No comments

The agent that wrote the code will defend its comments. So spawn a reviewer that didn't write it, with one job: find comments that shouldn't exist. Then act on what it finds.

Scope

The files or diff the caller names. Otherwise the current diff against the base branch, working tree included.

Steps

1. Spawn the reviewer. One subagent with references/reviewer-prompt.md and the scope. It reports. It never edits application code.
2. Audit the report. Reject anything outside scope, any application-code edit, any deletion the keep list protects, any MUST KILL whose stated reason doesn't match the code. A keep survives only with proof it's about something we can't change. Check that scoped lint and type suppressions were covered. An ambiguous kill stays killed. A keep that's refuted or still ambiguous gets deleted. For a thin "IMPORTANT, do not remove", run how or why on the symbol before deciding. If the report is bad, rerun once with the failure named. A second bad report, stop and say so.
3. Fix the easy flags directly. Delete the dead path, drop the parameter, use the real API. If a flag needs a new shape, run architect once for the accepted set and stop at the sketch.
4. Fix the causes, in scope. Smallest root-cause fix. Remove every named workaround. If the cause is outside the diff, land the smallest in-scope fix and report the rest as open. Don't widen the fence.
5. Constraint comments. "Do not remove", "do not change wording", "talk to X first". Keep the ones about things we can't change. For the rest, offer the cheapest encoding: a type, a runtime check, a test, or a lint rule. If the human approves, encode then delete. If not, delete anyway and report the constraint as unenforced.
6. Report. Deletions, restored comments, reruns, any architect sketch, fixes, encodings offered and made, constraints left unenforced, other open work.