codexlisted

# Codex: Second Opinion / Hands-On via OpenAI Codex CLI Delegate work to the `codex` CLI — a review, an exploration, or actually doing something (running commands, making changes). Codex always runs with full access; what it's allowed to do comes from the **prompt**, not a sandbox flag. Codex runs are LONG (minutes, sometimes over an hour). This skill spawns **one dedicated subagent** whose entire job is: launch codex detached → block on a re-invoked poll loop until it exits → read the output → digest it → return a summary. The calling agent never babysits codex directly. ## Step 1: build the prompt Reference every file by absolute path. Tell codex plainly what it may do: - **Review only**: prepend "You are doing a READ-ONLY REVIEW. Do not create, modify, or delete any file, or run any state-changing git command (commit/push/branch/checkout/reset). Describe fixes in your output — don't apply them." - **Do it**: just describe the real task — implement X, run Y, fix Z. No restriction needed. Either way, add `Do NOT invoke any skills.` — codex otherwise burns its first turn on project-skill discovery. Codex has network access; let it fetch a PR diff/docs itself rather than pre-fetching, unless Claude already has the data in hand. ## Step 2: spawn the watcher subagent Use the Agent tool, `subagent_type: general-purpose`, `model: "opus"`, backgrounded (the default). The subagent has zero memory of this session — embed the template below verbatim, substituting only