plan-review

Plan Review via Codex

Have Codex critique Claude's implementation plans for a second perspective.

CRITICAL: Default Model

ALWAYS use model: "gpt-5.2" unless the user explicitly requests a different model. Do NOT choose o3 or other models on your own.

CRITICAL: Instruct Codex

Every prompt sent to Codex MUST include these instructions:

"You are running non-interactively as part of a script. Do not ask questions or wait for input. Do not make any changes. Provide your complete feedback immediately."

Codex is a consultant. Claude Code handles all file modifications.

Quick Start (MCP)

If the codex MCP tool is available, read the plan and pass it to Codex:

First, read the plan file content, then:

mcp__plugin_codex_cli__codex({
  "prompt": "You are running non-interactively as part of a script. Do not ask questions or wait for input. Do not make any changes. Provide your complete feedback immediately.\n\nReview this implementation plan:\n\n[PLAN CONTENT HERE]\n\nConsider:\n1. Are there gaps or missing steps?\n2. Are there risks not addressed?\n3. Is the approach optimal?\n4. What alternatives should be considered?",
  "sandbox": "read-only",
  "model": "gpt-5.2"
})

Fallback (Bash)

If MCP is unavailable, tell Codex to read the file directly:

codex exec "You are running non-interactively as part of a script. Do not ask questions or wait for input. Do not make any changes. Provide your complete feedback immediately.

Review the implementation plan at path/to/plan.md

Consider:
1. Are there gaps or missing steps?
2. Are there risks not addressed?
3. Is the approach optimal?" --sandbox read-only -m gpt-5.2-codex 2>&1

Note: Do NOT use stdin piping with $(cat) - Codex doesn't expand shell command substitution. Instead, provide file paths in the prompt and let Codex read them directly.

With Source Context

Include source files for context in the prompt:

mcp__plugin_codex_cli__codex({
  "prompt": "You are running non-interactively as part of a script. Do not ask questions or wait for input. Do not make any changes. Provide your complete feedback immediately.\n\nReview this implementation plan:\n\n[PLAN CONTENT]\n\nAlso read these source files for context:\n- src/auth/login.ts\n- src/middleware/session.ts\n\nEvaluate if the plan addresses the actual codebase structure.",
  "sandbox": "read-only",
  "model": "gpt-5.2"
})

Focused Reviews

Risk assessment:

mcp__plugin_codex_cli__codex({
  "prompt": "You are running non-interactively as part of a script. Do not ask questions or wait for input. Do not make any changes. Provide your complete feedback immediately.\n\nReview this plan for risks:\n\n[PLAN CONTENT]\n\nEvaluate:\n- Breaking changes\n- Data loss potential\n- Rollback complexity\n- Dependencies that could fail",
  "sandbox": "read-only",
  "model": "gpt-5.2"
})

Completeness check:

mcp__plugin_codex_cli__codex({
  "prompt": "You are running non-interactively as part of a script. Do not ask questions or wait for input. Do not make any changes. Provide your complete feedback immediately.\n\nReview this plan for completeness:\n\n[PLAN CONTENT]\n\nEvaluate:\n- Are all edge cases covered?\n- Is testing addressed?\n- Are there missing steps?",
  "sandbox": "read-only",
  "model": "gpt-5.2"
})

Recommended Pattern

1. Use Claude's Read tool to get plan file content
2. Embed content directly in the Codex prompt
3. List additional source files for Codex to read from project

Performance

• MCP plan review: ~5-30 seconds
• MCP with source files: ~1-2 minutes
• Bash fallback: ~2-3 minutes

Notes

• Always use sandbox: "read-only" to prevent file modifications
• NEVER use sandbox: "danger-full-access" - this is forbidden
• Tool name may vary by installation. Check available tools for exact name.
• Read plan file content first, then include in prompt (Codex may not access ~/.claude/plans/)
• MCP is preferred; Bash fallback requires dangerouslyDisableSandbox: true
• See references/setup.md for troubleshooting