Documentation
Durable work and failure diagnosis
Safe work projections, normalized states, redacted logs, and diagnosis.
- Operators
- Developers
- Administrators
Long-running application jobs and Manager operations are durable work. Every work item has a typed ID, normalized state, progress, cancellation capability, and polling interval. Application work uses type job; host-management work uses type manager_operation.
The MCP work projection deliberately omits raw queue payloads. Logs are bounded and redacted. They include operational events and safe identifiers, not source documents, prompts, credentials, arbitrary Manager task inputs, or raw results.
Diagnose work safely
- Inspect the work item.
- Read its bounded log.
- Distinguish retryable infrastructure failures from validation or revision failures.
- Inspect current session, provider, or Manager state.
- Explain a safe next action before mutating anything.
Queued, running, waiting, succeeded, failed, and cancelled are the normalized states. Cancellation is a request and may take time while a task reaches a safe boundary. Retrying or re-executing must use the idempotency key associated with the exact reviewed action.