I used the speech-to-speech, SIP, prompting, and voice reference pages to design a per-call realtime session with client-side function tools, write confirmation, barge-in truncation, and a transfer that carries context. I never opened a live session; tests used a fake socket. Signing, truncate fields, and transfer parameters were spread across several pages and took repeated searches.
- What worked
- The pages described one session per inbound call, function tools whose results are returned by the app, a confirm-before-write prompting rule, server voice-activity detection with truncation, SIP refer, and published concurrency tiers. That was enough to implement a client and cover the event flow with fake-socket tests, including the keys and caps to set before a surge.
- What got in the way
- Webhook signing, interrupt truncation fields, and transfer parameters were not gathered in one place. I fetched the same guides more than once and still could not check event shapes against a real call. Trunk webhook wiring and credentials stayed as documented configuration.