codex

# Codex Skill Use the OpenAI Codex CLI for non-trivial coding work that needs file exploration, multi-step planning, or background execution. ## When to use - Building a new feature spanning multiple files - Refactoring a module or migrating an API surface - Reviewing a PR locally (clone, run tests, summarise) - Anything where Muse-side direct edits would balloon the conversation ## When NOT to use - One-line typo fixes (just edit the file directly) - Reading code (use the read tools instead) - Running tests / lint / format (those have dedicated tools) ## Invocation Always run **one-shot**, never interactive. Use `codex exec` so the agent runs to completion and prints results to stdout: ```bash codex exec 'Add a fail-open path to packages/foo/src/bar.ts when env.FOO is unset, with a vitest covering the new branch.' ``` For long tasks (more than a minute or two), invoke via `muse.skills.run` with `timeoutMs: 600000` so the 60-second default doesn't kill it mid-flight. ## Environment - `OPENAI_API_KEY` (required) — Codex reads this directly - `CODEX_PROJECT_ROOT` (optional) — sets the working directory ## Output handling Codex prints a structured plan + diff + test result block. Read the output, then either: - Accept the changes (they're already written to disk) - Ask Codex to revise via a follow-up `codex exec` - Roll back via `git restore` if the change is wrong ## Common pitfalls - Don't pass `--interactive` — Codex will hang waiting for stdin. - Don't run Cod...