claudex-loop

Why we included this project

The clean AI plan that only falls apart once real code starts is a common failure, and this skill goes after it directly. Installed into Claude Code as a plugin, claudex-loop runs a four-phase loop: Claude locks the plan with you, then OpenAI Codex attacks that plan in a read-only sandbox until there is nothing left to find. The core idea stays deliberately plain: the model that wrote the plan never gets to grade it. After the adversarial review, you choose which model builds and the other inspects the resulting diff before you approve anything, with the run shipped as named skills and a round-by-round review log. The added ceremony pays off most for teams doing non-trivial builds with AI pair-programming, where a single confident pass can quietly leave an unbuildable subsystem or a data-corrupting design in place.