crucible

Why we included this project

Most AI-assisted projects fail at the plan stage, and crucible is built around that specific failure. Claude scouts the codebase or prior art first, then walks you through one load-bearing question at a time until PLAN.md is locked. From there OpenAI Codex takes over as a read-only adversary, attacking the plan round after round while Claude arbitrates and revises; the loop ends only when Codex runs out of findings or hits the round cap. The point of the split is that the model which wrote the plan never gets to grade it. The optional build phase flips the roles, with Codex writing and Claude checking the diff. If you have been burned by a plan that sounded finished but wasn't, the Assumptions Ledger and the review log give you something concrete to work from.