This wraps OpenAI's Codex CLI so Claude can delegate entire coding tasks to GPT models instead of doing the work itself. You get three sandbox modes: read-only for analysis, workspace-write for actual implementation, and danger-full-access when you need network or system operations. The full-auto flag is your friend here since read-only is default. Honestly most useful when you want a second AI opinion on architecture decisions or need to parallelize work across different contexts. The code review subcommand against git diffs is solid. Just remember you're adding another model to the loop, which means more tokens and another set of quirks to manage.
npx skills add https://github.com/feiskyer/claude-code-settings --skill codex-skill