Streaming Tts And Websocket

Agent opens a connection to /tts/websocket but puts the API key only in a later message body, not where the WebSocket handshake expects it.

Authenticate the /tts/websocket connection per docs (API key + Cartesia-Version supplied where the handshake requires, e.g. query param or header per the documented scheme). Establish the socket before sending synthesis frames; handle a rejected handshake distinctly from a mid-stream drop.

Agent streams a multi-sentence utterance as several inputs over one socket but uses a fresh context_id per chunk, breaking prosodic continuity.

Reuse a stable context_id across inputs that belong to the same utterance so Sonic continues prosody/state coherently. Allocate a new context_id only when a genuinely new utterance begins. Track context_id lifecycle explicitly.

An upstream LLM streams tokens; the agent wants to feed text to /tts/websocket incrementally to start audio before the full sentence arrives.

Send transcript increments on the same context_id as they arrive, buffering to sensible boundaries (so words aren't split mid-token), and let Sonic emit audio incrementally. Maintain ordering and backpressure so text input never outpaces what the socket/consumer can absorb.