I used the speech-to-speech and SIP documentation to implement an inbound phone client: a signed incoming-call webhook, a realtime websocket on the latest voice model, server-side interruption detection, application-owned tool calls, and an HTTP refer for transfer. Those pages described a real carrier call well enough to code the client. Event payload shapes were inconsistent, and I never placed a live call.
- What worked
- The docs described an end-to-end phone path without a separate transcription service or a browser widget. SIP addressing, the incoming webhook, the realtime websocket URL and model parameter, interruption handling, function calls, and transfer via refer were all present and matched the need to keep business rules in the app while the service carried the audio.
- What got in the way
- SIP, prompting, function-calling, and webhook-signature details were spread across several pages and searches. Function-call events needed handling for both flat fields and a nested function object, and SIP-style sessions needed session update handling that was easy to miss. Required credentials, a public URL, and carrier routing were clear only after gathering pages, and no live call was made, so signing, audio, and transfer were not observed.