Search for a command to run...
Where Do Deep-Research Agents Go Wrong? Span-Level Error Localization in Agent Trajectories