Code Review
Scope
Run a bounded independent implementation review. This is not a per-task reviewer pipeline and never starts automatically from risk, difficulty, or change size.
Trigger
The only trigger is that the user explicitly requests implementation review. Security impact, compatibility risk, complexity, file count, line count, or the agent's own confidence must never substitute for that request. Without it, return to the root agent's normal work without invoking this Skill.
Review Loops
- The root agent identifies the exact task requirements and review scope from the conversation, plan, and current Git evidence.
- Spawn exactly one fresh reviewer for round one. Give it the requirements, scope, diff or commit evidence, and review rubric. It must be read-only and must not delegate to another agent.
- Round one is a full review of the requested implementation scope. Report findings first, ordered by severity and grounded in files and lines.
- If fixes are required, the root agent owns the fixes, integration, and tests.
- After blocking findings are fixed, spawn one fresh reviewer for a scoped re-review of those fixes and nearby regressions. Include the prior findings and the fix scope.
- Never start a third round. Report unresolved critical or major findings as blockers without claiming completion.
Minor observations may accompany a pass. A critical or major finding blocks a pass even if the reviewer labels its overall verdict inconsistently.
Review Handoff
Record the round, scope, review range, verdict, and blocking findings. Then return control to the root agent.