Discussion: Reduce dependency footprint

[outslept]

1. gunshi

Note

I wasn't around when the "core" deps were chosen. So I might be missing something on this.

gunshi is effectively only used to route 2 subcommands and call renderUsage; a whole framework looks like an overkill. if we want to change this there are 2 options:

• Option A: Replace with mri + a tiny manual dispatcher with lazy imports - smallest footprint and full control, but help/usage must be maintained manually
• Option B: Replace with a lightweight CLI library (e.g. sade) - keeps auto help and per‑command option parsing with only a small increase in dependency size (in the sade example, it weighs ~2.3 times less than gunshi)

2. module-replacements

I might be crazy, but.. We can host those three JSON manifests on "a registry" (like shadcn’s /r/.json concept, for example - https://ui.shadcn.com/r) and have the CLI fetch and cache them, so we avoid installing the whole package into node_modules as a result. This should be a lighter process, than a regular npm installation

3. module-replacements-codemods

Probably should open an issue upstream to track this. This is the biggest size eater rn. module-replacements-codemods currently pulls @ast-grep/napi, which brings prebuilt native bindings and adds a large install footprint (~80MB observed). Given that the repository only uses the ast-grep API in fewer than 10 places and a lot of the existing codemod base is jscodeshift-based, shipping those binaries by default feels disproportionate. Can we settle on one option or find a workaround there?

4. package-manager-detector

Check around process.env.npm_config_user_agent should be sufficient for our use case. It's not a lot of code to implement: https://github.com/antfu-collective/package-manager-detector/blob/main/src/detect.ts

5. premove

Since this is a dev dependency, it has the lowest priority. However, given that the test script already uses FORCE_COLOR=1 (which doesn't work properly on Windows without the tweaks), we could consider replacing premove with native rm -rf to eliminate this dependency entirely.

Branch - https://github.com/e18e/cli/compare/main...outslept:cli:replace-premove?expand=1

[43081j]

args parser / usage gen

i'm still a fan of gunshi but im not opposed to us dropping it in this particular project since we're not really making heavy use of its features

i've used sade before and that seems good

codemods

The ast-grep codemods are far better than the jscodeshift ones, much easier to maintain and better at retaining syntax (since they use tree-sitter), so i'd prefer the codemods repo to settle on that but im not sure that'll happen anytime soon

we should look into if we can help make ast-grep's JS bindings smaller. i'd be surprised if we couldn't

rest

for all the other suggestions i think some or all make sense but we can decide those once these bigger savings have been decided

[outslept]

I've opened the issue upstream. Preparing draft branches for dependency changes now

[43081j]

coming back to this:

i think we should prompt the user to install the optional dependencies if they configure the CLI in a way that requires them

e.g. ATTW should be off by default. if i turn it on, the CLI should warn and prompt me to install it as a dependency before running

[outslept]

nuxt cli does something similar. I love the concept!