codexlisted

# Codex: Read-Only Second Opinion via OpenAI Codex CLI Delegate read-only review work to the `codex` CLI. Codex gets its own look at the code/plan/diff and reports back. Claude stays in charge of any writes. Codex runs are LONG (often many 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 does not babysit codex and does not read raw codex output — it gets a finished summary back from the subagent. ## Mandatory flags - **`-C /tmp`** — always run codex from `/tmp`, never from the caller's repo. Running codex inside a real project (e.g. one with `keyshelf.config.ts`, `.env.keyshelf`, `package.json` triggers, etc.) makes codex burn its first turn auto-discovering project skills and it often exits mid-reasoning instead of doing the actual analysis. Pass any project file paths as **absolute paths inside the prompt** — codex's read-only sandbox can still read them. - **`--skip-git-repo-check`** — always set, since `/tmp` is not a git repo. - **`--sandbox read-only`** — never relax this. - **`-c approval_policy="never"`** — codex never prompts. - **`-o <file>`** — always write output to a file; never consume stdout directly. ## Step 1: build the review prompt Pick a recipe below (or write your own). Reference every project file by **absolute path** — codex is running from `/tmp` and has no ne