real-time

Real-time streaming across SSE, NDJSON, and WebSocket channels is coordinated by using an SSE broadcast registry for multi-user events, NDJSON for sender streaming, and WebSocket for voice audio integration. This separation prevents duplicate broadcasts and latency.

NDJSON streaming is needed for sender-side chunked AI responses, while SSE broadcasting is used for multi-user flow events and presence updates. Separating sender streaming from broadcasts avoids content duplication and reduces update latency.

Per-flow connection management and cancellation signals are handled through a dedicated flow-level connection registry. This registry controls runtimes by tracking active streaming sessions and issuing cancellation signals to terminate specific live flows.

WebSocket voice channels support real-time audio conversations and transcripts using proxy-friendly headers. This allows real-time voice integration to function correctly even when network traffic passes through intermediary proxy servers.

Live progress indicators during AI agent responses are implemented by starting an agent session and enabling NDJSON and SSE streams. This allows the application to observe and render chunked content updates as they are generated.

Duplicate live updates are avoided by maintaining a clear separation between sender-side NDJSON streaming and SSE broadcasting. This architectural distinction ensures that chunked content is sent once and broadcast appropriately without redundant transmissions.