Recovery Contract
Tentarc treats recovery as a first-class contract, not afterthought retries.
Recovery boundary
On crash/restart, runtime restores from persisted control/state plane data:
- pending control intents
- last committed status transition
- latest memory revision
- resumable execution cursor/checkpoint
Guarantees
- admitted duplicate requests stay blocked after restart
- completed side effects are not replayed blindly
- interrupted non-terminal runs can resume
- terminal runs remain terminal unless explicitly forked/replayed
Failure taxonomy
Classify failures explicitly:
- admission failure: rejected before execution
- execution failure: runtime/tool path failed during data plane
- projection failure: state write failed after execution step
Different class, different operator action.
Operator playbook
For production-like workflows:
- inspect status timeline
- locate last durable checkpoint
- decide resume vs replay vs fork
- preserve audit trail of control decisions
This keeps reliability high without hiding failure reality.
Share feedbackLast updated: Mar 6, 2026