review-changes（审查变更）

Review Changes

Review code changes in a feature branch and identify issues before merging.

Workflow

1. Determine Review Target

• Remote PR: If user provides PR number or URL (e.g., "review PR #123", "review https://github.com/org/repo/pull/123"):

1. Checkout the PR: gh pr checkout <PR_NUMBER> 2. Read PR context: gh pr view <PR_NUMBER> --json title,body,comments 3. Use the PR description and comments as additional context for the review

• Local Changes: If no PR specified, review current branch against default branch (continue to step 2)

1.5 Load Repository Guidelines

Search for repository-specific coding guidelines. These take precedence over built-in guidelines.

Discovery order (highest to lowest priority):

1. CLAUDE.md, .claude/CLAUDE.md
2. CODE_GUIDELINES.md, .github/CODE_GUIDELINES.md, docs/CODE_GUIDELINES.md
3. STYLE_GUIDE.md, .github/STYLE_GUIDE.md, docs/STYLE_GUIDE.md
4. CONTRIBUTING.md, .github/CONTRIBUTING.md, docs/CONTRIBUTING.md

Extract review-relevant content (coding standards, error handling, testing, security, naming conventions). Skip non-review content (issue templates, code of conduct). User guidelines take precedence over built-in guidelines and are additive. When reviewing a remote PR, load guidelines from the remote repository.

2. Detect Branches

# Get current branch
git branch --show-current

# Detect default branch (try remote HEAD, fall back to main)
git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's@^refs/remotes/origin/@@' || echo "main"

3. Get Changed Files

# Assess scope first
git diff --stat <default-branch>...HEAD

# Get full diff
git diff <default-branch>...HEAD

Performance guardrails:

1. Skip lock files, .min.js, .min.css, generated files, compiled output
2. If >30 changed files, prioritize source over tests/config; summarize skipped files
3. If a single file has >500 lines changed, summarize rather than line-by-line review
4. If total diff >3000 lines, select the ~20 most important files and summarize the rest

4. Analyze Changes

Use the diff as the primary input. Apply guidelines holistically across the diff rather than file-by-file. Only read full files when surrounding context is needed to understand a change.

Review process:

1. Identify applicable guidelines: repository-specific (from step 1.5), general (below), and language-specific (from reference files)
2. Check for critical issues first — report before scanning for minor ones
3. For each issue found, record: guideline violated, file and line number, problem description, fix suggestion
4. Skip inapplicable guidelines (no database operations → skip Database & Persistence)

5. Check Test Coverage

For each new or modified file containing business logic:

• Check if corresponding test file exists
• If tests exist, verify new code paths have coverage
• Flag missing tests for critical paths

6. Format Output

Present findings grouped by severity, ordered Critical → Major → Minor.

## Branch Review: `feature/xyz` → `main`

**Repository guidelines loaded:**
- `.github/CONTRIBUTING.md` - coding standards, testing requirements

*(These take precedence over built-in rules where they conflict)*

### 🔴 Critical (X issues)

**1. [Brief title]**
- **File**: `path/to/file.ts:42`
- **Problem**: Clear description of what's wrong
- **Fix**: Specific solution

### 🟠 Major (X issues)

**1. [Brief title]** `[repo]`
- **File**: `path/to/file.ts:87`
- **Problem**: Description
- **Fix**: Solution

### 🟡 Minor (X issues)

**1. [Brief title]**
- **File**: `path/to/file.ts:123`
- **Problem**: Description
- **Fix**: Solution

---
**Summary**: X critical, Y major, Z minor issues found.
[Ready to merge / Needs fixes before merge]

If no issues found in a category, omit that section. End with clear merge recommendation. Issues marked [repo] were flagged based on repository-specific guidelines. If no repository guidelines were loaded, omit the "Repository guidelines loaded" section.

Feedback tone: Be constructive — explain *why* a change is needed. Provide actionable suggestions. Assume positive intent. For approvals, acknowledge the specific value of the contribution.

General Guidelines

These apply to every review regardless of language.

API & Breaking Changes

• Removed or renamed public functions/methods without deprecation period
• Changed function signatures (new required parameters, changed return types)
• Modified response shapes in API endpoints (removed fields, changed types)
• Changed default values that alter existing behavior
• Database schema changes without migration scripts

Authentication & Authorization (Critical)

• Missing auth checks on new endpoints or routes
• Downgraded permissions (admin-only → public)
• Hardcoded credentials or API keys
• JWT/session issues: missing expiry, weak secrets, improper validation
• CORS misconfigurations: overly permissive origins (* in production)

Database & Persistence

• Missing transactions for multi-step atomic operations
• N+1 query patterns: fetching related data in loops instead of joins/eager loading
• Missing indexes on frequently queried columns
• Unbounded queries: SELECT * without LIMIT on large tables
• SQL injection: string concatenation in queries instead of parameterized queries

Concurrency (Critical for data corruption)

• Shared mutable state accessed without synchronization
• Missing locks/mutexes when modifying shared resources
• Check-then-act patterns without atomicity (TOCTOU)
• Deadlock potential: acquiring multiple locks in inconsistent order

Async Code

• After any await, verify assumptions are still valid — state may have changed
• Flag when code returns success without verifying the expected outcome of an async operation

External API Handling

• Missing timeouts on HTTP requests
• No retry logic for transient failures
• Missing circuit breakers for repeatedly failing services
• Not distinguishing between 4xx and 5xx errors
• Missing rate limiting awareness (no backoff on 429)

Edge Cases & Boundaries

• Code assumes arrays/lists are non-empty
• Missing null/undefined checks on optional values
• Off-by-one errors in loops, incorrect range checks
• Type coercion issues leading to unexpected behavior
• Unicode/encoding issues: assuming ASCII, incorrect string length

Defensive Coding

• Using external input (APIs/users) without validation
• Not handling failure cases for operations that can fail
• Array access without verifying index is valid
• Code relying on undocumented behavior or ordering

Input Sanitization (Critical)

• Command injection: unsanitized input in shell commands — use argument arrays, not string interpolation
• Path traversal: user input in file paths escaping intended directories — resolve and validate paths
• XSS: user input rendered as HTML — escape or use safe APIs, validate URL protocols
• Log injection: unsanitized input forging log entries — use structured logging
• SQL injection: string concatenation in queries — use parameterized queries

Memory & Performance

• Unbounded collections (arrays/maps growing without limits)
• Memory leaks: event listeners not removed, closures holding references
• Large allocations in loops that could be reused
• Blocking synchronous I/O or CPU-heavy work on main thread
• Missing pagination: loading entire datasets

File System Operations

• Reading files without verifying they exist
• Writing files without ensuring parent directory exists
• Missing error handling on file operations

Logging & Observability

• Insufficient logging for important operations
• Excessive logging creating noise or performance issues
• Missing correlation IDs for cross-service tracing
• Critical: Logging sensitive data (PII, passwords, tokens)
• New features without observability hooks

Testing Quality

• Tests that don't assert anything meaningful
• Missing edge case coverage (only happy path)
• Flaky tests: race conditions, time dependencies, external dependencies
• Test pollution: shared state, missing cleanup
• Mocking too much: tests don't exercise real code paths

Accessibility

• Missing alt text on images
• Non-semantic HTML (divs for buttons, missing form labels)
• Interactive elements not reachable via keyboard
• Missing ARIA labels on icon-only buttons
• Color-only indicators

Code Style

• Prefer early returns over nested if statements — flat code is easier to read

Error Messages

• Error messages must be actionable, contextual, and specific
• Include identifiers (order IDs, user IDs) for debugging
• Don't expose stack traces or internal details to end users
• Include error codes for programmatic handling

Language-Specific Guidelines

Based on file extensions in the diff, load the relevant reference:

• .ts, .js, .mjs, .cjs → Read references/typescript.md
• .tsx, .jsx → Read references/react.md (also load typescript.md)
• .py → Read references/python.md
• .go → Read references/go.md
• .rs → Read references/rust.md
• .java, .kt → Read references/java-kotlin.md

Only load references for languages present in the changed files.

Severity Definitions

• Critical — Block the merge. Broken code, security vulnerabilities, data leaks, runtime failures that will crash production.
• Major — Should fix. Unhandled async, missing error handling, resource leaks, race conditions, missing validation.
• Minor — Nice to fix. Code clarity, consistency, performance, dead code, duplication, unresolved TODOs.