Evidence label: SYNTHETIC design harness, not a production recovery result. No provider calls, provider tokens or production jobs are represented below.
View the full infographic → · View this article
A provider limit interrupts a model request. If the only copy of the task and recovery plan lives inside that request, the workflow has no reliable way to wait and resume. Put job ownership, state and scheduling outside the model turn.
Persist the work before waiting
A resumable job record needs a stable job ID, requested outcome, acceptance criteria, current stage, last verified checkpoint, attempt history and the dependency that caused the wait. Store only the minimal context needed to continue. Keep completed stages immutable unless new evidence or a failed verifier requires reopening them.
Use a bounded external wait
- Classify the response as resource limit, transient fault, authentication failure or permanent task error.
- For a resource wait, record provider retry-after guidance when available; otherwise use capped backoff.
- Pause outside the model call. A lightweight scheduler can check the dependency without asking a model to narrate the wait.
- At checkback, reconcile the prior attempt before retrying.
- Resume from the checkpoint and run the verifier against the durable result.
Keep WAIT_RESOURCE distinct from WAIT_AUTH. A quota wait does not repair expired credentials, and refreshing credentials does not create capacity.
What the harness says—and does not say
The KD11 V20 fake-adapter harness enumerated 1,000 simulated WAIT_RESOURCE lanes and 1,000 simulated WAIT_AUTH lanes, reporting 2,000 synthetic resumes with zero provider calls, zero provider tokens and zero production jobs. It exercises state-machine logic without contacting a provider. It cannot show that a live provider will accept a resumed request or that a job will pass quality checks.
Failure handling is part of the workflow
Fence retries with attempt IDs or idempotency keys. If an outcome is ambiguous, inspect the prior operation before replay. Record each transition, checkpoint and verification result. A heartbeat alone does not renew a semantic-progress deadline.
See the agent operating rules builder and canary incident. This is a design guide, not evidence of production recovery or savings.
Read the public-safe synthetic harness scope; it does not establish production recovery.