> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://infonite.dev/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://infonite.dev/_mcp/server.

# On Execution Started

POST 

Fired when the execution leaves the queue and the engine starts working.

It is the moment the run stops being a promise: from here on there is a session being opened against the source. `status_reason` is `RUNNING` — nothing has been retrieved yet, and nothing is readable yet.

**Subscribe to it for the clock, not for the data.** Nothing about this event requires you to act: it is what separates *waiting in the queue* from *working*, which is the difference between a slow platform and a busy one when somebody asks you why a case is taking so long.

Somebody is now logging in on your customer's behalf. If your product shows a state, this is when it stops saying *queued*.

The gap between `202 Accepted` and this event is **queue time** — it depends on load and on whether the execution can pause for a person. Measure the work from here, not from the request.

**Not every execution reaches this event.** One rejected at validation, or aborted before it was ever picked up, goes straight to `ended`. Treat `ended` as the only event you are guaranteed to receive.

Reference: https://infonite.dev/api-reference/direct-executions/direct-executions-api/webhooks/webhook-direct-executions-started

## Request

### Payload

- `app_id` (string, required) — The application this execution was launched with — the one your secret belongs to. Worth keeping when your product uses more than one, a sandbox and a production app being the usual case: every record and every event we send carries it.
- `customer_id` (string, required) — The `customer_id` you supplied when the execution was initialised, returned as you sent it — so an answer can be routed to the right case with no lookup on your side.
- `execution_id` (string, required) — The execution this belongs to — the value you poll for the state and address every results endpoint with. Always present here: an event exists because an execution does.
- `event` ("started", required) — The event type: `started`.
- `date` (string, required) — When the event was raised, in UTC. Deliveries can arrive out of order — a retry is delivered after whatever came next — so order your own records by this, not by arrival.
- `customer_available` (boolean, required) — Whether the execution was started as one a person can answer (`customer_interaction_available`). When `false`, a challenge from the source cannot be solved and the execution ends instead of waiting.
- `status_reason` (enum or enum or enum or enum or enum or enum or enum or enum or enum or enum or enum, required) — The precise reason for the state at the moment of the event — one of the execution reasons, or a feature's own when `producer` names a feature. Branch on this and not on the message: it is a stable code. [Every reason by family](/direct-executions/lifecycle#every-reason-by-family).
- `session_id` (string, optional, nullable)
- `external_execution_id` (string, optional, nullable)
- `locking_hash` (string, optional, nullable)
- `event_class` ("ExecutionEvent", optional, default: ExecutionEvent) — Payload class discriminator — always this value for an execution event. If the same endpoint also receives flow events, branch on it first.
- `producer` ("engine" or enum, optional, default: engine) — Who raised it: `engine` for the execution itself — the login, typically — or the feature's own code when a single feature is what needs something. An `action_required` from the accounts reader says `accounts_read`, so you know what is being asked for before reading the form.
- `status_message` (string, optional, nullable)
- `data` (map from string to any, optional, nullable)