Designing dynamic completions for each shell

[shannmu]

The current approach for dynamic completion is to write a glue script for each shell to call the COMPLETER implemented in Rust. The issue is that the built-in commands and variables differ between shells, making it difficult to pass consistent parameters to the COMPLETER via the glue script. For example, in bash, the COMP_WORDS variable is an array split by the characters in the COMP_WORDBREAKS variable, so --flag=value would be split into [--flag, =, value]. However, in Fish, they are not split. In summary, the behavior of each shell varies, and we aim to pass consistent parameters to the COMPLETER. If necessary, we might even need to obtain the entire command line and quote it.

There are already some command line argument parser projects:

• argcomplete
• cobra
• bpaf
• carapace
• argc-completions static completions
• inshellisense

How do they implement dynamic completion for various shells?

Related issues:
#1232
#3166

[shannmu]

This thread is for bash.

[shannmu]

argcomplete uses the same approach as us, a glue script for each shell to call external completer.
Its approach is to expose some custom environment variables: COMP_LINE, COMP_POINT, COMP_TYPE, _ARGCOMPLETE_COMP_WORDBREAKS, _ARGCOMPLETE, _ARGCOMPLETE_SHELL and _ARGCOMPLETE_SUPPRESS_SPACE, and it almost exposes the same environment variables for every shell. Then it has a lexer module that mainly uses the COMP_LINE, COMP_POINT, and _ARGCOMPLETE_COMP_WORDBREAKS environment variables to split the command line, which is then used for subsequent completion operations.

[epage]

So we've taken the approach to use COMP_WORDS and COMP_CWORD (array of words and index into that array) while they use a string with a byte(?) index into that string. My concern with that has been quoting. How do they deal with that?

Can you also link to the relevant sections for easier reference?

[epage]

Hmm, looks like they use a customized shlex so they can track absolute position as they deal with shell quoting
https://github.com/kislyuk/argcomplete/blob/6c9e540fdbdf331129a0100c3ed1a88730871634/argcomplete/lexers.py#L12-L14
https://github.com/kislyuk/argcomplete/blob/6c9e540fdbdf331129a0100c3ed1a88730871634/argcomplete/packages/_shlex.py#L251-L255

Shell quoting is shell specific. I expect most shells are similar to each other but I expect powershell to be very different.

I assume wordbreaks is needed because the shell expects us to only return values for what it thinks a word is. argcomplete integrates that into the lexer which looks a bit odd to me. My assumption is it'd be better to handle that in the shell-specific post-processing of completion results.

[shannmu]

Yes, I think so too. Shell quoting might be more troublesome for PowerShell, which is why I'm not sure if this approach is reasonable.

[shannmu]

This thread is for fish

[shannmu]

This thread is for zsh.

[shannmu]

This thread is for elvish

[shannmu]

This thread is for powershell.

[epage]

Meta thread

[shannmu]

There are too many scripts in their implementations. I think we don't need to consider them anymore.

• cobra
• argc-completions

[epage]

Could you clarify what you mean?

[shannmu]

In these two projects, the completions for each shell is implemented through many scripts. I think this doesn't match our original idea that we should implement the completion logic at a higher programming language level.

[epage]

I had missed that argc-completions is about code-genning more traditional static completions. I've crossed that off the list.

Looking at cobra, it seems like the generated bash code is still just glue code and would be relevant.

[shannmu]

Okay.