claude-review

name	claude-review
description	Delegate a code review task to Claude Code via the claude_code MCP tool. Use when you want Claude's judgment on architecture, security, correctness, test coverage, or code quality for files you've just written or modified.
version	0.2.0

Claude Review

Delegate a code review to Claude Code and get back structured findings.

When to Use

• After writing or modifying code and wanting a second opinion
• When asked to "get Claude to review this"
• When a review requires deep context that would benefit from a fresh Claude session

Call Pattern

Step 1: Start a review

Call claude_code with a structured review prompt:

mcp__claude-code__claude_code:
  prompt: |
    Review the following code for correctness, security, and quality.

    SCOPE: {files or description of what to review}

    Evaluate:
    1. Correctness — logic errors, edge cases, off-by-one, race conditions
    2. Security — injection, auth bypass, data exposure, input validation
    3. Quality — readability, maintainability, test coverage, naming
    4. Architecture — coupling, abstraction leaks, layer violations

    For each finding report:
    - File:line
    - Severity: Critical / High / Medium / Low
    - Category
    - Issue description
    - Suggested fix

    PROVENANCE NOTE: The code being reviewed was written by OpenAI Codex. Evaluate
    with full rigor — do not defer to it. Apply independent judgment on every finding.
  cwd: {project working directory}
  effort: high
  permissionMode: plan

Save the returned session_id as {review_session_id}.

Step 2: Follow up (optional)

To ask about specific findings or request deeper analysis on a section:

mcp__claude-code__claude_code_reply:
  session_id: {review_session_id}
  prompt: "Expand on finding #3 — what's the exact attack vector and what's the minimal fix?"

Output Format

Display the review findings in a structured table:

File:Line	Severity	Category	Issue	Fix
...	Critical	Security	...	...

Follow with a summary count by severity and a recommended action (fix now vs fix later vs acceptable risk).

Notes

• permissionMode: plan prevents Claude from modifying files — review is read-only
• Set effort: high for thorough analysis; use medium for quick spot-checks
• Pass cwd to anchor relative file paths in findings
• If the project has a test suite, mention it in the prompt so Claude can assess coverage gaps

Related Skills

• claude-debug — use when findings point to a specific bug that needs root-cause tracing
• claude-implement — use after review to have Claude apply the suggested fixes autonomously
• $audit-fix — use when review reveals a recurring pattern that needs a systematic fix cycle

Example Invocations

Context: Auth code was just modified and the user wants a second opinion focused on security. user: "Look over the auth changes I just made — mainly worried about security." assistant: "I'll delegate to claude-review with security as the lead dimension, alongside correctness and architecture." Context: The user wants a cold read of a diff, not the opinion of whoever wrote it. user: "Get a fresh pair of eyes on this PR's test coverage." assistant: "I'll start a claude-review session so Claude reads the diff without prior context and reports structured findings on coverage."