When code-repair agents fail, the final test suite shows that the patch is incorrect, but it does not reveal which earlier commitment or decision led to the failure. This makes debugging the agent's reasoning process difficult.

The new work, called Poison Valleys, aims to diagnose this path-dependent failure. It traces the sequence of steps in 1,844 model requests on two synthetic Python tasks, identifying which software obligation was not met.

By pinpointing the exact point of divergence, Poison Valleys could help developers understand and correct LLM-based code repair systems.