Untouched open Tcl bug tickets: Not a bug / RFE

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.

Not a bug / RFE 119

⬜bf2652aedf2026Tcl 9.1b1 with zipfs support: incorrect value for tcl_library variableNOT-A-BUG (TIP 732: embedding apps must call Tcl_FindExecutable or TclZipfs_AppHook; reporter confirmed that adding TclZipfs_AppHook fixes it)

Open ticket bf2652aedf · 07. Variables · filed 2026-09-29

Full verdict

NOT-A-BUG (TIP 732: embedding apps must call Tcl_FindExecutable or TclZipfs_AppHook; reporter confirmed that adding TclZipfs_AppHook fixes it)

  • Platform generic
  • Tested small C embedder against installed main 9.1 and 9.0.5: without the hook both fail Tcl_Init (9.1 also errors on info library); with TclZipfs_AppHook both give //zipfs:/lib/tcl/tcl_library. The wish symptom from the description is not reproducible (Jan, 9.1.0).
  • Cause documented 9.1 incompatibility (TIP 732), not a defect.
  • Prototype none (not a bug)
  • Suggested ticket comment Closing: as the reporter confirmed, calling TclZipfs_AppHook (or Tcl_FindExecutable) before Tcl_CreateInterp is required since TIP 732. Resolution: Works For Me / Not a bug.
⬜62144fdc022026"scan NaN %f" returns a list with an empty elementNOT-A-BUG (it is the normal inline result for a failed conversion; accepting NaN is a feature request, TIP 520)

Open ticket 62144fdc02 · 18. Commands M-Z · filed 2026-07-21

Full verdict

NOT-A-BUG (it is the normal inline result for a failed conversion; accepting NaN is a feature request, TIP 520)

  • Platform generic
  • Tested main (9.1.1), 9.0.5, 8.6.18, 8.5.19 all give the same output. scan NaN %f → {}, scan NaN %f fl → 0 with fl unset. For comparison scan x %f → {}, scan Nax %f → {}, scan "5 xyz" "%d %f" → {5 {}}, scan "" %f → `` (end of input).
  • Cause generic/tclScan.c Tcl_ScanObjCmd: TclParseNumber accepts "NaN", but Tcl_GetDoubleFromObj then rejects it unless ACCEPT_NAN is defined, so the conversion fails (goto done). In inline mode each specifier that was not converted gets an empty element; with variables the count 0 is returned. Both forms say "0 conversions", so they agree. The empty element is used for every failed conversion, not only NaN.
  • Prototype none (no bug; changing NaN handling is TIP 520 / ACCEPT_NAN material)
  • Suggested ticket comment The {} is the normal inline result for a conversion that failed ([scan x %f] gives the same). It matches the 0 returned in the varName form. Accepting NaN as a double is TIP 520 territory, so this is a feature request rather than a scan bug.
⬜ddd71e92fb2026External event loops cannot setup event sourcesNOT-A-BUG (missing API, RFE). The behaviour matches the current design.

Open ticket ddd71e92fb · 01. Notifier · filed 2026-06-18

Full verdict

NOT-A-BUG (missing API, RFE). The behaviour matches the current design.

  • Platform generic
  • Tested code reading of generic/tclNotify.c (main, 9.0, 8.6)
  • Cause The event source list (tsdPtr->firstEventSourcePtr) is private to tclNotify.c. The only public entry points that run the setupProcs are Tcl_DoOneEvent (setup, wait, check) and Tcl_ServiceAll (setup and check with no wait in between). A replacement notifier does see the Tcl_CreateFileHandler calls and Tcl_SetMaxBlockTime (through its createFileHandlerProc/setTimerProc hooks), but an external loop has no way to run the setup phase before it blocks and the check phase after it.
  • Prototype none (needs new public API, e.g. Tcl_SetupEventSources(flags)/Tcl_CheckEventSources(flags), i.e. a TIP, or the documented workaround of letting waitForEventProc run the external loop's single iteration)
  • Suggested ticket comment Confirmed by reading tclNotify.c: the setup/check phases are reachable only through Tcl_DoOneEvent or Tcl_ServiceAll, and the event source list is private. Splitting them out needs new public API, so this is a feature request (TIP). Until then, the usual approach is to make the replacement notifier's waitForEventProc run one blocking iteration of the external loop and call Tcl_DoOneEvent from the outside.
⬜7eb6b3a53f2026glob on UNC path requires additional slashNOT-A-BUG (documented: glob.n "WINDOWS PORTABILITY ISSUES" says \ before * escapes it; C:\\foo\\* is C:\foo\*) — confirm on Windows with wi…

Open ticket 7eb6b3a53f · 36. Pathname Management · filed 2026-05-29

Full verdict

NOT-A-BUG (documented: glob.n "WINDOWS PORTABILITY ISSUES" says \ before * escapes it; C:\\foo\\* is C:\foo\*) — confirm on Windows with win/7eb6b3a53f.tcl

  • Platform Windows
  • Tested code/doc inspection only (Unix has no backslash separator); repro script win/7eb6b3a53f.tcl
  • Cause generic/tclFileName.c TclGlob/DoGlob: only the leading volume/UNC prefix is split off without pattern processing; the rest is a glob pattern where \* is a literal *. apnadkarni's 2026-05-31 analysis is correct.
  • Prototype none (works as documented)
  • Suggested ticket comment This is the documented behaviour, see "WINDOWS PORTABILITY ISSUES" in glob.n: in the pattern a backslash escapes the next character, so \* is a literal asterisk. Use forward slashes or -directory {C:\WINDOWS} *. Propose closing as Works as Designed.
⬜0ecc9116b02025NOT-A-BUG (the porting guides are community wiki pages; the API is documented in doc/CrtObjCmd.3, which covers Tcl_CreateObjCommand2/Tcl_ObjCmdProc2)

Open ticket 0ecc9116b0 · 69. Other · filed 2025-06-07

Full verdict

NOT-A-BUG (the porting guides are community wiki pages; the API is documented in doc/CrtObjCmd.3, which covers Tcl_CreateObjCommand2/Tcl_ObjCmdProc2)

  • Platform docs (wiki porting guides, not in the repository)
  • Tested grep of doc/CrtObjCmd.3 on main (8 mentions of ObjCmdProc2/CreateObjCommand2)
  • Cause n/a. Ashok replied that anyone can edit the wiki guides.
  • Prototype none (not in the repository)
  • Suggested ticket comment The porting guides are wiki pages that anyone can edit. The Tcl_ObjCmdProc2/Tcl_CreateObjCommand2 API is documented in the Tcl_CreateObjCommand manual page. Closing; please add a pointer on the wiki if it helps.
⬜94e3289b792025MemoryModule doesn't handle TLS completely.NOT-A-BUG (obsolete: MemoryModule was never merged — TIP 709 was Rejected 1/5/0; branch tip-709 closed 2025-12-18, last check-in 6b3706cdb7 2025-11…

Open ticket 94e3289b79 · 40. Dynamic Loading · filed 2025-01-25

Full verdict

NOT-A-BUG (obsolete: MemoryModule was never merged — TIP 709 was Rejected 1/5/0; branch tip-709 closed 2025-12-18, last check-in 6b3706cdb7 2025-11-30; no MemoryModule code in main/9.0/8.6)

  • Platform Windows
  • Tested grep of main tree (no MemoryModule/MemoryLoadLibrary), branch and TIP state
  • Cause n/a
  • Prototype none (code not in any maintained branch)
  • Suggested ticket comment TIP 709 was rejected and the tip-709 branch is closed, so MemoryModule is not part of Tcl. Closing as Out of Date.
⬜edebb6a4ff2024lsearch -index requires nested items to existNOT-A-BUG (long-standing, deliberate behaviour shared with lsort -index; a change would be a feature request / TIP)

Open ticket edebb6a4ff · - New Builtin Commands · filed 2024-11-08

Full verdict

NOT-A-BUG (long-standing, deliberate behaviour shared with lsort -index; a change would be a feature request / TIP)

  • Platform generic
  • Tested main, 9.0.5, 8.6.18: lsearch -index 0 {{} one two three} one → element 0 missing from sublist "". lsort -index 0 {{} b a} gives the same error (both use SelectObjFromSublist), and lsort -index has done so since 8.4.
  • Cause generic/tclCmdIL.c SelectObjFromSublist raises "element N missing from sublist" for every lsearch/lsort -index use. Skipping such items in lsearch only (and not -1 vs empty-string matching, see pooryorick's comment) would change behaviour and needs a design decision, e.g. a new option.
  • Prototype none (behaviour change)
  • Suggested ticket comment This is shared with lsort -index (same SelectObjFromSublist) and has been consistent since lsearch -index appeared in 8.5. Ignoring short sublists would be a behaviour change and probably needs a TIP (perhaps an option), so this is a feature request rather than a bug.
⬜ad7bafd5e12024auto_path/tcl_library consistency for zipfs filesystemNOT-A-BUG (design/RFE: layout of //zipfs:/app and default auto_path); unchanged on main

Open ticket ad7bafd5e1 · - New Builtin Commands · filed 2024-10-06

Full verdict

NOT-A-BUG (design/RFE: layout of //zipfs:/app and default auto_path); unchanged on main

  • Platform generic
  • Tested code inspection of main: tclZipfs.c TclZipfsLocateTclLibrary still uses //zipfs:/app/tcl_library, init.tcl still adds only its parent, nothing adds //zipfs:/app/lib
  • Cause n/a (by design, see jan.nijtmans 2024-10-07)
  • Prototype none (behaviour change; jan.nijtmans wanted more opinions before adding //zipfs:/app/lib to auto_path)
  • Suggested ticket comment Still as designed on main. Adding //zipfs:/app/lib to auto_path is a feature request; the zipfs mkimg offset problem discussed here belongs to fc65ff1b66 or a new ticket. Propose changing type to RFE.
⬜e8664672cd2024tcllib packages in deeper subfolders won't load in safe interpNOT-A-BUG (works as documented; RFE at most)

Open ticket e8664672cd · 33. Safe Base · filed 2024-06-05

Full verdict

NOT-A-BUG (works as documented; RFE at most)

  • Platform generic
  • Tested synthetic package dir on auto_path whose pkgIndex.tcl sources sub/d.tcl two levels below the auto_path entry: in a safe::interpCreate child package require of the one-level package works and the two-level one fails with "permission denied" on main, 9.0.5 and 8.6.18.
  • Cause library/safe.tcl:AddSubDirs adds only first-level subdirectories of the parent's auto_path to the default access path; doc/safe.n ("the first-level sub directories of each directory in the parent auto_path will also be added") says exactly this. tcllib's page/plugins and term/ansi/code are two levels down.
  • Prototype none (widening the default access path to arbitrary depth is a security-policy change; workaround is safe::interpAddToAccessPath or -accessPath).
  • Suggested ticket comment This is the documented behaviour: only the first-level subdirectories of auto_path are added to the default access path. Use safe::interpAddToAccessPath (or -accessPath) for deeper package directories; recursive addition would need a policy decision (RFE).
⬜e7c63609752024http 2.10 meta changes dict keys to lowercaseNOT-A-BUG (documented incompatible change of http 2.10: http::meta is now an alias of http::responseHeaders, whose names are documented as conv…

Open ticket e7c6360975 · 81. Bundled Packages · filed 2024-05-10

Full verdict

NOT-A-BUG (documented incompatible change of http 2.10: http::meta is now an alias of http::responseHeaders, whose names are documented as converted to lower case)

  • Platform generic
  • Tested localhost server; http 2.10.3 (main) and 2.10.2 (9.0.5) return lower-case names, 2.9.9 (8.6.18) keeps server case
  • Cause library/http/http.tcl Event lower-cases header keys (set key [string tolower $key]); doc/http.n responseHeaders/meta
  • Prototype none (intended behaviour)
  • Suggested ticket comment This is intentional and documented in http 2.10: <tt>http::meta</tt> is an alias of <tt>http::responseHeaders</tt>, which returns lower-cased header names (and the result is not a dict, since names may repeat). Scripts should use <tt>http::responseHeaderValue</tt> or lower-case keys. Suggest closing as Works as designed.
⬜4b80f92abd2024cannot transform using unicode (maybe)NOT-A-BUG (transforms work on bytes by design)

Open ticket 4b80f92abd · 26. Channel Transforms · filed 2024-01-17

Full verdict

NOT-A-BUG (transforms work on bytes by design)

  • Platform generic
  • Tested 8.6.18, 9.0.5: a write-only chan push stdout transform receives puts "❽" as the 4 bytes e2 9d bd 0a (the channel's utf-8 encoding), and encoding convertfrom utf-8 gives back the character.
  • Cause transchan(n): the write method gets the data after the channel's encoding was applied, as a byte string. The wiki code passes those bytes straight to ::tk::ConsoleOutput, so one character shows up as 3 Latin-1 characters. The fconfigure ... -encoding utf-8 in initialize does not change that.
  • Prototype none (not a bug)
  • Suggested ticket comment Channel transforms always see bytes: the write method gets the data already encoded with the channel's -encoding. Decode it before passing it to the Tk console, e.g. ::tk::ConsoleOutput stdout &#91;encoding convertfrom &#91;fconfigure $chan -encoding&#93; $data&#93; (or fix the channel's encoding to utf-8 and always decode as utf-8). Propose closing as Works as designed.
⬜ab913b8e9e2023xterm window closes by itself when we run “ctrl + c”NOT-A-BUG (expected Unix signal semantics; Tcl has no job control)

Open ticket ab913b8e9e · - New Builtin Commands · filed 2023-11-16

Full verdict

NOT-A-BUG (expected Unix signal semantics; Tcl has no job control)

  • Platform unix
  • Tested 9.0.5 on Linux: exec sleep 30 & child is in tclsh's process group with no signals ignored; SIGINT to the group (what Ctrl+C on the terminal does) kills it. exit/Ctrl+D send no signal, so the child survives
  • Cause unix/tclUnixPipe.c forks background pipelines in the caller's process group and does not set SIGINT/SIGQUIT to SIG_IGN (POSIX sh does that for & without job control; interactive shells use a separate process group)
  • Prototype none (changing signal dispositions for exec ... & would be a behaviour change needing discussion)
  • Suggested ticket comment Ctrl+C sends SIGINT to the whole foreground process group, and <tt>exec ... &amp;</tt> children stay in tclsh's group, so they get it too. Use <tt>exec setsid xterm &amp;</tt> (or <tt>nohup</tt> plus <tt>trap '' INT</tt> via sh) to detach the child. Suggest closing as Works as designed.
⬜309cfea3a22023Append does not trigger variable read tracesNOT-A-BUG (long-standing, intentional and tested behaviour: append-7.5 "append var does not trigger read trace" since 2008, appendComp-7.9, trace-3…

Open ticket 309cfea3a2 · 07. Variables · filed 2023-10-03

Full verdict

NOT-A-BUG (long-standing, intentional and tested behaviour: append-7.5 "append var does not trigger read trace" since 2008, appendComp-7.9, trace-3.1/3.2 expect no read trace on append)

  • Platform generic
  • Tested reporter's script on main, 9.0.5, 8.6.18, 8.4.20: identical output (hello bar) everywhere; Ashok's 2026-05-24 output also shows the reported behaviour, he just did not see it as a problem
  • Cause generic/tclVar.c TclPtrSetVarIdx fires read traces only with TCL_TRACE_READS, which append (Tcl_AppendObjCmd, INST_APPEND_*) never passes; lappend reads the variable first (Tcl_ObjGetVar2 / TclPtrGetVarIdx, errors ignored, see bug 3057639). Practical consequence: append on a Tcl_LinkVar string variable uses the stale Tcl value and overwrites a newer C value (checked with testlink: C side "xyz", append string Z gives "abcZ").
  • Prototype none (behaviour change contradicting existing tests; a 10-line change in TclPtrSetVarIdx firing read traces and ignoring their errors like lappend works, but breaks append-7.5, appendComp-7.9, trace-3.1, trace-3.2, so needs a decision/TIP)
  • Suggested ticket comment This is the documented-by-tests behaviour (append-7.5, trace-3.1): &#91;append&#93; does not fire read traces, unlike &#91;lappend&#93;. Changing it is a behaviour change (it would also fix &#91;append&#93; on linked C variables using a stale value), so it needs a TIP rather than a bug fix.
⬜61f4c677702023The info command doesn't care of the working directory if changed by the programNOT-A-BUG (info script returns the path exactly as passed to source/tclsh; same since 8.4)

Open ticket 61f4c67770 · 17. Commands I-L · filed 2023-06-20

Full verdict

NOT-A-BUG (info script returns the path exactly as passed to source/tclsh; same since 8.4)

  • Platform generic
  • Tested main build, 8.6.18, 8.4.20: script run as tclsh t13d/sub/me.tcl does cd then file normalize [info script] → wrong path (found=0) on all
  • Cause by design: doc/info.n says info script returns the name of the file as given; a relative name becomes invalid after cd. Returning absolute paths would change long-standing behaviour (and the TIP 280 info frame file names)
  • Prototype none (intended behaviour; an RFE would need a TIP)
  • Suggested ticket comment <tt>info script</tt> returns the file name as it was given to <tt>source</tt>/tclsh, so a relative name must be normalized before any <tt>cd</tt>: <tt>set me &#91;file normalize &#91;info script&#93;&#93;; cd &#91;file dirname $me&#93;</tt>. Behaviour unchanged since 8.4; suggest closing as Works as designed.
⬜382e6c509c2023[read] loses data under strict encodingNOT-A-BUG (works as documented since TIP 653, 9.0.0): "AB" is not lost. The failing read $chan 5 returns it in the -data key of the error optio…

Open ticket 382e6c509c · - New Builtin Commands · filed 2023-04-14

Full verdict

NOT-A-BUG (works as documented since TIP 653, 9.0.0): "AB" is not lost. The failing read $chan 5 returns it in the -data key of the error options dictionary (catch {read $chan 5} m opts; dict get $opts -data → AB), and the channel stays at the invalid byte, so the next read with -profile replace correctly starts there.

  • Platform generic
  • Tested main (own build), 9.0.5, 8.7b1 (same output as the ticket); -data checked on main and 9.0. 8.6 has no profiles.
  • Cause generic/tclIOCmd.c:Tcl_ReadObjCmd stores the data decoded before the error in -data (TIP 653, d7fce1ea20/9a94125824, 2023-11); documented in doc/chan.n (read, ENCODING ERROR EXAMPLES).
  • Prototype none (not a bug)
  • Suggested ticket comment Since TIP 653 (9.0.0) the data decoded before the encoding error is returned in the -data key of the error options, catch {read $chan 5} m opts; dict get $opts -data gives "AB". The next read starts at the invalid byte, as doc/chan.n describes. Suggest closing as Works as Expected.
⬜dc23c641932023[read] fails, on encoding error, but current position advances anyway.NOT-A-BUG (resolved by TIP 653 in 9.0.0): the second option the ticket asks for is what was implemented. A blocking read that hits an encoding er…

Open ticket dc23c64193 · - New Builtin Commands · filed 2023-04-08

Full verdict

NOT-A-BUG (resolved by TIP 653 in 9.0.0): the second option the ticket asks for is what was implemented. A blocking read that hits an encoding error returns the data decoded so far ("hello") in the -data key of the error options, and the position is left at the invalid byte (tell = 5). This is documented in doc/chan.n.

  • Platform generic
  • Tested main (own build), 9.0.5, 8.7b1: tell=5, error options carry -data hello (main, 9.0); -profile replace then reads "U+FFFD world" (the replacement character, then "world").
  • Cause by design, generic/tclIOCmd.c:Tcl_ReadObjCmd (-data key, TIP 653, d7fce1ea20, 2023-11-06).
  • Prototype none (not a bug)
  • Suggested ticket comment Resolved by TIP 653 (Tcl 9.0): the data decoded before the error is returned in the -data key of the error options, and the position stays at the invalid byte, so nothing is lost on unseekable channels either (see doc/chan.n, ENCODING ERROR EXAMPLES). Suggest closing.
⬜9b63fd201b2023Fix Thread Join on WIndowsNOT-A-BUG (joinable threads work on Windows through the generic/tclThreadJoin.c emulation; the misleading doc sentence is gone in 9.0/main, 9183ace…

Open ticket 9b63fd201b · 80. Thread Package · filed 2023-02-20

Full verdict

NOT-A-BUG (joinable threads work on Windows through the generic/tclThreadJoin.c emulation; the misleading doc sentence is gone in 9.0/main, 9183ace241, 2025-09-26, but still in 8.6 doc/Thread.3:108). The reporter's native-HANDLE rewrite is in branch windows-thread-join (231c699c87 .. 9c9f39bc30, last change 2023-05-03), not merged.

  • Platform Windows
  • Tested source/timeline inspection; win/tclWinThrd.c:TclpThreadCreate on main still does CloseHandle(tHandle) and TclRememberJoinableThread
  • Cause design choice (emulated join), not a functional bug
  • Prototype none (the 8.6 doc line can simply be dropped, as was done on 9.0)
  • Suggested ticket comment Joining threads does work on Windows (emulated in tclThreadJoin.c), and the doc sentence saying it doesn't was removed on 9.0/main in &#91;9183ace241&#93;. The native-HANDLE implementation is in branch windows-thread-join, unmerged since 2023-05; it needs a rebase onto main and a review before it can be considered.
⬜b5bd08df8d2022The "int" type registration should be restored.NOT-A-BUG (design decision of TIP #484; still unregistered: main and 9.0 register only bytecode, cmdName, dict, double, list, procbody, regexp, str…

Open ticket b5bd08df8d · 10. Objects · filed 2022-11-17

Full verdict

NOT-A-BUG (design decision of TIP #484; still unregistered: main and 9.0 register only bytecode, cmdName, dict, double, list, procbody, regexp, string in generic/tclObj.c; 8.6 registers "int"). An RFE / TIP matter

  • Platform generic
  • Tested code reading of generic/tclObj.c TclInitObjSubsystem() on main and 9.0
  • Existing work none (Tcl_GetNumberFromObj, TIP #638, is the offered alternative; Python's _tkinter has worked around it)
  • Prototype none (needs a TIP / consensus, not a bug fix)
  • Suggested ticket comment none needed (Serhiy's 2026-09-20 comment is the latest); could be retyped as RFE.
⬜79a9ba48412022It may be error in "info".NOT-A-BUG (documented): -all adds inherited methods but still lists only exported ones, and -private adds the non-exported ones. To get every m…

Open ticket 79a9ba4841 · 17. Commands I-L · filed 2022-10-17

Full verdict

NOT-A-BUG (documented): -all adds inherited methods but still lists only exported ones, and -private adds the non-exported ones. To get every method, use -all -private (<cloned> destroy eval unknown variable varname for oo::object). doc/info.n describes this and also notes that the name -private clashes with true private methods (TIP 500, see -scope).

  • Platform generic
  • Tested main (own build), 8.6.18 — same output as the ticket
  • Cause by design (generic/tclOOInfo.c InfoClassMethodsCmd)
  • Prototype none
  • Suggested ticket comment Works as documented: -all means "include superclasses/mixins", -private means "include non-exported methods". Use info class methods ::oo::object -all -private to get all of them, or -scope (Tcl 9). Suggest closing as Works as Expected.
⬜c1c4daae0e2022may be error in "namespace eval"NOT-A-BUG (command resolution inside the namespace: the script's set/namespace resolve to the namespace's own commands)

Open ticket c1c4daae0e · 21. [namespace] · filed 2022-10-16

Full verdict

NOT-A-BUG (command resolution inside the namespace: the script's set/namespace resolve to the namespace's own commands)

  • Platform generic
  • Tested main build: in ::tcl::array namespace which set is ::tcl::array::set (= array set), in ::tcl::dict it is ::tcl::dict::set; in ::oo::InfoObject, namespace is the ::oo::InfoObject::namespace ([info object namespace]); ::mk is Metakit's own namespace with a set-like command
  • Cause by design, commands of the current namespace take precedence over global ones; the errors come from those commands
  • Prototype none
  • Suggested ticket comment Not a bug: inside &#91;namespace eval ::tcl::array&#93; the name set resolves to ::tcl::array::set (array set), in ::tcl::dict to ::tcl::dict::set, and in ::oo::InfoObject namespace is &#91;info object namespace&#93;. Use ::set / ::namespace in such scripts. Closing.
⬜49d6b5d7022022parray sort order is often not usefulNOT-A-BUG (feature request): library/parray.tcl sorts with the plain lsort (ASCII), as it has always done (main, 9.0, 8.6 identical). Switching t…

Open ticket 49d6b5d702 · 18. Commands M-Z · filed 2022-09-22

Full verdict

NOT-A-BUG (feature request): library/parray.tcl sorts with the plain lsort (ASCII), as it has always done (main, 9.0, 8.6 identical). Switching to -dictionary changes the output order (a small behaviour change), and passing extra lsort options changes the documented signature parray arrayName ?pattern? (library.n), so it needs a decision (TIP-level for the new arguments).

  • Platform generic
  • Tested code reading (main, 9.0, 8.6 library/parray.tcl are the same)
  • Cause library/parray.tcl uses lsort with no options
  • Prototype none (RFE; the -dictionary change would be a one-word edit if accepted)
  • Suggested ticket comment This is a feature request rather than a bug. Making -dictionary the default is a one-word change but changes the output, and extra sort arguments would need a TIP since the signature is documented in library.n. Suggest recategorising as a Feature Request.
⬜a7002736472022des package missing procNOT-A-BUG (Tcllib bug, wrong tracker; still present in Tcllib trunk)

Open ticket a700273647 · 22. [proc] and [uplevel] · filed 2022-04-10

Full verdict

NOT-A-BUG (Tcllib bug, wrong tracker; still present in Tcllib trunk)

  • Platform generic
  • Tested tcllib.fossil trunk: modules/des/des.tcl line 243 uses [namespace origin Chunk] for -in, but no Chunk proc exists in des.tcl, tcldes.tcl or tcldesjr.tcl. No matching Tcllib ticket found (closest: 3002617fff "des, des-mode parameter and key padding").
  • Cause modules/des/des.tcl: -in channel mode refers to a Chunk helper never defined (other Tcllib crypto modules, e.g. blowfish, have one)
  • Prototype none (not Tcl core)
  • Suggested ticket comment This is a Tcllib issue (modules/des/des.tcl still refers to an undefined Chunk in trunk); please report it at https://core.tcl-lang.org/tcllib. Closing here.
⬜f698462e652022Tcl_Parse migrationNOT-A-BUG (design RFE; the size_t migration it worried about shipped in 9.0 — Tcl_Token.size and Tcl_Parse.numWords/numTokens/tokensAvailable are T…

Open ticket f698462e65 · - New Stub Entries · filed 2022-03-06

Full verdict

NOT-A-BUG (design RFE; the size_t migration it worried about shipped in 9.0 — Tcl_Token.size and Tcl_Parse.numWords/numTokens/tokensAvailable are Tcl_Size in tcl.h; the requested new parse interface was never written)

  • Platform generic (API design)
  • Tested code reading main tcl.h; TIP 616 (Final, 9.0) is lists > 2^31, not a parser interface; dgp-refactor branch still exists unmerged
  • Cause n/a
  • Prototype none (new public API, needs a TIP)
  • Suggested ticket comment In 9.0 the Tcl_Token/Tcl_Parse fields became Tcl_Size, so the open structs now cover large scripts. A redesigned parse interface would need its own TIP; suggest converting this to an RFE or closing.
⬜a0377fd6952021Bug in trace add execution leave <non-existent-proc>NOT-A-BUG (the hang is in the third-party TclReadLine REPL, not in Tcl)

Open ticket a0377fd695 · - New Builtin Commands · filed 2021-07-07

Full verdict

NOT-A-BUG (the hang is in the third-party TclReadLine REPL, not in Tcl)

  • Platform generic
  • Tested main, 9.0.5, 8.6.18, 8.5.19 (script); main interactive tclsh on a pty
  • Cause after trace add execution puts leave foo, every [puts] raises invalid command name "foo" once it has done its work, as documented. TclReadLine prints errors with [puts], which fails again, and its error handler loops forever. Plain tclsh writes errors with Tcl_WriteObj, so the prompt keeps working (checked on a pty). Unrelated side note: in 8.6+ the errorInfo says (file "x.tcl" line 1) for a [puts] on line 2, while 8.5 says line 2.
  • Prototype none (not a Tcl bug)
  • Suggested ticket comment Plain tclsh just reports invalid command name "foo" and keeps running. The hang comes from TclReadLine, which prints errors with &#91;puts&#93;, and that &#91;puts&#93; fails again because of the trace. Please report it to the TclReadLine author (or have it use the channel commands only inside &#91;catch&#93;). Suggest closing as Invalid / Third Party.
⬜1e10b170232021make html How should it work?NOT-A-BUG (discussion/process ticket about the 8.7a5 release, a line that has since been abandoned). On main, make html-tcl DISTDIR=<dir> runs to…

Open ticket 1e10b17023 · - New Builtin Commands · filed 2021-06-18

Full verdict

NOT-A-BUG (discussion/process ticket about the 8.7a5 release, a line that has since been abandoned). On main, make html-tcl DISTDIR=<dir> runs to completion (257 pages, index.html, Keywords/TclCmd/TclLib/UserCmd, docs.css) with the --useversion/--htmldir/--srcdir arguments from unix/Makefile.in. The 9.0.x and 9.1.0 doc distributions have been built with this since.

  • Platform docs
  • Tested main make html-tcl into work-a6/dist (Tcl only, no Tk tree next to the wt)
  • Cause n/a (the tools/tcltk-man2html.tcl rewrite in 2021, which changed contents.htm → index.html and the CSS)
  • Prototype none
  • Suggested ticket comment make html works on main (and was used for the 9.0/9.1 releases); 8.7 is no longer developed. Suggest closing unless there is still a specific problem with the release workflow.
⬜8d7f605fc12021load-10.1 failsNOT-A-BUG (environment: temp dir mounted noexec; same in 8.5.19 per reporter)

Open ticket 8d7f605fc1 · 40. Dynamic Loading · filed 2021-06-07

Full verdict

NOT-A-BUG (environment: temp dir mounted noexec; same in 8.5.19 per reporter)

  • Platform unix
  • Tested main (30d8ccce48) load-10.1 passes here (/tmp is exec). Could not simulate a noexec /tmp: unprivileged user namespaces are blocked on this host (unshare -rm fails), so not reproduced directly.
  • Cause a library loaded from a VFS is copied to a native temporary directory before dlopen(). "failed to map segment from shared object: Operation not permitted" is the dlopen() error for a noexec mount (CentOS-7 hardening). Unchanged in all versions.
  • Prototype none (environment; a real fix would be a Linux memfd_create()+/proc/self/fd path or a fallback directory, i.e. a design change)
  • Suggested ticket comment The error is what dlopen() gives for a file on a noexec mount: the VFS library is copied to $TMPDIR (or /tmp) before loading. Running the tests with TMPDIR set to an exec-allowed directory should make load-10.1 pass; proposing to close as works-as-designed unless someone wants a memfd-based fallback.
⬜79c51f8af12020tcl::pkgconfig error messages inconsistent with the rest of TclNOT-A-BUG (cosmetic; still the same on main, 9.0, 8.6)

Open ticket 79c51f8af1 · 54. Configuration Reporting · filed 2020-10-25

Full verdict

NOT-A-BUG (cosmetic; still the same on main, 9.0, 8.6)

  • Platform generic
  • Tested main, 8.6.18: bad subcommand "z": must be get or list; get zz gives key not known (errorCode TCL LOOKUP CONFIG zz)
  • Cause generic/tclConfig.c QueryConfigObjCmd uses Tcl_GetIndexFromObj ("bad subcommand ..."), the standard message of non-ensemble commands (update foo → bad option, after foo → bad argument); "unknown or ambiguous subcommand" comes only from namespace ensembles. Messages are pinned by tests/config.test pkgconfig-2.x and Tk tests/pkgconfig.test.
  • Prototype none (matching the ensemble wording means turning each <pkg>::pkgconfig into an ensemble, and changes tested messages in Tcl and Tk)
  • Suggested ticket comment The message comes from Tcl_GetIndexFromObj, as in other non-ensemble commands (update, after); only ensembles say "unknown or ambiguous subcommand". Changing it would mean making pkgconfig an ensemble; low value, suggest closing (or an RFE to make key not known name the key).
⬜27944a36612020Taming test utf-6.88NOT-A-BUG (settled; behaviour defined and tested on all branches)

Open ticket 27944a3661 · 44. UTF-8 Strings · filed 2020-04-17

Full verdict

NOT-A-BUG (settled; behaviour defined and tested on all branches)

  • Platform generic
  • Tested main (9.1.1 build), 9.0 tip, 8.6 tip tcltest: testutfnext [testbytestring \xA0\xA0]1 → 2 on main/9.0, 1 on 8.6 (default TCL_UTF_MAX=3, coupled to Tcl_UtfToUniChar as dgp asked); 8.6 has utf-6.88.0 (ucs2 → 1) and utf-6.88.1 (fullutf → 2); main/9.0 utf-6.88 expects 2. testnumutfchars \xE0\x80\x80 → 3 everywhere.
  • Cause the two goals (skip to next char start vs. stay coupled with Tcl_UtfToUniChar) were split per branch; doc/Utf.3 now says Tcl_UtfAtIndex "is equivalent to calling Tcl_UtfToUniChar index times", resolving the doc conflict from the discussion.
  • Prototype none (nothing to fix; minor: doc/Utf.3 Tcl_UtfNext paragraph lacks a period after "next character")
  • Suggested ticket comment Settled: 8.6 keeps Tcl_UtfNext coupled to Tcl_UtfToUniChar (utf-6.88.0/.1), 9.0+ steps over trail bytes (utf-6.88 → 2), and Utf.3 no longer ties Tcl_UtfAtIndex to Tcl_UtfNext. Can be closed.
⏳b6d0d8cc2c2019(windows) socket close: graceful shutdown vs. lingering (eternal TIME_WAIT issue).NOT-A-BUG (TIME_WAIT on the active closer is normal TCP; Windows behaves exactly like Linux). Ticket status Pending/Fixed is wrong: branch bug-b6d0…

Open ticket b6d0d8cc2c · 25. Channel System · filed 2019-09-13 · Pending (resolved, waiting for confirmation)

⏳ Since triage (updated 2026-11-11)

Status now: Pending (Fixed)

    Full verdict

    NOT-A-BUG (TIME_WAIT on the active closer is normal TCP; Windows behaves exactly like Linux). Ticket status Pending/Fixed is wrong: branch bug-b6d0d8cc2c was never merged

    • Platform Windows
    • Tested win/b6d0d8cc2c.tcl on Linux 8.6.18 and 9.0.5: 50 localhost connections closed by the client (server closes on EOF) leave 50 TIME_WAIT entries, exactly what the ticket reports for Windows. TIME_WAIT on the side that closes first is required by TCP, not lingering.
    • Cause sebres' fix b960d1b71e (branch bug-b6d0d8cc2c, 2020-04-27: graceful disconnect, then hard reset/no linger) was never merged into main, 9.0 or 8.6 (win/tclWinSock.c still only has the plain closesocket path), so the Pending/Fixed status is not backed by a merge. Avoiding TIME_WAIT via an RST abort can lose data and differs from Unix.
    • Prototype none (behaviour is standard TCP; the branch changes semantics)
    • Windows repro win/b6d0d8cc2c.tcl (no Thread needed; counts TIME_WAIT for the port with netstat).
    • Suggested ticket comment Branch bug-b6d0d8cc2c was never merged, so this is not fixed. On Linux 8.6/9.0 the same test also leaves one TIME_WAIT per client-closed connection, which is normal TCP for the active closer; unless Windows shows something beyond that, I'd close this as not a bug rather than switch to a reset on close.
    Windows VM, run 2
    • Platform Windows
    • Verdict NOT-A-BUG (TIME_WAIT on the active closer is normal TCP; Windows behaves exactly like Linux). Ticket status Pending/Fixed is wrong: branch bug-b6d0d8cc2c was never merged
    • Tested VM win/b6d0d8cc2c.tcl: 50 localhost connections closed by the client, server closes on EOF → 50 of 50 TIME_WAIT for the port on 9.1b1, 9.0.5 and 8.6.18 — the same count as on Linux 8.6/9.0. The server side gets EOF each time (the close is graceful: FIN, not a reset).
    • Cause TCP's TIME_WAIT for the side that closes first; avoiding it needs an abortive close (SO_LINGER 0 / RST), which can drop unsent data
    • Prototype none
    • Suggested ticket comment Branch bug-b6d0d8cc2c was never merged, so this is not "Fixed". On Windows 11 (8.6.18, 9.0.5, 9.1) a client close leaves one TIME_WAIT per connection, exactly as on Linux; that is standard TCP for the active closer, and the peer gets a proper EOF. Propose closing as not a bug; an abortive-close option (-linger 0) would be a separate feature request.
    ⬜656fe3c8162019Tcl_DictObjPut() - poor usabilityNOT-A-BUG (documented behaviour; API RFE) — unchanged on main/9.0/8.6: an existing equal key keeps the old key object, a zero-ref new key is neithe…

    Open ticket 656fe3c816 · 15. Dict Object · filed 2019-07-25

    Full verdict

    NOT-A-BUG (documented behaviour; API RFE) — unchanged on main/9.0/8.6: an existing equal key keeps the old key object, a zero-ref new key is neither stored nor freed (leaks)

    • Platform generic
    • Tested code reading of tclDictObj.c Tcl_DictObjPut on main/9.0/8.6; doc/DictObj.3 "REFERENCE COUNT MANAGEMENT" says it "may increment the reference count of keyPtr"
    • Cause generic/tclDictObj.c:Tcl_DictObjPut — CreateChainEntry() only takes a reference when isNew. dkf's internal helpers TclDictPut/TclDictPutString/TclDictGet/TclDictRemove (branch more-dict-c-funcs) are on main, 9.0 and 8.6 as MODULE_SCOPE.
    • Prototype none (changing ownership of keyPtr is an API behaviour change, see sebres's example; public helpers need a TIP)
    • Suggested ticket comment Behaviour is as documented in DictObj.3. Internal TclDictPut()/TclDictPutString() avoid the pitfall in the core (8.6, 9.0, main); what remains is a TIP to make them public, so this is really an RFE.
    ⬜ed4c6100042019REPLNOT-A-BUG (RFE: line editing/history in the interactive tclsh)

    Open ticket ed4c610004 · 20. [interp] · filed 2019-06-08

    Full verdict

    NOT-A-BUG (RFE: line editing/history in the interactive tclsh)

    • Platform generic
    • Tested by reading; plain tclsh reads stdin with no line editor on main/9.0/8.6, so arrow keys insert escape sequences (^[[A), which then fail as an unknown command. The session is not actually broken.
    • Cause tclMain.c has no readline/linenoise integration by design; tclreadline, rlwrap or tkcon provide it.
    • Prototype none (feature request, would need a TIP and an external dependency)
    • Suggested ticket comment tclsh has no built-in line editor; use rlwrap tclsh, tclreadline or tkcon. Adding one would be a feature request (TIP), not a bug. Suggest closing as "Won't Fix" or moving to RFE.
    ⬜a64ef795182019Literal string representationNOT-A-BUG (by design; performance/memory RFE)

    Open ticket a64ef79518 · 47. Bytecode Compiler · filed 2019-05-23

    Full verdict

    NOT-A-BUG (by design; performance/memory RFE)

    • Platform generic
    • Tested main, 9.0, 8.6: variable L [list e1 e2 e3] inside namespace eval gives a list with string rep "e1 e2 e3"; with a non-constant word ([list e1 e2 [string cat x y]]) there is no string rep.
    • Cause generic/tclCompCmdsGR.c:TclCompileListCmd folds an all-literal [list] into one constant list object and pushes it with PUSH_OBJ -> TclAddLiteralObj, which hashes the string rep in the literal table (needed for literal sharing).
    • Prototype none (avoiding it means either not folding constant lists, i.e. emitting INST_LIST at run time, or dropping the string rep after hashing; both are design trade-offs)
    • Suggested ticket comment An all-literal &#91;list ...&#93; is folded at compile time into a shared literal, and literals are keyed by their string, so the string rep is kept. Workaround: build the value at run time (e.g. &#91;list {*}$x&#93; or a non-constant word). Treat as an RFE.
    ⬜9eac2d8a8d2019uplevel: level sniffing for single arg, default-level case (absolute level specifier only)NOT-A-BUG (documented behaviour; changing it needs a TIP)

    Open ticket 9eac2d8a8d · 22. [proc] and [uplevel] · filed 2019-05-07

    Full verdict

    NOT-A-BUG (documented behaviour; changing it needs a TIP)

    • Platform generic
    • Tested main, 9.0, 8.6. namespace eval ::ns1 {uplevel {#0 b}} -> bad level "#0 b" on all three. uplevel {1 b} -> invalid command name "1" on main/9.0 but bad level "1 b" on 8.6 (the digit case was already relaxed in 8.7+), so only the # case remains.
    • Cause generic/tclProc.c:TclObjGetFrame returns -1 for any argument starting with # that is not a valid #N, even when it is the only argument. uplevel.n on main says "Level cannot be defaulted if the first command argument is an integer or starts with #", so this is documented.
    • Prototype none (documented behaviour; dgp asked for a TIP. The fix would be small: in TclNRUplevelObjCmd, with objc == 2, treat a non-#N argument as the script)
    • Suggested ticket comment Still the case on 9.1 and 9.0 for the # form (the integer form already falls back to the script). It is documented in uplevel.n, so changing it needs a TIP.
    ⬜43a1af29d02019Add a description for the trap patternNOT-A-BUG (already documented, also in 8.6.9)

    Open ticket 43a1af29d0 · 18. Commands M-Z · filed 2019-05-04

    Full verdict

    NOT-A-BUG (already documented, also in 8.6.9)

    • Platform docs
    • Tested doc/try.n on main and in the 8.6.9 installed man page: trap matches when the list-length(pattern) prefix of -errorcode equals pattern (whitespace normalized), and "on error is equivalent to trap {}" answers the empty-list question. Checked on 9.0.5: trap {} catches any error, prefix match works.
    • Prototype none
    • Suggested ticket comment try.n describes it: the pattern is compared with the first <i>N</i> words of -errorcode, <i>N</i> being the list length of the pattern, and trap {} (empty pattern) is the same as on error. Suggest closing.
    ⬜204ac7df242019Incorrect handling of redirection in process pipingNOT-A-BUG (documented: exec redirections apply to the whole pipeline wherever they appear)

    Open ticket 204ac7df24 · 45. Parsing and Eval · filed 2019-04-30

    Full verdict

    NOT-A-BUG (documented: exec redirections apply to the whole pipeline wherever they appear)

    • Platform generic (unix tested)
    • Tested main, 9.0, 8.6 with exec echo BLA >ble | ./pe.sh etc. (pe.sh prints COUT/CERR). Same results on all three: >ble gets the last command's stdout (COUT), stderr goes to the exec error result. 1>ble is not an exec redirection, so it is passed to echo as an argument. Stderr only ended up in the file in the reporter's interactive session because unknown runs commands with >&@stdout.
    • Cause exec.n defines > fileName as "standard output from the last command" and 2> as "standard error from all commands"; redirection operators are pulled out of the arg list regardless of position (tclPipe.c:TclCreatePipeline). Per-stage redirection like sh's is not supported.
    • Prototype none (behaviour change would need a TIP)
    • Suggested ticket comment This is how &#91;exec&#93; is specified: redirections apply to the whole pipeline wherever they appear, > to the last command's stdout and 2> to all commands' stderr; 1> is not a recognised redirection. In interactive tclsh, unknown adds >&@stdout, which is why stderr seemed to land in the file. Per-stage redirection needs sh -c. Suggest closing as "Works as designed".
    ⬜de232b49f22019write-only nonblocking refchan and Tcl internal buffersNOT-A-BUG (behaviour by design; TCT-approved rollback, change needs a TIP)

    Open ticket de232b49f2 · 04. Async Events · filed 2019-04-26

    Full verdict

    NOT-A-BUG (behaviour by design; TCT-approved rollback, change needs a TIP)

    • Platform generic
    • Tested reporter's script with a 3 s timeout on main, 9.0.5, 8.6.18, 8.6.14, 8.6.9: all fail the same way, coroutine "::c1" is already running (the synchronous chan postevent from the watch proc runs the writable handler inside the coroutine), then no further writable events -> timeout.
    • Cause refchan writable events are only generated by chan postevent, which is synchronous; data parked in Tcl's output buffer does not trigger new events. pooryorick's d0dd6d19a4 (2019) generated them on a timer, caused 080f846fd5 and socket-14.11.0 failures and was rolled back in 2024 ("Rationale for rollback of refchan event generation in core" wiki page, TCT-approved). The ticket was closed as invalid/RFE by Jan and Harald several times and reopened by the reporter (last: 2024-11-15 "does not affect the unchained branch").
    • Prototype none (design dispute; reverting the rollback is not on the table without a TIP)
    • Suggested ticket comment none recommended; the core team's decision (2024-05/06 comments) stands. If anything, point to the rollback rationale page and suggest -buffering none on such refchans or a TIP for asynchronous chan postevent.
    ⬜15eb59f0502019Clock command stop workingNOT-A-BUG (application does not initialize the interpreter)

    Open ticket 15eb59f050 · 38. Init - Library - Autoload · filed 2019-01-06

    Full verdict

    NOT-A-BUG (application does not initialize the interpreter)

    • Platform generic (embedding)
    • Tested interp with unknown removed (what an interp without Tcl_Init looks like): clock format fails with "invalid command name ::tcl::clock::GetSystemTimeZone" on main/9.0.5, and with an auto-load error on 8.6.18; with normal initialization it works.
    • Cause since 8.5 parts of clock live in library/clock.tcl (8.6) or are auto-loaded/registered from Tcl scripts (9.x), which needs Tcl_Init (init.tcl, unknown/auto_load). 8.4's clock was C only. The reporter's tclEvalGetResult wrapper was never shown; no reply since 2019.
    • Prototype none
    • Suggested ticket comment Since 8.5 clock needs Tcl scripts from the library loaded on demand, so the embedding application must call Tcl_Init() (and check its result) after Tcl_CreateInterp(). No reply since 2019; suggest closing.
    ⬜67bab20ac62018if-construct: bizzare cases "without" body (mistake, typo, etc)NOT-A-BUG (works as documented; the proposal is an RFE)

    Open ticket 67bab20ac6 · 17. Commands I-L · filed 2018-12-21

    Full verdict

    NOT-A-BUG (works as documented; the proposal is an RFE)

    • Platform generic
    • Tested main build, 9.0.5, 8.6.18: if 0 else {...} runs the else-script, if 1 else {...} → invalid command name "else", as reported.
    • Cause then/else are optional noise words in if.n, so in if 0 else {...} the word else is the body. Making else in body position a keyword changes language semantics (a TIP matter). No branch with the announced work exists in the repository.
    • Prototype none (language change)
    • Suggested ticket comment This is how if is specified: else is optional, so in if c else {x} the word else is the then-body. Treating it as a keyword would be a language change needing a TIP; no branch for it was ever committed. Suggest closing or turning it into an RFE.
    ⬜c3e39977562018Cannot resolve hostnames after waking from suspend on linuxNOT-A-BUG (C library behaviour: glibc < 2.26 read /etc/resolv.conf only once per process; glibc 2.26 (2017) re-reads it when it changes)

    Open ticket c3e3997756 · 18. Commands M-Z · filed 2018-09-07

    Full verdict

    NOT-A-BUG (C library behaviour: glibc < 2.26 read /etc/resolv.conf only once per process; glibc 2.26 (2017) re-reads it when it changes)

    • Platform unix (Linux)
    • Tested not reproducible here without suspend/resume and a changing resolv.conf; code reading only (Tcl only calls getaddrinfo() in generic/tclIOSock.c:TclCreateSocketAddress, it keeps no resolver state of its own; local glibc is 2.43)
    • Cause after resume NetworkManager/dhclient rewrites /etc/resolv.conf (new nameserver); old glibc kept the stale nameserver in every running process, so getaddrinfo fails with EAI_AGAIN in the old tclsh while a new process works. Any long-running C program behaves the same.
    • Prototype none (Tcl could call res_init() on EAI_AGAIN, but that is a workaround for old C libraries and not needed with glibc >= 2.26)
    • Suggested ticket comment Tcl only calls getaddrinfo() and keeps no resolver state. The symptom matches glibc before 2.26, which read /etc/resolv.conf once per process; newer glibc reloads it when it changes. Please reopen if it still happens with a current glibc. Proposing to close.
    ⬜bb9129e46c2018TCL Memory raise up and didn't come down for everNOT-A-BUG (by design: zippy is a high-water-mark allocator; behaviour unchanged in main, 8.6)

    Open ticket bb9129e46c · 41. Memory Allocation · filed 2018-06-04

    Full verdict

    NOT-A-BUG (by design: zippy is a high-water-mark allocator; behaviour unchanged in main, 8.6)

    • Platform generic
    • Tested 12 rounds of 20 joinable threads that each build and free a 200k-element list. RSS after the rounds: main -O2 1.28 -> 1.59 GB, flattening after round 6; 8.6.18 0.87 -> 1.10 GB, same shape; main -O2 -DPURIFY (system malloc) stays at 47 MB.
    • Cause generic/tclThreadAlloc.c: when a thread exits, TclFreeAllocCache moves its blocks to the shared cache. Those blocks are reused, never returned to the OS, and there is no API to trim them. Growth is bounded (it flattens at the peak working set plus some fragmentation), so it is not a leak.
    • Prototype none (would need a trim API or allocator redesign, see also 3d1c409f1a)
    • Suggested ticket comment Memory freed by exiting threads goes to the allocator's shared pool and is reused, never given back to the OS. Repeating the workload does not grow it without bound (it flattens after a few rounds). A build with CFLAGS=-DPURIFY uses the system malloc and keeps RSS flat. Note that on Unix this is currently the only switch: -DUSE_THREAD_ALLOC=0 still selects zippy, because tclInt.h tests only defined(USE_THREAD_ALLOC).
    ⬜552ed5eac52018Strange discrepancy regarding INST_START_CMD in body of cycles if compiled with or without invocation in condition/iterator...patchNOT-A-BUG (intentional optimization) / DOCS for &#91;info cmdcount&#93;

    Open ticket 552ed5eac5 · 47. Bytecode Compiler · filed 2018-04-20

    🧪 Prototype check (2026-10-24)

    552ed5eac5.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before work-a2/v2/cc.tcl (two 9-iteration while loops with 4 sets, one with [t] in the condition) -> [info cmdcount] deltas a=61 b=4: the compactible loop (no invocation) is recompiled without INST_START_CMD, as triaged (intentional optimization); info.n still says "the total number of commands evaluated"
    • Patch applied cleanly (3 changed lines, unchanged)
    • After groff -man -Tutf8 doc/info.n renders the note correctly
    • Test none possible (docs only)
    Full verdict

    NOT-A-BUG (intentional optimization) / DOCS for &#91;info cmdcount&#93;

    • Platform generic
    • Tested main, 9.0, 8.6: the reporter's script gives 62 5 38 5 (8.5.19: 62 52 38 37). Command limits still work: a child interp with a body compiled before interp limit c commands was set still hits "command count limit exceeded" on main/9.0/8.6.
    • Cause generic/tclCompile.c:TclInitByteCodeObj/TclSetByteCodeFromAny recompiles "compactible" code (IsCompactibleCompileEnv: no invocations or other hazards) with atCmdStart = 2, i.e. without INST_START_CMD. This has been done since 2f5ee8efad (2013-08-07, 8.6.1), and since 2020 only in interps with no parent and no command/time limit. A body that contains an invocation is not compactible, so it keeps INST_START_CMD. That explains the discrepancy, and the only visible effect is the &#91;info cmdcount&#93; count.
    • Prototype prototypes/552ed5eac5.patch (3 changed lines, docs only: info.n says the cmdcount is approximate)
    • Suggested ticket comment The missing startCommand comes from the deliberate "compactible code" recompile in tclCompile.c (2f5ee8efad), which is skipped in child interps and whenever a command/time limit is set, so limits are not affected. Only &#91;info cmdcount&#93; undercounts; proposed a doc note in info.n saying the count is approximate.
    ⬜73384a82ad2018clock scan calendaric calculation errorNOT-A-BUG (by design: "ago" negates all relative units before it, including yesterday/tomorrow, which are ±1 day units; test clock-34.47 checks "2 …

    Open ticket 73384a82ad · 16. Commands A-H · filed 2018-04-16

    Full verdict

    NOT-A-BUG (by design: "ago" negates all relative units before it, including yesterday/tomorrow, which are ±1 day units; test clock-34.47 checks "2 days 2 hours ago" = -50 h). Same behaviour in main, 9.0, 8.6.

    • Platform generic
    • Tested main, 9.0 tip, 8.6 tip with -base 2018-04-16: "yesterday 3 months ago" -> 2018.01.17, "tomorrow 3 months ago" -> 2018.01.15; "yesterday -3 months" -> 2018.01.15. For comparison, GNU date applies "ago" to the preceding item only ("yesterday 3 months ago" -> 2018.01.15).
    • Cause generic/tclGetDate.y relspec relunits tAGO multiplies the accumulated yyRelSeconds/yyRelMonth/yyRelDay by -1, and "yesterday"/"tomorrow" are tDAY_UNIT entries (-1/+1 day) that add to yyRelDay
    • Prototype none (changing it would break the tested multi-unit "ago" semantics, and there is no bison here to regenerate tclDate.c; a doc sentence in clock.n would help: "ago" negates all preceding relative units, and yesterday/tomorrow count as -1/+1 day)
    • Suggested ticket comment This is how the free-form scanner works: "ago" negates all relative units before it (tested by clock-34.47), and yesterday/tomorrow are just -1/+1 day units. Write "3 months ago yesterday" or "yesterday -3 months" instead. Changing it would be incompatible; documenting it in clock.n seems the right fix.
    ⬜e593adf1032018Resources of a namespace are still available after the namespace is deleted.NOT-A-BUG (documented behaviour; the reporter says so too). Behaviour unchanged in main, 9.0, 8.6.

    Open ticket e593adf103 · 18. Commands M-Z · filed 2018-01-02

    Full verdict

    NOT-A-BUG (documented behaviour; the reporter says so too). Behaviour unchanged in main, 9.0, 8.6.

    • Platform generic
    • Tested main, 9.0 tip, 8.6 tip: the script prints "hello", namespace current is still ::n1 after namespace delete, and namespace exists n1 is 0 afterwards
    • Cause generic/tclNamesp.c keeps a namespace alive (marked dying, not findable by name) while a frame is still executing in it; namespace.n: "If a procedure is currently executing inside the namespace, the namespace will be kept alive until the procedure returns". The proposed redesign (branches bug-e593adf103 3825265b17, 2018-01-06, and bug-e593adf103-core-8, last 37f9d3e1fe, 2018-11-01) was never merged; the core-8 branch was closed by jan.nijtmans on 2024-12-13.
    • Prototype none (behaviour redesign, not a small fix)
    • Suggested ticket comment This is the documented behaviour (namespace.n: the namespace is kept alive until the procedure returns), unchanged in 8.6, 9.0 and main. The redesign branches were never merged and bug-e593adf103-core-8 has been closed. Proposing to close; a change of these semantics would need a TIP.
    ⬜85bbfda8522017Should [package prefer stable] prompt repeated searches?NOT-A-BUG (by design, per dgp's analysis; same on main, 9.0, 8.6)

    Open ticket 85bbfda852 · 39. Package Manager · filed 2017-11-14

    Full verdict

    NOT-A-BUG (by design, per dgp's analysis; same on main, 9.0, 8.6)

    • Platform generic
    • Tested main, 9.0 tip, 8.6 tip: with prefer stable and only 1.1a0 registered, package require p 1.0 returns 1.1a0 without calling the package unknown handler
    • Cause generic/tclPkg.c (PkgRequireCore) calls the unknown handler only when no registered version satisfies the requirement; package prefer only chooses among the versions already known
    • Prototype none (the question is a design decision; calling the handler on every require would be expensive)
    • Suggested ticket comment Unchanged in 8.6, 9.0 and main: &#91;package unknown&#93; is consulted only when no known version satisfies the request, and &#91;package prefer&#93; only chooses among the known versions. That is the intended trade-off; proposing to close as "works as designed" unless somebody wants to write a TIP.
    ⬜3ae12512642017string-29.6 fails due to interp aliasNOT-A-BUG (test-design issue; dgp's workaround is on main: string-29.5..29.15 run only in the $noComp (testevalex) pass, see if {$noComp} at te…

    Open ticket 3ae1251264 · 47. Bytecode Compiler · filed 2017-09-20

    Full verdict

    NOT-A-BUG (test-design issue; dgp's workaround is on main: string-29.5..29.15 run only in the $noComp (testevalex) pass, see if {$noComp} at tests/string.test:2452)

    • Platform generic
    • Tested main (string.test 1411 passed, 0 failed); try {string cat [list] [list x]} at top level still yields a string rep on main and 8.6, a proc-compiled call does not
    • Cause in the compiled path the [list x] argument goes through literal/bytecode handling that gives the result a string rep; the test is about TclStringCatObjv internals, which only the direct-eval path exercises cleanly
    • Prototype none (nothing to fix; the test suite passes)
    • Suggested ticket comment The efficiency tests string-29.5..29.15 now run only in the testevalex (non-compiled) pass of string.test, which is what they are meant to check; string.test passes on main. Closing.
    ⬜606b6bf44c2017documentation: chan configure -connecting, -errorNOT-A-BUG (already documented: socket.n "CONFIGURATION OPTIONS" describes -error, -sockname, -peername, -connecting; chan.n and fconfigure.n refer …

    Open ticket 606b6bf44c · 16. Commands A-H · filed 2017-09-10

    Full verdict

    NOT-A-BUG (already documented: socket.n "CONFIGURATION OPTIONS" describes -error, -sockname, -peername, -connecting; chan.n and fconfigure.n refer to socket(n) for socket-specific options)

    • Platform docs
    • Tested doc/socket.n, chan.n, fconfigure.n on main, 9.0, 8.6. -connecting was documented in socket.n by TIP 427 (b689b115bd, merged 9fcbf61071) before the ticket was filed; -error since dd5a31ec23.
    • Cause none; type-specific options are deliberately listed on the socket page, not in chan.n.
    • Prototype none (nothing to fix)
    • Suggested ticket comment Both options are documented in the CONFIGURATION OPTIONS section of socket(n), and chan(n)/fconfigure(n) refer there for socket-specific options. Closing as works-as-documented unless something specific is missing.
    ⬜fa5c8a82302017memory cost and execution time exploding of nested namespacesNOT-A-BUG (expected cost of the design: every namespace stores its fully-qualified name, so memory is quadratic in nesting depth)

    Open ticket fa5c8a8230 · 21. [namespace] · filed 2017-05-18

    Full verdict

    NOT-A-BUG (expected cost of the design: every namespace stores its fully-qualified name, so memory is quadratic in nesting depth)

    • Platform generic
    • Tested main, 8.6 tip: 3000 nesting levels take +69 MB / +67 MB RSS and about 1 s. The deepest name is 43890 chars, and the sum of all full names (~3000*43890/2 ≈ 66 MB) matches the RSS growth exactly. So there is no leak: the full names account for all of it.
    • Cause Namespace.fullName (tclInt.h) holds the complete qualified name of each namespace. Each namespace eval of an N-level name also parses the whole name, so time is quadratic as well.
    • Prototype none (computing fullName on demand would be a large redesign for a pathological case)
    • Suggested ticket comment The memory is the fully-qualified names stored in every namespace (Namespace.fullName). Their total grows quadratically with nesting depth, and it matches the measured RSS exactly, so nothing leaks. Fixing this would mean redesigning how namespaces store their names, which isn't worth it for 10000-deep nesting. Proposing to close.
    ⬜268108fc932017string map - Investigate Aho-Corasick as possible speed boosterNOT-A-BUG (performance RFE; unchanged on main: StringMapCmd in generic/tclCmdMZ.c still loops over every key at each position, on the Unicode rep)

    Open ticket 268108fc93 · 18. Commands M-Z · filed 2017-05-06

    Full verdict

    NOT-A-BUG (performance RFE; unchanged on main: StringMapCmd in generic/tclCmdMZ.c still loops over every key at each position, on the Unicode rep)

    • Platform generic
    • Tested code reading on main
    • Cause n/a. An Aho-Corasick version would also have to keep the documented "first key in list order that matches at this position wins" rule (not leftmost-longest), and the automaton build cost has to be cached in a map intrep to pay off.
    • Prototype none (RFE, beyond the size limit)
    • Suggested ticket comment Still valid as an RFE: &#91;string map&#93; is O(len × keys). An Aho-Corasick version must keep the "first matching key in list order" rule and cache the automaton in the map's intrep. Better retyped as an RFE than kept as a Bug.
    ⬜a0db647efa2017tcl_interactive wrong on windows when using mintty (MSYS2, win7x64)NOT-A-BUG (the reporter withdrew it on 2017-05-03: mintty without ConPTY gives the program named pipes, which no Windows API can tell apart from re…

    Open ticket a0db647efa · 07. Variables · filed 2017-05-03

    Full verdict

    NOT-A-BUG (the reporter withdrew it on 2017-05-03: mintty without ConPTY gives the program named pipes, which no Windows API can tell apart from redirection)

    • Platform Windows
    • Tested code review only. generic/tclMain.c still uses isatty(0). On Unix the behaviour is correct.
    • Cause in pipe mode, mintty's stdin is a pipe and not a console, so isatty()/GetConsoleMode() correctly report "not a tty". Newer MSYS2/Cygwin run console programs through a ConPTY pseudo console (no pipes), and there isatty() should succeed. The usual workarounds are winpty tclsh, or set tcl_interactive 1 in the shell.
    • Prototype none
    • Suggested ticket comment As the reporter says in the follow-up, mintty in pipe mode cannot be detected as a terminal: stdin really is a pipe. With ConPTY (current MSYS2/Cygwin) or winpty, isatty() works. Closing as Works as Expected.
    ⬜1cc44617e22017Mechanism with 64 bit support in tcl.h does not work outside of coreNOT-A-BUG (misunderstanding of TCL_WIDE_INT_IS_LONG; resolution already "Works For Me")

    Open ticket 1cc44617e2 · 69. Other · filed 2017-03-23

    Full verdict

    NOT-A-BUG (misunderstanding of TCL_WIDE_INT_IS_LONG; resolution already "Works For Me")

    • Platform generic
    • Tested code reading of generic/tcl.h on main, 9.0, 8.6.
    • Cause TCL_WIDE_INT_IS_LONG means "long is 64-bit" (LP64), so it is correctly undefined on Win64 (LLP64); Jan explained this in 2017 and added a Tk test case. The misleading "real 64-bit system" comment the reporter tripped over has since been rewritten in tcl.h on all three branches ("LP64 system such as modern Solaris or Linux ... not including Win64"). The C99 stdint.h modernization proposal is an RFE.
    • Prototype none
    • Suggested ticket comment Works as designed: TCL_WIDE_INT_IS_LONG means long is 64-bit (LP64), not "64-bit platform", so it is undefined on Win64; the tcl.h comment now says so. Closing.
    ⬜b9bfaf5fdb2016EAGAIN throw in reflected channel in blocking mode causes eof (data loss)NOT-A-BUG (no EOF and no data loss: a blocking [read] returns early with the data buffered so far, [eof] stays 0 and the rest arrives on the next […

    Open ticket b9bfaf5fdb · 25. Channel System · filed 2016-06-23

    Full verdict

    NOT-A-BUG (no EOF and no data loss: a blocking [read] returns early with the data buffered so far, [eof] stays 0 and the rest arrives on the next [read]; a handler told blocking 1 is expected to block instead of throwing EAGAIN)

    • Platform generic
    • Tested single-threaded refchan (work-a4/tb9.tcl: data, EAGAIN, EAGAIN, data, "") on main, 9.0.5, 8.6.18, 8.5.19 — all the same: first read → "1st data", eof=0; then "" eof=0; then ", 2nd data" eof=1. The attached threaded demo only does one [read] in blocking mode and never checks [eof], which is why it reports "ERROR".
    • Cause generic/tclIO.c treats EAGAIN from the driver as "no more data right now" even on a blocking channel and ends the read; it does not retry (which would busy-loop) or set EOF.
    • Prototype none (behaviour change/RFE; a refchan.n sentence that handlers in blocking mode should block, and that EAGAIN there makes [read] return short, would cover it)
    • Suggested ticket comment Not data loss: with EAGAIN in blocking mode &#91;read&#93; returns short, &#91;eof&#93; stays 0 and the remaining data is returned by the next &#91;read&#93; (same on 8.5–9.1). A handler in blocking mode should block; proposing to close, or to document this in refchan.n.
    ⬜85b87a75702016TclByteArrayMatch - unused? premature opt?NOT-A-BUG (code-quality question; the routine is still used)

    Open ticket 85b87a7570 · 10. Objects · filed 2016-03-07

    Full verdict

    NOT-A-BUG (code-quality question; the routine is still used)

    • Platform generic
    • Tested code review of main. TclByteArrayMatch (generic/tclUtil.c) is called from TclStringMatchObj and INST_STR_MATCH (tclExecute.c), i.e. compiled and direct string match and glob-reducible regexps, when both string and pattern are pure bytearrays and the match is case-sensitive, and recursively for *.
    • Cause n/a. The restriction to pure bytearrays dgp describes is already in place (TclIsPureByteArray checks).
    • Prototype none (removing it is optional cleanup; the 2923d52832 prototype at least leaves a single caller, TclStringMatchObj)
    • Suggested ticket comment TclByteArrayMatch is still reached for pure-bytearray string match, both compiled and direct (e.g. string match [binary format a* a*] [binary format a* ab]), and only for pure bytearrays. As dgp noted, it can stay until it gets in the way. Proposing to close as Works as Expected.
    ⬜cb7aac6f2c2015Alternate float format %#.0f returns "0." for "0.0"NOT-A-BUG ("0." is the C99 result for %#.0f); the MSVC %#g bug is NOT-REPRODUCED with the current UCRT (MSVC 2022)

    Open ticket cb7aac6f2c · 16. Commands A-H · filed 2015-08-31

    Full verdict

    NOT-A-BUG ("0." is the C99 result for %#.0f); the MSVC %#g bug is NOT-REPRODUCED with the current UCRT (MSVC 2022)

    • Platform Windows
    • Tested main (own build), 9.0.5, 8.6.18, 8.5.19 on Linux: 0.0000e+00 0.0000 0.000|0.e+00 0. 0., matches glibc printf and format-6.3/6.5. format.test passes on main (269/269).
    • Cause Tcl_AppendFormatToObj hands e/f/g conversions to the C library's snprintf. Old MSVC CRTs print one significant digit too many for %#.Ng of 0.0. format.test already skips format-6.3/6.5 on MSVC builds through the knownMsvcBug constraint ([tcl::build-info msvc] eq 0, check-in 30e09611ba, 2021).
    • Prototype none (C library behaviour; tests already skip it on MSVC). Windows repro: win/cb7aac6f2c.tcl (shows whether the current UCRT is still wrong, i.e. whether the constraint can go)
    • Suggested ticket comment %#.0f giving "0." is what C99 requires (# always keeps the decimal point). The real difference is MSVC's %#g of 0.0, which is a CRT bug; format-6.3/6.5 are skipped on MSVC builds through the knownMsvcBug constraint. Close as Won't Fix / third-party bug.
    Windows VM, run 4
    • Platform Windows (C runtime)
    • Verdict NOT-A-BUG ("0." is the C99 result for %#.0f); the MSVC %#g bug is NOT-REPRODUCED with the current UCRT (MSVC 2022)
    • Tested VM win/cb7aac6f2c.tcl: %#.4e %#.4f %#.4g|%#.0e %#.0f %#.0g of 0.0 gives 0.0000e+00 0.0000 0.000|0.e+00 0. 0. (same as glibc) on 9.1b1, 9.0.5 and 8.6.18 (all MSVC 1944, UCRT). format.test: 9.1 skips format-6.3/6.5 via knownMsvcBug (269 total, 0 failed); 8.6 runs them under notWinCI locally and passes (266, 0 failed).
    • Cause Tcl hands e/f/g conversions to the CRT; the old MSVC CRT printed one digit too many for %#.Ng of 0.0. The Windows 11 UCRT with MSVC 2022 is correct.
    • Prototype none (optional: drop or narrow the knownMsvcBug constraint in format.test, e.g. to msvc < 1900; the 2021 CI failure may have been an older toolset)
    • Suggested ticket comment %#.0f → "0." is correct C99 behaviour. The MSVC %#g problem is gone with the UCRT of MSVC 2022: format-6.3/6.5 give the expected results on Windows 11 with 8.6.18, 9.0.5 and 9.1, so the knownMsvcBug constraint in format.test could probably be relaxed. Close as Works For Me / third-party.
    ⬜448ba5cbfd2015MINGW build fails for 32bit on 64bit systemNOT-A-BUG (toolchain-specific; workaround exists)

    Open ticket 448ba5cbfd · 53. Configuration and Build Tools · filed 2015-07-10

    Full verdict

    NOT-A-BUG (toolchain-specific; workaround exists)

    • Platform build (Windows, TDM-gcc multilib with CFLAGS=-m32)
    • Tested main win/configure cross (x86_64-w64-mingw32 on Linux): RC="x86_64-w64-mingw32-windres -F pe-i386" configure ends up in the Makefile and produces a pe-i386 tcl.res.o; this mingw has no 32-bit multilib, so the full -m32 link was not tried
    • Cause win/Makefile.in still has no RCFLAGS, but RC is taken from the environment (AC_CHECK_TOOL), so RC="windres -F pe-i386" works. With -m32 the _WIN64 check leaves do64bit=no, so win32/zdll.lib is used, which i686 mingw-w64 binutils link. The supported route is a native i686 toolchain (MSYS2 mingw32 / i686-w64-mingw32), which works out of the box.
    • Prototype none
    • Suggested ticket comment For a multilib gcc, pass RC="windres -F pe-i386" (and CC="gcc -m32") to configure; RC is taken from the environment. The i686 mingw-w64 toolchains build 32-bit out of the box. Suggest closing.
    ⬜fd2d126f232015Wrong proc doesn't guives an errorNOT-A-BUG (documented: proc.n says defaulted arguments followed by required ones become de-facto required, "this may change in a future version")

    Open ticket fd2d126f23 · 22. [proc] and [uplevel] · filed 2015-05-16

    Full verdict

    NOT-A-BUG (documented: proc.n says defaulted arguments followed by required ones become de-facto required, "this may change in a future version")

    • Platform generic
    • Tested main build, 9.0.5, 8.6.18: proc f {a {b 1} c} is accepted, f x y fails with should be "f a ?b? c", f x y z works.
    • Cause tclProc.c TclCreateProc does not check the order of defaulted formals; by design (each formal is a list).
    • Prototype none (an error at definition time would break existing scripts; the misleading ?b? in the wrong-args message is pinned by test proc-old-30.8)
    • Suggested ticket comment This is documented in proc.n: a defaulted argument followed by required ones is de-facto required. Making it a definition-time error would be an incompatible change and needs a TIP; I suggest closing as "Works as designed", or converting it to a Feature Request.
    ⬜b32fffae812015clock scan returns wrong date if dayOfMonth is missingNOT-A-BUG (8.6 follows documented rule [7] of clock(n): without a complete date set the base date is used). 9.x changed: Apr12 -format %b%y now g…

    Open ticket b32fffae81 · 06. Time Measurement · filed 2015-05-05

    Full verdict

    NOT-A-BUG (8.6 follows documented rule [7] of clock(n): without a complete date set the base date is used). 9.x changed: Apr12 -format %b%y now gives the given year and month with the base's day (16 April 2012 for base 2026-10-16), still not day 1, and the docs don't describe that.

    • Platform generic
    • Tested main c3bf6dd052, 9.0.5, 8.6.18 (work-a6/t6.tcl, also %Y-%m and %m/%Y)
    • Cause library/clock.tcl / generic/tclClockFmt.c: no field set for "year + month" in the scan rules; 9.x fills the day of month from the base.
    • Prototype none (behaviour decision; sebres rejected day 1 as default)
    • Suggested ticket comment In 8.6 this is the documented fallback (rule &#91;7&#93;: the base date is used). 9.x now takes year and month from the input and the day from the base, which the manual does not describe. Use a format with %d, or add "01" to the input.
    ⬜cafef7c5852015Unexplained code in StringRplcCmd function (tclCmdMZ.c)patchNOT-A-BUG (second fetch is needed; already commented in 8.5/8.6, and the code is gone on 8.7+)

    Open ticket cafef7c585 · 18. Commands M-Z · filed 2015-02-17

    🧪 Prototype check (2026-10-24)

    cafef7c585.patch: VERIFIED (test-only patch; NOT-A-BUG ticket, can be closed after the test is added)

    • Base main (tip d6dd6d0f71); mutation check on core-8-6-branch (tip 326e1bf923)
    • Before not a bug (triage NOT-A-BUG): set s end-1; list [string replace $s $s $s X] [string replace $s 0 $s Y] -> endX1 Y1 on main and 8.6. The questioned second Tcl_GetUnicodeFromObj exists only in 8.6 (tclCmdMZ.c:2390, already commented); main uses TclStringReplace. The patch is test-only.
    • Patch applied cleanly on main (4 changed lines, unchanged). It does not apply to 8.6 string.test (hunk context differs); an 8.6 backport would need the same 4 lines placed by hand.
    • After string.test on main: 1421 total, 1413 passed, 8 skipped, 0 failed
    • Test string-14.25 passes on main. Mutation check on 8.6: with the second fetch at tclCmdMZ.c:2390 commented out the same script gives endX Y1 instead of endX1 Y1, so the test catches the regression the ticket asked about (on 8.6, where that code lives)
    Full verdict

    NOT-A-BUG (second fetch is needed; already commented in 8.5/8.6, and the code is gone on 8.7+)

    • Platform generic
    • Tested read the code on main, 9.0, 8.6; ran set s end-1; string replace $s $s $s X on main, 9.0.5, 8.6.18 (all give endX1)
    • Cause in 8.6, StringRplcCmd fetches the Unicode rep again because objv[1] may be the same Tcl_Obj as an index argument, and TclGetIntForIndexM then shimmers it. dgp added a comment saying so in dfe3934edb (8.5, 2018-03-13), merged to 8.6 in d50c1e61f1. On 8.7/9.x the command uses Tcl_GetCharLength + TclStringReplace (190272213a, 2018-03-14), with no duplicate fetch. dgp's requested test for the shimmering case was never added. (The string insert side request was done by TIP 504 in 8.7.)
    • Prototype prototypes/cafef7c585.patch (test only, 4 changed lines, test string-14.25; string.test passes on main)
    • Suggested ticket comment The second Tcl_GetUnicodeFromObj() is needed when the string is also an index argument (shimmering); 8.6 has a comment explaining this since dfe3934edb, and 8.7+ uses TclStringReplace(). string insert exists since 8.7 (TIP 504). Proposed test string-14.25 for the shared-object case; then close.
    ⬜e8efa378a92013Duplicate std* channels in different ThreadsNOT-A-BUG (documented in Tcl_StandardChannels(3): in a thread without platform std channels the first three channels opened become stdin/stdout/std…

    Open ticket e8efa378a9 · 25. Channel System · filed 2013-08-29

    Full verdict

    NOT-A-BUG (documented in Tcl_StandardChannels(3): in a thread without platform std channels the first three channels opened become stdin/stdout/stderr); the remaining request is an RFE needing a tcl-core decision

    • Platform Windows (wish, no console std handles); design question for all platforms
    • Tested Linux main build + Thread 3.0.7: a new thread has its own stdin/stdout/stderr, a channel opened there gets a normal name (file7). The wish-on-Windows case (no std handles) is as JKU describes and as documented.
    • Cause generic/tclIO.c Tcl_RegisterChannel/CheckForStdChannelsBeingClosed + TclpGetDefaultStdChannel returning NULL in GUI threads, so the next opened channels get the std names; standard channels refuse thread transfer.
    • Existing work none (Open; Serhiy and JKU agreed on 2026-10-14 that a change must be discussed on tcl-core first)
    • Prototype none (behaviour change, needs a decision)
    • Suggested ticket comment none for now; wait for the tcl-core discussion (or change Type to RFE).
    ⬜3612021fff2013info frame gives mismatching line number with info bodyNOT-A-BUG (by design; resolution is already "Wont Fix", the ticket was just never closed)

    Open ticket 3612021fff · 17. Commands I-L · filed 2013-04-27

    Full verdict

    NOT-A-BUG (by design; resolution is already "Wont Fix", the ticket was just never closed)

    • Platform generic
    • Tested main (own build), 9.0.5: unchanged, info frame reports lines 2 and 5 (file lines), while info body has the backslash-newline joined, so the second command is on line 4 of the body.
    • Cause the parser folds backslash-newline. Since 2009-08-25 (TIP 280 continuation-line work, 9d8e49c947) info frame deliberately counts physical source lines, so that debuggers point at the line in the editor.
    • Prototype none
    • Suggested ticket comment Works as designed: info frame line numbers follow the source text, not the folded info body. Resolution is already Won't Fix, so closing.
    ⬜3604162fff2013INST_STORE, TclObjLookupVarEx, createPart*NOT-A-BUG (the flags already agree)

    Open ticket 3604162fff · 47. Bytecode Compiler · filed 2013-02-11

    Full verdict

    NOT-A-BUG (the flags already agree)

    • Platform generic
    • Tested code reading of 8.6.0 (core-8-6-0), 8.6 tip and main
    • Cause doStoreStk (tclExecute.c, INST_STORE_STK etc.) calls TclObjLookupVarEx with createPart1=1. The uncompiled [set] write path goes through Tcl_ObjSetVar2, which also uses createPart1=1 (tclVar.c), and so do [append]/[lappend]. createPart1=0 shows up only on the read path (one-argument [set] -> Tcl_ObjGetVar2), and that matches the compiled INST_LOAD_STK (createPart1=0), not doStoreStk. This was already true in 8.6.0.
    • Prototype none (nothing to fix)
    • Suggested ticket comment The uncompiled write path (Tcl_SetObjCmd -> Tcl_ObjSetVar2) uses createPart1=1, the same as doStoreStk. createPart1=0 is only on the read path (Tcl_ObjGetVar2), and that matches INST_LOAD_STK. No mismatch in 8.6.0 or now, so I propose closing this.
    ⬜3588512fff2012expr { ( ! true ) == false }patchNOT-A-BUG (by design: == is numeric-preferring and falls back to string comparison; "false" is not numeric). The man page is vague, so a DOCS clari…

    Open ticket 3588512fff · 47. Bytecode Compiler · filed 2012-11-19

    🧪 Prototype check (2026-10-24)

    3588512fff.patch: VERIFIED

    • Base main (tip d6dd6d0f71)
    • Before work-a3/v3588.tcl -> (!true)==false -> 0, 0 == false -> 0, false==false -> 1 (by design; expr.n on the tip only says "Boolean equal and not equal")
    • Patch applied cleanly (6 changed lines)
    • After groff renders, no new warnings. Every statement checked on the tip: 1 == "1.0" -> 1 (numeric), 1 == "1x" -> 0 (string), bool(false)==(!true) -> 1, !!false == 0 -> 1
    • Test none possible (docs)
    Full verdict

    NOT-A-BUG (by design: == is numeric-preferring and falls back to string comparison; "false" is not numeric). The man page is vague, so a DOCS clarification is worth adding.

    • Platform docs
    • Tested main, 9.0.5, 8.6.18, 8.5.19: {(!true)==false} -> 0, {false==false} -> 1, {0==false} -> 0, {true==1} -> 0
    • Cause generic/tclExecute.c, the INST_EQ path: boolean words are not numbers, so a string comparison is done. doc/expr.n describes == only as "Boolean equal and not equal".
    • Prototype prototypes/3588512fff.patch (6 changed lines, docs only). It describes == / != as numeric-preferring with a string fallback, notes that true/false are not numeric, and recommends bool() or !!.
    • Suggested ticket comment This is by design: == compares numerically when both sides are numbers and otherwise as strings, and "false" is not a number. Proposed expr.n clarification (numeric-preferring, plus a hint to use bool() or !!) is available. Otherwise I'd close as Works as designed; changing == itself would need a TIP.
    ⬜3574242fff2012exact match is greedy by reluctant exact match quantifierNOT-A-BUG (documented in re_syntax.n: "{m} or {m}? has the same preference (possibly none) as the atom itself", while {m,m}? prefers shortest) — th…

    Open ticket 3574242fff · 43. Regexp · filed 2012-10-03

    Full verdict

    NOT-A-BUG (documented in re_syntax.n: "{m} or {m}? has the same preference (possibly none) as the atom itself", while {m,m}? prefers shortest) — the output is unchanged on main, 9.0, 8.6 and 8.4: abcd abc d vs abcd ab cd

    • Platform generic
    • Tested main build, 9.0.5, 8.6.18, 8.4.20 with all examples from the ticket. Side note: regexp -indices -inline {((a?)*|(b?)*)+} aaa → {0 2} {3 2} {3 2} {-1 -1} on main and 8.4, but {0 2} {3 2} {-1 -1} {-1 -1} on 9.0/8.6, so the recent regexp fixes on main (not yet in 9.0) changed this capture result.
    • Cause Spencer engine preference rules (the outermost preference wins); the "+" capture cases in the later comments follow the same rules and are the PostgreSQL issue sebres linked.
    • Prototype none (documented behaviour; the related {1,1} bug [524b35] was fixed on main by b0b515295b, 2026-10-14)
    • Suggested ticket comment The first example behaves as re_syntax.n says ({m}? keeps the atom's preference, {m,m}? forces shortest). The capture results with "+" follow the outer-preference rule, as in PostgreSQL. Suggest closing as documented behaviour.
    ⬜3547994fff2012Modernize Windows channel driversNOT-A-BUG (RFE; the concrete problem was fixed in 2012)

    Open ticket 3547994fff · 25. Channel System · filed 2012-07-24

    Full verdict

    NOT-A-BUG (RFE; the concrete problem was fixed in 2012)

    • Platform Windows
    • Tested code reading only (main, 9.0, 8.6)
    • Cause The original issue (TIP #398 fast exit blocked by the pipe writer thread, io-29.33b) was fixed by 9a396c910e (ferrieux, 2012-07-25, "Abandon the synchronous Windows pipe driver to its fate"); the ticket was then renamed into a general "move everything to overlapped I/O" plan. Since then win/tclWinPipe.c calls CancelSynchronousIo() before falling back to TerminateThread() (9.0, main; 8.6 looks it up dynamically in tclWinInit.c).
    • Prototype none (redesign, not a bug fix)
    • Suggested ticket comment The original TIP #398 problem was fixed in 9a396c910e, and the pipe reader thread now uses CancelSynchronousIo(). What remains is a general redesign of the Windows channel drivers, which belongs in a TIP/RFE; proposing to close.
    ⬜3531271fff2012glob with unreadable directoryNOT-A-BUG (RFE: behaviour unchanged and intentional since 8.5 - main, 9.0, 8.6, 8.5 all raise "couldn't read directory "tmp/": permission denied" e…

    Open ticket 3531271fff · 37. File System · filed 2012-06-01

    Full verdict

    NOT-A-BUG (RFE: behaviour unchanged and intentional since 8.5 - main, 9.0, 8.6, 8.5 all raise "couldn't read directory "tmp/": permission denied" even with -nocomplain; 8.4.20 returned an empty list, silently dropping also the readable tmp2/a.c match)

    • Platform generic
    • Tested main 9.1b1, 9.0.5, 8.6.18, 8.5.19, 8.4.20 on Linux with the reporter's layout (tmp mode 000, tmp2/a.c)
    • Cause unix/tclUnixFile.c TclpMatchInDirectory (and win/tclWinFile.c) return TCL_ERROR when opendir fails; -nocomplain only covers "no matches". doc/glob.n does not mention the case.
    • Prototype none (a new option or changing -nocomplain to skip unreadable intermediate directories is a behaviour change, would need agreement/TIP)
    • Suggested ticket comment Still the behaviour in 9.1/9.0/8.6: an unreadable directory matched by an intermediate pattern component is an error, -nocomplain only suppresses the "no match" error. Skipping such directories would need a new option (TIP), so this is a feature request.
    ⬜3439599fff2011thread::cancel not reporting errorNOT-A-BUG (works as designed in Thread 2.8.13 and 3.0.7: the errorproc is called through an event queued to the thread that registered it)

    Open ticket 3439599fff · 80. Thread Package · filed 2011-11-17

    Full verdict

    NOT-A-BUG (works as designed in Thread 2.8.13 and 3.0.7: the errorproc is called through an event queued to the thread that registered it)

    • Platform generic (Thread package)
    • Tested dgp's demo on main + Thread 3.0.7 and 8.6.18 + Thread 2.8.13, 40 runs each. As written (no event loop after thread::join), "NO ERROR THROWN" 40/40. With one update after thread::join, the errorproc gets "eval canceled ... while executing "while 1 {}"" 20/20. Canceling a script sent with -async, and cancel -unwind of a thread in thread::wait, are reported as well.
    • Cause thread generic/threadCmd.c ThreadErrorProc: when the error handler belongs to another thread, the report is sent asynchronously to that thread, so the demo exits before the event is serviced
    • Prototype none (at most a doc note in thread::errorproc that the handler runs from the event loop of its owning thread)
    • Suggested ticket comment The cancellation error is reported: errorproc runs from the event loop of the thread that registered it, so the demo needs an update (or vwait) after thread::join to see it. Works with Thread 2.8.13 and 3.0.7. Proposed to close, possibly with a note added to the thread::errorproc docs.
    ⬜3385749fff2011File open failures auditedNOT-A-BUG (RFE; not doing a check before open is intentional)

    Open ticket 3385749fff · 38. Init - Library - Autoload · filed 2011-08-03

    Full verdict

    NOT-A-BUG (RFE; not doing a check before open is intentional)

    • Platform generic
    • Tested not run (request to add file exists/stat() before every open in init.tcl/auto.tcl and C code)
    • Cause by design: hobbs noted that ACL systems fool file readable, and dkf noted the check-then-open race. Failed opens of optional files (auto_path probing, tclIndex, pkgIndex.tcl) are expected. A stat() would itself be an audited access on most systems.
    • Prototype none
    • Suggested ticket comment Pre-checking with file exists/stat() is racy and, as noted, unreliable with ACLs, so Tcl deliberately just tries the open. Closing as Won't Fix; specific noisy probes (e.g. at startup) can be reported separately.
    ⬜3382532fff2011Synchronous geturl uses vwaitNOT-A-BUG (by design, now documented)

    Open ticket 3382532fff · 29. http Package · filed 2011-07-29

    Full verdict

    NOT-A-BUG (by design, now documented)

    • Platform generic
    • Tested code reading, library/http/http.tcl on main (geturl without -command calls http::wait, which does vwait ${token}(status))
    • Cause http::geturl is event-driven internally (coroutine Event, non-blocking connect, pipelining), as kjnash explained in 2022. doc/http.n now says "The event queue is even used without the -command option. As a side effect, arbitrary commands may be processed while http::geturl ..." The 2011 patch no longer applies.
    • Prototype none (would need a redesign of the http package)
    • Suggested ticket comment Synchronous http::geturl uses the event loop by design since the HTTP/1.1 rewrite, and http.n documents it. A fully synchronous mode would be a feature request; proposing to close.
    ⬜3304643fff2011User object info lost on more aggressive numeric conversionNOT-A-BUG (Tcl may re-type any value; an ObjType must be able to rebuild its intrep from the string, as dgp said). Partly improved anyway: the repo…

    Open ticket 3304643fff · 10. Objects · filed 2011-05-19

    Full verdict

    NOT-A-BUG (Tcl may re-type any value; an ObjType must be able to rebuild its intrep from the string, as dgp said). Partly improved anyway: the reporter's "INF" == "" case no longer shimmers.

    • Platform generic
    • Tested 8.6.18, 9.0.5, main with a list-typed value (work-a3/330464.tcl). $v == "" keeps the list intrep for 123 and INF. 0.5 == abc converts it to double, 123 == 456 to int (8.4 would also convert in the second case). 8.4/8.5 lack ::tcl::unsupported::representation, so they were not compared.
    • Cause generic/tclExecute.c INST_EQ/NEQ calls GetNumberFromObj on both operands before falling back to string comparison. An empty operand short-circuits this.
    • Prototype none (by design; changing it would need a TIP, and extensions should not rely on intreps surviving)
    • Suggested ticket comment Values may change their internal type at any time; an extension type needs a setFromAnyProc or a handle scheme. Comparing with "" no longer converts the other operand, and "eq" never does. I suggest closing as Works as designed.
    ⬜3184384fff2011The thread package can't be installed on SunOSNOT-A-BUG (local setup: reporter found LD_LIBRARY_PATH pointed to an old non-threaded libtcl; works after fixing it)

    Open ticket 3184384fff · 49. Threading · filed 2011-02-17

    Full verdict

    NOT-A-BUG (local setup: reporter found LD_LIBRARY_PATH pointed to an old non-threaded libtcl; works after fixing it)

    • Platform unix (SunOS 5.10/sparc, Thread 2.6.6, Tcl 8.4.19)
    • Tested not run (old Solaris/8.4 setup; attached config.log is clean as dkf noted). Tcl 9 is always threaded, so the "Tcl core wasn't compiled for threading" check cannot trigger with a correct libtcl.
    • Cause wrong libtcl picked up at runtime (reporter's comment of 2011-02-25).
    • Prototype none
    • Suggested ticket comment The reporter found the cause: LD_LIBRARY_PATH pointed to an old non-threaded Tcl library. Not a Thread bug; closing.
    ⬜3105247fff2010-NaN?NOT-A-BUG (intentional: kennykb, dkf, hobbs 2010; the ticket was auto-closed as Pending/Invalid on SourceForge, then reopened by the migration)

    Open ticket 3105247fff · 48. Number Handling · filed 2010-11-08

    Full verdict

    NOT-A-BUG (intentional: kennykb, dkf, hobbs 2010; the ticket was auto-closed as Pending/Invalid on SourceForge, then reopened by the migration)

    • Platform generic
    • Tested main build, 8.6.18: 0xfff8000000000000 formats as -NaN and 0xfff8000000000001 as -NaN(1), and both round-trip bit for bit through [binary format q]; IEEE 754 allows an optional sign on NaN.
    • Cause generic/tclStrToD.c / tclUtil.c Tcl_PrintDouble print the sign bit of NaN on purpose so that distinct bit patterns keep distinct string reps (follows David Gay's dtoa).
    • Prototype none (behaviour change rejected by the maintainers)
    • Suggested ticket comment The sign of NaN is printed so the bit pattern survives a round trip through its string rep (still so in 8.6 and 9.1); IEEE 754 allows a signed "nan". Resolution Invalid was already set in 2010; closing.
    ⬜3102850fff2010HTTP does not follow redirectsNOT-A-BUG (RFE)

    Open ticket 3102850fff · 29. http Package · filed 2010-11-04

    Full verdict

    NOT-A-BUG (RFE)

    • Platform generic
    • Tested code reading, library/http/http.tcl on main: no redirect handling (only the 30x reason phrases); the caller must check http::ncode and the Location meta.
    • Cause by design. kjnash (http maintainer) argues redirects belong in a wrapper (RFC 2616/9110 requires user confirmation for some redirects of POST), not in a single-request http::geturl. Duplicate [3612478] was already closed.
    • Prototype none (feature, would need a TIP-level option such as -followredirect)
    • Suggested ticket comment This is a feature request rather than a bug: http::geturl performs one request by design, and following redirects can be done by the caller (see the wiki recipe in the comments). Proposing to close it, or retype it as an RFE.
    ⬜3026795fff2010clock format %x returns wrong formatNOT-A-BUG (by design; RFE to build a clock locale from the Windows regional settings)

    Open ticket 3026795fff · 06. Time Measurement · filed 2010-07-08

    Full verdict

    NOT-A-BUG (by design; RFE to build a clock locale from the Windows regional settings)

    • Platform Windows (behaviour is generic)
    • Tested main build and 8.6.18 with LANG=en_GB.UTF-8: %x gives 07/08/2010 (the {} locale), -locale current gives 08/07/10 (en_gb.msg DATE_FORMAT %d/%m/%y)
    • Cause [clock] is locale-independent unless -locale is given, and -locale current uses only the msgcat catalog for the language and variant, not the Windows regional settings. This is the design described by kennykb in 2012.
    • Prototype none (feature request)
    • Suggested ticket comment Works as designed: without -locale, %x uses the C locale. Reading Windows regional settings into a clock locale would be a feature request, so suggest closing this or converting it to an RFE.
    ⏳3018050fff2010fcopy-from-starkit hangs on WindowsNOT-A-BUG (VFS channel driver without readable events; Pending/Works For Me, Serhiy could not reproduce on Windows and Linux with 8.5.8/8.6.18/9.0)

    Open ticket 3018050fff · 25. Channel System · filed 2010-06-18 · Pending (resolved, waiting for confirmation)

    ⏳ Since triage (updated 2026-11-11)

    Status now: Pending (Works For Me)

      Full verdict

      NOT-A-BUG (VFS channel driver without readable events; Pending/Works For Me, Serhiy could not reproduce on Windows and Linux with 8.5.8/8.6.18/9.0)

      • Platform Windows (actually generic: channel-driver dependent)
      • Tested Linux, work-a3/fc62.tcl (async fcopy from a file to a localhost socket): 8.6.18 real file OK; 9.0.5 from //zipfs: times out (zipfs channels post no events, 0654cecae1); main (own build) from //zipfs: OK. No starkit/tclkit 8.5 available; Serhiy already tested reflected channels (tclvfs style) on Windows and Linux.
      • Cause async fcopy needs readable events from the source channel; tclvfs/zip channel drivers that don't post them stall on every platform (cjmcdonald's comment). Not a Tcl core bug.
      • Existing work none for this ticket; related zipfs fix zipfs-channel-events (0654cecae1, 1ae447d307) merged into main (fadaf68f34, 2026-10-16), not in 9.0.
      • Prototype none (not a core bug)
      • Suggested ticket comment none new; Pending since Serhiy's 2026-10-12 comment, can be closed as Works For Me if nobody objects.
      ⬜2981283fff2010Tcl_CreateInterp() thread unsafe?NOT-A-BUG (the Thread commit that claimed it was revoked by its author)

      Open ticket 2981283fff · 45. Parsing and Eval · filed 2010-04-02

      Full verdict

      NOT-A-BUG (the Thread commit that claimed it was revoked by its author)

      • Platform generic
      • Tested history only. Thread check-in b4fd5b228d (2010-04-02 16:05, "serialize Tcl_CreateInterp() as it is thread-unsafe") was reverted 44 minutes later by 821ceb6d69 "Revoked last checin (wrong turn!)", same author. Thread trunk still calls Tcl_CreateInterp() unserialized in NewThread (threadCmd.c), and Thread's thread-creating tests have run that way since then. No concurrent stress test run (no Thread build here).
      • Cause n/a
      • Prototype none
      • Suggested ticket comment The Thread commit that serialized Tcl_CreateInterp() (b4fd5b228d) was revoked by its author the same day ("wrong turn"), and Thread has called it from many threads without a lock ever since. No known unsafety; suggest closing.
      ⬜2960042fff2010high water mark memory not shared for file I/O needsNOT-A-BUG (by design: Tcl_Obj memory is cached by the allocator, not returned to malloc; reported behaviour unchanged)

      Open ticket 2960042fff · 42. Memory Preservation · filed 2010-02-27

      Full verdict

      NOT-A-BUG (by design: Tcl_Obj memory is cached by the allocator, not returned to malloc; reported behaviour unchanged)

      • Platform generic
      • Tested main (9.1.1 build), 9.0.5, 8.6.18: RSS 7 MB → 389 MB after building and freeing a 5M-element [split] list inside a proc, → 580 MB after a further 200 MB string, i.e. the freed Tcl_Obj memory is not reused for ordinary allocations.
      • Cause tclThreadAlloc.c (threaded builds; tclObj.c block allocator in non-threaded ones) keeps freed Tcl_Objs on per-thread/shared free lists and never frees the blocks. Returning them would need a new allocator design (the per-page bookkeeping discussed in the comments).
      • Prototype none (allocator redesign, RFE)
      • Suggested ticket comment Still the same in 8.6–9.1: freed Tcl_Objs stay in the allocator's free lists and are not returned for other allocations. This is a design limitation of the object allocator, not a leak. Building with -DPURIFY (system malloc) avoids it. Suggest turning this into an RFE or closing it.
      ⬜2944777fff2010Tcl_WaitForEvent: CFRunLoop finished on MacOSXNOT-A-BUG (calling Tcl_DoOneEvent on an interp/notifier from another thread; das explained the threading model in 2010)

      Open ticket 2944777fff · None · filed 2010-02-02

      Full verdict

      NOT-A-BUG (calling Tcl_DoOneEvent on an interp/notifier from another thread; das explained the threading model in 2010)

      • Platform macOS
      • Tested none (macOS-only, application misuse)
      • Cause application called Tcl_DoOneEvent from worker threads; the panic in tclMacOSXNotify.c Tcl_WaitForEvent is a safety check for CFRunLoopStop by a third party.
      • Prototype none
      • Suggested ticket comment As explained by das, Tcl_DoOneEvent must only be called from the thread that owns the interp/notifier; the reporter's code used it from several threads. Closing as not a bug.
      ⬜2883036fff2009upvar-7.1 leaves irremovable variableNOT-A-BUG (documented behaviour; RFE for a way to remove a non-local link)

      Open ticket 2883036fff · 07. Variables · filed 2009-10-21

      Full verdict

      NOT-A-BUG (documented behaviour; RFE for a way to remove a non-local link)

      • Platform generic
      • Tested main, 9.0.5, 8.6.18, 8.5.19: after upvar 0 x uv; set uv 34; unset uv, info globals uv still lists uv and a second unset fails. Same on all versions.
      • Cause unset acts on the link target. upvar.n says "There is no way to unset an upvar variable except by exiting the procedure in which it is defined", and that text is still in doc/upvar.n. For global/namespace scope the link lives until the namespace is deleted.
      • Prototype none (a behaviour change needs a TIP, as nijtmans/ferrieux said in the comments)
      • Suggested ticket comment Works as upvar.n documents (unset acts on the target; the link stays until its frame or namespace goes away). Removing global/namespace links would need a TIP (e.g. &#91;unset -link&#93;). Suggest closing as Works As Documented, or moving it to an RFE.
      ⬜2878945fff2009Tcl_ExternalToUtf(...dst, dstLen,...) writes to dst[dstLen]NOT-A-BUG (dgp withdrew the claim himself on 2009-10-15; the docs now cover the buffer size)

      Open ticket 2878945fff · 10. Objects · filed 2009-10-14

      Full verdict

      NOT-A-BUG (dgp withdrew the claim himself on 2009-10-15; the docs now cover the buffer size)

      • Platform docs
      • Tested code reading: main/9.0 Tcl_ExternalToUtfEx (generic/tclEncoding.c) subtracts the terminator length from dstLen first and returns TCL_CONVERT_NOSPACE without writing when dstLen is too small; 8.6 Tcl_ExternalToUtf decrements dstLen and returns early if dstLen < 1.
      • Cause none. doc/Encoding.3 (9.x) now says dstLen is the size of the output buffer, describes NUL termination and TCL_ENCODING_NO_TERMINATE, and notes the TCL_UTF_MAX slack quirk under TCL_CONVERT_NOSPACE.
      • Prototype none
      • Suggested ticket comment As noted in the 2009-10-15 comment, dstLen already reserves room for the terminator, and since 9.0 dstLen==0 does not write either. Encoding.3 now documents dstLen as the buffer size, plus termination and the TCL_UTF_MAX slack. Can be closed.
      ⬜2837163fff2009explicit request for hidden filesNOT-A-BUG (documented in glob.n, same as sh; unchanged since 8.5)

      Open ticket 2837163fff · 37. File System · filed 2009-08-13

      Full verdict

      NOT-A-BUG (documented in glob.n, same as sh; unchanged since 8.5)

      • Platform unix (generic semantics)
      • Tested Linux main, 8.6.18: .hid, \.hid, {.}hid match a dotfile; [.]hid, [.h]hid, ?hid, *hid match it only with -types hidden. bash and Python's glob also do not match [.]hid.
      • Cause unix/tclUnixFile.c:TclpMatchInDirectory (and win/tclWinFile.c) treat a pattern as asking for hidden files only if it starts with . or \.; {.} works because braces are expanded before. POSIX 2.13.3 leaves it unspecified whether [.] in a bracket expression matches a leading period.
      • Prototype none (behaviour choice; changing it would differ from sh)
      • Suggested ticket comment This matches sh: bash also doesn't match .mozilla with [.]mozilla, and POSIX leaves it unspecified. glob.n says a leading . "must be matched explicitly or with a {} construct" unless -types hidden is given. Closing as works-as-designed.
      ⬜2818765fff2009tclConfig.sh references undefined var ${CC} for TCL_SHLIB_LDNOT-A-BUG (by design since 7c64b38c55, 2007). The behaviour is unchanged on main: TCL_SHLIB_LD='${CC} ${CFLAGS} ${LDFLAGS} -shared'.

      Open ticket 2818765fff · 53. Configuration and Build Tools · filed 2009-07-08

      Full verdict

      NOT-A-BUG (by design since 7c64b38c55, 2007). The behaviour is unchanged on main: TCL_SHLIB_LD='${CC} ${CFLAGS} ${LDFLAGS} -shared'.

      • Platform build
      • Tested main build's unix/tclConfig.sh; tclConfig.sh.in substitutes @SHLIB_LD@ unchanged (the same on 8.6/9.0).
      • Cause SHLIB_LD intentionally keeps ${CC}/${CFLAGS} unexpanded so that CC can be overridden at make time. Consumers have to provide CC (e.g. CC=$TCL_CC) or eval with their own compiler. TEA does not use TCL_SHLIB_LD.
      • Prototype none
      • Suggested ticket comment This is intentional since 7c64b38c55: TCL_SHLIB_LD is a template expanded with the consumer's own CC/CFLAGS/LDFLAGS; set CC=$TCL_CC before evaluating it. Closing as "Won't Fix".
      ⬜2796593fff2009clock broken when time_t tracks leapsecondsNOT-A-BUG (non-POSIX system time; ticket resolution is already Invalid)

      Open ticket 2796593fff · 06. Time Measurement · filed 2009-05-25

      Full verdict

      NOT-A-BUG (non-POSIX system time; ticket resolution is already Invalid)

      • Platform unix
      • Tested main build, 9.0.5, 8.6.18 with TZ=right/America/New_York: clock format 1243137600 gives 05/23/09 23:59:36 (24 s off), the reporter's symptom
      • Cause Tcl assumes POSIX time_t without leap seconds. ReadZoneinfoFile ignores the leap-second records of the "right/" zone files, so on a system set up for TAI-like time_t, local times are off by the leap-second count. kennykb explained this in 2009; supporting it would need time2posix/posix2time, which glibc does not provide.
      • Prototype none
      • Suggested ticket comment Tcl requires POSIX time_t (no leap seconds). "right/" zones are unsupported by design, and the resolution is already Invalid, so suggest closing.
      ⬜2789295fff2009Please add the file COPYING to tclX.X.X source distributionNOT-A-BUG (licence is in license.terms in every source distribution, as dkf said)

      Open ticket 2789295fff · 69. Other · filed 2009-05-09

      Full verdict

      NOT-A-BUG (licence is in license.terms in every source distribution, as dkf said)

      • Platform site (distribution)
      • Tested license.terms present at the top of main, 9.0, 8.6 trees
      • Cause n/a
      • Prototype none
      • Suggested ticket comment The licence is in license.terms at the top of every source distribution and referenced by every source file; there is no plan to rename it. Closing.
      ⬜2685152fff2009Fedora systemtap: can't compile with dtraceNOT-A-BUG (systemtap limitation, fixed in systemtap 0.9.5 per the reporter on 2009-05-18). CANNOT-TEST here (no systemtap/sys/sdt.h installed).

      Open ticket 2685152fff · 53. Configuration and Build Tools · filed 2009-03-12

      Full verdict

      NOT-A-BUG (systemtap limitation, fixed in systemtap 0.9.5 per the reporter on 2009-05-18). CANNOT-TEST here (no systemtap/sys/sdt.h installed).

      • Platform build (Linux systemtap SDT)
      • Tested none (no dtrace/stap tools). The macros in question (TclDecrRefCount → TCL_DTRACE_OBJ_FREE inside STORE_ARY, TCL_DTRACE_INST_NEXT) still exist, but the duplicate-label error came from old systemtap's label-based probe macros.
      • Cause systemtap < 0.9.5 generated a _stapprobe1_<name>_<__LINE__> label per probe, so two probes expanded on one source line collided.
      • Prototype none
      • Suggested ticket comment The reporter confirmed that systemtap 0.9.5 fixed the duplicate-label problem, so nothing to change in Tcl. Closing.
      ⬜2544673fff2009[interp create foo -safe] acceptedNOT-A-BUG (documented-by-usage compatibility behaviour; removal needs a TIP)

      Open ticket 2544673fff · 20. [interp] · filed 2009-01-29

      Full verdict

      NOT-A-BUG (documented-by-usage compatibility behaviour; removal needs a TIP)

      • Platform generic
      • Tested main, 8.6.18: interp create foo -safe creates a safe interp "foo"; interp create a b errors correctly
      • Cause tclInterp.c:Tcl_InterpObjCmd OPT_CREATE still has the "Weird historical rules" loop accepting -safe after the path. The test suite itself relies on it (interp.test lines 124, 127, 143, 202, 502-528; safe.test 150, 161), and it is surely used in the wild.
      • Prototype none (incompatible change, would need a TIP plus test edits)
      • Suggested ticket comment Still accepted on 9.1. The trailing <tt>-safe</tt> form is used throughout interp.test and safe.test and very likely in user code, so removing it is an incompatible change that would need a TIP. Suggest closing as Won't Fix, or turning it into a TIP proposal (deprecate first).
      ⬜2282547fff2008load of ill-behaved shared library from VFS: core dumpNOT-A-BUG (works as designed since TIP #357 80aa157f45: libraries are unloaded at finalization only with TCL_UNLOAD_DLLS; an ill-behaved library …

      Open ticket 2282547fff · 40. Dynamic Loading · filed 2008-11-14

      Full verdict

      NOT-A-BUG (works as designed since TIP #357 80aa157f45: libraries are unloaded at finalization only with TCL_UNLOAD_DLLS; an ill-behaved library needs --disable-dll-unloading; on Unix a VFS library's temporary copy is deleted right after loading, so nothing ties its lifetime to the VFS). No reproducer was ever given, the ticket is an RFE-like note from 2008

      • Platform generic
      • Tested code reading of generic/tclLoad.c (TclFinalizeLoad) and unix/tclLoadDl.c on main; no test run (no ill-behaved library to test with)
      • Existing work none
      • Prototype none (no bug)
      • Suggested ticket comment Serhiy's 2026-10-12 comment already explains it; propose closing as "Works as designed" (close only after a maintainer agrees).
      ⬜2170040fff2008integrate iocpsock to 8.6NOT-A-BUG (feature request; the work branch tip-162-branch was abandoned: last check-in acab51185e 2009-01-06, marked Closed by kbk 2011-02-28; I…

      Open ticket 2170040fff · 25. Channel System · filed 2008-10-16

      Full verdict

      NOT-A-BUG (feature request; the work branch tip-162-branch was abandoned: last check-in acab51185e 2009-01-06, marked Closed by kbk 2011-02-28; IPv6 sockets (TIP 162) went in by another route in 8.6)

      • Platform Windows
      • Tested repository history only
      • Cause n/a — win/tclWinSock.c still uses the WSAAsyncSelect-style notifier thread, not IOCP
      • Prototype none (RFE, large redesign; would need a TIP)
      • Suggested ticket comment The iocpsock integration branch tip-162-branch was closed in 2011 and never merged; IPv6 support came with TIP 162 in 8.6 separately. Suggest closing as a stale RFE; replacing the Windows socket driver with IOCP would need a new TIP.
      ⬜2094946fff2008glob behavior change in Tcl 8.4.14NOT-A-BUG (intentional since 8.4.14, consequence of the fix for Bug 943995; dgp said so in 2008)

      Open ticket 2094946fff · 36. Pathname Management · filed 2008-09-05

      Full verdict

      NOT-A-BUG (intentional since 8.4.14, consequence of the fix for Bug 943995; dgp said so in 2008)

      • Platform generic
      • Tested glob /tmp/ gives /tmp on 8.4.11, /tmp/ on 8.4.15, 8.6.18, 9.0.5, main (Linux)
      • Cause glob keeps the trailing slash of the pattern since the 943995 fix; that's the intended behaviour.
      • Prototype none (no change wanted; restoring the old result would break 943995)
      • Suggested ticket comment The trailing slash has been kept since 8.4.14 as a result of the fix for 943995, and nothing has changed since (8.6.18, 9.0, 9.1 are the same). Closing as Works As Expected.
      ⬜2037611fff2008NRE: streamline Tcl's evaluation stacks?NOT-A-BUG (refactoring/design idea, no defect described; msofer's 2010 comment says the coroutine blocker is gone)

      Open ticket 2037611fff · 45. Parsing and Eval · filed 2008-08-04

      Full verdict

      NOT-A-BUG (refactoring/design idea, no defect described; msofer's 2010 comment says the coroutine blocker is gone)

      • Platform generic
      • Tested n/a
      • Cause n/a
      • Prototype none (major redesign that would break Itcl/XOTcl/tbcload, as the ticket says)
      • Suggested ticket comment This is a design note rather than a bug. Suggest moving it to a Feature Request or closing it; the coroutine concern was resolved in 2010.
      ⬜2017162fff2008NRE and debugging toolsNOT-A-BUG (vague reminder/RFE; the TclX crash in the comments was TclX using 8.5 internals, msofer 2010)

      Open ticket 2017162fff · 45. Parsing and Eval · filed 2008-07-13

      Full verdict

      NOT-A-BUG (vague reminder/RFE; the TclX crash in the comments was TclX using 8.5 internals, msofer 2010)

      • Platform generic
      • Tested n/a (not tested with a TCL_COMPILE_DEBUG build)
      • Cause n/a
      • Prototype none
      • Suggested ticket comment There is no concrete defect here; the TclX profile crash came from TclX using changed internal structures. Suggest closing, or moving it to a feature request for NRE debugging aids.
      ⬜1972867fff2008invalid encoding cachingNOT-A-BUG (an already loaded encoding stays usable while anything, including a literal's Tcl_Encoding intrep, holds a reference; [encoding dirs] on…

      Open ticket 1972867fff · 10. Objects · filed 2008-05-26

      Full verdict

      NOT-A-BUG (an already loaded encoding stays usable while anything, including a literal's Tcl_Encoding intrep, holds a reference; [encoding dirs] only affects future file loads, as documented)

      • Platform generic
      • Tested main, 9.0, 8.6: after encoding convertto jis0208 ... in a proc, encoding dirs {} and the same literal still converts (0 8C); a fresh name gives "unknown encoding". encoding-2.2 and 11.1 were made robust in 2008 by dgp (llength shiftjis shimmer) and pass.
      • Cause generic/tclEncoding.c caches Tcl_Encoding in the obj intrep (GetEncodingFromObj); no epoch tied to the encoding search path. Only visible as differing lifetime of loaded encodings, which only the test suite depended on.
      • Prototype none (behaviour by design; adding an epoch would only change when a loaded encoding gets freed)
      • Suggested ticket comment Still the same in 8.6/9.0/9.1, but [encoding dirs] only governs loading of new *.enc files; an encoding that is still referenced (here via a literal's internal rep) stays valid. The tests were adjusted in 2008. Propose closing as Works as designed.
      ⬜1816166fff2007[file link -symbolic] fails on some relative targetsNOT-A-BUG (documented: on Unix a relative target is relative to the link's location, file.n "file link"). Changed between 8.4.15 (accepts cwd-relat…

      Open ticket 1816166fff · 37. File System · filed 2007-10-18

      Full verdict

      NOT-A-BUG (documented: on Unix a relative target is relative to the link's location, file.n "file link"). Changed between 8.4.15 (accepts cwd-relative targets) and 8.5.

      • Platform unix
      • Tested main, 9.0.5, 8.6.18, 8.5.19 refuse file link -symbolic td/l1 td/td1; 8.4.15 accepts it.
      • Cause unix/tclUnixFile.c:TclpObjLink checks the target relative to the link's directory, matching what the kernel will resolve.
      • Prototype none
      • Suggested ticket comment This is the documented behaviour since 8.5 (file.n: relative symlink targets "must be relative to the actual linkName's location (not to the cwd)"), and that's what the kernel resolves. Closing as Works As Expected.
      ⬜1777366fff2007changes to /etc/resolv.conf not handledNOT-A-BUG (on current systems: glibc >= 2.26 (2017) reloads /etc/resolv.conf itself when it changes; other libcs handle it in their resolver, which…

      Open ticket 1777366fff · 25. Channel System · filed 2007-08-20

      Full verdict

      NOT-A-BUG (on current systems: glibc >= 2.26 (2017) reloads /etc/resolv.conf itself when it changes; other libcs handle it in their resolver, which is the right layer)

      • Platform unix
      • Tested code reading only (no res_init() anywhere in unix/ on main; changing /etc/resolv.conf is out of scope here); glibc here is 2.43
      • Cause the old glibc stub resolver read resolv.conf once per process; the Debian patch called res_init() from Tcl's hostname lookup when the file's mtime changed.
      • Prototype none (libc's job; the Debian patch was non-portable, as noted in the comments)
      • Suggested ticket comment glibc 2.26 and later reload /etc/resolv.conf automatically when it changes, so Tcl doesn't need to call res_init() itself (it was also non-portable). Closing as out of date.
      ⬜1733117fff2007auto_path changes and avail package cacheing issuespatchNOT-A-BUG (documented semantics) / DOCS

      Open ticket 1733117fff · 39. Package Manager · filed 2007-06-07

      🧪 Prototype check (2026-10-24)

      1733117fff.patch: VERIFIED

      • Base main (tip d6dd6d0f71)
      • Before work-a4/v4docs2.tcl (pkg vfoo 1.0 in pk1 registered by a failed scan, then pk2 with 2.0 appended to auto_path): main and 8.6 -> known: 1.0; require -> 1.0; after package forget vfoo -> 2.0. Exactly the behaviour the new text describes.
      • Patch applied cleanly (10 changed lines)
      • After groff 0 warnings (0 before)
      • Test none possible (docs only)
      Full verdict

      NOT-A-BUG (documented semantics) / DOCS

      • Platform generic
      • Tested main and 8.6.18: dir1 with foo 1.0 scanned by an earlier failed require, then dir2 with foo 2.0 appended to auto_path: [package require foo] gives 1.0 ([package versions] knows only 1.0). As dgp said, [package require] promises only to satisfy the requirements; [package unknown] runs only when no known version does.
      • Cause by design (tclPkgUnknown is invoked only when the requirement cannot be satisfied).
      • Prototype prototypes/1733117fff.patch (10 changed lines, doc only): package.n says the unknown handler should register ifneeded scripts ("to find", not "to load the package"), and notes the auto_path case and the workarounds ([package forget] or tighter requirements), the doc improvement dgp left the ticket open for.
      • Suggested ticket comment Behaviour is as designed; proposed a package.n clarification: the package unknown handler only registers ifneeded scripts and runs only when no known version satisfies the requirement, so directories added to auto_path later are not searched.
      ⬜1669392fff2007TIP 280 differenceNOT-A-BUG (Tcl_EvalObjv runs a single pre-parsed command, not a script, so there is no script location; behaviour unchanged and codified by tests i…

      Open ticket 1669392fff · 45. Parsing and Eval · filed 2007-02-26

      Full verdict

      NOT-A-BUG (Tcl_EvalObjv runs a single pre-parsed command, not a script, so there is no script location; behaviour unchanged and codified by tests info-41.0/41.1)

      • Platform generic
      • Tested C harness on main, 9.0, 8.6: Tcl_Eval gives type eval line 1 cmd {info frame 0} level 0; Tcl_EvalObjv (flags 0 or TCL_EVAL_GLOBAL) gives type precompiled.
      • Cause generic/tclCmdIL.c TclInfoFrame falls back to the interp's root CmdFrame (type precompiled) when the command was not invoked from a script.
      • Prototype none (optional doc tweak: info.n describes precompiled only as tbcload code; it is also the result for commands invoked directly via Tcl_EvalObjv/invokehidden with no script)
      • Suggested ticket comment Still the same in 8.6/9.0/9.1 and expected (see jenglish): Tcl_EvalObjv has no script, so [info frame 0] reports "type precompiled". At most info.n should say that precompiled also covers commands invoked directly from C. Propose closing.
      ⬜1665441fff2007Generate an error on default arg followed by non-default argNOT-A-BUG (RFE; behaviour now documented in proc.n since c992833adf, 2008-01-17: such defaulted arguments "become de-facto required arguments ... p…

      Open ticket 1665441fff · 22. [proc] and [uplevel] · filed 2007-02-21

      Full verdict

      NOT-A-BUG (RFE; behaviour now documented in proc.n since c992833adf, 2008-01-17: such defaulted arguments "become de-facto required arguments ... portable code should ensure that all optional arguments come after all required arguments")

      • Platform generic
      • Tested main, 9.0, 8.6: proc prov {{opt 23} man} ... is accepted; prov 1 → wrong # args: should be "prov ?opt? man"; prov 1 2 works; same for [apply].
      • Cause by design in generic/tclProc.c; raising an error now would break existing code.
      • Prototype none (an error would be an incompatible change needing a TIP). Small possible cleanup: the wrong-#-args message still shows ?opt? although opt is effectively required (ProcWrongNumArgs).
      • Suggested ticket comment proc.n documents this case since 2008 (defaulted args before required ones become required). Making it an error would be incompatible and would need a TIP. Propose closing; only the "?opt?" in the wrong # args message is still misleading.
      ⬜1617990fff2006Incorrect result of access() for NFS mounted filesNOT-A-BUG (NFS client access() behaviour; resolution already Wont Fix), CANNOT-TEST (NFS)

      Open ticket 1617990fff · 37. File System · filed 2006-12-18

      Full verdict

      NOT-A-BUG (NFS client access() behaviour; resolution already Wont Fix), CANNOT-TEST (NFS)

      • Platform unix
      • Tested none (no NFS); code reading: TclpObjAccess still uses access(), which is correct for ACLs (dkf's point)
      • Cause NFS access() semantics (server-side checks, root squash, etc.), not Tcl.
      • Prototype none
      • Suggested ticket comment Resolution is already Wont Fix: access() is the right call (it honours ACLs), and NFS-specific test failures belong under the notNetworkFilesystem constraint. Closing.
      ⬜1555698fff2006socket name corruptionNOT-A-BUG (in the tcllib dns package: it closes the socket by a stale name)

      Open ticket 1555698fff · 24. Channel Commands · filed 2006-09-10

      Full verdict

      NOT-A-BUG (in the tcllib dns package: it closes the socket by a stale name)

      • Platform generic
      • Tested tcllib trunk modules/dns/dns.tcl: ::dns::Finish does close $state(sock) and then calls the -command callback, but leaves state(sock) set; ::dns::cleanup closes $state(sock) again. A socket opened inside the callback can get the freed name (names come from the fd on 8.4/8.6, from the state pointer on 9.x), so cleanup closes the user's new socket and its writable handler never fires. Tcl-only demo (work-a5/sn.tcl): close a socket, open another (same name on main), [close] the old name: closes the new one. There are never two open channels with the same name.
      • Cause tcllib dns::Finish/dns::cleanup double close by name, not Tcl.
      • Prototype none (tcllib: unset state(sock) after closing it in Finish, or close only once in cleanup)
      • Suggested ticket comment Not a Tcl bug: the dns package closes its socket before calling -command and dns::cleanup closes it again by name; a socket opened in the callback can reuse that name, so cleanup closes it. Channel names of closed channels may be reused. Should go to tcllib (dns::Finish should forget state(sock) after closing it). Closing here.
      ⬜1539898fff2006[file copy] doesn't handle relative symlinksNOT-A-BUG (documented: file.n says "file copy will copy soft links (i.e. the links themselves are copied, not the things they point to)"; same as c…

      Open ticket 1539898fff · 37. File System · filed 2006-08-14

      Full verdict

      NOT-A-BUG (documented: file.n says "file copy will copy soft links (i.e. the links themselves are copied, not the things they point to)"; same as cp -a / cp -R)

      • Platform unix
      • Tested main, 8.6.18 (directory with ../fred link copies the link text verbatim)
      • Cause TclpCopyFile/DoCopyFile copy the link with readlink/symlink, by design.
      • Prototype none (documented behaviour)
      • Suggested ticket comment This is the documented behaviour: file.n states that &#91;file copy&#93; copies symbolic links themselves, not their targets, as cp -R does, so a relative link keeps its relative text. Closing as works-as-documented.
      ⬜1510060fff2006[string is double -strict] versus [expr int()]NOT-A-BUG (NaN is a valid double; int() of NaN has no integer value)

      Open ticket 1510060fff · 18. Commands M-Z · filed 2006-06-21

      Full verdict

      NOT-A-BUG (NaN is a valid double; int() of NaN has no integer value)

      • Platform generic
      • Tested main, 9.0, 8.6, 8.5: string is double -strict NaN = 1, expr {int(NaN)} gives "floating point value is Not a Number" (the same for Inf). Behaviour unchanged and consistent with the docs.
      • Cause by design; callers that need int() must also exclude NaN/Inf (e.g. $x == $x, or [string is entier] / [string is integer] in 9.x).
      • Prototype none
      • Suggested ticket comment Works as designed: NaN (and Inf) are doubles but have no integer value, so int() must fail. Code that needs int() should check $x == $x (and finiteness) as well. Suggest closing.
      ⬜1469327fff2006Tcl_DoWhenIdle can't rescheduleNOT-A-BUG (documented limitation; the recommended zero-timeout timer works for what the BUGS section is about)

      Open ticket 1469327fff · 02. Event Loops · filed 2006-04-12

      Full verdict

      NOT-A-BUG (documented limitation; the recommended zero-timeout timer works for what the BUGS section is about)

      • Platform generic
      • Tested 8.6.18, 9.0.5, main: an idle callback that reschedules itself makes update idletasks and update loop forever (as documented in DoWhenIdle.3 BUGS); an after 0 callback that reschedules itself does not block update idletasks (returns at once, callback not run) but does make a plain update loop forever, on all versions (8.4.20 same).
      • Cause Tcl_ServiceIdle/update idletasks loop until no idle handlers remain; zero timers are not idle handlers, so they don't block Tk's "wait for idle" paths. Tk bug 1348333 (text widget creating too many timers) is a Tk-side problem and is still open in the Tk tracker.
      • Prototype none
      • Suggested ticket comment The workaround in Tcl_DoWhenIdle(3) does hold: a self-rescheduling zero-timeout timer doesn't block update idletasks (only a full update, which by definition processes all pending events). The Tk text widget problem is Tk ticket 1348333. Suggest closing this one.
      ⬜1432269fff2006[file copy] fails to recursively copy directory if it existsNOT-A-BUG (documented: "Trying to overwrite a non-empty directory ... will all result in errors even if -force was specified"; resolution is alread…

      Open ticket 1432269fff · 37. File System · filed 2006-02-15

      Full verdict

      NOT-A-BUG (documented: "Trying to overwrite a non-empty directory ... will all result in errors even if -force was specified"; resolution is already Rejected, reporter said in 2008 "if not just close it")

      • Platform generic
      • Tested main (third copy of foo into bar fails with "file exists", with and without -force)
      • Cause by design, [file copy] does not merge directory trees.
      • Prototype none (documented behaviour)
      • Suggested ticket comment Works as documented: file.n says overwriting a non-empty directory is an error even with -force. Merging trees would be a new feature. Closing (resolution is already Rejected).
      ⬜1367506fff2005Inline scan returns too many itemspatchNOT-A-BUG (intentional, per hobbs 2005) / DOCS: scan 0.99 %d.%d.%d.%d returns 0 99 {} {} on main and 8.6.18 (one element per conversion, empty …

      Open ticket 1367506fff · 23. [scan] · filed 2005-11-27

      🧪 Prototype check (2026-10-24)

      1367506fff.patch: VERIFIED

      • Base main (tip d6dd6d0f71)
      • Before scan 0.99 %d.%d.%d -> 0 99 {}, scan {} %d.%d -> `` (work-a4/v4docs1.tcl); scan.n describes only the empty-input case.
      • Patch applied cleanly (6 changed lines); renders "... For example, scan 0.99 %d.%d.%d returns “0 99 {}”." (example matches real output)
      • After groff 0 warnings (0 before)
      • Test none possible (docs only)
      Full verdict

      NOT-A-BUG (intentional, per hobbs 2005) / DOCS: scan 0.99 %d.%d.%d.%d returns 0 99 {} {} on main and 8.6.18 (one element per conversion, empty for unperformed ones); scan "" %d.%d returns ""; scan.n only describes the second case.

      • Platform docs
      • Tested main build, 8.6.18
      • Cause generic/tclScan.c fills unconverted inline positions with empty strings by design; doc/scan.n does not say so.
      • Prototype prototypes/1367506fff.patch (6 changed lines, docs only): sentence plus example in scan.n.
      • Suggested ticket comment Inline [scan] intentionally returns one element per conversion, with empty strings for the conversions not reached; changing that would break scripts. Proposed doc clarification in scan.n.
      ⬜1329754fff2005sockets lose data on WindowsNOT-A-BUG (TIP #398: exit does not wait for nonblocking output; the reporter's 10 x 10 KB scenario is NOT-REPRODUCED)

      Open ticket 1329754fff · 27. Channel Types · filed 2005-10-18

      Full verdict

      NOT-A-BUG (TIP #398: exit does not wait for nonblocking output; the reporter's 10 x 10 KB scenario is NOT-REPRODUCED)

      • Platform Windows
      • Tested Linux 8.6.18, 9.0.5, main: the reporter's scenario (non-blocking client writes 10 x 10 KB, then flush+close+exit, or just exit) delivers all data in 10/10 runs.
      • Cause unknown on Windows; suspects are win/tclWinSock.c close of a non-blocking socket with pending output (closesocket without lingering/flush of the background-flush queue) and exit-time finalization not completing background flushes.
      • Prototype none
      • Repro win/1329754fff.tcl (self-contained: server + exec'd client, variants "close" and "exit", 5 runs each, prints VERDICT). Run with 9.0.5, 8.6.18 and main on Windows 11.
      • Suggested ticket comment (after the Windows run) Not reproducible on Linux; on Windows 11 with 8.6.18/9.0.5/main the reporter's scenario ...
      Windows VM, run 3
      • Platform Windows (generic behaviour, visible on Windows because of smaller socket buffers)
      • Verdict NOT-A-BUG (TIP #398: exit does not wait for nonblocking output; the reporter's 10 x 10 KB scenario is NOT-REPRODUCED)
      • Tested VM win/1329754fff.tcl (10 x 10 KB, close and exit variants, 5 runs each): all data arrives on 9.1b1, 9.0.5, 8.6.18. Scaled to 200 x 10 KB (work-win3/1329v.tcl, 10 runs per variant): nonblocking flush; close; exit loses the tail (received 183–191 of 200) in 1–3 of 10 runs on 9.1b1, 9.0.5, 8.6.18; plain exit loses 1 message in 0–2 of 10; fconfigure -blocking 1 before flush, or keeping the event loop alive 1 s after close: 0/10 lost; with TCL_FLUSH_NONBLOCKING_ON_EXIT=1: 0/20 lost on 9.1b1 and 8.6.18. Linux 9.1/8.6 with 200 x 10 KB: 0/20 lost (bigger localhost buffers).
      • Cause generic/tclIO.c TclFinalizeIOSubsystem: since TIP #398 (8.6) exit no longer switches nonblocking channels to blocking, so output still queued in Tcl's background flush is dropped; documented in chan.n (close), with TCL_FLUSH_NONBLOCKING_ON_EXIT as the opt-out.
      • Prototype none
      • Suggested ticket comment The original 10 x 10 KB demo delivers all data on Windows 11 with 8.6.18, 9.0.5 and 9.1. With more data a nonblocking &#91;close&#93; followed at once by &#91;exit&#93; can still drop the queued tail, but since TIP #398 that is documented behaviour (chan.n): switch the socket to blocking before the final flush, keep the event loop running, or set TCL_FLUSH_NONBLOCKING_ON_EXIT=1 (all tested, no loss). Propose closing.
      ⬜1107673fff2005tclIO.c (CVS): Tcl_NextHashEntry should be usedNOT-A-BUG (deliberate idiom: each iteration closes/deletes the entry, so re-fetching the first entry is the safe way to drain the table; andreas_ku…

      Open ticket 1107673fff · 25. Channel System · filed 2005-01-23

      Full verdict

      NOT-A-BUG (deliberate idiom: each iteration closes/deletes the entry, so re-fetching the first entry is the safe way to drain the table; andreas_kupries said so in 2005)

      • Platform generic
      • Tested code reading, generic/tclIO.c:DeleteChannelTable (main), unchanged.
      • Cause n/a. The loop body can delete entries other than the current one (closing a channel can run scripts and close other channels), so re-fetching with Tcl_FirstHashEntry is the safe choice.
      • Prototype none
      • Suggested ticket comment The loop deletes the current entry each time and may delete others, so re-fetching the first entry is intentional and safe. Closing as not a bug.
      ⬜1084795fff2004Release notes hard to findNOT-A-BUG (website issue from 2004; www.tcl-lang.org now has release pages and changes files; not a Tcl source bug)

      Open ticket 1084795fff · 69. Other · filed 2004-12-14

      Full verdict

      NOT-A-BUG (website issue from 2004; www.tcl-lang.org now has release pages and changes files; not a Tcl source bug)

      • Platform site
      • Tested none (no network); the source tree ships changes.md / README.md release notes
      • Cause n/a
      • Prototype none
      • Suggested ticket comment The website has been reorganised since 2004 (each release has its own page with release notes, and the sources ship the changes file). Closing as out of date; website issues go to the tcl-website tracker.
      ⬜1034983fff2004man2help: .AS macro not supportedNOT-A-BUG (WinHelp tooling removed in 8.7/9.0 by fadc186aac, 2020-11-25; in 8.6 tools/man2help*.tcl remain but no makefile uses them)

      Open ticket 1034983fff · 53. Configuration and Build Tools · filed 2004-09-26

      Full verdict

      NOT-A-BUG (WinHelp tooling removed in 8.7/9.0 by fadc186aac, 2020-11-25; in 8.6 tools/man2help*.tcl remain but no makefile uses them)

      • Platform docs (Windows WinHelp build)
      • Tested code reading: tools/man2help2.tcl in 8.6 still has an empty AS {} handler; 9.0/main have no man2help*.tcl, win/makefile.vc builds .chm via tcltk-man2html.tcl (htmlhelp target).
      • Cause tools/man2help2.tcl ignored .AS; WinHelp (.hlp) itself is unsupported since Windows Vista.
      • Prototype none (obsolete tool)
      • Suggested ticket comment WinHelp generation was removed in 8.7/9.0 (fadc186aac); the HTML/CHM docs are made by tcltk-man2html.tcl, which handles .AS. Closing as out of date.
      ⬜992108ffff2004TclFinalizeLock on Unix not correct.NOT-A-BUG (deliberate design, resolution already "Wont Fix"; ticket just never closed)

      Open ticket 992108ffff · 49. Threading · filed 2004-07-16

      Full verdict

      NOT-A-BUG (deliberate design, resolution already "Wont Fix"; ticket just never closed)

      • Platform unix
      • Tested code reading on main (unix/tclUnixThrd.c TclFinalizeLock still only unlocks initLock; 9.0/8.6 the same). No leak on Linux/glibc (static mutexes allocate nothing). FreeBSD libthr allocates on first use of a PTHREAD_MUTEX_INITIALIZER mutex, so 3 small allocations stay alive after Tcl_Finalize there — only visible when libtcl is unloaded and reloaded (CANNOT-TEST here; would need the FreeBSD VM + valgrind).
      • Cause destroying globalLock/initLock in Tcl_Finalize would break re-initialization and races with concurrent Tcl_Finalize (vasiljevic's analysis). Note: the comment is now partly stale: allocLock is a PMutex initialized with pthread_once + pthread_mutex_init, not PTHREAD_MUTEX_INITIALIZER, and the once-guard cannot be reset, so it must not be destroyed either.
      • Prototype none
      • Suggested ticket comment As discussed, the statically initialized locks are intentionally kept across Tcl_Finalize so Tcl can be re-initialized; the resolution is already Wont Fix. Closing.
      ⬜971976ffff2004Inconsistent naming in [file exists]NOT-A-BUG (documented in filename.n: "Any number of trailing slash characters ... are simply ignored ... foo/ does not necessarily mean a directory…

      Open ticket 971976ffff · 37. File System · filed 2004-06-13

      Full verdict

      NOT-A-BUG (documented in filename.n: "Any number of trailing slash characters ... are simply ignored ... foo/ does not necessarily mean a directory"); the "odder" case from the comments is fixed

      • Platform generic
      • Tested main, 8.6.18: [file exists plainfile/] → 1 (shell test -e says no), [file exists plainfile/.] → 0 (was 1 in 8.4.6), [file exists dir/] → 1
      • Cause path normalization strips trailing separators before access(); a change would need a TIP (Vince, 2004).
      • Prototype none (behaviour change; needs a TIP)
      • Suggested ticket comment The trailing-slash behaviour is documented in filename.n since 2004; &#91;file exists file/.&#93; now correctly returns 0. Making "file/" mean "must be a directory" would need a TIP. Closing as documented.
      ⬜941872ffff2004path -->FS function limitationsNOT-A-BUG (design limitation / RFE: one owning filesystem per path, linear registration-order lookup)

      Open ticket 941872ffff · 37. File System · filed 2004-04-25

      Full verdict

      NOT-A-BUG (design limitation / RFE: one owning filesystem per path, linear registration-order lookup)

      • Platform generic (VFS design)
      • Tested code reading (generic/tclIOUtil.c Tcl_FSGetFileSystemForPath still walks the list in registration order)
      • Cause Tcl_Filesystem design. zipfs (9.0) mounts archives with the same model, so stacked or nested filesystems are still not possible.
      • Prototype none (would need a TIP-level VFS redesign)
      • Suggested ticket comment This is a VFS design limitation, unchanged in 9.x. Propose to close as an RFE / TIP material, or move it to the feature-request tracker.
      ⬜753306ffff2003Tcl_TraceCommand docs not up to date (TIP 62)NOT-A-BUG (obsolete: TCL_TRACE_ENTER_EXEC/LEAVE_EXEC were public in 8.4's tcl.h, but since 8.5 they are internal, defined only in generic/tclInt.h;…

      Open ticket 753306ffff · 46. Traces · filed 2003-06-12

      Full verdict

      NOT-A-BUG (obsolete: TCL_TRACE_ENTER_EXEC/LEAVE_EXEC were public in 8.4's tcl.h, but since 8.5 they are internal, defined only in generic/tclInt.h; the C API for execution traces is Tcl_CreateObjTrace, documented in CrtTrace.3)

      • Platform docs
      • Tested grep of tcl.h/tclInt.h in 8.4.2, 8.4 tip, 8.5, 8.6, 9.0, main; doc/TraceCmd.3 documents exactly the public flags TCL_TRACE_RENAME/TCL_TRACE_DELETE
      • Cause n/a
      • Prototype none (nothing to document in the public API)
      • Suggested ticket comment Since 8.5 the TIP 62 exec flags are no longer in tcl.h (only tclInt.h), so Tcl_TraceCommand's doc listing RENAME/DELETE is complete. Execution tracing from C goes through Tcl_CreateObjTrace. Proposing to close as out of date.
      ⬜748952ffff2003get.test failedNOT-A-BUG (obsolete: since 8.5 number parsing is Tcl's own (TclParseNumber/bignums), not libc strtol; get-1.10 now tests `testgetint -1844674407370…

      Open ticket 748952ffff · 52. Portability Support · filed 2003-06-04

      Full verdict

      NOT-A-BUG (obsolete: since 8.5 number parsing is Tcl's own (TclParseNumber/bignums), not libc strtol; get-1.10 now tests testgetint -18446744073709551614 and expects the "integer value too large" error)

      • Platform unix (Tru64)
      • Tested code reading of tests/get.test in main and 8.6; Tru64 not available.
      • Cause Tru64 libc strtol overflow handling in 8.4.x.
      • Prototype none
      • Suggested ticket comment Tcl 8.5+ no longer uses strtol for integer parsing and get-1.10 was rewritten, so this 8.4/Tru64 failure no longer applies. Closing.
      ✅653218ffff2002Channels CloseProc() not called on closepatchNOT-A-BUG as reported (by design, close(n): a non-blocking channel with pending output is closed in the background once the flush completes; serial…

      Open ticket 653218ffff · 25. Channel System · filed 2002-12-13

      ✅ Since triage (updated 2026-11-11)

      Status now: Closed (Invalid)

      • 💬 2026-10-16 serhiy.storchaka: The deferred close is by design (see close(n)). Testing this found that on Linux in 8.7 and later a background flush to a pipe whose reader has exited never ends: 7aa8fd34c4.
      Full verdict

      NOT-A-BUG as reported (by design, close(n): a non-blocking channel with pending output is closed in the background once the flush completes; serial "discard on close" was added by TIP 160 -closemode, 9.0, but that applies only inside the driver close, not to Tcl-level buffers). But it uncovered a REPRODUCED regression (8.7b1, 9.0, main; not 8.6) in the Linux epoll notifier: if the reader goes away during the background flush, the flush never finishes, the closeProc is never called and the process spins at 100% CPU until exit.

      • Platform generic (+ unix regression on 8.7/9.x)
      • Tested open "|sleep 1" w, non-blocking, 300 KB written, close: on 8.6 the fds are released once sleep exits (EPIPE). On main/9.0 they stay open for 17+ s, and strace shows epoll_wait returning EPOLLERR with timeout 0 in a loop. Simpler: chan event $p writable on a pipe whose child exited without reading fires after 7 ms on 8.6; on 8.7b1/9.0/main it never fires (3 s timeout, ~2 s CPU burnt), and close then reports no error, so the data loss goes unreported.
      • Cause unix/tclEpollNotfy.c:PlatformEventsTranslate maps EPOLLERR only to TCL_EXCEPTION; the pipe's write handler waits for TCL_WRITABLE, so it never runs and the level-triggered EPOLLERR wakes epoll_wait forever. select() reports such an fd as writable. The code dates from the epoll notifier's introduction (858569fb04, 2017-05-18).
      • Prototype prototypes/653218ffff.patch (26 changed lines: EPOLLERR also sets TCL_READABLE|TCL_WRITABLE; test io-78.1). Test fails without the fix (timeout 0 {}) and passes with it (writable 1 {broken pipe}). io, event, socket, chanio, ioCmd, exec, unixNotfy and zlib tests pass. The kqueue notifier (BSD/macOS) looks fine (EV_ERROR is folded into the filter's own mask), not tested.
      • Suggested ticket comment The deferred close itself is by design (close(n)), and TIP 160 -closemode discard covers the serial OS buffers. But on 8.7/9.x Linux, a background flush to a pipe whose reader died never ends: the epoll notifier maps EPOLLERR only to TCL_EXCEPTION, so the writable handler never runs and the notifier spins. Proposed fix: map EPOLLERR to readable+writable too (test io-78.1). That one probably deserves its own ticket.
      ⬜579633ffff2002why are there static paths in libtcl?NOT-A-BUG (design question, answered in the 2002 comments; superseded by zipfs)

      Open ticket 579633ffff · 38. Init - Library - Autoload · filed 2002-07-10

      Full verdict

      NOT-A-BUG (design question, answered in the 2002 comments; superseded by zipfs)

      • Platform build
      • Tested main build: info library is //zipfs:/lib/tcl/tcl_library (script library embedded in libtcl, TIP 430, i.e. the "internal VFS" idea from the comments); unix/tclUnixInit.c still compiles in TCL_LIBRARY (defaultLibraryDir) and TCL_PACKAGE_PATH as fallbacks/tcl_pkgPath.
      • Cause libtcl cannot find its own install location on Unix, so configure-time paths are the last-resort search location. The tkFont.c part belongs to Tk.
      • Prototype none (question, not a defect)
      • Suggested ticket comment Since Tcl 9.0 the script library and encodings are embedded in libtcl via zipfs, so the compiled-in TCL_LIBRARY is only a fallback; TCL_PACKAGE_PATH remains as the default tcl_pkgPath. Suggest closing.
      ⬜579310ffff2002io.test child processesNOT-A-BUG (test-suite refactoring suggestion, not a bug)

      Open ticket 579310ffff · 25. Channel System · filed 2002-07-09

      Full verdict

      NOT-A-BUG (test-suite refactoring suggestion, not a bug)

      • Platform generic
      • Tested code reading; tests/io.test on main still opens 77 open "|..." pipelines to child tclsh processes
      • Cause n/a. io.test uses child processes to feed data slowly or in pieces into Tcl_GetsObj etc. Rewriting them to use in-process channels (reflected channels / testchannel) is possible but is a test-suite rewrite with no user-visible bug behind it.
      • Prototype none (RFE-like, large rewrite)
      • Suggested ticket comment This is a test-suite improvement idea, not a bug. io.test still uses child processes; rewriting them on top of reflected channels would be welcome, but this could be closed or moved to RFE.
      ⬜533301ffff2002fconfigure -peername very slowNOT-A-BUG (reverse DNS lookup by design; synchronous getnameinfo() has no timeout)

      Open ticket 533301ffff · 24. Channel Commands · filed 2002-03-21

      Full verdict

      NOT-A-BUG (reverse DNS lookup by design; synchronous getnameinfo() has no timeout)

      • Platform generic
      • Tested code reading on main (no outside network for a slow-DNS repro)
      • Cause unix/tclUnixSock.c:TcpHostPortList (and win/tclWinSock.c) still call getnameinfo() without NI_NUMERICHOST for -peername/-sockname, so a slow resolver blocks [fconfigure]. The workaround dkf mentioned is still there and works on Unix and Windows: set ::tcl::unsupported::noReverseDNS (undocumented). A real fix (numeric-only option, async lookup) needs a TIP.
      • Prototype none (API/behaviour design)
      • Suggested ticket comment Still by design in 9.x: -peername/-sockname do a synchronous reverse lookup. Setting ::tcl::unsupported::noReverseDNS turns it off (Unix and Windows). Anything better (an option that returns only the address) needs a TIP, so this could be moved to RFE.
      ⬜439114ffff2001Channel buffering affects [close] statusNOT-A-BUG (documented: no exit status from closing a non-blocking pipe)

      Open ticket 439114ffff · 25. Channel System · filed 2001-07-06

      Full verdict

      NOT-A-BUG (documented: no exit status from closing a non-blocking pipe)

      • Platform generic
      • Tested main, 9.0.5, 8.6.18, 8.4.20 on Linux, with open "|sh -c {exit 33}" read via fileevent until eof. Blocking: close raises CHILDSTATUS pid 33. -blocking 0: close returns 0 with no error. The trigger is -blocking, not buffering.
      • Cause unix/tclUnixPipe.c:PipeClose2Proc (and the Windows equivalent) detaches and reaps the children without collecting their status when isNonBlocking, so close never blocks. doc/open.n says: "unless the pipeline is in non-blocking mode then no exit status is returned (a silent close with -blocking 0)". Getting the status of children that have already exited (as the reporter suggested) would be a behaviour change.
      • Prototype none (documented behaviour; collecting status of already-exited children would be a design change)
      • Suggested ticket comment Works as documented in open.n: closing a pipeline in non-blocking mode returns no exit status. Use fconfigure $fid -blocking 1 before &#91;close&#93; to get the CHILDSTATUS error. Can be closed (or moved to RFE for "report status if the children have already exited").
      ⬜408568ffff2001Variable substitution parsing limited to ASCII alphanumericsNOT-A-BUG (documented: Tcl.n says "Letters and digits are only the standard ASCII ones" for $name and $name(index), in 8.6, 9.0 and main)

      Open ticket 408568ffff · 45. Parsing and Eval · filed 2001-03-14

      Full verdict

      NOT-A-BUG (documented: Tcl.n says "Letters and digits are only the standard ASCII ones" for $name and $name(index), in 8.6, 9.0 and main)

      • Platform generic
      • Tested "$пи", "$é" stay literal on main build, 9.0.5, 8.6.18, 8.5.19; ${пи} works; expr {пи()} gives "invalid character" everywhere.
      • Cause generic/tclParse.c ParseVarName / TclIsBareword accept only ASCII alnum, _ and ::.
      • Prototype none (language change, needs a TIP). sebres' branch bug-408568 (745068c247..32568cdfa1, 2019-03-08, based on the 8.5 branch sebres-8-5-comp-8-6-fix) was never merged.
      • Suggested ticket comment The ASCII-only rule is now explicitly documented in Tcl.n (rule 8). Branch [/timeline?r=bug-408568|bug-408568] (8.5-based) was never merged; changing it needs a TIP. Suggest converting to RFE or closing.
      ⬜404906ffff2001file exists com1 behaves differentlyNOT-A-BUG (obsolete: the difference was between Windows 95/98 and NT/2000; Tcl has not supported 9x for a long time)

      Open ticket 404906ffff · 37. File System · filed 2001-02-28

      Full verdict

      NOT-A-BUG (obsolete: the difference was between Windows 95/98 and NT/2000; Tcl has not supported 9x for a long time)

      • Platform Windows (95/98 vs NT)
      • Tested none (no Win9x)
      • Cause Win9x API differences for reserved device names
      • Prototype none
      • Suggested ticket comment Windows 95/98 is no longer supported, and on NT-based Windows &#91;file exists com1&#93; is consistent. Propose to close as Out of Date.
      ⬜219289ffff2000Tcl doesn't read utf-16 files properlyNOT-A-BUG (BOM detection is an application concern; explicit-order utf-16le/utf-16be encodings exist since 8.7/9.0, covering hobbs' request)

      Open ticket 219289ffff · 10. Objects · filed 2000-10-26

      Full verdict

      NOT-A-BUG (BOM detection is an application concern; explicit-order utf-16le/utf-16be encodings exist since 8.7/9.0, covering hobbs' request)

      • Platform generic
      • Tested big-endian file with BOM read on main build and 9.0.5 with -encoding unicode/utf-16/utf-16le (first char U+FFFE, wrong order) and utf-16be (correct, BOM kept as U+FEFF). 8.6.18 has no utf-16be/le.
      • Cause "utf-16"/"unicode" are fixed native(little)-endian; no BOM sniffing in channels.
      • Prototype none (BOM handling would be a new feature/TIP)
      • Suggested ticket comment Since 8.7/9.0 there are explicit utf-16le and utf-16be encodings, so the application can check the first two bytes and pick one. Automatic BOM detection would need a TIP; suggest closing (or moving to RFE).