The title is the whole problem. One agent writes the code and the same agent tells you it looks good, which is not review, it is a receipt. I have tried a lot of Claude + Codex setups. The handoff was always the weak point: Claude writes a plan, you paste it into Codex, and Codex builds a summary of reasoning it never did. Then you paste the diff back for review. Half a day of tab-switching per feature. This closes it into one loop. Codex opens a persistent thread and writes the plan as the implementer. Claude approves it or sends it back. Codex resumes that same thread and implements its own approved plan. Claude reviews the diff against that plan file. The part I did not expect to matter as much as it did: keeping the thread. When Codex implements a plan it authored, the reasoning is still in context, so it stops re-deciding what the plan already settled. The gates refuse things on purpose. Implementation will not start unless the plan is approved, and the approval is bound to the plan text that was read, so editing the plan afterwards voids it. Release refuses unless the review approved the tree as it is right now. An "approve" that carries blocking items is rewritten to "revise" on disk. I found several of those the honest way: I handed the whole thing to Codex as an adversarial reviewer, and it opened with "not release-ready, both hard gates can report approval for state they did not validate." It was right. v0.2.0 is what came out of that. It ran 45 gated waves across 20 lanes of a production codebase before I extracted it. MIT, free, no signup, nothing to buy. Tests are 53 assertions that run against a stub codex, so you can verify the gates without spending quota. Happy to answer anything, and I would genuinely like to hear where it breaks for you.

codex-claude-loop

Your AI approves its own work. This one can't.

Your AI approves its own work. This one can't.