What problem does it solve? Reviewing someone else's pull request in the mj-agent repository requires checking module boundaries, in-source canonical SKILL.md and prompt changes, biz_catalog drift, and SQL guardrail red lines—checks that are easy to miss manually. This Skill runs a structured five-stage audit so reviewers can decide whether a PR should be merged on architectural grounds. ## Core Features & Use Cases - Five-stage PR audit: Locates the PR via gh CLI, classifies changed files by mj-agent category, runs fixed checks (branch sync, commit convention) plus dynamically triggered checks (D1-D9) based on what changed. - mj-agent-specific risk lenses: Enforces mandatory stops for in-source SKILL.md canonical changes, system.md version bumps, biz_catalog drift against upstream schema, and SQL guardrail boundary expansion (ADR-006/009). - Human-in-the-loop publishing: Presents a severity-grouped review summary (Critical/Important/Suggestion) and asks before posting a comment to GitHub or merging, with double confirmation required for merges. - Use Case: A teammate asks "评审 PR #65". The Skill fetches the PR, detects it touches src/mj_agent/skills/**/SKILL.md, triggers the mandatory D3 canonical check, flags a missing eval_references update as Critical, and posts a structured review comment after your approval. ## Quick Start Ask the assistant to review PR #65 for architecture compliance in the mj-agent repository.
