CUADebug is a framework for diagnosing and repairing computer-use agent failures across perception, spatial grounding, interaction, reasoning, and environment dynamics. It combines a CUA-specific taxonomy, CUAErrorBench with 204 human-annotated failed OSWorld trajectories, and a tool-augmented debugger that inspects suspicious steps using paired screenshots and action traces. On the main Claude-agent split, Gemini 2.5 Pro raises joint subtype-and-step diagnosis from 11.2% to 19.6%. Continual re-execution success improves from 12.2% to 25.86%, approaching the reported 29.21% with human-oracle guidance.
No heat snapshots are available in the last 24 hours.