Checkpoint Resume
Persist the minimum state needed to continue correctly after an interruption.
Checkpoint
- Read the active task contract.
- Record:
- current stage
- completed steps that have evidence
- unresolved blockers
- one concrete next action
- relevant artifact or commit references
- Prefer
hermes-durable checkpointwhen available. - Write checkpoints atomically. Never destroy the last known-good checkpoint before the replacement is complete.
- Do not call a step complete only because the current agent remembers doing it.
Resume
- Read
task.json, thencheckpoint.json, then the evidence ledger. - Reconstruct the current lane from durable state before reading old chat history.
- Confirm that referenced files, commits, or test outputs still exist when they materially affect the next action.
- Treat side effects as already completed only when durable evidence or repository state proves they occurred.
- Continue from
next_action; do not restart the whole plan by default. - If durable state conflicts with current repository reality, stop and surface the conflict instead of guessing.
Handoff
Keep the handoff provider-neutral. A different coding agent should be able to resume the task from repository state alone.