Kestrel stores run state separately from the live connection that created it. That lets a person inspect or continue work after a terminal, browser, or network request has ended.
Durable records
The store keeps:
- session, thread, and run identities
- ordered events and terminal outcomes
- versioned session memory and task state
- tool results and linked artifacts
- workspace checkpoints and operator actions
Replay
Replay reads recorded evidence to explain or compare what happened. It does not rerun the model, replace the original terminal result, or turn a partial live stream into a new source of truth.
When investigating an outcome, inspect the terminal event first, then the events, memory revision, tool results, artifacts, and checkpoints attached to the same run. This keeps diagnosis tied to recorded work rather than to a new prompt that may behave differently.
Continue with Reliability, Artifact inspection, and Evaluations.