Auto Code Review

Auto Code Review

Mandatory Entry Point

When this skill is triggered, you MUST read references/auto_code_review.md in full and execute according to its rules.

• Do NOT substitute the full specification with preambles, Cursor rule summaries, or other secondary summaries.
• Do NOT probe reviewer CLIs, invoke reviewers, or create review archives without explicit authorization in the current request.
• Runtime prerequisites (not distributed with the skill sync package; must be provided by the host environment): env/review.json (template: env/review.json.example), in-project .auto-review-config.json, and AUTO_REVIEW_* environment variables. See AGENT-BRIEF.md and docs/auto-code-review.md for configuration loading priority and field semantics.

Nine Core Rules

• [ACR-001] Explicit Authorization Gate: The skill is entered ONLY when the user explicitly triggers it; completion of code changes is NOT a trigger. Configuration can only control capability availability — it cannot represent authorization for the current request.
• [ACR-002] Traceable Scope: Prefer reviewing changes precisely recorded in the current request. When the boundary cannot be proven, ask the user to choose staged or worktree. Do NOT present git diff HEAD as "this round's changes".
• [ACR-003] Reviewer Read-Only: The reviewer ALWAYS runs in read-only mode, outputting review comments without modifying files.
• [ACR-004] Layered Write Permissions: Default is review-only; the main agent only triages and reports. Only when the user explicitly specifies --fix or "review and fix" may the main agent apply fixes and re-review.
• [ACR-005] MAX_ROUNDS=3: review-only runs exactly one round; review-and-fix runs at most 3 rounds. On non-convergence, output deadlock — do NOT fake a pass.
• [ACR-006] Post-Authorization Closed Loop: After explicit trigger, execute review → archive → sync → merge. History recall is now handled by the global historical-recall skill, so this skill no longer recalls inline. Archives are written to .plan-reviews/ and belong only to the authorized review session.
• [ACR-007] Configurable Reviewer: Reviewer, rounds, and single-model fallback are all configurable. AUTO_REVIEW_ENABLED=false is the capability-level disable switch; true does NOT constitute user authorization.
• [ACR-008] Single-Model Fallback Requires Explicit Permission: Same-model self-review is NOT performed by default. Fallback occurs only when explicitly allowed by configuration, and logs must note reduced credibility.
• [ACR-009] Execution Package and Quorum Proof: Before invoking reviewers, create one shared review package. Each round MUST record selected reviewers, expected reviewer count, raw output paths, legal verdicts, and pass/fail reasons. A round passes only when every selected reviewer returns a legal standalone VERDICT: APPROVED; missing reviewers, missing raw output, timeout, or invalid verdict all fail closed. In review-only mode, do NOT declare the gate passed; report "reviewers approved, no code changes made" instead.

Modes

• /auto-review: Read-only review; no workspace modifications.
• /auto-review --fix: Review, main agent fixes adopted issues, then re-review.
• Normal implementation requests: do NOT trigger this skill.

Relationship with Adjacent Skills

Workflow

Implementation complete → User explicitly triggers → Select scope/mode → Reviewer read-only review
                                                         ├─ review-only: report and archive
                                                         └─ review-and-fix: fix → re-review → archive