Local Code Review

Review uncommitted changes like a pull request, using sub-agents for analysis.

Usage

/local-code-review                        # Review all changes
/local-code-review security, performance  # Focus on specific areas

Focus areas via $ARGUMENTS: security, performance, architecture, readability, error-handling, testing

Workflow

1. Gather Changes (parallel)

Run in parallel via Bash:

• git status — changed files overview
• git diff — unstaged changes
• git diff --cached — staged changes
• git log -5 --oneline — recent commits for context

If no changes found, report "No uncommitted changes to review" and stop.

2. Detect Project Context

Auto-detect project conventions by checking for:

• CLAUDE.md (project root and .claude/ directory)
• docs/ or doc/ directory for architecture docs
• Linter/formatter configs (.eslintrc, pyproject.toml, analysis_options.yaml, etc.)
• Package manifests (package.json, pubspec.yaml, Cargo.toml, go.mod, etc.)

Read available files to understand project language, framework, and conventions.

3. Evaluate Scope

Changed files count?
    │
    ├─ ≤10 files ──► Single agent review
    └─ >10 files ──► Staged review by directory

Staged review: Group files by top-level directory, launch one sub-agent per group.

4. Delegate to Sub-agent

Launch Task with the following:

Scenario	Agent
Standard review (≤10 files)	general-purpose
Deep analysis or complex logic	advanced-general-purpose

Sub-agent prompt must include:

• Full diff output
• File list with change summary
• Project context (language, framework, conventions detected in Step 2)
• Focus areas from $ARGUMENTS (if provided)
• The 8 review categories below

Review Categories

1. Correctness — Logic errors, edge cases, off-by-one, null/undefined handling
2. Security — Injection, auth issues, secrets exposure, OWASP Top 10
3. Readability — Naming, complexity, dead code, unclear intent
4. Architecture — Separation of concerns, dependency direction, project conventions
5. Error Handling — Missing catches, swallowed errors, unclear error messages
6. Performance — N+1 queries, unnecessary allocations, blocking operations
7. Testing — Missing test coverage for changed logic, test quality
8. Project Conventions — Style guide compliance, patterns used elsewhere in codebase

If $ARGUMENTS specifies focus areas, prioritize those categories but still scan all.

5. Structured Output

The sub-agent must return this format:

## Review Summary

**Files Changed**: N files
**Language/Framework**: [detected]
**Overall Assessment**: Approved / Changes Requested / Needs Major Revision

## Findings

### Critical (Must Fix)
- [ ] Description — `file:line`

### Warnings (Should Fix)
- [ ] Description — `file:line`

### Suggestions (Consider)
- [ ] Description — `file:line`

### Good Practices Observed
- Positive observation

## Recommendations
- Actionable next steps

Agent Selection

Signal	Agent
Standard review	general-purpose
>10 files, staged	general-purpose per group
"thoroughly", deep analysis	advanced-general-purpose

Integration

• Use after intent-first to clarify review scope if request is vague
• Pairs with complex-orchestrator for staged reviews of large changesets
• Follow delegation-triggers guidelines for agent selection