Triage of all 509 open and Pending tickets of tcllib (bugs, patches and feature requests) on 2026-11-15. Each ticket was checked against tcllib trunk on Linux with Tcl 9.1, 9.0, 8.6 and 8.5, using local servers for the network modules and a Windows 11 VM for Windows-only tickets. Outdated, superseded, duplicate and misfiled tickets are listed separately. Rows marked patch link to a prototype fix with a regression test (147 in all, in tcllib-triage/prototypes). Updated 2026-11-16.
Tcl_InitStubs(interp, "8.6", 0), which fails in Tcl 9 (package require Tcl 8.6 -> version conf…Status now: Open
practcl-initstubs-tcl9, not merged: Fix d3d4117803: practcl generated Tcl_InitStubs(interp, "8.6", 0) and Tk_InitStubs with "8.6", which fails on Tcl/Tk 9. Require "8.6-" instead.REPRODUCED (9.1, 8.6). Generated loader code has Tcl_InitStubs(interp, "8.6", 0), which fails in Tcl 9 (package require Tcl 8.6 -> version conflict). Tk_InitStubs is the same.
practcl::module ... generate-loader output on 9.1 and 8.6. package require Tcl 8.6 in tclsh9.1 gives "version conflict".
::dns::N array behind. Proposed fix in bran…REPRODUCED (9.1, 8.6): a bad option, a missing nameserver, missing udp support or missing tls leave a ::dns::N array behind. Proposed fix in branch bug-f6624cf52ba4 (f8ed54f73b), not merged. That branch has a bug: in the bad-option path it unsets state before building the option list, so the error ends with "must be " and nothing after it.
::dns::resolve error returns do not unset state
package require tls error message kept.
unset state after set opts), add the tests, merge
package require mime creates them, as …Status now: Open
mime-catch-options, not merged: Fix 0534941436: mime read errorCode and errorInfo after a successful catch, so a stale value made the first parse of a multipart message return text/plain. UsREPRODUCED (9.1, 8.6), when ::errorCode/::errorInfo do not exist. Normally hidden because something during package require mime creates them, as the reporter says.
unset ::errorCode ::errorInfo: the first parse gives text/plain, the second gives multipart/mixed, on both 9.1 and 8.6
$errorCode/$errorInfo after catch even on success, in initializeaux, parsetype, getbody, copymessage, buildmessage, parseaddress and addr_next. parsepart swallows the resulting error.
catch ... result opts / return -options $opts $result and drops global errorCode errorInfo. Tests mime-15.0 (multipart parse) and mime-15.1 (parseaddress/buildmessage/copymessage/getbody) fail without the fix. The full mime suite passes on 9.1 and 8.6.)
Status now: Open
graph-tarjan-node-spaces, not merged: Fix 0f65bfa2d3: struct::graph::op::tarjan hung and isCutVertex? gave wrong answers for node names containing spaces. The fix for Bug 2815302 (746fd7f965) was REPRODUCED (9.1, 8.6). This is a regression of Bug 2815302.
{beamsect_C C} with an arc makes tarjan hang (killed by timeout). isCutVertex? gives wrong results for such names (1 instead of 0).
TarjanSub and isCutVertex? use struct::set subtract (set operand) instead of exclude (single element). This was fixed in 746fd7f965 (2009-07-10) and reintroduced by the GSoC integration 42c21b6c75 (2009-09-15). Fleury still has the fix.
exclude in both places. graphops.test passes 434/434 on 9.1/8.6. Without the fix, tarjan-3.9 hangs and isCutVertex?-3.4 fails. The reporter's attached diff is not in the local repo (attachment not synced), so the prototype was written independently.
Status now: Open
pkgindex-version-fixes, not merged: Fix 61e49798ee, 523f7e0556, cdbdd399e4: pkgIndex.tcl entries that do not match the code. nettool's index sourced nexttool.tcl, so package require nettool REPRODUCED (9.1, 8.6). All five mismatches are still on trunk.
package require -exacts every tcllib package from the source tree in a fresh child interp, on 9.1 and 8.6. Got the "provided instead" errors for doctools::toc, math::filters, tcl::chan::std, and nexttool.tcl not found for nettool (and udpcluster, which requires it). rest's index says 1.7 while rest.tcl provides 1.8 (bumped in 0b3dd048f8 without the index). Extra finding: struct 1.5 (struct1.tcl) requires struct::matrix 1.2.1, which no longer exists.
package provide. The nettool typo is also 61e49798ee.
Status now: Open
autoproxy-tls-socket-tcl9, not merged: Fix 29ff5c3abd: autoproxy::tls_socket did not work with Tcl 9. http 2.10 calls it at global level, so its upvar of host and state saw the wrong variables. TakREPRODUCED (9.1, 9.0; works on 8.6)
can't read "uhost" with and without a proxy. Passes on 8.6.
tls_socket uses upvar host/upvar state, assuming that http::geturl calls it directly. http 2.10 calls the socket command from http::OpenSocket via namespace eval ::, so the upvars point to global variables. This also means tls_status was being written into a global ::state.
http::geturl or http::OpenSocket, takes the target from state(socketinfo), and does nothing if neither is found). For the test: autoproxytls.test now simulates the tls package when it is missing, so the existing tests autoproxy-tls-1.1.0/1.1.1/1.1.3 run everywhere. 1.1.0 and 1.1.1 fail on 9.1 without the fix. With the fix the module passes on 9.1, 9.0, 8.6 and 8.5. The second test named 1.1.1 ("with proxy and http proxy config") stays limited to have_tls: it fails with Tcl 8.6 http (Tcl 2d9b31130b), and on 9.x it would need http::register ... {} 0 1.
namespace eval ::, so tls_socket's upvar no longer finds host/port. Proposed fix: tls_socket finds the http state array and takes the target from state(socketinfo). The tls tests now run without a real tls package.
REPRODUCED (9.1, 8.6): two bugs
mime::encodingasciiP "Simple line\r" → 0. A parsed message (mime::initialize -string of a full mail, no CTE header) gets a computed "Content-Transfer-Encoding: quoted-printable" (base64 with the reporter's charset=Cp1252) from buildmessage, but the body is copied unencoded. That is the garbled mail the reporter saw.
$r == {[string length $line] - 1}, a literal string. (2) mime.tcl:getTransferEncoding (new in 1.7) computes [encoding] for non-canonical (parsed) parts too, but copymessageaux only encodes canonical bodies. In 1.6 only the received CTE was written for parsed parts.
REPRODUCED (9.1, 8.6), regression in 1.7 (1.6.2 buildmessage built the string by hand, with no channel translation)
REPRODUCED (9.1, 8.6)
findLineSegmentIntersection {0 0 10 0} {5 1e-8 15 1e-8} and {0 0 10 0} {10 1e-8 20 1e-8} print the message to stderr and return none (9.1, 8.6).
Status now: Open
pkgindex-version-fixes, not merged: Fix 61e49798ee, 523f7e0556, cdbdd399e4: pkgIndex.tcl entries that do not match the code. nettool's index sourced nexttool.tcl, so package require nettool REPRODUCED (9.1, 8.6). Also in the tcllib 2.0 release.
package require nettool from the source tree gives couldn't read file ".../nettool/nexttool.tcl" on 9.1 and 8.6. The tcllib-2-0 tag has the same pkgIndex.tcl.
nexttool.tcl came in with 94a3829047 (2024-09-13, "Tcl 9 portability fixes"). The test suite does not catch it because nettool.test sources nettool.tcl directly. Side note: build/build.tcl, which regenerates pkgIndex, still writes the stale versions 0.5.2 / available_ports 0.1, and ::nettool::version says 0.5.3.
REPRODUCED (9.1, 8.6)
too many nested evaluations) because of the recursion.
::yaml::_parsePlainScalarInBlock: leftover if {$loop == 5} return from fc59390525 (2009), plus one recursion level per line.
REPRODUCED (9.1, 8.6)
"key: some long text\n - text\nother: val\n" and "a: x\n y" also lose the continuation line.
::yaml::_parseBlockNode passes the scalar's own start column as base to _parsePlainScalarInBlock, so continuation lines indented less than the first word end the scalar. The spec only requires them to be more indented than the parent node. The * is then parsed as an alias.
$indent + 1 as the base. Results match PyYAML.
REPRODUCED (9.1, 8.6); fix exists on branch tkt-4dd68ce4e76841f (02df8c0ed9, 2026-03-25), reviewed OK by Harald, not merged
$A/$seA and $B/$seB by an extra sqrt(df-2) (sqrt(df-1) without intercept); seA/seB are already standard errors.
REPRODUCED (9.1, 8.6)
::fileutil::magic::cfront::generate compressed 0 -- Magdir → about 90 rejected lines ("bad modifier" 56, "unknown type" 14 e.g. guid, "type not implemented" 12, "bad indirect offset type" 8) plus about 295 warnings. Then installed the generated script and ran filetype /bin/ls: it returns the same nonsense NAXOS CAD match as in the report, and some inputs throw can't read "typematch(1)". The shipped 5.34-based recognizer gives the correct ELF answer.
cfront::compile opens the make magic file with r+ and rewrites it in place, so it modifies the user's input tree. That should be fixed too.
Status now: Open
zipfile-encode-mtime, not merged: Fix 1ae6b4775e: zipfile::encode stored the ctime of files instead of the modification time. Remove the knownBug constraint from zipfile-encode-encode-6.3.REPRODUCED (9.1, 8.6; existing knownBug test zipfile-encode-encode-6.3)
file: stores $s(ctime) (inode change time) instead of $s(mtime) for the zip timestamp.
file: recorded the ctime instead of the mtime. Proposed fix: use mtime. Test 6.3 is no longer knownBug.
split2matrix -alternate m line sep expand fails with "wrong#args", because the 5-argument branch is labelled 4.REPRODUCED (9.1, 8.6): split2matrix -alternate m line sep expand fails with "wrong#args", because the 5-argument branch is labelled 4.
::csv::split2matrix has a duplicate 4 switch arm. read2matrix has the correct 5.
5.
Status now: Open
yencode-ydecode-fixes, not merged: Fix 69d33816bb: ydecode accepted only all-digit file names, so it nearly always returned default.bin; ydecode -file read in text mode (an error on Tcl 9, corrREPRODUCED (9.1, 8.6). Three bugs:
name=(\d+), so it always returns default.bin unless the name is all digits.
ydecode -file opens the file in text mode. On 9.x this throws on 8-bit data; on 8.6 utf-8 sequences in the data are silently corrupted.
opts array as a scalar ($opts).
regexp {name=(\d+)}, open $opts(filename) r, must be -$opts
REPRODUCED (9.1, 8.6), regression in 1.7.x
-encoding 8bit → two CTE lines. Also happens without -encoding whenever an encoding is computed (e.g. application/octet-stream → base64), including inside multipart.
REPRODUCED (9.1, 8.6)
*a\b* gives <p>*a\b*</p>, and *a\b\** gives <p>*a\b**</p> (9.1, 8.6). Torsten's branch bug-markdown-a028a78b5b7e (33e0f416bb, 2026-01-30) adds only tests and no fix. Some of its expected results (*a\b* giving <em>ab</em>, \c giving c) contradict CommonMark, where a backslash before a non-punctuation character is kept.
(?:[^X\\]|\\X)*. A backslash is allowed only before the delimiter X.
(?:[^X\\]|\\.)*, so any escape pair is consumed and parse_inline then handles the escape. The tests cover *a\b* → <em>a\b</em>, *a\b\** → <em>a\b*</em>, _a\\_ → <em>a\</em>, and *a\* → no emphasis. 2.1–2.3 fail without the fix; all pass on 9.1 and 8.6.
\* case. Prototype with tests available. Some expectations in the branch's tests (*a\b* → ab) differ from cmark.
auth bearer for create_interface, and rest::simple supports it, but create_interface ignores it…REPRODUCED (9.1, 8.6): rest.man already documents auth bearer for create_interface, and rest::simple supports it, but create_interface ignores it. No Authorization header is sent.
auth {bearer ...}.
basic and sign.
lappend proc [list lappend headers Authorization "Bearer $token"]. The ticket's version puts the token into a double-quoted string inside the generated proc body, so a token containing $, [ or " is substituted or evaluated at call time. The test uses such a token. It fails without the fix and passes on 9.1 and 8.6. This patch and the 3613725fff patch both create rest.test (and the cdbdd399e4 patch too), so the tests need merging if more than one is applied.
distribution.snapshot ScmUnpack fails with unknown encoding "binary" for .tgz/.gz archives. Open ticket 146…Status now: Open
tcl9-encoding-binary, not merged: Fix 8ac37e504d, 146ae829a9: -encoding binary is gone in Tcl 9. practcl's snapshot unpacking and doctools::tcl::parse file still used it, so they failed on Ttcl9-encoding-binary, not merged: Close the archive channel that ScmUnpack leaves open before deleting the test files, so that practcl-snapshot-unpack-001 also passes on Windows.REPRODUCED (9.0, 9.1; 8.6 is fine). distribution.snapshot ScmUnpack fails with unknown encoding "binary" for .tgz/.gz archives. Open ticket 146ae829a9 (doctools) reports the same leftover in doctools2base/tcl_parse.tcl:44: doctools::tcl::parse file always fails on Tcl 9. These were missed by the 5da15eba71 cleanup (2024).
fconfigure $fh -encoding binary -translation lf in practcl, and fileutil::cat -translation binary -encoding binary in doctools2base
-translation binary in practcl (both copies) and drops the redundant -encoding binary in tcl_parse.tcl. Tests: practcl-snapshot-unpack-001 (practcl.test) and doctools-tcl-parse-*-12.0 (tcl_parse.test). Two leftovers are not touched: amazon-s3/S3.test:517,543 (constraint-gated tests) and the example in jpeg/jpeg.man:63.
Status now: Open
tcl9-encoding-binary, not merged: Fix 8ac37e504d, 146ae829a9: -encoding binary is gone in Tcl 9. practcl's snapshot unpacking and doctools::tcl::parse file still used it, so they failed on Ttcl9-encoding-binary, not merged: Close the archive channel that ScmUnpack leaves open before deleting the test files, so that practcl-snapshot-unpack-001 also passes on Windows.REPRODUCED (9.1, 9.0; works on 8.6)
fileutil::cat -translation binary -encoding binary fails on 9.0/9.1 with unknown encoding "binary", so doctools::tcl::parse file always fails on Tcl 9. The test suite only exercised parse text.
typemethod file still passes -encoding binary. The 2024 sweep (ticket 5da15eba71, closed Fixed) missed this one.
-encoding binary (-translation binary alone already selects iso8859-1, so the behaviour is the same on 8.6). The new tests run parse file over the existing tcl_data inputs. Without the fix all 18 fail on 9.1; with it they pass on 9.1 and 8.6. The only remaining failure, doctools-msgcat-2.0, also fails unpatched (see 1968777fff).
-encoding binary, other batch). 0c118eed38 (tclvfs, closed Invalid). 0e69b86c2e (Tkhtml, not tcllib).
doctools::tcl::parse file fails on Tcl 9 because of -encoding binary in tcl_parse.tcl. Fix: drop it, since -translation binary already implies iso8859-1. A test is added.
Status now: Open
pluginmgr-cwd-dir-collision, not merged: Fix 1a7976b873: pluginmgr failed to load a plugin when the current directory had a directory with the plugin's name. Use file isfile instead of file exists. AREPRODUCED (9.1, 8.6) for the directory case. The file case is documented behaviour: "if the string contains the path to an existing file then this file is taken as the implementation". The reporter says part of the report was pilot error.
load simple fails with couldn't read file "simple": is a directory.
file exists $name, which is also true for directories
file exists -> file isfile, .man wording, new pluginmgr.test with test pluginmgr-load-1.0, which fails without the fix). The patch adds the test file; the module had no tests.
load, because LoadPlugin checks file exists. A proposed fix uses file isfile, with a test. Loading a file path relative to the cwd is documented behaviour.
REPRODUCED (9.1, 8.6)
pm load x; pm list; pm plugin → y (the last listed plugin) on 9.1 and 8.6, using Tcl-module plugins. Side finding: package (pkgIndex) based plugins do not load at all on 9.x. TIP 579 made the safe base stop syncing the child's auto_path with interpAddToAccessPath, which SetupIp/path rely on. It works on 8.6. No ticket exists for this.
method list uses the instance variable plugin as the regexp capture variable.
::safe::interpConfigure -autoPath).
list captured into the instance variable plugin. A proposed fix, with a new pluginmgr.test, is ready.
REPRODUCED (9.1, 8.6)
pmsub/plugins/{p,q}-1.0.tm with -pattern pmsub::plugins::*: load p works but list returns nothing on 9.1 and 8.6. The "proj* gives ect" remark is expected behaviour: list returns the part matched by *, which load maps back.
method list triggers the package scan with package require bogus-package. The tm unknown handler then only scans the top level of the module roots, never pmsub/plugins/.
string map {* bogus-package} $pattern). The test fails without the fix on 9.1 and 8.6. This patch is stacked on 8ab3be2e6a.patch (it includes it, because both use the new pluginmgr.test).
crc::crc32 -filename of a 4800-byte file equals the strin…Status now: Open
crc32-critcl-negative-sum, not merged: Fix 7bd2d439ca: with tcllibc on Tcl 9, crc32 of data over one chunk (4 KiB for files) failed where C long is 32 bits (Windows): Crc32_c returns a negative valREPRODUCED (emulated, critcl path only); the pure Tcl path works on Windows (9.1, 8.6): crc::crc32 -filename of a 4800-byte file equals the string CRC (D1A741BC). The critcl path (tcllibc), which the triage found to be the cause, could not be tested because tcllibc is not built on the VM.
Crc32_c with a mock that behaves like the C code with a 32-bit long under Tcl 9's Tcl_GetLongFromObj range (LONG_MIN..ULONG_MAX). It fails without the fix and passes with it on 9.1 and 8.6.
Crc32_c returns Tcl_NewLongObj(t ^ 0xFFFFFFFF), which is negative where long has 32 bits. crc32.tcl Crc32Update then stores $sum ^ 0xFFFFFFFF without masking, which gives values like -3632003085. On the next chunk Tcl 9 rejects such a seed with "integer value too large to represent". Tcl 8.6 accepted -ULONG_MAX..ULONG_MAX, which is why 8.6 worked. Harald probably had no tcllibc loaded, so he got the pure Tcl path. In crc32 -filename, an error also left the file open and the token array behind.
crc::crc32 -filename of a 4800-byte file equals the string CRC (D1A741BC). The critcl path (tcllibc), which the triage found to be the cause, could not be tested because tcllibc is not built on the VM.
Status now: Open
ftpd-tcl9-load, not merged: Fix 34a589f660: package require ftpd failed on Tcl 9 with can't read "tcl_platform(user)". The namespace body used 'global tcl_platform', which does nothing oREPRODUCED (9.1, 9.0; 8.6 loads fine)
package require ftpd on 9.1 and 9.0 fails with can't read "tcl_platform(user)".
namespace eval ::ftpd uses global tcl_platform, which does nothing outside a proc. Tcl 9 no longer falls back to global variables in namespace context.
REPRODUCED (9.1, 8.6)
field_decode "=?UTF-8?Q?a_?= =?UTF-8?Q?b?=" → "ab"
REPRODUCED (9.1, 8.6)
(efattr >> 16) & 0x1ff as permissions whenever efattr != 0. In DOS/Windows zips only the low byte is used. The condition should test the MSW. Side finding: VER uses >> 16 instead of >> 8, so the "version made by" host is always reported as MS-DOS. That would need fixing before the host byte can be used instead.
mkTestZip helper; tests zipfile-decode-unzip-8.0, 8.1 in encode.test, constraint unix). Permissions are applied only when efattr >> 16 is non-zero. 8.0 fails without the fix. The full zip suite passes on 9.1 and 8.6.
gcc -std=c99 -pedantic warns "redefinition of typedef 'YYSTYPE'". clang -std=c99 warns that this is a C11 feature [-Wtypedef-redefi…REPRODUCED (gcc -std=c99 -pedantic warns "redefinition of typedef 'YYSTYPE'". clang -std=c99 warns that this is a C11 feature [-Wtypedef-redefinition]. Older clang makes it an error.)
typedef union YYSTYPE YYSTYPE; and bison emits the same typedef again
union YYSTYPE; and uses union YYSTYPE * in the jsonlexp prototype, in both json.y and the generated json.tab.c. Clean under -std=c99 -pedantic with gcc and clang. No regression test possible (compile-time only; critcl build is not covered by the Tcl testsuite).
REPRODUCED (9.1, 8.6)
dict2yaml {k {{val}}} emits k: {val}, which is a flow mapping in YAML, so yaml2dict returns k {}. Quoted "{val}" parses fine.
::yaml::_dumpScalar emits plain scalars even when they start with an indicator character ({ [ & * ! # " ' % @ , | >`).
\ and ". Side finding: huddle create k {[a "b"]} itself errors (huddle::isHuddle runs lindex on non-list strings); not fixed here.
{val} unquoted, which YAML reads as a mapping. Proposed fix quotes values that start with an indicator; tests yaml-36.*.
REPRODUCED (9.1, 8.6)
d/sub without a trailing slash, marked as a directory either by the MS-DOS attribute (0x10) or by the unix mode S_IFDIR, plus a file d/sub/f.txt. unzipfile fails with couldn't open "out/d/sub": is a directory on 9.1 and 8.6. The reporter's archive (from GiD on Windows) probably has such entries, or entries ending in a backslash.
/. Any other zero-size entry is "touched" as a file, and that fails when the directory already exists because a child was extracted first (array names order).
mkTestZip that writes a stored zip with chosen vmb/efattr; tests zipfile-decode-unzip-7.0, 7.1 in encode.test). An entry is also treated as a directory when its name ends in / or \, or efattr has 0x10, or the unix file type is S_IFDIR. Both tests fail without the fix. The full zip suite passes on 9.1 and 8.6.
REPRODUCED (9.1, 8.6)
coroutine::util gets_safety is "unknown or ambiguous subcommand"; no errorcode. Patched: 9/9 pass on both
chan pending input before reading (always 0 at first, and counts data after the newline), never checks the line length, and checks pending only after being woken; gets_safety is not in the ensemble export list
solveGauss {{0 1} {1 0}} {1 2} gives "domain error", because solveGauss never p…REPRODUCED (9.1, 8.6). Worse than reported: even the non-singular solveGauss {{0 1} {1 0}} {1 2} gives "domain error", because solveGauss never pivots.
{{1 2} {2 4}} and {{0 1} {1 0}} with solveGauss, and the ticket example with solvePGauss. All give "domain error: argument not in valid range" (9.1, 8.6).
REPRODUCED (9.1, 8.6)
dict2yaml/huddle2yaml of indented code loses all indentation in the | block.
::yaml::_doLiteralBlock applies string trim to every line.
string trimright; output parses back unchanged in tcllib and PyYAML. Limitation: a value whose *first* line is indented would still need a block indentation indicator; not handled.
REPRODUCED (9.1, 8.6)
::httpd::coro::<uuid> timeout after-script is fired by hand. Result: can't read "query": no such variable, exactly as in the report. The SSL part is not a bug: the man page server is plain HTTP, so curl https:// hangs in the TLS handshake until the 120 s read timeout fires.
::httpd::server method Connect: query is only created after the request line is read. The on error handler uses [dict getnull $query REQUEST_URI], so a read timeout (or any gets error) raises a background error instead of answering 400. The handler also used the wrong key (it should be http REQUEST_URI).
set query [dict create UUID $uuid] before the try and fixes the debug key. The test fails without the fix. Module tests pass on 9.1 and 8.6.
coroutine::util::gets_safety rethrows read errors with return -code $result -options $opts, which loses the error message (separate small bug in modules/coroutine).
query before it exists, after the 120 s read timeout. Your curl used https against a plain-HTTP server, hence the stall. Proposed fix initialises query first (patch + test httpd-client-0009).
REPRODUCED (9.1, 8.6)
huddle equal on two yaml2huddle results fails with Invalid subcommand 'equal' for type '::huddle::Type_str'.
::yaml::_makeChildType writes superClass string; huddle looks for superclass.
superclass the yaml scalar types inherit huddle string methods (e.g. huddle equal). Proposed fix with test yaml-33.1.
REPRODUCED (9.1, 9.0: error; 8.6: wrong hash). The test md5crypt-5.0/5.1 is now marked knownBug, so it is skipped in normal runs.
md5crypt::md5crypt "пароль" abcdefgh → "expected code point values below 0xff"; the test's salt "\0\1\2" gives the same error. 8.6: returns $1$abcdefgh$Oi/N6O.fyd5T7zv71xnU10, but openssl passwd -1 (UTF-8) gives $1$abcdefgh$QfFdDZYjOC5k44ZZ.ueGu0. A Latin-1 "pässwörd" also differs from openssl on both versions.
encoding convertto utf-8 in md5crypt_tcl; tests md5passwd-8.1–8.3 (Ukrainian and Latin-1 passwords, md5crypt and aprcrypt, expected values from openssl), md5passwd-9.* for the pure-Tcl path when C is loaded; 3 fail without fix on 9.1 and 8.6, all pass with it)
REPRODUCED (9.1, 8.6)
package require yaml, huddle null fails with Invalid subcommand 'null' for type '::huddle::Type_null', and so does huddle::json2huddle {{"a": null}}. Without yaml both work.
!!null type from namespace ::yaml::types::null. huddle addType derives the ensemble name from the namespace tail, so it overwrites huddle's own ::huddle::Type_null. (The reporter's patches are phantom artifacts in the repo, so I couldn't read them.)
yaml_null (tag still !!null). Side effect: huddle type of a parsed YAML null becomes yaml_null instead of null. Needs a human decision on the name.
null type replaced huddle's own null type. Proposed fix registers it under a distinct name; tests yaml-34.*.
REPRODUCED (9.1, 8.6)
path set to a directory containing sub/index.html (and sub/index.md). GET /docs/sub/index.html returns "Listing contents of /sub/index.html/" instead of the file (standalone script and module test, 9.1 and 8.6).
content.file method content: the test [file isdirectory $local_file] || [file tail $local_file] in {index index.html index.tml index.md} treats any existing index file as a directory. It then searches for index files inside the file and falls back to DirectoryListing. FileName already maps a top-level "index.html" to the directory, so the tail test is never needed. The fix hypnotoad suggested in the ticket is wrong, because it would treat every non-index file as a directory.
[file isdirectory $local_file]. The test uses index.md because in the test harness ::fileutil::magic::filetype is missing, so .html files return a 500 error there (side issue). The test fails without the fix (0 1). Module tests pass on 9.1 and 8.6.
file isdirectory only (patch + test httpd-client-0010). Note that the workaround suggested in the 2024-01-09 comment inverts the logic for all other files.
REPRODUCED (9.1, 8.6)
compare [fromstr 1.000001] [fromstr 1.000_001] gives 1 (9.1, 8.6). fromstr accepts any garbage silently: abc → {0 abc 0}, 0x10, 12a, 1.2.3 → 123e-3, 1e1_0 → exponent 1. On 8.6 fromstr 09 then fails in arithmetic with an octal error.
_ stays inside the mantissa and is counted as a fraction digit.
_ between digits, fraction, exponent, Inf/NaN), strips the underscores, strips leading zeros of integers (the octal issue), and errors on anything else. decimal.man documents the accepted syntax. 9–10 tests fail without the fix; all pass on 9.1 and 8.6.
%F0%A0%AE%B7 decodes to 4 Latin-1 characters (re-encoded as c3b0c2a0c2aec2b7) instead of U+20BB7.REPRODUCED (9.1, 8.6). %F0%A0%AE%B7 decodes to 4 Latin-1 characters (re-encoded as c3b0c2a0c2aec2b7) instead of U+20BB7.
DES::des -key … -in $chan gives invalid command name "Chunk".Status now: Open
des-chunk-proc, not merged: Fix a1dec8173c: DES::Update and the -chunksize option called the missing proc ::DES::Chunk. Add it, in the style of blowfish: process whole blocks, keep the rREPRODUCED (9.1, 8.6): DES::des -key … -in $chan gives invalid command name "Chunk".
::DES::des installs a fileevent handler Chunk that was never written (blowfish and aes have one).
::DES::Chunk in the style of blowfish: it processes whole blocks, keeps the remainder, zero-pads only the final partial block (same as data mode), and reports errors. It is a bit over the 60-line guideline, mostly because of the tests.
Status now: Open
smtpd-deliver-command-prefix, not merged: Fix 1dd19ccd14: the smtpd -deliver callback worked only with a single-word command. Call it as a command prefix with {*}. Add smtpd.test.REPRODUCED (9.1, 8.6)
-deliver [list deliverto delivered]. Delivery fails with 554 (invalid command name "deliverto delivered").
$deliver ... as a single word, unlike the other callbacks. The ticket's proposed eval $deliver [state ...] would break on data with spaces.
{*}$deliver. Passes on 9.1 and 8.6, fails without the fix.
{*} (the suggested eval would break on message data with spaces). Patch with test in prototypes/1dd19ccd14.patch.
REPRODUCED (9.1, 8.6)
message:PHYS-279; "a:b: c" gives the scalar a:b: c instead of {a:b c}; "http://x: 1" is not a mapping.
_parsePlainScalarInBlock and _parsePlainScalarInFlow: the plain-scalar regexps let :[^ ]+ swallow the key's final :, so the key indicator is lost.
a::b, http://x.org/a, e:f still work). It touches the same line as a23348be0b.
Status now: Open
huddle-compile-bool-tag, not merged: Fix e2ba08b285: huddle compile bool used the tag "bool" instead of "b".REPRODUCED (9.1, 8.6)
huddle compile bool true gives HUDDLE {bool true}, and jsondump then fails with "is not a huddle". The same happens nested (compile {dict a bool} ...). The bug is present since 1.18.
::huddle::compile, branch bool: it wraps with tag bool, but the boolean type's tag is b.
b. The test fails without the fix. Yaml module tests pass on 9.1 and 8.6. Test-isolation side note: huddle-12.0 deletes ::map but leaves the type registered, so any jsondump in a later test fails with invalid command name "::huddle::Type_map". New tests must therefore go before huddle-12.0.
REPRODUCED (9.1, 8.6)
knownbug constraint since bf7b7ef0cc). <div allowfullscreen></div> on a paragraph line becomes <div allowfullscreen></div> inside <p>. With ="1" the tag is kept but stays inside the paragraph.
re_htmltag in parse_inline does not accept attributes without a value (the block regexp was fixed for that in 31b56b6b47; the inline one was not).
<div|p|h1-6|blockquote|pre|table|dl|ol|ul|form|fieldset|iframe|math now ends the paragraph, with a blank-line separator as the tests expect. The inline tag regexp now accepts valueless attributes. The patch removes the knownbug constraint from div-1.3/1.4 and adds div-1.5 (<span hidden> inline) and div-1.6 (an inline tag at line start does not break the paragraph). div-1.3/1.4/1.5 fail without the fix; on 9.1 and 8.6 the result is 89 passed, 6 skipped.
REPRODUCED (9.1, 8.6)
Markdown::convert on cases 9, 11, 16, 17 from the ticket (9.1, 8.6). They are still marked .FAIL and skipped in markdown.test (since 4a1e5f6187, 2022-05-03). x __\.__ x gives x <em></em>.<em></em> x, and [foo_bar](sna\_fu) gives href="sna\_fu".
\\$chr), so __\.__ does not match as strong emphasis. (b) Inline-link URLs are only html-escaped; backslash escapes are not removed.
.FAIL cases basic-1.9/11/16/17 are re-enabled). It allows \\. (any escaped char) in the emphasis body and adds a small unescape helper for link destinations. The emphasis part is the same change as in a028a78b5b.patch. Without the fix the four tests fail; with it markdown.test passes on 9.1 and 8.6 (89 passed, 8 skipped).
REPRODUCED (9.1, 8.6)
"alphabet:\n \"abcdefghi\\\n jklmnopqr\\\n stuvwxyz\"\n" gives abcdefghi jklmnopqr stuvwxyz; PyYAML gives abcdefghijklmnopqrstuvwxyz.
::yaml::_parseDoubleQuoted folds line breaks before handling escapes, so an escaped line break (\ at end of line) becomes a space.
\+newline+indentation before folding, keeping \\ pairs intact. Matches PyYAML on these cases.
starttls and info whoami now always fail with `invalid command name "asnEnc…Status now: Open (Fixed)
ldap-starttls-asncontext, not merged: Fix af927e957f: ldap::starttls and ldap::info whoami always failed with 'invalid command name' since tcllib 1.21. 495ba2b8f5 called asnEncodeString, which asnREPRODUCED (9.1, 8.6). The 2022 fix 495ba2b8f5 made things worse: starttls and info whoami now always fail with invalid command name "asnEncodeString". This affects tcllib 1.21, 2.0 and trunk.
asnEncodeString is not found, because ::asn::asnEncodeString is not in asn's export list. Wrote a mock-server test with ldap::info whoami: it fails on trunk with that error on 9.1 and 8.6. With the fix the full ldap module passes 186/186 on 9.1 and 8.6.
ldap::Whoami and ldap::starttls call the unqualified, unexported asnEncodeString 80 .... The exported asnContext 0 $oid produces the same bytes (tag 0x80 = context-specific primitive [0]).
7719 80 17 ... and the returned authzId)
package require asn, as the reporter suggested. The prototype needs no asn change.
asnEncodeString, which asn does not export, so starttls and whoami fail with "invalid command name" in 1.21/2.0. Proposed fix: use the exported asnContext 0 $oid (same encoding), with a mock-server test.
REPRODUCED (9.1, 8.6)
ftp::Put $h -data {} put.txt returns 1 but uploads the *string* "put.txt" to the remote file ftp.tmp. put.txt keeps its old contents. Append behaves the same. The attached ftp.patch is not in the local repo, so the fix was written from the code.
::ftp::Put/::ftp::Append argument loops use $source == "" to decide whether a positional argument was already seen, so an empty -data value is overwritten by the remote name.
-data {} the remote name was taken as the data and stored in ftp.tmp. Proposed fix with tests in the prototype.
cron::every test 10 mycron called from inside the task leaves the frequency at the old value. A one-shot cron::in task t…Status now: Open
cron-task-reschedule, not merged: Fix 9833d8dca1: a cron task could not change its own schedule, because the command ran inside dict with and its changes were overwritten. Run the command outsREPRODUCED (9.1, 8.6): cron::every test 10 mycron called from inside the task leaves the frequency at the old value. A one-shot cron::in task that reschedules itself is deleted after it runs.
sak test run modules/cron on 9.1 and 8.6
runTasksCoro runs plain commands inside dict with processTable($task), which writes the old values back afterwards (as hypnotoad noted in 2024). One-shot tasks are also put on the cancel list before they run.
dict with and stores err/result afterwards. One-shot tasks are cancelled only if they were not rescheduled. running 0 is no longer written into a task that was deleted.
dict with, which restores the old settings. Proposed patch runs it outside and keeps one-shot tasks that reschedule themselves.
Status now: Open
sasl-ntlm-nonce-dash, not merged: Fix 66b79317f3: SASL NTLM authentication failed when the server nonce started with "-": DES::des took it for an option. Add -- to the DES::des calls.REPRODUCED (9.1, 8.6)
SASL::NTLM::LMhash/NThash with a nonce starting with "-" (new test) on 9.1 and 8.6. Both fail with bad option "-...": must be one of -chunksize, ....
DES::des ... -key $key $nonce has no --. A server nonce that starts with "-" is parsed as an option. DES::des already accepts --.
-- to the three DES::des calls, as the ticket proposes. sasl tests: 21/21 pass on 9.1 and 8.6. The new test fails without the fix.
--, so a server nonce that starts with "-" is taken as an option. Proposed fix with a test in prototypes/66b79317f3.patch.
REPRODUCED (9.1, 8.6) with a strict mock server; pyftpdlib itself not available (no network)
Put -data, Append -data and Put file. tcllib's own ftpd tolerates it, because it reads the already-buffered data later.
InitDataConn installs the writable handler (HandleOutput/HandleData) as soon as the data connection exists, before the server has accepted STOR/APPE. -buffering line is not the cause: full buffering only delays the same early write until the close.
StartPut, started on the 1xx reply to STOR/APPE or when the connection arrives later; error branches unset state.data in active mode too; CloseDataConn drops a pending writer. Tests ftp-4.0/4.1/4.2 with the strict server, plus the regular ftpd Get tests)
REPRODUCED (9.1, 8.6)
<img src="/images/etoyoc-banner.jpg">. Tested with sak on 9.1 and 8.6.
::httpd::reply method html_header emits the author's site banner when no banner option is given. DirectoryListing and the default error templates call it without that option.
<img> only when banner is given, and changes both the amalgamated httpd.tcl and build/reply.tcl. The test fails without the fix (1 1 instead of 1 0). Module tests pass on 9.1 and 8.6 (27/27).
banner option is given (patch + test httpd-client-0011).
REPRODUCED (9.1, 8.6)
Rapha{\"{e}}lle -> Rapha"elle, {{Oxytocin}} -> Oxytocin
\ as an escape that is removed. Inner braces and TeX commands are lost.
-raw) and a decision.
REPRODUCED (9.1, 8.6)
-multiple variant (-named_arguments_first 0). mytest -myarg a gives mydefault a. The full tepam test suite passes with and without the patch on 9.1 (non-Tk files) and 8.6 (all files, Tk ones included).
lappended to it, for named arguments and for leading unnamed -multiple arguments. The named-arguments-first unnamed path already uses set ... [lrange ...] and is correct.
set), later ones are appended. Tests .2, .3 and .5 fail without the fix.
method variant and the Ensemble variant both u…REPRODUCED (9.1, 8.6): still 4x tip479-001..004 on trunk. Two sets are real duplicate names (the method variant and the Ensemble variant both use tip479-001..004). The other two copies come from the intended "Repeating tests with 8.7 features" block that build.tcl appends. All 8 titles say "a later ensemble definition trumps a more primitive one", which does not match what the tests do (dictargs argument handling).
REPRODUCED (9.1, 8.6)
ip::version 10.0.0.1/64 gives 4. ip::version 10.0.0.1/dsffds/44 gives 4. ip::mask 10.0.0.1/dsffds/44 gives dsffds. ip::prefix 10.0.0.1/64 errors with "negative shift argument". The attached diff is not retrievable (fossil artifact cannot resolve 927b8447…).
version returns -1 for more than one "/", an empty mask, a bit count above 32/128, or a non-numeric mask that is not an address of the same family. ip::mask is unchanged (making it return -1 would change the API). dns module tests pass on 9.1 and 8.6.
ip::mask should validate.
Status now: Open
tool-uuid-seed-no-user, not merged: Fix 4c3e532500: package require tool failed when none of USERNAME, USER or user was set in the environment. tool::uuid_seed read env(user) unconditionally; drREPRODUCED (9.1, 8.6)
env -u USER -u USERNAME -u user tclsh with package require tool gives can't read "::env(user)": no such variable on 9.1 and 8.6. The package fails to load because uuid_seed runs at load time.
::tool::uuid_seed: the code sets a local user from env(USERNAME), env(USER) or env(user) without checking the last one, and never uses it. The seed uses the literal string user@.
huddle wrap s abc gives "wrong # args: should be "huddle wrap node"", while huddle.man still documents only `huddle wrap t…REPRODUCED (9.1, 8.6). huddle wrap s abc gives "wrong # args: should be "huddle wrap node"", while huddle.man still documents only huddle wrap tag src. The man page's type-callback example (huddle to_node, huddle wrap D ..., setting) also describes the 0.1 API.
::huddle::wrap {node} was changed to one argument in the 0.2 rewrite.
wrap accepts both node and the 0.1 form tag src (an empty tag gives HUDDLE src, as in 0.1), and huddle.man documents both forms. The test fails without the fix. Yaml module tests pass on 9.1 and 8.6.
REPRODUCED (9.1, 8.6)
lo {ipv4 {{} 255.0.0.0} ...}, with no addresses and no MACs. broadcast_list gives only 127.0.0.1. hwid_list falls back to the fake 0x010203040506. ip_list fails with invalid command name "ldelete".
::nettool::dump only understands the old inet addr:X Bcast:Y Mask:Z / HWaddr format. The new format uses inet X netmask M broadcast B, ether MAC, and iface: with a trailing colon. Also ldelete, used by every ip_list, is never defined.
::nettool::ifconfig_parse and accepts both formats. It maps the new format to the old keys (addr:, Mask:, Bcast:, ether:), so callers are unchanged. The patch also defines ::ldelete next to ::ladd. With it, dump/ip_list/mac_list/hwid_list/broadcast_list return real data here. The new tests fail without the fix. 28/28 pass on 9.1 and 8.6. The attached nettool.diff (switch to ip address show and /proc/net/arp) is not in the local repo (blob missing), so I could not compare it.
ip/proc as the ticket suggests is a possible later step, because net-tools is deprecated.
ldelete. Proposed patch parses both formats and adds the missing helper, with tests.
REPRODUCED (9.1, 8.6); the 2021 fix sits on unmerged branch tkt-d74e6561a2-yaml-huddle (f4046fd6a8)
yaml::huddle2yaml [yaml::yaml2huddle "a: 1\n..."] prints only --- and then the raw HUDDLE {!!map ...} text.
::yaml::_imp_huddle2yaml knows only the huddle types string/list/dict, not yaml's mapping/sequence/str/int/…, so everything falls through to the default branch. Huddle number/boolean values are also emitted raw.
REPRODUCED (9.1, 8.6)
invalid command name "::tcl::clock::SetupTimeZone" (auto_load); 8.6: can't upvar from variable to itself from global env in msgcat::Init
global inside a coroutine to coroutine::util::global (upvar #1). This includes Tcl's own globals (auto_index, auto_path, env, tcl_*), so auto_load, msgcat, etc. lose them. Qualified names (global ::foo) are mapped wrongly too.
global. Other names stay coroutine-local (1.2 checks this). 3 of the 4 tests fail without the fix. The patch adds the test file; the module had no tests.
global override entirely (a design decision)
csv::split -alternate {a;b""b;c} \; gives a b {;c}, expected a b\"b c.REPRODUCED (9.1, 8.6): csv::split -alternate {a;b""b;c} \; gives a b {;c}, expected a b\"b c.
::csv::Split, alternate DFA. A " in the middle of an unquoted value switches to the quoted-value state. The following " then closes the value, and the token after it is dropped.
"" inside unquoted values. Lone quotes are handled as before.
"" there as a literal quote, as the regular parser and the documented format do.
Status now: Open
bibtex-comment-anchor, not merged: Fix 6bef0b5d0f: bibtex silently dropped a record that contained the word "comment" or "preamble" in a field. Anchor the record-type regexps.REPRODUCED (9.1, 8.6)
title = {No comment on this} is silently dropped (taken as @comment). note = {See preamble {x} here} is returned as a preamble.
comment and preamble regexps are not anchored with ^. The string one was fixed in 1.17. The attached patch is not in the repository (its artifact is missing).
^ to both regexps. 7.0 and 7.1 fail without the fix. All 24 tests pass on 9.1 and 8.6.
([^\n])* captures only the last character, so -commentcommand gets e.g. } instead of the comment text.
REPRODUCED (9.1, 8.6)
MySubclass clm works, MySubSubclass clm → unknown method. On 9.x the oo::util version replaces the native oo::define::classmethod (TIP 478), which does not have this bug.
oo::class.Delegate create mixes in only the delegates of the direct superclasses. The delegates do not inherit from each other, so grandparent class methods are not seen.
next also work (A/B/C chain checked). The test fails without the fix on 9.1 and 8.6. Conflicts textually with 3609183fff.patch, see there.
argv0 does not exist (embedders like eggdrop) or when package require imap4 runs inside a proc.Status now: Open
imap4-load-without-argv0, not merged: Fix dbde846f53: package require imap4 failed when argv0 did not exist, e.g. in an embedded interpreter. Check for ::argv0 before comparing it with the script REPRODUCED (9.1, 8.6). Loading fails when argv0 does not exist (embedders like eggdrop) or when package require imap4 runs inside a proc.
unset ::argv0; package require imap4 gives can't read "argv0": no such variable on 9.1 and 8.6.
if {[info script] eq $argv0} uses an unqualified argv0 (local scope inside a proc) and does not check that it exists. The reporter's info exists argv0 attempt failed because it was not namespace-qualified.
[info exists ::argv0] && [info script] eq $::argv0. Both tests fail without the fix and pass on 9.1 and 8.6. Moving the demo to examples/, as aku suggested, would be the larger alternative.
fossil add)
::argv0 guarded by info exists, with a new imap4.test covering loading without argv0 and from inside a proc.
REPRODUCED (9.1, 8.6)
columns/rows under util-linux script(1) with stty cols 123 rows 45 (ncurses 6.6 tput). With stdin on the tty: 123 45. With stdin a pipe or /dev/null (the reporter's echo ... | tclsh case): 80 24.
exec tput cols <@stdin captures tput's stdout and stderr, so tput sees a terminal only through stdin. When stdin is not a tty it falls back to 80x24.
2>@stderr so tput can use the terminal on stderr. Test 1.1 fails without the fix. Side effect: a tput error message now goes to stderr instead of raising a Tcl error.
state(size) is never set and the query fails with an "Event error". The same happens…Status now: Open
dns-tcp-partial-reads, not merged: Fix 187a6fa7ae, 022bca0ff5: the dns client did not handle TCP replies that arrive in several reads. Buffer until the 2-byte length has arrived, read the lenREPRODUCED (9.1, 8.6): if the first TCP read returns 1 byte, state(size) is never set and the query fails with an "Event error". The same happens after an empty read. The id check also scans the latest chunk instead of the reply, and a length above 32767 reads as negative.
::dns::TcpEvent
Status now: Open
comm-lost-conn-peer, not merged: Fix 4811399edc, 1491894fff: comm raised an error when a client disconnected after the "commIncoming race condition" warning. commLostConn now clears the penREPRODUCED (9.1, 8.6)
websocket::close never sends the close frame: the peer receives nothing and the socket is just closed.REPRODUCED (9.1, 8.6). websocket::close never sends the close frame: the peer receives nothing and the socket is just closed.
websocket::takeover on the server side, websocket::close $srv 1000. The client read 0 bytes.
::websocket::close sets Connection(state) CLOSED before calling send $sock 8, and send refuses to write unless the state is CONNECTED. The second half of the ticket (waiting for the peer's close reply instead of disconnecting at once) is a bigger change and is not in the prototype.
close now sets a transient state CLOSING, and send allows only opcode 8 in that state. The early return now also covers CLOSING. That prevents recursion when the close frame cannot be written: send fails, calls close 1001, and that nested call returns immediately; I checked this with a peer that had already gone. The new test passes on 9.1 and 8.6 and fails without the fix. I ran it locally, without the network tests.
REPRODUCED (9.1, 8.6)
PATH=/usr/bin:/bin, then nettool::dump/ip_list/hwid_list/mac_list/broadcast_list/arp_table on 9.1 and 8.6. All fail with couldn't execute "ifconfig" (or "arp"). ifconfig/arp are in /usr/sbin here. Package loaded with the pkgIndex typo from 61e49798ee fixed locally.
::nettool::dump and ::nettool::arp_table (Linux and macOS variants) run exec ifconfig and exec arp -a without a path. /sbin and /usr/sbin are often not in PATH for ordinary users.
::nettool::sbin_exec: it uses auto_execok and falls back to /sbin, /usr/sbin, /usr/local/sbin. The 4 exec calls now go through it. The test drops sbin from PATH and calls dump. It fails without the fix. The module's tests pass 26/26 on 9.1 and 8.6.
after canary tic…REPRODUCED (9.1, 8.6). A peer that sends 1 byte and stalls freezes the server's event loop until it closes. In my test, a 100 ms after canary ticked twice in about 2 s instead of about 20 times.
websocket::takeover; a second tclsh sends 1 byte, sleeps 2 s, closes.
::websocket::takeover sets -blocking on, and ::websocket::Receiver does read $sock N for the header, length, mask and payload. Writes in send are blocking too.
Status now: Open
tar-gnu-longlink, not merged: Fix 7446477f89: tar::create rejected names that do not fit into the prefix/name fields or are longer than 255 characters. Write a GNU ././@LongLink entry for tar-gnu-longlink, not merged: Run tar-tkt-7446477f89 only on Unix: the names are longer than MAX_PATH on Windows, and the result is compared with the output of the system tar.REPRODUCED (9.1, 8.6)
tar::create with a 154-char name without /, and with a 271-char path. It fails with "path name cannot be split into prefix and name" or "path name over 255 chars". GNU tar handles both with a ././@LongLink (type L) entry, which tar.tcl already reads (HandleLongLink).
formatHeader supports only the ustar prefix/name split. Fix: when the name cannot be split or exceeds 255 chars, emit a GNU @LongLink entry with the full name, followed by the normal header with a truncated name.
tar tf). The new test fails without the fix. The tar module passes on 9.1 and 8.6 (13/13). Pre-existing limitation, not addressed: tar::remove does not understand LongLink entries.
././@LongLink entry for them. Since tar already reads that format, proposed fix: write it too (patch with test).
REPRODUCED (9.1, 8.6; snit 1 and 2)
err=dtor error, and the constructor error is lost. Also, the half-destroyed object is left behind: the instance command still exists and dog destroy then fails with "Called 'destroy' method in constructor".
RT.DestroyObject uncaught after a constructor error. RT.DestroyObject aborts in the middle of its cleanup if the user destructor throws.
RT.DestroyObject finishes the cleanup (instance command, hull, namespace) even if the destructor fails, then rethrows. snit.test passes on tclsh9.1/8.6 and wish9.1/8.6 (Xvfb). Without the fix, error-1.3a/b fail, and on tclsh 82 later tests cascade because of the leftover object.
$obj destroy with a failing destructor now really destroys the object before rethrowing)
Status now: Open
huddle-isHuddle-nonlist, not merged: Fix e3a6e3e02a: huddle list with a quote as element failed, because isHuddle applied list commands to strings that are not lists. Check string is list first.REPRODUCED (9.1, 8.6)
huddle list \" raises "unmatched open quote in list" on trunk (9.1, 8.6) and in 1.17 and 1.18. huddle list \{ fails the same way.
::huddle::isHuddle runs lindex $obj 0 on every argument before checking that it is a list.
![string is list $obj] || [llength $obj] != 2 || [lindex $obj 0] ne "HUDDLE". The test fails without the fix. The yaml module tests pass on 9.1 and 8.6 (157/157).
string is list first (patch + test huddle-3.7).
websocket::test without upgrade, ::websocket::Client_<srv>_<cli> stays set.REPRODUCED (9.1, 8.6). After a successful websocket::test without upgrade, ::websocket::Client_<srv>_<cli> stays set.
websocket::test s0 c0 /x {...} followed by info vars ::websocket::Client_*.
::websocket::test creates the context; only ::websocket::upgrade unsets it.
websocket::discard cliSock.
live /x h1 a and live /x h2 b, a client offering b, a gets a fro…REPRODUCED (9.1). Behaves as described; it is really an RFE. With handlers live /x h1 a and live /x h2 b, a client offering b, a gets a from h1: the order in which handlers were registered wins.
websocket::test with two handlers, inspecting Client(protocol).
::websocket::test loops over the server's handlers and takes the first one that matches any client protocol.
REPRODUCED (9.1, 8.6; design limitation)
rest::get $url {} {format json} returns the raw string. Only result is honoured in rest::simple, although rest.man lis…REPRODUCED (9.1, 8.6): rest::get $url {} {format json} returns the raw string. Only result is honoured in rest::simple, although rest.man lists format for simple (and uses it in the Twitter example). create_interface accepts both.
format json returned raw text, result json returned a dict.
dict exists $config result.
format like create_interface does. rest-1.0 fails without the fix and both pass on 9.1 and 8.6. The rest module had no test file before. The test stubs tdom if it is missing, since only JSON is used.
format although the docs and create_interface treat it as a synonym of result. Proposed fix accepts both, with a test.
fileutil::lexnormalize ../toto gives toto; a/../../b gives bREPRODUCED (9.1, 8.6): fileutil::lexnormalize ../toto gives toto; a/../../b gives b
../a -> a) never run, because they are guarded by the non-existent internal LexNormalize.
.. for relative paths when nothing is left to remove. Note: fileutil::jail relies on the current behaviour, so the attached fileutil.tcl must not be applied as is; it is also wrong for absolute /...
.. are kept for relative paths (absolute paths unchanged). jail now strips leading .. itself, so its behaviour is unchanged (jail-2.3 and the new jail-2.3.1 pass). Without the fix, fu-lexnormalize-2.3..2.7 fail on 9.1 and 8.6. With it, fileutil passes 463/463 on both.
REPRODUCED (9.1, 8.6)
oo::class create example {self mixin ooutil::singleton; ...}) returns two different objects on 9.1 and 8.6; ooutil::singleton create ... works. Applying the mixin after creation (oo::objdefine example mixin ooutil::singleton) also works.
oo::class.Delegate create (classmethod support, added in the same release) runs the definition script, then calls oo::objdefine $cls mixin {*}$superdelegates. This replaces all object mixins, so the self mixin singleton set by the script is lost. The doc example also uses the unqualified name singleton, which does not resolve.
mixin {*}$superdelegates {*}[info object mixins $cls]) and changes the example in ooutil.man and the code comment to ooutil::singleton. The test fails without the fix on 9.1 and 8.6. Conflicts textually with f990aa37bc.patch (same line): if both are applied, the line becomes oo::objdefine $cls mixin $delegate {*}[info object mixins $cls].
ooutil::singleton.
REPRODUCED (9.1, 8.6). Partly addressed by the 472259ffff prototype.
typedGetoptions with -n abc for n.integer gives only the usage text. The specific "Option requires integer argument -- n" is dropped.
[typedUsage] and discards $arg. The attached patch is reversed (p4 "#7 - #1"), against 1.3.3, and bundles the file/dir and .must features plus a fileutil dependency.
CMDLINE USAGE INVALID {Option requires integer argument -- n}, test typed-cmdline-8.1).
Status now: Open
generator-exists-colorful-names, not merged: Fix 3578420fff: generator exists failed for generator names with glob or other special characters. Use namespace which -command. Add generator.test.REPRODUCED (9.1, 8.6)
::g[1] used with generator foreach gives no items on 9.1 and 8.6. generator exists ::g? returns 1 if a command ::gx exists.
exists uses info commands $generator, which treats the name as a glob pattern.
exists now uses namespace which -command; adds the module's first test file, generator.test, with tests generator-exists-1.0/1.1). Both tests fail without the fix and pass with it on 9.1 and 8.6.
exists passes the name to info commands as a pattern. Proposed fix using namespace which -command, with a new generator.test.
REPRODUCED (9.1, 8.6)
Invalid level 'none' - levels are ... on trunk with 9.1 and 8.6. With the fix the child is created at level none and all its level commands are no-ops. The log module passes 372/372 on 9.1 and 8.6.
::logger::init, the parent walk calls enable $enabled, and the tree setlevel calls disable/enable with the parent's currentloglevel. Both reject the pseudo-level none.
setlevel accepts none and disables every level. The parent walk skips enable for none. Side effect: ${log}::setlevel none now works too.
none is passed to enable/setlevel. Proposed fix with test logger-18.0: setlevel treats none as "disable all", and the init walk skips it.
REPRODUCED (9.1, 8.6)
{list-no-comma: [ "a" "b" "c" ]} gives {list-no-comma {a b c}}; {x: [ "a": "b", "c": "d" ]} gives {x {a b c d}}. Compared with PyYAML: the first is an error, the second is a list of single-pair mappings.
::yaml::_parseFlowNode ignores commas between flow entries. A : inside a flow sequence just adds another sequence element.
k: v inside [...].
[a: b, c: d] yields single-pair mappings as the spec requires; tests yaml-41.*.
REPRODUCED (9.1, 8.6)
yaml2dict "a: |\n x", "a: >\n x" and "a: x # c" (no final newline) hang on both versions. With a final newline they work.
::yaml::_getLine sets data(start) to length+1 when the last line has no newline. _eof tests start == length and never becomes true, so _parseBlockScalar and the block parser loop forever.
_getLine moved past the end of a buffer without a final newline, so the EOF check never fired. Proposed fix clamps the position; tests yaml-40.*.
CWD sub, absolute names such as SIZE /sub/f.txt are resolved against the current directory instead of the root.REPRODUCED (9.1, 8.6): after CWD sub, absolute names such as SIZE /sub/f.txt are resolved against the current directory instead of the root.
Status now: Open
smtp-rejected-message-error, not merged: Fix 3308105fff: since tcllib 1.21 (1df85e69fb) a message rejected by the server made smtp::sendmessage return code 7 with a raw dict instead of raising an errREPRODUCED (9.1, 8.6; changed form)
sendmessage now returns code 7 with the raw response dict (diagnostic {...} args {} code 451), not an error. With -Bcc, a rejected second (Bcc) copy is not reported at all.
return -options $copts $cres. That leaks the internal code 7, and copts2/cres2 of the Bcc send are dropped. Side note, not fixed: the header-restore code from 8340b8ed2f is dead because of typos (dict set origheaderx, dict get orignames without $).
REPRODUCED (9.1, 8.6)
nmea::configure rate N with no open file fails (can't read nmea(fh)), although the man page says no file needs to be open.
::nmea::configure the rate branch schedules read_file on $nmea(fh) unconditionally.
[info exists nmea(fh)]. Documents close and configure. nmea had no test file. 3 of the 4 new tests fail without the fix, and all pass on 9.1 and 8.6. Use fossil diff --new-file / fossil add for nmea.test.
configure rate without an open file. It adds the first nmea tests.
REPRODUCED (9.1, 8.6): the probable cause is reproduced, not the CPU load itself, because the reporter gave no script
interact on a chan pipe input (fed "\n"). Afterwards the input channel stays non-blocking (fconfigure -blocking = 0). After interact on stdin, any later gets stdin/read stdin loop in the application returns at once instead of waiting, which spins at 100% CPU.
term::receive::listen sets -blocking 0 and unlisten never restores it. menu/pager interact call listen/unlisten on -in (stdin by default).
configure -proxy_host alone still does not install http::config -proxyfilter, and the docs do not say so (they do not do…REPRODUCED (9.1, 8.6): configure -proxy_host alone still does not install http::config -proxyfilter, and the docs do not say so (they do not document init's arguments either)
GET https://host/... (absolute URL). That is an http package issue (Tcl 2d9b31130b; on Tcl 9 you need http::register https 443 ... {} 0 1). The tls tests also need a reset of -proxyfilter.
init calls http::config -proxyfilter. (The attached patch also has a bug: string length options(proxy_host) is missing the $.)
REPRODUCED (9.1, 8.6)
"a: |\n line\n # mid\n end\n" gives a {line\nend\n}; the # line is dropped (only the first content line survives). Compared with PyYAML.
::yaml::_parseBlockScalar skips every line starting with #, whatever its indentation.
# lines are treated as comments; test yaml-39.*.
REPRODUCED (9.1, 8.6) / PARTLY-FIXED (indented comments are kept since c443dd347c, 2014, ticket 3612465)
lsort -dictionary order. A comment after the last key of a section ends up above whatever section is written next. A trailing comment at end of file is dropped.
array names sections and sorts keys. _loadfile drops pending comments at EOF. The comment for "after last key" is attached to the following section header, so it moves when sections are reordered.
sections()) and for keys (new order array). commit writes in file/creation order. Trailing comments are kept in a trailer variable and written at the end. The return order of ini::sections/keys is unchanged (test inifile-1.2 depends on hash order).
ini::sections/ini::keys in file order.
ini::revert passes the handle name instead of the channel to _setfileenc (it fails when -encoding was used). namespace eval ... {array set data {}} does not clear the arrays, so revert does not actually discard changes.
parseMimeValue {form-data; name="f"; filename="a;b.txt"} gives filename "a, and importFile -client returns "a.REPRODUCED (9.1, 8.6). parseMimeValue {form-data; name="f"; filename="a;b.txt"} gives filename "a, and importFile -client returns "a.
Finish closes state(sock) but keeps the name, and a later dns::cleanup closes it again. The attached…Status now: Open
dns-async-double-close, not merged: Fix 2853152fff: an async dns query closed its socket twice. Clear state(sock) after closing, and skip an empty socket in cleanup.REPRODUCED (9.1, 8.6), by code and test. Finish closes state(sock) but keeps the name, and a later dns::cleanup closes it again. The attached patch was never applied. A real name reuse could not be forced here (Unix socket names are pointers), so the test watches the close calls.
::dns::Finish and ::dns::cleanup.
state(sock) after closing, and cleanup skips an empty socket.
tcl_platform(wordSize) is 4, ::crc::signbit is 2147483648, and CRC-32/BZIP2 of "123456789" through `::crc::Cr…Status now: Open
crc-32bit-wordsize, not merged: Fix 2848168fff: ::crc::Crc gave wrong results for non-reflected 32-bit CRCs where tcl_platform(wordSize) is 4 (Windows): the signbit mask cleared bit 24. ValuREPRODUCED on Windows (9.1, 8.6): tcl_platform(wordSize) is 4, ::crc::signbit is 2147483648, and CRC-32/BZIP2 of "123456789" through ::crc::Crc gives F9E2E667 instead of FC891918.
::crc::signbit 2147483648
::crc::Crc and crc32.tcl Crc32_tcl mask with ~signbit>>7, which clears bit 24 on Windows. Since Tcl 8.5 has bignums and the values are always masked non-negative, the mask is not needed at all.
tcl_platform(wordSize) is 4, ::crc::signbit is 2147483648, and CRC-32/BZIP2 of "123456789" through ::crc::Crc gives F9E2E667 instead of FC891918.
::crc::Crc_table 32 0x04C11DB7 0 + ::crc::Crc 123456789 32 tbl 0xFFFFFFFF 0xFFFFFFFF 0 with tclsh91 and tclsh86t. The prototype prototypes/2848168fff.patch removes the mask; its test crc16-10.0 emulates this signbit on Linux.
REPRODUCED (9.1, 8.6)
dns::configure -protocol tcp. spf::SPF returns oogle-site-verification=abc=spf1 ip4:... and spf::spf returns none for every address.
::spf::SPF strips the first character of each rdata (dns.tcl ReadString already removes the length byte) and concatenates all TXT records, including non-SPF ones. Its error path also uses undefined $log/$r.
v=spf1 records (RFC 7208 4.5), returns none if there is none, and raises a permanent error if there are several. It also fixes the error-path variables. spf.test passes on 9.1/8.6; the new tests fail without the fix. Verified end-to-end with the local DNS server: + / - results.
Status now: Open
spf-ipv4-ipv6-mismatch, not merged: Fix 2526633fff: SPF checks raised an error when an IPv4 address was compared with an IPv6 one. Such addresses now simply do not match.REPRODUCED (9.1, 8.6)
spf::Spf 2001:db8::1 ... "v=spf1 ip4:... ip6:... -all" gives error in "ip4": type mismatch. Same for an IPv4 sender with ip6 first, and for a/mx with an IPv6 sender.
_ip4, _ip6, _a, _mx and _ptr call ip::equal on addresses of different families, which throws.
IpMatch returns 0 when the families differ, and _ptr gets the same check. spf.test passes on 9.1/8.6; the new tests fail without the fix. The patch is independent of 2526629fff; both apply together (47/47 pass).
a/mx still only look up A records (no AAAA), a separate limitation.
SID=foo ; path=/ ; ... secure , with spaces before ";" and a trailing space. That is not RFC 6265 sy…REPRODUCED (9.1, 8.6). setCookie still emits SID=foo ; path=/ ; ... secure , with spaces before ";" and a trailing space. That is not RFC 6265 syntax (name=value; attr=val). Modern browsers trim the value, but the output is still malformed.
ncgi::setCookie ...; set ncgi::cookieOutput and the existing ncgi-11.2/12.4 header tests, on 9.1 and 8.6
Status now: Open
matrix-link-stale-elements, not merged: Fix 2426988fff: an array linked to a struct::matrix kept stale elements after deleting rows or columns or transposing. Pass unsets on to the linked array.REPRODUCED (9.1, 8.6)
a and, with -transpose, to b, then delete row 1. The linked arrays keep stale elements for the old last row (0,2, 1,2). The same happens after delete column, delete rows/columns n and transpose of a non-square matrix.
_link puts only a write trace on the internal data array. Moving cells down is propagated, but unsetting the vacated last row/column is not. Fix: trace {write unset} and make MatTraceOut unset the array element on an element unset. A whole-array unset is ignored, and the existing cross-array unset propagation is unchanged.
REPRODUCED (9.1, 8.6; snit 1 and 2)
::snit::RT.from)
dog create spot -foo -bar -snarf 0 -foo {a b} with [from args -bar xyzzy] returns -snarf and leaves {-foo 0 -foo {a b}}; tclsh9.1, tclsh8.6, snit.tcl + snit2.tcl.
RT.from uses lsearch -exact on the whole list, so it matches values at odd positions too, and it only removes the first occurrence.
from on lists with a positional prefix of odd length would stop finding the option.
REPRODUCED (9.1, 8.6; Xvfb)
myWidget .x twice, then myWidget info instances gives can't read "instance": no such variable. Also when the name collides with a plain frame. snit 1 and 2.
installhull, which runs outside the catch that cleans up failed constructions. A failed hull creation leaves a Snit_inst* namespace without Snit_instance.
REPRODUCED (9.1, 8.6)
typemethod {a_b c} / {a b_c} and the same for methods. The second overwrites the first (both return 2). snit 1 and 2.
[join $method _].
RT.MangleMethod escapes %→%25 and _→%5f inside tokens before joining with _. Names without _/% keep the same proc names. Adds an internal %J substitution, so the documented %j (underscore join in delegate patterns) is unchanged. It covers compile-time and runtime snit::method/snit::typemethod, dispatch, and info args/body/default. snit.test (snit 2): 0 failures on 9.1/8.6. snit 1 failure counts unchanged and the new tests pass. The new tests fail without the fix. Larger than the usual limit, but mechanical.
_ and % inside tokens in the generated proc names; documented %j unchanged (prototypes/2063144fff.patch, with tests).
Status now: Open
jpeg-bad-exif-offsets, not merged: Fix 2055837fff: jpeg::getExif raised an error for Exif entries whose offsets point past the end of the data, e.g. a bad InteropIFD. Skip such entries.REPRODUCED (9.1, 8.6)
expected integer but got "" or can't read "num": no such variable.
REPRODUCED (9.1, 8.6, 8.5)
unknown error code "deflist" (for locale pl_pl). LANG=C and de_DE give real messages. doctools2 has the same problem: doctools::idx::parse text foo under pl_PL gives error on line 1.3: docidx/plaintext instead of the message.
mcload the mpformats/*.msg catalogs, but those exist only for c/en/de/fr. There is no ROOT catalog, so other locales find nothing, and the overridden mcunknown prints the code.
mcload, each checker reads c.msg and defines its messages for the root locale {}. msgcat then falls back to C for unknown locales or incomplete catalogs. Changing mclocale inside mcunknown does not work: msgcat 1.6+ reloads package locales and calls file, which is a formatter command in the checker interp. The new tests fail on 9.1 and 8.6 without the fix and pass with it. The doctools module is 602/602 on both. The doctools2 (doctools::msgcat::init) side is not covered; fixing it there needs a root-locale fallback in doctools2base/msgcat.tcl.
doctools-msgcat-2.0 fails when the user's LANG is not C (here uk_UA). The test sets env(LANG) after msgcat was already loaded. This is a test-environment problem.
Status now: Open
tiff-addentry-overwrite, not merged: Fix 1924065fff: tiff::addEntry did not replace an existing entry with the same tag, and nametotag matched only part of the name. Remove the knownBug constrainREPRODUCED (9.1, 8.6)
-constraints knownBug: tiff-23.* return {Artist Andreas} {Artist Andreas}. Run directly, tiff.test needs TCLLIBPATH, because its support block lacks cmdline for fileutil. sak works.
ifd($e), where $e is the whole {tag format value} list, not the tag. The new entry therefore lands under a separate key, the file gets two entries with the same tag, and getEntry returns the old one. Also, nametotag checks the hex-tag pattern against $x (the whole entry) instead of $y, so addEntry $f {{013b 2 AK}} fails with "unknown tag".
ifd([lindex $e 0]), fixes the nametotag pattern, removes knownBug from tiff-23.* and adds tiff-23a.* (overwrite by numeric tag). Without the fix 6 tests fail. With it, 285 tests, 0 failures on 9.1 and 8.6. Unrelated, not fixed: _copyData's EXIF branch uses a stale $x in lindex $ifd($x) 0 and should use ifd(8769).
REPRODUCED (9.1, 8.6)
option -foo[clock format 0 ...] in a type body. 9.1: unknown subcommand "getsystemlocale". 8.6: can't read "TclLibDir". Both snit 1 and 2.
$compiler alias clock ::clock in both snit versions, as kennykb suggested. snit.test: 0 failures on 9.1 and 8.6 (snit 2). snit 1 runs (copy of snit.test switched to snit 1) have the same pre-existing failures as before (27 on 9.1, 3 on 8.6). The new test fails without the fix. clock was already a reserved word, so macros are unaffected.
ncgi::value as the 3-character byte string …REPRODUCED (9.1, 8.6) for multipart/form-data text fields: a field sent as UTF-8 "йq" comes back from ncgi::value as the 3-character byte string "йq". Percent-encoded urlencoded data is fine (decode handles UTF-8).
-translation binary. That is required, because CONTENT_LENGTH counts bytes and file uploads must stay binary. But ncgi::multipart never converts non-file parts with the part's charset (or UTF-8).
encoding convertfrom utf-8 themselves; possible options are an opt-in flag or decoding only parts without a filename)
REPRODUCED (9.1, 8.6)
ip::reduceToAggregates {1.2.0.4/255.255.0.255 1.2.3.0/255.255.255.0} gives 1.2.0.4/16 on 9.1 and 8.6. Also found: a single-element list fails ("cannot use non-numeric string "" as left operand of "&"").
ip::collapse also turns the same input into 1.2.0.0/16. A non-contiguous mask cannot be expressed there; consider rejecting such masks.
ftp::OpenREPRODUCED (9.1, 8.6), partly: the leak is real, but only for failed ftp::Open
ftp::Open (connection refused, or the server closes before the greeting) returns -1 but leaves the ::ftp::ftpN state array behind, one per attempt. The reporter's script retries Open in a loop and never cleans up after a failure, so it leaks exactly this way.
::ftp::Open. When OpenControlConn fails, only ftp(State) is unset. When the login fails, Close is called, but it returns early ("Not connected!") because ftp(State) is already gone. Neither path unsets the array or closes the control socket.
REPRODUCED (9.1, 8.6)
ftp::Get $h nosuchfile -variable v against a local tcllib ftpd, in passive and active mode. It returns "", but the manual says 0. Also found: in active mode, against a server that opens the data connection right after PORT (ftpd, pyftpdlib), an error reply to RETR (e.g. 550) hangs ftp::Get forever. The control timeout fires inside a nested vwait on state.data and cannot release it.
::ftp::Get has an inline special case that returns "" and skips CloseDataConn. In the get_sent/reget_sent error branches, state.data is unset only in passive mode, so WaitComplete waits for a data connection that the server never closes. In active mode the data socket was also never recorded, so it could not be closed.
Get -variable returned "" instead of 0 and did not close the data connection. In active mode a refused RETR could even hang. Proposed fix with tests in the attached prototype (branch to be created).
Status now: Open
pop3d-dbox-skip-dirs, not merged: Fix 1749395fff: pop3d::dbox counted subdirectories of a mailbox, e.g. CVS, as messages. Glob only files.REPRODUCED (9.1, 8.6)
CVS/ subdirectory. STAT reports 2 messages (the directory counts, with its 4096-byte size), and RETR of that "message" fails.
glob -nocomplain [file join $dir *]. AK's 2007 objection (Tcl 8.3 lacks -types/-directory) no longer applies, since tcllib needs 8.5+.
glob -nocomplain -types f -directory $dir *. The test fails without the fix, and the pop3d suite passes on 9.1 and 8.6.
glob -types f -directory; it has a test.
statTime($name) still grows by one element per call, so memory grows without bound in long runs.REPRODUCED (9.1, 8.6): statTime($name) still grows by one element per call, so memory grows without bound in long runs.
foo called 100 times. The total list length in the profiler arrays grows by 100.
lappend ::profiler::statTime($name) $t, and stats recomputes from all samples at print/dump time.
statSumSq($name) and adds StatsFromSums N sum sumsq, which reproduces the existing stats result exactly, including the integer mean, using sum((v-m)^2) = S - 2mT + Nm^2. The public stats {args} proc is unchanged. Tests: profiler-12.0 (state does not grow) and profiler-12.1 (same result as stats). Both fail without the fix; the suite passes on 9.1 and 8.6. The attached 2007 patch changes the numbers (float mean) and the stats signature, so it was not used. The open branch pbrooks-2026-11-windows-profiler-fixes only touches tZero/tMark, so there is no overlap.
REPRODUCED (9.1, 8.6)
CR CR LF. Because of that, TOP n k never finds the header/body separator and sends the whole message.
mime::buildmessage (mime 1.7.2, memchan with -translation crlf) already returns CRLF lines, and pop3d writes them to a socket configured with -translation crlf. The TOP branch splits on \n and looks for an empty line, but the separator is \r. Two more TOP problems turn up once the separator is found: incr limit -1 makes TOP n k send only k-1 body lines, and H_top treats TOP n 0 as "whole message" while RFC 1939 says headers only.
TOP n 0 return headers only. The new tests fail without the fix (3 failures) and the pop3d suite passes on 9.1 and 8.6 (155/155). The attached 2007 patch (reversed diff) instead switched the socket to lf and added \r by hand everywhere. The prototype is less invasive.
REPRODUCED (still inconsistent)
REPRODUCED (9.1, 8.6)
<script>...i<window.document...</script> on 9.1 and 8.6: the parser reports a bogus tag window.document.forms.length;.
parse finds tags with one regsub over the whole document. The raw text of <script>/<style> is not treated specially.
PrepareHtml converts </> inside script/style to </>, the same way {, } and \ are already turned into entities in all text; test htmlparse-12.0). The test fails without the fix. The 59 tests pass on 9.1 and 8.6. Limitation: in incremental (-incvar) mode, a script element split across chunks is not handled.
</> in them are delivered as entities, like the existing brace escapes.
REPRODUCED (9.1, 8.6)
record delete instance $obj.sub1.reserved1, which names a field, not an instance. It silently does nothing. record delete instance junk and record delete record junk also return without error. struct_record.man says both "will return an error if the ... does not exist".
Delete has #error "$item_ is not a instance" commented out, and the record branch has no existence check. Fields cannot be deleted at all, by design. Separately, deleting a nested instance ($obj.sub1) directly leaves the parent broken (show values then fails); the prototype does not change that.
record delete.
$objTest.sub1.reserved1), not an instance. Record fields cannot be deleted, but the call should have raised an error, as documented, instead of silently doing nothing. Proposed fix: raise an error for unknown instances and records (patch with tests).
REPRODUCED (9.1 with the 34a589f660 fix, 8.6)
STOR newfile.txt gives "550 Permission denied". MKD newdir and RMD dir are always denied too.
::ftpd::fileAuth. Fs exists is false for directories, so the "check the parent directory" branch for new paths never runs, and the final Fs exists $path check denies everything that is not an existing file.
fsCmd exists is false for directories. Proposed fix with tests in the prototype.
Status now: Open
comm-lost-conn-peer, not merged: Fix 4811399edc, 1491894fff: comm raised an error when a client disconnected after the "commIncoming race condition" warning. commLostConn now clears the penREPRODUCED (9.1, 8.6)
-hist 10 gives 0 2, expected -1 1 0 1. Negative integers are bucketed correctly, because in…REPRODUCED (9.1, 8.6): counting 9.1 and -9.2 into -hist 10 gives 0 2, expected -1 1 0 1. Negative integers are bucketed correctly, because integer division already floors.
::counter::count, branch -hist: int($delta / bucketsize) truncates toward zero for reals. The ticket's change to the HTML labels is not needed: with floor, label i*bucketsize is the correct lower bound of bucket i.
int(floor(...)).
counter::histHtmlDisplay fails with invalid command name "ncgi::value" unless the caller has already loaded ncgi and html.REPRODUCED (9.1, 8.6): counter::histHtmlDisplay fails with invalid command name "ncgi::value" unless the caller has already loaded ncgi and html.
::counter::histHtmlDisplayRow calls ncgi::value and html::font but never requires either package.
counter::init ::foo and counter::init a::b fail with "parent namespace doesn't exist". counter::init a(b) fails with "…REPRODUCED (9.1, 8.6): counter::init ::foo and counter::init a::b fail with "parent namespace doesn't exist". counter::init a(b) fails with "variable isn't array".
counter::T-$tag, H-$tag, …), so :: and a trailing (…) break them.
counter::init now rejects such names with a clear error, and counter.man documents the restriction.
:: and a trailing (x) fail with obscure errors. Proposed patch: reject such names in counter::init with a clear message and document the restriction.
counter::get foo -maxPerMinute on a -hist counter gives "can't read counter(maxPerMinute)". The options are still undocu…REPRODUCED (9.1, 8.6): counter::get foo -maxPerMinute on a -hist counter gives "can't read counter(maxPerMinute)". The options are still undocumented and missing from the "Invalid option" error message.
::counter::get reads counter(maxPer*) without checking the type. Only -timehist counters set those values.
REPRODUCED (9.1, 8.6)
ftp::Quote $h SITE {CHMOD 644} {my file} sends SITE {CHMOD 644} {my file} to the server.
::ftp::Quote uses the list $args as the command string.
join $args " ", as suggested in the ticket, and ftp.man says the arguments are joined with spaces.
REPRODUCED (9.1, 8.6)
Open -command) where the 226 reply arrives before the data EOF, then Close. About 1 s later: bgerror can't read "ftp(Wait)": no such variable. Before Close, the same stray timer fires a spurious timeout callback.
::ftp::CopyNext re-arms the Timeout timer even for the final chunk. In async mode nothing cancels it once the reply has already been handled. ::ftp::Timeout also assumes the connection still exists.
Status now: Open
pop3-eof-busy-loop, not merged: Fix 815405ffff: pop3::top spun forever at full CPU when the server closed the connection in the middle of a reply, and pop3 send failed with an empty message.REPRODUCED (9.1, 8.6), for the "100% CPU" part. The Windows 2000/Tcl 8.4.2 wish specifics are not reproducible here: 50 TOP calls under wish on Linux work.
pop3::top spin forever at ~75–100% CPU (pop3.test hangs) on 9.1 and 8.6. After a connection drop, pop3::send fails with an empty message ("POP3 STAT ERROR: "), which matches the reporter's "unknown (no error string) error out of pop3::send".
::pop3::RetrSlow loops on gets without checking for EOF, so after the server closes, gets returns "" forever. ::pop3::send also ignores EOF.
socket -backlog)REPRODUCED (9.1; 8.6 by code, the test needs socket -backlog)
-backlog 1 and a full queue, so SYNs are dropped and connects hang. ftp::Open ... -timeout 1 returns -1 only after 2m13s.
OpenControlConn uses a blocking socket connect, which ignores -timeout.
backlog so it is skipped on 8.6). In synchronous mode it connects with -async and waits for writable or the timeout. Async (-command) mode is left unchanged. Refused-connection, unknown-host and ftpd transfer checks still pass.
REPRODUCED (9.1, 8.6)
RR instead of RR , and drops the single space between </b> and <i>.
2treeCallback runs string trim on all text. (The set text [mapEscapes ...] next to it is dead code.)
</title>) now becomes PCDATA nodes. This is what the reporters of this ticket and 3091244 asked for, but a human must decide whether it is acceptable.
REPRODUCED (9.1, 8.6)
ftp::Delete with -timeout 1 returns 1, a false success. With the server not answering RETR, ftp::Get -variable throws can't read "ftp(Start_Time)" from ElapsedTime, which is the crash in the 2007 comment and in ticket 1709566.
::ftp::Timeout sets ftp(state.control) 1, so WaitOrTimeout reports success.
Close after the timeout no longer hangs. Real transfers via ftpd still pass (active and passive).
REPRODUCED (9.1, 8.6; still by design)
package require stooop, info commands ::_proc → ::_proc and info procs ::proc → ::proc.
rename proc _proc, then defines ::proc. stooop needs this so that proc Class::member {this ...} works transparently.
stooop::import / stooop::proc, and a compatibility decision)
REPRODUCED (8.6 for mime and many other modules; 9.1 and 8.6 for doctools2base). The argv0 part was fixed in 2003.
source, so on 8.6 the UTF-8 test files are read in the system encoding.
doctools::msgcat::init idx looks for the uk/pl catalogs.
source -encoding utf-8 in all.tcl, and msgcat::mclocale c in msgcat.test. The existing tests serve as the regression tests: mime 4.4/5.x/6.2/7.1, textutil, base64, yaml and doctools-msgcat-2.0 fail without the fix and pass with it under C/koi8u/iso8859-2 on 8.6 and 9.1.
tclsh mime.test).
REPRODUCED (9.1, 8.6)
-?, -help, -bogus and -f (missing argument) all give the same message and the same CMDLINE USAGE errorcode. The caller cannot tell help from a real error, and the bad option is not named. The same is true for getKnownOptions and typedGetoptions.
result(?) on any getopt error. typedGetoptions throws the plain usage error. The specific message ("Illegal option ...") is discarded.
CMDLINE USAGE, and invalid input gets CMDLINE USAGE INVALID <specific message>. trap {CMDLINE USAGE} still catches both. All tests pass on 9.1 and 8.6 (139 passed, 9 skipped). The 3 new tests fail without the fix. Prefixing the message itself with the problem would break 8 existing tests, so it was not done.
INVALID <message> to the errorcode for real errors.
Status now: Open
ftp-hpupdate-tk9, not merged: Fix 450914ffff: the ftp hpupdate example failed at startup on Tk 9: it exposed the private tkButtonInvoke command, which is gone. Use a small helper that flasREPRODUCED (Tk 9.0, 9.1)
invalid command name "::tk::unsupported::ExposePrivateCommand" (hpupdate.tcl line 36). It still works on 8.6.
tkButtonInvoke via ::tk::unsupported::ExposePrivateCommand, and Tk 9 removed that command.
ButtonInvoke proc (flash + invoke, skipped when disabled) used by the 6 key bindings. No regression test: examples/ has no test suite. Checked by hand on Xvfb with 9.1 and 8.6: the app starts, the bindings use the new proc, and invoking it fires the button command.
::tk::unsupported::ExposePrivateCommand is gone). The patch replaces tkButtonInvoke with a small public-API helper (flash + invoke).
package require Tcl 8.5 9+ (invalid version syntax, added in 0b3dd048f8 2025-07-16) was fixed on trunk in [f14416cf2…Status now: Open
pkgindex-version-fixes, not merged: Fix 61e49798ee, 523f7e0556, cdbdd399e4: pkgIndex.tcl entries that do not match the code. nettool's index sourced nexttool.tcl, so package require nettool PARTLY-FIXED. The reported package require Tcl 8.5 9+ (invalid version syntax, added in 0b3dd048f8 2025-07-16) was fixed on trunk in f14416cf2a (2025-08-19). Tcllib 2.0 itself shipped rest 1.7 with "8.5 9" and was never affected; the reporter had a trunk snapshot. On trunk package require rest still fails on all Tcl versions: "attempt to provide package rest 1.7 failed: package rest 1.8 provided instead". pkgIndex.tcl and rest.man were not bumped along with rest.tcl. sak validate versions rest flags it.
package require rest fail; proposed fix syncs them.
PARTLY-FIXED (692adc1efb, 2024-09-17, in tcllib 2.0)
sha256c before tcllibc (692adc1efb). If other code loads tcllibc first, the tcllibc shared library has already run package provide sha256c, so the sha256c.tcl critcl definitions (critcl::cheaders) never run. The reporter's last comment says the package forget sha256c workaround then fails with "does not support stubs". aku explains why the provide call in tcllibc cannot simply be dropped.
PARTLY-FIXED: on trunk with Tcl 9.1 on Windows, the fileutil, fumagic, math, mime, report, tar, textutil, virtchannel_base and yaml failures of the report are gone. Still failing on Windows only: httpd (4), irc picoirc (20), pop3d (1), profiler (1) and uri (1). crc (7) and hook (2) fail on 9.1 on Linux too.
sak test run of trunk with tclsh91 on the Windows VM (websocket.test left out because it contacts an external server).
\x0D; the test channel keeps the CR of the CRLF translation on Windows.
uri::geturl file://C:/... drops the drive letter (:/Users/...).
tr\u0413\xA9s for très. The prototype for 509776ffff (source test files as UTF-8) is the fix. fumagic.filetype-1.25.1/1.27 recognise the PE test file only as MS-DOS on Windows.
file:// with drive letters) and httpd CGI/proxy on Windows. Merge the profiler branch.
PARTLY-FIXED (the actual bug, a stray "]", was fixed in 42266acec4 on 2023-08-16, in tcllib 2.0. The rest is a refactor of error handling and whitespace for Jim Tcl, not a bug.)
encode "(" error is fixed (7804eb94e9 2023-10-06, merged 758418619e 2023-11-14, tcllib 2.0; test ncgi-4.5). Sti…PARTLY-FIXED. The Tcl 9 / TIP 465 encode "(" error is fixed (7804eb94e9 2023-10-06, merged 758418619e 2023-11-14, tcllib 2.0; test ncgi-4.5). Still broken on 9.1 and 8.6: any character above U+00FF raises "can't read map(й): no such element in array", U+0080–U+00FF are encoded as Latin-1 (ö gives %F6), U+0100 gives the bogus %100, and NUL is not encoded.
ncgi::encode with й, 中, U+1F600, ö, NUL, on 9.1 and 8.6
map, which is built for code points 1..256
Status now: Pending (Fixed)
PARTLY-FIXED (fixed with http 2.10 in Tcl 9.0/9.1; still sent twice with http 2.9.9 in Tcl 8.6)
websocket::open. 9.1.1 and 9.0.5 (http 2.10.3) send one header, Connection: upgrade, close. 8.6.18 (http 2.9.9) sends Connection: close and Connection: Upgrade.
geturl forces -keepalive 0 for upgrade requests and writes its own Connection: close line, followed by the user-supplied Connection: Upgrade header. websocket cannot avoid this from tcllib.
nettool::ip_list works on Windows 9.1 (returns the VM's two IPv4 addresses). package require nettool itself fails …PARTLY-FIXED: without twapi, nettool::ip_list works on Windows 9.1 (returns the VM's two IPv4 addresses). package require nettool itself fails (pkgIndex typo, 61e49798ee), so nettool.tcl was sourced directly. The twapi variants, which call the undefined ldelete, could not be tested (twapi not installed).
::twapi::get_ip_addresses, and if it is missing, falls back to ::twapi::get_system_ipaddrs (-version 4 or -ipversion 4). This is in the tcllib-1-19 tag and not in tcllib-1-18. NEEDS-WINDOWS to run it.
ldelete result 127.0.0.1, and nettool never defines ldelete. On Linux ip_list fails with invalid command name "ldelete", which I reproduced. The twapi variants should fail the same way.
ldelete definition is part of prototypes/9a21637273.patch (test ip-list-ldelete).
ldelete is undefined. That is fixed together with 9a21637273.
nettool::ip_list works on Windows 9.1 (returns the VM's two IPv4 addresses). package require nettool itself fails (pkgIndex typo, 61e49798ee), so nettool.tcl was sourced directly. The twapi variants, which call the undefined ldelete, could not be tested (twapi not installed).
PARTLY-FIXED. Fixed: try/throw (4e8a0be16c) and tcldes/tclDESjr (7f225e0906) in 2016. oometa and udpcluster (formerly nameserv::cluster) now have pages. tool.man now says "TclOO Library (TOOL) Framework", and "Dictionary Tools" is the legit tool::dict_ensemble page. Still open: processman titledesc is still the cron text, and modules/tool/meta.man is still a stale copy of ooutil.man (oo::util 1.2.2 vs 1.2.3), so the TOC lists oo::util twice.
websocket::validate …PARTLY-FIXED (Connection check added in aa43bc825c, tcllib 1.21; the Upgrade check is still missing). REPRODUCED (9.1, 8.6): websocket::validate accepts a request that has no Upgrade header at all. An Upgrade header without websocket is rejected.
websocket::validate {Connection Upgrade Sec-WebSocket-Version 13 Sec-WebSocket-Key ...} returns a dict instead of raising an error.
::websocket::validate: $websocket is checked only inside the upgrade branch of the header loop, never after the loop.
logger::appender::fileAppend -outputChannel, d01676c313, 2005-11-14, but it is undocumented…PARTLY-FIXED (writing to a file already works through logger::appender::fileAppend -outputChannel, d01676c313, 2005-11-14, but it is undocumented. Rotation is not implemented)
logger::utils::applyAppender -appender fileAppend -serviceCmd $log -appenderArgs [list -outputChannel $f] writes [date] [foo] [global] [warn] to file to the file on 9.1 and 8.6. fileAppend is missing from loggerAppender.man.
-outputChannel $chan), which needs documenting. The attached patch is reversed and has no rotation; rotation can be done by reopening the channel.
PARTLY-FIXED (9.1, 8.6)
couldn't open socket: .... A closed local port gives error reading "sockNN": transport endpoint is not connected (9.1) or socket is not connected (8.6). A host that accepts and closes gives "premature end-of-file from server" (accurate).
socket -async connect, gets reports ENOTCONN. The real error ("connection refused") is only in fconfigure -error, which is never queried. On 8.5, gets even returns -1 without eof, so the connect wait runs until the timeout.
fconfigure -error first in readable. mime tests: 104/104 on 9.1 and 8.6. smtp-2.1 fails without the fix.
PARTLY-FIXED (.CS/.CE for examples and whitespace compaction were added in 2010, ChangeLog [Bug 3058654]). REPRODUCED (9.1, 8.6): a multi-line [titledesc]/[moddesc] still gives a multi-line NAME line and a broken .TH line. The NAME-lists-commands part is still an open RFE.
[titledesc {A long\n multi-line description}] to nroff on 9.1/8.6. NAME is spread over 2 lines, and the moddesc in .TH is cut at the newline.
[ \t]+, not newlines. Also, the loop variable text stays upvar-linked to the last variable, so later set text [c_held precomments] overwrites it. That is harmless today (copyright is used before), but a trap when extending the loop.
PARTLY-FIXED (the missing upvar was fixed in 2010, ChangeLog 2010-01-20, tcllib 1.13). Renaming is still broken on 9.1 and 8.6.
::ftpd::fsFile::fs rename with a docRoot that is not the process cwd. The target name is not joined with docRoot, so the file is renamed relative to the server's cwd, or the call fails. When the target directory is missing it replies "550" and then still tries the rename, which errors, and RNTO sends a second reply "500". RNTO to a new name was also denied by the default fileAuth (keithv's comment), which is 1508656fff.
::ftpd::fsFile::fs rename: $newname lacks file join $docRoot, and there is no return after the 550.
-command mode still affected on 9.1, 8.6)PARTLY-FIXED (sync mode fixed since tcllib 1.17 / ftp 2.4.13; async -command mode still affected on 9.1, 8.6)
Get -variable returns the full 8000 bytes (WaitComplete waits for the data EOF, and WaitDataConn 6474b535bb, tkt eb0b15d598, 2014-02-21 covers late data connections). In async mode (Open -command), the get callback is called with 0 bytes and the variable stays empty, because the data arrives after completion. With normal ordering (data, then 226) async works.
StateHandler get_close. In async mode the state machine completes on 226 without waiting for the data connection EOF (WaitComplete only waits when ftp(Command) is empty).
get_close, list_close and reget_close. That is more than a local fix.
ftp::Open -command: if the server sends 226 before the data, the get callback receives empty data.
PARTLY-FIXED. decode handles 2- and 3-byte UTF-8 since ncgi 1.4/1.4.1 (a6430d4122 2012-05, b6632970f6 2013-01; tcllib 1.15). The remaining gaps are 4-byte sequences in decode, and encode, which fails on any character above U+00FF ("can't read map(й)") and encodes U+0080–U+00FF as Latin-1.
ncgi::encode й gives an error on 9.1 and 8.6. ncgi::decode %F0%A0%AE%B7 returns 4 Latin-1 characters.
map array, which only covers characters 1–256. decode has no 4-byte regsub.
package require dtplite 1.0.5, so an older installed …PARTLY-FIXED (9.1, 8.6). The "picks up an older tcllib" half is gone: apps/dtplite now does package require dtplite 1.0.5, so an older installed tcllib cannot satisfy it. The other half is reproduced: after installer.tcl -pkg-path <dir not on auto_path> the installed dtplite fails with "can't find package dtplite 1.0.5". The same applies to all the apps (mkdoc, pt, page, nns...).
ainstall copies apps verbatim and does not record the package install path. AK's attached auto_path snippet was never applied.
lappend auto_path <pkg-path> to the apps it installs, but AK pointed out it conflicts with relocatable/batteries-included distributions, so the maintainer must decide. Setting TCLLIBPATH, the standard Tcl way to find extensions in a non-default location, works.
PARTLY-FIXED (asn 0.8, 2007): asnPeekTag/asnTag/asnRetag handle multi-byte and non-universal tags; asnChoice/asnChoiceConstr exist for encoding. Status of the other points:
string range is enough.
PARTLY-FIXED (snit 2.x accepts a command prefix; snit 1.x does not)
option -text -configuremethod [list SetText 1]. snit 2.3.4 works on 9.1/8.6 ({*} expansion). snit 1.4.3 fails with "::t1 SetText 1" is not defined.
{*}. snit.man still documents methodName only.
PARTLY-FIXED (9.1, 8.6)
sendmessage with tcl_platform(user) set to "" and no From, and with -header {From ""}, on 9.1 and 8.6.
array set aprops {error "invalid address \"$from\""} was added at some point, so it no longer fails with "can't read aprops(error)". But the default is in braces, so the message shows a literal $from: error in -originator: invalid address "$from".
[list ...] and the right variable. mime module tests: 104/104 pass on 9.1 and 8.6. The new tests fail without the fix. All four smtp prototypes create smtp.test with the same preamble. Applied together, merge them into one file.
PARTLY-FIXED (fixed for the default TclOO implementation; still REPRODUCED with stack_tcl.tcl, used when TclOO is absent, e.g. tclsh8.5, and in the critcl implementation)
::oo::ObjNN, no collision. 8.5 (pure Tcl impl): with proc stack2 {} defined, struct::stack fails with command "::stack2" already exists. The "always global namespace" part was fixed long ago: names are now created in the caller's namespace.
stack_tcl generates stack$counter once and errors on a clash. stack_c.tcl SDnewName does the same. queue_tcl/queue_c, tree_c and graph_tcl use the same pattern.
stack_tcl). Fix: skip generated names that already exist. The new test runs only for the pure Tcl impl (tclsh8.5). It fails there without the fix and passes with it. On 9.1/8.6 the struct suite is unchanged (test skipped). On 8.5 the pre-existing failure set-tcl-1.0 is unrelated. The critcl impl was not changed or tested (no tcllibc build).
stackN already exists. Proposed fix for the Tcl fallback: skip existing names (patch with test).
unset ftp(state.data) error was fixed in ftp 2.4.5, tcllib 1.10 (2006-10-20, bugs 1234831/1581…PARTLY-FIXED (9.1, 8.6). The reported passive-mode unset ftp(state.data) error was fixed in ftp 2.4.5, tcllib 1.10 (2006-10-20, bugs 1234831/1581453). Active mode still hangs.
-timeout cannot break it. A leaked data socket also raises a background error (can't read "ftp(Start_Time)") after Close.
*_sent error branches, ftp(state.data) is only unset in passive mode. In active mode WaitComplete sits in a nested vwait ftp(state.data) that nothing ends. Also, inline Get/Put never record the data socket, so CloseDataConn cannot close it.
ftp(DataSock), and makes inline Get call CloseDataConn on failure.
HostAsDriveLetter 0. With that default the W…PARTLY-FIXED. Since uri 1.2.7 (tcllib 1.19, kjnash 2016-12/2017-01, e.g. 0b8557b27d), the default is HostAsDriveLetter 0. With that default the Windows-only file://localhost:/a/b/c join is no longer used, so the canonicalize tests no longer depend on the platform. uri::join scheme file path {c:\windows} still gives file:///c:\windows on trunk. pbrooks is working on a fix in branch pbrooks-ticket-664392 (2026-10-31 .. 2026-11-13, not merged): it converts backslashes on Windows in join, and makes geturl percent-decode the path and convert drive-letter paths to native filenames. That branch bumps uri to 1.2.9.
file://localhost/a/b/c round-trips. Read the branch diff.
::uri::JoinFile passes backslashes through.