The Debugger Looked at the Step Before the Click

CUADebug introduces a failure taxonomy, a human-annotated set of 204 failed OSWorld trajectories and a debugger that inspects suspicious before-and-after screenshots with action traces. Task reasoning and control accounted for 110 failures, more than perception, grounding or external-system categories. On the reported split, structured root-cause guidance roughly doubled continual re-execution success from 12.2 to 25.86 percent, still below human-oracle guidance at 29.21 percent. The evidence is benchmark-specific and leaves most failed tasks unresolved.