[PATCH] completion: complete paths for git send-email

[PATCH] completion: complete paths for git send-email

[Yury Norov (NVIDIA)]

From: Yury Norov <ynorov@nvidia.com>

git send-email accepts either revisions or paths to patch files, but its
Bash completion only offers revisions. This prevents patch files from
being completed. It can also make a prefix such as "0" expand to an
unrelated hexadecimal ref even when matching 0001-*.patch files exist.

In my Linux tree, an attempt to autocomplete the standard-named patch
brings a random hashtag:

 $ ls 0*
 0001-bitmap-drop-bitmap_next_set_region.patch
 $ git send-email 0<Tab>
 $ git send-email 05c69d298c96703741cac9a5cbbf6c53bd55a6e2

Introduce an append variant of __gitcomp_file() and use it to add
filesystem candidates after the existing revision candidates.  Keep the
latter because revisions remain valid send-email arguments.

Add a regression test covering patch files alongside a 40-hex ref.

Assisted-by: Codex <codex@openai.com>
Signed-off-by: Yury Norov <ynorov@nvidia.com>
---
 contrib/completion/git-completion.bash | 29 +++++++++++++++++++-------
 t/t9902-completion.sh                  | 12 ++++++++++-
 2 files changed, 33 insertions(+), 8 deletions(-)

diff --git a/contrib/completion/git-completion.bash b/contrib/completion/git-completion.bash
index e87578771..b7017488d 100644
--- a/contrib/completion/git-completion.bash
+++ b/contrib/completion/git-completion.bash
@@ -579,21 +579,18 @@ __gitcomp_file_direct ()
 }
 
 # Generates completion reply with compgen from newline-separated possible
-# completion filenames.
+# completion filenames by appending them to the existing list of completion
+# candidates, COMPREPLY.
 # It accepts 1 to 3 arguments:
 # 1: List of possible completion filenames, separated by a single newline.
 # 2: A directory prefix to be added to each possible completion filename
 #    (optional).
 # 3: Generate possible completion matches for this word (optional).
-__gitcomp_file ()
+__gitcomp_file_append ()
 {
 	local IFS=$'\n'
 
-	# XXX does not work when the directory prefix contains a tilde,
-	# since tilde expansion is not applied.
-	# This means that COMPREPLY will be empty and Bash default
-	# completion will be used.
-	__gitcompadd "$1" "${2-}" "${3-$cur}" ""
+	__gitcompappend "$1" "${2-}" "${3-$cur}" ""
 
 	# use a hack to enable file mode in bash < 4
 	compopt -o filenames +o nospace 2>/dev/null ||
@@ -601,6 +598,23 @@ __gitcomp_file ()
 	true
 }
 
+# Generates completion reply with compgen from newline-separated possible
+# completion filenames.
+# It accepts 1 to 3 arguments:
+# 1: List of possible completion filenames, separated by a single newline.
+# 2: A directory prefix to be added to each possible completion filename
+#    (optional).
+# 3: Generate possible completion matches for this word (optional).
+__gitcomp_file ()
+{
+	# XXX does not work when the directory prefix contains a tilde,
+	# since tilde expansion is not applied.
+	# This means that COMPREPLY will be empty and Bash default
+	# completion will be used.
+	COMPREPLY=()
+	__gitcomp_file_append "$@"
+}
+
 # Find the current subcommand for commands that follow the syntax:
 #
 #    git <command> <subcommand>
@@ -2634,6 +2648,7 @@ _git_send_email ()
 		;;
 	esac
 	__git_complete_revlist
+	__gitcomp_file_append "$(compgen -f -- "$cur")"
 }
 
 _git_stage ()
diff --git a/t/t9902-completion.sh b/t/t9902-completion.sh
index 55dc9eabf..e87827f21 100755
--- a/t/t9902-completion.sh
+++ b/t/t9902-completion.sh
@@ -2777,7 +2777,17 @@ test_expect_success PERL 'send-email' '
 	test_completion "git send-email --val" <<-\EOF &&
 	--validate Z
 	EOF
-	test_completion "git send-email ma" "main "
+	test_completion "git send-email ma" "main " &&
+
+	git tag 05c69d298c96703741cac9a5cbbf6c53bd55a6e2 &&
+	test_when_finished "git tag -d 05c69d298c96703741cac9a5cbbf6c53bd55a6e2 &&
+		rm -f 0001-example.patch 0002-example.patch" &&
+	touch 0001-example.patch 0002-example.patch &&
+	test_completion "git send-email 0" <<-\EOF
+	0001-example.patch
+	0002-example.patch
+	05c69d298c96703741cac9a5cbbf6c53bd55a6e2 Z
+	EOF
 '
 
 test_expect_success 'complete files' '
-- 
2.53.0

Re: [PATCH] completion: complete paths for git send-email

[Junio C Hamano]

"Yury Norov (NVIDIA)" <yury.norov@gmail.com> writes:

> From: Yury Norov <ynorov@nvidia.com>
>
> git send-email accepts either revisions or paths to patch files, but its
> Bash completion only offers revisions. This prevents patch files from
> being completed. It can also make a prefix such as "0" expand to an
> unrelated hexadecimal ref even when matching 0001-*.patch files exist.
>
> In my Linux tree, an attempt to autocomplete the standard-named patch
> brings a random hashtag:
>
>  $ ls 0*
>  0001-bitmap-drop-bitmap_next_set_region.patch
>  $ git send-email 0<Tab>
>  $ git send-email 05c69d298c96703741cac9a5cbbf6c53bd55a6e2

Wow.  Even though I use nothing but 'git send-email' when sending my
own patches, I have never noticed this behavior.  I guess that is
primarily because I only use the command via my own wrapper script,
so the usual bash completion kicks in only for filenames in my
workflow.  Since I store my patches two levels deep in my working
tree (for example, '+outgo/topic/0000-cover-letter.txt'), I suspect
that even if I got rid of my wrapper, I would not suffer from this
issue.  An attempt to run 'git send-email +outgo/contrib-doc/0<TAB>'
expanding the trailing '0' into a hexadecimal object name would
indeed be quite annoying.

Good find.

> Introduce an append variant of __gitcomp_file() and use it to add
> filesystem candidates after the existing revision candidates.  Keep the
> latter because revisions remain valid send-email arguments.

OK.  I will need help from those who are more familiar with our
completion code than I am to properly assess this change.  Any
assistance in reviewing this would be appreciated.

> diff --git a/t/t9902-completion.sh b/t/t9902-completion.sh
> index 55dc9eabf..e87827f21 100755
> --- a/t/t9902-completion.sh
> +++ b/t/t9902-completion.sh
> @@ -2777,7 +2777,17 @@ test_expect_success PERL 'send-email' '
>  	test_completion "git send-email --val" <<-\EOF &&
>  	--validate Z
>  	EOF
> -	test_completion "git send-email ma" "main "
> +	test_completion "git send-email ma" "main " &&
> +
> +	git tag 05c69d298c96703741cac9a5cbbf6c53bd55a6e2 &&
> +	test_when_finished "git tag -d 05c69d298c96703741cac9a5cbbf6c53bd55a6e2 &&
> +		rm -f 0001-example.patch 0002-example.patch" &&

If the initial 'git tag' fails, 'test_when_finished' is never
registered, and we end up failing to remove the '000?-example.patch'
files.  The usual way to write this is:

 - set up 'test_when_finished' with a body that is written to
   succeed even if the clean-up target is not present (your '-f' in
   'rm -f' is good, as it prevents 'rm' from failing even if
   '0001-example.patch' does not get created); then

 - write the test code that dirties the state (requiring clean-up)
   after registering the 'test_when_finished' handler.

That is, "Prepare the clean-up first, and then you do not have to
worry about making a mess."

By the way, the use of a purely hexadecimal string as a tag or
branch name is highly misleading.  What happens if an object exists
whose name is identical to that tag?  Git offers ways to
disambiguate if you really want to, but I do not see any reason for
a sensible person or workflow to deliberately place oneself in a
situation where such disambiguation becomes necessary.

Of course, that is no excuse for the bug.  Our completion script
should not misbehave, even when confronted with a workflow that uses
funny-looking tags.

Re: [PATCH] completion: complete paths for git send-email

[D. Ben Knoble]

On Sun, Jul 19, 2026 at 9:45 AM Yury Norov (NVIDIA)
<yury.norov@gmail.com> wrote:
>
> From: Yury Norov <ynorov@nvidia.com>
>
> git send-email accepts either revisions or paths to patch files, but its
> Bash completion only offers revisions. This prevents patch files from
> being completed. It can also make a prefix such as "0" expand to an
> unrelated hexadecimal ref even when matching 0001-*.patch files exist.
>
> In my Linux tree, an attempt to autocomplete the standard-named patch
> brings a random hashtag:

It is unusual to call this a "hashtag." Perhaps "hash" or "object
name" (or id) based on the glossary and datamodel docs?

>  $ ls 0*
>  0001-bitmap-drop-bitmap_next_set_region.patch
>  $ git send-email 0<Tab>
>  $ git send-email 05c69d298c96703741cac9a5cbbf6c53bd55a6e2
>
> Introduce an append variant of __gitcomp_file() and use it to add
> filesystem candidates after the existing revision candidates.  Keep the
> latter because revisions remain valid send-email arguments.
>
> Add a regression test covering patch files alongside a 40-hex ref.
>
> Assisted-by: Codex <codex@openai.com>
> Signed-off-by: Yury Norov <ynorov@nvidia.com>
> ---
>  contrib/completion/git-completion.bash | 29 +++++++++++++++++++-------
>  t/t9902-completion.sh                  | 12 ++++++++++-
>  2 files changed, 33 insertions(+), 8 deletions(-)
>
> diff --git a/contrib/completion/git-completion.bash b/contrib/completion/git-completion.bash
> index e87578771..b7017488d 100644
> --- a/contrib/completion/git-completion.bash
> +++ b/contrib/completion/git-completion.bash
> @@ -579,21 +579,18 @@ __gitcomp_file_direct ()
>  }
>
>  # Generates completion reply with compgen from newline-separated possible
> -# completion filenames.
> +# completion filenames by appending them to the existing list of completion
> +# candidates, COMPREPLY.
>  # It accepts 1 to 3 arguments:
>  # 1: List of possible completion filenames, separated by a single newline.
>  # 2: A directory prefix to be added to each possible completion filename
>  #    (optional).
>  # 3: Generate possible completion matches for this word (optional).
> -__gitcomp_file ()
> +__gitcomp_file_append ()
>  {
>         local IFS=$'\n'
>
> -       # XXX does not work when the directory prefix contains a tilde,
> -       # since tilde expansion is not applied.
> -       # This means that COMPREPLY will be empty and Bash default
> -       # completion will be used.
> -       __gitcompadd "$1" "${2-}" "${3-$cur}" ""
> +       __gitcompappend "$1" "${2-}" "${3-$cur}" ""
>
>         # use a hack to enable file mode in bash < 4
>         compopt -o filenames +o nospace 2>/dev/null ||
> @@ -601,6 +598,23 @@ __gitcomp_file ()
>         true
>  }
>
> +# Generates completion reply with compgen from newline-separated possible
> +# completion filenames.
> +# It accepts 1 to 3 arguments:
> +# 1: List of possible completion filenames, separated by a single newline.
> +# 2: A directory prefix to be added to each possible completion filename
> +#    (optional).
> +# 3: Generate possible completion matches for this word (optional).
> +__gitcomp_file ()
> +{
> +       # XXX does not work when the directory prefix contains a tilde,
> +       # since tilde expansion is not applied.
> +       # This means that COMPREPLY will be empty and Bash default
> +       # completion will be used.
> +       COMPREPLY=()
> +       __gitcomp_file_append "$@"
> +}
> +

Curious; the diff itself is much more readable for me when applied
locally (it shows the addition of __gitcomp_file_append and the
replacement of a few lines in __gitcomp_file).

Nonetheless, this follows the pattern established by __gitcompadd and
__gitcompappend, so that part at least looks like it functions as
expected. (I can't comment too much on the code that existed there
already.)

>  # Find the current subcommand for commands that follow the syntax:
>  #
>  #    git <command> <subcommand>
> @@ -2634,6 +2648,7 @@ _git_send_email ()
>                 ;;
>         esac
>         __git_complete_revlist
> +       __gitcomp_file_append "$(compgen -f -- "$cur")"

At least with Bash with compgen, this looks to me like it does append
file names to the COMPREPLY.

But, with the "hack" comment in the modified function, do we also need
to account for older bash? It looks like that comes from 3ffa4df4b2
(completion: add hack to enable file mode in bash < 4, 2013-04-27).
After studying a bit more, that hack is to make Bash do the right
thing during file completion, not to workaround different methods of
generating filenames (unlike Zsh, which has a newer and an older
completion system, Bash's seems relatively stable?).

So, I think this looks good.

>  }
>
>  _git_stage ()
> diff --git a/t/t9902-completion.sh b/t/t9902-completion.sh
> index 55dc9eabf..e87827f21 100755
> --- a/t/t9902-completion.sh
> +++ b/t/t9902-completion.sh
> @@ -2777,7 +2777,17 @@ test_expect_success PERL 'send-email' '
>         test_completion "git send-email --val" <<-\EOF &&
>         --validate Z
>         EOF
> -       test_completion "git send-email ma" "main "
> +       test_completion "git send-email ma" "main " &&
> +
> +       git tag 05c69d298c96703741cac9a5cbbf6c53bd55a6e2 &&
> +       test_when_finished "git tag -d 05c69d298c96703741cac9a5cbbf6c53bd55a6e2 &&
> +               rm -f 0001-example.patch 0002-example.patch" &&
> +       touch 0001-example.patch 0002-example.patch &&
> +       test_completion "git send-email 0" <<-\EOF
> +       0001-example.patch
> +       0002-example.patch
> +       05c69d298c96703741cac9a5cbbf6c53bd55a6e2 Z
> +       EOF
>  '
>
>  test_expect_success 'complete files' '
> --
> 2.53.0

Junio commented on the test, so I'll stop here.

Pending a commit message tweak for "hashtag," I'm satisfied enough for

Reviewed-by: D. Ben Knoble <ben.knoble@gmail.com>

(Or feel free to use "Acked-by" if this is not a strong enough review
for you/the project!)

-- 
D. Ben Knoble

Re: [PATCH] completion: complete paths for git send-email

[Junio C Hamano]

"D. Ben Knoble" <ben.knoble@gmail.com> writes:

> On Sun, Jul 19, 2026 at 9:45 AM Yury Norov (NVIDIA)
> <yury.norov@gmail.com> wrote:
>>
>> From: Yury Norov <ynorov@nvidia.com>
>>
>> git send-email accepts either revisions or paths to patch files, but its
>> Bash completion only offers revisions. This prevents patch files from
>> being completed. It can also make a prefix such as "0" expand to an
>> unrelated hexadecimal ref even when matching 0001-*.patch files exist.
>>
>> In my Linux tree, an attempt to autocomplete the standard-named patch
>> brings a random hashtag:
>
> It is unusual to call this a "hashtag." Perhaps "hash" or "object
> name" (or id) based on the glossary and datamodel docs?

Very good point, but I am not sure if the author truly meant object
names here.  The reproduction test uses a long hexadecimal string,
but that is not an object name; it is an unusual-looking tag name.
It is like naming a topic branch '012345' and complaining that:

    $ git send-email 0<TAB>

completes the input to the branch name while ignoring the
0001-changes.patch file.

When you have a branch named '0-tolerance-policy' and:

    $ git send-email 0<TAB>

completes to that branch name, you would not dream of complaining
about the completion.  IOW, I think the complaint is somewhat unfair
to begin with.

Actually, I do not know if the completion script really expands an
abbreviated object name to a full one.  I tried:

    $ git rev-parse seen^2
    179eccf0d01729c19a3238905b951b1880aa4ba1
    $ git checkout master
    $ . contrib/completion/git-completion.bash
    $ git send-email 17<TAB>

and waited for some time, but it did not complete to anything.

In any case, when both a '0001-my-changes.patch' file and a
'0-tolerance-policy' branch exist in your repository and current
working directory, running:

    $ git send-email 0<TAB>

should offer both as candidates, I thihk.  Since I only ever pass
filenames to the command, I personally do not think it is a huge
loss if the completion script stops looking at refs and sticks to
filenames only, but others may have a use for that feature.

Re: [PATCH] completion: complete paths for git send-email

[Yury Norov]

On Tue, Jul 21, 2026 at 10:09:56AM -0700, Junio C Hamano wrote:
> "D. Ben Knoble" <ben.knoble@gmail.com> writes:
> 
> > On Sun, Jul 19, 2026 at 9:45 AM Yury Norov (NVIDIA)
> > <yury.norov@gmail.com> wrote:
> >>
> >> From: Yury Norov <ynorov@nvidia.com>
> >>
> >> git send-email accepts either revisions or paths to patch files, but its
> >> Bash completion only offers revisions. This prevents patch files from
> >> being completed. It can also make a prefix such as "0" expand to an
> >> unrelated hexadecimal ref even when matching 0001-*.patch files exist.
> >>
> >> In my Linux tree, an attempt to autocomplete the standard-named patch
> >> brings a random hashtag:
> >
> > It is unusual to call this a "hashtag." Perhaps "hash" or "object
> 
> Very good point, but I am not sure if the author truly meant object
> names here.   > name" (or id) based on the glossary and datamodel docs?

I said hashtag because for me it's a hash of the tag:

git send-email 0<TAB>
git send-email 05c69d298c96703741cac9a5cbbf6c53bd55a6e2

But also it's a name of the tag, and git warns about it:

  $ git show 05c69d298c96703741cac9a5cbbf6c53bd55a6e2
  warning: refname '05c69d298c96703741cac9a5cbbf6c53bd55a6e2' is ambiguous.
  Git normally never creates a ref that ends with 40 hex characters
  because it will be ignored when you just specify 40-hex. These refs
  may be created by mistake. For example,
  
    git switch -c $br $(git rev-parse ...)
  
  where "$br" is somehow empty and a 40-hex ref is created. Please
  examine these refs and maybe delete them. Turn this message off by
  running "git config set advice.objectNameWarning false"
  commit 05c69d298c96703741cac9a5cbbf6c53bd55a6e2 (tag: 05c69d298c96703741cac9a5cbbf6c53bd55a6e2)
  Author: Tejun Heo <tj@kernel.org>
  Date:   Tue May 15 08:22:04 2012 +0200
  ...

I have no local branch or local file with that name, but the tag exists
for 14 years, and will not go away. And yes, it breaks autocompletion.

So, after rethinking, the problem looks like this: if autocompletion
logic finds a tag beginning with that pattern, it doesn't attempt to
search for the matching files, which is wrong

> The reproduction test uses a long hexadecimal string,
> but that is not an object name; it is an unusual-looking tag name.
> It is like naming a topic branch '012345' and complaining that:
> 
>     $ git send-email 0<TAB>
> 
> completes the input to the branch name while ignoring the
> 0001-changes.patch file.
> 
> When you have a branch named '0-tolerance-policy' and:
> 
>     $ git send-email 0<TAB>
> 
> completes to that branch name, you would not dream of complaining
> about the completion.  IOW, I think the complaint is somewhat unfair
> to begin with.
> 
> Actually, I do not know if the completion script really expands an
> abbreviated object name to a full one.  I tried:
> 
>     $ git rev-parse seen^2
>     179eccf0d01729c19a3238905b951b1880aa4ba1
>     $ git checkout master
>     $ . contrib/completion/git-completion.bash
>     $ git send-email 17<TAB>
> 
> and waited for some time, but it did not complete to anything.
> 
> In any case, when both a '0001-my-changes.patch' file and a
> '0-tolerance-policy' branch exist in your repository and current
> working directory, running:
> 
>     $ git send-email 0<TAB>
> 
> should offer both as candidates, I thihk.  Since I only ever pass
> filenames to the command, I personally do not think it is a huge
> loss if the completion script stops looking at refs and sticks to
> filenames only, but others may have a use for that feature.

Agree. The test should create a file 0001.patch, then a tag
0-tag, then a branch 0-branch, maybe something else that is
relevant; and then make sure every option is correctly offered
by autocompletion.

Guys please let me know if everything else is needed before I send v2.

Thanks,
Yury

Re: [PATCH] completion: complete paths for git send-email

[Junio C Hamano]

Yury Norov <ynorov@nvidia.com> writes:

>> In any case, when both a '0001-my-changes.patch' file and a
>> '0-tolerance-policy' branch exist in your repository and current
>> working directory, running:
>> 
>>     $ git send-email 0<TAB>
>> 
>> should offer both as candidates, I thihk.  Since I only ever pass
>> filenames to the command, I personally do not think it is a huge
>> loss if the completion script stops looking at refs and sticks to
>> filenames only, but others may have a use for that feature.
>
> Agree. The test should create a file 0001.patch, then a tag
> 0-tag, then a branch 0-branch, maybe something else that is
> relevant; and then make sure every option is correctly offered
> by autocompletion.
>
> Guys please let me know if everything else is needed before I send v2.

So in short, we want the problem description updated to something
like:

   When branches and tags whose names share the same prefix as a
   file (or a directory???) that stores a patch exist, the attempt
   to complete that shared prefix

       $ git send-email that-shared-prefix<TAB>

   should offer both branches, tags, and files (and directories???).
   But the completion only offers branches and tags and fails to
   offer files.

And the description of the solution would follow after that in the
proposed log message.

As to the tests, using 40-hex is misleading, and 0-branch as you
said would be sufficient to reproduce and demonstrate the issue, and
that your code change fixes it.

Ben, anything I missed?

Thanks.