Assistant request

When a call starts, OnCore POSTs an assistant-request event to your server. It carries the same identifiers the end-of-call report will use, so it's the moment to pre-load customer context: look the caller up in your CRM, warm caches, or open a record you'll complete when the call ends.

Delivery#

Sent as an HTTP POST with a JSON body to your configured webhook URL (the assistant's serverUrl if set, otherwise your tenant-level server URL), verified with the plaintext x-sadie-core-secret header. See the webhooks overview. Deliveries are not retried; the call proceeds regardless of your response.

Payload#

{
  "type": "assistant-request",
  "timestamp": 1753872000000,
  "status": "ringing",
  "customer_number": "+15555550111",
  "tenant_id": "0195f1e2-1111-7000-8000-000000000000",
  "assistant_id": "0195f1e2-2222-7000-8000-000000000000",
  "call_id": "0195f1e2-3333-7000-8000-000000000000",
  "restaurant_number": "+15555550123",
  "call_source": "pstn"
}

Responding#

Respond with a 2xx quickly. Your JSON response body can optionally customize the call that is starting:

{
  "data": {
    "custom_prompts": [
      { "title": "Customer context", "content": "The caller has an active booking." }
    ],
    "overrides": {
      "first_message": "Welcome back. How can I help with your booking?",
      "prompt_variables": { "customer_tier": "gold" },
      "voice": {
        "id": "0195f1e2-4444-7000-8000-000000000000",
        "settings": {
          "speed": 1.1,
          "locale": "en-GB",
          "accent": "general-american",
          "normalization": "auto"
        }
      }
    }
  }
}

Custom prompt sections are appended to the call's system prompt. prompt_variables are applied after the prompt is assembled, so they can resolve placeholders in saved and custom prompt text. A non-empty first_message replaces the saved greeting.

Voice, transcriber, and model overrides are looked up by ID. A valid ID replaces the saved choice for this call; an unknown ID is ignored and the saved configuration remains in use. The same three channels also accept a settings object, which merges provider settings over whatever that channel resolved to, so a single response can swap a transcriber and retune it at once. The keys each model accepts are listed under Providers, and every one of them is overridable per call. The example above uses the Cartesia voice settings: speed, plus the locale, accent and normalization of sonic-3.5 and sonic-3.6. A key the chosen model does not accept is passed through to the provider as sent, so a wrong value fails at the provider rather than here. These response fields are read only for assistant-request; response bodies from end reports do not alter a completed call.

Next steps#

• Verifying webhooks: confirm the delivery really came from OnCore.
• End-of-call report: the delivery at the other end of the call.