Get Codex's review of Claude's implementation plans. Trigger when user wants a second opinion on a plan ("have Codex review this plan", "get second opinion from Codex", "critique this plan with Codex"), or after Claude creates a plan file that needs validation before implementation.
Get Codex's review of implementation plans for validation and second opinion. Trigger when user requests a plan critique or after Claude creates a plan file that needs review before implementation.
/plugin marketplace add robbyt/claude-skills/plugin install codex@robbyt-claude-skillsThis skill inherits all available tools. When active, it can use any tool Claude has access to.
Have Codex critique Claude's implementation plans for a second perspective.
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.
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"
})
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.
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"
})
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"
})
sandbox: "read-only" to prevent file modificationssandbox: "danger-full-access" - this is forbiddendangerouslyDisableSandbox: truereferences/setup.md for troubleshooting