Patch Changes
-
7e55599: feat(gateway): support Live models over WebSocket
gateway.experimental_realtime()returns a continuous Live model foropenai/gpt-live-1andopenai/gpt-live-1.1. It connects to the Gateway Live route and speaks the normalized Live protocol (session-start, PCM16 audio,context-append,session-closewith final usage). Pass{ api: 'live' }for other Live model IDs or{ api: 'realtime' }to force the realtime route. Before, these model IDs opened the realtime route, which rejected them withinvalid_model.gateway.experimental_realtime.getToken()mints Live-bound client secrets for Live models.experimental_useRealtimeand the realtime session can connect a continuous model straight to the provider with a short-lived token (api.token) when the model explicitly declares theclient-secret-websocketconnection. Before, continuous WebSocket sessions required an application relay.
-
beccad3: chore(provider/gateway): update gateway model settings files