WebSocket endpoints

Routed in src/ws/attach.ts by URL path. Only one voice/call connection can be active at a time; the chat WS allows multiple clients.

Path	Audio codec	Handler	Use
/chat	—	src/session/chat.ts	Text chat
/browser-call	PCM16 24 kHz	src/session/browserCall.ts	Browser voice
/call	G.711 µ-law	src/session/call.ts	Twilio phone calls

/chat

Client → server

{ "type": "chat.message", "content": "..." }
{ "type": "history.more", "cursor": { "createdAtMs": 1700000000000, "id": "event-id" } }

Server → client

{ "type": "chat.working", "request_id": "..." }
{ "type": "chat.response", "content": "...", "conversation_id": "..." }
{ "type": "chat.done", "request_id": "..." }
{ "type": "history.page.start", "mode": "prepend" }
{ "type": "history.page", "mode": "prepend", "next_cursor": { "createdAtMs": 1699999999000, "id": "older-event-id" }, "has_more": true }

Timeline replay and live observability also use conversation.item.*, timeline.span.*, memory.*, inner.activation, context.usage, context.compacted, and context.capsule events. Initial replay is a bounded global event page; older cursor pages are bracketed by history.page.start and history.page so the client can prepend them atomically. Context and memory events render as Inner Plane activity rather than user messages. See src/session/chat.ts and src/session/history.ts for the complete mapping.

/browser-call

Binary audio frames in both directions (base64-wrapped JSON messages), plus control frames for VAD, barge-in, and response lifecycle. The pipeline lives in src/voice/.

Key control messages:

• session.start (client) — opens a Realtime session
• input_audio_buffer.append (client) — incoming mic chunk
• media (server) — outbound speech chunk
• response.audio.done (model event handled by the server) — assistant finished generating audio
• input_audio_buffer.speech_started (model event) — realtime VAD interruption trigger

/call

Twilio-specific framing: each frame is { event: "media", media: { payload: <base64 µ-law> }}. The server declares the stream as audio/pcmu and passes the µ-law payload directly to and from the OpenAI Realtime API. Same barge-in logic as /browser-call.

Barge-in invariant

In both voice paths, barge-in is handled natively by the Realtime API (interrupt_response: true). The server no longer drives interruption by proactively sending response.cancel as primary behavior; instead it flushes downstream playback buffers and suppresses stale deltas while the model auto-cancels in-flight output.