Author here. This blog post is the result of my puzzling over how auto-complete works in the shell, and subsequently whether shell completion scripts could be written in Rust. It also introduces a new Rust crate, shell_completion[0], which is a library that exposes some low level primitives for writing shell completion scripts in Rust. shell_completion is certainly in its infancy, so I'd be happy to have contributors of any kind (both code and opinions are welcome).
These [0] two [1] sections of the GNU Bash man pages were very useful. I also found that this tutorial [2] provided a useful perspective. I spent a small amount of time looking through the bash-completion [3] project as well as the git bash completions [4].
Generally, my process for writing is google around until I have some base level of knowledge, then building one or more demo applications to ensure my understanding is correct. Then as I'm writing the blog post I'll be refining the demo applications into examples for the blog post and/or library code that I want to open source.
I've actually written two separate articles on shell completion, explicitly for subcommands, because the resources regarding subcommands is so lacking.
I need to improve them and write a bit more information, I was writing a library to generate completion scripts for both bash and zsh, so I had to know the differences quite well.
That sounds useful. I used to write tons of ZSH scripts when I used to obsessively make my Linux desktop experience nice and automated. Writing custom completions always felt awkward for me and unintuitive (like most ZSH programming).
Does your shell_completion library work with ZSH? I'm not familiar with how sh/bash differs here from ZSH but I'd image it'd be easy to make compatible.
I'm waiting patiently for Evlish to mature so I can use their Go-like (and Go-powered) scripting language and burn my old ZSH scripts in a fire. https://github.com/elves/elvish
I talked with the author of Elvish about it, and we both want it to happen, but are both working on other things now AFAIK.
If you managed to write a working bash/zsh completion script, then kudos :) It's not easy. The basic idea is that shell completions could be both (1) shared among shells and (2) moved into the same language that the binary is written in (optionally).
Seeing this got me curious to know if any work has been done on generating completions for applications that make use of the clap command-line argument parser.
I did come across this work by clap while doing research for this post. The major difference in strategy is that clap is generating completion scripts in Bash, while shell_completion is a library which helps writing completion scripts in Rust. I think one exciting bit of future potential for shell_completion is it could potentially allow building something like what clap has done using the standard Rust macro system.
I actually wrote a library that generates completions for both zsh and bash for the Ruby Thor CLI library. https://github.com/snowe2010/fylla/
This could be useful in research to make it work for clap.
I wrote two articles that kind of highlight the differences in completions between the two, but only in regards to subcommands, since that's what I was trying to cover.
If I'm understanding correctly, this is a shim that:
1) serializes the fields the bash completion API gives you (when you hit TAB)
2) exports them to Rust, then
3) passes the Rust program's results back to bash?
I've been working on similar problems, and I understand why that's appealing. Bash is a pretty bad language to write completion scripts in! It's really bad at string and array processing [1].
----
Though on a quick glance, if you're exporting $COMP_LINE and $COMP_POINT, then it's making the job of the completion script very hard, since you're passing off the responsibility of parsing shell to it:
In other words, if you have something like:
mycommand --flag arg\ with\ spaces --flag2=<TAB>
Then it's not desirable to pass the entire string. It would be better to parse it and pass:
["mycommand", "--flag", "arg with spaces", "--flag2="] # and indicate index 3 being completed
Another common case is:
echo $(dirname <TAB> # I don't need echo $( to complete
I noticed this deficiency of bash while essentially cloning the bash API for Oil: https://www.oilshell.org/
One consequence of this is that the bash-completion project (maintained separately from bash) actually has an ad hoc bash parser in bash! This is really bad -- unsurprisingly it has obvious bugs, like not being able to complete inside command subs. An ad hoc bash parser in Rust would also be bad.
So instead Oil parses shell for you -- since it has a shell parser! -- and then it gives you the partial argv array. So then the plugin is only responsible for what it can reasonably do.
-----
Additional, I proposed that this could be made into a protocol that other shells consume. So instead of writing bash completions in Rust, you could write zsh/fish/bash/Oil completions in Rust.
A lot of people are already "scraping" bash completions this way, i.e. to use in zsh or their own shells, so it should not be hard to come up with something "standard", given those working examples.
I wouldn't say it's the #1 priority now because Oil already emulates bash quite well, but I think a bunch of people are motivated because they 1) don't want to write completions N times, for every shell and 2) Don't want to write completions in the bash language.
Your characterizations are spot on. One major potential improvement for the shell_completion crate is to provide a higher level API for use by completion script writers. This will of course make writing completion scripts easier, but it will also allow for abstraction over various shells.
I would have gone with an out-of-band way to signal that the application supports a completion mechanism;
One way would be to add a new section to the binary (though that would only work with binaries), that could contain standardized data, such as a json tree of the possible completions, or the proper syntax to use the executable as a "completion server" as in your proposal.
In shell scripts, it could be one of the header lines starting with a magic comment such as "#!completion: shellac_v0.1
IMO having the shell detect that the executable supports a standardized completion format gets you halfway there ;)
Yes, the detection is definitely an important part of it that I've idly thought about, but haven't written anything about. The proposal is still very early, and not implemented anywhere.
The Github wiki is open to edit if you want to write up a proposal (on a new page) and discuss it on https://oilshell.zulipchat.com/ . A prototype is even better.
Although I think that perhaps running the binary with `-h` or `--help` and an env variable like SHELLAC_FEATURE_DETECT=1 might be a more backward compatible detection mechanism. If the binary doesn't support shellac, then it will probably exit quickly with help, or an error message. (And shells could cache this detection info.)
If it does support Shellac, then it will print some kind of "spec" to stdout describing where the completion is. Is it in the binary itself written in Rust/Python/etc., or is it in a separate completion script ?
The conversation has been dormant lately, but like I said I think there are a few groups of people motivated to do this (e.g. alternative shell users who don't want to rewrite completions themselves).
These shouldn't be taken for granted, I think the feature should be opt-in (and quite easy to contribute).
I would have executed the binary if there is a segment saying that it supports Shellac. The binary then detects the environment variable that's in the current spec, and prints completion information to stdout.
I did this https://github.com/xliiv/fui
This is something different, but maybe someone could say it's sort of "completions"?
Fui also has a feature which allow to copy/dump a whole form (ctrl-k) string which can be pasted to bash.
It is not possible for a shell to know how to perform completions for every binary on a users machine. Shell completion scripts define the auto-completion for each specific binary, so for autocomplete to work (well) it needs to be supported both by the shell and by the application. This post is about writing shell completion scripts (the application side, not the shell side).
[0]: https://github.com/joshmcguigan/shell_completion