What problem does it solve? Managing AI code review on GitHub pull requests is repetitive and error-prone: you must classify the PR by branch hierarchy, request a Codex review, wait for results, triage P2+ findings, fix or defend them, and repeat until the latest head passes. This Skill turns that into a deterministic loop with explicit exit criteria. ## Core Features & Use Cases - Hierarchy-aware PR classification: Distinguishes layer-0 common changes from project-layer changes, validates target/base branches, and prevents direct commits to project main branches. - Codex review loop with timeouts: Calls @codex review, watches for eyes reactions, retries unacknowledged calls up to 3 times, enforces 15-minute response timeouts, and can delegate waiting to a review-waiter agent. - P2+ finding triage: Classifies Codex findings on the current head as fix-required, defensible (with recorded evidence), or user-decision-required, and resolves stale review threads with documented justification. - Fallback and handoff paths: When Codex review is not configured, records the fallback evidence and runs runtime health checks and silo runtime handoff for product-code PRs. - Use Case: After pushing a feature branch, ask the agent to open a PR and drive it to a Codex pass; the loop fixes actionable findings, defends justified ones with evidence, and only exits when the latest head is clean. ## Quick Start Ask the agent to create a pull request for the current branch and run the Codex review loop until the latest head has no remaining P2 or higher findings.
