Review Loop

Review, fix, re-review until a round finds nothing worth fixing. Fixes are new code — they need their own review.

Rules

1. Review in a fresh subagent every round. Invoke review-code with <target> --sub (default target diff). If review-code is missing, spawn a subagent to review for correctness, security, resource, and performance defects, rating each finding Yes / Judgment call / No for worth fixing. Never review your own fixes inline.

2. Use diff from round 2 for every target, including paths. GitHub does not see unpushed fixes, re-reviewing the whole codebase wastes context, and path-scoped reviews can miss fixes in related files. For pr <number>, run gh pr checkout <number> first.

3. Fix every Yes; fix a Judgment call only if Trivial or Small. Skip No. Do not re-fix or re-argue a skipped item when a later round raises it again.

4. Run the repo's checks after each round's fixes (make build / make test / make lint, else native commands). Fix breakage before the next review.

5. Stop when a round has nothing new to fix. Cap at 5 rounds; round 5 is review-only, even if it finds actionable issues. Do not edit after that review — hand the open findings to the user. If a fixed finding comes back in any round, stop fixing and report the open findings.

6. Commit each fix unless told otherwise. --commit each (default) — one commit per fixed finding. --commit end — one commit after the loop stops. Never push. On the default branch, branch before the first commit — otherwise diff anchors at HEAD and committed fixes drop out of the next review.

7. Report at the end: rounds run and why it stopped, fixes (path:line — what changed), skipped findings with a reason each, open findings, and check results.