interfaceVersion
1.0.0
description
Reliability
displayName
Reliability
concern
How agent runs survive failure, retry, and resume to a known-good state.
surface
Implementations expose:
- **Checkpoint write/read** — `checkpoint(runId, payload)` persists
a serializable run snapshot; `resume(runId)` reconstructs it.
- **Retry policy** — `retryPolicy(failure)` returns a backoff plan
(attempts, delay, jitter) given a typed failure.
- **Idempotency keys** — every side-effecting tool call is guarded
by an idempotency key sourced from this interface.
The interface guarantees at-least-once delivery of effects with
de-duplication on resume.
builtInDefault
The default implementation uses an on-disk JSON journal under
`.a5c/runs/<runId>/` and a fixed exponential-backoff retry policy
(3 attempts, base 500ms, jitter 0.2). No idempotency-key store —
keys are derived from `(toolId, inputHash)`.
notableImpls