Gemini Live API Expert - Fix WebSocket Disconnects & Silent Sessions

Project Overview We run a real-time conversational voice application powered by the Gemini Live API. Our architecture is: client app « our Python backend (WebSocket relay) « Gemini Live API. We also use function calling / tools integrated into the backend. The product works, but we're hitting reliability issues during live sessions that we need an experienced engineer to diagnose and fix. The Problems 1. Frequent 1007 disconnects. Sessions drop with errors like: Connection lost (received 1007 (invalid frame payload data) Request contains an invalid argument.; then sent 1007 ...) — attempting reconnect We need to find what argument/payload is malformed (audio format, config, tool schema, message framing, etc.) and stop these from happening. 2. Silent sessions. Mid-session, Gemini sometimes simply stops responding — the WebSocket stays open but no audio/text comes back to the user. We need to understand why it goes silent (tool-call handling, turn completion, timeouts, backpressure on our relay) and make sessions recover gracefully. 3. Other reliability issues. These are the two we've identified so far, but there may be additional problems we haven't pinpointed yet — intermittent latency, edge cases in tool calls, or other instability under real usage. We'd like you to surface and address any related issues you find while investigating, not just the two above. What We're Looking For: • Proven experience with the Gemini Live API (or strong real-time streaming LLM voice experience) • Strong Python skills (async/concurrency, WebSocket servers) • Deep WebSocket knowledge — framing, close codes, keep-alive, reconnection strategy • Experience with function calling / tool use over a streaming connection • Comfortable debugging audio streaming (encoding, sample rate, chunking) • Able to read our backend code, reproduce the issues, and propose concrete fixes Deliverables • Root-cause diagnosis of the 1007 errors and the silent-session issue • Identification of any other reliability issues found during the investigation • Code fixes (or a clear, actionable plan if a quick fix isn't possible) • A more robust reconnection / error-handling approach so users aren't interrupted To Apply Please briefly describe a Gemini Live or real-time voice issue you've debugged before, and your initial hypothesis on what might cause a 1007 "invalid argument" in this setup.