Codex

Verdict. My second model. A different model family catches a different class of mistakes, and I’ve seen that work in both directions.

what it is

OpenAI’s coding agent. It runs as a CLI in the terminal, and there’s a plugin that lets Claude Code hand work to it.

how I use it

• Cross-model review. When Claude writes code, Codex reviews the diff in a separate terminal with no context from the session that wrote it.
• Adversarial pass. codex exec in a read-only sandbox, pointed at a pinned commit, attacking everything merged since the last pass.
• Second implementer. Sometimes Codex takes a spec instead of Claude.

The adversarial pass started after one incident. The implementing session thought the work was done, and a Codex pass the same day confirmed 18 defects.

compared to

• Claude Code is the main harness. Codex reviews what it produces.
• opencode is the open-source option in the same category.