review（审查）

Review

Purpose

Use this as the entrypoint for local code review. Assume the implementation is wrong until it survives scrutiny. Default to finding concrete bugs, weak reasoning, unproven assumptions, and missing coverage over giving praise.

Routing Rules

• Do not read this handbook linearly.
• Pick exactly one review type first.
• Default to robustness.md when the user asks for a generic review.
• If the user wants multiple review types, handle each review type separately. When subagents are available and explicitly authorized, split the types across parallel agents; otherwise review them one at a time. For formal multi-type reviews, write one review file per type.
• Follow the shared rules here plus exactly one type-specific file.
• If the user asks to review against commits, plans, files, or commands, capture that in ## Scope; do not create a special filename format for it.
• If the first type-specific file is insufficient, only then inspect another review type for overlap.

Workflow

1. Determine the review scope from the user's request: files, diff, commits, plan, command output, or changed behavior.
2. Route to exactly one review type, then read this file plus the chosen type-specific file.
3. Inspect the target and relevant surrounding code, tests, docs, configs, and history needed to evaluate the requested scope.
4. Stress the highest-risk assumptions for that review type and keep only findings backed by concrete observed or strongly inferred evidence.
5. Report findings in severity order. Use chat for informal or lightweight requests; write a standalone review artifact for formal reviews, multi-type reviews, or when durable review context is requested or clearly useful.

Quick Picks (Task -> File)

• Generic local code review -> robustness.md
• Backwards compatibility / change safety -> stability.md
• Security audit -> security.md
• Performance review -> performance.md
• Docs review -> docs.md
• Automated test review -> automatic-testing.md

Review Types

• robustness.md - correctness, failure handling, maintainability, coupling, developer experience, and overall implementation quality
• stability.md - backwards compatibility, contract drift, migrations, rollout safety, and user-visible behavior changes
• security.md - trust boundaries, auth, input handling, secret exposure, and exploitability
• performance.md - latency, throughput, memory, concurrency, unnecessary work, and hot-path regressions
• docs.md - incorrect, missing, stale, or incomplete documentation and upgrade guidance
• automatic-testing.md - broken tests, weak tests, and missing automated coverage for changed behavior

Reviewer Posture

• Be professionally adversarial. Treat claims, abstractions, tests, and comments as suspect until verified.
• Prefer concrete breakage, missing evidence, and incorrect assumptions over style feedback.
• Distinguish observed evidence from inference. Report inferred concerns only when the reasoning is strong enough to survive challenge.
• Do not pad the review with praise, reassurance, or generic quality language.
• If there are no findings, say so explicitly under ## Findings.

Announce at start: "I'm using the review skill to perform a <review-type> review."

Save reviews to: <repository-root>/reviews/YYYY-MM-DD_HH:MM_<review-type>_<review-name>.md

Create reviews/ if it does not exist when writing an artifact.

Review Structure

# Review: <Review Name>

## Review Type
<robustness | stability | security | performance | docs | automatic-testing>

## Scope
<Standalone description of what was reviewed. Include commits, plans, files,
commands, constraints, and assumptions when they materially affect the review.>

## Findings
### <High|Medium|Low>: <Short issue title>
**Category:** <specific category within the review type>
**Impact:** <why this matters>
**Location:** `path/to/file.ts:42`, `path/to/other.ts:10`
**Evidence:** <Observed|Inferred>. <brief proof: code path, command result,
log, failing test, type error, missing assertion, docs drift, etc.>
**Suggested Fix:** <concrete repair or next step; use `None` if no responsible
fix is clear yet>

Review Rules

• Formal review artifacts MUST be standalone; do not rely on the prior chat.
• For informal chat reviews, keep the same evidence standard and severity ordering, but do not write a file unless durable context is warranted.
• ## Scope MUST restate the exact review target and any constraints needed to understand the findings.
• Order findings by severity: High, then Medium, then Low.
• Every finding MUST include Category, Impact, Location, Evidence, and Suggested Fix.
• Evidence MUST explicitly say Observed or Inferred.
• Use exact file paths and line numbers when the location is known.
• Do not invent evidence. If a suspicion is weak, drop it.
• Use only the sections shown in ## Review Structure.
• If no findings remain after review, write No findings. under ## Findings.

Examples

• "Review this change for robustness."
• "Review commit abc123 for stability."
• "Audit these auth changes for security."
• "Review the docs changes in docs/."
• "Review the automated tests for this feature."