Core Concepts

Core concepts

English | 简体中文 | Home

Concept	Responsibility
Dispatcher / Task	The 0.7 managed API for submitting a local task, waiting for its result and reopening its durable handle after restart.
Handler	Application function that receives payload and execution context.
Kernel / Runtime	Kernel stores execution authority; Runtime executes registered handlers.
Execution command	Identifies one execution, handler binding, payload, timeout and retry policy.
Run / task	A Run groups application work. A task records its dependencies and business attempts.
Host	Drives execution and delivery while its process remains alive.
Effect	Records an external operation's intent and receipt for explicit recovery.
Recovery / generation	Recovery reopens an eligible failed or cancelled Run; generation separates new work from its history.
Notification / inbox	Delivers outcomes to the application; a durable inbox deduplicates receipt.
Observation / budget	Observations report saved execution facts; budgets carry the deadlines that constrain work.

With a started Dispatcher, Dispatcher.submit() creates the Run and task, wakes the background Host and returns a durable Task handle. Start the Dispatcher with app.start() or a with block. The application can wait for the result or consume saved result notifications. See managed tasks for the short path.

The lower-level workflow creates a Run, records a task and dispatch intent, delivers the command, executes it, then synchronizes results. Use it when the application needs dependencies, waits, explicit recovery or control over Run decisions. The application validates the result and explicitly chooses the next operation or finishes the Run.

submit_task() persists intent; it does not execute the handler. Manual loops use flush(), runtime.run_once() and sync(). OrchestratorHost drives the background loop. Creating an Orchestrator alone starts no scheduler.

Execution success, valid output structure and business acceptance are separate. Dependencies must be settled before dispatch, but the application must check whether their outcomes are acceptable. Success does not automatically dispatch a successor or finish a Run. An open wait prevents finish; it does not automatically block task dispatch.

Handle submission replay, execution retries, business rework and notification redelivery separately. Replay requests with their original identity and content; use a new decision and identity for new business work. Reconcile uncertain external effects using evidence before retrying them.

Effect recovery and Run recovery solve different problems. Effect recovery settles one uncertain external operation. Run recovery retains the Run's identity and history, then reopens an eligible failed or cancelled Run under its next generation. Events, attempts, results and notifications carry the generation that produced them, so late data from generation 0 remains distinguishable from new work.

Task.observe() and Task.events() read persisted execution facts without claiming work or changing its deadline. Handlers report activity through HandlerContext.activity; see execution activity and supervision for the limits of those facts and the optional stall workflow.

See SDK contracts, recovery and output validation for the detailed rules.