Learning-path discovery can now accept natural follow-up messages, preserve confirmed learner context across restarts and resume work without duplicating a turn. Clarification questions are bounded, corrections take effect on later turns and unrelated requests return a redirect while keeping the useful conversation state intact.
The backend now records agent history, public conversation state, model usage and generation work as one durable operation. Replayed requests return the original result, worker interruptions are recovered with fencing and bounded cleanup, and allowance checks happen before generation work is admitted. The conversation routes and learner-facing progress stream remain scheduled for later checkpoints.