Codex CLI

If .claude/skills/codex-cli/SKILL.md.disabled exists, stop before reading a fallback. Project configuration: If the current project contains .claude/skills/codex-cli/SKILL.md, read and apply it instead of this bundled default. The project copy is the capability source of truth. Repository context: Read ../setup/references/task-context.md when resolving a task, project commands, named agent roles, or legacy .claude/ resource paths. Use repository evidence and applicable project instructions; a missing local workflow copy does not require setup. Resolve bundled resources from the installed skill, never from the target cwd.

Codex CLI

Announcement: Begin with: "I'm using the codex-cli skill for cross-AI validation with Codex."

Use this wrapper for one-shot Codex review, approach validation, or a delegated second opinion. Read the shared lifecycle in references/cross-ai-run.md before running it.

Provider adapter

1. Check command -v codex, then codex --version. Read codex --help and codex exec --help when the installed version or requested scope is unclear.
2. Use the locally supported non-interactive command and output option. Choose the caller's model, or omit a model override for the configured default. Never pass a hardcoded model from this skill.
3. Prefer the provider's least-privilege mode for review. Do not use an autonomous write or permission-bypass mode unless explicitly requested.
4. Capture stdout/stderr according to the installed CLI's help, read the result, and return the shared receipt. Historical codex review versus codex exec review flag differences belong in reference.md; verify them locally before use.

If codex is absent or needs an update, report the exact missing prerequisite and stop. Installing or updating it is a separate user action; this skill never runs npm i -g, curl | bash, or changes Codex settings.

Prompt requirements

The prompt is a cold-start handoff. Include the repository root, task path when task-attached, changed files or diff scope, requirements, and the specific questions to answer. Ask for findings or an approach only; the result does not authorize repository edits.

For task-attached calls, resolve ../setup/references/task-context.md and save the result receipt in that task's evidence area. Standalone calls return the result without creating a task.

When not to use

• The user asked for the primary review or implementation workflow only
• The request needs an interactive conversation
• The question is trivial or can be answered from local evidence

References

• references/cross-ai-run.md — shared preflight, context, capture, and evidence lifecycle
• reference.md — exact flags only after checking the installed CLI; historical compatibility notes
• templates.md — prompt shapes, with placeholders that must be resolved before execution