use-codex

Development

Delegate work to OpenAI's Codex from inside Claude Code — a generic task (implement a feature, investigate a bug, refactor) or a code review as a cross-model second opinion. Trigger ONLY when the user explicitly names Codex — e.g. "use codex to implement this", "have codex fix this", "delegate this to codex", "review this with codex", "codex adversarial review", "set up codex". Do NOT trigger on a plain "implement…" / "fix…" / "review…" with no mention of codex — those are Claude's own job (reviews go to the built-in /code-review). Also covers installing/setting up Codex.

View on GitHub ↗

QUICK START

How to use this skill

Bring this guide into your coding agent with a prompt tailored to the tool you use.

Open your project in Codex.
Copy and paste the prompt below.
Review files and risks before approving any action.

Prompt to paste

Inspect the Agent Skill at https://github.com/LibreYOLO/libreyolo/blob/HEAD/skills/use-codex/SKILL.md and all supporting files. Report provenance, permissions, scripts, dependencies, and security risks. Propose an exact installation into .agents/skills/use-codex, but wait for my approval before writing files or running scripts.

Copying this prompt does not install or run the skill. Review third-party files before use.

Use Codex

Codex is a separate agent with its own model and reasoning. You orchestrate it and own the final integration: Codex drafts → you review and reconcile → user approves. Never auto-apply its output. It starts cold on this repo — always include the house rules (bottom) in the brief.

Setup

Requires the codex CLI: npm install -g @openai/codex, then the user runs codex login (interactive, ChatGPT/OpenAI account). On Windows, npm globals often aren't on the shell PATH — if codex isn't found but node is installed, prepend them first:

$env:Path = "C:\Program Files\nodejs;$env:APPDATA\npm;$env:Path"

Delegate a generic task

codex exec --sandbox workspace-write -a never "<task, acceptance criteria, constraints>"

Codex edits the working tree directly. Show the user its plan and the resulting git diff before integrating.
Analysis only, no edits: use --sandbox read-only.
Model / effort per run: -m <model> and -c model_reasoning_effort=<minimal|low|medium|high>. Deep pass: codex exec -m gpt-5-codex -c model_reasoning_effort=high …. Persistent defaults live in ~/.codex/config.toml (or -p <profile>).

Delegate a code review

If the official plugin is installed, tell the user to run /codex:review (--base <ref>, --wait/--background) or /codex:adversarial-review (challenges the design, good before big merges) — you cannot type slash commands yourself. Otherwise run it directly:

codex exec review --base dev

Feature branches review against dev. Require file:line citations; priority correctness > security > spec > performance > maintainability > tests; no style nits. Render verdict + findings for the user; do not apply fixes unasked.

Other routes

MCP server (autonomous multi-turn hand-off): user registers once with claude mcp add codex -- codex mcp-server; you then get mcp__codex__codex / mcp__codex__codex-reply tools to drive Codex sessions yourself.
Official plugin (codex-plugin-cc) — user installs with: /plugin marketplace add openai/codex-plugin-cc, then /plugin install codex@openai-codex, /reload-plugins, /codex:setup. Beyond reviews it adds /codex:rescue (delegate a fix), /codex:transfer (persistent Codex thread), /codex:status|result|cancel (background jobs). Optional --enable-review-gate blocks Claude until Codex signs off — off by default, suggest only if asked. Plugin commands take model/effort from ~/.codex/config.toml, not flags.

House rules (include in every brief)

Follow the repo's licensing policy: no third-party CV library names in committed text; respect clean-room boundaries.
Branch flow is dev → release; feature branches base on dev.
No new dependencies without surfacing them; match existing style; leave tests runnable.