/codex — Peer AI via OpenAI Codex

You have access to OpenAI's Codex CLI as a peer AI. Use it to brainstorm, delegate tasks, get second opinions, or ask questions from a different AI perspective.

How to invoke

Run the wrapper script via Bash:

bash "${CLAUDE_PLUGIN_ROOT}/scripts/codex-run.sh" \
  --prompt "YOUR PROMPT HERE" \
  --cwd "/path/to/relevant/directory" \
  --mode exec

Then read the output file path it prints and present the results to the user.

Step-by-step workflow

1. Parse the user's request. The user says /codex "some prompt" or /codex followed by a question.
2. Determine the working directory. Default to the current project root. If the user references a specific repo or directory, use that.
3. Run codex-run.sh with --mode exec and the user's prompt via the Bash tool.
4. Read the output file using the Read tool. The script prints the output file path to stdout.
5. Present the result to the user. Frame it as "Codex's perspective" or "Codex says:" — make it clear this came from the peer AI.
6. Synthesize if appropriate. If the user asked for a comparison or second opinion, provide your own perspective alongside Codex's.

Options

Flag	Default	Description
--prompt	(required)	The question or task for Codex
--cwd	current dir	Working directory for Codex
--mode	exec	Always exec for this skill
--timeout	from config (120s)	Max seconds to wait
--model	from config	Override the Codex model
--sandbox	full-auto	Sandbox mode: full-auto or suggest

Example invocations

• /codex "What's the best way to implement a rate limiter in Go?"
• /codex "Review this architecture: we have a React frontend calling a FastAPI backend with Redis caching"
• /codex "Write a bash script that finds duplicate files by hash"
• /codex "Debug this error: TypeError: Cannot read property 'map' of undefined"

Rules

• Always use --ephemeral and --full-auto — these are hardcoded in the wrapper, never override them.
• Never pass secrets or API keys in the prompt.
• Respect timeouts — if Codex times out, tell the user and suggest a simpler prompt.
• Frame the output clearly — the user should always know which AI produced which response.
• Don't modify files based on Codex's suggestions without user confirmation.