInitPlatform resets an inherited SIG_IGN on SIGCHLD to SIG_DFL, as it already does for its children in RestoreSignals(). Test exec-23.1.
            ⬜2980143fff2010support return optionspatchREPRODUCED (Thread 3.0.7 on main and 9.0.5, Thread 2.8.13 on 8.6.18)

            Open ticket 2980143fff · 80. Thread Package · filed 2010-03-31

            🧪 Prototype check (2026-10-24)

            2980143fff.patch: PATCH-INCOMPLETE (fixes only the errorcode double-wrapping, which it does correctly; the ticket's actual request, transporting the full return-options dict, is not done, as the triage says: feature, beyond the size limit)

            • Base Thread trunk (tip 39d9f08aa0, 3.0.7), against own main build
            • Before tclsh r298.tcl (catch {thread::send $t {error ERR INFO {MY CODE}}} m o) -> errorCode=<{MY CODE}> -errorcode=<{MY CODE}> (list errorcode wrapped once more)
            • Patch applied cleanly
            • After errorCode=<MY CODE> -errorcode=<MY CODE>; Thread tests/all.tcl 144 total, 0 failed
            • Test existing thread-11.13 fails before / passes after
            Full verdict

            REPRODUCED (Thread 3.0.7 on main and 9.0.5, Thread 2.8.13 on 8.6.18)

            • Platform generic (Thread package)
            • Tested catch {thread::send $t {return -code error -errorcode {MY CODE} -foo bar msg}} m o: only -errorcode/-errorinfo are carried, custom options are lost; return -code break / -code 7 in the target come back as plain code 2 with -code 0 -level 1. Additionally a real bug: a list errorcode is wrapped once more: -errorcode {MY CODE} arrives as {{MY CODE}} (sync send; same on 8.6/9.0/main).
            • Cause generic/threadCmd.c ThreadSetResult only copies errorCode/errorInfo strings; ThreadSend applies the errorcode with Tcl_SetErrorCode(interp, str, NULL), which treats the whole string as one list element.
            • Prototype prototypes/2980143fff.patch (Thread trunk; 10 changed lines, test thread-11.13) — fixes only the errorcode double-wrapping (Tcl_SetObjErrorCode). Full return-options transport (new result field, Tcl_GetReturnOptions/Tcl_SetReturnOptions in the sync and both async callback paths, behaviour change for break/continue/return) not prototyped: feature, beyond the size limit.
            • Suggested ticket comment Still only -errorcode/-errorinfo are transported, and a multi-word -errorcode comes back wrapped in an extra list level (&#91;thread::send $t {error E I {MY CODE}}&#93; gives errorCode {MY CODE} instead of MY CODE). Proposed fix for the latter: use Tcl_SetObjErrorCode in ThreadSend (test thread-11.13); full return options support remains open.
            ⬜2978770fff2010invokehidden -namespace|global: no [info level]patchREPRODUCED (main, 9.0, 8.6)

            Open ticket 2978770fff · 20. [interp] · filed 2010-03-29

            🧪 Prototype check (2026-10-24)

            2978770fff.patch: VERIFIED (the -errorstack side effect CALL err CALL err stays as the triage noted)

            • Base main (tip d6dd6d0f71)
            • Before interp-20.51 FAILED: {1 lv lv} {2 {} {lv x}} {2 {} lv} 1 {INNER {returnImm boom {}} CALL err UP 1} (info level 1 empty with -namespace/-global)
            • Patch applied cleanly (25 changed lines)
            • After {1 lv lv} {2 {lv x} {lv x}} {2 lv lv} 1 {INNER {returnImm boom {}} CALL err CALL err}; tests interp 385/0 failed, info 285/0, error 318/0, safe 153/0, namespace 312/0
            • Test existing interp-20.51 fails before / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested reporter's script on main (9.1.1 build), 9.0.5, 8.6.18: with -namespace/-global, info level is 2 and info level 1 is empty.
            • Cause tclInterp.c TclObjInvokeNamespace pushes a namespace CallFrame (TclPushStackFrame) but leaves objc/objv at 0/NULL, while [namespace eval] fills them in.
            • Prototype prototypes/2978770fff.patch (25 changed lines, test interp-20.51). Stores the hidden command's words in the frame, so info level 1 shows the hidden command (hideme). Side effect: -errorstack then contains that frame too (CALL err CALL err; [namespace eval] frames also show up in -errorstack). interp/info/error/safe.test pass. Showing the full interp invokehidden ... words instead would need the caller's objv passed down (signature change of the internal function).
            • Suggested ticket comment Still present: the namespace frame pushed by invokehidden -namespace/-global has no objv. Proposed fix: fill it with the hidden command's words, as &#91;namespace eval&#93; does with its own; test interp-20.51.
            ⬜2960862fff2010loaded library paths should be normalizedpatchREPRODUCED (main, 9.0, 8.6)

            Open ticket 2960862fff · 40. Dynamic Loading · filed 2010-02-28

            🧪 Prototype check (2026-10-24)

            2960862fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before work-a4/v4load.tcl (cd dltest; load ./tcl9pkga.so Pkga c; cd ..; load dltest/tcl9pkga.so Pkga c; load $abs/./tcl9pkga.so Pkga c) -> Pkga entries: 3
            • Patch applied cleanly (59 changed lines). Checked that every exit after the new Tcl_DStringInit(&fileName) in Tcl_LoadObjCmd/Tcl_UnloadObjCmd goes through done (the remaining early returns precede the init).
            • After same script -> Pkga entries: 1; make test TESTFLAGS="-file 'load.test unload.test safe.test interp.test'" -> 604 total, 595 passed, 9 skipped, 0 failed
            • Test existing load-12.1 fails before (unpatched tclLoad.c: load.test 30 passed, 1 failed) / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested Linux main build, 9.0 and 8.6 tip tcltest with dltest/pkga: cd dltest; load ./pkga.so; cd ..; load dltest/pkga.so; load $abs/dltest/./pkga.so gives three info loaded entries, and Pkga_Init runs three times.
            • Cause generic/tclLoad.c:Tcl_LoadObjCmd/Tcl_UnloadObjCmd match libraries with strcmp on the file name as typed and store that string, so relative names also go stale after cd.
            • Prototype prototypes/2960862fff.patch (59 changed lines, test load-12.1). New helper NormalizedFileName() uses the normalized path (in a DString) when the file exists, and the given name otherwise (search-path loads). Used by load and unload. One early return in unload became goto done to free the DString. load.test, unload.test, safe*.test, interp.test pass. Visible change: info loaded now shows absolute paths for files loaded by relative name.
            • Suggested ticket comment Still present: load/unload compare library names as given. Proposed fix: record the normalized path when the file exists, so the same library under another path is found as already loaded; test load-12.1. Note that info loaded then reports normalized paths.
            ⬜2935779fff2010Incorrect -errorline from compiled [catch]REPRODUCED (main, 9.0, 8.6, 8.5)

            Open ticket 2935779fff · 47. Bytecode Compiler · filed 2010-01-20

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5)

            • Platform generic
            • Tested main (9.1.1 build), 9.0.5, 8.6.18, 8.5.19: same error at line 3 of the catch body gives -errorline 4 (line in the proc body) with compiled [catch] and compiled [try], 3 with set c catch; $c {...} (not compiled).
            • Cause tclExecute.c/tclNamesp.c TclLogCommandInfo counts lines from codePtr->source, the start of the whole compiled script (proc body), because a compiled catch body is inlined in the enclosing bytecode.
            • Prototype none. Fixing it means recording the catch/try body's source start in the ExceptionRange at compile time and using it as the base when logging. That changes visible -errorline values of compiled code, and the reporter prefers the current proc-relative numbers (comment 2010-01-21), so it needs a decision first.
            • Suggested ticket comment Still present in 8.5–9.1: compiled &#91;catch&#93;/&#91;try&#93; report -errorline relative to the enclosing proc body, uncompiled ones relative to the body script (as documented). A fix would record the body start in the catch ExceptionRange; first decide whether the documented behaviour is still wanted, since this changes results for compiled code.
            ⬜2927851fff2010Tcl dies when STDIN is closedpatchREPRODUCED (main, 9.0, 8.6)

            Open ticket 2927851fff · 50. Embedding Support · filed 2010-01-07

            🧪 Prototype check (2026-10-24)

            2927851fff.patch: VERIFIED

            • Base main (tip c092dae4af)
            • Before embed91.exe (work-win1/emb/embed2.c built against tcl91: sets STD_INPUT_HANDLE to a file, creates the interp, CloseHandle()s it, then [open]s files) -> BUG REPRODUCED: [open] returned the stdin channel (iteration 1)
            • Patch applied cleanly, 21 changed lines, not changed
            • After NOT-REPRODUCED: 200 opens returned new channels; io, ioCmd, chanio, fCmd, winPipe, winFCmd, winConsole: 2684 total, 3 failed (winFCmd-6.1, 6.13, 9.3, as baseline)
            • Test none possible (needs a host that closes an OS std handle behind Tcl's back)
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform Windows
            • Tested code reading of main; Unix check (tclsh x.tcl <&- on main and 8.6.18: [open] never returns a std channel; Unix Tcl_MakeFileChannel does not look up existing channels by fd).
            • Cause win/tclWinChan.c OpenFileChannel returns an existing channel when a FileInfo with the same HANDLE exists. If the host closes the OS stdin handle after Tcl made the stdin channel on it, CreateFile later reuses the handle value and [open]/Tcl_FSEvalFileEx get the stdin channel back; closing it gives "called Tcl_Close on channel with refCount > 0". Unchanged since 8.5; Tcl_MakeFileChannel's DuplicateHandle check only rejects handles that are already invalid when the channel is created.
            • Prototype none (needs a Windows reproduction first; a likely fix is to drop or validate the handle-reuse lookup in OpenFileChannel for handles from CreateFile)
            • Repro win/2927851.tcl (runs win/2927851/embed.exe, prebuilt with mingw from embed.c, or rebuilds it with cl). The embedder loads the tclsh's tclNN.dll, sets STD_INPUT_HANDLE to a file, creates an interp and the stdin channel, CloseHandle()s it, then checks whether [open] returns "stdin"; prints VERDICT: REPRODUCED/NOT-REPRODUCED.
            • Suggested ticket comment (after the Windows run) On Windows OpenFileChannel reuses an existing channel with the same HANDLE, so a std handle closed behind Tcl's back is later matched with an unrelated CreateFile handle, and [open] returns stdin.
            Windows VM, run 1
            • Platform Windows
            • Verdict REPRODUCED (main, 9.0, 8.6)
            • Tested embedding program (work-win1/emb/embed2.c, built against each version's tcl.h/import lib; the prebuilt win/2927851/embed.exe works only with 8.6, Tcl 9 does not export Tcl_FindExecutable/Tcl_GetStringResult): it sets STD_INPUT_HANDLE to a file, creates an interp and the stdin channel, CloseHandle()s it; the next [open] of a file returns "stdin" (iteration 1) on 9.1b1, 9.0.5 and 8.6.18. [source] then reads through that channel. No panic seen with this sequence.
            • Cause win/tclWinChan.c OpenFileChannel returns an existing channel whose FileInfo has the same HANDLE value; CreateFile reuses the closed stdin handle value
            • Prototype prototypes/2927851fff.patch (main, 21 changed lines, no test: no script-level way to close an OS handle behind Tcl's back): OpenFileChannel gets a reuse flag; TclpOpenFileChannel (fresh CreateFile handle) passes 0, Tcl_MakeFileChannel keeps 1. With it the embed repro gives 200 fresh channels; io, ioCmd, chanio, fCmd, winPipe tests pass; winFCmd has the same 3 failures as the unpatched build (6.1, 6.13, 9.3, VM environment).
            • Suggested ticket comment Still reproducible with 8.6, 9.0 and 9.1 by closing the std handle in the host: OpenFileChannel (win/tclWinChan.c) matches the recycled HANDLE value and &#91;open&#93; returns stdin. Proposed fix: skip the lookup of existing channels for handles just returned by CreateFile.
            ⬜2889293fff2009[info exists array(elem)] deletes array searchespatchREPRODUCED (main, 9.0, 8.6, 8.5)

            Open ticket 2889293fff · 07. Variables · filed 2009-10-30

            🧪 Prototype check (2026-10-24)

            2889293fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before tclsh r288.tcl (the ticket's script) -> couldn't find search "s-1-tcl_platform" from array anymore
            • Patch applied cleanly (64 changed lines, of them ~12 code, the rest tests info-42.1..42.4)
            • After same -> result: user; tests info (292/0 failed), var, set-old, trace, upvar, env, set, incr, append, execute, compile: 0 failed
            • Test existing info-42.1, 42.2, 42.4 fail before / pass after; 42.3 (read trace may still create the element) passes both ways, as intended
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5)

            • Platform generic
            • Tested main (own build), 9.0.5, 8.6.18, 8.5.19. Fails at global level and with non-constant names in procs (INST_EXIST_STK); a proc with a constant local array name already works (INST_EXIST_ARRAY has a no-trace shortcut).
            • Cause generic/tclTrace.c:TclVarTraceExists and the INST_EXIST_STK path in generic/tclExecute.c look up with createPart2=1, so a missing element is created in the hash table (TclLookupArrayElement calls DeleteSearches on a new entry) and then cleaned up again. Creating it is only needed when the array has a read trace that could define the element.
            • Prototype prototypes/2889293fff.patch (55 changed lines, about 12 of them code; tests info-42.1..42.3). Look up with createPart2=0 first, and retry with createPart2=1 only when the array is read-traced. info-42.1/42.2 fail on 9.0 without the patch; info, var, set-old, trace, upvar, env, set, incr, append tests pass. (pooryorick's 2013 alternative at ynform.org is not in tcl.fossil.)
            • Suggested ticket comment Still present in 9.1. &#91;info exists a(x)&#93; makes a temporary element for a missing key, and that kills any active search. Proposed fix: create it only when the array has a read trace (patch + tests info-42.*).
            ⬜2859027fff2009Bytecode Compiler & tcl_precisionpatchREPRODUCED (8.6; 8.5 by report). Not applicable to 9.0/main: tcl_precision was removed by TIP 488 (f4efe03192, 2017-12-01).

            Open ticket 2859027fff · 47. Bytecode Compiler · filed 2009-09-15

            🧪 Prototype check (2026-10-24)

            2859027fff.patch: VERIFIED

            • Base core-8-6-branch (tip 326e1bf923)
            • Before proc p {} {expr {1.1+2.2}}; puts [p]; set tcl_precision 3; puts [p] -> 3.3000000000000003 twice; util-11.24 FAILED (3.3000000000000003 3.3000000000000003)
            • Patch applied cleanly (23 changed lines)
            • After repro prints 3.3000000000000003 then 3.3; tests util 2289/0 failed, expr 14163/0, compile 160/0, execute 154/0, basic 118/0
            • Test existing util-11.24 fails before / passes after
            Full verdict

            REPRODUCED (8.6; 8.5 by report). Not applicable to 9.0/main: tcl_precision was removed by TIP 488 (f4efe03192, 2017-12-01).

            • Platform generic
            • Tested 8.6.18 and own core-8-6-branch build: proc p {} {expr {1.1+2.2}}; p; set tcl_precision 3; p still gives 3.3000000000000003, while a non-folded expr {$x+2.2} gives 3.3. main and 9.0.5 always give 3.3000000000000003 (tcl_precision has no effect there).
            • Cause constant expressions are folded at compile time (TclCompileExpr). In generic/tclUtil.c:TclPrecTraceProc a change of tcl_precision does not invalidate the existing bytecode, so the folded literal keeps the string from the old precision.
            • Prototype prototypes/2859027fff.patch (core-8-6-branch, 23 changed lines, test util-11.24). Bump iPtr->compileEpoch when the precision really changes; the reporter's recursive script then prints the right value at every level. util, expr, compile, execute and basic tests pass. Limit: precision is per thread, and only the interp that writes the variable gets its epoch bumped.
            • Suggested ticket comment Only 8.6 is affected (TIP 488 removed tcl_precision in 9.0). Proposed fix for 8.6: bump the compile epoch when tcl_precision changes, so folded constant expressions are recompiled (patch + test util-11.24). Alternatively close as Won't Fix, since tcl_precision is deprecated.
            ⬜2837916fff2009clock scan 01/2010 -format %m/%Y returns wrong resultREPRODUCED (8.6, 8.5, 8.7b1: year/month ignored, the base date is used; main, 9.0: behaviour changed, still not the expected 1 January)

            Open ticket 2837916fff · 06. Time Measurement · filed 2009-08-14

            Full verdict

            REPRODUCED (8.6, 8.5, 8.7b1: year/month ignored, the base date is used; main, 9.0: behaviour changed, still not the expected 1 January)

            • Platform generic
            • Tested main build, 9.0.5, 8.7b1, 8.6.18, 8.5.19 with -base 2026-10-16: 01/2010 %m/%Y gives 2026-10-16 on 8.x and 2010-01-16 on 9.x. With -base on the 31st, 9.x gives 02/2010 -> 2010-03-03, even with -validate 1.
            • Cause in 8.6 clock.tcl, rule [7] of the doc ("if none of the above rules results in a usable date, the date of the base time is used") drops a lone year+month. The 9.0 C engine (generic/tclClockFmt.c ClockScan, CLF_DATE now includes CLF_MONTH/CLF_YEAR) uses the year and month it found and takes the missing day of month from -base. The day is never checked against the month, so it overflows into the next month, and doc/clock.n was not updated.
            • Prototype none (choosing day 1 vs. the base day vs. an error is a behaviour decision; the 9.x overflow needs a decision too)
            • Suggested ticket comment 8.6 still ignores year+month without a day. 9.0 now uses them but takes the day from -base, so clock scan 02/2010 -format %m/%Y on the 31st gives 3 March 2010, even with -validate. Proposal: default the missing day to 1 when a month is given, and document it.
            ⬜2837204fff2009glob on realtive patterns and dirs is thread "unsafe"REPRODUCED (main, 9.0, 8.6; by code reading, not run)

            Open ticket 2837204fff · 37. File System · filed 2009-08-13

            Full verdict

            REPRODUCED (main, 9.0, 8.6; by code reading, not run)

            • Platform generic
            • Tested code reading of generic/tclFileName.c:DoGlob/TclGlob and generic/tclIOUtil.c:Tcl_FSMatchInDirectory. No threaded race run; it needs another thread to cd while a recursive glob is running.
            • Cause DoGlob calls Tcl_FSMatchInDirectory once per directory level with relative path objects. Each call resolves them (normalization, or Tcl_FSGetCwd for an empty prefix) against the process-wide cwd at that moment, so one [glob] can mix results from two cwds.
            • Prototype none (needs a redesign: resolve cwd once in TclGlob, glob absolute paths and strip the prefix from the results, as Tcl_FSMatchInDirectory already does for the cwd case)
            • Suggested ticket comment Still the case: each level of a relative [glob] resolves the cwd again. A fix would get the cwd once in TclGlob and search with absolute paths, stripping the prefix from the results. Low priority.
            ⬜2827015fff20092>@stdout does 2>@1patchREPRODUCED (main, 9.0, 8.6, 8.4)

            Open ticket 2827015fff · 27. Channel Types · filed 2009-07-25

            🧪 Prototype check (2026-10-24)

            2827015fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before work-a4/v4exec.tcl (middle tclsh runs exec tclsh << {puts stdout OUT; flush stdout; puts stderr ERR} >@stderr 2>@stdout, its stdout captured, its stderr to a file) -> middle stdout (captured): {} middle stderr: {OUT\nERR} (both streams went to stderr). Same with the fork path forced (use_spawn = 0, local experiment only).
            • Patch ported: TclpCreateProcess now creates the error pipe via local pipeIn/pipeOut, so the dup block (hunk 2) is inserted by hand after errPipeOut = pipeOut;; code unchanged otherwise. 47 changed lines.
            • After -> middle stdout (captured): ERR middle stderr: OUT, both on the posix_spawn path (glibc default) and the fork path (use_spawn forced to 0 temporarily); make test TESTFLAGS="-file 'exec.test io.test ioCmd.test pid.test process.test'" -> 1484 total, 1260 passed, 224 skipped, 0 failed; exec.test with fork path forced: 99 passed, 0 failed
            • Test existing exec-15.8 fails before (exec.test unpatched: 98 passed, 1 failed) / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4)

            • Platform unix
            • Tested reporter's nested >@stderr 2>@stdout script on main (own build), 9.0.5, 8.6.18, 8.4.20: prints ---- then OUT1/ERR1/OUT2/ERR2 (both inner streams to the middle stderr)
            • Cause unix/tclUnixPipe.c:TclpCreateProcess sets up the child's fds 0, 1, 2 in order (SetupStdFile in the fork path, posix_spawn_file_actions_adddup2 in the spawn path, which is what glibc >= 2.24 uses), so an errorFile that is fd 1 (or 0), or an outputFile that is fd 0, has already been overwritten by the time it is dup'ed. The parent now dups such fds to >= 3 (F_DUPFD, close-on-exec) before spawn/fork and closes the copies afterwards; covers both paths. (The 2009 dup-2.patch only patched the fork child.)
            • Prototype prototypes/2827015fff.patch (47 changed lines, test exec-15.8); exec.test, io.test, ioCmd.test, pid.test, process.test pass
            • Suggested ticket comment TclpCreateProcess sets up the child's stdin/stdout/stderr in sequence, so with 2>@stdout fd 1 has already been replaced by the exec pipe when stderr is dup'ed from it. Proposed fix: dup such fds above 2 in the parent first (fork and posix_spawn paths), test exec-15.8. Potential incompatibility: scripts that (mis)used 2>@stdout as 2>@1 will change behaviour.
            ⬜2826430fff2009Channel name recycling is dangerousREPRODUCED (main, 9.0, 8.6) — feature request; TIP 355 (this ticket's TIP) is still Draft, Tcl-Version 9.1

            Open ticket 2826430fff · 24. Channel Commands · filed 2009-07-24

            Full verdict

            REPRODUCED (main, 9.0, 8.6) — feature request; TIP 355 (this ticket's TIP) is still Draft, Tcl-Version 9.1

            • Platform unix (generic request)
            • Tested open/close a file and a server socket 3 times: main/9.0 give file5 file5 file5 (8.6: file3 x3); sock names are pointer-based and repeat as well (sock5d5e33caadd0 x3 on 9.0)
            • Cause unix/tclUnixChan.c names file channels after the fd (file%d), sockets after the pointer (sock%p), so names are reused as soon as the fd/memory is.
            • Prototype none (needs TIP 355 to be finished and voted; the uniq.patch attached here is the starting point)
            • Suggested ticket comment Still the case in 9.1: file channel names follow the fd and socket names the (reused) pointer. This needs TIP 355 (still Draft) to be completed and voted.
            🔧2811381fff2009unexpected 'result' in trace execution leavepatchREPRODUCED (main, 9.0, 8.6, 8.5); regression since 8.5.0 (8.4.2-8.4.20 pass the original result to every trace)

            Open ticket 2811381fff · 46. Traces · filed 2009-06-24

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-17 check-in 6fd61eb7c4 on leave-trace-own-result, not merged: Fix 2811381fff: since 8.5 each leave or leavestep trace got the result of the previous trace instead of that of the traced command. Restore the saved interpre
            • 💬 2026-10-17 serhiy.storchaka: 8.4 passed the result of the traced command to every leave trace; since 8.5 TclCheckExecutionTraces() and TclCheckInterpTraces() restore the saved result only after the last trace, so each trace gets the result of the previous one. The documentation describes the 8.4 behaviour. Proposed fix in branch leave-trace-own-result: restore the result after each trace. Test trace-40.2.
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5); regression since 8.5.0 (8.4.2-8.4.20 pass the original result to every trace)

            • Platform generic
            • Tested main (own build), 9.0.5, 8.6.18, 8.5.0, 8.5.1, 8.5.19 show second: first; 8.4.2, 8.4.11, 8.4.15, 8.4.20 show second: hello. leavestep traces behave the same (s2: s1).
            • Cause generic/tclTrace.c:TclCheckExecutionTraces and TclCheckInterpTraces save the interp state once, before the first trace, and restore it only after the last one. So trace N+1 gets trace N's result in its result argument. The final command result is still restored, which is why traced returns hello.
            • Prototype prototypes/2811381fff.patch (33 changed lines, test trace-40.2). After each successful trace, restore the saved state, so the next trace sees the command's own result (the 8.4 behaviour, and what trace.n says). trace, basic, proc, var, info and oo tests pass; trace-40.2 fails on 9.0 without the patch. The alternative is to document the chaining in trace.n, as the reporter asked.
            • Suggested ticket comment Since 8.5.0 each leave/leavestep trace gets the previous trace's result instead of the command's (8.4 passed the original). Proposed fix: restore the saved interp state after each trace (patch + test trace-40.2).
            ⬜2762041fff2009Fix (too) early recognition of EOF in a channel stackREPRODUCED (main, 9.0, 8.6) for the generic problem; the original zlib/http symptom is gone (zbug.tcl gives 5000/5000 on all three, also with `zlib…

            Open ticket 2762041fff · 26. Channel Transforms · filed 2009-04-14

            Full verdict

            REPRODUCED (main, 9.0, 8.6) for the generic problem; the original zlib/http symptom is gone (zbug.tcl gives 5000/5000 on all three, also with zlib push ... -limit 4096)

            • Platform generic
            • Tested main (own build), 9.0.5, 8.6.18: attached zbug.tcl; plus work-a4/2762041-refchan.tcl — a [chan push] transform that buffers until EOF and returns 5000 bytes from drain: the first read $pr (read-to-EOF) returns only 4096 bytes, [eof] is already 1, and a second read returns the missing 904
            • Cause As Andreas analysed in 2010: CHANNEL_EOF lives in the shared ChannelState, so when the transform's Tcl_ReadRaw on the base channel hits EOF, generic/tclIO.c:DoReadChars (and GetInput callers) treats the whole stack as at EOF and stops, though the top transform still holds data (tclIORTrans.c keeps the rest of the drain result in its result buffer). The zlib channel avoids it by its read-ahead limit / own buffering, not by a generic fix.
            • Prototype none (needs a per-channel EOF flag or a "transform has pending data" check in the generic read loops; a design change beyond the size limit)
            • Suggested ticket comment The zlib/http symptom no longer reproduces, but the generic problem remains on 9.1/9.0/8.6: with a reflected transform whose drain returns more than one buffer, read-to-EOF returns early and [eof] is already true while 904 bytes are still in the transform; a second [read] returns them. Keep open as the reminder for a per-channel EOF.
            🔧2577420fff2009Tcl_GetInterpPath produces invalid path.patchREPRODUCED (main with ASan; 8.5 shows garbage paths; 8.6/9.0 have the same code, and the freed memory just still holds the old text). Heap use-afte…

            Open ticket 2577420fff · 20. [interp] · filed 2009-02-07

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-17 check-in 49e00a9cd6 on interp-delete-callbacks-path, not merged: Fix 2577420fff: Tcl_GetInterpPath() and Tcl_GetParent() used freed memory in a Tcl_CallWhenDeleted() callback of a child interpreter. Unlink the child from it
            • 💬 2026-10-17 serhiy.storchaka: Still a use-after-free in 9.0 and 9.1 (ASan): the child's entry in the parent's child table is removed before the child's Tcl_CallWhenDeleted() callbacks run, and the InterpInfo may be freed before them too, because it is freed by another delete callback. 8.6 and 9.0 print the right path only because the freed memory still holds it. Proposed fix in branch interp-delete-callbacks-path: unlink after Tcl_DeleteInterp…
            Full verdict

            REPRODUCED (main with ASan; 8.5 shows garbage paths; 8.6/9.0 have the same code, and the freed memory just still holds the old text). Heap use-after-free.

            • Platform generic
            • Tested C extension work-a1/gip/gip.c (Tcl_CallWhenDeleted callback that calls Tcl_GetInterpPath(Tcl_GetParent(i), i)) with script work-a1/t10.tcl. 8.5.0, 8.5.6, 8.5.19 print garbage. 8.6.0, 8.6.18, 9.0.5 and main print the right paths, but main with ASan+PURIFY reports a heap-use-after-free in Tcl_GetInterpPath (freed hash entry). The rename child {} path also has a use-after-free in Tcl_GetParent (InterpInfo already freed).
            • Cause generic/tclInterp.c:ChildObjCmdDeleteProc deletes the child's entry in the parent's childTable before Tcl_DeleteInterp runs the child's delete callbacks. Also, InterpInfoDeleteProc is registered with Tcl_CallWhenDeleted, so depending on hash order it can free the InterpInfo before user callbacks run.
            • Prototype prototypes/2577420fff.patch (34 changed lines, no test: there is no script-level way to register a delete callback that calls Tcl_GetInterpPath; checked with the ASan repro). Delete the hash entry after Tcl_DeleteInterp. Replace the InterpInfoDeleteProc callback with TclInterpInfoDelete, which DeleteInterpProc calls after all assocData callbacks. ASan clean on all three deletion paths. Full suite passes (521526 tests, 0 failures, without http*/zipfs), and interp, safe and basic tests pass under ASan.
            • Suggested ticket comment Still a use-after-free on 9.1 (ASan): the child's childTable entry is removed, and the InterpInfo may be freed, before the child's Tcl_CallWhenDeleted callbacks run. Proposed fix: unlink after Tcl_DeleteInterp, and free InterpInfo after all delete callbacks.
            ⬜2549621fff2009[interp create] fails to normalize pathspatchREPRODUCED (main, 9.0, 8.6)

            Open ticket 2549621fff · 20. [interp] · filed 2009-01-30

            🧪 Prototype check (2026-10-24)

            2549621fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before interp create a; interp create { a }; interp create {}; interp children -> both creates succeed, children {{} { a } a}
            • Patch applied cleanly (28 changed lines, unchanged; also covers 2549598fff)
            • After same script -> interpreter named "a" already exists, cannot create and interpreter named "" already exists, cannot create, children a; tests interp, safe, safe-zipfs, oo: 959 total, 934 passed, 25 skipped, 0 failed
            • Test existing interp-2.14 and interp-2.15 fail before (C fix reverted; the leftover children then also break interp-3.1/3.7 and the file aborts) / pass after
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested main (own build), 9.0.5, 8.6.18: interp create a; interp create { a } creates a second child literally named " a ", which other subcommands can't reach (they parse the path as a list and find "a").
            • Cause generic/tclInterp.c:ChildCreate: for a 1-element list it uses TclGetString(pathPtr), the whole string, instead of objv[0]. Every other place (GetInterp) uses the list elements.
            • Prototype prototypes/2549621fff.patch (28 changed lines, tests interp-2.14, interp-2.15; this patch also covers 2549598fff). Use objv[0] when objc == 1, and reject objc == 0. Full suite passes (521547 tests, 0 failures, without http*/zipfs); the new tests fail on 9.0.
            • Suggested ticket comment Still present. ChildCreate uses the whole string of a one-element path instead of its element. Proposed fix: use the list element, as GetInterp does (patch + test interp-2.14). Incompatibility: interp create {{x y}} now makes a child named "x y", reachable as {{x y}}, consistent with the other subcommands.
            ⬜2549598fff2009[interp create] accepts empty pathpatchREPRODUCED (main, 9.0, 8.6)

            Open ticket 2549598fff · 20. [interp] · filed 2009-01-30

            🧪 Prototype check (2026-10-24)

            2549598fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before tcltest 2549.tcl -> interp create {} returns the empty string and succeeds, interp children -> {}, interp exists {} -> 1; also interp create { a } creates a child named a (children {} { a }), the companion bug 2549621
            • Patch applied cleanly (28 changed lines; also covers 2549621)
            • After interp create {} -> error interpreter named "" already exists, cannot create; tests interp.test 386 passed/7 skipped/0 failed, safe.test 153/2/0, cmdMZ.test 96/1/0
            • Test existing interp-2.15 (and interp-2.14 for 2549621) fail before / pass after
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested main, 9.0.5, 8.6.18: interp create {} succeeds, interp children gives {{}}, the child is reachable only as {{}}, and interp exists {} still means the current interp.
            • Cause generic/tclInterp.c:ChildCreate accepts an empty list (objc == 0) and uses the empty string as the child name.
            • Prototype prototypes/2549598fff.patch (the same patch as 2549621fff, 28 changed lines, test interp-2.15). An empty path now fails with interpreter named "" already exists, cannot create.
            • Suggested ticket comment Still present. Proposed fix (together with 2549621fff): reject the empty path in &#91;interp create&#93;, because it names the current interpreter (test interp-2.15). A child named by the empty string can still be created explicitly with interp create {{}}.
            ⬜2449154fff2008chan event exception anomalypatchREPRODUCED (main, 9.0, 8.6, 8.5) — the message is no longer "can't read ::errorInfo", but a stale errorInfo from whatever error happened before is …

            Open ticket 2449154fff · 02. Event Loops · filed 2008-12-19

            🧪 Prototype check (2026-10-24)

            2449154fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before work-a3/v2449.tcl (catch {error stale}, then a chan event ... readable {...; return} handler, no bgerror) -> prints the stale couldn't open ".../tclIndex": no such file or directory ... "open [file join $dir tclIndex]" errorInfo instead of the bad-code message
            • Patch applied cleanly (23 changed lines)
            • After same -> command returned bad code: 2; after 0 return still gives command returned bad code: 2 ("after" script). Tests event (58 passed/0 failed), io, ioCmd, interp, timer, chanio: 0 failed
            • Test existing event-5.16 and event-7.8 (from the patch) fail on the unpatched tip / pass after
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5) — the message is no longer "can't read ::errorInfo", but a stale errorInfo from whatever error happened before is printed instead of "command returned bad code: 2"

            • Platform generic
            • Tested main (own build), 9.0.5, 8.6.18, 8.5.19: chan event $f readable {...; return} (also break) with no bgerror proc: prints the previous error (e.g. foo ... "error foo", or an auto_load "tclIndex" error caught long before); after 0 return prints "command returned bad code: 2 ("after" script)" correctly
            • Cause generic/tclEvent.c:TclDefaultBgErrorHandlerObjCmd sets the result to "command returned bad code: N" for non-error codes, but only touches errorInfo when the options carry -errorinfo, which Tcl_GetReturnOptions omits for non-error codes; the no-bgerror fallback then writes the stale ::errorInfo. [after] works only because AfterProc calls Tcl_AddErrorInfo first, which seeds errorInfo from the result; TclChannelEventScriptInvoker does not.
            • Prototype prototypes/2449154fff.patch (23 changed lines, tests event-5.16, event-7.8): for a non-error code without -errorinfo, seed errorInfo from the message with Tcl_AddErrorInfo(interp, ""). event, io, ioCmd, interp, timer tests pass; both new tests fail on 9.0.
            • Suggested ticket comment Still present: for a non-error exception from a [chan event] handler the default bgerror handler reports a stale errorInfo (from some earlier error) instead of "command returned bad code: 2". Proposed fix: TclDefaultBgErrorHandlerObjCmd seeds errorInfo from the message when the options have no -errorinfo; tests event-5.16, event-7.8.
            ⬜2433843fff2008return options can leak to next bytecodeREPRODUCED (main, 9.0, 8.6)

            Open ticket 2433843fff · 47. Bytecode Compiler · filed 2008-12-16

            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested main (own build), 9.0.5, 8.6.18: the ticket script still shows -leaky option in the error options; it also leaks into a later *successful* proc result (proc p {} {return -level 0 -leaky x; set y 1}; catch p m o gives o with -leaky). Not reproduced through direct evaluation ([testevalex]), only in bytecode.
            • Cause tclResult.c:TclProcessReturn stores the options in iPtr->returnOpts; with -level 0 -code ok execution continues, and TEBC (tclExecute.c, the TCL_OK path after invoke, "We do not call Tcl_ResetResult") never resets returnOpts, while compiled commands such as loadScalar never invoke anything. Clearing after INST_INVOKE would break catch {uplevel 0 {return -level 0 -foo bar}} m o, which must keep -foo; a correct fix needs a per-command reset point in the bytecode engine (INST_START_CMD is not always emitted).
            • Prototype none (not a small change; needs a design decision in the bytecode engine, see also 2007198)
            • Suggested ticket comment Still reproducible in 9.1, 9.0 and 8.6, including leaks into a later successful result. Direct evaluation is fine; the bytecode engine never resets returnOpts between commands after a level-0 &#91;return&#93;.
            ⬜2392973fff2008ERR_LEGACY_COPY not Save/RestoredREPRODUCED (main, 9.0, 8.6; by code reading, no script demonstrates it, as in the report)

            Open ticket 2392973fff · 45. Parsing and Eval · filed 2008-12-05

            Full verdict

            REPRODUCED (main, 9.0, 8.6; by code reading, no script demonstrates it, as in the report)

            • Platform generic
            • Tested generic/tclResult.c on main, 9.0, 8.6: Tcl_SaveInterpState stores only iPtr->flags & ERR_ALREADY_LOGGED; Tcl_RestoreInterpState restores only that bit. The workaround in TclSafeCatchCmd (tclResult.c, end of file) sets ERR_LEGACY_COPY by hand with the comment "todo: unless ERR_LEGACY_COPY not set in restore (branch [bug-554117edde] not merged yet)".
            • Cause tclResult.c:Tcl_SaveInterpState / Tcl_RestoreInterpState ignore ERR_LEGACY_COPY.
            • Prototype none (an existing fix is on branch bug-554117edde, sebres 2020 9cace6cc78, rebased onto 8.6 by jan.nijtmans 2023-09-12 45213b1bc0, not merged into core-8-6-branch, 9.0 or main; it saves and restores ERR_LEGACY_COPY and reads ::errorInfo/::errorCode when the interp fields are NULL, with tests result-1.9.*)
            • Suggested ticket comment Still true in 9.1/9.0/8.6. Same root cause as 554117edde; branch bug-554117edde (not merged) saves and restores ERR_LEGACY_COPY along with ERR_ALREADY_LOGGED and has tests. Suggest handling both tickets together.
            ⬜2345638fff2008Tcl_GetMemoryInfo aborts w/threaded memory allocatorpatchREPRODUCED (main, 9.0, 8.6; only with --enable-symbols=mem or -DPURIFY) — documented limitation rather than a bug

            Open ticket 2345638fff · 49. Threading · filed 2008-11-25

            🧪 Prototype check (2026-10-24)

            2345638fff.patch: VERIFIED+TEST-ADDED (only with --enable-symbols=mem / TCL_MEM_DEBUG or PURIFY builds; 8.6 has the same stub, not ported)

            • Base main (tip d6dd6d0f71). Reproduced in an out-of-tree build work-a1/build-mem (wt-a1/unix/configure --enable-symbols=mem, so -DTCL_MEM_DEBUG=1 while the Makefile still has -DUSE_THREAD_ALLOC=1; tclInt.h #undef's USE_THREAD_ALLOC for TCL_MEM_DEBUG and PURIFY).
            • Before work-a1/build-mem/gmi (work-a6/tsan/gmi.c: Tcl_FindExecutable + Tcl_GetMemoryInfo) -> Tcl_GetMemoryInfo called when threaded memory allocator not in use, SIGILL (rc 132). Same through the new test command: build-mem/tcltest work-a1/gmi.tcl -> same panic, rc 132; build-mem/tcltest tests/thread.test -match thread-9.* (new tests, fix removed) -> panic, rc 132. The normal build returns the stats (2 caches: thread0x..., shared).
            • Patch 79 changed lines (fix 10 + doc 6; the rest is the test command). generic/tclThreadAlloc.c: the non-threaded-allocator Tcl_GetMemoryInfo stub appends nothing instead of panicking. Why empty: the result is documented as a list-of-lists with one list per allocator cache; without the threaded allocator there are no caches, so an empty list is exactly true, and callers that parse the list (AOLserver/NaviServer stats pages) keep working; a "note" element would break such parsers, and a panic kills a server for asking for statistics. doc/Alloc.3: one list per cache; without the threaded allocator (TCL_MEM_DEBUG, PURIFY) nothing is appended (also fixes the missing full stop in the old sentence). generic/tclTest.c: new testmemoryinfo command; it calls (Tcl_GetMemoryInfo)(&ds) directly, because tclTest.c is stub-enabled and the Tcl_GetMemoryInfo stub macro aborts with "Cannot call Tcl_GetMemoryInfo from stubbed extension" (the tcltest executable links libtcl, so the direct call resolves). tests/thread.test: constraints testmemoryinfo, memory; tests thread-9.1, 9.2.
            • After build-mem: gmi -> len=0, rc 0; tcltest gmi.tcl -> 0; thread-9.1 and 9.2 pass. Normal build: thread-9.1 passes, 9.2 skipped (memory); thread.test 57 total, 4 passed, 53 skipped (no Thread package), 0 failed. Alloc.3 troff warnings 0/0.
            • Test added thread-9.1 (testmemoryinfo returns a list whose elements are shared/thread... lists, including shared if non-empty; constraint testmemoryinfo) and thread-9.2 (empty result when the threaded allocator is not used; constraints testmemoryinfo memory). In the --enable-symbols=mem build both fail before (process panics) and pass after; in the normal build 9.1 passes before and after.
            Full verdict

            REPRODUCED (main, 9.0, 8.6; only with --enable-symbols=mem or -DPURIFY) — documented limitation rather than a bug

            • Platform generic
            • Tested C program calling Tcl_GetMemoryInfo: main PURIFY build panics "Tcl_GetMemoryInfo called when threaded memory allocator not in use"; normal main build returns the stats. Same stub in 8.6 tclThreadAlloc.c:1152 and 9.0.
            • Cause generic/tclInt.h #undef's USE_THREAD_ALLOC for TCL_MEM_DEBUG and PURIFY builds, so tclThreadAlloc.c compiles the panicking stub although tclConfig.sh still lists -DUSE_THREAD_ALLOC=1. doc/Alloc.3 already says it is only available with the threaded allocator; in 9.x a script can check tcl::build-info memdebug/purify (Tcl 9 is always threaded). A configure switch for the allocator does not exist (RFE part not done).
            • Prototype none (would be a behaviour decision: e.g. append nothing instead of panicking)
            • Suggested ticket comment Still as designed: with TCL_MEM_DEBUG or PURIFY the threaded allocator is disabled and Tcl_GetMemoryInfo panics (documented in Alloc.3). In 9.x the build can be checked with &#91;tcl::build-info memdebug&#93; / &#91;tcl::build-info purify&#93;. Making it a no-op instead of a panic would be a small change if wanted.
            🔧2339907fff2008execing command that starts with 2 failspatchREPRODUCED (main, 9.0, 8.6, 8.4.15+) — regression in 8.4.13 from 47d7a23bdf (core-8-4-branch, 2006-03-16; trunk 99ff6f81ba, same day), fix for …

            Open ticket 2339907fff · 24. Channel Commands · filed 2008-11-24

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-17 check-in 622c092a53 on exec-pipeline-words, not merged: Fix 2339907fff and 1862356fff, regressions from 47d7a23bdf: in the first pass of TclCreatePipeline() a command word starting with "2" (but not "2>") was n
            • 💬 2026-10-17 serhiy.storchaka: Regression since 8.4.13 from 47d7a23bdf (the fix for [768659]): in the first pass of TclCreatePipeline() a word starting with "2" that is not a "2>" redirection does not count as a command, so [exec 2DSlider] fails with "illegal use of | or |& in command". Proposed fix in branch exec-pipeline-words, test exec-23.1. It also fixes 1862356fff.
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4.15+) — regression in 8.4.13 from 47d7a23bdf (core-8-4-branch, 2006-03-16; trunk 99ff6f81ba, same day), fix for Bug 768659; 8.4.2 and 8.4.11 work

            • Platform generic
            • Tested exec 2DSlider (script in PATH) on main, 9.0.5, 8.6.18, 8.4.20: "illegal use of | or |& in command"; with a path (bin/2DSlider) it works; 8.4.2/8.4.11 OK
            • Cause generic/tclPipe.c:TclCreatePipeline, first pass: case '2': breaks out for a word that is not 2>... without clearing needCmd, so a command word starting with "2" right after the start or a | is not counted as a command and the final if (needCmd) check reports the bar error.
            • Prototype prototypes/2339907fff.patch (7 changed lines, test exec-23.1); exec.test passes
            • Suggested ticket comment Regression from 47d7a23bdf (Bug 768659): in TclCreatePipeline a word starting with "2" but not "2>" does not clear needCmd. Proposed fix: set needCmd = 0 there; test exec-23.1.
            ⬜2166501fff2008fCmd-6.24,26 fail when [pwd] eq $HOMEpatchREPRODUCED (main; same test code in 9.0/8.6)

            Open ticket 2166501fff · 24. Channel Commands · filed 2008-10-14

            🧪 Prototype check (2026-10-24)

            2166501fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before cd work-a3/fhome; HOME=$PWD tcltest tests/fCmd.test -> fCmd-6.24 and fCmd-6.26 FAILED (Total 305 Passed 258 Skipped 45 Failed 2)
            • Patch applied cleanly (6 changed lines, test only)
            • After same command -> Total 305 Passed 260 Skipped 45 Failed 0; normal cwd: Failed 0
            • Test the patch fixes the tests themselves; the cwd=HOME run is the proof (fails before / passes after)
            Full verdict

            REPRODUCED (main; same test code in 9.0/8.6)

            • Platform unix (test suite)
            • Tested main own build, fCmd.test run with cwd = HOME = work-a4/fhome: fCmd-6.24 and 6.26 fail ("error copying .../td1" to "td1/td1"); with a normal cwd they pass
            • Cause tests/fCmd.test 6.24/6.26 create [file home]/td1 and copy it to the relative target "td1"; when cwd is HOME the target is the source itself, so the copy goes into td1/td1 and the error message differs.
            • Prototype prototypes/2166501fff.patch (6 changed lines, test-only: copy target renamed to td1copy, still removed by cleanup's td* glob); fCmd.test passes both with cwd = HOME and normally
            • Suggested ticket comment The tests copy [file home]/td1 to "td1", which is the same directory when the suite runs from $HOME. Proposed fix: use the target name td1copy in fCmd-6.24 and 6.26.
            ⬜2098955fff2008TCL_COMPILE_STATS lacking a mutexpatchREPRODUCED (main, only with -DTCL_COMPILE_STATS; 9.0/8.6 same code by reading)

            Open ticket 2098955fff · 41. Memory Allocation · filed 2008-09-07

            🧪 Prototype check (2026-10-24)

            2098955fff.patch: VERIFIED (GCC/clang only; MSVC threaded builds still use the plain increments, as the triage says)

            • Base main (tip d6dd6d0f71)
            • Before TSan build of the tip with -DTCL_COMPILE_STATS (work-a6/tsanstats, rebuilt from scratch), Thread 3.0.7, tclsh cs3.tcl (4 threads allocating objects), 3 runs -> 346/384/393 TSan reports, of them 190/204/208 on global tclObjsAlloced and 133/157/168 on tclObjsFreed
            • Patch applied cleanly (offset 1 line)
            • After same build rebuilt from scratch with the patch, 3 runs -> 14/12/12 reports, 0 on either counter (remaining ones are the threaded allocator's, unrelated); obj.test 84/0 failed and compile.test 171/0 failed on the stats build
            • Test none possible (needs a COMPILE_STATS + TSan build)
            Full verdict

            REPRODUCED (main, only with -DTCL_COMPILE_STATS; 9.0/8.6 same code by reading)

            • Platform generic (debug setting)
            • Tested TSan build of main with -DTCL_COMPILE_STATS, 4 Thread-package threads allocating objects (work-a6/cs3.tcl): 200 races on 'tclObjsAlloced' and 173 on 'tclObjsFreed' per run; with the prototype none (the remaining reports are the threaded allocator's sharedCache/cachedInExit, unrelated). tclObjsShared is already updated under tclObjMutex.
            • Cause generic/tclInt.h TclIncrObjsAllocated/TclIncrObjsFreed are plain ++ on process-wide counters.
            • Prototype prototypes/2098955fff.patch (15 changed lines, no test: needs a COMPILE_STATS+TSan build) — relaxed __atomic_fetch_add for GCC/clang threaded builds, plain increments otherwise (MSVC would need InterlockedIncrement, not done). A mutex per object alloc/free seemed too costly.
            • Suggested ticket comment Still present: with TCL_COMPILE_STATS, tclObjsAlloced/tclObjsFreed are incremented without synchronisation (ThreadSanitizer confirms). Proposed fix: atomic relaxed increments in TclIncrObjsAllocated/TclIncrObjsFreed (GCC/clang), patch available.
            ⬜2015360fff2008untested Tcl_ResetResult() callspatchREPRODUCED (main; the calls are still not covered by tests)

            Open ticket 2015360fff · 47. Bytecode Compiler · filed 2008-07-10

            🧪 Prototype check (2026-10-24)

            2015360fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before not a behaviour bug. On the tip the call in TclCompileIfCmd (tclCompCmdsGR.c:261) still follows Tcl_GetBooleanFromObj(NULL, boolObj, ...), and TokenToObj takes no interp, so nothing can have written the result: the call is dead. The tclCompile.c one is still commented out.
            • Patch applied cleanly (3 changed lines)
            • After tests if, compile, subst, basic, expr, expr-old, execute: 0 failed
            • Test none possible (dead-code removal, no observable behaviour)
            Full verdict

            REPRODUCED (main; the calls are still not covered by tests)

            • Platform generic
            • Tested main (own build): with all three remaining Tcl_ResetResult calls in the compilers disabled (tclCompCmdsGR.c:261 in TclCompileIfCmd, tclCompCmdsGR.c:2841 in TclCompileReturnCmd for bad options, tclCompCmdsSZ.c:1510 in TclSubstCompile on a parse error) the full suite (all.tcl -notfile 'http*.test') shows no new failures (only the 25 known zipfs.test failures from TCL_LIBRARY), and probe scripts (bad [return] options, broken [subst]) behave the same as 9.0.5. The tclCompile.c call from the ticket was already commented out in 2009 (7748dccef5, Tcl_NRSubstObj, "TODO: Confirm we do not need this").
            • Cause the if-compiler one is dead code: since the constant check uses Tcl_GetBooleanFromObj(NULL, ...), nothing writes the result there (same in 8.6/9.0). The return and subst ones clear an error message that a compile function itself wrote, so they are defensive, not proven unnecessary.
            • Prototype prototypes/2015360fff.patch (3 changed lines: drops the dead call in TclCompileIfCmd and the commented-out call with its TODO in Tcl_NRSubstObj; full suite passes)
            • Suggested ticket comment The &#91;if&#93; one is dead now (the constant test passes a NULL interp); proposed removal, together with the commented-out call in Tcl_NRSubstObj. The calls in the &#91;return&#93; and &#91;subst&#93; compilers clear messages the compilers set themselves; the suite passes without them, but they are kept as cheap safety.
            ⬜2006342fff2008[expr] can bypass [interp limit]REPRODUCED (main, 9.0, 8.6)

            Open ticket 2006342fff · 56. LibTomMath · filed 2008-06-30

            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested work-a2/t2006.tcl: a child interp with interp limit time 1 s ahead runs expr {330000000 > 0}. main (debug build) ran 20.5 s, 9.0.5 4.0 s, 8.6.18 11.0 s, and every one returned rc=0 '1': the limit was neither enforced during the computation nor raised afterwards. The original 29999999 is now fast (it is a shift), so a non-power-of-2 base is needed.
            • Cause limit/cancel checks happen only at bytecode instruction boundaries (TclLimitReady/Tcl_LimitCheck in TEBC); a single INST_EXPON on bignums runs entirely inside libtommath (mp_expt_n / mp_mul) with no hook. Because the eval ends right after the instruction, nothing checks the limit afterwards either. [interp cancel] has the same problem.
            • Prototype none (needs limit hooks inside libtommath loops, or a size cap on ** results in safe interps: a design decision)
            • Suggested ticket comment Still present in 8.6.18/9.0.5/main: expr {3**30000000 > 0} in a time-limited interp runs 4–20 s and returns normally. A fix needs either limit checks inside the bignum loops or a cap on result size (e.g. for safe interps).
            ⬜2001389fff2008empty string as path: troubling exceptionsREPRODUCED (main, 9.0, 8.6, 8.4): this is a design question, not a defect with a local fix

            Open ticket 2001389fff · 36. Pathname Management · filed 2008-06-24

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4): this is a design question, not a defect with a local fix

            • Platform generic
            • Tested main, 9.0.5, 8.6.18, 8.4.11/8.4.15: file pathtype {} gives relative, file normalize {} gives {}, file system {} raises "unrecognised path", file dirname {} gives ., file type {} raises ENOENT. All versions behave the same.
            • Cause generic/tclPathObj.c special-cases "" on purpose (comment in Tcl_FSGetNormalizedPath area: "valid path ... but doesn't exist").
            • Prototype none (making "" consistent changes documented/observable behaviour; needs a decision/TIP)
            • Suggested ticket comment Unchanged in 8.6.18, 9.0 and 9.1. Making {} behave like other relative paths (or like no path at all) is a behaviour change that needs a decision, maybe a TIP. Is anyone interested?
            ✅1967630fff2008tclLoadDl.c dlerror() is not thread safepatchREPRODUCED at the libc level (NetBSD 10.1 often, OpenBSD 7.9 rarely); NOT-REPRODUCED through Tcl [load] (main, 4 threads, both OSes)

            Open ticket 1967630fff · 40. Dynamic Loading · filed 2008-05-20

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Fixed)

            • 🔧 2026-10-25 check-in 1773409c76 on unix-dlerror-mutex, in main: Fix 1967630fff: tclLoadDl.c dlerror() is not thread safe. On NetBSD and OpenBSD the dlerror() state is process-wide, so a failing [load] in one thread could g
            • 🔧 2026-10-28 check-in e0e0341124 on trunk, in main, 9.0: Fix 1967630fff: tclLoadDl.c dlerror() is not thread safe
            • 🔧 2026-10-28 check-in 803bbbf4d9 on core-9-0-branch, in main, 9.0, 8.6: Fix 1967630fff: tclLoadDl.c dlerror() is not thread safe
            • 🔧 2026-10-28 check-in d870c50070 on core-8-6-branch, in main, 9.0, 8.6: Fix 1967630fff: tclLoadDl.c dlerror() is not thread safe
            • 💬 2026-10-25 serhiy.storchaka: Reproduced on NetBSD 10.1 and OpenBSD 7.9, where the dlerror() state is process-wide: a failing [load] concurrent with a successful [load] in another thread gets the other thread's message or NULL, and NULL crashes in Tcl_ObjPrintf(). Linux, FreeBSD and macOS keep it per thread. Proposed fix in branch unix-dlerror-mutex, based on the attached patch: a mutex from dlopen()/dlsym() until dlerror() has been read, and …
            • 💬 2026-10-28 jan.nijtmans: Fixed in e0e0341124, core-9-0-branch and core-8-6-branch Many thanks!
            🧪 Prototype check (2026-10-24)

            1967630fff.patch: VERIFIED (fix proven with an emulating shim, 40 runs before/after; no test in the suite)

            • Base main (tip d6dd6d0f71). Threaded build (wt-a1/unix), Thread 3.0.7 built in work-a1/thread/build-a1 against it. Hunks 3-5 do not apply to core-8-6-branch (different code around dlopen); same approach would port by hand.
            • Before on glibc dlerror() is thread-local, so the race cannot occur natively: dl/runmany.sh tclsh 40 400 "" -> SUMMARY runs=40 ok=40 wrong=0 crash=0. To reproduce the NetBSD/OpenBSD behaviour (process-global dlerror, confirmed there in result-unix-1.md) on Linux I wrote an LD_PRELOAD shim work-a1/dl/globaldlerror.c: dlopen()/dlsym() failures store the message in one global buffer, dlerror() returns and clears it, plus usleep(20) after each call. Note that [load] already serialises Tcl_LoadFile with tclLoad.c's libraryMutex, which is why plain failing-[load] stress (dl/dlerr.tcl, 3 threads) never showed anything even with the shim. But the Tcl_FindSymbol calls for *_SafeInit/*_Unload/*_SafeUnload after a successful load run outside that mutex. work-a1/dl/dlerr2.tcl: thread L loads 400 distinct copies of dltest/tcl9pkga.so (which lacks those symbols, so dlsym fails), threads AAAA/BBBB [load] missing files 1600 times each. With the shim: dl/runmany.sh tclsh 40 400 dl/globaldlerror.so -> SUMMARY runs=40 ok=0 wrong=0 crash=40: every run SEGFAULTs. gdb: AppendPrintfToObjVA <- Tcl_ObjPrintf("couldn't load file \"%s\": %s") <- TclpDlopen (tclLoadDl.c:132) <- Tcl_LoadObjCmd: the other thread's failing dlsym() + dlerror() consumed the message, so this thread's dlerror() returned NULL and TclpDlopen passed NULL to %s (FindSymbol already guards NULL with "unknown", TclpDlopen does not). So on NetBSD/OpenBSD a failed [load] concurrent with a successful [load] in another thread can crash Tcl, not just give a wrong message (not run on the VMs; the shim emulates their documented behaviour).
            • Patch 16 changed lines, unix/tclLoadDl.c only: Tcl-private TCL_DECLARE_MUTEX(dlMutex) held from the first dlopen() until the error message has been copied into the interp result (unlocked on the encoding-error early return and on both exits), and in FindSymbol from the first dlsym() until after dlerror()/the error message; plus the NULL guard ("unknown") for dlerror() in TclpDlopen, like FindSymbol has. Tcl mutexes are reentrant on main (TIP 509), so a library constructor that calls Tcl_LoadFile in the same thread does not deadlock. tclLoadDyld.c (macOS) not changed: Apple's dlerror() keeps per-thread state, so it is not affected. As noted in triage, the mutex only covers Tcl's own calls, not dlopen() by other code in the process.
            • After with the shim dl/runmany.sh tclsh 40 400 dl/globaldlerror.so -> SUMMARY runs=40 ok=40 wrong=0 crash=0 (each run: L 0, AAAA 0, BBBB 0 wrong); without the shim 40/40 ok. After fossil revert + rebuild the shim run crashes again (3/3). load.test 30/30 passed, unload.test 27/27 passed.
            • Test none added. Without the shim it never fails on Linux (40/40 before), and a suite test would need the Thread package, an LD_PRELOAD shim and hundreds of library copies, so it is neither cheap nor meaningful on the platforms where the suite usually runs. The reproducer is kept in work-a1/dl (globaldlerror.c, dlerr2.tcl, runmany.sh; copies/ is 47 MB and can be regenerated with a cp loop). On NetBSD 10.1 dlerr2.tcl without the shim should show the crash/wrong messages before the fix.

            1967630fff.patch (native NetBSD/OpenBSD): VERIFIED natively. The bug is real on NetBSD 10.1 and OpenBSD 7.9: in 40 runs, 33/40 (NetBSD) and 24/40 (OpenBSD) fail before the patch, half of them by SIGSEGV. With the patch, 0/40 fail on both.

            • Base main (tip d6dd6d0f71, fossil tarball trunk), unpatched and with prototypes/1967630fff.patch (patch -p0), built natively on NetBSD 10.1 (gcc) and OpenBSD 7.9 (clang) in /tmp: unix/configure (threaded by default), gmake, gmake tcltest, unix/dltest built (tcl9pkga.so). Thread trunk (39d9f08aa0, 3.0.7) with tclconfig trunk in its tclconfig/, built against each Tcl build dir. 400 copies of the unpatched build's dltest/tcl9pkga.so made with a cp loop on each VM.
            • dlerror() is process-wide on both: work-a1/dl/dlerrt.c (a second thread makes dlopen() fail and is joined, then main calls dlerror()) prints the other thread's message on both: NetBSD Cannot open "/nonexist-B.so", OpenBSD File not found -> VERDICT: dlerror state is PROCESS-WIDE.
            • Repro work-a1/dl/dlerr2-bsd.tcl = dlerr2.tcl with one change. The glibc message pattern (...: /nonexist-X.so: *) does not fit the BSD loaders, and OpenBSD's message has no file name at all. So each failing load is now compared with the message the same [load] gives single-threaded before the threads start. Thread L loads pkga1..400.so; AAAA/BBBB each [load] a missing file 1600 times. Driver work-a1/dl/runvm-bsd.sh (ulimit -c 0, timeout 120 per run; reports ok/wrong/crash/hang). Each build ran 40 times with the same parameters (n=400).
            • Before (unpatched), reproduced natively on both, no shim:
              • NetBSD: SUMMARY runs=40 ok=7 wrong=13 crash=20 hang=0. All 20 crashes are rc=139 (SIGSEGV). Wrong messages are another thread's dlsym() error delivered as a failed [load] result, e.g. couldn't load file "/nonexist-AAAA.so": Undefined symbol "_Pkga_SafeUnload". Some are garbled, because the shared buffer was overwritten while being read: Undefined symbol "xist-ASafeUnl and Undefined symbol exist-AAAA.so".
              • OpenBSD: SUMMARY runs=40 ok=16 wrong=4 crash=20 hang=0. All 20 crashes are rc=139. Wrong messages: couldn't load file "/nonexist-AAAA.so": Unable to resolve symbol.
              • The crashes fit the NULL-to-%s path found with the Linux shim (TclpDlopen <- Tcl_ObjPrintf). I could not get a native backtrace: under gdb the race never fired (3+8 tries on OpenBSD, gdb 6.3; 3 tries on NetBSD), and one NetBSD gdb run hung (killed by PID). No core files (ulimit -c 0).
            • Patch unchanged prototypes/1967630fff.patch (16 changed lines, unix/tclLoadDl.c).
            • After (patched): NetBSD SUMMARY runs=40 ok=40 wrong=0 crash=0 hang=0; OpenBSD SUMMARY runs=40 ok=40 wrong=0 crash=0 hang=0. In the patched tree, gmake test TESTFLAGS="-file \"load.test unload.test\"" gives Total 57 Passed 57 Skipped 0 Failed 0 on both (load.test + unload.test).
            • Test none added, for the same reason as above: it needs the Thread package and hundreds of library copies. The repro is work-a1/dl/dlerr2-bsd.tcl + runvm-bsd.sh + dlerrt.c.
            Full verdict

            REPRODUCED at the libc level (NetBSD 10.1 often, OpenBSD 7.9 rarely); NOT-REPRODUCED through Tcl [load] (main, 4 threads, both OSes)

            • Platform unix
            • Tested code reading of unix/tclLoadDl.c on main: dlerror() is still called without a mutex after dlopen() (line ~129) and dlsym() (line ~226). The second half of the ticket (TclpFindSymbol reporting into interp) is already done: the error now goes into interp.
            • Cause unix/tclLoadDl.c:TclpDlopen/FindSymbol: dlerror() isn't required to be thread-safe by POSIX (before POSIX.1-2024).
            • Prototype none (a Tcl-private mutex would only protect Tcl's own load calls, not dlopen() done by extensions or libraries. Low practical value.)
            • Suggested ticket comment The interp-error half was done long ago. The remaining issue only matters on libcs with a global dlerror buffer (glibc, musl and FreeBSD use thread-local state). A Tcl-internal mutex would not cover dlopen() calls by other code. To check on the NetBSD/OpenBSD VMs whether it's still relevant there.
            Unix VMs, run 1
            • Platform unix (NetBSD, OpenBSD)
            • Verdict REPRODUCED at the libc level (NetBSD 10.1 often, OpenBSD 7.9 rarely); NOT-REPRODUCED through Tcl [load] (main, 4 threads, both OSes)
            • Tested C tests with two pthreads calling dlopen() on files that fail differently and checking dlerror() (work-unix1/scripts/dlerr.c; dlerr2.c for OpenBSD, whose messages have no file name, so it pairs a missing file with a non-ELF file). NetBSD 10.1: about 7–9% of 200000 calls per thread get NULL or the other thread's message (e.g. Cannot open "/nonexist-BBBB.so" in thread AAAA). OpenBSD 7.9: 1 NULL in 200000 (2 runs; ld.so serializes dlopen, so the window is tiny). Tcl tests (dlerr.tcl / dlerr2.tcl): 3–4 Thread-package threads load failing files 100000–300000 times each and check that the message matches their own failure. 0 wrong on NetBSD (3.6 million loads) and OpenBSD (3 runs × 400000).
            • Cause unix/tclLoadDl.c:TclpDlopen/FindSymbol read dlerror() right after dlopen()/dlsym() without a lock. On NetBSD and OpenBSD the error buffer is process-global, and dlerror() clears it.
            • Prototype none (a Tcl-private mutex around dlopen()+dlerror() and dlsym()+dlerror() would be about 15 lines, but it would not cover dlopen() calls from other code, and Tcl never hit the race in these runs, so a fix could not be verified)
            • Suggested ticket comment Confirmed at the C level: dlerror() is process-global on NetBSD 10.1 (concurrent dlopen() failures often get the other thread's message or NULL) and on OpenBSD 7.9 (very rare). Through &#91;load&#93; with 4 threads it never showed up. A mutex in tclLoadDl.c would only cover Tcl's own calls. Low priority.
            ✅1943546fff2008Strangely failing [puts]REPRODUCED (Tk 9.1b1, Tk 8.6.18; not with Tk 9.0.5, which wraps this line differently) — a Tk bug, belongs in the Tk tracker

            Open ticket 1943546fff · 24. Channel Commands · filed 2008-04-16

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Duplicate)

            • 💬 2026-10-18 serhiy.storchaka: This is a Tk text widget bug: display lineend can land inside a multi-byte character, so [.t get] returns the lone first byte of "ä" and [puts] fails. Moved to the Tk tracker as 18f35ed566, with a smaller reproducer. The script attached here no longer fails with the current branches, since Tk 7d8d10e4a9 lays out an unmapped text widget at its requested size, but the bug itself remains.
            Full verdict

            REPRODUCED (Tk 9.1b1, Tk 8.6.18; not with Tk 9.0.5, which wraps this line differently) — a Tk bug, belongs in the Tk tracker

            • Platform generic (Tk text widget, not Tcl)
            • Tested Xvfb :74, installed wish9.1 (9.1b1, under dbus-run-session), wish9.0 (9.0.5), wish8.6 (8.6.18) with the attached pitsana.tk; work-a4/1943546-pit4.tcl wraps [puts] to dump the code points: 9.1b1 and 8.6 get "U+00C3" from $t get "1.63 + 1 display lines display linestart" "... display lineend" (indices 1.64-1.65, the a-umlaut), i.e. the lead byte of the UTF-8 "ä" as a character of its own. 9.1 then fails in puts with "invalid or incomplete multibyte or wide character" (strict profile), 8.6 with "invalid argument". The never-ending loop is the script's own end condition.
            • Cause In Tk's text display-line index code (tkTextDisp.c / tkTextIndex.c, "+N display lines" with -wrap word breaking inside a long word) the computed byteIndex lands inside a multi-byte UTF-8 character; TextGetText then copies half a character (as the 2008 comment found: byteIndex 64..65 covers 0xC3 only).
            • Prototype none (Tk; not investigated further in this Tcl batch)
            • Suggested ticket comment Still reproducible with Tk 9.1b1 and 8.6.18: "+1 display lines" in a -wrap word text returns an index in the middle of the UTF-8 sequence of "ä", so [.t get] returns the lone lead byte (U+00C3) and a strict [puts] fails. This is a Tk text widget bug; suggest moving it to the Tk tracker.
            ⬜1902648fff2008Do not use strcasecmppatchREPRODUCED (main, 9.0, 8.6); the -nocase correctness half is FIXED

            Open ticket 1902648fff · 52. Portability Support · filed 2008-02-27

            🧪 Prototype check (2026-10-24)

            1902648fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before work-a4/v4strict.sh (the Makefile's gcc line for tclClock.o, tclClockFmt.o, tclUnixChan.o plus -std=c11 -D_POSIX_C_SOURCE=200809L) -> tclClock: error: implicit declaration of function 'strcasecmp'; tclUnixChan: implicit declaration of function 'strncasecmp' (and cfmakeraw)
            • Patch applied cleanly (3 changed lines)
            • After tclClock and tclClockFmt compile cleanly; tclUnixChan only keeps implicit declaration of function 'cfmakeraw' (a BSD/glibc extension, not POSIX, outside this ticket); full rebuild without new warnings; clock.test 8756 passed / 7 skipped / 0 failed; unixChan.test + chanio.test 766 passed / 0 failed
            • Test none possible (compile-time portability; the strict compile above is the check)
            Full verdict

            REPRODUCED (main, 9.0, 8.6); the -nocase correctness half is FIXED

            • Platform unix (portability)
            • Tested the TIP 241 [lsort/lsearch/switch -nocase] uses now go through TclUtfCasecmp (8.6+), so -nocase no longer uses strcasecmp. Remaining strcasecmp/strncasecmp calls: main generic/tclClock.c (locale names), tclClockFmt.c, unix/tclUnixChan.c (serial options), tclEnv.c (Windows-only). No Tcl header includes <strings.h>, where POSIX declares them. Compiling main tclClock.c with -std=c11 -D_POSIX_C_SOURCE=200809L (glibc) → error: implicit declaration of function 'strcasecmp' (gcc 14+ makes this an error). Default glibc builds work only because <string.h> also declares it under _DEFAULT_SOURCE.
            • Cause unix/tclUnixPort.h does not include <strings.h>, although configure already defines HAVE_STRINGS_H.
            • Prototype prototypes/1902648fff.patch (3 changed lines: #ifdef HAVE_STRINGS_H / #include <strings.h> / #endif in unix/tclUnixPort.h). With it, the strict compile of tclClock.c, tclClockFmt.c and tclUnixChan.c is clean; the main build succeeds; clock.test 8753 passed / 0 failed.
            • Suggested ticket comment The -nocase uses were replaced by TclUtfCasecmp long ago, but strcasecmp/strncasecmp are still used (clock, serial channel options) without <strings.h>, so strict POSIX compiles (or dietlibc) fail. Proposed fix: include <strings.h> in tclUnixPort.h when HAVE_STRINGS_H.
            🔧1862356fff2008arguments to TclCreatePipeline cannot start with '|'patchREPRODUCED (main, 9.0, 8.6, 8.4.15+) — same regression as 2339907: 47d7a23bdf (core-8-4-branch, 2006-03-16; trunk 99ff6f81ba), fix for Bug 7686…

            Open ticket 1862356fff · 25. Channel System · filed 2008-01-02

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-17 check-in 622c092a53 on exec-pipeline-words, not merged: Fix 2339907fff and 1862356fff, regressions from 47d7a23bdf: in the first pass of TclCreatePipeline() a command word starting with "2" (but not "2>") was n
            • 💬 2026-10-17 serhiy.storchaka: Regression since 8.4.13 from 47d7a23bdf (the fix for [768659]): the first pass of TclCreatePipeline() takes any word starting with "|" for a bar, although only "|" and "|&" are, so [exec echo |foo] fails with "illegal use of | or |& in command". Proposed fix in branch exec-pipeline-words, tests exec-23.2 and exec-23.3. It also fixes 2339907fff.
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4.15+) — same regression as 2339907: 47d7a23bdf (core-8-4-branch, 2006-03-16; trunk 99ff6f81ba), fix for Bug 768659; 8.4.2/8.4.11 print "|foo"

            • Platform generic
            • Tested exec echo |foo and exec echo a |foo on main (own build), 9.0.5, 8.6.18, 8.4.20: "illegal use of | or |& in command"; 8.4.2, 8.4.11: OK
            • Cause generic/tclPipe.c:TclCreatePipeline, first pass: case '|': treats any word starting with "|" as a bar (lastBar, cmdCount++, needCmd = 1) even when more characters follow, while the second pass (segment split) only treats exactly "|" / "|&" as bars. Since 47d7a23bdf the needCmd check turns that into an error.
            • Prototype prototypes/1862356fff.patch (27 changed lines, tests exec-23.2, exec-23.3): a word like "|foo"/"|&foo" is an ordinary word (needCmd = 0, no bar counted); exec.test passes (exec-10.* bar errors unchanged). Applies independently of the 2339907 patch (different hunk, both add tests after exec-22.1 — the second one applied needs its test hunk re-anchored).
            • Suggested ticket comment Regression from 47d7a23bdf: the first pass of TclCreatePipeline counts any word starting with "|" as a bar, although only "|" and "|&" are. Proposed fix: treat "|foo" as an ordinary word; tests exec-23.2, exec-23.3.
            ⬜1835292fff2007Safe interp problems with namespaces and ensemblesREPRODUCED (main, 9.0, 8.6) for item (4) only; (2)/(3) fixed long ago (chan is safe and exposed), (1) is by design (dkf, 2012)

            Open ticket 1835292fff · 20. [interp] · filed 2007-11-20

            Full verdict

            REPRODUCED (main, 9.0, 8.6) for item (4) only; (2)/(3) fixed long ago (chan is safe and exposed), (1) is by design (dkf, 2012)

            • Platform generic
            • Tested main, 9.0, 8.6: interp hide foo info then foo eval {::tcl::info::commands i*} still works; same for interp hide bar string + ::tcl::string::length, interp hide bar chan + ::tcl::chan::names.
            • Cause generic/tclInterp.c Tcl_HideCommand only moves the global ensemble command into the hidden table; the ::tcl::<ens>::* implementation commands stay visible. Hiding requires the user to also hide/rename the subcommands, or the ensemble machinery to be namespace-aware. Unsafe subcommands are already hidden individually (tcl:file:*, tcl:info:cmdtype, ...), so the default safe profile is not affected; only user-made restrictions are.
            • Prototype none (design issue: namespace-aware hiding or "hide ensemble with its implementation namespace" semantics needed)
            • Suggested ticket comment Items (2) and (3) are fixed; (1) is by design. (4) still holds in 8.6/9.0/9.1: hiding an ensemble leaves ::tcl::<name>::* callable. The default safe profile hides the unsafe subcommands individually, so this only affects users who hide extra ensembles; documenting it in interp(n) would be the minimal fix.
            🔧1827008fff2007delete and copy fail in linux when path contains ".."patchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4), now worse: since 8.5 file delete -force d1/d2/d3/.. returns success but only empties d2 (d2 itself remains…

            Open ticket 1827008fff · 37. File System · filed 2007-11-06

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-17 check-in 50534654de on delete-refuse-dot-dot, not merged: Fix 1827008fff: on Unix [file delete -force d1/d2/d3/..] deleted the contents of d2 but not d2, and reported success. Refuse to remove a path ending in "." or
            • 💬 2026-10-17 serhiy.storchaka: Still there, and since 8.5 the failure is silent: [file delete -force d1/d2/d3/..] deletes the contents of d2 but not d2 itself, and returns success (the final rmdir() fails with ENOENT, which is ignored). Proposed fix in branch delete-refuse-dot-dot: on Unix, refuse to remove a path ending in "." or "..", as rm(1) does. Normalizing the path, as on Windows, would be wrong on POSIX: ".." is resolved physically ther…
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4), now worse: since 8.5 file delete -force d1/d2/d3/.. returns success but only empties d2 (d2 itself remains). 8.4 at least reported an error.

            • Platform unix
            • Tested main, 9.0.5, 8.6.18, 8.5.19, 8.4.15 (strace on main)
            • Cause unix/tclUnixFCmd.c:TclpObjRemoveDirectory passes the unnormalized native path. rmdir("d2/d3/..") fails with ENOTEMPTY, the traversal deletes d2's contents (d3 included). The final rmdir("d2/d3/..") then fails with ENOENT, which TclFileDeleteCmd ignores as "already deleted". file copy d1/d2/d3/.. x targets x/.. (file tail is ".."), giving "file exists". That part is not addressed.
            • Prototype prototypes/1827008fff.patch (29 changed lines, test fCmd-5.13): use the normalized path when the path ends in "..". "." is left alone on purpose: file delete -force . would then remove the cwd. fCmd, unixFCmd, fileSystem and fileName tests pass.
            • Suggested ticket comment Still there, and since 8.5 the failure is silent: d2's contents get deleted but d2 stays, and the final ENOENT is ignored. Proposed fix: TclpObjRemoveDirectory on Unix uses the normalized path for paths ending in ".." (test fCmd-5.13). file copy with a ".." tail still copies to target/...
            ⬜1812624fff2007overeager compilationREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            Open ticket 1812624fff · 47. Bytecode Compiler · filed 2007-10-13

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            • Platform generic
            • Tested set x 1; set x [proc set args "return $x"] in a child interp and in a proc: x ends up "" (the bytecoded [set] runs), not 1 as Rule 2 of Tcl.n would require.
            • Cause generic/tclCompile.c compiles [set] inline (INST_STORE_*); the compiled code is not invalidated mid-command when a command is redefined during argument substitution (the compile-epoch check happens only between bytecode executions). Same for any compiled command.
            • Prototype none (fixing needs either a per-instruction epoch check, i.e. a large slowdown, or a change to Tcl.n Rule 2; msofer and dkf both judged it won't-fix in 2007/2008)
            • Suggested ticket comment Still the same in 8.6/9.0/9.1: bytecoded commands are bound before their arguments are substituted. The realistic resolution is a note in Tcl.n that redefining a command while its own arguments are substituted has undefined effect; otherwise Won't Fix.
            ⬜1805928fff2007[package require] in index script protectionpatchREPRODUCED (main, 9.0, 8.6, 8.5)

            Open ticket 1805928fff · 39. Package Manager · filed 2007-10-01

            🧪 Prototype check (2026-10-24)

            1805928fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before work-a3/circ/r.tcl (an auto_path dir whose pkgIndex.tcl begins with package require circ1805928) -> 141 identical lines error reading package index file .../pkgIndex.tcl: can't find package circ1805928, then can't find package circ1805928. work-a3/circ/u.tcl (a [package unknown] handler that requires its own package) -> too many nested evaluations (infinite loop?).
            • Patch applied cleanly (58 changed lines)
            • After r.tcl -> a single error reading package index file ...: circular package dependency: attempt to find circ1805928 requires circ1805928 ("package require" in an index script?) line plus the final "can't find package"; u.tcl -> circular package dependency: attempt to find foo requires foo (...). Tests package (404/0 failed), pkgMkIndex, tm, safe, safe-zipfs, init, autoMkindex, interp, unknown, platform: 0 failed.
            • Test existing package-3.30.1 (from the patch) fails before / passes after; package-3.30.2 passes both (it is a regression guard)
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5)

            • Platform generic
            • Tested auto_path dir whose pkgIndex.tcl starts with [package require <itself>]: no stack overflow any more, but every version logs the same "error reading package index file ...: can't find package circ1805928" 170 times (recursive tclPkgUnknown) and ends with a plain "can't find package". A [package unknown] handler requiring its own package gives "too many nested evaluations (infinite loop?)" (9.0).
            • Cause generic/tclPkg.c PkgRequireCoreStep1 invokes the package unknown script again for a package whose unknown script is already running.
            • Prototype prototypes/1805928fff.patch (58 changed lines, tests package-3.30.1, package-3.30.2): mark Package.clientData while the unknown script runs; a nested require that has to run it again fails with "circular package dependency: attempt to find X requires X ("package require" in an index script?)" (errorcode TCL PACKAGE CIRCULARITY). Requiring a version that the index already registered via ifneeded still works (3.30.2). The repro now logs one line with that message. package.test, pkgMkIndex.test, tm, safe*, init, autoMkindex, interp, unknown tests pass. Modern port of the 2007 circ.patch.
            • Suggested ticket comment Still present: no stack overflow now, but the index is re-sourced recursively (170 identical tclLog lines) and the error is unhelpful. Proposed fix, a modern port of circ.patch: detect re-entry of [package unknown] for the same package and report a circular dependency (tests package-3.30.1/3.30.2).
            🔧1802895fff2007AfterDelay overflowpatchREPRODUCED (9.0, 8.6; on 64-bit Linux only for delays within ~1 s of 2^63 ms); FIXED on main since 9.1a1 (TIP 723 acb0a91c2c, 2026-01-08: deadlin…

            Open ticket 1802895fff · 03. Timer Events · filed 2007-09-26

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-19 check-in 42f3de85fb on after-overflow-9-0, not merged: Fix 1802895fff: [after] with a delay close to 2**63 ms could overflow the time arithmetic, so a blocking [after] returned at once. Add the delay with a satura
            • 🔧 2026-10-19 check-in d7e3050ca6 on after-overflow-8-6, not merged: Fix 96c3f3b47d and 1802895fff: [after] with a delay over about 12 years overflowed the seconds where long is 32 bits (Windows, 32-bit systems), so later tim
            • 💬 2026-10-19 serhiy.storchaka: Fixed on main by TIP 723 (too large delays give "time too far away"). Still present in 9.0 and 8.6: a blocking after 0x7fffffffffffffff returns at once on 64-bit systems (the millisecond arithmetic overflows), and where long is 32 bits any delay over about 11 years does. Proposed fix in branches after-overflow-9-0 and after-overflow-8-6: both forms of [after] add the delay with a saturating TimeAddMs(). Tests time…
            Full verdict

            REPRODUCED (9.0, 8.6; on 64-bit Linux only for delays within ~1 s of 2^63 ms); FIXED on main since 9.1a1 (TIP 723 acb0a91c2c, 2026-01-08: deadlines are long long µs and too-large delays raise "time too far away")

            • Platform generic (worst on Windows, where Tcl_Time.sec is a 32-bit long)
            • Tested Linux x86_64: after 0x7fffffffffffffff (blocking) returns at once on 8.6.18, 9.0.5 and the 9.0 tip whenever the current µs field is >= 193000 (about 80% of runs); main errors "time too far away". The reporter's value 0b1101111... (~9.6e11 ms) waits correctly on 64-bit Linux. No 32-bit build available (no multilib); Windows overflow (now.sec + ms/1000 > LONG_MAX, i.e. any delay over ~3.9e11 ms, ~12 years, for both the blocking and the script form) is from reading the code.
            • Cause generic/tclTimer.c: Tcl_AfterObjCmd and AfterDelay add ms/1000 to the long Tcl_Time.sec without a range check; on LP64 the usec carry can push the seconds past WIDE_MAX/1000 so TCL_TIME_DIFF_MS_CEILING overflows negative and AfterDelay breaks out at once.
            • Prototype prototypes/1802895fff.patch (against core-9-0-branch; 86 changed lines, of which ~40 are tests; new static TimeAddMs() that saturates the deadline, used by both paths; tests timer-11.3 (blocking form, in a child process, fails without the fix) and timer-11.4 (script form, matters on Windows)). timer.test passes 3/3.
            • Suggested ticket comment Fixed on main (9.1) by the TIP 723 timer rewrite, which reports "time too far away". 9.0/8.6 still overflow the long seconds field (immediately on Windows for delays over ~12 years, and on 64-bit for delays near 2^63 ms). Proposed fix for 9.0: saturate the deadline in a TimeAddMs() helper, tests timer-11.3/11.4.
            ⬜1777301fff2007TEA: --exec-prefix not set from --prefixpatchREPRODUCED (tclconfig trunk bed05771cd, 2026-08-20)

            Open ticket 1777301fff · 85. tclconfig · filed 2007-08-19

            🧪 Prototype check (2026-10-24)

            1777301fff.patch: VERIFIED

            • Base tclconfig trunk (tip bed05771cd), own checkout work-a1/v1/tclconfig
            • Before minimal TEA project (TEA_INIT, TEA_PATH_TCLCONFIG, TEA_LOAD_TCLCONFIG, TEA_PREFIX; configure regenerated with autoconf 2.73) against ~/tcltk/9.0/lib (TCL_EXEC_PREFIX=~/tcltk/9.0): ./configure --prefix=/some/path -> prefix=/some/path exec_prefix=~/tcltk/9.0
            • Patch applied cleanly (5 changed lines)
            • After --prefix=/some/path -> exec_prefix=/some/path; unchanged: no options -> both ~/tcltk/9.0, --exec-prefix=/ep -> prefix=~/tcltk/9.0 exec_prefix=/ep, both options -> as given
            • Test none possible (TEA has no test suite); the four configure runs above (script work-a1/v1/tea.sh) stand in for one
            Full verdict

            REPRODUCED (tclconfig trunk bed05771cd, 2026-08-20)

            • Platform build (tclconfig/TEA)
            • Tested minimal TEA configure.ac (TEA_INIT/TEA_LOAD_TCLCONFIG/TEA_PREFIX) in work-a2/tea against my main build (TCL_EXEC_PREFIX=/usr/local): --prefix=/some/path gives exec_prefix=/usr/local; with the prototype it gives /some/path. No-option, --exec-prefix only, and both-options cases unchanged.
            • Cause tclconfig tcl.m4: TEA_INIT sets exec_prefix=$prefix and exec_prefix_default=yes; TEA_PREFIX then sees exec_prefix_default=yes and overwrites exec_prefix with TCL_EXEC_PREFIX even when --prefix was given explicitly.
            • Prototype prototypes/1777301fff.patch (5 changed lines, against tclconfig.fossil trunk tcl.m4; take TCL_EXEC_PREFIX only when prefix was defaulted too; no test suite for TEA)
            • Suggested ticket comment Still true: with only --prefix, TEA_PREFIX replaces exec_prefix with TCL_EXEC_PREFIX. Proposed fix (tclconfig tcl.m4, TEA_PREFIX): use TCL_EXEC_PREFIX only when --prefix was also defaulted, otherwise exec_prefix follows --prefix as configure --help says.
            ⬜1743302fff2007[evalstats] auditpatchREPRODUCED (main; 9.0/8.6 same code) for the remaining review item; the original crash was fixed by dgp in 2007 (resolution already "Fixed")

            Open ticket 1743302fff · 47. Bytecode Compiler · filed 2007-06-26

            🧪 Prototype check (2026-10-24)

            1743302fff.patch: VERIFIED+TEST-ADDED (with the crash fix added; note evalstats is now also available in safe interps of debug builds: an information leak only, debug builds only)

            • Base main (tip d6dd6d0f71); verified in an out-of-tree --enable-symbols=all build (work-a5/dbg) plus the normal build
            • Before (debug build): puts [info commands evalstats]; interp create c; puts child:[c eval {info commands evalstats}],[c eval {info exists tcl_traceExec}] -> evalstats / child:,0
            • Patch applied cleanly; extended. With the original patch, interp create -safe s; s eval evalstats SEGFAULTS (EvalStatsCmd, tclExecute.c:10256). It is a pre-existing bug: the literal-table loop reads globalTablePtr->buckets[i] with a stale i instead of buckets[ui]. In the main interp it happens not to crash, but the patch makes evalstats reachable in every child/safe interp. Fixed i -> ui in the patch. Added test execute-13.1 and a constraint evalstats ([info commands evalstats]). Now 46 changed lines
            • After (debug): child:evalstats,1, safe: tcl_traceExec unset, s eval evalstats works; tests interp 384/0 failed, execute 158/0, coroutine 82/0, nre 29/0, safe 153/0, safe-zipfs 22/0. Normal build: interp, execute, coroutine, nre, safe 0 failed
            • Test added execute-13.1 (constraint evalstats, so it runs only in TCL_COMPILE_STATS builds); fails before / passes after in the debug build
            Full verdict

            REPRODUCED (main; 9.0/8.6 same code) for the remaining review item; the original crash was fixed by dgp in 2007 (resolution already "Fixed")

            • Platform build (TCL_COMPILE_STATS / TCL_COMPILE_DEBUG builds only)
            • Tested main built with --enable-symbols=all: evalstats and tcl_traceExec exist only in the first interp; interp create c; c eval {info commands evalstats} → empty.
            • Cause generic/tclExecute.c TclCreateExecEnv calls InitByteCodeExecution (create [evalstats], link tcl_traceExec) only once per process, guarded by execInitialized. tcl_traceCompile (tclCompile.c, traceInitialized) has the same first-interp-only pattern.
            • Prototype prototypes/1743302fff.patch (27 changed lines, no test: needs a debug build). Calls InitByteCodeExecution for every interp's first ExecEnv (not for coroutine ExecEnvs, which otherwise panic in postinit-1.10), drops execInitialized/execMutex, and unlinks/unsets tcl_traceExec in safe interps (keeps safe-6.1 passing). Verified in the debug build: interp, execute, coroutine, nre, safe, safe-zipfs tests pass; normal build: interp, coroutine, nre, safe pass.
            • Suggested ticket comment The crash was fixed long ago; the remaining issue is that [evalstats] and tcl_traceExec are only created in the first interpreter. Proposed fix: create them per interpreter (hidden from safe interps), patch available.
            ⬜1706840fff2007LITERAL_ON_HEAP flag unusedREPRODUCED (main, 9.0, 8.6): still no caller in the core passes LITERAL_ON_HEAP (grep of generic/, unix/, win/, pkgs/)

            Open ticket 1706840fff · 47. Bytecode Compiler · filed 2007-04-24

            Full verdict

            REPRODUCED (main, 9.0, 8.6): still no caller in the core passes LITERAL_ON_HEAP (grep of generic/, unix/, win/, pkgs/)

            • Platform generic (internal code cleanup)
            • Tested source grep of main, 9.0, 8.6
            • Cause generic/tclLiteral.c TclCreateLiteral/TclRegisterLiteral still carry the ownership-transfer branches (3 + 1 places) and tclCompile.h still defines the flag.
            • Prototype none (not a bug; TclRegisterLiteral is in the internal stubs table (tclInt.decls), so an extension such as tbcload could in principle pass the flag — removing it is a maintainer API decision, best done on main only)
            • Suggested ticket comment Still unused in 8.6/9.0/9.1. Since TclRegisterLiteral is an internal stub, the flag could be dropped on main (9.1/9.2) only; otherwise close as Won't Fix. This is a cleanup, not a bug.
            ✅1696642fff2007sigpipes in test suitepatchREPRODUCED (main; same Makefile target in 9.0/8.6) — only under gdb; not a Tcl bug

            Open ticket 1696642fff · 25. Channel System · filed 2007-04-08

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Fixed)

            • 🔧 2026-10-20 check-in b0abe7aa01 on gdb-ignore-sigpipe, in main: Fix 1696642fff: "make gdb-test" and "make gdb" tell gdb not to stop on SIGPIPE, which Tcl ignores.
            • 🔧 2026-10-26 check-in 4576331fff on trunk, in main, 9.0, 8.6: Fix 1696642fff: "make gdb-test" and "make gdb" tell gdb not to stop on SIGPIPE
            • 🔧 2026-10-26 check-in 59a640d4e4 on core-9-0-branch, in main, 9.0: Fix 1696642fff: "make gdb-test" and "make gdb" tell gdb not to stop on SIGPIPE
            • 🔧 2026-10-26 check-in 3dd530527b on core-8-6-branch, in main, 9.0, 8.6: Fix 1696642fff: "make gdb-test" and "make gdb" tell gdb not to stop on SIGPIPE
            • 💬 2026-10-20 serhiy.storchaka: Still the case on main. Not a Tcl bug: Tcl ignores SIGPIPE (unix/tclUnixInit.c) and handles EPIPE, but gdb stops on the signal anyway, e.g. in io-29.27. Proposed in branch b0abe7aa01: the gdb-test and gdb targets of unix/Makefile.in pass -ex "handle SIGPIPE nostop noprint" to gdb (the workaround from this ticket). With it, io-29.27 runs to completion under make gdb-test.
            • 💬 2026-10-26 jan.nijtmans: Fixed in 4576331fff, core-9-0-branch and core-8-6-branch Many thanks!
            Full verdict

            REPRODUCED (main; same Makefile target in 9.0/8.6) — only under gdb; not a Tcl bug

            • Platform unix (build/test tooling)
            • Tested main: gdb -batch -ex run --args ./tcltest ../tests/io.test -match io-29.27 stops with "Program received signal SIGPIPE"; with handle SIGPIPE nostop noprint the test passes.
            • Cause Tcl sets SIGPIPE to SIG_IGN (unix/tclUnixInit.c) and expects EPIPE; gdb intercepts the signal anyway, so make gdb-test stops at io-29.27 and similar tests.
            • Prototype prototypes/1696642fff.patch (8 changed lines, unix/Makefile.in: gdb-test and gdb pass -ex "handle SIGPIPE nostop noprint"; no test, verified with echo run | make gdb-test TESTFLAGS="-file io.test -match io-29.27")
            • Suggested ticket comment Not a Tcl bug: Tcl ignores SIGPIPE, but gdb stops on it. Proposed: make gdb-test/gdb pass handle SIGPIPE nostop noprint to gdb (the workaround from this ticket), then close.
            ⬜1695203fff2007[info frame] reports wrong linepatchREPRODUCED (main, 9.0, 8.6, 8.5) for direct (non-compiled) evaluation only; the continuation-line case from the 2007 comment is fixed (correct in 8…

            Open ticket 1695203fff · 45. Parsing and Eval · filed 2007-04-05

            🧪 Prototype check (2026-10-24)

            1695203fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before tcltest tip280.tcl (the ticket script) -> type source line 1 file .../tip280.tcl cmd {info frame 0} level 0
            • Patch applied cleanly (5 changed lines)
            • After same -> ... line 3 ...; tests info, basic, parse, subst, source, proc, trace, eval: all pass (0 failed)
            • Test existing info-42.0 (added by the patch) fails on the unpatched tip / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5) for direct (non-compiled) evaluation only; the continuation-line case from the 2007 comment is fixed (correct in 8.5.19 and later)

            • Platform generic
            • Tested main, 9.0, 8.6, 8.5. Wrong (word start line) for the main script file of tclsh (Tcl_FSEvalFileEx → TclEvalEx) and Tcl_EvalEx/testevalex; correct in procs, [eval], [uplevel], and [source] (these compile the script).
            • Cause generic/tclParse.c TclSubstTokens advances line past nested commands and counts backslash-newlines, but not past newlines in TCL_TOKEN_TEXT parts, so a [cmd] after a literal newline inside the same word gets the word's start line. The compiler (tclCompile.c TclCompileTokens) already does TclAdvanceLines for text tokens.
            • Prototype prototypes/1695203fff.patch (5 changed lines, test info-42.0, added at the end of info.test). Test fails before/passes after; full suite passes except the known 25 zipfs.test failures under TCL_LIBRARY.
            • Suggested ticket comment Still present for direct evaluation (main script file, Tcl_EvalEx): TclSubstTokens() does not count newlines in text tokens before a nested command. Compiled paths are correct. Proposed fix: TclAdvanceLines() for TCL_TOKEN_TEXT, test info-42.0.
            ⬜1669424fff2007trusted info leaks into safe interp via TIP 280REPRODUCED (main, 9.0, 8.6, 8.5)

            Open ticket 1669424fff · 20. [interp] · filed 2007-02-26

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5)

            • Platform generic
            • Tested interp create -safe s; s eval {info frame 0} from a sourced file returns type source ... file /full/path/of/parent/script.tcl on all versions with [info frame].
            • Cause TIP 280 location info (CmdFrame with type source and file) of a literal script in the parent is passed into the child by [interp eval] (literal compiled with the parent's location), and generic/tclCmdIL.c TclInfoFrame reports it without a Tcl_IsSafe check. The same applies to scripts a safe-base child [source]s through its alias (real paths, which safe.tcl otherwise hides behind tokens).
            • Prototype none (needs a decision on what a safe interp may see: drop/replace file for safe interps, or not propagate location across the interp boundary; either changes documented [info frame] keys)
            • Suggested ticket comment Still present in 8.6/9.0/9.1: [info frame] in a safe child reports the parent's full script path. Possible fix: in TclInfoFrame, for safe interps report type eval (or omit file) for frames whose location comes from outside the child; needs agreement on the policy.
            ⬜1667660fff2007[lsort -integer] not updated for bignumspatchREPRODUCED (main, 9.0, 8.6, 8.5)

            Open ticket 1667660fff · 17. Commands I-L · filed 2007-02-24

            🧪 Prototype check (2026-10-24)

            1667660fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before tclsh r166.tcl (lsort -integer {1 2 1000000000000000000000000}) -> integer value too large to represent
            • Patch applied cleanly (offset 8 lines)
            • After same -> 1 2 1000000000000000000000000; tests cmdIL (171, 0 failed), lsearch (165, 0), cmdMZ (97, 0)
            • Test existing cmdIL-1.16.1/1.16.2/1.16.3 fail before (all three) / pass after
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5)

            • Platform generic
            • Tested lsort -integer {1 2 1000000000000000000000000} gives "integer value too large to represent" everywhere. In 9.x [string is integer] accepts that value, and unsorted lsearch -integer already handles it, while lsearch -sorted -integer fails too.
            • Cause generic/tclCmdIL.c Tcl_LsortObjCmd keys -integer elements with TclGetWideIntFromObj into collationKey.wideValue.
            • Prototype prototypes/1667660fff.patch (45 changed lines, tests cmdIL-1.16.1-1.16.3), main only (uses Tcl_GetNumberFromObj): keep the wide fast path; on the first bignum switch to a new SORTMODE_BIGINTEGER (keys are the Tcl_Obj, compared with TclCompareTwoNumbers) and restart the collection loop (also resets numElements for -unique). Non-integers still give "expected integer but got ...". The 1M-int lsort -integer benchmark is unchanged (240 vs 250 ms, -O0). cmdIL/lsearch tests pass. lsearch -sorted -integer is not covered.
            • Suggested ticket comment Still present in 9.x, where [string is integer] accepts bignums. Proposed fix: lsort -integer falls back to bignum comparison when an element does not fit in a wide int, with no cost for the all-wide case (tests cmdIL-1.16.*). lsearch -sorted -integer has the same limitation.
            ⬜1655294fff2007command found in deleted namespaceREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            Open ticket 1655294fff · 21. [namespace] · filed 2007-02-08

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            • Platform generic
            • Tested dgp's script: after namespace delete of the current namespace, namespace which test still returns ::x::test while namespace exists ::x is 0, namespace which ::x::test is empty and ::x::test gives "invalid command name".
            • Cause generic/tclNamesp.c: a namespace with active call frames is only marked NS_DYING; lookups by name skip it, but relative lookups from a frame inside it (TclGetNamespaceForQualName with the current nsPtr) still find its commands, which is intended so that running code keeps working. [namespace which] then returns a name that cannot be resolved.
            • Prototype none (needs the namespace-deletion redesign msofer proposed in [2724403] approach #1, or a special case in [namespace which] for dying namespaces; dgp's 2007 patch is against 8.4)
            • Suggested ticket comment Still the same in 8.4–9.1: in a dying namespace, unqualified lookup still finds its commands, so [namespace which] returns a name that no longer resolves. A small option would be to make [namespace which] return "" when the command's namespace is NS_DYING; a full fix needs the deletion-order change from [2724403].
            ⬜1592212fff2006[interp alias] and [unknown] cannot work togetherREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            Open ticket 1592212fff · 45. Parsing and Eval · filed 2006-11-07

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            • Platform generic
            • Tested reporter's script on main (9.1.1 build c3bf6dd052), 9.0 tip, 8.6.18, 8.5.19, 8.4.20: namespace eval a x gives ::a, namespace eval a y (alias to x) gives ::.
            • Cause generic/tclInterp.c AliasNRCmd/AliasObjCmd invoke the target with TCL_EVAL_INVOKE (lookup in the global namespace), but when the target is missing, ::unknown runs and its uplevel 1 sees the caller's frame, so lookup context and execution context differ. Same family as 1590483fff and 1444291.
            • Prototype none (needs a design choice: FQ alias targets, or no unknown for alias targets; both change visible behaviour)
            • Suggested ticket comment Still the same in 9.1, 9.0 and 8.6. Fixing it needs a decision about how alias targets are resolved (see also &#91;1590483fff&#93;), not a local fix.
            ⬜1590483fff2006alias can break [uplevel 1 [info level 0]] idiomREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            Open ticket 1590483fff · 45. Parsing and Eval · filed 2006-11-04

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            • Platform generic
            • Tested reporter's corrected script (from the 2006-11-04 comment) prints x :: ::a on main, 9.0, 8.6.18, 8.5.19, 8.4.20.
            • Cause generic/tclInterp.c: the alias passes the unqualified target name x, so [info level 0] is x, and uplevel 1 re-resolves it in the caller's namespace ::a (where a::x shadows it).
            • Prototype none (msofer's approach, qualifying the target at alias creation, changes [info level 0] and error traces; dkf questioned in 2016 whether this should be supported at all)
            • Suggested ticket comment Still reproducible in 9.1/9.0/8.6. Fully qualifying the target would change &#91;info level 0&#93; output for every alias, so this needs a decision rather than a bug fix; suggest closing as Won't Fix unless someone wants to TIP it.
            💬1577730fff2006interp transfer leaks information when transfering channelsREPRODUCED (main, 9.0, 8.6) (RFE-like)

            Open ticket 1577730fff · 20. [interp] · filed 2006-10-15

            💬 Since triage (updated 2026-11-11)

            Status now: Open

            • 💬 2026-10-17 serhiy.storchaka: The heap address in socket channel names (Unix since 8.6.0, Windows since 8.6.2) is now tracked separately in d6c1d2f78e.
            Full verdict

            REPRODUCED (main, 9.0, 8.6) (RFE-like)

            • Platform generic
            • Tested main, 9.0 tip, 8.6.18: after interp transfer {} $chan $safe the safe interp sees file5 / sock5aa65bfc3110 in chan names, so the channel type shows. Socket names also contain the channel's heap address (a pointer leak, defeats ASLR in principle). Inside the safe interp fconfigure/chan configure are hidden, so nothing more leaks.
            • Cause channel names are type prefix + number/pointer (generic/tclIO.c, unix/tclUnixSock.c sock%p); interp share/transfer keep the name. No chan rename exists.
            • Prototype none (needs a new feature: opaque per-interp channel names or a chan rename)
            • Suggested ticket comment Still true in 9.1/9.0/8.6; socket names even include a heap address. Fixing it needs new API (opaque per-interp names or channel renaming), so this is an RFE.
            ⬜1563538fff2006append+env bugREPRODUCED (main, 9.0, 8.6)

            Open ticket 1563538fff · 07. Variables · filed 2006-09-22

            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform Windows
            • Tested Linux main, 9.0, 8.6.18: env names are case-sensitive there, so the ticket's case does not apply. Generic part confirmed on all three (and 8.5): append on an array element does not fire read traces (append a(k) 123 with a read trace that would set a(k) leaves it 123), which is what makes append env(a) skip the case-insensitive lookup in EnvTraceProc. Repro: win/1563538fff.tcl (child tclsh started with TRIAGE_X=VALUE runs append env(triage_x) 123; Windows verdict REPRODUCED if the result is 123).
            • Cause generic/tclEnv.c: Windows env names are looked up case-insensitively only in the read trace (TclGetEnv); append (TclPtrSetVar with TCL_APPEND_VALUE) does not fire read traces, so an element a that does not exist yet starts out empty.
            • Prototype none (needs a Windows test first; a fix would either fold env names to their existing spelling in TclSetupEnv/EnvTraceProc write branch or make append fire read traces on env)
            • Suggested ticket comment Still expected on Windows, since append does not fire read traces in 9.1/9.0/8.6 either. Needs a check on Windows (repro script available).
            Windows VM, run 2
            • Platform Windows
            • Verdict REPRODUCED (main, 9.0, 8.6)
            • Tested VM win/1563538fff.tcl: child started with TRIAGE_X=VALUE runs append env(triage_x) 123 → 123 on 9.1b1, 9.0.5 and 8.6.18 (expected VALUE123)
            • Cause generic/tclEnv.c looks up Windows names case-insensitively only in the read trace (TclGetEnv), and append does not fire read traces (tclVar.c TclPtrSetVarIdx only does with TCL_TRACE_READS, deliberately since [3057639]); the element triage_x starts empty and the write then replaces TRIAGE_X. PATH and COMSPEC are not affected: TclSetupEnv uppercases them.
            • Prototype none (needs a decision: fold names to the existing spelling in the write trace, or fire read traces for append on env)
            • Suggested ticket comment Still in 8.6, 9.0 and 9.1 on Windows: &#91;append env(a) x&#93; ignores an existing variable spelled A, because append does not fire env's read trace, which is where the case-insensitive lookup happens. PATH and COMSPEC are safe (normalized to upper case at startup).
            ✅1528500fff2006outputChannel incompatible with reflected channelspatchREPRODUCED (main, 9.0, 8.6)

            Open ticket 1528500fff · 34. tcltest Package · filed 2006-07-25

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Fixed)

            • 🔧 2026-10-20 check-in ca00499284 on tcltest-output-open-channel, in main: Fix 1528500fff: tcltest::outputChannel and errorChannel use an already open channel as is instead of opening a file with its name.
            • 🔧 2026-10-22 check-in 891b7e7921 on trunk, in main, 9.0: Many tcltest improvements, 773924ffff, 9f4feaaaa1, 811659ffff, 572400ffff, 8413e79232 and 1528500fff
            • 🔧 2026-10-22 check-in e44f1d0e87 on core-9-0-branch, in main, 9.0: Many tcltest improvements, 773924ffff, 811659ffff, 572400ffff, 8413e79232 and 1528500fff
            • 💬 2026-10-20 serhiy.storchaka: Still the case on main, 9.0 and 8.6: outputChannel and errorChannel open any name except stdout/stderr as a file, so with a reflected channel the output goes to a new file "rc0" in the current directory. Proposed fix in branch ca00499284: a name in [chan names] is used as the channel (tcltest does not close it); other names are still file names, which -outfile/-errfile rely on. New test tcltest-6.7.1.
            • 💬 2026-10-20 dgp: will have a look
            • 💬 2026-10-22 jan.nijtmans: Fixed in 891b7e7921. Not backported yet to 9.0 and 8.6 Don, if you have any remarks, feel free to follow-up!
            • 💬 2026-10-28 jan.nijtmans: Now also backported to core-9-0-branch and core-8-6-branch Closing. Many thanks!
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic (tcltest)
            • Tested tcltest::outputChannel $f with $f an open channel ([file tempfile]) creates a file named "file5" in the cwd and writes there, on main, 9.0.5 and 8.6.18; errorChannel does the same.
            • Cause library/tcltest/tcltest.tcl outputChannel/errorChannel: everything except stdout/stderr is [open]ed as a filename.
            • Prototype prototypes/1528500fff.patch (34 changed lines, test tcltest-6.7.1): if the argument is in [chan names], use it as is (not added to ChannelsWeOpened, so tcltest never closes it). tcltest.test passes; without the fix 6.7.1 fails and leaves file5 behind. A tcltest version bump is not included.
            • Suggested ticket comment Still present: outputChannel/errorChannel treat any name except stdout/stderr as a file name. Proposed fix: accept an already open channel as is (test tcltest-6.7.1).
            🔧1523045fff2006file cmd fails to copy to VFATpatchREPRODUCED (main, 8.6; simulated: LD_PRELOAD shim makes chmod() fail with EPERM in the target dir, as on VFAT; no real VFAT mount, that needs root/…

            Open ticket 1523045fff · 37. File System · filed 2006-07-15

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-19 check-in 6ffdcda1bf on copy-eperm-attrs, not merged: Fix 1523045fff and 731062ffff: [file copy] to a filesystem that cannot store the mode or the times (VFAT, CIFS with a fixed owner) failed with "operation no
            • 💬 2026-10-20 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1: after copying the data, [file copy] sets the mode and the times of the new file, and when that fails with EPERM (VFAT and similar filesystems) it removes the copy and reports an error. Copying a directory fails the same way and leaves a partial tree. Proposed fix in branch copy-eperm-attrs, also for 731062ffff: EPERM from chmod() and utime() of the newly created file is ignored…
            Full verdict

            REPRODUCED (main, 8.6; simulated: LD_PRELOAD shim makes chmod() fail with EPERM in the target dir, as on VFAT; no real VFAT mount, that needs root/udisks)

            • Platform unix
            • Tested main (own build), 8.6.18 with work-a5/vf/shim.so
            • Cause unix/tclUnixFCmd.c:CopyFileAtts returns TCL_ERROR when chmod() of the freshly created copy fails, and TclUnixCopyFile then unlinks the copy, so [file copy] fails although the data was copied. GNU cp keeps the file.
            • Prototype prototypes/1523045fff.patch (7 changed lines, no test: needs a filesystem that refuses chmod) — treats EPERM from the second chmod() as "filesystem cannot store the mode" and keeps the copy; repro passes, fCmd.test passes.
            • Suggested ticket comment Still present: CopyFileAtts() fails on EPERM from chmod() and the finished copy is deleted. Proposed fix: ignore EPERM there (we own the new file, so EPERM means the filesystem can't store the mode), as cp does.
            ⬜1452969fff2006performance: regexp 20sec delay from 8.0 to 8.3REPRODUCED (main, 9.0, 8.6, 8.4)

            Open ticket 1452969fff · 43. Regexp · filed 2006-03-18

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4)

            • Platform generic
            • Tested the attached jdReproduceIssue.tcl string (59201 chars): regexp {^(.*)=(.*)$} with capture vars takes 3.2 s on 8.4.20/8.6.18/9.0.5 and 5.2 s on the unoptimized main build; without captures 15–240 µs, with non-greedy (.*?) ~0.7–1.2 ms.
            • Cause generic/regexec.c: with capturing subexpressions the engine runs cdissect/ccondissect, which for a concatenation of two greedy .* tries split points one by one and re-runs the DFA for each, quadratic in the string length. Inherent to the Spencer engine.
            • Prototype none (an engine redesign, or a special case like the one dkf suggested in 2006, is far beyond a small fix)
            • Suggested ticket comment Still 3 s in 8.6/9.0/9.1 for the attached data: dissecting (.*)=(.*) is quadratic in the Spencer engine. Workaround: (.*?) or [string first]. No small fix; keep open or close as a known limitation.
            ⬜1281584fff2005-soname missing from GNU buildsREPRODUCED (main, 9.0, 8.6) — RFE/policy rather than a bug

            Open ticket 1281584fff · 53. Configuration and Build Tools · filed 2005-09-04

            Full verdict

            REPRODUCED (main, 9.0, 8.6) — RFE/policy rather than a bug

            • Platform build (Linux/GNU)
            • Tested readelf -d on my main build (wt-a2/unix/libtcl9.1.so), installed 9.0.5 and 8.6.18 libtcl*.so: no DT_SONAME on any of them.
            • Cause unix/tcl.m4 SC_CONFIG_CFLAGS, Linux*|GNU*... case sets SHLIB_LD='${CC} ${CFLAGS} ${LDFLAGS} -shared' with no -Wl,-soname; TCL_SHLIB_LD_EXTRAS is only set for Darwin. Distros (Debian, Fedora) add the soname by setting TCL_SHLIB_LD_EXTRAS/patching.
            • Prototype none (which soname to use — libtcl9.1.so vs libtcl9.1.so.0 as Debian does — is a packaging-policy decision, and it would change what distros already override)
            • Suggested ticket comment Still true for 8.6 and 9.x: the Linux libraries are linked without DT_SONAME. Distributions add it via TCL_SHLIB_LD_EXTRAS. If we want a default, -Wl,-soname,${TCL_LIB_FILE} (libtcl9.1.so) in the Linux section of tcl.m4 would be the minimal change; opinions welcome.
            ⬜1224888fff2005possible to have an empty info nameofexecutablepatchREPRODUCED (main, 9.0, 8.6)

            Open ticket 1224888fff · 36. Pathname Management · filed 2005-06-21

            🧪 Prototype check (2026-10-24)

            1224888fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before PATH=/usr/bin:/bin bash -c "exec -a mytclsh wt-a2/unix/tclsh ne.tcl" with ne.tcl = puts "<[info nameofexecutable]>" -> <>
            • Patch applied cleanly (23 changed lines, unchanged)
            • After same command -> <wt-a2/unix/tclsh>; tests unixFile.test: 7/7 passed
            • Test existing (updated by the patch) unixFile-1.2 and unixFile-1.5 fail before (C fix reverted: result empty) / pass after; Linux-specific (the expected value falls back to {} where /proc/self/exe is missing)
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform unix
            • Tested bash -c "exec -a mytclsh /path/tclsh x.tcl" → [info nameofexecutable] is empty on main, 9.0.5, 8.6.18
            • Cause unix/tclUnixFile.c:TclpFindExecutable only uses argv[0] and PATH; a bare argv[0] that is not on PATH gives an empty name. Branch bug-1224888 (dladdr / /proc/self/exe / _NSGetExecutablePath, last check-in d6a83ef702 2017-11-17) was never merged; main only uses dladdr() for the shared-library name.
            • Prototype prototypes/1224888fff.patch (23 changed lines, tests unixFile-1.2/1.5 updated) — last-resort readlink("/proc/self/exe") only when the PATH search failed, so the existing argv[0]/PATH precedence is unchanged; elsewhere it falls back to empty as before. Repro now gives the full path; unixFile.test passes. Does not cover macOS/BSD without /proc (the branch has those).
            • Suggested ticket comment Still present in 9.1. Proposed fix: when argv&#91;0&#93; has no slash and isn't found on the PATH, use readlink("/proc/self/exe") as a last resort (unixFile-1.2/1.5 updated). The older bug-1224888 branch, which also does macOS, was never merged.
            ⬜1214461fff2005encoding names case senstivitypatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4 on Unix)

            Open ticket 1214461fff · 10. Objects · filed 2005-06-03

            🧪 Prototype check (2026-10-24)

            1214461fff.patch: VERIFIED (after rework; the original patch was PATCH-INCOMPLETE because the mixed-case mac* encodings stayed case-sensitive)

            • Base main (tip d6dd6d0f71)
            • Before tcltest 1214.tcl -> UTF-8: 1 unknown encoding "UTF-8", same for Iso8859-1, CP1251, and encoding system Utf-8
            • Patch applied cleanly, but INCOMPLETE as written: it only retried with the lower-cased name, and its comment ("all encodings that come with Tcl have lower-case names") is wrong. The 15 mac* encodings (macRoman, macJapan, ...) have mixed-case names, so MacRoman/macroman/MACROMAN still failed with the original patch. Reworked (same idea, a case-insensitive fallback after the exact lookup): new static FindEncodingNameNoCase() looks first through the loaded encodings, then, only after the exact file load fails, through the encoding file map (FillEncodingFileMap, so only on the miss path), and loads the canonical name. Errors from LoadEncodingFile (e.g. encoding-11.6 "invalid file") are kept. 89 changed lines now (about 75 in C, most of them the helper and its header comment), which is over the size limit
            • After UTF-8/Iso8859-1/CP1251: 0 abc, encoding system Utf-8 -> utf-8; MacRoman/macroman/MACROMAN/MACJAPAN: 0 abc; an unknown name still gives unknown encoding "Nope" with errorCode TCL LOOKUP ENCODING Nope. Tests: encoding.test 232 passed/5 skipped/0 failed (also with make test, i.e. zipfs library), io.test 863/21/0, chanio.test 766/15/0, cmdAH.test 16971/30/0, utf.test 399/0/0, source.test 23/0/0, fCmd.test 260/45/0
            • Test existing encoding-1.4, extended with MACROMAN/macroman; fails before / passes after (encoding-1.5 passes on both, as it should)
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4 on Unix)

            • Platform generic
            • Tested encoding convertto UTF-8|Iso8859-1|CP1251 abc and encoding system Utf-8 give unknown encoding on main, 9.0, 8.6.18, 8.5.19, 8.4.20 (Linux). On Windows the .enc file lookup goes through the case-insensitive file system, so CP1251 works there (and creates a second encoding named CP1251); not re-checked on Windows with the 9.x zipfs library.
            • Cause generic/tclEncoding.c Tcl_GetEncoding: exact hash lookup, then LoadEncodingFile opens <name>.enc verbatim, so case sensitivity depends on the file system.
            • Prototype prototypes/1214461fff.patch (26 changed lines, tests encoding-1.4, encoding-1.5): if the exact name is not found and it contains upper-case letters, retry Tcl_GetEncoding with the lower-cased name before loading <name>.enc verbatim. Encodings registered with mixed-case names by extensions still match exactly. encoding.test, io.test, cmdAH.test, utf.test, fCmd.test, source.test pass.
            • Suggested ticket comment Still case-sensitive on Unix in 9.1/9.0/8.6. Proposed fix: Tcl_GetEncoding retries with the lower-cased name (all bundled encodings are lower case, IANA names are case-insensitive); tests encoding-1.4/1.5.
            ⬜1157137fff2005Cannot open file if parent directory unreadable on SolarisREPRODUCED (main, 8.6 on OpenIndiana)

            Open ticket 1157137fff · 37. File System · filed 2005-03-05

            Full verdict

            REPRODUCED (main, 8.6 on OpenIndiana)

            • Platform unix (Solaris)
            • Tested Linux main/8.6 with a real chmod 111 parent: works (Linux getcwd() is a syscall). With getcwd() failing: [open file] and [glob *] fail with "error getting working directory name", and [file exists file] silently returns 0.
            • Cause generic/tclIOUtil.c:Tcl_FSGetCwd re-checks the native cwd on every relative-path normalization and, when getcwd() fails, deliberately resets the cached cwd and errors (pwd error is tested). Relative opens therefore depend on getcwd().
            • Prototype none (needs a design decision: fall back to the cached cwd for normalization while keeping [pwd]'s error)
            • Suggested ticket comment Still in the code: any relative path needs Tcl_FSGetCwd(), which fails if getcwd() fails, as it does on old Solaris with an unreadable parent. Modern Linux is not affected; someone with Solaris/illumos should check whether getcwd() still fails there.
            • To run on OpenIndiana mkdir -p /tmp/d1/d2; echo J > /tmp/d1/d2/file; chmod 111 /tmp/d1; cd /tmp/d1/d2; echo 'puts [gets [open file]]' | tclsh
            Unix VMs, run 1
            • Platform unix (Solaris/illumos)
            • Verdict REPRODUCED (main, 8.6 on OpenIndiana)
            • Tested OpenIndiana, as a normal user: mkdir -p /tmp/d1/d2; echo J >/tmp/d1/d2/file; chmod 111 /tmp/d1; cd /tmp/d1/d2. /usr/bin/pwd fails ("cannot determine current directory: Permission denied"). Both main and 8.6: pwd errors with "error getting working directory name: permission denied", open file fails with the same error, file exists file returns 0, and glob * gives "couldn't determine the current working directory". Opening by absolute path works.
            • Cause generic/tclIOUtil.c:Tcl_FSGetCwd: illumos getcwd() still needs read access to every ancestor, and Tcl resolves every relative path through the cwd. When getcwd() fails, relative paths fail (and [file exists] wrongly says 0).
            • Prototype none (needs a design decision: e.g. pass relative native paths through unchanged, or fall back to "." when getcwd() fails, while [pwd] keeps its error)
            • Suggested ticket comment Still happens on OpenIndiana with main and 8.6: in a directory with an unreadable parent, &#91;open file&#93; and &#91;glob *&#93; fail with "error getting working directory name: permission denied", and &#91;file exists file&#93; returns 0, because every relative path goes through getcwd(). Fixing this needs a fallback for relative paths when the cwd is unknown.
            ⬜1146047fff2005Unexpected [glob] result, when match is partly inaccessablepatchREPRODUCED (main, 9.0, 8.6, 8.4)

            Open ticket 1146047fff · 37. File System · filed 2005-02-22

            🧪 Prototype check (2026-10-24)

            1146047fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before g1/boot/a, g1/root/x, chmod 0 g1/root; glob {g1/[rb]oot/*} -> couldn't read directory "g1/root/": permission denied; fileName-20.12 fails with an error
            • Patch applied cleanly (hunk offset 17 lines; 25 changed lines)
            • After fileName-20.12 passes; tests fileName 260/0 failed, cmdAH 16971/0, fCmd 260/0, unixFCmd 30/0, fileSystem 141/0
            • Test existing fileName-20.12 (unix notRoot) fails before / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4)

            • Platform unix
            • Tested main, 9.0.5, 8.6.18, 8.4.20 with gl/boot readable, gl/root mode 000: glob {gl/[rb]oot/*} and glob gl/*/* fail with "couldn't read directory gl/root/: permission denied" (with or without -nocomplain since 8.5; 8.4 returned empty with -nocomplain). sh/csh list gl/boot/*.
            • Cause generic/tclFileName.c:DoGlob stops at the first error while it recurses into the subdirectories matched by a wildcard, so one unreadable match discards everything.
            • Prototype prototypes/1146047fff.patch (25 changed lines, test fileName-20.12) — in the wildcard recursion loop, skip a subdirectory whose scan fails with EACCES; literal paths (glob root/*) still error. Behaviour change: glob {[r]oot/*} now returns {} (the Tcl 9 no-match result) instead of erroring. fileName/cmdAH/fCmd/unixFCmd tests pass; test fails without the fix.
            • Suggested ticket comment Still present: one unreadable directory matched by a wildcard makes the whole &#91;glob&#93; fail, even with -nocomplain. Proposed fix: skip such directories (EACCES) as sh does, keep the error for literal path components; test fileName-20.12.
            ⬜1111134fff2005[unload] and shared lib detachingpatchREPRODUCED (main; 8.6 has the same code in generic/tclLoad.c)

            Open ticket 1111134fff · 40. Dynamic Loading · filed 2005-01-27

            🧪 Prototype check (2026-10-24)

            1111134fff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before work-a4/v4unload.tcl (load dltest/tcl9pkgua.so into interps i and j; unload -keeplibrary in i; unload in j) -> mapped after: 0 (the .so is gone from /proc/self/maps although i asked to keep it)
            • Patch applied cleanly (21 changed lines)
            • After same script -> mapped after: 1; make test TESTFLAGS="-file 'unload.test load.test'" -> 58 total, 58 passed, 0 failed
            • Test existing unload-6.1 (in the patch) fails before (unpatched tclLoad.c: 57 passed, 1 failed, unload-6.1) / passes after
            Full verdict

            REPRODUCED (main; 8.6 has the same code in generic/tclLoad.c)

            • Platform generic
            • Tested main (own build) with dltest/tcl9pkgua.so: i loads, j loads, unload -keeplibrary in i, unload in j → the .so disappears from /proc/self/maps and from [info loaded], so j's unload detached the library i asked to keep.
            • Cause generic/tclLoad.c:UnloadLibrary decrements interpRefCount/safeInterpRefCount even when keepLibrary is set, so the next plain unload elsewhere sees a zero count and unloads the file.
            • Prototype prototypes/1111134fff.patch (21 changed lines, test unload-6.1) — with -keeplibrary, return before decrementing the refcount, so the library stays pinned in the process. unload.test and load.test pass; unload-6.1 fails without the fix. Semantics question for review: the library then stays loaded for the life of the process (which is what -keeplibrary asks for).
            • Suggested ticket comment Still present: &#91;unload -keeplibrary&#93; drops the interp's reference, so a later &#91;unload&#93; in another interp detaches the library. Proposed fix: don't drop the reference with -keeplibrary; test unload-6.1.
            ⬜1084013fff2004resolver: performance bugREPRODUCED (main, 9.0, 8.6) (code reading; performance only; same epoch bump in 9.0 and 8.6 tclResolve.c)

            Open ticket 1084013fff · 07. Variables · filed 2004-12-12

            Full verdict

            REPRODUCED (main, 9.0, 8.6) (code reading; performance only; same epoch bump in 9.0 and 8.6 tclResolve.c)

            • Platform generic
            • Tested code reading on main: Tcl_SetNamespaceResolvers (tclResolve.c) still bumps nsPtr->resolverEpoch, and proc/eval/subst bytecode is recompiled on an epoch mismatch (tclProc.c:1556, tclExecute.c, tclCompile.c). The second remark (resolver called per local at each proc call) is mostly addressed: resolved locals are cached and refreshed only when TCL_BYTECODE_RESOLVE_VARS is set (once after compile or an epoch change, tclProc.c InitResolvedLocals).
            • Cause resolverEpoch is overloaded to signal "compiled command shadowed" (tclNamesp.c TclResetShadowedCmdRefs) as well as resolver changes; a new command resolver can also change which commands get compiled inline, so recompiling is not always wasted.
            • Prototype none (performance RFE; nobody changes resolvers often)
            • Suggested ticket comment Still the same design in 9.1: changing a namespace resolver forces a recompile, and since a command resolver can change inlined commands this is partly needed. Local variable resolution is cached since 8.5. Suggest closing as Won't Fix or turning it into an RFE.
            ⬜1078030fff2004DdeUninitialize can be called with null ddeInstancepatchREPRODUCED by code reading (main, 9.0, 8.6); no visible symptom (DdeUninitialize(0) just fails)

            Open ticket 1078030fff · 28. dde Package · filed 2004-12-03

            🧪 Prototype check (2026-10-24)

            1078030fff.patch: VERIFIED

            • Base main (tip c092dae4af)
            • Before new helper ddehook.exe (scratch C program: an embedded Tcl loads tcl9dde15.dll into 4 interps, IAT-hooks its DdeNameService/DdeUninitialize imports, then Tcl_Finalize) -> no server name: DdeNameService(inst=0, ...) -> 0; DdeUninitialize(inst=0) -> 0 4 times; with dde servername: one correct unregister and DdeUninitialize(16777344) -> 1, followed by 3 more calls of each with inst=0. The bug is confirmed at run time (no visible symptom; win/1078030fff.tcl exits cleanly)
            • Patch applied cleanly, 14 changed lines, not changed
            • After ddehook -> no server name: no DDE calls at exit; named: exactly one DdeNameService(DNS_UNREGISTER) + one DdeUninitialize(inst) -> 1; win/1078030fff.tcl both modes rc=0; nmake test TESTFLAGS=-file winDde.test 50 total, 47 passed, 3 skipped, 0 failed (registry.test run directly: registry-2.7 fails, a test-text bug from b6340f48b4 that expects "configuration tcl::registry key"; it is unrelated to the patch)
            • Test none possible (no script-visible symptom; the check needs the IAT hook of an embedding program)
            Full verdict

            REPRODUCED by code reading (main, 9.0, 8.6); no visible symptom (DdeUninitialize(0) just fails)

            • Platform Windows
            • Tested win/tclWinDde.c on main/9.0/8.6: DdeExitProc is registered by Dde_Init for every interp that loads dde, and again by Initialize; it calls DdeNameService/DdeUninitialize unconditionally, so on the 2nd+ run (or when dde was never initialized) it passes ddeInstance == 0. Repro: win/1078030fff.tcl (loads dde in several interps, with and without servername, and checks a clean exit; the null calls need a breakpoint on DdeUninitialize to be seen).
            • Cause win/tclWinDde.c DdeExitProc has no ddeInstance/ddeServiceGlobal checks.
            • Prototype prototypes/1078030fff.patch (14 changed lines, no test; not compiled, Windows only): return early if ddeInstance == 0; unregister and free ddeServiceGlobal only if it was created, and reset it.
            • Suggested ticket comment Still true: DdeExitProc is registered once per interp plus once by Initialize and does not check ddeInstance. Proposed fix: return early when ddeInstance is 0 and unregister/free the service name only when it was registered.
            Windows VM, run 4
            • Platform Windows
            • Verdict REPRODUCED by code reading (main, 9.0, 8.6); no visible symptom (DdeUninitialize(0) just fails)
            • Tested VM win/1078030fff.tcl (dde loaded in 4 interps, with and without servername): clean exit on 9.1b1, 9.0.5, 8.6.18, and with the prototype
            • Cause win/tclWinDde.c DdeExitProc is registered per interp and by Initialize and has no ddeInstance/ddeServiceGlobal checks
            • Prototype prototypes/1078030fff.patch (existing, 14 changed lines, no test) applied to tcl-wt91 and built with MSVC 2022: compiles cleanly, the repro exits cleanly, winDde.test 47 passed / 3 skipped (debug) / 0 failed (same as unpatched), the 708166 blocked-server scenario still exits at once. Tested on Windows.
            • Suggested ticket comment Still true: DdeExitProc runs once per registration and calls DdeNameService/DdeUninitialize with ddeInstance 0 on later runs (harmless, but wrong). Proposed fix: return early when ddeInstance is 0 and unregister/free the service name only when it was registered (winDde.test passes on Windows).
            ⬜1059897fff2004[info procs] & [info commands]: inconsistent searchpatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            Open ticket 1059897fff · 17. Commands I-L · filed 2004-11-04

            🧪 Prototype check (2026-10-24)

            1059897fff.patch: VERIFIED+TEST-ADDED (doc + behaviour-pinning tests; the knownBug test is turned into a test of the documented rule, which closes the ticket by decision "document, not change")

            • Base main (tip d6dd6d0f71)
            • Before info.n: "Returns the names of all visible procedures", which suggests global procs are included like info commands. Behaviour (work-a1/1059897fff.tcl, 1059b.tcl; identical on main and 8.6 tip): in namespace eval ns with ::string_cmd, ::string_cmd2, ::ns::string_cmd and an imported op: info procs string* -> string_cmd; info procs -> op string_cmd; info procs ::string* -> ::string_cmd ::string_cmd2; info procs ::ns::* -> ::ns::op ::ns::string_cmd; info commands string_* -> string_cmd string_cmd2. Test info-15.8 expected the global procs too and was knownBug.
            • Patch 40 changed lines. doc/info.n info procs: names of the procedures in the current namespace incl. imported ones; qualifiers select a namespace and the results are qualified the same way; unlike info commands, an unqualified or empty pattern does not search the global namespace (use ::*). tests/info.test: info-15.8 rewritten in place to the documented behaviour (knownBug removed, result string_cmd; line count kept, since the info frame tests hard-code line numbers), new info-42.1 at the end.
            • After man page renders; troff warnings 21/21 (all pre-existing). info.test on main: 289 total, 287 passed, 2 skipped, 0 failed (info-15.8 now runs and passes instead of being skipped as knownBug).
            • Test info-15.8 (now pins current behaviour), added info-42.1 (in a namespace: info procs with no/unqualified pattern lists only local and imported procs, ::-qualified pattern finds the global one, namespace-qualified pattern gives qualified names, info commands contrast). Pins current behaviour: pass on main; both extracted into work-a1/p1059.test pass on 8.6 tip.
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            • Platform generic
            • Tested work-a1/1059897fff.tcl on main (c3bf6dd052), 9.0.5, 8.6.18, 8.5.19, 8.4.20: in a namespace with a local string_cmd, info procs string* -> string_cmd, info commands string_* -> string_cmd string_cmd2 (global one found too). Test info-15.8 is still knownBug.
            • Cause tclCmdIL.c InfoProcsCmd searches only the current (or named) namespace for an unqualified pattern, while InfoCommandsCmd also searches the global namespace (dkf's 2005 summary). info.n says "all visible procedures", which matches [info commands], not the code.
            • Prototype none (behaviour decision: either change [info procs] to also look in :: like [info commands], or document the current-namespace-only rule in info.n and drop info-15.8)
            • Suggested ticket comment Still the same in 9.1/9.0/8.6: &#91;info procs&#93; with an unqualified pattern looks only in the current namespace, &#91;info commands&#93; also in ::. Needs a decision: change info procs to match the "visible procedures" wording in info.n, or document the current behaviour and remove knownBug test info-15.8.
            ✅1056311fff2004async-4.3: huge memory hogpatchREPRODUCED (8.6: async-4.3 alone takes 383 MB RSS; main/9.0: 430 MB, but only on MSVC builds: since 8.7/9.0 the test doesn't run on Unix/MinGW at a…

            Open ticket 1056311fff · 04. Async Events · filed 2004-10-28

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Fixed)

            • 🔧 2026-10-17 check-in fdf3d4acf3 on async-test-msvc-constraint, in main: Fix the knownMsvcBug constraint in async.test, inverted since c0c61fa7cf: async-4.3 ran only on MSVC builds, where it is known to fail, and was skipped everyw
            • 💬 2026-10-17 serhiy.storchaka: Since c0c61fa7cf async-4.3 runs only on MSVC builds (the knownMsvcBug constraint was inverted), so it did not use memory elsewhere; fixed in branch async-test-msvc-constraint. The memory is for compiling the 1500000 commands. With 150000 they run in 0.7 ms here, less than the 1 ms after which [testasync marklater] fires the event, so the test would no longer check the delivery inside the sequence.
            Full verdict

            REPRODUCED (8.6: async-4.3 alone takes 383 MB RSS; main/9.0: 430 MB, but only on MSVC builds: since 8.7/9.0 the test doesn't run on Unix/MinGW at all, see below)

            • Platform generic (test suite)
            • Tested main (own build + Thread 3.0.7), 9.0 and 8.6 tcltest, /usr/bin/time maxrss.
            • Cause the test builds a 1.5M-command straight-line body (string repeat {;incr i;} 1500000). The parse/compile of that body costs ~400 MB. Also, tests/async.test:24 has testConstraint knownMsvcBug [expr {[tcl::build-info msvc]>0}], inverted compared with format.test/winFCmd.test (eq 0). The inversion dates from c0c61fa7cf (2021-07-23, "Re-implement testConstraint knownMsvcBug in terms of tcl::build-info"; before it was "not Travis-Windows"), so on 8.7/9.x async-4.3 runs only on MSVC builds, the one platform it was meant to skip.
            • Prototype prototypes/1056311fff.patch (4 changed lines, main; fixes the constraint and uses 150000 repetitions). async.test: 12/12 pass in 0.73 s, maxrss 52 MB (was 3.5 s / 430 MB). async-4.3 passed 40/40 at N=150000 (it even passes 40/40 at N=1000, 0/40 at N=0). For 8.6, only the repeat count applies.
            • Suggested ticket comment The constraint for async-4.3 is inverted on 8.7+ (&#91;tcl::build-info msvc&#93;>0), so the test runs only on MSVC builds. Proposed fix: invert it back and cut the body from 1.5M to 150000 commands: 52 MB instead of 430 MB, still 40/40 passes.
            🔧1030133fff2004more leave trace tests neededpatchREPRODUCED (main, 9.0, 8.6) (the save/restore now lives in tclTrace.c TclCheckExecutionTraces/TclCheckInterpTraces)

            Open ticket 1030133fff · 46. Traces · filed 2004-09-17

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-17 check-in 3cb057b2bd on leave-trace-tests, not merged: Add tests for 1030133fff: leave and leavestep traces must not change the result, -errorcode, -errorinfo or custom return options of the traced command (trace-
            • 💬 2026-10-17 serhiy.storchaka: The save and restore are now in TclCheckExecutionTraces() and TclCheckInterpTraces() (tclTrace.c). Existing tests catch their removal only partly, and none checks -errorcode, -errorinfo or custom return options. Proposed tests in branch leave-trace-tests: trace-42.1 to trace-42.5 check that leave and leavestep traces keep the result and return options of the traced command; they fail when Tcl_RestoreInterpState() …
            Full verdict

            REPRODUCED (main, 9.0, 8.6) (the save/restore now lives in tclTrace.c TclCheckExecutionTraces/TclCheckInterpTraces)

            • Platform generic
            • Tested mutation check on main: replacing Tcl_RestoreInterpState with Tcl_DiscardInterpState in TclCheckExecutionTraces makes trace-23.2/3, 25.6 and 28.2-28.4 fail, and in TclCheckInterpTraces trace-21.5, 23.x, 25.7, 25.10, 28.1-28.5 fail. So existing tests partly cover it, but none checks -errorcode or custom return options explicitly.
            • Cause missing tests
            • Prototype prototypes/1030133fff.patch (54 changed lines, tests only: trace-41.1..41.5; leave/leavestep traces that raise and catch an error and return custom options must not change the traced command's result, -errorcode, -errorinfo or custom options. They fail under the mutation above and pass on main, 9.0 tip, 8.6 tip)
            • Suggested ticket comment Proposed tests trace-41.1..41.5 check that leave and leavestep traces preserve the result, -errorcode, -errorinfo and custom return options; they fail when the Tcl_SaveInterpState/RestoreInterpState pair in tclTrace.c is removed.
            ⬜1017004fff2004[namespace import] / alias foil each others loop preventionpatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            Open ticket 1017004fff · 20. [interp] · filed 2004-08-26

            🧪 Prototype check (2026-10-24)

            1017004fff.patch: VERIFIED (only the ticket's first case; dgp's second variant, an import loop through an alias, is deliberately not handled, as the triage says)

            • Base main (tip d6dd6d0f71)
            • Before tcltest 1017.tcl (ticket's first script) -> interp alias {} ns::target {} alias returns 0 ns::target, then alias -> 1 {too many nested evaluations (infinite loop?)}
            • Patch applied cleanly
            • After same -> 1 {cannot define or rename alias "target": would create a loop}; tests interp.test 385 passed/7 skipped/0 failed, namespace.test 312/2/0, namespace-old.test 126/0/0, rename.test 19/0/0, safe.test 153/2/0, basic.test 118/29/0
            • Test existing interp-17.7 fails before / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            • Platform generic
            • Tested both scripts from the ticket on main, 9.0, 8.6.18, 8.5.19, 8.4.20: neither is rejected at creation; calling alias ends in "too many nested evaluations (infinite loop?)" (no crash).
            • Cause generic/tclInterp.c TclPreventAliasLoop follows only alias-to-alias links, not imported commands; Tcl_Import's loop check (tclNamesp.c) follows only import links, not aliases.
            • Prototype prototypes/1017004fff.patch (19 changed lines, test interp-17.7): TclPreventAliasLoop resolves each target through TclGetOriginalCommand, so the first case is now rejected with "would create a loop". The second case (import loop that goes through an alias) is left alone, since Tcl_Import would need to know alias internals, and as dgp notes the recursion limit catches it at runtime. interp, namespace, namespace-old, rename, safe and basic tests pass.
            • Suggested ticket comment Still in 9.1/9.0/8.6 (the recursion limit catches it at runtime). Proposed fix for the alias side: TclPreventAliasLoop follows &#91;namespace import&#93; links (test interp-17.7). The import-through-alias case is left to the recursion limit.
            ⬜994454ffff2004False [info exists] breaks [array anymore]patchREPRODUCED (main, 9.0, 8.6, 8.4)

            Open ticket 994454ffff · 46. Traces · filed 2004-07-20

            🧪 Prototype check (2026-10-24)

            994454ffff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before reporter's script at top level (tcltest 9944.tcl) -> after info exists a(dummy), array anymore a s-1-a -> couldn't find search "s-1-a" (compiled local arrays in a proc and global a in a proc are not affected)
            • Patch applied cleanly (64 changed lines, about 15 in C)
            • After -> array anymore returns 0 and the search survives; tests info.test 289 passed/3 skipped/0 failed, var.test 217/4/0, set-old.test 153/0/0, trace.test 286/4/0, env.test 26/6/0, upvar.test 70/0/0
            • Test existing info-42.1, 42.2, 42.4 fail before / pass after; info-42.3 (a read trace still creates the element) passes on both, as intended. (The triage's Prototype line names set-old-11.3/11.4; the patch actually has info-42.1..42.4, placed at the end of info.test as required)
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4)

            • Platform generic
            • Tested reporter script on main build (9.1.1), 9.0.5, 8.6.18, 8.4.20; also with a non-empty array
            • Cause generic/tclTrace.c:TclVarTraceExists and the INST_EXIST_* path in generic/tclExecute.c look up with createPart2=1, so a missing element is created (TclLookupArrayElement calls DeleteSearches on a new element) and then cleaned up; this is only needed when the array has read traces (env, user traces) that may define it.
            • Prototype prototypes/994454ffff.patch (36 changed lines, tests set-old-11.3, set-old-11.4) — look up without creating first, retry with createPart2=1 only if the array has a read trace (essentially msofer's 2006 idea). Full suite passes (except network/zipfs files not run).
            • Suggested ticket comment Still present in 9.1/9.0/8.6. &#91;info exists&#93; creates the missing element to fire read traces, which kills active searches. Proposed fix: create the element only if the array has read traces (prototype with tests set-old-11.3/11.4); searches on traced arrays are still terminated.
            ⬜953284ffff2004normalized path cache gets outdatedpatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4): the cached [file normalize] result goes stale. Since 8.5 [open] no longer reads the wrong file (8.4.20 did);…

            Open ticket 953284ffff · 37. File System · filed 2004-05-13

            🧪 Prototype check (2026-10-24)

            953284ffff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before filesystem-1.55 FAILED: wrong right (p1, normalized before relinking link -> wrong to link -> right, keeps the stale normalized path)
            • Patch applied cleanly (tclPathObj.c hunks offset 6 lines; 39 changed lines)
            • After right right; tests fileSystem 142/0 failed, fCmd 260/0, cmdAH 16971/0, fileName 259/0, unixFCmd 30/0, io 863/0; make test -file zipfs.test 498/0
            • Test existing filesystem-1.55 (hasLinks) fails before / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4): the cached [file normalize] result goes stale. Since 8.5 [open] no longer reads the wrong file (8.4.20 did); only the normalized value is wrong.

            • Platform generic
            • Tested main (own build), 9.0.5, 8.6.18, 8.5.19, 8.4.20 on Linux, using the reporter's scenario (relink lnk from wrong to right; p1 normalized before, p2 after)
            • Cause generic/tclPathObj.c:Tcl_FSGetNormalizedPath caches normPathPtr in the FsPath internal rep, but absolute paths (SetFsPathFromAny with transPtr==pathPtr) and joined paths (TclNewFSPathObj) carry filesystemEpoch 0, which means "never stale". Also, nothing bumps the epoch when a link is created.
            • Prototype prototypes/953284ffff.patch (39 changed lines, test filesystem-1.55). Tcl_FSLink calls Tcl_FSMountsChanged() after it creates a link, and Tcl_FSGetNormalizedPath stamps the current epoch whenever it computes normPathPtr. fileSystem/fCmd/cmdAH/fileName/unixFCmd pass, the full suite has 0 failures (zipfs checked via make test). Limitations: links changed by other processes or by [file delete]/[file rename] are still not noticed, and every [file link] now invalidates all cached path reps.
            • Suggested ticket comment Still reproducible on 9.0/8.6: &#91;file normalize&#93; keeps the old target after a link is replaced; &#91;open&#93; is fine since 8.5. Proposed fix: bump the filesystem epoch in Tcl_FSLink() and stamp the epoch on cached normalized paths, test filesystem-1.55. External changes to links can still not be detected.
            🔧903803ffff2004win build/testsuite should build pkga-pkgfREPRODUCED (main, 9.0, 8.6) by code reading — RFE

            Open ticket 903803ffff · 53. Configuration and Build Tools · filed 2004-02-24

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-20 check-in 4cf416d187 on win-dltest, not merged: Fix 903803ffff: the Windows builds (win/Makefile.in and win/makefile.vc) build the unix/dltest packages, so load.test, unload.test and pkgMkIndex.test run the
            • 💬 2026-10-20 serhiy.storchaka: Still the case on main, 9.0 and 8.6. Proposed in branch 4cf416d187: nmake -f makefile.vc tcltest and make tcltest (win/Makefile.in) also build the packages from unix/dltest into a dltest directory next to the test shell, as on Unix; the sources stay in unix/dltest. With them load.test, unload.test and pkgMkIndex.test run 78 tests instead of 39, all passing, with MSVC 2022, MSYS2 UCRT64 and a MinGW-w64 cross buil…
            Full verdict

            REPRODUCED (main, 9.0, 8.6) by code reading — RFE

            • Platform build (Windows)
            • Tested win/Makefile.in, win/makefile.vc and win/rules.vc in main and 8.6 have no dltest rules; load.test/unload.test look for [file dirname [info nameofexecutable]]/dltest and their $dll constraint is false on Windows, so load/unload of pkga..pkgua is untested there.
            • Cause unix/dltest/ has its own Unix-only configure/Makefile.in; nothing equivalent for Windows.
            • Prototype none (needs new nmake and MinGW build rules for ~10 DLLs, beyond a small fix)
            • Suggested ticket comment Still the case: the Windows builds don't build unix/dltest, so load.test and unload.test are skipped there. Would need dltest targets in win/Makefile.in and makefile.vc; patches welcome.
            ⬜886457ffff2004mode of transforms does not restrict channel opspatchREPRODUCED (main, 9.0, 8.6)

            Open ticket 886457ffff · 26. Channel Transforms · filed 2004-01-28

            🧪 Prototype check (2026-10-24)

            886457ffff.patch: VERIFIED (behaviour change for C transforms stacked with a narrower mask: needs core review, as triage said)

            • Base main (tip d6dd6d0f71)
            • Before work-a4/v4trans.tcl: read-only reflected transform pushed on an r+ file: puts $f xyz; flush $f -> write not supported by Tcl driver; write-only transform: gets -> read not supported by Tcl driver (the channel stays R/W)
            • Patch applied cleanly (38 changed lines). Rebuilt all objects including tclIO.h (tclIO/IOCmd/IOGT/IORChan/IORTrans/IOUtil/CmdAH/Zlib/UnixChan/Test).
            • After -> channel "file5" wasn't opened for writing / ... wasn't opened for reading; after chan pop writing works again; make test TESTFLAGS="-file 'io.test ioCmd.test iogt.test ioTrans.test chan.test chanio.test zlib.test socket.test ioUtil.test'" -> 2473 total, 2284 passed, 189 skipped, 0 failed
            • Test existing iortrans-bug-886457 fails before (Result was 0 {} 0 {}: the buffered puts does not even report an error) / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested chan push of a read-only reflected transform (methods initialize/finalize/read) on an r+ file: puts goes to the transform driver and fails with "write not supported by Tcl driver" instead of the generic "channel wasn't opened for writing"; the same for a write-only transform and gets. The same on main, 9.0, 8.6. ChnlStack.3 says "a read-write channel may become read-only after the Tcl_StackChannel call", but Tcl_StackChannel only checks that the mask overlaps.
            • Cause generic/tclIO.c:Tcl_StackChannel never applies mask to statePtr->flags; Tcl_UnstackChannel resets the flags to maxPerms.
            • Prototype prototypes/886457ffff.patch (38 changed lines: new Channel.permMask field in tclIO.h; Tcl_StackChannel clears the R/W flags the new layer doesn't allow; Tcl_UnstackChannel restores the flags allowed by the remaining layer after the close; test iortrans-bug-886457). The test fails without the fix. With it: io, ioCmd, iogt, ioTrans, chan, chanio, zlib and socket tests all pass, and the full suite has only the usual 25 zipfs failures (TCL_LIBRARY run). A visible behaviour change for C transforms registered with a narrower mask than the channel, so it needs core review.
            • Suggested ticket comment Still true: Tcl_StackChannel only checks that the mask overlaps the channel mode and never restricts the channel, so writes reach a read-only transform's driver. Proposed fix: record the mask per layer, clear the disallowed flags on stack and restore them on unstack (test iortrans-bug-886457).
            ✅845241ffff2003Unsetting env vars in slave interpspatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4) — [info exists env(X)] stays 1 in the parent after a child unsets env(X); the upvar part (stale value through…

            Open ticket 845241ffff · 08. Environment Variables · filed 2003-11-19

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Fixed)

            • 🔧 2026-10-19 check-in f1438aeb70 on env-stale-element, in main: Fix 3602289fff, 67fd4f973a and 845241ffff: after another interp unset an env variable, a read failed but left the old element, so [info exists ::env(X)] s
            • 🔧 2026-10-30 check-in 1f1167cd23 on trunk, in main, 9.0: Fix [845241]: Unsetting env vars in slave interps
            • 🔧 2026-10-30 check-in ec3f6936f2 on core-9-0-branch, in main, 9.0, 8.6: Fix [845241]: Unsetting env vars in slave interps
            • 🔧 2026-10-30 check-in 86ce5ad9dd on core-8-6-branch, in main, 9.0, 8.6: Fix [845241]: Unsetting env vars in slave interps
            • 💬 2026-10-19 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1: after a child interp unsets an env variable, [info exists env(X)] stays 1 in the parent. Proposed fix in branch env-stale-element, together with 3602289fff and 67fd4f973a: the read trace removes the stale element. Test env-11.1.
            • 💬 2026-10-30 jan.nijtmans: Fixed in 1f1167cd23, core-9-0-branch and core-8-6-branch Many thanks!
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4) — [info exists env(X)] stays 1 in the parent after a child unsets env(X); the upvar part (stale value through an alias) is gone in 9.x but still present in 8.6

            • Platform generic
            • Tested reporter script (with a private env var) on main, 9.0, 8.6, 8.5.19, 8.4.20
            • Cause generic/tclEnv.c:EnvTraceProc read branch returns "no such variable" when getenv() fails, but leaves the stale value in the parent's env array element; [info exists] ignores trace errors and sees a defined element
            • Prototype prototypes/845241ffff.patch (26 changed lines, test env-11.1) — on a failed getenv(), make the element undefined (keeping its traces; a plain Tcl_UnsetVar2 broke env-7.3). env/var/trace/interp/safe/info/upvar tests pass
            • Suggested ticket comment Still reproducible: the parent's env array keeps the stale element, so &#91;info exists&#93; says 1. Proposed fix: EnvTraceProc clears the element's value when the variable is gone from the environment (test env-11.1).
            ⬜824251ffff2003async client sockets block for DNSREPRODUCED (main, 9.0, 8.6)

            Open ticket 824251ffff · 27. Channel Types · filed 2003-10-15

            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested LD_PRELOAD shim making getaddrinfo("slow.test") sleep 2 s (work-a4/dns/slowgai.c): socket -async slow.test $port returns after 2001 ms on main, 9.0 and 8.6.
            • Cause TclCreateSocketAddress() (generic/tclIOSock.c) calls getaddrinfo() synchronously in Tcl_OpenTcpClient before the asynchronous connect starts. Only http has a workaround (http::config -threadlevel, 608f2fbab4, 8.7/9.0), which runs ::socket in a Thread.
            • Prototype none (needs a resolver thread in the socket driver on Unix and Windows; a design change, RFE-sized)
            • Suggested ticket comment Still true in 8.6/9.0/9.1: socket -async resolves the host synchronously, since TclCreateSocketAddress() calls getaddrinfo() before the async connect. http works around it with -threadlevel (608f2fbab4); a real fix needs a resolver thread in the socket driver.
            ✅811659ffff2003Skipped command should report skip in all casespatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            Open ticket 811659ffff · 34. tcltest Package · filed 2003-09-24

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Fixed)

            • 🔧 2026-10-20 check-in 1c58514c1d on tcltest-limitconstraints-verbose, in main: Fix 811659ffff: tests skipped by -limitconstraints are reported with -verbose s.
            • 🔧 2026-10-22 check-in 891b7e7921 on trunk, in main, 9.0: Many tcltest improvements, 773924ffff, 9f4feaaaa1, 811659ffff, 572400ffff, 8413e79232 and 1528500fff
            • 🔧 2026-10-22 check-in e44f1d0e87 on core-9-0-branch, in main, 9.0: Many tcltest improvements, 773924ffff, 811659ffff, 572400ffff, 8413e79232 and 1528500fff
            • 💬 2026-10-20 serhiy.storchaka: Still the case on main, 9.0 and 8.6. Skips by -skip and -match are silent by design (tests tcltest-3.* and 4.* check that, and reporting them would flood the output with -match), but tests skipped by -limitconstraints were not reported with -verbose s either, unlike those skipped by an unmet constraint. Proposed fix in branch 1c58514c1d (stacked on the other tcltest fixes): that branch of tcltest::Skipped uses _…
            • 💬 2026-10-28 jan.nijtmans: Now also backported to core-9-0-branch and core-8-6-branch Closing. Many thanks!
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            • Platform generic
            • Tested work-a1/811659ffff.tcl with -verbose s: only the constraint skip prints ++++ a2 SKIPPED: no; tests skipped by -skip, -match and -limitconstraints print nothing, on all versions.
            • Cause library/tcltest/tcltest.tcl tcltest::Skipped calls _noticeSkipped (which prints the SKIPPED line) only for unmet constraints. The -skip/-match silence is enshrined in tcltest-3.1..3.4 and 4.1..4.5 (they check that skipped names do not appear with -verbose ps), so it looks deliberate; the -limitconstraints branch is the real odd one (it also calls AddToSkippedBecause at every test level, unlike the other branches).
            • Prototype prototypes/811659ffff.patch (10 changed lines, test tcltest-5.2a): the -limitconstraints branch now uses _noticeSkipped, so -verbose s reports SKIPPED: userSpecifiedLimitConstraint. -skip/-match left silent (changing them breaks 9 existing tests and floods output with -match). tcltest.test passes (131/129 passed, 2 skipped); new test fails without the fix.
            • Suggested ticket comment -skip and -match skips are silent by design (tests tcltest-3.* and 4.* depend on it), but tests skipped by -limitconstraints were not reported with -verbose s either. Proposed fix: use _noticeSkipped there too, with test tcltest-5.2a.
            ✅773924ffff2003nested tests and -matchpatchREPRODUCED (main, 9.0, 8.6)

            Open ticket 773924ffff · 34. tcltest Package · filed 2003-07-18

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Fixed)

            • 🔧 2026-10-20 check-in e055230591 on tcltest-nested-match, in main: Fix 773924ffff: -match and -skip select top-level tests only; tests nested in a selected test always run.
            • 🔧 2026-10-22 check-in 891b7e7921 on trunk, in main, 9.0: Many tcltest improvements, 773924ffff, 9f4feaaaa1, 811659ffff, 572400ffff, 8413e79232 and 1528500fff
            • 🔧 2026-10-22 check-in e44f1d0e87 on core-9-0-branch, in main, 9.0: Many tcltest improvements, 773924ffff, 811659ffff, 572400ffff, 8413e79232 and 1528500fff
            • 💬 2026-10-20 serhiy.storchaka: Still the case on main, 9.0 and 8.6: 14 tests in tcltest.test fail when selected alone with -match (tcltest-21.4/5/6/10/11/12, 24.9/10/13/15/17/18/20, 25.3), because their nested tests do not match. Proposed fix in branch e055230591: -match and -skip are applied only to top-level tests, so tests nested in a selected test always run. Documented in tcltest.n, new test tcltest-28.1. All 131 tcltest.test tests now p…
            • 💬 2026-10-22 jan.nijtmans: Fixed in 891b7e7921, not yet backported to 9.0 and 8.6
            • 💬 2026-10-28 jan.nijtmans: Now also backported to core-9-0-branch and core-8-6-branch Closing. Many thanks!
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested tcltest.test -match tcltest-21.4 fails on main, 9.0 tip and 8.6 tip (nested tcltest-21.4.0 is skipped by -match). Running each of the 130 tcltest.test tests alone with -match: 14 fail on main (tcltest-21.4/5/6/10/11/12, 24.9/10/13/15/17/18/20, 25.3).
            • Cause library/tcltest/tcltest.tcl tcltest::Skipped applies -skip and -match to tests at every testLevel, so a selected test's nested tests are dropped.
            • Prototype prototypes/773924ffff.patch (22 changed lines, test tcltest-28.1): -skip/-match are applied only to top-level tests (testLevel 1); nested tests always run. After it all 130 tcltest.test tests pass when run alone with -match, full tcltest.test passes; new test fails without the fix.
            • Suggested ticket comment Still true: 14 tests in tcltest.test fail when selected alone with -match. Proposed fix: apply -match/-skip only to top-level tests, so a selected test's nested tests always run; test tcltest-28.1.
            ⬜768678ffff2003open "|prog >file" throws error but still executes progpatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            Open ticket 768678ffff · 25. Channel System · filed 2003-07-09

            🧪 Prototype check (2026-10-24)

            768678ffff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before work-a3/v768.tcl: open "|echo foo >bar768" r -> "can't read output from command: standard output was redirected", yet bar768 exists: 1; open "|cat << hello768 >bar768" w -> stdin error, and the file exists too
            • Patch applied cleanly (78 changed lines)
            • After both errors are unchanged and bar768 exists: 0. The allowed forms still work (w with >file, r with 2>@stderr, [list | sh -c ... 2> file]). Tests ioCmd (278/0 failed), exec (98 passed, 0 failed; the 101 skips are win/haveTestApp etc.), io, chanio: 0 failed
            • Test existing iocmd-11.5 and 11.6 (from the patch) fail on the unpatched tip / pass after
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            • Platform generic
            • Tested open "|echo foo >bar" r errors with "standard output was redirected", yet bar gets created and written; the same for >>, >&, r+. With w, open "|cat << x" w runs cat (its output shows up) before the stdin error. Same on all versions tested.
            • Cause generic/tclPipe.c:Tcl_OpenCommandChannel checks TCL_ENFORCE_MODE only after TclCreatePipeline() has opened the redirect files and spawned the processes (it then just detaches the pids).
            • Prototype prototypes/768678ffff.patch (78 changed lines, about 35 of them code, the rest a function header comment and tests iocmd-11.5/11.6; over the ~60 guideline only because of comments and tests). A new static RedirectedStreams() pre-scans argv for <…/>… operators (skipping the operand of bare operators and 2>… stderr redirections) so the mode error is raised before anything is opened or started. TclCreatePipeline's signature (used by BLT) is unchanged and the old post-checks stay as a fallback. ioCmd.test and exec.test pass. The new tests fail without the fix (file created).
            • Suggested ticket comment Still true in 8.6/9.x: the access-mode check runs after the pipeline has been started. Proposed fix: scan the arguments for stdin/stdout redirections in Tcl_OpenCommandChannel before calling TclCreatePipeline, which keeps the BLT-visible signature (tests iocmd-11.5/11.6).
            ⬜764634ffff200364bit support for Tru64Unix 5.1ApatchREPRODUCED (main, 9.0, 8.6) by code reading — cosmetic configure warning

            Open ticket 764634ffff · 52. Portability Support · filed 2003-07-02

            🧪 Prototype check (2026-10-24)

            764634ffff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before configure run in work-a2/v2/osf with a fake uname first in PATH (work-a2/v2/fakebin/uname: OSF1 / V5.1 / alpha), configure --enable-64bit -> checking system version... OSF1-V5.1 then WARNING: 64bit support being disabled -- don't know magic for this platform
            • Patch applied cleanly (4 changed lines, unchanged); ~/tcltk/autoconf-2.73/bin/autoconf -f on the patched tcl.m4 regenerates a configure byte-identical to the patched one
            • After same fake-OSF1 configure -> no warning; normal Linux configure + make still OK
            • Test none possible (build system; no Tru64 machine); the fake-uname configure run is what was checked
            Full verdict

            REPRODUCED (main, 9.0, 8.6) by code reading — cosmetic configure warning

            • Platform build (Tru64/Alpha)
            • Tested unix/tcl.m4 SC_CONFIG_CFLAGS: the OSF1-V* case never sets do64bit_ok, so --enable-64bit gives "64bit support being disabled -- don't know magic for this platform" although Alpha is always LP64. Can't run on Tru64.
            • Cause unix/tcl.m4 OSF1-V* case lacks do64bit_ok=yes.
            • Prototype prototypes/764634ffff.patch (4 changed lines: 2 in tcl.m4, 2 in configure regenerated with autoconf 2.73; sh -n configure ok; untested on Tru64)
            • Suggested ticket comment Proposed fix: set do64bit_ok=yes in the OSF1-V* case of tcl.m4, since Tru64 on Alpha is always 64-bit, so --enable-64bit no longer warns. Untested on Tru64 (nobody has one); fine to just close as out of date instead.
            🔧731062ffff2003file copy reports "not owner" although rights sufficepatchREPRODUCED (main, 9.0, 8.6), emulated: an LD_PRELOAD shim made chmod()/utime() return EPERM on the target dir, which is what a vfat mount with uid=…

            Open ticket 731062ffff · 37. File System · filed 2003-05-01

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-19 check-in 6ffdcda1bf on copy-eperm-attrs, not merged: Fix 1523045fff and 731062ffff: [file copy] to a filesystem that cannot store the mode or the times (VFAT, CIFS with a fixed owner) failed with "operation no
            • 💬 2026-10-20 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1, as ferrieux's strace showed: the data is copied, but chmod() or utime() of the new file fails with EPERM when the filesystem has a fixed owner (vfat mounted with uid=, CIFS), and then the copy is removed. Proposed fix in branch copy-eperm-attrs, together with 1523045fff: ignore EPERM there and keep the copy, as cp does.
            Full verdict

            REPRODUCED (main, 9.0, 8.6), emulated: an LD_PRELOAD shim made chmod()/utime() return EPERM on the target dir, which is what a vfat mount with uid=65534 does for non-owners

            • Platform unix
            • Tested main (own build), 9.0.5, 8.6.18 on Linux with work-a5/731062-shim.c and work-a5/731062.tcl; cp(1) succeeds under the same shim
            • Cause unix/tclUnixFCmd.c:TclUnixCopyFile copies the data, then CopyFileAtts() does chmod()+utime(). If that fails (EPERM: we do not own the new file), it unlinks the copy and the whole [file copy] fails. cp without -p does not copy the mode or times, and cp -p only warns.
            • Prototype prototypes/731062ffff.patch (5 changed lines, no test: needs a filesystem with forced ownership). On EPERM from attribute copying, keep the file and succeed. fCmd/unixFCmd/fileSystem/cmdAH pass. This changes behaviour (mode/mtime silently not preserved), so it needs a decision; an alternative is to keep failing but leave the copied file in place.
            • Suggested ticket comment Still present: &#91;file copy&#93; deletes the copy and fails with EPERM when it cannot chmod/utime the new file (e.g. on vfat mounted with a fixed uid). Proposed fix: keep the copy when attribute copying fails with EPERM, as cp does.
            ⬜703709ffff2003'TCL_REG_CANMATCH' flag returns incorrect matchesREPRODUCED (main, 9.0 tip, 8.6 tip)

            Open ticket 703709ffff · 43. Regexp · filed 2003-03-14

            Full verdict

            REPRODUCED (main, 9.0 tip, 8.6 tip)

            • Platform generic
            • Tested main tcltest build and the 8.6 tip tcltest: testregexp -xflags -- c .x "asd asd" v gives 0 0 (expected 6); reg-32.4–32.9 (knownBug) all fail on main (7 failures).
            • Cause generic/regexec.c reports the DFA "cold start" position (from shortest()) as the canmatch position; as tgl explained in 2015, with overlapping partial matches there is no cold start, so the value is meaningless outside the engine. Used by Tk's text widget regexp search (tkText.c, TCL_REG_CANMATCH).
            • Prototype none (needs tracking match-start per NFA state; tgl judged it impractical)
            • Suggested ticket comment Still present; reg-32.4–32.9 (knownBug) fail on 9.1. Per tgl's analysis the cold-start position can't give a correct canmatch answer, so a fix needs engine work. Perhaps document TCL_REG_CANMATCH as a heuristic and leave open.
            ⬜696707ffff2003trace command on 'env' bugREPRODUCED (main, 9.0, 8.6, 8.4) — but it follows from documented trace semantics (traces on a variable are disabled while one of its traces runs)

            Open ticket 696707ffff · 08. Environment Variables · filed 2003-03-03

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4) — but it follows from documented trace semantics (traces on a variable are disabled while one of its traces runs)

            • Platform generic
            • Tested reporter script on main, 9.0, 8.6, 8.4
            • Cause traces fire newest first, so the user's read trace runs before Tcl's own EnvTraceProc; its [set ::env($e) ...] does not fire EnvTraceProc's write trace (traces are inactive during a trace), so setenv() is never called, and then EnvTraceProc's read branch finds no such variable in the environment and fails the read. Afterwards env(TEST2) holds a value that is not in the real environment ([info exists] 1). With the 845241 prototype the stale value is cleared instead ([info exists] 0), which is more consistent but still fails the read
            • Prototype none (would need trace-order or nested-trace changes; a "keep the Tcl-side value" rule in EnvTraceProc conflicts with the 845241 fix)
            • Suggested ticket comment Still the case: the user's read trace runs before Tcl's env trace, and a write made inside a trace does not fire the env write trace, so the variable never reaches the environment and the env read trace reports it missing. This follows from the documented rule that traces are disabled while a trace runs; suggest documenting it in tclvars.n (env) rather than changing the trace machinery.
            ⬜652765ffff2002errors disappear in trace callbackspatchREPRODUCED (main, 9.0, 8.6) — errors in rename/delete command-trace callbacks are swallowed; still undocumented

            Open ticket 652765ffff · 46. Traces · filed 2002-12-12

            🧪 Prototype check (2026-10-24)

            652765ffff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before v4docs3.tcl: trace add command p delete {apply {args {error boom}}}; catch {rename p {}} -> 0 {}, command gone; same for a rename trace (0 {}, q2 exists); main and 8.6.
            • Patch applied cleanly (4 changed lines)
            • After groff 0 warnings (0 before)
            • Test none possible (docs only)
            Full verdict

            REPRODUCED (main, 9.0, 8.6) — errors in rename/delete command-trace callbacks are swallowed; still undocumented

            • Platform generic
            • Tested rename and delete trace that throws, on main, 9.0, 8.6: [rename] returns 0, no errorInfo
            • Cause generic/tclBasic.c CallCommandTraces ignores the callback result by design; Tcl_DeleteCommandFromToken cannot return an error (dgp 2005: reporting needs a TIP)
            • Prototype prototypes/652765ffff.patch (4 changed lines, docs only) — doc/trace.n says errors from command (rename/delete) traces are ignored. Reporting them (e.g. via Tcl_BackgroundException as dkf proposed) is a behaviour change for a TIP
            • Suggested ticket comment Still the case in 9.1/9.0/8.6. Proposed doc fix stating that errors in rename/delete traces are ignored; reporting them as background errors would need a TIP.
            🔧623585ffff2002[file mkdir] does not use default aclpatchREPRODUCED (main, 8.6; same code in 9.0)

            Open ticket 623585ffff · 37. File System · filed 2002-10-15

            🔧 Since triage (updated 2026-11-11)

            Status now: Open

            • 🔧 2026-10-17 check-in 5d7e74108f on mkdir-no-umask-call, not merged: Fix cd609219f5 and 623585ffff: [file mkdir] read the umask with umask(0) + umask(mode), which threads could leave at 0 for the whole process, and masked the
            • 💬 2026-10-17 serhiy.storchaka: Still present on Linux: DoCreateDirectory() applies the umask to the mode itself, and when the parent has a default ACL (so the kernel does not apply the umask) this mode cuts down the inherited permissions. Proposed fix in branch mkdir-no-umask-call, with cd609219f5: call mkdir(path, 0777) and let the kernel apply the umask or ACL, as mkdir(1) does. Test unixFCmd-21.1.
            Full verdict

            REPRODUCED (main, 8.6; same code in 9.0)

            • Platform unix
            • Tested main (own build), 8.6.18 on Linux ext4 with setfacl -d u::rwx,g::rwx,o::rwx and umask 022. mkdir(1) gives group::rwx other::rwx, [file mkdir] gives r-x/r-x.
            • Cause unix/tclUnixFCmd.c:DoCreateDirectory masks the umask into the mode passed to mkdir(). When the parent has a default ACL the kernel ignores the umask, and the mode argument then cuts down the inherited ACL. The masking is redundant otherwise (mkdir applies the umask anyway, and the "force owner rwx" part never worked: umask 0700 still gives 0077). The umask(0)/umask(old) pair is also racy in threaded programs.
            • Prototype prototypes/623585ffff.patch (24 changed lines, test unixFCmd-21.1, constraint defaultAcl = setfacl works). Calls mkdir(path, 0777) and lets the kernel apply umask/ACL. Results for umask 077/022/0700 are unchanged; unixFCmd/fCmd/fileSystem/cmdAH pass.
            • Suggested ticket comment Still present on Linux: DoCreateDirectory() pre-applies the umask, which overrides the parent's default ACL. Proposed fix: call mkdir(path, 0777) and let the kernel apply umask/ACL (this also removes a umask() race between threads), test unixFCmd-21.1.
            ⬜573292ffff2002prefix not honoured on make installpatchREPRODUCED (main, 9.0, 8.6) — by design; unix/README is wrong

            Open ticket 573292ffff · 53. Configuration and Build Tools · filed 2002-06-24

            🧪 Prototype check (2026-10-24)

            573292ffff.patch: VERIFIED (wording corrected)

            • Base main (tip d6dd6d0f71)
            • Before make -n install prefix=/v4/p -> libtcl, tclsh, tclConfig.sh still to /usr/local/..., but headers to /v4/p/include; make -n install exec_prefix=/v4/ep -> tclsh to /v4/ep/bin, lib still /usr/local/lib; make -n install DESTDIR=/v4/stage -> everything under /v4/stage/usr/local.
            • Patch ported (rewording): the original said setting prefix/exec_prefix on the make command line "does not change them", which is wrong (includedir/mandir follow prefix, bindir follows exec_prefix, libdir is expanded by configure). Now: "Do not set prefix or exec_prefix on the make command line: configure has already expanded them in some of the directory variables (e.g. libdir), so only part of the files would move." DESTDIR sentence kept. 8 changed lines.
            • After text matches the make -n results above
            • Test none possible (docs only)
            Full verdict

            REPRODUCED (main, 9.0, 8.6) — by design; unix/README is wrong

            • Platform build (unix)
            • Tested main build in wt-a2: make -n install-binaries prefix=/X/pfx still installs to /usr/local/bin and /usr/local/lib (configure writes exec_prefix = /usr/local and libdir = /usr/local/lib expanded into the Makefile). make install DESTDIR=/X/stage correctly installs to /X/stage/usr/local/...; overriding prefix, exec_prefix and libdir together also works. unix/README step (f) in main, 9.0 and 8.6 still says you can override by "modifying the prefix and exec_prefix variables in the Makefile".
            • Cause unix/configure.ac expands exec_prefix/libdir at configure time; DESTDIR (INSTALL_ROOT) is the supported way to stage an install.
            • Prototype prototypes/573292ffff.patch (7 changed lines, unix/README only; docs fix pointing to DESTDIR)
            • Suggested ticket comment make install DESTDIR=/staging/dir is the supported way to stage an install (for stow or packaging); setting prefix at make time is not. Proposed fix: correct unix/README, which still claims prefix/exec_prefix can be overridden in the Makefile.
            ✅572400ffff2002constraint exprs defeat limitConstraintspatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            Open ticket 572400ffff · 34. tcltest Package · filed 2002-06-22

            ✅ Since triage (updated 2026-11-11)

            Status now: Closed (Fixed)

            • 🔧 2026-10-20 check-in 241d536f30 on tcltest-limitconstraints-expr, in main: Fix 572400ffff: with -limitconstraints, a constraint expression is satisfied only if it is true and names one of the selected constraints.
            • 🔧 2026-10-22 check-in 891b7e7921 on trunk, in main, 9.0: Many tcltest improvements, 773924ffff, 9f4feaaaa1, 811659ffff, 572400ffff, 8413e79232 and 1528500fff
            • 🔧 2026-10-22 check-in e44f1d0e87 on core-9-0-branch, in main, 9.0: Many tcltest improvements, 773924ffff, 811659ffff, 572400ffff, 8413e79232 and 1528500fff
            • 💬 2026-10-20 serhiy.storchaka: Still the case on main, 9.0 and 8.6: -limitconstraints forces the unselected constraints to 0, so expressions like !foo or $::tcl_version > 0 become true. With -constraints knownBug -limitconstraints 1 the whole test suite runs 76 tests. Proposed fix in branch 241d536f30 (stacked on the other tcltest fixes): in that mode an expression constraint must also name one of the selected constraints. The knownBug run th…
            • 💬 2026-10-20 dgp: Thanks for the work on this. I'm putting it on my list of fixes to examine. Just a comment that for me a desired long-range fix for these issues would be to eliminate support for constraint expressions. They've always been short of working well, reliably and robustly, and the gain from trying to offer it is just not compelling. At least the ideal tcltest, version 3 in my head does not have it.
            • 💬 2026-10-20 dgp: One consequence of constraint expression support is you cannot have an expression with a name anything like this-should-only-run-in-this-situation because tcltest insists on thinking that is a constraint subtraction. Just not good.
            • 💬 2026-10-22 jan.nijtmans: Fixed in 891b7e7921, not backported yet to 9.0 and 8.6 I don't think I would want constraints with minus-signs in it. What I see now is using camelCase, like thisShouldOnlyRunInThisSituation
            • 💬 2026-10-28 jan.nijtmans: Now also backported to core-9-0-branch and core-8-6-branch Closing. Many thanks!
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            • Platform generic
            • Tested work-a1/572400ffff.tcl with -constraints knownBug -limitconstraints 1: tests constrained by !foo and $::tcl_version > 0 run on all versions. Full suite on main with -constraints knownBug -limitconstraints 1: 70 tests run (20 pass, 50 fail) before, 23 (7 pass, 16 fail) after the prototype.
            • Cause library/tcltest/tcltest.tcl: -limitconstraints is implemented by forcing all unlisted constraints to 0 (ClearUnselectedConstraints/testConstraint), which makes negated or arbitrary expressions true; Skipped's expression branches never check the selected list.
            • Prototype prototypes/572400ffff.patch (32 changed lines, test tcltest-5.2b): in limitConstraints mode an expression constraint is satisfied only if it is true and mentions at least one of the selected constraints. tcltest.test passes; new test fails without the fix.
            • Suggested ticket comment Still true: with -limitconstraints, constraints like !foo become true because unselected constraints are forced to 0. Proposed fix: an expression constraint must also name one of the -constraints to run in that mode; test tcltest-5.2b.
            ⬜556837ffff2002nested tests are not counted correctlyREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            Open ticket 556837ffff · 34. tcltest Package · filed 2002-05-16

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            • Platform generic
            • Tested reporter's script (work-a1/556837ffff.tcl): Total 1 ... Failed 1 on all versions.
            • Cause tcltest counts only testLevel 1 tests (by design); documented under KNOWN ISSUES in tcltest.n. tcltest.test itself relies on it (its nested tests are expected to fail without failing the outer run).
            • Prototype none (documented limitation; counting nested tests would break tcltest's own test suite and change totals for existing users)
            • Suggested ticket comment Unchanged in 9.1/9.0/8.6, and documented in tcltest.n KNOWN ISSUES. tcltest.test depends on nested tests not being counted, so this is unlikely to change; suggest closing as Won't Fix (use child interps for meta-tests).
            💬553544ffff2002spaces in paths break stuffREPRODUCED (main, core-8-6-branch)

            Open ticket 553544ffff · 53. Configuration and Build Tools · filed 2002-05-08

            💬 Since triage (updated 2026-11-11)

            Status now: Open

            • 💬 2026-10-20 serhiy.storchaka: Rechecked on main (unix): an out-of-tree build directory with a space and a --prefix with a space now work (build, install, tests). Only a space in the source directory still fails: configure succeeds, but make stops at once with "No rule to make target '.../macosx/configure.ac'", since unix/Makefile.in uses the source paths as make targets and prerequisites in some 220 rules. Building through a symlink to the sou…
            Full verdict

            REPRODUCED (main, core-8-6-branch)

            • Platform build (unix)
            • Tested tarball of main (c3bf6dd052) and of core-8-6-branch extracted under work-a2/sp ace/; configure succeeds, make fails at once: "No rule to make target 'ace/tcl/macosx/configure.ac'" (logs in work-a2/sp ace/).
            • Cause unix/Makefile.in uses $(TOP_DIR)/$(UNIX_DIR) etc. as make targets/prerequisites (e.g. the configure regeneration rules); GNU make cannot handle spaces in target names, and many shell commands use the paths unquoted.
            • Prototype none (would need quoting throughout Makefile.in/tcl.m4/tclConfig.sh and dropping path-valued targets; not a small change)
            • Suggested ticket comment Still true in 8.6 and 9.x: with a space in the source path, configure works but make fails immediately (the configure-regeneration rules use the source path as a make target). A full fix means reworking Makefile.in quoting; patches welcome.
            ⬜547989ffff2002Undocumented public API: threadsREPRODUCED (main, 9.0, 8.6): Tcl_GetAllocMutex is still stub slot 387 in tcl.decls, exported, and not documented in doc/*.3

            Open ticket 547989ffff · 49. Threading · filed 2002-04-24

            Full verdict

            REPRODUCED (main, 9.0, 8.6): Tcl_GetAllocMutex is still stub slot 387 in tcl.decls, exported, and not documented in doc/*.3

            • Platform generic (API/docs)
            • Tested grep of tcl.decls/doc on main, 9.0, 8.6. dkf's branch bug-547989 (942397a2a7, 2013-09-10, on 8.6-era code: removes slot 387, renames to TclpGetAllocMutex) was never merged. No user found in Thread, Tk or Itcl sources; core uses it in tclAlloc.c, tclCkalloc.c, tclThreadAlloc.c.
            • Cause internal helper exported through the public stub table since 1999.
            • Prototype none (removing it is an ABI change for 10.0 or needs a TIP; documenting it is a policy decision)
            • Suggested ticket comment Still public and undocumented in 9.1. The branch bug-547989 was never merged and Tcl 9.0 shipped with slot 387; removal (or a "deprecated, internal" note) now has to wait for Tcl 10 / a TIP. Lowering to a Tcl 10 cleanup item.
            ⬜528448ffff2002Tcl_UpdateLinkedVar doesn't report errorREPRODUCED (main, 9.0, 8.6)

            Open ticket 528448ffff · 09. Linked C Variables · filed 2002-03-11

            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested tcltest testlink update with a write trace on the linked var that throws (work-a3/528448.tcl): update returns OK, empty result, error lost, on main, 9.0 and 8.6 tip tcltest
            • Cause generic/tclLink.c:Tcl_UpdateLinkedVar ignores the result of Tcl_ObjSetVar2 (no TCL_LEAVE_ERR_MSG) and returns void (tcl.decls slot), so it has no way to report the error
            • Prototype none (needs a public API change: int return type or a new function, i.e. a TIP; void→int in the stub slot would be binary compatible)
            • Suggested ticket comment Still the case: Tcl_UpdateLinkedVar returns void and drops errors from write traces. Fixing it means changing the public signature (e.g. return int with TCL_LEAVE_ERR_MSG), so it needs a TIP; void→int keeps the stub table binary compatible.
            ⬜526163ffff2002tty not reset on exitREPRODUCED (main, 9.0, 8.6, 8.4); intended since TIP #35

            Open ticket 526163ffff · 25. Channel System · filed 2002-03-05

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4); intended since TIP #35

            • Platform unix
            • Tested Python pty harness: open /dev/pts/N r+; exit (also after -mode, -blocking 0) on main, 9.0.5, 8.6.18, 8.4.20. termios of the pty is changed after exit in all of them; file status flags (O_NONBLOCK) are not, since they belong to Tcl's own open file description.
            • Cause unix/tclUnixChan.c:TclpOpenFileChannel calls TtyInit(), which switches the tty to raw (iflag=IGNBRK, oflag=lflag=0, VMIN=1), and only *after* that saves initState. So TtyCloseProc's doReset (TIP #160) only undoes later fconfigure changes, never the raw mode from the open itself. Not restoring at close was intended by TIP #35 (so Tcl can be used like stty), so changing it would change behaviour.
            • Prototype none (behaviour change; saving initState before TtyInit would also change what -inputmode reset and close restore)
            • Suggested ticket comment Still the case: opening a tty applies TtyInit (raw mode) before initState is saved, so neither close nor -inputmode reset brings back the state from before the open. Keeping settings after close is what TIP #35 wanted. Restoring the state from before the open needs a decision (maybe a TIP).
            ⬜502197ffff2002SetEnvironmentVariable() currupts listpatchREPRODUCED (main, 9.0, 8.6)

            Open ticket 502197ffff · 08. Environment Variables · filed 2002-01-11

            🧪 Prototype check (2026-10-24)

            502197ffff.patch: VERIFIED+TEST-ADDED (on Wine, not real Windows; triage had already run the repro on a Windows VM)

            • Base main (tip d6dd6d0f71); Windows build cross-compiled with mingw-w64 in work-a4/winbuild (wt-a4/win/configure --host=x86_64-w64-mingw32 --enable-64bit), run under Wine with a private WINEPREFIX=work-a4/wineprefix
            • Before win/502197ffff.tcl (helper winenv90.dll) under Wine -> 1. after SetEnvironmentVariable: env()=first OS=second child=second / 2. new via SetEnvironmentVariable: env()=<unset> ... / VERDICT: REPRODUCED (2 of 3 checks out of sync)
            • Patch applied cleanly (tclEnv.c part: 28 changed lines; triage said 33), plus the test added below: 90 changed lines now
            • After -> env()=second, env()=fresh (in names=0, as triage noted: array names env still misses OS-only variables), VERDICT: NOT-REPRODUCED; env.test under Wine 33 total, 32 passed, 1 skipped, 0 failed; winPipe.test under Wine has the same 2 failures (winpipe-4.5, winpipe-8.4: Wine quirks) with and without the patch; unix env.test 26 passed / 7 skipped / 0 failed; unix build unaffected (code is #ifdef _WIN32)
            • Test added test command testwinsetenv name ?value? (win/tclWinTest.c, calls SetEnvironmentVariableW) and env-12.0 (constraints win testwinsetenv) in tests/env.test; under Wine it fails before (Result was: first 0 {}) / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform Windows
            • Tested code reading only. Unix has no equivalent (setenv/putenv and environ are the same store). Repro: win/502197ffff.tcl with helper win/502197ffff/winenv90.dll / winenv86.dll (mingw cross-built with the stub lib linked in, source winenv.c; wsetenv/wgetenv call SetEnvironmentVariableW/GetEnvironmentVariableW). It checks reading after SetEnvironmentVariable (existing and new var) and writing from Tcl
            • Cause generic/tclEnv.c on Windows reads the CRT's _wenviron (TclSetupEnv, TclGetEnv) and writes with _wputenv. _wputenv also updates the OS block, but SetEnvironmentVariableW does not update the CRT copy, so changes made by an embedder or extension through the Win32 API never show up in env(); child processes, which get the OS block, do see them. A fix would be to read through GetEnvironmentVariableW/GetEnvironmentStringsW on Windows
            • Prototype none (not testable here; an untested rewrite of the Windows read path is risky)
            • Suggested ticket comment Still expected on Windows: Tcl's env reads the CRT copy (_wenviron), which SetEnvironmentVariable does not update. Repro script with a helper DLL prepared; a fix would read via GetEnvironmentVariableW on Windows.
            Windows VM, run 2
            • Platform Windows
            • Verdict REPRODUCED (main, 9.0, 8.6)
            • Tested VM win/502197ffff.tcl with the helper DLLs (winenv90.dll loads into 9.1 too): after SetEnvironmentVariableW, env() keeps the old value (env()=first OS=second child=second), a variable created only by SetEnvironmentVariableW is not visible (<unset>), Tcl → OS direction works. Same on 9.1b1, 9.0.5, 8.6.18.
            • Cause generic/tclEnv.c reads the CRT copy (_wenviron) in TclGetEnv/TclSetupEnv; SetEnvironmentVariableW updates only the OS block
            • Prototype prototypes/502197ffff.patch (main, 33 changed lines, no test: a test needs a C command calling SetEnvironmentVariableW): TclGetEnv on Windows reads with GetEnvironmentVariableW, falling back to _wenviron only on an unexpected error. With it checks 1 and 2 pass (env()=second, env()=fresh); array names env still does not list the new variable (TclSetupEnv would need GetEnvironmentStringsW). env.test, exec.test, winPipe.test pass on Windows. Tested on Windows.
            • Suggested ticket comment Still in 8.6, 9.0 and 9.1: env() reads the C runtime's copy of the environment, which SetEnvironmentVariable() does not update. Proposed fix: read single variables with GetEnvironmentVariableW on Windows (tested); &#91;array names env&#93; would additionally need GetEnvironmentStringsW.
            ⬜488530ffff2001setuid Tcl scripts don't work on SolarispatchREPRODUCED (main on OpenIndiana; 8.6 has the same code, not run)

            Open ticket 488530ffff · 37. File System · filed 2001-12-03

            🧪 Prototype check (2026-10-24)

            488530ffff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before emulated the Solaris start-up on Linux: LD_PRELOAD shim work-a2/v2/devfd_shim.c makes open("/dev/fd/N") return dup(N) (shared offset, as on Solaris); runfd.sh opens foo.tcl on fd 3, consumes 10 bytes (dd bs=10 count=1 <&3), then runs tclsh /dev/fd/3 and source /dev/fd/3 -> invalid command name "-----..."" (file "/dev/fd/3" line 1), exactly the ticket's error
            • Patch applied cleanly (12 changed lines, unchanged)
            • After same -> both lines printed for tclsh and for source; printf 'puts piped\n' | tclsh /dev/stdin (non-seekable) still works; tests source, main, interp, fileSystem, ioUtil, io: 1527 total, 1478 passed, 49 skipped, 0 failed
            • Test none possible in tcltest (needs the offset-sharing /dev/fd semantics; on Linux /dev/fd/N reopens with a fresh offset); the shim repro above stands in. Real Solaris/illumos run still advisable.
            Full verdict

            REPRODUCED (main on OpenIndiana; 8.6 has the same code, not run)

            • Platform unix (Solaris/illumos)
            • Tested code reading of main generic/tclIOUtil.c. Prototype checked on Linux for regressions only (source/main/interp/fileSystem tests, script from a pipe via /dev/stdin still works).
            • Cause Solaris starts setuid scripts as /dev/fd/N, sharing the offset the shell already advanced; Tcl reads from that offset.
            • Prototype prototypes/488530ffff.patch (12 changed lines, no test; Solaris only). (void) Tcl_Seek(chan, 0, SEEK_SET) right after open in both Tcl_FSEvalFileEx and TclNREvalFile; seek errors on non-seekable channels are ignored. To verify on the OpenIndiana VM: script with #!/bin/sh + exec tclsh "$0" and a long first line, chmod 4755 or 2755, run ./foo.tcl.
            • Suggested ticket comment Still applies by code reading: &#91;source&#93;/tclsh never rewind the script channel. Proposed fix: seek to 0 after opening (ignoring failure), needs confirmation on Solaris/illumos.
            Unix VMs, run 1
            • Platform unix (Solaris/illumos)
            • Verdict REPRODUCED (main on OpenIndiana; 8.6 has the same code, not run)
            • Tested OpenIndiana, main installed with --prefix (runpath, so LD_LIBRARY_PATH is not needed under set-id). Script: #!/bin/sh, set top 1 ;# \, exec .../tclsh9.1 "$0" ${1+"$@"}, then puts "script=[info script] top=[info exists top]". Root was used only to give the script group other (setgid to one's own group does not trigger /dev/fd), then mode 2755. Run as the user: Tcl gets /dev/fd/3 as the script. Unpatched main prints top=0, i.e. it starts reading after the shell's exec line, with both /bin/sh (ksh93) and bash. With the prototype it prints top=1. The ticket's visible error needs a shell that leaves the offset mid-line; ksh93 and bash seek back to the end of the line they consumed, so the remaining text still parses, but the script still runs from the wrong offset. Without set-id both builds read from the start.
            • Cause illumos runs set-id scripts as /dev/fd/N, which shares the offset the shell advanced; generic/tclIOUtil.c:Tcl_FSEvalFileEx/TclNREvalFile never rewind.
            • Prototype prototypes/488530ffff.patch (12 changed lines, no test; unchanged). Verified on OpenIndiana (top=1). source/main/interp/fileSystem/cmdAH/fileName/unixFile/fCmd tests pass with it there.
            • Suggested ticket comment Reproduced on OpenIndiana: a setgid script is run as /dev/fd/3, and Tcl starts reading at the shell's offset (after the exec line), not at the top. Proposed fix: seek to 0 after opening the script (failure ignored), verified there.
            ⬜472113ffff2001[namespace which -var] fails for ary(i)patchREPRODUCED (main, 9.0, 8.6)

            Open ticket 472113ffff · 21. [namespace] · filed 2001-10-17

            🧪 Prototype check (2026-10-24)

            472113ffff.patch: VERIFIED (documented behaviour change, may still want a TIP / core vote)

            • Base main (tip d6dd6d0f71)
            • Before namespace-34.8 FAILED: {} {} {} {} {} {} (namespace which -variable ary(i) empty although ary(i) exists)
            • Patch applied cleanly; added the doc/namespace.n sentence the triage said was missing (array element: array looked up, index appended whether or not the element exists; empty for a missing array or a defined scalar), checked with man -l. Now 53 changed lines
            • After ::test_ns_1::ary(i) {::test_ns_1::ary(no such)} ::test_ns_1::ary(i) ::test_ns_1::undef(x) {} {}; tests namespace 313/0 failed, namespace-old 126/0, var 217/0, upvar 70/0, info 285/0, oo 373/0
            • Test existing namespace-34.8 fails before / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6)

            • Platform generic
            • Tested reporter script on main, 9.0, 8.6: [namespace which -variable ary] gives ::a::ary, [namespace which -variable ary(i)] gives "" although ary(i) exists
            • Cause generic/tclNamesp.c:NamespaceWhichCmd passes the whole "ary(i)" to Tcl_FindNamespaceVar, which looks it up as a scalar name; Tcl_GetVariableFullName also refuses array elements. msofer's 2006 objection (elements don't know their array) does not apply if the name is split before lookup
            • Prototype prototypes/472113ffff.patch (49 changed lines, test namespace-34.8) — if the name ends in "(...)", resolve the array part and append the index (element need not exist, as oehhar asked); a defined scalar or a missing variable still gives "". Full suite passes. It changes documented behaviour ("returns empty if the variable does not exist"), so namespace.n needs a sentence too and it may want a TIP or at least a core vote; 8.6 code is the same
            • Suggested ticket comment Still present. Proposed fix: &#91;namespace which -variable a(i)&#93; resolves the array part and returns its qualified name plus the index (empty if the array part does not resolve or is a scalar), test namespace-34.8. It changes documented behaviour, so it needs a doc update and agreement.
            ⬜458892ffff2001Legal but irregular names ??REPRODUCED (main, 9.0, 8.6, 8.4)

            Open ticket 458892ffff · 21. [namespace] · filed 2001-09-05

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4)

            • Platform generic
            • Tested reporter's script and the ::foo::bar:::baz case from the comments on main build (c3bf6dd052), 9.0.5, 8.6.18, 8.4.20: identical output everywhere: (1 1 1 2 2 1), namespace children ::foo::bar: returns ::foo::bar:::baz, and that name can't be used.
            • Cause generic/tclNamesp.c TclGetNamespaceForQualName: any run of 2+ colons is one separator, so a trailing : of a namespace/var name merges with the next separator. Documented in namespace.n ("two or more colons").
            • Prototype none (language-level design question; needs a TIP; msofer's 2007 note says "live with this")
            • Suggested ticket comment Still the same in 9.1/9.0/8.6. Names with leading/trailing single : merge with the :: separator (documented in namespace.n); changing it needs a TIP. Suggest closing as "Won't Fix" or moving to RFE.
            ⬜449279ffff2001[namespace which -var] ignores aliasesREPRODUCED (main, 9.0, 8.6, 8.4)

            Open ticket 449279ffff · 21. [namespace] · filed 2001-08-08

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4)

            • Platform generic
            • Tested reporter's script on main build, 9.0.5, 8.6.18, 8.4.20: all print x is ::bar::x; for a plain local var namespace which -variable y returns "".
            • Cause generic/tclNamesp.c NamespaceWhichCmd uses Tcl_FindNamespaceVar, which never looks at the proc's local (compiled) variables; resolving a link to its target name would also need a Var->name mapping (the "analog of namespace origin for variables" is an RFE).
            • Prototype none (behaviour change + new feature; also problematic for links to array elements / other frames, as kbk noted)
            • Suggested ticket comment Unchanged in 9.1/9.0/8.6. Making &#91;namespace which -variable&#93; frame-aware is a behaviour change and the "origin for variables" part is a feature request; suggest converting to an RFE/TIP.
            ⬜219297ffff2000euc-jp encoding ignores euc-3 extensionREPRODUCED (main, 9.0, 8.6, 8.4)

            Open ticket 219297ffff · 10. Objects · filed 2000-10-26

            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4)

            • Platform generic
            • Tested encoding convertfrom euc-jp "\x8f\xb0\xa1" (JIS X 0212 U+4E02): main/9.0.5 error "unexpected byte sequence ... '\x8F'" (strict profile); 8.6.18/8.4.20 give U+008F U+4E9C. encoding convertto euc-jp 丂 gives ?.
            • Cause library/encoding/euc-jp.enc is a table ("M") encoding, which supports at most 2-byte sequences, so the 3-byte SS3 (0x8F) range for JIS X 0212 can't be expressed; jis0212.enc exists but only iso2022-jp uses it.
            • Prototype none (needs either 3-byte table support in tclEncoding.c or a C-coded EUC-JP converter combining jis0201/0208/0212; well over 60 lines)
            • Suggested ticket comment Still present in 9.1: euc-jp.enc is a 2-byte table encoding, so SS3 (0x8F) JIS X 0212 sequences aren't decoded (error with the 9.x strict profile). Fixing needs a dedicated EUC-JP converter using jis0212.enc.
            ⬜219279ffff2000pkg_mkIndex incorrectly handles namespace import commandpatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4) (only with pkg_mkIndex -lazy; the default -direct index does not list commands)

            Open ticket 219279ffff · 39. Package Manager · filed 2000-10-26

            🧪 Prototype check (2026-10-24)

            219279ffff.patch: VERIFIED

            • Base main (tip d6dd6d0f71)
            • Before tclsh r219.tcl (reporter's foo.tcl; pkg_mkIndex -lazy, then package require and plain foo in a new interp) -> index {foo.tcl source ::JCS_task_service::foo}, plain foo: 1 invalid command name "foo"
            • Patch applied cleanly
            • After index {foo.tcl source {::JCS_task_service::foo foo}}, FOO!, plain foo: 0; pkgMkIndex.test 27/0 failed (also via make test with a rebuilt zip), autoMkindex 11/0, package 403/0
            • Test existing pkgMkIndex-11.1 (updated expectation) fails with the unpatched package.tcl / passes after
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4) (only with pkg_mkIndex -lazy; the default -direct index does not list commands)

            • Platform generic
            • Tested reporter's file (work-a1/p219279/foo.tcl, work-a1/219279ffff.tcl): index lists only ::JCS_task_service::foo, not the global foo the package imports itself, so lazy loading via plain foo fails. Same on all versions; test pkgMkIndex-11.1 asserts this output.
            • Cause library/package.tcl pkg_mkIndex: after sourcing, it maps every new global command to its namespace origin and drops the global name, without telling apart its own namespace import -force from imports done by the package.
            • Prototype prototypes/219279ffff.patch (11 changed lines, test pkgMkIndex-11.1 updated): remember the global commands right after sourcing; for those, keep the global name in addition to the origin. Index becomes {::fubar::foo foo} (the ticket's first desired form). pkgMkIndex.test, autoMkindex.test pass.
            • Suggested ticket comment Still true with pkg_mkIndex -lazy. Proposed fix: commands the package itself imports into :: are indexed under both names; test pkgMkIndex-11.1 updated (it encoded the old output).
            ⬜219254ffff2000[file executable file_name] always returns 1 when working frpatchREPRODUCED (main, 8.6 as root on OpenIndiana)

            Open ticket 219254ffff · 37. File System · filed 2000-10-26

            🧪 Prototype check (2026-10-24)

            219254ffff.patch: VERIFIED (emulated; real check on the OpenIndiana VM as root is still pending)

            • Base main (tip d6dd6d0f71)
            • Before LD_PRELOAD shim emulating Solaris root (geteuid()=0, access() always 0; work-a5/219254-shim.c), files 0644 noexec.txt / 0700 uexec.txt, work-a5/219254.tcl -> uid=0 noexec(644)=1 uexec(700)=1 glob-x=noexec.txt uexec.txt
            • Patch applied cleanly (21 changed lines)
            • After with the shim uid=0 noexec(644)=0 uexec(700)=1 glob-x=uexec.txt; without it unchanged (noexec(644)=0 uexec(700)=1); tests fileName 259/0 failed, fCmd 260/0, cmdAH 16971/0, fileSystem 141/0, unixFile 7/0
            • Test none possible (needs a real superuser on a system whose access() grants X_OK to root; Linux does not; tested only with the shim)
            Full verdict

            REPRODUCED (main, 8.6 as root on OpenIndiana)

            • Platform unix (Solaris; POSIX allows it)
            • Tested main (own build) patched vs 9.0.5 with work-a5/219254-shim.c and work-a5/219254.tcl
            • Cause unix/tclUnixFile.c:TclpObjAccess and NativeMatchType rely on access(X_OK), which POSIX lets succeed for the superuser regardless of mode bits
            • Prototype prototypes/219254ffff.patch (21 changed lines, no test: needs root on Solaris). For euid 0 with X_OK, also require an x bit on non-directories in TclpObjAccess; glob -types x checks the already-available stat buffer. fileName/fCmd/cmdAH/fileSystem/unixFile pass. To verify on the OpenIndiana VM: as root, touch f; chmod 644 f; tclsh: file executable f should give 0.
            • Suggested ticket comment Still applies by code reading: access(X_OK) may succeed for root without any execute bit (Solaris). Proposed fix: for root, also check the mode bits in TclpObjAccess and glob -types x; needs confirmation on Solaris/illumos.
            Unix VMs, run 1
            • Platform unix (Solaris/illumos)
            • Verdict REPRODUCED (main, 8.6 as root on OpenIndiana)
            • Tested OpenIndiana via pfexec (root): for a 0644 file, file executable f returns 1 and glob -types x * lists it on main (installed build) and core-8-6-branch. As a normal user it is 0. With prototypes/219254ffff.patch (main): root gets 0 for the 0644 file and 1 for 0744 and a directory, and glob -types x lists only d g. Non-root is unchanged. cmdAH, fileName, unixFile, fileSystem and fCmd tests pass with the patch (as a normal user).
            • Cause unix/tclUnixFile.c:TclpObjAccess and NativeMatchType use access(X_OK), which illumos grants to the superuser without any x bit (POSIX allows this).
            • Prototype prototypes/219254ffff.patch (21 changed lines, no test since it needs root; unchanged). Verified on OpenIndiana as root.
            • Suggested ticket comment Reproduced on OpenIndiana as root with main and 8.6: &#91;file executable&#93; and glob -types x report a 0644 file as executable, since access(X_OK) succeeds for root. Proposed fix: for root, also require an x bit on non-directories; verified there.
            🟢219233ffff2000string match oddity when - is the last charpatchREPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            Open ticket 219233ffff · 18. Commands M-Z · filed 2000-10-26

            🟢 Since triage (updated 2026-11-11)

            Status now: Open (Remind)

            • 🔧 2026-10-19 check-in 7ca44862f9 on glob-bracket-posix, not merged: Fix 219233ffff and 3556218fff: in glob patterns ([string match], [glob], [lsearch -glob], [switch -glob]), a "-" right before the closing bracket of a set n
            • 🔧 2026-10-19 check-in 2e0da1e14e on glob-match-linear, in main: Fix a4a7f2abd3: [string match], [glob] and [lsearch -glob] took exponential time for patterns like a*a*a*...b. When the string runs out before the pattern, an
            • 💬 2026-10-19 serhiy.storchaka: Still present in 8.6, 9.0 and 9.1. Proposed fix for 9.1 in branch glob-bracket-posix, on top of glob-match-linear (the matcher change for a4a7f2abd3): in all four matchers, a - right before the closing bracket is a literal, as in POSIX and Python's fnmatch. So {[.-]} matches only "." and "-", and {*[.-]*} matches "2.1-Beta". Test string-11.57; util-5.40, 5.41, 5.43 and 5.44 tested the old behaviour and are chang…
            • 💬 2026-10-19 dgp: Alert that if this change is accepted it will require bold labels of POTENTIAL INCOMPATIBILITY in the release notes. I'm unsettled by changing the functioning of RE processing in order to "fix" a "bug", most especially one that has survived for 25 years and more. Would help to have more people examine the impact.
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.5, 8.4)

            • Platform generic
            • Tested work-a1/219233ffff.tcl, all reported cases on all versions: string match {[.-]} - -> 0, {[.-]} A -> 1, {*[.-]*} 2.1-Beta -> 0, {[a-z0-9_/-]} \\ -> 1, {[a-]} \] -> 1.
            • Cause generic/tclUtil.c Tcl_StringCaseMatch and TclByteArrayMatch, generic/tclUtf.c TclUniCharCaseMatch and TclUniCharMatch: a - right before ] starts a range whose end is the ], so the set's closing bracket is used up.
            • Prototype prototypes/219233ffff.patch (31 changed lines; tests string-11.56, string-11.57; util-5.40/41/43/44 updated, they encoded the old range-to-] behaviour): a - followed by ] is a literal -, as in POSIX brackets. All four matchers fixed (byte array and Unicode paths checked under gdb). Full suite on main (without http*/zipfs) passes: 521549 total, 0 failed. Incompatible for patterns like [A-]] (was range A..]), so main/9.1 only; the unclosed-bracket case {[.-B} (sebres' comment) is not touched.
            • Suggested ticket comment Still present. Proposed fix: treat - right before ] as a literal in all four glob matchers (string, byte array, Unicode); new tests string-11.56/11.57, util-5.40/41/43/44 changed as they tested the old behaviour. Incompatible, so main only.
            💬219164ffff2000various aspects of error handling fail with interp aliasesREPRODUCED (main, 9.0, 8.6, 8.4)

            Open ticket 219164ffff · 20. [interp] · filed 2000-10-26

            💬 Since triage (updated 2026-11-11)

            Status now: Open

            • 💬 2026-10-17 serhiy.storchaka: interp-26.8 (the parent's errorInfo and errorstack leaking into a safe interp through an alias) is now tracked separately in a63f5a4fed.
            Full verdict

            REPRODUCED (main, 9.0, 8.6, 8.4)

            • Platform generic
            • Tested interp.test -constraints knownBug -match interp-26.* on main build, 9.0 tip and 8.6 tip tcltest: interp-26.4, 26.6, 26.8 still fail (26.5 has since had its expected result changed to the actual behaviour and passes). 26.4 also checked on 8.4.20 (2 2 2 2 2 2 2).
            • Cause 26.4: return -code N evaluated at level 0 via invokehidden yields TCL_RETURN (same as at top level anywhere); 26.6: break/continue returned through a proc-based alias target become "invoked break outside of a loop" by proc semantics; 26.8: errorInfo from the parent's alias implementation is copied into a safe child (generic/tclInterp.c AliasObjCmd / Tcl_TransferResult).
            • Prototype none (26.4/26.6 are arguably correct behaviour and the tests' expectations should change; 26.8 is a design/security decision about hiding parent errorInfo from safe interps)
            • Suggested ticket comment interp-26.4/26.6/26.8 still fail in 9.1. 26.4 and 26.6 follow normal return-code rules (suggest updating the tests' expected results); 26.8 (parent errorInfo leaking into safe interps) is the only part worth deciding on, maybe as a separate ticket.