review

name	review
description	Use when you want to review code changes made by Claude Code — standard review for bugs, or adversarial review to challenge implementation choices

Review Claude Code Output

Overview

Review code changes produced by a Claude Code task. Two modes:

• Standard review: Check for bugs, style issues, security problems, and performance concerns
• Adversarial review: Challenge implementation choices, question assumptions, identify failure modes and alternatives

Both modes produce structured JSON output with verdict, findings (severity-ranked), and next steps.

Workflow

1. Determine review mode:

   • User says "adversarial", "challenge", "质疑" → adversarial=true
   • Default → standard review

2. Call cc_review with:

   • cwd (required): absolute path to the user's current workspace
   • job (optional): job ID or prefix to review (default: latest completed job)
   • adversarial (optional): set to true for adversarial mode
   • focus (optional): specific aspect to focus on (e.g., "security", "performance")
   • base (optional): git base ref for diff (default: auto-detect)
   • scope (optional): review scope — auto (default: working-tree if dirty, else branch), working-tree (staged+unstaged+untracked), branch (diff against default branch from merge-base)

3. The tool returns the diff, git context, and review instructions. You (Codex) execute the review — apply the review criteria to the diff and produce a structured JSON result.

4. Parse the review result JSON and present findings to the user:

   • Overall verdict (approve / needs-attention / request_changes / reject)
   • Findings ordered by severity (critical → low)
   • Specific file and line references
   • Concrete recommendations for each finding
   • For adversarial: assumptions identified and their validity

5. If issues are found, suggest next steps:

   • Delegate fixes back to Claude Code: /claude:delegate
   • Fix directly in Codex

Review Output Schema

All reviews produce JSON matching this schema:

{
  "verdict": "approve|needs-attention|request_changes|reject",
  "summary": "terse ship/no-ship assessment",
  "findings": [{ "severity": "critical|high|medium|low", "title": "...", "body": "...", "file": "...", "line_start": 1, "line_end": 1, "confidence": 0.8, "recommendation": "..." }],
  "next_steps": ["step 1", "step 2"]
}

Adversarial Review Dimensions

When adversarial=true, the review examines:

1. Assumptions & Preconditions — what does the code assume? What breaks if violated?
2. Failure Modes & Degradation — when does this fail silently? At scale? Under concurrency?
3. Alternative Approaches — were simpler alternatives considered? Known patterns/libraries?
4. Technical Debt & Maintenance — coupling, rewrite risk, known incompleteness markers
5. Boundary Conditions — empty inputs, null, race conditions, platform-specific behavior

Git Context

The review tool automatically provides:

• Changed files list
• Working tree state (modified/added/deleted/untracked)
• Diff with size measurement (truncated if >512KB)
• Untracked file contents (for new files)
• Diff stat summary
• Auto-detected base ref (merge-base on feature branches, HEAD~1 on default)

Examples

• "Review the Claude Code changes" → standard review
• "Adversarial review of what Claude Code did" → adversarial review
• "Review focusing on security" → cc_review with focus="security"
• "Review job cc-abc" → cc_review with job="cc-abc" (prefix matching)