That's an interesting idea. Do you have a link to show how you did this?

The completions don't work sometimes (for instance you don't get completion for some commands when running them with sudo) and it doesn't fall back on providing file completions if it doesn't know what other completions to provide, which is a bit annoying, but it generally works and I'm sure there are ways to fix that too.

https://github.com/sigoden/argc-completions

https://github.com/microsoft/inshellisense

One design aspect I do question is that completions are baked into the library. It'd be neat to have them in SQLite or something that could be updated independently (not sure how feasible that is though. I've never been any good at querying graphs like flag/command completions in SQLite).

I'm curious where did you get that from.

Edit: I read the note in your dotfiles repo. Yes calling `compinit` on each shell invocation is going to be really slow. That's not how you're supposed to do it, you could at least add the `-C` flag to cache the completions. Ideally you'd also use `zcompile` to compile the cache to ZSH word code. This puts my completions initializing time at ~20ms on a lower/mid-end laptop. Additionally you can do the trick `fish` does and defer the initialization of completions until the first hit of Tab key, so the impact on shell startup time is exactly 0.

    autoload -U compinit && compinit
    {
      # Compile the completion dump to increase startup speed. Run in background.
      zcompdump="${ZDOTDIR:-$HOME}/.zcompdump"
      if [[ -s "$zcompdump" && (! -s "${zcompdump}.zwc" || "$zcompdump" -nt "${zcompdump}.zwc") ]]; then
        # if zcompdump file exists, and we don't have a compiled version or the
        # dump file is newer than the compiled file
        zcompile "$zcompdump"
      fi
    } &!

Do you have any pointers for the "load on tab" idea? I didn't turn up any good results in DDG and LLMs were just hallucinating.

BTW I believe `-C` will disable some cache checking, caching is enabled by default

> To speed up the running of compinit, it can be made to produce a dumped configuration that will be read in on future invocations; this is the default, but can be turned off by calling compinit with the option -D.

> ...

> ... The check performed to see if there are new functions can be omitted by giving the option -C. In this case the dump file will only be created if there isn't one already.

    {
      # load compinit and rebind ^I (tab) to expand-or-complete, then compile
      # completions as bytecode if needed.
      lazyload-compinit() {
        autoload -Uz compinit
        # compinit will automatically cache completions to ~/.zcompdump
        compinit
        bindkey "^I" expand-or-complete
        {
          zcompdump="${ZDOTDIR:-$HOME}/.zcompdump"
          # if zcompdump file exists, and we don't have a compiled version or the
          # dump file is newer than the compiled file, update the bytecode.
          if [[ -s "$zcompdump" && (! -s "${zcompdump}.zwc" || "$zcompdump" -nt "${zcompdump}.zwc") ]]; then
            zcompile "$zcompdump"
          fi
        } &!
        # pretend we called this directly, instead of the lazy loader
        zle expand-or-complete
      }
      # mark the function as a zle widget
      zle -N lazyload-compinit
      bindkey "^I" lazyload-compinit
    }

This is bash, not zsh, but I have this working in my dotfiles by just telling bash where to look for my custom on-demand completions:

https://github.com/NateEag/dotfiles/blob/6862726ad2ecaa3a30e...

I imagine something similar works for zsh.

Do you by chance have any good resources on optimising my config further?

You could peek at something like zprezto https://github.com/sorin-ionescu/prezto or pure https://github.com/sindresorhus/pure for tips.

Fetching git/hg/... info is always slow, so try and speed that up where you can (as to how to do that, uhh... I know my prompt has a dirty-state check nicked from pure for speed reasons). You can also cache any `asdf init zsh` or similar to a file and do the same "run in background" trick so the next shell will have any changes.

The biggest improvement I can remember was dropping zprezto for my own much smaller config, I really did not need much comparatively. Mostly some git info and "good default" options. I use zgenom for a plugin manager but only have 3 plugins, probably I should just dump it and inline the plugins to avoid getting owned one day.

You're right, my memory has let me down.

> Do you have any pointers for the "load on tab" idea? I didn't turn up any good results in DDG and LLMs were just hallucinating.

The simplest implementation would be something like

    bindkey ^I init_completions

    init_completions () {
      # ... init logic here ...

      # rebind tab to complete
      bindkey ^I complete-word
      # actually do complete the initial request
      zle complete-word
    }

The big gain from something like carapace or my theoretical SQLite-based completion system is faster startup time. I had to remove zsh-completions from my shell setup as it added too much to the startup time (https://github.com/bbkane/dotfiles/blob/master/zsh/README_no...)

To me it make sense that completion for program P should be handled by program P itself. That way, completions are unlikely to get out of sync with the application, and the completion handler can use the same option parser as the application. A way to do this is to use a special "hidden" switch to request completion.

Specifically the DomTerm terminal emulator (https://domterm.org) handles its own completions. Bash allows you to register a command that handles completions for some other command. The following tells bash that to handle completions for the domterm command it should call domterm with the magic "#complete-for-bash" option followed by the existing line and position.

    complete -o nospace -C 'domterm "#complete-for-bash" "$COMP_LINE" "$COMP_POINT"' domterm

I don't know how to do something similar for other shells; suggestions welcome. Some other shells (such as Fish) include help text as part of the suggested completions. Handling that would require a more complex protocol. Perhaps one that returs a result in JSON format?

If a such a protocol were standardized (perhaps invoked by a "#completions-as-json" option?), we just need to solve a final piece of the puzzle: How does a shell know if a command understands the "#completions-as-json" option? My suggestion is that we define a new keyword-value pair in the "desktop" file specification; the shell would search for a P.desktop file,

The concept is awesome but it's curious to see YAML at the core of this project, rather than a Turing-complete, embeddable language like Lua.

A specific example would be the Carapace (armor) upgrades in Starcraft 2 (one faction, the Zerg, are basically insects).

Also sometimes used for (humanoid) armor crafted from shells (fantasy settings).

I think this is MUCH more prevalent in scifi/fantasy targeted at male audiences, and the data confirms it [1]

[1]: https://osf.io/g4xrt/#!