On a commandline like "ls arg" (cursor at end) we do not expand
abbrevations on enter. OTOH, on "ls " we do expand. This can be
frustrating because it means that the two obvious ways to suppress
abbrevation expansion (C-Space or post-expansion C-Z) cannot be used to
suppress expansion of a command without arguments. (One workaround is
"ls #".)
Only expand-on-execute if the cursor is at the command name (no space
in between).
This is a strict improvement for realistic scenarios, because if there
is a space, the user has already expressed the intent to not expand
the abbreviation. (I hope no one is using recursive abbreviations.)
Closes#8423
This was supposed to act like `type -q` or `command -q`, in that it
returns 0 if at least 1 exists.
But because it used the wrong variable it didn't.
Fixes#8431.
This allows rebinding escape in the user list without breaking e.g.
arrow keys (which send escape and then `[A` and similar, so escape is
a prefix of them).
Fixes#8428.
This fixes printing octal and hex values that are negative or larger
than UINT_MAX.
Negative values get a leading -, like:
> math --base hex -10
-0xa
Fixes#8417.
Commit ec3d3a481 (Support "$(cmd)" command substitution without line
splitting, 2021-07-02) started treating an input string like
"a$()b" as if it were "a"$()"b". Yet, we do not actually insert the
virtual quotes. Instead we just adapted the definition of when quotes
are closed - hence the changes to quote_end().
parse_util_locate_cmdsubst_range() is aware
of the changes to quote_end() but some of its
callers like parse_util_detect_errors_in_argument() and
highlighter_t::color_as_argument() are not. They split strings at
command substitution boundaries without handling the special quoting
rules. (Only the expansion logic did it right.)
Fix this by handling the special quoting rules inside
parse_util_locate_cmdsubst_range(). This is a bit hacky since it
makes it harder for callers to process some substrings in between
command substitutions, but that's okay because current callers only
care about what's inside the command substitutions.
Fixes#8394
Since #4376, for-loops would set the loop variable outside, so it
stays valid.
They did this by doing the equivalent of
```fish
set -l foo $foo
for foo in 1 2 3
```
And that first imaginary `set -l` would also fire a set-event.
Since there's no use for it and the variable isn't actually set, we
remove it.
Fixes#8384.
widechar_width no longer classifies U+1F41F as widened-in-9, so the
width no longer changes.
Since we're interested in testing the change here, we need a different
emoji.
Just use 🥁, which was introduced in 9 as wide, and therefore widened
in 9.
fish might use XDG_RUNTIME_DIR for the uvar notifier fifo, so this
makes sure that tests are isolated.
Also set permissions to comply with the XDG basedir spec.
Like the $status commit, this would add the offset to already existing
errors, so
```fish
(foo)
(bar)
something
```
would see the "(foo)" error, store the correct error location, then
see the "(bar)" error, and *add the offset of (bar)* to the "(foo)"
error location.
Solve this by making a new error list and appending it to the existing
ones.
There's a few other ways to solve this, including:
- Stopping after the first error (we only display the first anyway, I
think?)
- Making it so the source location has an "absolute" flag that shows
the offset has already been added (but do we ever need to add two offsets?)
I went with the simpler fix.
This would break the location of any prior errors without doing
anything of value.
E.g.
```fish
echo foo | exec grep # this exec is not allowed!
$status
somethingelse # The error might be found here!
```
Would apply the offset of `$status` to the offset of `exec`, locating
the error for `exec` somewhere after $status!
Prior to this change, tmux based tests would call 'isolated-tmux' which would
initialize tmux on first call, an admitted "evil hack." Switch to requiring
an explicit call to 'isolated-tmux-start' which then defines 'isolated-tmux'
and other functions. Add some loop-until-prompt logic into
'isolated-tmux-start'. This improves reliability of the tmux tests on systems
under load; at least it makes the tests pass in the background on my Mac.
Remove the '$sleep' variable, to be replaced with 'tmux-sleep'.
This makes it so we treat backspaces as width -1, but never go below a
0 total width when talking about *lines*, like in screen or string
length --visible.
Fixes#8277.
When cd is passed a broken symlink, this changes the error message from
"no such directory" to "broken symbolic link". This scenario probably
won't happen very often since completion won't suggest broken symlinks
but it can't hurt to give a good error.
Fish used to do this until 7ac5932. This logic used to be in
path_get_cdpath, however, that is only used for highlighting, so we
don't need error messages there. Changing cd is enough.
Reword from "rotten" to "broken" since that's what file(1) uses.
Clean-up leftovers from old "rotten" code (nomen est omen).
See #8264
This currently changes builtin realpath with the "-s" option:
builtin realpath -s ///tmp
previously would print "///tmp", now it prints "/tmp".
The only thing "allow_leading_double_slashes" does is allow *two*
slashes.
This is important for `path match`, to be introduced in #8265.
This lets us run non-fish targets (such as `fish_tests`) under a clean
test environment without running into the fish-specific payload
configuration now carried out by `test_driver.sh` which expects a
`.fish` payload that it will run under a deterministically configured
instance of fish, running in an environment initialized by
`test_env.sh`.
This should fix the problem with in-tree builds leaving detritus behind
after a `make test` when `fish_tests` would be executed without
`test_driver.sh` - it is now executed under `test_env.sh` instead.
The tmux-prompt test would sometimes fail because the first call was:
isolated-tmux capture-pane -p
this would run a capture-pane which would race with starting fish
itself; occasionally the pane would be empty since fish has not yet
drawn a prompt. Add a loop to give fish time to draw the prompt.
On macOS, the tests would often fail because calls to `pkill` would "leak"
across tests: kill processes run by other tests. This is because on macOS,
the -P argument to pkill must come before the process name. On Linux it
doesn't matter.
This improves test reliability on Mac.
For littlecheck/pexpect this just unconditionally enables color.
I have no idea what happens if you run cmake outside of a terminal
, but the worst that can happen is that *errors* have color
escapes in them.
If someone figures out how to get cmake to tell us if it's running in
a terminal, we can add a check.
This used the *logical* $PWD, but realpath would operate on the
physical $PWD if given ".", even with -s. This makes this test fail if the $PWD is
logically different from physical.
This was long overdue since the setup logic is much more complex than
the actual tests.
tmux-prompt.fish had extra logic to protect against XDG_CONFIG_HOME
with leading double double-dot. I believe this is no longer necessary
with the new test driver.
We still use our own temp dir because we want to be able to run this
independently of the test driver, This can be useful for debugging
tests. For example we can insert a "$tmux attach" command in a test,
and then run
build/fish -C 'source tests/test_functions/isolated-tmux.fish' tests/checks/tmux-bind.fish
This allows to inspect the state of the test and debug interactively.
Attaching to the terminal doesn't work when running inside littlecheck
because littlecheck consumes our output and doesn't give us a terminal.
(Maybe there's an easy way to fix that?)
On request of a team member, this patches `basic.fish` to no longer
depend on being invoked by the test driver and started up in a $PWD that
points to a clean temporary directory.
This was requested by a team member who would like for some tests to
remain invokable (in thier own $HOME) directly via littlecheck without
relying on the test driver to prep the environment.
A comment explaining the rationale is also added so this doesn't get
passed down as folklore "you need to include this for tests to run" even
though no one understands why.
Tests are now executed in a test-specific temporary directory, so test
output on failure should be reproducible/reusable as-is without needing
to have TMPDIR defined (as it only exists by default under macOS).
Instead of trying to assert that there are no zombies when the test
starts (which often fails) and to prevent conflating existing or
irrelevant zombies with the ones we are interested in checking for,
have `ps` also emit the parent process id and filter its output to
include only children of the current fish instance.
Aside from the fact that the shared state could cause problems, tests
were randomly assuming it would be created where that wasn't the case.
In particular, `redirect.fish` and `basic.fish` were failing on only
macOS because `../test/temp` didn't exist yet - it would be created by
other tests later.
Even though we are using CMake's ctest for testing, we still define our
own `make test` target rather than use its default for many reasons:
* CMake doesn't run tests in-proc or even add each tests as an
individual node in the ninja dependency tree, instead it just bundles
all tests into a target called `test` that always just shells out to
`ctest`, so there are no build-related benefits to not doing that
ourselves.
* CMake devs insist that it is appropriate for `make test` to never
depend on `make all`, i.e. running `make test` does not require any
of the binaries to be built before testing.
* The only way to have a test depend on a binary is to add a fake test
with a name like "build_fish" that executes CMake recursively to
build the `fish` target.
* It is not possible to set top-level CTest options/settings such as
CTEST_PARALLEL_LEVEL from within the CMake configuration file.
* Circling back to the point about individual tests not being actual
Makefile targets, CMake does not offer any way to execute a named
test via the `make`/`ninja`/whatever interface; the only way to
manually invoke test `foo` is to to manually run `ctest` and specify
a regex matching `foo` as an argument, e.g. `ctest -R ^foo$`... which
is really crazy.
With this patch, it is now possible to execute any single test by name,
by invoking the build directly, e.g. to run the `universal.fish` check:
`cmake --build build --target universal.fish` or
`ninja -C build universal.fish`. Unfortunately, this is not integrated
into the Makefile wrapper, so `make universal.fish` won't work (although
this can potentially be hacked around).
Fixes#8232.
Note that this needed to have expect_prompt used in the pexpect test -
we might want to add a "catchup" there so you can just ignore the
prompt counter for a bit and pick it back up later.
This disables job control inside command substitutions. Prior to this
change, a cmdsub might get its own process group. This caused it to fail
to cancel loops properly. For example:
while true ; echo (sleep 5) ; end
could not be control-C cancelled, because the signal would go to sleep,
and so the loop would continue on. The simplest way to fix this is to
match other shells and not use job control in cmdsubs.
Related is #1362
* commandline: Add --is-valid option to query whether it's syntactically complete
This means querying when the commandline is in a state that it could
be executed. Because our `execute` bind function also inserts a
newline if it isn't.
One case that's not handled right now: `execute` also expands
abbreviations, those can technically make the commandline invalid
again.
Unfortunately we have no real way to *check* without doing the
replacement.
Also since abbreviations are only available in command position when
you _execute_ them the commandline will most likely be valid.
This is enough to make transient prompts work:
```fish
function reset-transient --on-event fish_postexec
set -g TRANSIENT 0
end
function maybe_execute
if commandline --is-valid
set -g TRANSIENT 1
commandline -f repaint
else
set -g TRANSIENT 0
end
commandline -f execute
end
bind \r maybe_execute
```
and then in `fish_prompt` react to $TRANSIENT being set to 1.
Because we are, ultimately, interested in how many cells a string
occupies, we *have* to handle carriage return (`\r`) and line
feed (`\n`).
A carriage return sets the current tally to 0, and only the longest
tally is kept. The idea here is that the last position is the same as
the last position of the longest string. So:
abcdef\r123
ends up looking like
123def
which is the same width as abcdef, 6.
A line feed meanwhile means we flush the current tally and start a new
one. Every line is printed separately, even if it's given as one.
That's because, well, counting the width over multiple lines
doesn't *help*.
As a sidenote: This is necessarily imperfect, because, while we may
know the width of the terminal ($COLUMNS), we don't know the current
cursor position. So we can only give the width, and the user can then
figure something out on their own.
But for the common case of figuring out how wide the prompt is, this
should do.
* Add `set --function`
This makes the function's scope available, even inside of blocks. Outside of blocks it's the toplevel local scope.
This removes the need to declare variables locally before use, and will probably end up being the main way variables get set.
E.g.:
```fish
set -l thing
if condition
set thing one
else
set thing two
end
```
could be written as
```fish
if condition
set -f thing one
else
set -f thing two
end
```
Note: Many scripts shipped with fish use workarounds like `and`/`or`
instead of `if`, so it isn't easy to find good examples.
Also, if there isn't an else-branch in that above, just with
```fish
if condition
set -f thing one
end
```
that means something different from setting it before! Now, if
`condition` isn't true, it would use a global (or universal) variable of
te same name!
Some more interesting parts:
Because it *is* a local scope, setting a variable `-f` and
`-l` in the toplevel of a function ends up the same:
```fish
function foo2
set -l foo bar
set -f foo baz # modifies the *same* variable!
end
```
but setting it locally inside a block creates a new local variable
that shadows the function-scoped variable:
```fish
function foo3
set -f foo bar
begin
set -l foo banana
# $foo is banana
end
# $foo is bar again
end
```
This is how local variables already work. "Local" is actually "block-scoped".
Also `set --show` will only show the closest local scope, so it won't
show a shadowed function-level variable. Again, this is how local
variables already work, and could be done as a separate change.
As a fun tidbit, functions with --no-scope-shadowing can now use this to set variables in the calling function. That's probably okay given that it's already an escape hatch (but to be clear: if it turns out to problematic I reserve the right to remove it).
Fixes#565
Fixes some regressions from 35ca42413 ("Simplify some parse_util functions").
The tmux tests are not beautiful but I find them easy to write.
Probably a pexpect test would also be enough here?
for PWD in foo; true; end
prints:
>..src/parse_execution.cpp:461: end_execution_reason_t parse_execution_context_t::run_for_statement(const ast::for_header_t&, const ast::job_list_t&): Assertion `retval == ENV_OK' failed.
because this used the wrong way to see if something is read-only.
This allows us to test that `test` takes numbers with decimal point even in comma-using locales,
to stop those pesky americans from breaking everything again.
(and yes, we use french to keep myself honest)
Through a mechanism I don't entirely understand, $PWD is sometimes
writable (so that `cd` can change it) and sometimes not.
In this case we ended up with it writable, which is wrong.
See #8179.
This didn't do all the syntax checks, so something like
fish -c 'echo foo; and $status'
complained of a missing command `0` (i.e. $status), and
fish -c 'echo foo | exec grep'
hit an assert!
So we do what read_ni does, parse each command into an ast, run
parse_util_detect_errors on it if it worked and then eval the ast.
It is possible to do this neater by modifying parser::eval, but I
can't find where.
This is slightly unclean. Even tho it would otherwise be syntactically
valid, using $status as a command is very very very likely to be an
error, like
if not $status
We have reports of this surprisingly regularly, including #2773.
Because $status can only ever be a value from 0 to 255, it is also
very unlikely to be an actual command, and that command is very
unlikely to do what you want.
So we simply point the user towards the "conditions" help section,
that should explain things.
This is opt-in through a new feature flag "ampersand-nobg-in-token".
When this flag and "qmark-noglob" are enabled, this command no longer
needs quoting:
curl https://example.com/thing?foo=bar&duran=duran
Compared to the previous approach e1570a4 ("Let '&' only separate as
the first char of a word"), this has some advantages:
1. "&&" and "&>" are no longer affected. They are still special, even
if used between tokens without spaces, like "echo bar&>foo".
Maybe this is not really *better*, but it avoids risking to annoy
users by breaking the old variant.
2. "&" is still special if at the end of a token, like in "sleep 1&".
Word movement is not affected by the semantics change, so Alt-F and
friends still stop at every "&".
Currently, if a "return" is given outside of a function, we'd just
throw an error.
That always struck me as a bit weird, given that scripts can also
return a value.
So simply let "return" outside also exit the script, kinda like "exit"
does.
However, unlike "exit" it doesn't quit an interactive shell - it seems
weird to have "return" do that as well. It sets $status, so it can be
used to quickly set that, in case you want to test something.
Today the reader exposes its internals directly, e.g. to the commandline
builtin. This is of course not thread safe. For example in concurrent
execution, running `commandline` twice in separate threads would cause a
race and likely a crash.
Fix this by factoring all the commandline state into a new type
'commandline_state_t'. Make it a singleton (there is only one command
line
after all) and protect it with a lock.
No user visible change here.
In the variable handler, we just go through the entire thing and keep
every element once.
If there's a duplicate, we set it again, which calls the handler
again.
This takes a bit of time, to be paid on each startup. On my system,
with 100 already deduplicated elements, that's about 4ms (compared to
~17ms for adding them to $PATH).
It's also semantically more complicated - now this variable
specifically is deduplicated? Do we just want "unique" variables that
can't have duplicates?
However: This entirely removes the pathological case of appending to
$fish_user_paths in config.fish (which should be an FAQ entry!), and the implementation is quite simple.
This adds a hack to the parser. Given a command
echo "x$()y z"
we virtually insert double quotes before and after the command
substitution, so the command internally looks like
echo "x"$()"y z"
This hack allows to reuse the existing logic for handling (recursive)
command substitutions.
This makes the quoting syntax more complex; external highlighters
should consider adding this if possible.
The upside (more Bash compatibility) seems worth it.
Closes#159
This apparently doesn't work at all under Github Actions with tsan, so let's skip it.
If anyone feels the need to dig deeper into this, have at it. I find
this distracting.
When the user presses control-C, fish marks a cancellation signal which
prevents fish script from running, allowing it to properly unwind.
Prior to this commit, the signal was cleared in the reader. However this
missed the case where a binding would set $fish_bind_mode which would
trigger event handlers: the event handlers would be skipped because of
the cancellation flag was still set. This is similar to #6937.
Let's clear the flag earlier, as soon as we it's set, in inputter_t.
Fixes#8125.
In some setups (eg. macports) $tmpdir can expand to more than
100 symbols and tests fail with 'socket file name too long'
errors.
Using relative path to socket file fixes the issue.
* string: Allow `collect --no-empty` to avoid empty ellision
Currently we still have that issue where
test -n (thing | string collect)
can return true if `thing` doesn't print anything, because the
collected argument will still be removed.
So, what we do is allow `--no-empty` to be used, in which case we
print one empty argument.
This means
test -n (thing | string collect -n)
can now be safely used.
"no-empty" isn't the best name for this flag, but string's design
really incentivizes reusing names, and it's not *terrible*.
* Switch to `--allow-empty`
`--no-empty` does the exact opposite for `string split` and split0.
Since `-a`/`--allow-empty` already exists, use it.
The tmux-prompt test was failing when run more than once, because
XDG_DATA_HOME has a leading double-dot, causing the uvars file to
leak across sessions. Descend more deeply into our tmpdir to isolate
our XDG_DATA_HOME.
This reverts commit b56b230076.
which somehow made us miss repaints on uvar notifications.
The commit was a workaround for a polling bug which was later properly
fixed by 7c5b8b855 ("Use the uvar notifier pipe timestamp to avoid
excessive polling"), so it's no longer necessary.
Add a system test. If I had a better understanding of the bug I could
probably write a better test.
Fixes#8088
We used to warn about PATH and CDPATH that are not valid directories,
but only if they contain colons.
However, the warning was a false positive because we would split
those values by colons anyway. So there is nothing left we want to
warn about.
Fixes#8095
sigint2 would hang (probably because of different semantics in signal
delivery?)
wcstod isn't implemented correctly, so math can't do hex numbers.
OpenBSD only passes the filename as argv[0] and doesn't give us another feature I know of, so status fish-path can't work.
* Try to set LC_CTYPE to something UTF-8 capable
When fish is started with LC_CTYPE=C (even just effectively, often via
LC_ALL=C!), it's basically broken. There's no way to handle non-ASCII
characters with a C locale unless we want to write our
locale-independent replacements for all of the system functions.
Since we're not going to do that, let's try to find *some locale* for
LC_CTYPE.
We already do that in __fish_setlocale, but that's
- a bit of a weird thing that reads unstandardized system
configuration files
- allows setting locale to C explicitly
So it's still easily possible to end up in a broken configuration.
Now, the issue with this is that there is (AFAICT) no portable way to
get a list of all allowed locales and C.UTF-8 is not standardized, so
we have no one locale to fall back on and are forced to try a few. The
list we have here is quite arbitrary, but it's a start.
Python does something similar and only tries C.UTF-8, C.utf8 and
"UTF-8".
Once C.UTF-8 is (hopefully) standardized, that will just start
working (tm).
Note that we do not *export* the fixed LC_CTYPE variable, so external
programs still have to deal with the C locale, but we have no real
business messing with the user's environment.
To turn it off: $fish_allow_singlebyte_locale, if set to something true (like "1"),
will re-run the locale initialization and skip the bit where we force
LC_CTYPE to be utf8-capable.
This is mainly used in our tests, but might also be useful if people
are trying to do something weird.
The hope is that the noshebang test was fixed on old glibc
through e74b9d53df. Revert the previous optimistic attempts to
fix these through adding sleeps and subshells.
This reverts commit b3da0bd5a2.
This reverts commit 8a86d3452f.
This is an attempt to solve the test failures on Launchpad's CI.
I'm assuming when we do a redirection like
foo > file
and then try to execute `file` immediately afterwards, we either
haven't written it soon enough or closed the file, so we get a "text
file busy" error.
So, when we do that in a new fish the file should be closed once it
quits.
See #8021.
When you try to execute a file directly after you've written to it,
you might, on some systems, get a "text file busy" error.
So we unfortunately have to sleep to avoid it.
See #8021 for where this was added,
537b3f6cb1 for the same problem.
Now that `$last_pid` is never fish's pid, we no longer need to force
jobs to run in their own pgroup. Restore the job control behavior to
what it was prior, so that signals may be delivered properly in
non-interactive mode.
This reverts commit 3255999794
Prior to this change, a function with an on-job-exit event handler must be
added with the pgid of the job. But sometimes the pgid of the job is fish
itself (if job control is disabled) and the previous commit made last_pid
an actual pid from the job, instead of its pgroup.
Switch on-job-exit to accept any pid from the job (except fish itself).
This allows it to be used directly with $last_pid, except that it now
works if job control is off. This is implemented by "resolving" the pid to
the internal job id at the point the event handler is added.
Also switch to passing the last pid of the job, rather than its pgroup.
This aligns better with $last_pid.
It is possible to run a function when a process exits via `function
--on-process-exit`, or when a job exits via `function --on-job-exits`.
Internally these were distinguished by the pid in the event: if it was
positive, then it was a process exit. If negative, it represents a pgid
and is a job exit. If zero, it fires for both jobs and processes, which is
pretty weird.
Switch to tracking these explicitly. Separate out the --on-process-exit
and --on-job-exit event types into separate types. Stop negating pgids as
well.
This switches builtin_wait from waiting on jobs in the active job list, to
waiting on the wait handles. The wait handles may be either derived from
the job list itself, or from saved wait handles from jobs that exited in
the background.
Fixes#7210
Prior to this fix, an escaped character like \x41 (hex for ascii A)
was interpreted the same was as A, so that $\x41 would be the same
as $A. Fix this by inserting an INTERNAL_SEPARATOR before these escapes,
so that we no longer treat it as part of the variable name.
This also affects brackets; don't treat echo $foo\1331\135 the same as
echo $foo[1].
Fixes#7969
This is the last time I'm doing this before I rip these particular
tests out.
As far as I know there is no actual *problem* here, this is just
failing through a combination of macOS and Github Actions being slow
as molasses.
So it is wasting our time and therefore worse than not having these
tests at all, especially since they very rarely fail for good reasons.
We would leave some escape delay tests intact with generous timeouts, which would provide 90%
of the coverage with 10% of the hassle.
This simply checks if the parser requested exit after running any
binding scripts (in read_normal_chars).
I think this means we no longer need the `exit` bind function.
Fixes#7967.
Just add some extra sleep time so it hopefully also works when the
CI system is overloaded. This succeeded >60 times in the CI, without
a single failure.
In case it legitimately fails again, we should provide simple steps
to reproduce the failure interactively (using "tmux attach").
The uvar issue only triggered because two fish are started - one is
running the tmux-complete script, the other one is running inside tmux.
We could reduce the complexity of this test by writing it in a
different language, like sh or python.
Reproducible at least on Linux, where the "named pipe" universal
variable notifier is used:
rm -rf build/test/xdg_config
XDG_CONFIG_HOME=build/test/xdg_config ./build/fish -c "xterm -e ./build/fish"
The child fish reacts to keyboard input with a noticeable initial
delay. This is because the universal variable file is polled over
a million times, even when I immediately press Control-D. This polling
prevents readb() from handling keyboard input.
Before commit 939aba02d ("Refactor input_common.cpp:readb"), readb()
reacted to keyboard input even when there were universal variable
notifications. Restore this behavior, but make sure to call the
universal variable notifier after the new "prepare_to_select" logic.
Maybe the problem is in the notifier but the old behavior was sane.
Fixes the problems described in
7a556ec6f2 (commitcomment-49773677)
Adding "-d uvars-file" to the reproducesr shows that we are checking
the uvar file repeatedly:
uvar-file: universal log sync
uvar-file: universal log sync elided based on fast stat()
uvar-file: universal log no modifications
This only uses the functions fish ships with, but still doesn't allow
any *customization*, which is the point of no-config.
This makes it a lot more usable, given that the actual normal prompt
and things are there.
This still doesn't set any colors, because we don't run
__fish_config_interactive because we don't read config.fish (any
config.fish), because that would run the snippets.
In many cases we currently discard escaped newlines, since they
are often unnecessary (when used around &|;). Escaped newlines
are useful for structuring argument lists. Allow them for variable
assignments since they are similar.
Closes#7955
This would print the default "Argument is invalid" error string, which
is *true* but not super obvious, because `test` doesn't always perform
numeric conversion, and that's the bit that failed here.
This refactors the behavior of string match with capture groups to
correctly handle multiple arguments. Now the variable capture applies to
the first match, as documented. Fixes#7938.
string match is documented as setting an unset variable if a capture group
is unmatched in an otherwise matched regex, and if the `--all` flag is not
provided. However prior to this fix, it instead set a variable containing
the empty string as a single value. Correct the implementation to match
the documentation.
Note that if the `--all` flag is provided we continue to set empty
strings, which is documented.
This removes the relative XDG paths, which could have potentially
confused tmux, and also starts the window with the correct size
instead of adjusting the size afterwards.
Make it an ordinary struct wrapping a vector, instead of a template.
This is in preparation for using it more widely, for matching bindings
as well as mouse CSI sequences.
Also add some mouse-disabling tests.
This runs in 100ms increments, so there's not a lot of harm in trying
longer - it should take the same time everywhere it succeeded before.
But I've reproduced failures on FreeBSD 13 on sr.ht, so there's at
least one platform where a total time of 1 second isn't enough.
Now we do 50 tries, which is 5 seconds.
This could have been one iteration off, e.g.
```fish
function on-winch --on-signal winch
echo $LINES
end
```
Resize the terminal, it'll print e.g.
24
then run `echo $LINES` interactively, it might have a different answer.
This isn't beautiful, but it works. A better solution might be to make
the termsize vars electric and just always update them on read?
Fixes#7926.
Also switches the default status order for non-informative to the informative one:
stagedstate invalidstate dirtystate untrackedfiles stashstate
instead of
dirty staged stash untracked
Things like
```fish
complete command -n '__fish_seen_subcommand_from subcommand'
--force-files
```
would not be obeyed because we only checked force-files when there was
an option.
Fixes#7920.
When fish starts, it notices which pgroup owns the tty, and then it
restores that pgroup's tty ownership when it exits. However if fish does
not own the tty, then (on Mac at least) the tcsetpgrp call triggers a
SIGSTOP and fish will hang while trying to exit.
The first change is to ignore SIGTTOU instead of defaulting it. This
prevents the hang; however it risks re-introducing #7060.
The second change somewhat mitigates the risk of the first: only do the
restore if the initial pgroup is different than fish's pgroup. This
prevents some useless calls which might potentially steal the tty from
another process (e.g. in #7060).
This correctly sets $status when a builtin succeeds but its output fails;
for example if the output is redirected to a file and that write fails.
Fixes#7857
* math: Make function parentheses optional
It's a bit annoying to use parentheses here because that requires
quoting or escaping.
This allows the parens to be omitted, so
math sin pi
is the same as
math 'sin(pi)'
Function calls have the lowest precedence, so
math sin 2 + 6
is the same as
math 'sin(2 + 6)'
* Add more tests
* Add a note to the docs
* even moar docs
Moar docca
* moar tests
Call me Nikola Testla
It's not super clear what $SHLVL is useful for, but the current
definition is essentially
"number of shells in the parent processes + 1"
which isn't *super useful*?
Bash's behavior here is a bit weird in that it increments $SHLVL
basically always, but since it auto-execs the last process it will
decrement it again, so in practice it's often not incremented.
E.g.
```
> echo $SHLVL
1
> bash -c 'echo $SHLVL; bash'
2
>> echo $SHLVL
2
```
Both bashes here end up having the same $SHLVL because this is
equivalent to `echo $SHLVL; exec bash`. Running `echo $SHLVL` and then
`bash -c 'echo $SHLVL'` in an interactive bash will have a different
result (1 and 2) because that doesn't *exec* the inner bash.
That's not something we want to get into, so what we do is increment
$SHLVL in every interactive fish. Non-interactive fish will simply
import the existing value.
That means if you had e.g. a bash that runs a fish script that ends up
opening a new fish session, you would have a $SHLVL of *2* - one for the
bash, and one for the inner fish.
We key this off is_interactive_session() (which can also be enabled
via `fish -i`) because it's easy and because `fish -i` is asking for
fish to be, in some form, "interactive".
That means most of the time $SHLVL will be "how many shells am I deep,
how often do I have to `exit`", except for when you specifically asked
for a fish to be "interactive". If that's a problem, we can rethink it.
Fixes#7864.
- Check for special characters *before* attempting to parse
- Also ignore lines with `{` and `*`
- Also skip lines with `<<` because that might be a heredoc (or a
- `<<<` herestring)
Fixes#7874.
This cleans up some exit code processing. Previously a failed exec
would produce exit code 125 unconditionally, while a failed posix_spawn
would produce exit code 1 (!).
With this change, fish reports exit code 126 for not-executable, and 127
for file-not-found. This matches bash.
We have no idea why this was even a thing. For now simply set it to
"all"/"full" (why these two names? no idea) at startup and allow
changing it later.
Settting it *immediately* when defining the variable sets it too soon
because we don't have the interactive signal handlers
enabled (including the one for SIGTTOU), so let's first settle for
this little piece of awkwardness.
This needs widespread testing, so we merge it early, immediately after
the release.
Fixes#5036Fixes#5832Fixes#7721
(and probably numerous others)
I believe they are both equivalent for our particular purpose, since we
only care about enforcing the size fish sees.
`resize-window` was only introduced in tmux 2.9, which isn't available
at least on Ubuntu 18.04 LTS (currently using tmux 2.6) and probably
many others.
(Clever idea to use tmux here!)
Consider
$ complete -c foo -a 'aab aaB' -f
$ foo A<TAB>
since 28d67c8 we would insert the common prefix AND show the pager.
Due to case-insensitive comparison, "b/B" was considered to be part
of the prefix. Since the prefix is added to each pager item [1]
we get wrong results. Fix this by removing the insensitive comparison
between completions - I don't think it was of much use anyway.
Commandline tokens are still matched case-insensitively, this is
just about completions.
Test this by running interactive fish inside tmux (pexpect's terminal
emulation not have enough capabilities). Also add tests for recent
interactive regressions #7526 and #7738.
Closes#3978
[1]: b38a23a would solve this differently by giving every pager item
its own prefix, but was reverted since it needs more fixes.
This should cover most cases - the user didn't install the docs and is
trying to view the man page via __fish_print_help, so we don't have a
way to show anything.
But `help thing` will fall back to the online version of the docs,
which should work if there's an internet connection.
See #7824.
This concerns the behavior when running an external command from a key
binding. The history is:
Prior to 5f16a299a7, fish would run these external commands in shell
modes. This meant that fish would pick up any tty changes from external
commands (see #2114).
After 5f16a299a7, fish would save and restore its shell modes around
these external commands. This introduced a regression where anything the
user typed while a bound external command was executing would be echoed,
because external command mode has ECHO set in c_lflag. (This can be
reproed easily with `bind -q 'sleep 1'` and then pressing q and typing).
So 5f16a299a7 was reverted in fd9355966.
This commit partially reverts fd9355966. It has it both ways: external
commands are launched with shell modes, but/and shell modes are restored
after the external command completes. This allows commands to muck with
the tty, as long as they can handle getting shell modes; but it does not
enable ECHO mode so it fixes the regression found in #7770.
Fixes#7770. Fixes#2114 (for the third time!)
This partially reverts commit fd9355966e.
Unfortunately this causes input coming in while bind functions are
running to show up on screen.
Since the cure is worse than the disease let's just stop doing it.
My guess is this needs to *only* be done while running an external
command.
Fixes#7770
Reintroduces #2114
Partially reverts 5f16a299a7
This is broken on OpenBSD because it apparently doesn't have a /proc
we can query, so it just gives "fish".
Since it's unnecessary in this context just skip it.
For reasons unclear to me, fish enables bold mode unconditionally if
the background is set.
However, this called a background "set" if it wasn't exactly the
"normal" color, whereas set_color --print-colors would set a color
of *none*.
We have three special non-color colors:
- "normal"
- "reset"
- "none"
All of these specify some form of absence of background color, so all
of them should be checked.
Fixes#7805
* Rewrite the real file if history file is a symlink
When the history file is a symbolic link, `fish` used to overwrite
the link with a real file whenever it saved history. This makes
it follow the symlink and overwrite the real file instead.
The same issue was fixed for the `fish_variables` file in 622f2868e
from https://github.com/fish-shell/fish-shell/pull/7728.
This makes `fish_history` behave in the same way. The implementation
is nearly identical.
Since the tests for the two issues are so similar, I combined them
together and slightly expanded the older test.
This also addresses https://github.com/fish-shell/fish-shell/issues/7553.
* Add user-facing error when history renaming fails
Currently, when history file renaming fails, no message is shown to the
user. This happens, for instance, if the history file is a symlink
pointing to another filesystem.
This copies code (with a bit of variation, after reviewer comments) from
589eb34571/src/env_universal_common.cpp (L486-L491)
into `history.cpp`, so that a message is shown to the user.
* fixup! Rewrite the real file if history file is a symlink
This reverts commit e240d81ff8 and
introduces a more compatible method of finding newly added fish scripts
to syntax check.
`find -newer` is the original and is supported by everything under the
sun (including FreeBSD, NetBSD, Solaris, OpenIndiana, macOS 10.10, WSL,
and more), and if not, the tests will succeed anyway. `find -mnewer` was
added later around the time `find -cnewer` and co (which checks the
creation date rather than the modification date) was introduced, but
apparently the GNU version of coreutils never introduced the `-mnewer`
alias for `-newer`.
Yes, this is hacky and yes it would be ideal if the build system is the
one that picked which tests to run rather than the test itself picking.
But let's not pretend that our tests are idealogically ideal or pure
right now and until we fix the mess that is our CMake test integration
(e.g. use ctest and configure each test to be run separately with
configurable payloads, etc) eight seconds is still eight seconds, and
again, the CI isn't affected.
My find (GNU findutils 4.8.0) prints
> find: unknown predicate `-mnewer'
So we would have to test for support.
Also this is *super* hacky - tests aren't supposed to keep files
around, this is something you would do in the build system.
This reverts commit ddd0e28b4f.
Only check fish files that have been modified since the last time they
were checked. (This continues with the assumption that we are testing
for broken /usr/share fish scripts and not breakage of the fish parser,
which is covered by all the other tests.)
This saves 8 seconds on an NVMe disk under WSL. Won't affect integrity
of CI runs, which start with a blank slate each time.
Apparently the grep on FreeBSD doesn't do \s or \t. Since we're
looking for an actual tab, just give it an actual tab.
See https://builds.sr.ht/~faho/job/448496.
This `set -e` had a cartesian product that caused it to remove the
indexes separately, so the later indexes were off - removing the first
and then the second ends up removing the first and then the
old-*third* which is now the second.
Just quote the expansion so it runs in one go.
Fixes#7776
NetBSD's sleep quits when foregrounded sometimes. I'm not entirely
sure *why*, but this is reproducible with the default /bin/sh, so it's
not our fault.
Because this fails our tests, go back to using cat *there*, because we
can't use it on macOS - 4c9d01cab0.
The user may write for example:
echo foo >&5
and fish would try to output to file descriptor 5, within the fish process
itself. This has unpredictable effects and isn't useful. Make this an
error.
Note that the reverse is "allowed" but ignored:
echo foo 5>&1
this conceptually dup2s stdout to fd 5, but since no builtin writes to fd
5 we ignore it.
fish_indent used to increment the indentation level whenever we saw an escaped
newline. This broke because of recent changes to parse_util_compute_indents().
Since parse_util_compute_indents() function already indents continuations
there is not much to do for fish_indent - we can simply query the indentation
level of the newline. Reshuffle the code since we need to pass the offset
of the newline. Maybe this can even be simplified further.
Fixes#7720
This fails on FreeBSD on sr.ht and NetBSD on my own VM, but it works manually.
It also fails on macOS but I have no way to confirm.
I think it might be a problem in pexpect's platform support?
Either way, the test is valuable so just skip it there and solve it later.
Otherwise this would look weird if you had, say, a tab in there.
See #7716.
(note that this doesn't handle e.g. zero-width-joiners, because those
aren't currently escaped. we might want to add an escape mode for
unprintable characters, but for combining codepoints that's tricky!)
After commit 6dd6a57c60, 3 remaining
builtins were affected by uint8_t overflow: `exit`, `return`, and
`functions --query`.
This commit:
- Moves the overflow check from `builtin_set_query` to `builtin_run`.
- Removes a conflicting int -> uint8_t conversion in `builtin_return`.
- Adds tests for the 3 remaining affected builtins.
- Simplifies the wording for the documentation for `set --query`.
- Does not change documentation for `functions --query`, because it does
not state the exit code in its API.
- Updates the CHANGELOG to reflect the change to all builtins.
This was lost in
6bdbe732e40c2e325aa15fcf0f28ad0dedb3a551..c7160d7cb4970c2a03df34547f357721cb5e88db.
Note that we only print a term-support flog message for now, the
warning seems a bit much.
Fixes#7709.
Prior to this fix, if stdin were explicitly closed, then builtins would
silently fail. For example:
count <&-
would just fail with status 1. Remove this limitation and allow each
builtin to handle a closed stdin how it sees fit.
builtin_set_query returns the number of missing variables. Because the
return value passed to the shell is an 8-bit unsigned integer, if the
number of missing variables is a multiple of 256, it would overflow to 0.
This commit saturates the return value at 255 if there are more than 255
missing variables.
Prior to this change, the checks/git.fish test would fail if run from a
git interactive rebase (such as via `git rebase -i --exec 'ninja test'`),
because git itself would inject stuff into the environment. Teach the git
test how to clean up its environment first before running.
This needs to be rewritten, I'm pretty sure we have like 6 of these
kinds of ad-hoc "is this quoted" things lying around.
But for now, at least don't just check if the *previous* character was
a backslash.
Fixes#7685.
This goes to a separate file because that makes option parsing easier
and allows profiling both at the same time.
The "normal" profile now contains only the profile data of the actual
run, which is much more useful - you can now profile a function by
running
fish -C 'source /path/to/thing' --profile /tmp/thefunction.prof -c 'thefunction'
and won't need to filter out extraneous information.
Expansion parses slices like "$PATH[1..2]", but so does "set" when assigning
"set PATH[1..2] . .". Commit be06f842a ("Allow to omit indices in index
range expansions") forgot the latter.
This makes the fish_git_prompt variable handlers kick in, meaning we
see the informative chars.
The big question here is what happens if there's a non-UTF-8 locale in
the test.
Theoretically we set LC_CTYPE, but.....
This used to print a literal DEL character in the output for `bind`,
which wouldn't actually show up and made it hard to figure out what
the key was.
So we just escape it back to how we actually used it - `\x7f`.
Fixes#7631.
A weird interaction between grouped short options and our weird option
parsing that puts unknown options back:
```
echo "-n foo"
```
would see the `-n`, turn off printing newlines, interpret the " " as
another grouped short option, see that there is no short option for
space and put the entire token back on the arguments pile.
So it would print "-n foo" *without a newline*.
Fix this by keeping an old state of the options around and reverting
it when putting options back.
The alternative is *probably* to forbid the " " short option in
wgetopt, then check if an option group contains it and error out, but
this should only really be a problem in `echo` because that is,
AFAICT, the only thing that puts the options back.
Fixes#7614
This adds a test to ensure that if a long running background process is
launched from a command substitution, that process does not cause the
cmdsub to hang. That could easily happen if we just wait for the pipe to
close; this is verifying that we are also checking for the job to complete.
Prior to this change, `fish_private_mode` worked by just suppressing
history outright. With this change, `fish_private_mode` can be toggled on
and off. Commands entered while `fish_private_mode` is set are stored but
in memory only; they are not written to disk.
Fixes#7590Fixes#7589
Don't go into implicit interactive mode without ever executing
anything - not even `exit` or reacting to ctrl-d. That just renders
the shell useless and unquittable.
It was always a bit ridiculous that argparse required `X-longflag` if
that "X" short flag was never actually used anywhere.
Since the short letter is for getopt's benefit, we can hack around
this with our old friend: Unicode Private Use Areas.
We have a counter, starting at 0xE000 and going to 0xF8FF, that counts
up for all options that don't have a short flag and provides one. This
gives us up to 6400 long-only options.
6.4K should be enough for everybody.
Prior to this change, a glob like `**/file.txt` would only match
`file.txt` in subdirectories; the `**` must match at least one directory.
This is historical behavior.
With this change we move a little closer to bash's implementation by
allowing a literal `**` segment to match in the current directory. That
is, `**/foo` will match both `foo` and `bar/foo`, while `b**/foo` will
only match `bar/foo`.
Fixes#7222.
Before running a command, or before importing a command from bash history,
we perform error checking. As part of error checking we expand commands
including variables and globs. If the glob is very large, like `/**`, then
we could hang expanding it.
One fix would be to limit the amount of expansion from the glob, but
instead let's just not expand command globs when performing error checking.
Fixes#7407
This would tell you a function was "Defined in - @ line 1" for every
function defined via `source`.
Really, ideally we'd figure out where the *source* call was, but that'
much more complicated, so we just give a comprehensible message.
This matches what we do in --profile's output:
```
> source /home/alfa/.config/fish/config.fish
--> set -gx XDG_CACHE_HOME /home/alfa/.cache
--> set -gx XDG_CONFIG_HOME /home/alfa/.config
--> set -gx XDG_DATA_HOME /home/alfa/.local/share
```
instead of
```
+ source /home/alfa/.config/fish/config.fish
+++ set -gx XDG_CACHE_HOME /home/alfa/.cache
+++ set -gx XDG_CONFIG_HOME /home/alfa/.config
+++ set -gx XDG_DATA_HOME /home/alfa/.local/share
```
This increases a 100ms timeout to 200ms, because we've hit it on
Github Actions:
```
INPUT 3904.65 ms (Line 223): set -g fish_escape_delay_ms 100\n
OUTPUT +1.74 ms (Line 224): \rprompt 25>
INPUT +0.71 ms (Line 230): echo abc def
INPUT +0.57 ms (Line 231): \x1b
INPUT +0.57 ms (Line 232): t\r
OUTPUT +2.41 ms (Line 234): \r\ndef abc\r\n
OUTPUT +1.63 ms (Line 234): \rprompt 26>
INPUT +0.75 ms (Line 239): echo ghi jkl
INPUT +0.57 ms (Line 240): \x1b
INPUT +134.98 ms (Line 242): t\r
```
In other places it decreases sleeps where we just wait for a timeout to elapse, in which case we don't need much longer than the timeout.
And again clang-format does something I don't like:
- if (found != end && std::strncmp(found->name, name, len) == 0 && found->name[len] == 0) return found;
+ if (found != end && std::strncmp(found->name, name, len) == 0 && found->name[len] == 0)
+ return found;
I *know* this is a bit of a long line. I would still quite like having
no brace-less multi-line if *ever*. Either put the body on the same
line, or add braces.
Blergh
E.g. if we do `string match -q`, and we find a match, nothing about
the input can change anything, so we quit early.
This is mainly useful for performance, but it also allows `string`
with `-q` to be used with infinite input (e.g. `yes`).
Alternative to #7495.
Currently a bit limited, unfortunately printf's `%a` specifier is
absolutely unreadable.
So we add `hex` and `octal` with `0x` and `0` prefixes respectively,
and also take a number but currently only allow 16 and 8.
The output is truncated to integer, so scale values other than 0 are
invalid and 0 is implied.
The docs mention this may change.
The '--import' flag was used for importing named capture groups, but it
was decided to always import them unconditionally. This flag was causing
the tests to fail.
Prior to this fix, when key binding is a script command (i.e. not a
readline command), fish would run that key binding using fish's shell
tty modes. Switch to using the external tty modes. This re-fixes
issue #2214.
With the upcoming fix to place the tty in external-proc mode, add a sleep
which resolves a race between emitting a newline and restoring it to shell
mode.
Prior to this change, when a process resumes because it is brought back
to the foreground, we would reset the terminal attributes to shell mode.
This fixed#2114 but subtly introduced #7483.
This backs out 9fd9f70346, re-introducing #2114 and re-fixing #7483.
A followup fix will re-fix #2114; these are broken out separately for
bisecting purposes.
Fixes#7483.
I *think* this might sometimes (on CI) be eating the prompt, so that the actual `prompt`
part of `expect_prompt` doesn't find anything.
On Github Actions we see things like:
```
Testing file pexpects/generic.py ... Failed to match pattern: prompt 5
generic.py:35: timeout from expect_prompt("echo .history.*")
[...]
OUTPUT +1.08 ms (Line 31): \rprompt 4>
INPUT +0.35 ms (Line 34): echo $history[1]\n
OUTPUT +1.58 ms (Line 35): echo $history[1]\r\necho $history[1]\r\n⏎ \r⏎ \r\rprompt 5>
```
so the prompt *is* printed, it's just not correctly matched.
Apparently on macOS SIGTSTP (from control-Z) causes `read()` to return
EINTR.
This means `cat | cat` will exit as soon as it's backgrounded and
brought back.
So instead we use `sleep`, which won't read(), and therefore is
impervious to these puny attacks.
See discussion in #7447.
The classic mistake: Some of these have a bit of a delay, but it's supposed to
be *under* the timeout, so it needs to be *shorter* not longer to
increase the slack.
We just had the following output on Github Actions:
INPUT +0.94 ms (Line 34): echo ghi jkl
INPUT +0.72 ms (Line 35): \x1b
INPUT +63.12 ms (Line 37): t\r
The default escape delay is 30ms, that had 60ms between an escape and
a tab, so it missed it.
So: We have to increase the delay for CI's benefit. Let's try with
80ms, because otherwise we'd have to bump up other timeouts and the
bind tests take long enough as it is.
Github Action's macOS builds are even more resource-starved (even tho
they use the same provider?) than
Travis, but Travis is unusable to us now, so....
In some cases the completion we come up with may be unexpected, e.g.
if you have files like
/etc/realfile
and
/etc/wrongfile
and enter "/etc/gile", it will accept "wrongfile" because "g" and
"ile" are in there - it's a substring insertion match.
The underlying cause was a typo, so it should be easy to go back.
So we do a bit of magic and let "cancel" undo, but only right after a
completion was accepted via complete or complete-and-search.
That means that just reflexively pressing escape would, by default, get you back to
the old token and let you fix your mistake.
We don't do this when the completion was accepted via the pager,
because 1. there's more of a chance to see the problem there and 2.
it's harder to redo in that case.
Fixes#7433.
This was typically overridden by "too many/few arguments", but it's
actually incorrect:
sin(55
has the correct number of arguments to `sin`, but it's lacking
the closing `)`.
We've heard news of this regressing, so let's add the test that should
have been there already (mea culpa!).
Because we now use POSIX_VDISABLE, this should also work in tandem
with ctrl-space (which sends NUL), but we can't test *that* because
some systems might not have POSIX_VDISABLE.
This switch is no longer necessary when only one command is given.
Internally completions are stored separately for each command,
so we only every print one command name per "complete" line anyway.
These aliases seem to be common, see #7389 and others. This prevents
recursion on that example, so `alias ssh "env TERM=screen ssh"` will just
have the same completions as ssh.
Checking the last token is a heuristic which hopefully works for most
cases. Users are encouraged to use functions instead of aliases.
Ensure that the increment= param is set via keyword, not via positional arg.
This mistake was masking a bug where the "^a b c" match was not being tested,
because it was being set as the value for increment!
This switches the 'increment' param from "after" to "before." Instead
of expect_prompt saying if the next prompt will be incremented, each
call site says if it should have been incremented sinec the last prompt.
I am not sure why this worked, actually.
These tests did not have $fish set anywhere, and on my fresh OpenBSD
VM it ended up calling whatever that calls "fish" (I think it's that
"Go fish!" game?).
If the padding is not divisible by the char's width without remainder,
we pad the remainder with spaces, so the total width of the output is correct.
Also add completions, changelog entry, adjust documentation, add examples
with emoji and some tests. Apply some minor style nitpicks and avoid extra
allocations of the input strings.
Prior to this change, tab completing with a variable assignment like
`VAR=val cmd<tab>` would parse out and apply VAR=val, then recursively
invoke completions. This caused some awkwardness around the wrap chain -
if a wrapped command had a variable completion we risked infinite
recursion. A secondary problem is that we would run any command
substitutions inside variable assignment, which the user does not expect
to run until pressing enter.
With this change, we explicitly track variable assignments encountered
during tab completion, including both those explicitly given on the
command line and those found during wrap chain walk. We then apply them
while suppressing command substitutions.
In preparation for applying variable assignments (VAR=VAL cmd), separate
them out from the command when performing completions. This includes both
those that the user typed, and any that come about through
completion --wraps.
The "wrap chain" refers to a sequence of commands which wrap other
commands, for completion purposes. One possibility is that a wrap chain
will produce a combinatorial explosion or even an infinite loop, so there
needs to be logic to prevent that. Part of that logic is encapsulated in a
visited set (wrap_chain_visited_set_t) to prevent exploring the same item
twice.
Prior to this change, we stored pairs (command, wrapped_command). But we
only really need to store the wrapped command. Switch to that.
One consequence is that if a command wraps another command in more than
one way, we won't explore both ways. This seems unlikely in practice.
Detect recursive calls to builtin complete and the internal completion in
the same place.
In 0a0149cc2 (Prevent infinite recursion when completion wraps variable assignment)
we don't print an error when completing certain aliases like:
alias vim "A=B vim"
But we also gave no completions.
We could make this case work, but I think that trying to salvage situations
like this one is way too complex. Instead, let the user know by printing an
error. Not sure if the style of the error fits.
We could add some heuristic to alias to not add --wraps in some cyclic cases.
