What problem does it solve? After implementing a plan, code often contains logic defects, scope drift, and missed edge cases that a single review pass misses. This Skill runs a bounded review-fix-verify-commit loop that repeatedly checks implementation against the plan until quality converges, without scope creep. ## Core Features & Use Cases - Multi-persona review dispatch: First cycle fans out parallel reviewer personas (correctness, scope-guardian, security, adversarial, coherence) via subagents, with a local sequential fallback when subagent tools are unavailable. - Plan-bounded fixing: Findings are deduplicated, severity-classified (critical/high/medium/low), and fixed surgically; out-of-scope improvements are accumulated as Notes instead of being auto-fixed. - Convergence-based termination: The loop stops after two consecutive cycles with no material findings, runs a full test suite before finishing, and reports residual risks plus accumulated Notes. - Use Case: After finishing a feature branch, ask the AI to run the feedback loop; it reviews the diff against the plan, fixes a race condition and an async contract mismatch, verifies with targeted tests, commits each cycle, and stops when two clean reviews pass. ## Quick Start Run the self-feedback loop on my current implementation against the latest plan in docs/plans, fixing and committing any findings until the review comes back clean.
