Dual plan (Claude + Codex/ChatGPT)

Feature workflow with two models: Claude drives, Codex independently critiques at two gates, the plan and the final diff. The value is adversarial: Codex is prompted to find what the plan gets wrong, not to agree with it.