rubber-duck

A fast second opinion on every edit Claude makes.

After each successful Edit, MultiEdit or Write, the change goes to a small, fast model (Haiku by default) with one question: is there a real problem here? It looks for removed error handling, a deleted or disabled test, a swallowed exception, a hardcoded secret or URL, an off-by-one, a changed public signature, or a TODO left in. If it finds one you get a toast:

🦆 Removed error handling around fetch (api.ts)

Otherwise it stays quiet. /duck opens a pane with the last 20 concerns (time, file, concern) and a Clear button.

It never blocks or slows the edit: the review runs after the tool call has returned, and if it fails, nothing happens (details go to the debug log). At most one review per file runs at a time; if Claude edits a file again meanwhile, only the latest edit gets reviewed next. Lockfiles, build output (dist/, node_modules/, ...), minified or generated files, and files over 200 KB are skipped.

Install

claude --plugin-dir /path/to/claude-mods/rubber-duck

Or add the repo as a marketplace and install it with /plugin.

Options

| Option | Default | What it does | | --- | --- | --- | | model | haiku | the review model: an alias or a full model id | | tellClaude | false | also pass new concerns to Claude with your next prompt, so it can fix them or call them false alarms |

How it works

A tool.call hook awaits next(e), takes the diff from the tool's own patch, and starts the review on a $.clock.after timer, which calls $.model.complete. Concerns live in $.state, which the /duck pane reads, and with tellClaude on a prompt.submit hook adds them to your next prompt as context.