A `def --env` completion does not update the environment when run as a completion

[drbrain]

Describe the bug

I want to cache completion results in $env but nushell does not allow this with dev --env.

Here is part of a completion for git diff tool --tool [COMPLETE-TOOL-HERE. bug-tools is the completion function for the --tool argument and it takes some time for git to respond. I want to cache the output in $env to make future completions fast.

# Show diff tools
export def --env bug_tools [] {
  if ("GIT_NU_DIFF_TOOLS" in $env) {
    $env.GIT_NU_DIFF_TOOLS
  } else {
    let tools = run-external "git" "difftool" "--tool-help"
      | parse -r "\t\t(?<value>.+?)\\s+(?<description>.*)"

    $env.GIT_NU_DIFF_TOOLS = $tools

    $tools
  }
}

# Show changes using common diff tools
export extern main [
  ...pathspec: path
  --tool(-t): string@bug_tools
]

How to reproduce

1. Save the above file to bug.nu
2. use bug.nu *
3. Note the time taken for bug --tool <TAB>
4. Note the time taken for bug --tool <TAB> (a second time)
5. Note the time taken for bug_tools
6. Note the time taken for bug_tools (a second time)

Expected behavior

The second invocations of bug --tool <TAB> should be significantly faster than the first run.

I expect invocation of bug --tool <TAB> to behave like this:

❯ use bug.nu *
❯ timeit bug_tools
1sec 706ms 167µs 417ns
❯ timeit bug_tools
72µs 125ns

Screenshots

No response

Configuration

key	value
version	0.93.0
major	0
minor	93
patch	0
branch
commit_hash
build_os	macos-aarch64
build_target	aarch64-apple-darwin
rust_version	rustc 1.77.2 (25ef9e3d8 2024-04-09) (Homebrew)
cargo_version	cargo 1.77.2
build_time	2024-04-30 22:51:13 +00:00
build_rust_channel	release
allocator	mimalloc
features	dataframe, default, sqlite, system-clipboard, trash, which
installed_plugins	dns, example, highlight, net

Additional context

No response

[kubouch]

This works as intended IMO. Running completions shouldn't have side effects. There could be an error if you attempt to use def --env as a completion function, though, to make it more clear.

You could try using the stor commands if you really need to store some data from the completion function.

[paaloeye]

A working example of using stor in the context of completion:

Yeah, Nushell doesn't cache, so anything that needs expensive recalculations (network requests which carapace does sometimes, for example) will be slow.

I decided to use stor for my own completions:

def all-packages [] {
	try {
		stor open | query db "select * from __comp_apk" | length
	} catch {
		stor create --table-name __comp_apk --columns {name: str, version: str}
		get-all-packages | stor insert --table-name __comp_apk
	}

	stor open | query db "select * from __comp_apk" | get name
}

Where get-all-packages is a slow command. One added benefit is that the cache persists between different completion invocations.

It's not optimal, however. This cache doesn't handle updates, for example. I feel like it'd be better to either have a more sophisticated cache implementation in std for use in completion scripts or a built-in completion caching mechanism.

Originally posted by @kaathewisegit in #11733