workflow

Add the "use workflow" directive as the first line of an async function to make it durable, and use "use step" for individual retryable units. Start the workflow from an API route with start() from workflow/api, which returns a run ID immediately.

Use createHook() inside the workflow with a deterministic token, then call resumeHook(token, data) from an API route to deliver the event. Hooks also implement AsyncIterable, so for await...of loops can receive multiple events over time.

Workflow functions run in a sandboxed VM without fetch, setTimeout, or Node.js modules. Import fetch from the workflow package and assign it to globalThis.fetch, use sleep() instead of timers, or move the logic into a "use step" function with full Node.js access.

Yes, use DurableAgent from @workflow/ai/agent, which handles the sandbox fetch setup automatically. Tool execute functions needing Node.js access should use "use step", while those using workflow primitives like sleep or createHook should run at the workflow level.

Use the @workflow/vitest plugin for integration tests, which runs workflows in-process without a server. Helpers like waitForHook, waitForSleep, resumeHook, and getRun(runId).wakeUp() let you control pause points and assert on run.returnValue.

All arguments and return values must be serializable: strings, numbers, booleans, plain objects, arrays, Date, Map, Set, URL, Headers, streams, and typed arrays are supported. Functions, class instances, Symbols, and WeakMap/WeakSet are not supported.