review-local（评论本地）

Local Review

Run the local AI review convergence loop on the current branch. This skill dispatches domain-specific reviewers against your code, produces a structured review summary, and runs a bounded fix convergence loop — all without GitHub or a PR.

Use this as a quality checkpoint anytime you have code on a feature branch: before pushing, mid-development, or as a standalone review outside /ship.

How to invoke

Do not over-prepare. When the user asks you to run /review-local, run the script. Do not create branches, stage files, or commit on behalf of the user — the script works on the current branch state (committed + staged + unstaged).

The script path is relative to the /ship skill directory (this skill shares scripts with /ship): plugins/eng/skills/ship/scripts/run-local-review.sh

Always run with run_in_background: true — the review dispatches 17 parallel reviewers and runs fix passes, routinely exceeding the 600-second Bash timeout:

Bash(command: "<absolute-path-to-skill>/scripts/run-local-review.sh [options]",
     run_in_background: true,
     description: "Local review")

Use the absolute script path. Do not rely on CWD being correct — your shell CWD may differ from the target repo. The script resolves the repo root via git rev-parse --show-toplevel from the CWD where it is invoked, so if your CWD is a different repo, the review will run against the wrong codebase. Always invoke with the full absolute path and ensure you are in the correct repo directory.

What it runs

The script:

1. Stages the portable PR review bundle into tmp/ship/pr-review-plugin/
2. Dispatches 17 parallel domain-specific reviewers (via the Agent tool in their own sessions)
3. Aggregates findings into a structured review summary
4. Runs a bounded fix convergence loop when findings are blocking

Options

Output files

All output goes to tmp/ship/:

Standalone usage (no /ship, no state.json)

The script works without any /ship artifacts:

• Creates tmp/ship/ if missing
• Detects quality gates from package.json when no state.json exists (looks for test, typecheck, lint scripts)
• Gracefully handles missing /ship-specific files (spec.json, progress.txt, codebase-context.md — all produce _Not available._ in prompts)
• Assessment protocol still loads from the staged review bundle

Examples

# Default: review + auto-fix loop (up to 5 passes)
<ship-skill-base>/scripts/run-local-review.sh

# Review against a specific branch
<ship-skill-base>/scripts/run-local-review.sh --target develop

# Review + fix loop (up to 2 passes)
<ship-skill-base>/scripts/run-local-review.sh --max-fix-passes 2

# Run inside Docker sandbox
<ship-skill-base>/scripts/run-local-review.sh --docker

Interpreting results

The script emits a structured return payload to stdout at exit with three delimited sections:

1. Exit envelope (=== LOCAL REVIEW EXIT ===): Always present. Contains exit_code, exit_reason, pass counts, fix commit SHAs, last_recommendation, blocking status, duration, and file pointers.
2. Review status (=== REVIEW STATUS ===): Parsed review-status.json content — recommendation, risk, issue counts, blocking reasons. Present on all non-crash exits.
3. Iteration log (=== REVIEW ITERATION LOG ===): Full history of review passes and fix responses. Only included on non-zero exits (blocking/fatal) to provide remediation context.

Exit reasons:

You can also check review-status.json directly for the gate status:

• APPROVE: Review passed — gate is green, script exits 0.
• APPROVE WITH SUGGESTIONS: Suggestions remain. The script treats this as not-yet-converged and will attempt fix passes. Only pure APPROVE exits green.
• REQUEST_CHANGES with Critical/Major findings: Review is blocking.

Do not blindly apply every suggestion. Validate each finding against the diff, codebase patterns, and your intent before acting.

If you are assessing review findings yourself (e.g., the automated fix loop exited non-zero and you are triaging the remaining findings manually), use the Skill tool to invoke /assess-findings before assessing any finding. Do not evaluate findings without it.

Persisting declined findings

After assessment completes (whether by the automated fix loop or manual triage), append a deferredFindings array to review-status.json containing findings from the /assess-findings Declined Findings Summary (Phase 7) where Future-relevant: Yes. Each entry:

{
  "finding": "writeTracker coupling across modules",
  "classification": "Pre-existing, out of scope",
  "evidence": "Valid encapsulation concern, not caused by this PR"
}

If no future-relevant declined findings exist, set "deferredFindings": []. This array is consumed by /ship, which persists items to state.json's deferredScope[] and surfaces them in the Ship Summary.

Relationship to /ship

/ship invokes the same run-local-review.sh script directly in Phase 4: Review gate — pre-QA (/review-local) and Phase 6: Review gate — post-QA (/review-local). /ship does not load this skill — it calls the script as an orchestrator. This skill exists for standalone human invocation outside /ship.