dunkeln/crucible

Crucible is a plugin for Codex, an AI coding agent, that changes how the results of an agent's work session are kept and used afterward. Normally when an AI agent fixes a bug or completes a task, the details of that work disappear into a chat transcript once the session ends. Crucible instead turns each attempt into a structured record with four parts: the attempt itself, a check that verifies whether it actually worked, the evidence produced by that check, and a decision about whether to promote the result as something worth keeping. The core rule behind the project is that nothing counts as a success unless a verifier confirms it. A verifier is a test or check that stays close to the specific project, since the project itself is what defines what counts as correct. Even a failed attempt is treated as useful, because it shows exactly where the agent's approach diverged from what the verifier expected, and that failure becomes information a future attempt can learn from. The project includes a benchmark comparing a plain task folder against a Crucible-managed task pack that includes prior failing evidence and a follow-up brief. On this test, the Crucible-assisted run showed a meaningful drop in wasted input tokens and billable cost alongside a higher-quality output, with the verifier passing on the final result. Crucible is installed as a Codex plugin through Codex's plugin marketplace and then used directly inside the repository where the work is happening. Humans stay responsible for deciding what gets promoted or kept, while Codex can propose tasks, verifiers, and repairs, and Crucible runs the checks and records what happened.