v4.7.4 — Direct Live Voice

Google Live Voice (Direct — No Gateway)

• Direct WebSocket to Google BidiGenerateContent API
• 34 curated voice tools with full execution
• Session resumption (instant reconnect, 2hr token validity)
• Context window compression (unlimited session duration)
• 20ms heartbeat ping (prevents NAT drops)
• Interruption handling (barge-in clears playback)
• thinkingLevel: minimal (lowest latency)
• Input/output audio transcription
• Multimodal during voice (send images/text while on call)
• GoAway handling for graceful server disconnects

Optimizations

• System buffer audio capture (device-native, reliable)
• No client-side VAD (Google server VAD handles detection)
• Trimmed system prompt for voice (skip UI docs)
• Pre-connect WebSocket on chat load
• Binary frame handling (Gemini 3.1 format)
• All parts processed per event (audio + transcript simultaneous)

Also in this release

• Cloudflare Workers AI provider + image gen
• File server: WiFi lock, 256KB buffers
• Windows deploy: SFTP fix
• KeyGen: RSA fallback, off main thread
• DeepSeek DSML loop detection
• Gemini thought_signature for tool calls