Closes #18.

Only the prompt fix action was implemented; a command action rendered nothing. This runs the script per docs/guide.md: the smell's whole bag as JSON on stdin, once per smell, stdout/stderr streamed to the agent, exit code folded into the run's exit code.

Locked decisions

• 18.1 (spawn/timeout): an unrunnable fixer is a blocking config error — it hard-fails the run (exit 1) regardless of severity, with a clear notice in the agent output. Consistent with Sensor spawn/timeout failure should fail the run (exit 1), not produce a false-clean #25's sensor policy.
• 18.2 (zero-issue smells): the script runs only when the smell actually has issues.

Exit folding per the docs table: command exit 0 never blocks (even for an enforced smell — it overrides the default "enforced + issues = 1" rule); non-zero blocks only an enforced smell. prompt behaviour is unchanged.

Atomic commits

1. #18: let runTool write stdin to the spawned process — optional stdin with an EPIPE guard; existing callers unaffected.
2. #18: add the command fix runner — runFixCommand + unit tests (exit 0 / non-zero / enforced vs suggested / spawn-failure).
3. #18: run command fix actions from the guide — guide becomes async, runs each command action, streams output into the section, folds the exit code; runner threads cwd and awaits.

Gates

• pnpm build, pnpm lint — clean (all functions within max-lines/max-params)
• pnpm test — 392 passed (+7)

🤖 Generated with Claude Code

Generated by Claude Code