Define, deploy, and trigger a durable multi-step Cloudflare Workflow with retries, sleeps, and human-in-the-loop event waits

Documented steps

1. Workflows ship with the Workers runtime — no extra npm package is needed. Work inside a Wrangler-based Workers project (`npm create cloudflare@latest`).
2. Define the workflow class: `export class MyWorkflow extends WorkflowEntrypoint<Env, Params> { async run(event: WorkflowEvent<Params>, step: WorkflowStep) { ... } }`. `event.payload` holds the params passed at creation; `event.timestamp`, `event.instanceId`, and `event.workflowName` are also available.
3. Wrap EVERY side-effecting or non-deterministic operation (API calls, DB writes, Date.now(), Math.random()) inside `await step.do('unique step name', async () => { ... return serializableResult; })`. Step names act as cache keys and must be deterministic — never interpolate a timestamp or random value into a step name.
4. Add per-step retry and timeout config: `await step.do('name', { retries: { limit: 5, delay: '10 seconds', backoff: 'exponential' }, timeout: '5 minutes' }, async () => {...})`.
5. Pause without burning compute using `await step.sleep('name', '1 hour')` or `await step.sleepUntil('name', someDate)`. Instances in the waiting state do not consume a concurrency slot, so millions can sleep simultaneously.
6. Implement human-in-the-loop with `await step.waitForEvent('name', { type: 'approved', timeout: '24 hours' })`, then unblock it externally via `instance.sendEvent({ type: 'approved', payload })`.
7. Abort retries immediately for permanent failures by throwing `new NonRetryableError('message')` from inside the `step.do()` callback — the instance transitions to errored instead of consuming its retry budget.
8. Register the binding in wrangler.jsonc: `"workflows": [{ "name": "my-workflow", "binding": "MY_WORKFLOW", "class_name": "MyWorkflow", "limits": { "steps": 25000 } }]`.
9. Trigger an instance from a Worker via the binding: `const instance = await env.MY_WORKFLOW.create({ id: 'user-123', params: { hello: 'world' } })`. For bulk triggering use `createBatch(arrayOfOptions)` (up to 100 per call) rather than looping `create()`.
10. Fetch an existing instance with `await env.MY_WORKFLOW.get(id)` (throws if the id does not exist), then `await instance.status()` returns `{ status, error, output }` where status is one of queued | running | paused | errored | terminated | complete | waiting | waitingForPause | unknown.
11. Control a running instance with `await instance.pause()`, `await instance.resume()`, `await instance.terminate()`, or `await instance.restart({ from: { name: 'stepName', count: N } })` to replay from a specific step while reusing cached earlier results.
12. To manage instances over HTTP instead of a binding: `POST /accounts/{account_id}/workflows/{workflow_name}/instances` to create, `POST .../instances/batch` for batch create, `GET .../instances/{instance_id}` for status, `PATCH .../instances/{instance_id}/status` to pause/resume/terminate, and `POST .../instances/{instance_id}/events/{event_type}` to send an event.
13. Keep each step's return value under 1 MiB (2^20 bytes); for larger binary output return a `ReadableStream<Uint8Array>` instead. Event payloads sent via `sendEvent` share the same 1 MiB cap.
14. Always `await` every step.* call. An unawaited step promise silently swallows errors and loses its cached return value; wrap any `Promise.race()` / `Promise.any()` inside a single `step.do()` so caching stays consistent across restarts and hibernation.
15. Official documentation (verify before relying on limits, which change): https://developers.cloudflare.com/workflows/ | https://developers.cloudflare.com/workflows/build/workers-api/ | https://developers.cloudflare.com/workflows/build/rules-of-workflows/ | https://developers.cloudflare.com/workflows/reference/limits/ | https://developers.cloudflare.com/workflows/build/trigger-workflows/