Triage of 818 open and Pending Bug tickets of Tcl with no comment or check-in from me on 2026-10-16 (in Tcl's workflow Pending means "resolved, waiting for confirmation"). Reproduced on Linux (main 9.1 tip, 9.0 tip, 8.7b1, 8.6 tip, 8.5, 8.4 and old releases), on FreeBSD, NetBSD, OpenBSD and OpenIndiana VMs for platform tickets, and on a Windows 11 VM. Rows marked patch link to a prototype patch (287 in all, in tcl-triage/prototypes). Updated 2026-11-11: the emoji before each ticket id shows its progress.
e5c21e566f.patch: NOT-REPRODUCED-ON-TIP (the reported SMB case cannot be triggered here; the patch builds and is harmless)
e5c.tcl gives empty lists for a missing local dir, //localhost/C$/<missing>, a missing share and a deep missing path, both as admin and as basic user. New with the basic-user token: a dir with a DENY(RD) ACE gives couldn't read directory ".../priv/*.ext": permission denied (EACCES) for priv, priv/. and priv/sub, local and via \\localhost\C$. That is consistent on Windows, unlike the Unix stat/opendir difference; as admin it is listed anyway
NEEDS-WINDOWS (unresolved: the SMB ENOENT error does not occur against a Windows 11 SMB server; needs the reporter's server type); Windows EACCES case not testable as the elevated ssh admin (a DENY list-directory ACE does not stop listing)
priv mode 000: glob -nocomplain -directory priv * → "couldn't read directory "priv/": permission denied"; -directory priv/. and priv/sub → empty, no error (main, 9.0.5, 8.6.18, 8.5.0, 8.6.0); 8.4.20 → empty for all.
glob -nocomplain -directory <d> *.ext returns an empty list without error for a missing local dir, //localhost/C$/<missing>, //localhost/C$/no/such/deep/path, //localhost/nosuchshare/x, and a dir with a DENY(RD) ACE (listed anyway). Related: for paths longer than MAX_PATH 9.x glob silently returns nothing (see 3eace1b21f).
NEEDS-WINDOWS (unresolved: Offline Files needs the CSC service and an offline-available share, i.e. system configuration changes not allowed on this VM; CscService is stopped/manual)
<file-on-offline-share> ?<dir>? (needs a share made "Always available offline", synced, then Work offline).
NEEDS-WINDOWS (unresolved: the VM has no serial port — all four VirtualBox UARTs are off and no com0com is installed; adding one needs a VM configuration change while the user is running it)
?COMn? — needs a COM port with no peer (VirtualBox serial port in "Disconnected"/host pipe mode with no reader, or com0com with the other side closed). It runs the write+close in a child with -handshake rtscts, blocking 0 and 1, and reports HANG / REOPEN-FAILED; prints a VERDICT line and exits by itself.
NEEDS-WINDOWS (by code reading, still present in main, 9.0, 8.6)
set env(TZ) :UTC; clock format 0; unset env(TZ).
Status now: Closed (Fixed)
unix-locale-init-lock, in main: Merge trunkTcl_Init being part of it) as performant as possible. Part of that is minimizing calls as well as synchronization. To that end, is there a reason for retrieving the locale from the environment on a per interp basis and not once per process (or per thread)? If not, could this be subsumed into Tcl_InitSubsystems? Thoughts? …CANNOT-TEST (OpenBSD) — does not reproduce on Linux/glibc (expected); Serhiy reproduced it on OpenBSD 7.9 amd64 (see comments). The original Tcl_Init() trigger is gone on trunk/9.0 since the fix for 63941fee4f (00b1026efe trunk, 27f849b6b3 9.0, 2026-10-17, unreleased); [encoding user] / Tcl_GetEncodingNameForUser() still call setlocale() unlocked
tclsh c35e347c6f-encoding-user-race.tcl 10x each (main should crash with a double free in free(), the branch should not)
CANNOT-TEST (macOS). This looks like the user's environment, plus a side bug in tcltk-man2html that reproduces on Linux.
html-tk to make install. BUILD_HTML runs $(BUILD_TCLSH) = ${TCL_BIN_DIR}/tclsh. With --with-tcl=/usr/local/tcl8.6.16/lib (an install dir) no tclsh exists there, so it falls back to TCL_EXE (SC_PROG_TCLSH = the first tclsh* on $PATH). On this machine that is an old 8.6.6 tclsh, with no TCL_LIBRARY set, and it finds /usr/local/lib/tcl8.6/init.tcl from a newer install: "have 8.6.6, need exactly 8.6.16". Tk itself is installed; only the HTML docs step is affected. Side bug: tools/tcltk-man2html.tcl takes the *newest* tk* sibling of --srcdir (lsort ... end in 8.6, findversion in 9.x). That is why the 8.6.16 install says "using Tk source directory tk9.0.2" and builds 9.0.2 HTML docs. Reproduced on Linux with sibling tk8.6.18 and tk9.0.5 trees: both the 8.6 and the 9.0 scripts pick tk9.0.5.
--useversion (or the script to prefer the srcdir it is run from), which is a Tk/Tcl Makefile change to decide on.
make install TCL_EXE=/usr/local/tcl8.6.16/bin/tclsh8.6 (or point --with-tcl at the Tcl build dir). Separately, tcltk-man2html.tcl picks the newest tk* directory next to the source tree, which is why the tk9.0.2 docs were built.
CANNOT-TEST (macOS). Same cause as 15d60c0d2e (duplicate).
make install runs html-tk through TCL_EXE = the first tclsh on PATH (an old 8.6.6), because --with-tcl=/usr/local/tcl9.0.2/lib has no build tclsh. Not a Tk build failure as such.
make install TCL_EXE=/usr/local/tcl9.0.2/bin/tclsh9.0, or configure Tk with --with-tcl pointing at the Tcl build directory. Suggest closing as a duplicate of 15d60c0d2e.
CANNOT-TEST (needs the external test server test-tcl-http.kerlin.org, constraint serverNeeded; no outside network allowed)
CANNOT-TEST (macOS)
${CONFIGURE_ARGS} ${EXTRA_CONFIGURE_ARGS}; options given to a separate macosx/configure run are not used by it (the GNUmakefile is meant to be driven via make EXTRA_CONFIGURE_ARGS=..., see macosx/README). The "unrecognized options framework/dtrace" warnings in sub-makefiles come from unix/configure.ac forwarding all of its arguments to bundled packages via PKG_CFG_ARGS (same mechanism as 17ffc49790) - harmless warnings.
make EXTRA_CONFIGURE_ARGS="..." rather than from a separate macosx/configure run; the framework/dtrace warnings come from passing Tcl's configure arguments on to the bundled packages and are harmless. Would documenting EXTRA_CONFIGURE_ARGS more prominently be enough?
CANNOT-TEST (no network; the website source is not in any local repository)
bbe7c6ff9e.patch: VERIFIED (mechanism only, under emulation; macOS root cause unconfirmed)
make test TESTFLAGS="-file zipfs.test -match *password*" -> 10 failed, all invalid password. Without the shim the tip passes on Linux. tests/zipfs.test on the tip still disables these tests on Darwin.
make test TESTFLAGS="-file zipfs.test": 535 total, 498 passed, 37 skipped, 0 failed; builds without warnings
CANNOT-TEST (macOS) — still present: tests/zipfs.test on main still skips the password tests on Darwin (constraint bbe7c6ff9e)
-I$(ZLIB_DIR)), but on macOS it is linked with the system libz; the traditional PKWARE decryption keys use crc32tab = get_crc_table(), and only the password paths (init_keys/zdecode) use that table, which matches exactly the failing tests (wrong keys → garbage → "invalid CRC"/"decompression error"). Apple's libz need not return a standard 256-entry table there.
crc_table from the bundled crc32.h that tclZipfs.c already includes instead of get_crc_table(), and re-enable the tests on Darwin so a CI run on macOS (core-* tag) validates it.
Status now: Open
double-checked-lock-ordering, not merged: Fix 112af76b09: use an acquire load for the unlocked check and a release store for publishing in the lazy initialization of mutexes and condition variables (TCANNOT-TEST (memory-ordering race, not observable on x86-64; confirmed by code reading: Tcl_MutexLock()/Tcl_ConditionWait2() on main read the lazily created pointer without a lock and without acquire ordering)
CANNOT-TEST (macOS 13 on ARM)
_dyld_fork_child atfork handler in the child after fork() in TclpCreateProcess, i.e. in the OS/expect environment, not in Tcl code. Since 8.6.14 (42d3f0cdb9, 2023-07-26, fix for c54e4a1aeb) TclpCreateProcess prefers vfork() when available, which macOS has, so the fork()+atfork path from the backtrace is no longer taken by Tcl's exec.
./configure && make shell on macOS without macher; unix/Makefile.in still appends the zip to the dyli…CANNOT-TEST (macOS) — needs a non-framework ./configure && make shell on macOS without macher; unix/Makefile.in still appends the zip to the dylib with cat when macher is not found, so it is probably still the case (code reading on main 2026-10-22)
CANNOT-TEST (8.5 only — 8.5 is no longer maintained; needs a host with a noexec temp dir, which requires root to set up)
CANNOT-TEST (AIX/xlc)
pc, gcc/clang/MSVC are fine. The reporter's workaround (disable the peephole for _AIX) points to an xlc optimizer/codegen bug rather than a Tcl bug; the reported CFLAGS was -g only, so it may also be the computed-goto-free switch in xlc.
#if !defined(TCL_COMPILE_DEBUG) && !defined(_AIX) is the obvious candidate if someone with AIX confirms on 8.6.18/9.0)
!defined(_AIX) guard (or rather an xlc version check) is the right fix.
7cd721e9ea.patch: VERIFIED
CFLAGS="-g -Dmkdtemp=no_such_mkdtemp" -> linking tclsh fails, undefined reference to «no_such_mkdtemp» (the tip's configure.ac has no mkdtemp check)
~/tcltk/autoconf-2.73/bin/autoconf -f regenerates a configure byte-identical to the patched one
ac_cv_func_mkdtemp=no (work-a2/v2/nomk2): builds, compat/mkdtemp.c compiled without warnings, libtcl exports no_such_mkdtemp; file tempdir -> /tmp/tcl_XXXXXX mode 0o40700; file tempdir <dir>/tdx works; error path can't create temporary directory: no such file or directory; cmdAH + fCmd tests: 17306 total, 17231 passed, 75 skipped, 0 failed. Normal in-tree build: libtcl uses glibc mkdtemp@GLIBC_2.2.5 as before
CANNOT-TEST (Solaris 8; illumos has mkdtemp() in stdlib.h). Prototype verified on illumos with the fallback forced.
file tempdir works (dir created, mode 0700, error path OK), cmdAH.test passes (17001/0 failed). Normal build still uses glibc's mkdtemp.
ac_cv_func_mkdtemp=no: compat/mkdtemp.c is compiled without warnings and libtcl defines mkdtemp. In both builds file tempdir creates /var/tmp/tcl_XXXXXX with mode 0700, file tempdir /tmp/tdx works, the error path says "can't create temporary directory: no such file or directory", and cmdAH.test passes (17001 total, 0 failed).
mkdtemp symbol from libtcl.
ac_cv_func_mkdtemp=no): file tempdir and cmdAH.test work. Solaris 8 itself was not tested.
CANNOT-TEST (no network; website source not available locally)
CANNOT-TEST (needs a serial port forwarded over RDP to Windows)
Status now: Closed (Fixed)
fCmd-test-constraints, in main: Rebase fCmd-test-constraints to main. Fix [219158], 300724b80c and 852a17eeb9: fCmd.test constraints (read-only directory rename, the lost notWine definitiofCmd-test-constraints, in main: Rebase Fix [219158], 300724b80c and 852a17eeb9: fCmd.test constraitstrunk, in main, 9.0: Fix [219158], 300724b80c and 852a17eeb9: fCmd.test constraitscore-9-0-branch, in 9.0: Fix [219158], 300724b80c and 852a17eeb9: fCmd.test constraitsCANNOT-TEST (the VM's page file is at the default location, so the failure does not occur; reproducing needs the page file moved off C:\ or disabled, a system setting I was told not to change)
pagefile; conflicts with current trunk in tests/fCmd.test (see 219158ffff)
\??\C:\pagefile.sys. win/852a17eeb9.tcl with tclsh91: "fCmd-30.3 PASSES here"; fCmd.test -match fCmd-30.3* on 9.1: Passed 1, Failed 0. (The 9.0/8.6 uninstalled builds could not load the registry package from library/, so the script ran on 9.1 only; the test assumption is the same everywhere.)
CANNOT-TEST (no tdbcmysql sources/repository and no MySQL server available offline)
CANNOT-TEST (macOS)
-error after the async connect to an IPv4-only server via "localhost", and no writable event on refused connect) were seen on High Sierra with 8.7 at 84de05583b.
localhost_v4 constraints.
make test TESTFLAGS="-file socket.test -match socket-14.*" still fail on a current macOS with 9.0/main? If not, this can be closed.
Status now: Pending (Works For Me)
CANNOT-TEST (no 32-bit toolchain on the VM: C:\msys64\mingw32 contains only empty bin/include/lib skeleton dirs, no gcc/g++; the only g++ is ucrt64 x86_64 16.2.0, and no 32-bit tclsh is built either; installing mingw-w64-i686-gcc was not allowed)
CXX=i686-w64-mingw32-g++ ./build.sh) and run with a 32-bit tclsh; also cppext64.dll from mingw64 for 64-bit. Three variants (load without Init, with TCL_FINALIZE_ON_EXIT=1, with an init name), prints VERDICT.
pacman -S mingw-w64-i686-gcc in MSYS2 plus a 32-bit Tcl build, e.g. nmake MACHINE=IX86 or the mingw32 configure build).
CANNOT-TEST (IRIX/MIPSPro)
1 -NaN* on main, 9.0 and 8.6; x86-64 passesbf029ced86.patch: VERIFIED (on x86-64 and by pattern check; the MIPS/RISC-V run is still CANNOT-TEST)
binary scan \xFF\xFF\xFF\xFF f1 a -> -NaN(7ffffe0000000), sign kept); no qemu-riscv64/qemu-mips here. Checked the patterns instead (work-a4/v4nan.tcl): the old glob 1 -NaN* rejects 1 NaN(7fffffe0000000) and 1 NaN, the reported MIPS/RISC-V results
^1 -?NaN accepts 1 -NaN(...), 1 NaN(...), 1 NaN and still rejects 1 1.0; binary.test 754 total, 673 passed, 81 skipped, 0 failed on x86-64
CANNOT-TEST (MIPS/RISC-V hardware; no qemu-riscv64 here) — test still expects 1 -NaN* on main, 9.0 and 8.6; x86-64 passes
Tcl_NewDoubleObj(fvalue); the float→double widening may drop the sign of a NaN (RISC-V fcvt.d.s always gives the canonical +NaN), and IEEE 754 leaves that sign unspecified.
-match regexp -result {^1 -?NaN} plus a comment); binary.test passes on x86-64. (A code alternative, building the double bits by hand for NaN floats, would also keep the sign and payload, but seems unnecessary.)
CANNOT-TEST (needs trofs, not available offline). Core behaviour unchanged.
Status now: Open
oo-define-list-overflow, not merged: Fix 40f628e8e3 in oo::define: check the result of Tcl_ListObjReplace() in MagicDefinitionInvoke(), and pass the real length of the list to Tcl_EvalObjv().CANNOT-TEST (the remaining oo::define case needs objc > LIST_MAX: unreachable on 64-bit 9.x; on 8.6 it needs two expanded lists of 268M elements, ~2 GB each plus a 4+ GB eval stack, not run on the shared machine). The 2015 linsert/concat cases were fixed by dgp in 2015. Code confirmed unchanged: "TODO: overflow?" unchecked Tcl_ListObjReplace() in MagicDefinitionInvoke() on main, 9.0 and 8.6.
oo::define C {*}[lrepeat 50000000 x] under ulimit -v): allocation failures panic ("list creation failed: unable to alloc"), before the unchecked error path can be reached, so no crash repro. Branch tip: oo.test, ooNext2.test, ooProp.test pass; oo62.tcl with 1000 args OK.
/* TODO: overflow? */ at a different place).
CANNOT-TEST (AIX/xlc not available)
make test TESTFLAGS="-verbose bpe" to find the test and attach a dbx backtrace; otherwise suggest closing.
CANNOT-TEST (macOS) — no normalization code in glob/filesystem paths on main (unix/, macosx/, tclFileName.c, tclIOUtil.c), so behaviour is unchanged
17a26ddf66.patch: VERIFIED
LD_PRELOAD=chgrp_shim.so ./tcltest ../tests/fCmd.test -match 'fCmd-27.[56]' -> both FAILED, could not set group for file "foo.tmp": operation not permitted
CANNOT-TEST (needs a file whose group cannot be re-set by the owner, e.g. inherited from a setgid dir of a foreign group on NFS/idmapped storage; on local Linux fs a no-op chgrp to the current gid is always allowed); tests unchanged on main, 9.0, 8.6
CANNOT-TEST (iOS SDK) — first half FIXED since 8.6.2 (a41a431706, 2014-07-31)
--disable-load: compiles cleanly, load gives "dynamic loading is not currently available". tclLoadNone.c TclpLoadMemory now matches the tclInt.h prototype on main, 9.0 and 8.6.
#if defined(__APPLE__) && defined(__DYNAMIC__) → <crt_externs.h>/_NSGetEnviron, is unchanged on main/9.0/8.6. The reporter's TARGET_OS_MAC guard would not help anyway: TARGET_OS_MAC is 1 on iOS too, so the test would have to be !TARGET_OS_IPHONE / TARGET_OS_OSX. The ios branch (ba55cebfc9, 2014) was never merged and was closed in 2022.
!TARGET_OS_IPHONE guard is still needed.
CANNOT-TEST (macOS)
-constraints nonPortable passes on Linux main. On macOS the test is still constrained {testfork nonPortable}, so it is skipped by default and no longer hangs the suite.
make test TESTFLAGS="-file unixForkEvent.test -constraints nonPortable"; if it passes, this can be closed.
3077617fff.patch: VERIFIED (configure level only; the AIX link itself is still untested)
tcl_cv_sys_version=AIX-5 unix/configure (gcc) -> Makefile LDFLAGS = $(LDFLAGS_OPTIMIZE), no -brtl (the xlc branch has it, the gcc branch does not)
~/tcltk/autoconf-2.73/bin/autoconf -f regenerates a configure byte-identical to the patched one
LDFLAGS = $(LDFLAGS_OPTIMIZE) -Wl,-brtl; a plain Linux configure is unaffected (no brtl). Linking on AIX is not tested
CANNOT-TEST (AIX); the cause is still present on main/9.0/8.6 by code reading
AIX-* (non-ia64) on main. For gcc, SHLIB_LD is ${CC} -shared -Wl,-bexpall with no -brtl. LDFLAGS="$LDFLAGS -brtl" is added only in the non-gcc (xlc) branch.
-ltcl8.5 only to libtcl8.5.a, never .so. So gcc ... -ltcl8.5 fails although libtcl8.5.so was built (reporter's ls confirms). The xlc path works because it adds -brtl.
LDFLAGS="$LDFLAGS -Wl,-brtl" in the gcc branch). Untested; needs an AIX box.
-ltcl8.5 cannot find libtcl8.5.so. Proposed fix: add -Wl,-brtl to LDFLAGS for gcc too. Needs testing by someone with AIX.
CANNOT-TEST (WinXP/MSVC6 test failures from 2010; partly fixed: filesystem-1.3/1.4 fixed by a47641a031, 8.6.6, per apnadkarni)
CANNOT-TEST (the sampleextension repository and the old CVS nightly tarballs are not available here). This is also not a Tcl core issue.
CANNOT-TEST (Expect for Windows 5.43 binary from 8.4 copied into ActiveState 8.5.4; third-party package, error only in an attached screenshot)
CANNOT-TEST (macOS); the cause is still present in main by code reading
#include "tcl.h", tclInt.h "tclPort.h"), and for the framework build unix/configure.ac puts them in a separate PrivateHeaders directory (PRIVATE_INCLUDE_DIR="${libdir}/PrivateHeaders"; install-private-headers in unix/Makefile.in copies them unchanged).
CANNOT-TEST (Interix)
2003130fff.patch: VERIFIED
LD_PRELOAD=eisdir.so tcltest fCmd.test -constraints notWine -> fCmd-6.16, 6.18, 9.9, 9.13 FAILED (error renaming "td2" to "td1/td2": is a directory), 273 passed/28 skipped/4 failed. (-constraints notWine is needed: on the tip notWine is still undefined in fCmd.test, so these tests are always skipped, 300724b80c.)
CANNOT-TEST (NFS server that returns EISDIR); reproduced with an LD_PRELOAD shim that maps rename()'s ENOTEMPTY/EEXIST to EISDIR: fCmd-6.16, 6.18, 9.9, 9.13 fail exactly as reported
notWine isn't defined in fCmd.test, so they are always skipped. That's 300724b80c, fixed only on the unmerged branch fCmd-test-constraints (0049270a8b, 2026-10-12). 8.6's fCmd.test has no notWine constraint.
notWine definition 300724b80c.
CANNOT-TEST (needs the TLS extension, not installed here; the reporter's stall scripts live on the old SourceForge TLS tracker #1945535, not in this repository)
CANNOT-TEST (QNX6)
make test results for 8.6.18 or 9.0.4 in a new ticket. Closing.
CANNOT-TEST (needs Tcl embedded in a thread of Macromedia Director MX 2004; nothing to reproduce with plain tclsh/wish)
CANNOT-TEST (Tru64 UNIX; obsolete)
TclpTime_t tt did not match the prototypes (promotmatch). TclpTime_t is gone, and the functions now have ANSI definitions taking const time_t *, so the error cannot recur. Tru64 is long dead and its cc is not supported by Tcl 8.6/9.
CANNOT-TEST (OSF1/Tru64 is a dead platform with no VM)
unix notRoot constraints.
CANNOT-TEST (needs nested non-native VFSes whose normalizePathProcs hand off to each other). By code reading the limitation is still there: TclFSNormalizeToUniquePath makes a single pass (native first, then each other FS once) and callers ignore the returned startAt.
CANNOT-TEST (LynxOS)
timezone manipulation were removed by the 8.5 clock rewrite). tclPosixStr.c in 8.6 and main still has unguarded case EBADMSG: and case ENOTSUP:, so a platform where EBADMSG == ENOTSUP would still fail with "duplicate case value". The -lbsd link need is not handled in tcl.m4 (no LynxOS case).
CANNOT-TEST (Tru64)
CANNOT-TEST (needs an NFS-mounted build dir; no NFS here and mounting needs root)
CANNOT-TEST (macOS)
CANNOT-TEST (HP-UX)
$(mandir)/mann with .n suffix (MANN_INSTALL_DIR), no HP-UX special case.
mannew/.new for section n; Tcl installs mann/.n everywhere.
Status now: Closed (Fixed)
fCmd-test-constraints, in main: Rebase fCmd-test-constraints to main. Fix [219158], 300724b80c and 852a17eeb9: fCmd.test constraints (read-only directory rename, the lost notWine definitiofCmd-test-constraints, in main: Rebase Fix [219158], 300724b80c and 852a17eeb9: fCmd.test constraitstrunk, in main, 9.0: Fix [219158], 300724b80c and 852a17eeb9: fCmd.test constraitscore-9-0-branch, in 9.0: Fix [219158], 300724b80c and 852a17eeb9: fCmd.test constraitsCANNOT-TEST (OpenBSD) — fCmd-9.4 passes on Linux (main 2026-10-22); the failure needs a rename(2) that requires write permission on a renamed directory
file rename of a 555 directory to succeed; POSIX allows EACCES (OpenBSD does this on purpose since 2013)
fossil merge --dry-run into current trunk gives 1 conflict in tests/fCmd.test, so it needs a rebase
make test TESTFLAGS="-file fCmd.test" on main (fCmd-9.4 should fail) and on fCmd-test-constraints (fCmd-9.4.2 runs with renameNeedsDirWrite, no failure)