zsh/Src, branch master

55128: socket: unmetafy the socket name

2026-09-01T05:22:46Z

This also led me to discover that if you type a leading nul in zsh, it passes through correctly to linux and it gives us an abstract socket instead, so document that that works, and print them properly. Also add support for reporting the autobind socket name if you pass an empty path. The more specific size in the bind and connect calls is due to abstract sockets using the entire specified length, regardless of any embedded 0 bytes.

55109: remove obsolete nameref-specific code from typeset_single

2026-08-23T22:18:45Z

Test changes here are for extra coverage, previous tests still pass

55061 (first part): break/continue errors in a subshell with no local loop

2026-08-23T00:18:22Z

54940: Forbid (!v) substitution flags for symmetry with (v!)

2026-08-23T00:10:58Z

55047 (first part): Revert ${!var} in ksh mode to remove namref special case

2026-08-23T00:00:11Z

55026: Fix ${(!)ref::=var} substitutions for symmetry with other flags

2026-08-22T23:44:43Z

55022: descriptive names for parameter declaration CPP macros

2026-08-22T23:15:29Z

55020: fix changing flags (not type) of a TYPESET_TO_UNSET parameter

2026-08-22T23:08:55Z

54810: use realparamtab to create and scope private variables

2026-08-22T22:58:38Z

55082: fix zlecs bounds checks in vimatchbracket and alignmultiwordleft

2026-08-14T03:10:41Z

55076: _exit in zexecve when the interpreter doesn't work

2026-08-13T12:26:00Z

55065: fix null deref in havefiles

2026-08-13T12:25:59Z

This was introduced in 51404.

55055: break: more completely enumerate loop types we are not in

2026-08-05T09:41:33Z

55021: prevent segfault upon "typeset -U" of a tied array hidden by a local

2026-07-27T04:22:48Z

55030: allow "typeset -hn", issue error on useless no-op "typeset -gh"

2026-07-27T03:58:50Z

Includes unposted K01nameref test

54994: use gdbm_fd_open so we can move the fd to >9

2026-07-25T05:03:05Z

This function is available since gdbm 1.13 and even debian old old stable has 1.17, so this should be fine.

54986: actually prevent FDT_MODULE fds from being closed by user

2026-07-25T05:03:05Z

The stated goal of FDT_MODULE is: /* * Entry visible to other processes but controlled by a module. * The difference from FDT_EXTERNAL is that closing this using * standard fd syntax will fail as there is some tidying up that * needs to be done by the module's own mechanism. */ #define FDT_MODULE 3 but this wasn't actually enforced. By accident, some fds ended up as FDT_INTERNAL because movefd clobbers the FDT_MODULE value and those were protected from closing. I think this straightens out all the bits properly, at least the ones I found. (Note that zsocket is using FDT_EXTERNAL purposefully because the user *is* expected to close those fds with the {myfd}>&- syntax, it has no specific interface to close opened sockets since it has no internal state for them.) Both zsocket and ztcp forgot to movefd the some of their fds which meant it was potentially in the 0-9 range, which is also not protected from being closed by the user, but: Since the -d flag allows the user to put module fds at low fds, I made the check forbid closing FDT_MODULE fds in that range too, I think that's correct? zftp did some very questionable things, relying on FDT_UNUSED fds to act approximately like it expected, instead use FDT_MODULE there too. Also make redup ensure the fdtable is big enough for the assignments after, just for consistency, since the old addmodulefd call before didn't necessarily grow it to the moved-to fd number. Remove some misleading comments like /* move the fd since no one will want to read from it */ the reason we move the fd is because we have promised the user we will not open fds in the 0-9 range. Also added maximum macro, and converted existing users of MIN/MAX to them, since we have not, as far as I can tell, made any effort to check that those macros are in fact defined by anyone.

54985: fix crash with bracketed-paste-magic and undo

2026-07-21T14:20:48Z

In particular, fc -P/-p didn't interact well with undo. Make sure we store/restore histline across those calls.

54943: Do not export named references even when ALLEXPORT is enabled

2026-07-19T21:13:59Z

54809: ZLE specials are always in realparamtab and can't be namerefs

2026-07-19T20:34:52Z

54941: Fix saving and restoring of subscripted inline assignments

2026-07-16T01:45:06Z

54956: init: don't reset ZLE_RPROMPT_INDENT on term init

2026-07-16T01:40:33Z

54927: checkptycmd: remove select()+FIONREAD variant, re-enable poll on osx

2026-07-11T23:57:56Z

54930: zexit: don't write history in sub-shells

2026-07-11T23:00:59Z

54909: Untokenize before quoting unquoted literal values

2026-07-08T18:40:29Z

% print -r - ${(q+):-foo = bar}|xxd 00000000: 2766 6f6f 208d 2062 6172 270a 'foo . bar'. % print -r - ${(q+):-foo * bar}|xxd 00000000: 2766 6f6f 2087 2062 6172 270a 'foo . bar'. % print -r - ${(qqqq):-foo = bar} $'foo \215 bar'

54783: improve checkptycmd

2026-07-04T17:56:26Z

With the previous (and now fallback) version of checkptycmd(), doing this: zpty foo 'while read; do echo hi; done' zpty (or zpty -t foo) would hang forever, which seems counterintuitive to the purpose of listing processes (or testing if a process is running).

54889: comparguments: always parse optargs as optargs

2026-07-03T11:59:42Z

54876: comparguments: search more thoroughly for foreign-set options

2026-07-03T11:59:40Z

54831: comparguments: treat empty description as if missing

2026-06-24T00:36:14Z

54792: prevent named references from being entered as named directories

2026-06-23T04:29:59Z

47546: builtin: trivial cleanup

2026-06-23T00:15:14Z

54822: the ::= fix in 54674 was incomplete

2026-06-21T07:26:14Z

% foo=xxx; () { echo ${2::=$foo} } yyy yyy We get the argv parameter returned when asking for "2" so use getvalue() to get a vbuf with the proper start/end set instead of making up bogus values.

54773: params: make out-of-range (w) subscript index work like others

2026-06-21T00:48:39Z

54767: zformat: replace -qn by spec quoting indicator syntax

2026-06-21T00:17:54Z

54801: fix two more warnings from 54710

2026-06-18T04:15:00Z

54776: fix warnings from 54710

2026-06-16T08:39:58Z

Some compilers warn if a label immediately precedes a variable declaration.

54771: fix a bug in nosort deduplication

2026-06-16T08:39:58Z

Introduced in 49915. Confusion in pointers could mark the wrong entries in some cases. Make sure asp is incremented to the correct entry by the outer loop's 'asp++', and don't touch sp at all. In particular, the third added test was missing the 'foo' entry: Y01completion.ztst: starting. --- expected +++ actual @@ -1,4 +1,3 @@ line: {tst }{} -NO:{foo} NO:{bar} NO:{baz} Was testing: nosort dedup with three same-string non-duplicates

54764: zerrmsg: handle null pointers with %s/%l

2026-06-16T03:15:01Z

54762: make curjob and prevjob work better

2026-06-13T23:57:17Z

In particular, these two sequences were still broken after 54294: % v() { vim "$@"; }; v [suspend vim] % urxvt & [3] 12707 % bg %- bg: job already in background % fg [1] continued [exit vim] % fg before: fg: no current job after: [3] + running urxvt % sleep 100 ^Z zsh: suspended sleep 100 % bg [1] continued sleep 100 % fg before fg: no current job after: [1] - running sleep 100 The changed test expected the sole remaining job to be %- which should not be possible, and now it is properly %+ instead.

54753: zpty: register pty fd as FDT_MODULE so it isn't closed when forking

2026-06-13T23:57:17Z

This lets us remove these mysterious braces too.

54721: illegal -> invalid, invalid option -> bad option

2026-06-13T23:57:17Z

Combinations of options can remain invalid, though.

54715: compctl: print multibyte errors correctly (phew)

2026-06-13T23:57:17Z

54714: add MB_NICECHAR and use it

2026-06-13T23:57:17Z

54713: fix multibyte error reporting in zparseopts

2026-06-13T23:57:08Z

This one is complicated by the fact it's using fprintf. If the user passed a nonprintable char as an option, we shouldn't print that, which nicechar handles for us. Fixes 54077.

54712: handle the Pound token in error message

2026-06-13T23:56:15Z

otherwise we print this: parse error in ${...\M-D...} substitution In practice, I'm not sure how to hit this path, I uncommented the condition to test it.

54711: print invalid option to getopts correctly

2026-06-13T23:56:15Z

54710: handle multibyte arguments in error messages

2026-06-13T23:56:15Z

A lot of places check option strings char by char, then send that char to zwarn/zerr, instead of reading out the full multibyte character. Doing so is quite easy with unmeta_one in most cases, so just go ahead and do that. The print functions in fact expect a wchar_t for %c anyway. Also renames "unknown option" to "bad option" because that's what every other place calls it (also one place was missing the colon, added that). To reproduce the math errors, since it's slightly less obvious (you need a multibyte character that is also an invalid parameter name, otherwise it actually works fine) % zsh -c '(( ¯ ))' zsh:1: bad math expression: illegal character: ^ % zsh -c 'zmodload zsh/mathfunc; (( sin(1¯) ))' zsh:1: bad math expression: illegal character: ^ % zsh -c 'の=5; echo $(( の ))' # unchanged because it is working 5 Also adds some tests because B12 was looking very empty, not sure if this is the correct way to handle the intersection of possibly not compiled module with possibly not available UTF-8 locale, but X02 and some others do a similar thing with ZSH_TEST_LANG, though in those cases that's handled specially by comptest, not the test itself.

unposted: kill: fix -l argument error, add tests

2026-06-11T01:47:14Z

fixes this issue: % kill -l 6term kill: unknown signal: SIGterm

54731: Adapt gdbm code to named references

2026-06-10T04:06:12Z

54727: Implement references to positional parameters

2026-06-10T03:47:57Z

54722: Fix nameref issues in builtin "unset"

2026-06-10T03:47:57Z

54707: fix autoload +X nonexist leaving funcdef at dummy

2026-06-10T03:20:52Z

% autoload +X nonexist zsh:999: nonexist: function definition file not found % nonexist exec.c:6086: function definition nonexist has zero reference count All the paths that set shf->funcdef in loadautofn call freeeprog on the old one first anyway, so rather than make a new mkautofn if loadautofn fails, just keep the old one.

54705: ksh93: fix use after free of .sh.edchar

2026-06-10T03:20:52Z

% bindkey T foo % zle -N foo % foo() { emulate -L ksh; () { zle -M ${.sh.edchar} } } % zmodload zsh/ksh93 % [press T twice] \M-(\M-^Ak\M-<\M-G^? or alternatively, emulate ksh in the main shell and just define this, but then you have to make sure you're in a zsh -f shell so random things aren't entering functions in the background. % foo() { zle -M ${.sh.edchar} } the code tries to restore sh_edchar = sh_unsetval; in the else branch, but this doesn't ever run in the above scenario because we never enter or leave a function with ksh emulation enabled (or indeed at all).

54704 + 54728: remove obsolete workaround in ksh93 module

2026-06-10T03:20:52Z

After 54475 changed deleteparamdef to use getnode2 instead, this workaround isn't needed anymore.

54696: unset REPLY before e:: and oe:: so things don't break if it had acquired a weird type

2026-06-10T03:20:52Z

Reported in 43032.

54693: show ARGV0 and - precommand more clearly in xtrace output

2026-06-10T03:20:51Z

Reported in 48901. Also has a secondary effect of correcting the string sent to the command_not_found_handler. If you don't have such a handler, then the errors are correct before the patch as well. command_not_found_handler() { echo -E >&2 command not found: $1 } Before: % ARGV0=foobar blurg zsh:1> ARGV0=foobar foobar command not found: foobar % - blurg +zsh:1> -blurg command not found: -blurg % ARGV0=foobar - blurg +zsh:1> ARGV0=foobar foobar command not found: foobar After: % ARGV0=foobar blurg +zsh:1> ARGV0=foobar blurg command not found: blurg % - blurg +zsh:1> - blurg command not found: blurg % ARGV0=foobar - blurg +zsh:1> ARGV0=foobar - blurg command not found: blurg

54692: untokenize string sent to matheval for (#) flag

2026-06-10T03:20:51Z

Fixes 49409.

54691: fix :W and :w duplicating input

2026-06-10T03:20:51Z

Reported in 50991 and introduced in 22525.

54688: handle Stéphane correctly

2026-06-10T03:20:51Z

In the original report we had [[ é = *$'\xa9' ]] which has op == P_STAR, in this case charstart is filled out inline, in the other cases it's done by the patrepeat function. If ptlen > P_LS_LEN(next) we have returned, if it is < then we have backtracked into the region where charstart is valid and we can check that, and finally if they are equal we know we are on a valid character boundary (but charstart might not be valid here, in fact multiple tests fail if you leave out the ptlen check). Separately, % [[ éé = é#$'\xa9' ]]; echo $? pattern.c:4113: closure following more than one character 0 % [[ éé = (é)#$'\xa9' ]]; echo $? 1 so just treat multibyte single characters as "not simple", which fixes that case too.

54687: fix brace completion

2026-06-10T03:20:51Z

% mkdir 'a b c'; touch 'a b c'/{'a b c','d e f'} % ls a\ b\ c/{ corrupts the command line into % ls a\ b\ {c/ and the effect is worse the more backslashes there are. It seems like this code should simply always use bp->qpos so there is no reason to keep hasunqu around either. For the compctl path, the curpos members need to be whatever they're determined to be at the start by addwhat for the comparison with ppl and psl, so just fix them at the end before add_match_data. The quotename in get_comp_string fixes subsequent completions, eg % ls a\ b\ c/{a\ b\ c,

54686: count lines in showmsg more correctly

2026-06-10T03:20:51Z

If you do this, wide() { zle -M ${(l:$COLUMNS::a:)} }; zle -N wide; bindkey W wide then we would move the cursor up one row too many after printing the message. The hunks with 1 + are triggered by this (one is MULTIBYTE the other not), wide() { zle -M ${(l:$COLUMNS::a:)}$'\n'b }; zle -N wide; bindkey W wide

54683: fix for ctrl-r with combining characters or nonprintable wide chars

2026-06-10T03:20:50Z

Typing one of these characters in isearch caused the terminal cursor to jump to the wrong row, and various other problems, because we totally lost track of things. Make use of addmultiword() to draw the combining characters, analogously to how the main input line does it. This resolves ccs/vcs getting out of sync with the display, because there are no top level glyphs of zero width anymore. The code seems to simply have forgotten about nonprintable wide characters that weren't ascii control characters, so handle those like in the main line redraw code. I didn't bother handling the ZSH_INVALID_WCHAR range because I don't see how you could possibly insert one of those in the isearch buffer. If you manage to do this, congratulations.

54674: ::= didn't respect parameter flags

2026-06-10T03:20:50Z

The documentation says ${name::=word} In the first form, if name is unset then set it to word; in the second form, if name is unset or null then set it to word; and in the third form, unconditionally set name to word. In all forms, the value of the parameter is then substituted. Particularly that "the value of the parameter is substituted", not "the word is substituted". This is what happens before the fix, and similarly for array flags like "unique": % typeset -E3 fl; hi=5; echo ${fl::=hi} hi % echo $fl 5.00e+00

54676: preserve COLUMNS + LINES from the environment

2026-06-09T02:11:58Z

54659: comparguments: support zsh-style '-' terminator with -S -S

2026-06-09T00:53:13Z

54718: Remove support for named references to subscripted variables

2026-06-09T00:39:39Z

54475: Don't perform dereferencing when looking up certain special parameters

2026-06-09T00:19:08Z

Tweaked to account for workers/54328

54700: Refine string and array slice assignments

2026-06-08T23:59:37Z

54294: Fix %- (prevjob) picking wrong job after resuming

2026-06-08T22:29:22Z

Also corrects exit status of "wait %%" (TBD from workers/43945)

54278: Detect invalid variable names in reference initializers

2026-06-08T22:01:06Z

54277: reference loops are detected at create time, remove loop tagging

2026-06-08T21:46:19Z

unposted: fix NULL termination in ksh93

2026-06-08T14:37:19Z

I'm not sure how I messed this up in 54558, but now it should be actually correct. I even tested it: % [[ foo = (#m)(#b)f(o)o ]] % zmodload -F zsh/ksh93 p:.sh.match % () { setopt localoptions ksharrays; typeset -p .sh.match } typeset -g -ar .sh.match=( foo o )

54672: use bangchar instead of this hardcoded !

2026-06-07T10:45:30Z

54671: fix anchored pattern with --disable-multibyte

2026-06-07T10:45:30Z

Just copied this code over from the multibyte branch.

54669: add --disable-link-all-libs configure option

2026-06-07T10:45:29Z

When this is used, libraries only needed by specific modules will only be linked to those modules (currently libgdbm and libpcre2). When --enable-cap is used, the main shell also uses libcap functions so we want to link that either way. Speaking of the cap module, it did some weird thing where it always compiles but creates useless builtins that only error, which means you can't rely on zmodload to check if they'll work. I'm sure someone will complain about me changing this but I'm going to at least try it and see if someone has a good reason for it. The gdbm module also has a big ifdef for the whole module, but instead of dummy builtins it just does #error, so remove that. We won't try to compile it if the user didn't ask for it, and if the user asked for it and it wouldn't compile, configure won't create a Makefile. Using the ac_cv_ cache variables in db_gdbm.mdd also doesn't actually work properly, those will remain set and enable the module to be linked after you pass --disable-gdbm and you will get the aforementioned #error. Use the enable_gdbm variable instead, like pcre.mdd does (and also cap.mdd now that I evilly made it conditional on --enable-cap).

54328: Only ever turn global parameters into named directories

2026-06-04T04:26:38Z

54340: Remove partial duplication of strsetfn and arrsetfn

2026-06-04T04:22:44Z

44874: fix a memory leak and potential heap corruption when rotating lines

2026-06-04T03:46:11Z

54624: parse_class would underallocate if posix character classes are used

2026-06-04T03:46:11Z

54620: fix check for when to hash commands

2026-06-04T03:46:09Z

54617: fix lookup in $commands array when hashlistall is unset

2026-06-04T03:45:50Z

While the documentation for the zsh/parameter module does state The zsh/parameter module gives access to some of the internal hash tables and commands This array gives access to the command hash table. , the latter entry does *also* say values are the pathnames of the files that would be executed when the command would be invoked. Change getpmcommand to always do a lookup via findcmd() if it the entry is not found in the hash and HASHLISTALL is unset. This will indirectly respect HASHCMDS and add the single entry if it is set, look it up in the hash if it was added and then HASHCMDS was unset, etc.

unposted: remove further remnant of restricted mode

2026-06-04T02:14:47Z

54644: init: add patchlevel to --version output

2026-06-02T00:00:21Z

49374: Show patchlevel in version string

2026-06-01T23:55:33Z

54625: remove .cvsignore

2026-05-30T04:23:34Z

54603 (tweaked): zsh_eval_context: add math context

2026-05-28T03:47:19Z

tweaked to add void to prototype + NEWS update

unposted: zformat: pass through %) with -q

2026-05-28T03:42:51Z

the behaviour in the original patch was not actually desirable

54602: zformat: add -qQ options to auto-escape %s

2026-05-28T03:35:22Z

+ NEWS update

54589: remove dead code identified by scan-build

2026-05-26T15:02:58Z

54615: fix ambiguous dangling from 54595

2026-05-25T23:56:02Z

54609: fix accept-and-hold completion for words with leading quotes

2026-05-25T05:10:42Z

The intention here was probably to skip leading braces, but the code is a no-op because m->qipl is 0 in that case. When there is a quote, we don't want to skip it, so just remove that thing. There can never be a combination of both cases, "{ will quote the brace and not actually do brace completion, and for {" other code will remove the " because we only handle leading quotes (I think?).

54609 + 54606: additional fixes for interactive menu selection

2026-05-25T05:10:42Z

Both of these involve we being incorrect, the we = wb + modelen hunk fixes this problem: % : ; foobar start interactive menu selection, type doc/z/ and hit tab, then select one of the files, you end up with % : Doc/Zsh/arith.yo[suffix space]h/; foobar The other hunk is for when there is a leading quote in the word being completed, like "NE being incorrectly completed to "NEWS"E During completion, the quote tokens are temporarily removed, and we is decremented (this all happens in get_comp_string, around the parts involving skipchars, I would advice not looking too closely). Then the complist code restores the original string via setmline but forgot to restore we. I don't think there's any point in trying to track exactly how many characters were deleted to add them back, just set it back to the cursor position.

54601: fix complist interactive mode overwriting buffer

2026-05-25T05:10:42Z

This has been discussed a few times starting with 49242, and continued at 50412, 50422, 52312 and 52657. reproduce recipe: % zmodload zsh/complist % bindkey '^I' menu-select % MENUMODE=interactive % touch test{1,2} % : ; foobar ^ 1. Type the above line in its entirety. 2. Place the cursor before the ; 3. Press Tab. 4. Press Enter. % : test1 bar ^ Completion is written over existing buffer contents. The main confusing thing to note is that do_single looks at minfo.cur, whether you pass it or not, because minfo is a global variable. When it is not NULL, it uses minfo.pos and minfo.len which are stale in these particular situations. When it is instead NULL, it recalculates them correctly for us, which is what we want. The two different hunks are separate code paths for accepting the completion or navigating to another match.

54600: fix metafy stuff in add_autoload_function

2026-05-25T05:10:42Z

54599: use correct buf size in cancd

2026-05-25T05:10:42Z

When path is metafied, it can be longer than PATH_MAX, so use MAXCMDLEN. Using ztrlen >= PATH_MAX instead of strlen >= MAXCMDLEN is just a minor optimization to avoid the syscall if we know the path is too long anyway, but it's not worth doing this extra check on the result of the snprintf call (probably).

54598: fix getfpfunc metafication

2026-05-25T05:09:00Z

It wouldn't load functions if they or the directory they were in had a metafied byte in the name, for example れ has a 0x82 byte. One detail to note here is the try_dump_file call using the unmetafied ppu path, and then metafied function name s. That's correct because it's comparing s with the function name inside the dumpfile, which is metafied.

54597: hashcmd could reject valid paths to hash if they had metafied characters

2026-05-25T05:07:39Z

And the iscom takes care of paths that are actually too long for us.

54596: pathprog used a signed char as a buffer size

2026-05-25T05:05:10Z

54595: overflow check in isreallycom

2026-05-25T05:05:10Z

54594: fix two very similar coverity issues

2026-05-25T05:05:09Z

Coverity CID 1255838 stack buffer overflow in execute() Coverity CID 1255846 stack buffer overflow in findcmd() This can be triggered by inserting long entries manually, then using =foo expansion. Unfortunately, it is also possible to overflow buf by importing a PATH entry that's over 4xPATH_MAX characters long. There was also some related confusion about how long a path can be and when/how to check the length of metafied strings, so fix all that up too. Use MAXCMDLEN (which is PATH_MAX*4) for metafied strings, though this is a little overkill, metafication can only double the length. We do not at least need the +1 at the end. PATH_MAX includes the terminating nul byte, so we make sure to not be off-by-one on our ztrlen checks too. The hunks in execute() are quite specific about the ztrlen length so it can continue checking or return the correct error, while findcmd() just relies on RET_IF_COM to work correctly instead. Somewhat unrelated change that snuck in, return 127 when the command is too long like the other exec failures, because I had to add a new return anyway.