What problem does it solve? Internal self-review tends to rubber-stamp its own work, and a single external review pass can produce false positives or truncated results that look like a clean pass. This Skill orchestrates an independent external LLM reviewer, verifies every finding against the actual code or documents, and loops until the review converges or a circuit breaker forces a decision. ## Core Features & Use Cases - External review orchestration: Dispatches review briefs to codex CLI, codex-mcp, or a fallback sub-agent with a four-tier degradation chain and explicit failure classification (quota exhaustion, truncation, protocol incompatibility). - Per-finding verification: Each external finding is independently confirmed, rejected with evidence, or marked partial, with severity ratings (P1/P2/P3) and concrete fix proposals. - Convergence tracking and circuit breakers: A weighted health score per round detects non-converging review loops, triggers trend warnings, and forces a retrospective with root-cause analysis after repeated stalls or a 5-round safety limit. - Use Case: After finishing a feature branch, run the skill to have an external model attack your uncommitted diff, then walk through each reported issue with evidence-based confirmation or rebuttal before merging. ## Quick Start Ask the AI to run an adversarial review on the current uncommitted changes and verify each finding before proposing fixes.
