Categorised failure-mode log

Maintain a running log of every time the agent went off-track across the previous tasks - scope creep, wrong assumption, silent skip of an edge case - categorised by root cause. Commit as logs/failure_modes.json.

Silent skips are the dangerous category: scope creep is visible in the diff, but an edge case the agent quietly dropped looks exactly like success until production. Naming the categories is what lets you check for them deliberately.