code-review（代码审查）

Code Review

Expert code reviewer with automated pre-review checks (tests, linting, formatting) and optional auto-fix capabilities.

Parameters

The skill accepts different modes and options:

• mode: changes | target

- changes: Review uncommitted changes (git diff + status) - target: Review specific files/directories matching glob pattern

• target: File path or glob pattern (required if mode=target)

- Examples: src/auth, **/*.ts, lib/utils.ts

• intent: Optional description or file reference

- Plain text: "add OAuth2 authentication" - File reference (example): @docs/requirements.md (placeholder - replace with actual requirements file) - Used for Requirements Alignment section

• save: Boolean flag (--save)

- If true: write review to docs/CODE_REVIEW.md - Default: show in chat only

Process Workflow

1. Gather Code

changes mode:

git status
git diff HEAD

target mode:

• Use Glob to find files matching target pattern
• Read matched files
• Exit gracefully if nothing found

2. Load Intent (if provided)

• If starts with @: Read file at path
• Otherwise: use string as-is
• Store for Requirements Alignment section

3. Pre-Review Checks

Run scripts/pre-review.sh to check tests, linting, formatting:

bash scripts/pre-review.sh

Exit codes (bitwise OR):

• 0: All passed → proceed to review
• 1: Lint failed → auto-fix
• 2: Format failed → auto-fix
• 4: Tests failed → auto-fix
• 8: No package.json → skip to review

Handle failures:

Exit code 8 (no package.json):

• Skip all checks
• Proceed directly to review

Exit code 1 (lint failed):

• Spawn Task with fix-eslint skill
• Wait for completion
• If auto-fix fails or times out, note the failure and proceed to review anyway
• Re-run pre-review.sh once. If still failing after the second run, proceed to review and note failures in output

Exit code 2 (format failed):

• Run pnpm format directly
• Re-run pre-review.sh once. If still failing after the second run, proceed to review and note failures in output

Exit code 4 (tests failed):

• Spawn Task: "Fix failing tests - review test output and fix the code to make tests pass"
• Wait for completion
• If auto-fix fails or times out, note the failure and proceed to review anyway
• Re-run pre-review.sh once. If still failing after the second run, proceed to review and note failures in output

Multiple failures (e.g., 7 = lint + format + tests):

• Handle in sequence: format first, then lint, then tests
• Re-run checks after each fix

Max attempts:

• Limit to 2 full pre-review cycles (1 initial run + 1 retry)
• If checks still fail after second run, proceed to review anyway
• Note failures in review output

4. Execute Review

Follow methodology.md for complete review process and output format.

Process:

1. Scan for critical safety/security issues
2. Verify tests & edge cases
3. If intent provided: validate implementation accomplishes goals
4. Note improvements & positives
5. Summarize decision with next steps

Output format: Use exact headings from methodology.md:

• Critical Issues - with line numbers (L42, L42-47, file.ts:42)
• Functional Gaps - missing tests/handling
• Requirements Alignment - only if intent provided
• Improvements Suggested
• Positive Observations
• Overall Assessment - Approve | Request Changes | Comment Only

5. Present Review

Always show review in chat, formatted with markdown headings.

6. Save (if requested)

If save=true:

• Use Write tool to create/overwrite docs/CODE_REVIEW.md
• Include all sections with exact headings
• Preserve markdown formatting

Examples

Review uncommitted changes:

mode: changes
intent: null
save: false

Review with intent:

mode: changes
intent: "Implement OAuth2 authentication with secure token storage"
save: false

Review specific files:

mode: target
target: "src/auth/**/*.ts"
intent: "security review"
save: true

Review with file-based requirements:

mode: changes
intent: "@docs/requirements/auth.md"  # Example: replace with actual requirements file path
save: true

Notes

• Use this skill immediately after writing or modifying code
• Gracefully handles projects without package.json (skips checks)
• Auto-fixes: delegates to fix-eslint skill for lint errors, runs pnpm format for format errors
• Intent can be plain text or file reference with @ prefix
• Re-runs checks after auto-fixes, max 2 cycles
• Pre-review script (scripts/pre-review.sh) returns bitwise exit codes: 1=lint, 2=format, 4=test, 8=no package.json