Skip to main content

Implement recovery actions

Stub

This How-to is a stub. JARVIS models recovery actions as artifacts/events, but the supported recovery catalog is not yet stabilized.

Goal​

You will implement bounded recovery actions (retry, remap, escalate, abort) and record them as durable artifacts.

When to use this​

  • A subtask fails and you need deterministic “what happens next”.
  • You want replayable debugging (recovery decisions are durable, not ephemeral).

Prerequisites​

  • A recovery action schema
  • Budgets for recovery (max retries, max remaps, max total node runs)
  • An enforcement point that prevents infinite loops

Steps​

  1. Define the supported recovery actions for the executor (start small).
  2. On failure, choose a recovery action within budget.
  3. Emit a recovery artifact/event and apply it deterministically:
    • retry → re-run the same node run (idempotency required),
    • remap → pick another candidate and regenerate args,
    • escalate → produce an “operator required” artifact,
    • abort → fail the composite node run.

Verify​

  • Failures always produce a recovery artifact/event.
  • Recovery does not exceed budgets and cannot loop indefinitely.

Troubleshooting​

  • Infinite retries/remaps → enforce max retries/remaps and fail closed.
  • Recovery is non-deterministic → record the action + rationale in the artifact payload.
  • Recovery needs approval semantics → integrate PDP “require approval” later.

Cleanup / Rollback​

  • None.

Next steps​