On Arch Linux, jre-openjdk-headless creates a symlink from /usr/lib/jvm/java-20-openjdk/manto /usr/share/man/ therefore the command fish_update_completions finds all man pages twice. The first time it can generate the completions, but with the second file it ignores all commands because they are already added to already_output_completions.

The simplest way to reproduce this is to pass the same manpage file twice to create_manpage_completions.py.

In most cases this is no issue and it just ignores the last file, because it fails to generate the completions.

But for The Silver Searcher (ag) the second pass falls back to the TypeDeroffManParser and is able to generate a different, but incomplete, completion file and overwrites the previous one.

Reproducing the errors:

• Ignores failure of second file:
  • python share/tools/create_manpage_completions.py --progress /usr/share/man/man1/grep.1.gz /usr/share/man/man1/grep.1.gz --stdout
• Creates two different completion files:
  • python share/tools/create_manpage_completions.py --progress /usr/share/man/man1/ag.1.gz /usr/share/man/man1/ag.1.gz --stdout

I'd be happy to create a pull request to fix this but I see multiple options:

1. Skip generating the completion if the command was already created, unlike already_output_completions which is done per argument
2. Reset already_output_completions at the start of parse_manpage_at_path
3. Add all values in already_output_completions to the built_command_output to ensure that all options are written to the completions file.
4. Make the list of manpage files unique by using their checksums

I'm not sure how those workarounds would change the behaviour mentioned in

System Information

• fish --version: fish, version 3.6.1
• uname -a: Linux ng-lenovo-2 5.15.110-2-MANJARO #1 SMP PREEMPT Thu May 4 06:51:53 UTC 2023 x86_64 GNU/Linux
• echo $TERM: xterm-kitty