Codex Code Reviewer

Codex Code Reviewer

Overview

This skill provides a systematic, iterative code review workflow powered by zen mcp's codex tool. It automatically checks recently modified code files against project standards (CLAUDE.md requirements), presents identified issues to users for approval, applies fixes, and re-validates until code quality standards are met or the maximum iteration limit is reached.

Operation Modes:

• Interactive Mode (Default): Present issues to user, wait for approval before applying fixes
• Full Automation Mode: Automatically select and apply all suggested fixes without user approval (activated when user initially requests "full automation" or "complete automation")

When to Use This Skill

Trigger this skill when the user says:

• "Use codex to check the code"
• "Please help me check if the recently generated code has any issues"
• "Check the code after each generation"
• Similar requests for code quality validation using codex

Workflow: Iterative Code Review Cycle

Prerequisites

Before starting the review cycle:

1. ** Read Operation Mode from Context (automation_mode - READ FROM SSOT):**
   • automation_mode definition and constraints: See CLAUDE.md「📚 共享概念速查」
   • This skill's role: Skill Layer (read-only), read from context [AUTOMATION_MODE: true/false], true → auto-fix / false → ask user

1a. Read Coverage Target from Context (coverage_target - READ FROM SSOT):

• coverage_target definition and constraints: See CLAUDE.md「📚 共享概念速查」
• This skill's role: Skill Layer (read-only), read from context [COVERAGE_TARGET: X%], validate coverage (≥ target pass, < 70% reject)

2. Identify Files to Review:

   • Use git status or similar to identify recently modified files
   • Interactive Mode (automation_mode = false): Confirm with user which files should be reviewed
   • Automated Mode (automation_mode = true): Auto-select all recently modified files, log decision to auto_log.md

3. Initialize iteration counter: current_iteration = 1

4. Set maximum iterations: max_iterations = 5

5. Initialize review tool tracker: first_review_done = false

6. Detect Review Context:

   • Check if this is final quality validation (project completion/final verification)
   • If YES: Enable Final Validation Mode (requires minimum 2 passes: codereview + clink)
   • If NO: Use Standard Review Mode

Review Cycle Loop

Execute the following loop until termination conditions are met:

Step 1: Select and Invoke Appropriate Review Tool

Tool Selection Logic:

if current_iteration == 1 and not first_review_done:
    # First review: Use mcp__zen__codereview
    review_tool = "mcp__zen__codereview"
    first_review_done = True
else:
    # Second and subsequent reviews: Use mcp__zen__clink with codex CLI
    review_tool = "mcp__zen__clink"

A) First Review: Call mcp__zen__codereview

Tool: mcp__zen__codereview
Parameters:
- step: Detailed review request (e.g., "Review the following files for code quality, security, performance, and adherence to project standards...")
- step_number: current_iteration
- total_steps: Estimate based on findings (start with 2-3)
- next_step_required: true (initially)
- findings: Document all discovered issues
- relevant_files: [list of recently modified file paths - MUST be absolute full paths]
- review_type: "full" (covers quality, security, performance, architecture)
- model: "codex" (or user-specified model)
- review_validation_type: "external" (for expert validation)
- confidence: Start with "exploring", increase as understanding grows

Important: Use absolute full paths; follow zen mcp's codereview workflow (2-3 steps); reference references/quality_standards/README.md (on-demand loading)

B) Second and Subsequent Reviews: Call mcp__zen__clink with codex CLI

Tool: mcp__zen__clink
Parameters:
- prompt: "Review files for: 1) Code quality 2) Security 3) Performance 4) Architecture 5) Documentation. Categorize by severity (critical/high/medium/low), provide file:line locations and fix recommendations."
- cli_name: "codex"
- role: "code_reviewer"
- files: [absolute full paths]
- continuation_id: [optional, for context continuity]

Important: Use absolute full paths for files; use continuation_id for context continuity; supported params: prompt, cli_name, role, files, images, continuation_id

Step 2: Present Findings and Decision Making

After codex returns the review results:

1. Summarize Issues Clearly:

   • Group by severity (critical/high/medium/low)
   • Provide specific file locations and line numbers
   • Explain the impact of each issue

2. Decision Making Based on automation_mode (READ FROM CONTEXT):

   • automation_mode behavior and decision logic: See CLAUDE.md「📚 共享概念速查」and「G11 Automated Repair Skills」
   • [AUTOMATION_MODE: false] → Interactive: Ask user approval
   • [AUTOMATION_MODE: true] → Automated: Auto-fix Critical/High, conditional Medium, auto-fix Low style issues
   • Output [Automated Decision Record] fragments for auto_log.md (collected by router, generated by simple-gemini)

Step 3: Apply Fixes

Interactive Mode: If user approves Automation Mode: Proceed directly based on auto-decision logic

1. Apply fixes to the identified issues
2. Use appropriate tools (Edit, Write, etc.) to modify code
3. Follow project coding standards from CLAUDE.md
4. Document changes made

Automation Mode Transparency:

• auto_log mechanism and format: See CLAUDE.md「📚 共享概念速查 → auto_log」and skills/shared/auto_log_template.md
• Example output:

  [Automated Fix Record]
   Fixed: 3 critical, 2 medium, 1 low | Skipped: 1 medium (business logic)
   1. SQL injection (critical) → Fixed | 2. N+1 query (medium) → Fixed
   3. Variable naming (low) → Fixed | 4. Logic optimization (medium) → Skipped

Step 4: Increment and Validate

current_iteration = current_iteration + 1

Check termination conditions:

Standard Review Mode:

• Success: Review tool reports no issues → Exit loop, report success
• Max iterations reached: current_iteration > max_iterations → Exit loop, report remaining issues
• User cancellation: User declined fixes → Exit loop
• Continue: Issues remain and iterations < max_iterations → Go to Step 1

Final Validation Mode (Project Completion/Final Verification):

• Minimum Pass Requirement: MUST complete at least 2 passes
  • Pass 1: mcp__zen__codereview (first_review_done = true)
  • Pass 2: mcp__zen__clink with codex CLI
• Early Exit Prevention:
  • If current_iteration < 2 → MUST continue to Step 1 (even if no issues found)
  • If current_iteration >= 2 AND no issues found in both passes → Exit loop, report success
• Max iterations reached: current_iteration > max_iterations → Exit loop, report remaining issues
• User cancellation: User declined fixes → Exit loop
• Continue: Issues remain OR minimum passes not met → Go to Step 1

Termination and Reporting

When the cycle terminates, provide a final report:

Code Review Completion Report:

Review rounds: X / 5
Reviewed files: [list files]

Tools used:
- Round 1: mcp__zen__codereview (codex workflow validation)
- Round 2: mcp__zen__clink (codex CLI direct analysis)
- Round 3+: mcp__zen__clink (continued)

Final status:
- All issues fixed / Max reviews reached / User cancelled
- [Final Validation Mode] Completed minimum 2 passes / Did not meet minimum 2 pass requirement

Fix summary:
- Fixed issues: X
- Remaining issues: Y (if any)

Recommendations: [next steps]