Replaces batch AI responses with real-time SSE streaming from OpenRouter.
Tokens are buffered client-side and drained via requestAnimationFrame for
a smooth typing effect instead of choppy chunk dumps.
Backend:
- Rewrite openrouter service for SSE streaming with incremental tool call accumulation
- Add AiStreamChunk/AiStreamEnd WebSocket event types
- Stream content deltas to clients during all tool call rounds
- Increase broadcast channel capacity (256 -> 4096) and handle Lagged errors gracefully
Frontend:
- Add StreamBuffer utility with adaptive rAF-based character draining
- Show streaming message-bubble with blinking cursor during generation
- Clean up buffer on room switch and final message replacement
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Jason Tudisco
4a002c85d4