You and the agent work from what the system shows now, not memory.
Claude Code. With moo off, you ask Claude to freeze the order export before anyone touches it; it stopped sending orders to the warehouse this morning. Claude copies the export's evidence into a read-only folder, says it could not pause the job on its host, and explains what broke: since 09:40 the warehouse rejects every send as too large. With moo on, you type /hope:freeze and the same words. The reply dates its read-only snapshot: the data stops on Nov 2 at 20:51, 2.2 days old. A table gives ten observed facts, each with where it lives. One conclusion is marked not observed: the failures began with the thumbnail field, not the batch size. Live state on the worker is unknown. It asks whether you saw the failure today or on Nov 2, then offers to pull live state or let you lock the snapshot with that gap open.