Durable execution
Durable commands and bounded reasoning stages exist; general dynamic agent orchestration is target work.
The implemented product path commits a user-visible task and a dispatch command before the worker begins. DBOS checkpoints execution; product PostgreSQL stores authority, accounting and accepted results. This separation lets a browser reconnect without creating a second task.
sequenceDiagram participant UI as Browser participant API as Product API participant DB as Product PostgreSQL participant D as Dispatcher participant W as DBOS worker UI->>API: Command with idempotency key API->>DB: Commit task and dispatch intent API-->>UI: Task ID D->>DB: Claim committed intent D->>W: Enqueue invocation W->>DB: Save stages and progress UI->>API: Read events after saved cursor API-->>UI: Persisted progress or result
Invariants with enforcement
Section titled “Invariants with enforcement”An idempotency key cannot be reused with different input, so a network retry cannot silently change accepted work. Dispatch is recorded with the command before the worker starts; a crash between acceptance and enqueue can be recovered from that record. Clarification continues the same task, and replay reuses completed stages instead of repeating reasoning against changed inputs. Cancellation is persisted rather than inferred from a closed browser tab.
See Workers & recovery for stale-attempt publication and Model gateway for external paid-call uncertainty. Those pages own the guarantees and their failure cases.
A dynamic graph planner and general-purpose child-agent orchestration are target work. The current capability code uses a bounded research/synthesis/assessment sequence; the diagram does not imply a completed autonomous agent platform.