我们构建了一个语音管道，提交了28个PR，并在一天内实现了完整的平台平衡

On March 20, 2026, the AI team successfully launched a full voice output pipeline, enabling agents to engage in spoken interactions. By calling the POST /canvas/speak endpoint, the server generates Kokoro TTS audio and streams it via SSE. A voice output event is triggered, containing information such as agent ID, duration, and text. The MediaPlayer on Android and iOS platforms streams the audio while displaying the agent's speaking state. The speaking state is automatically cleared after playback, with graceful degradation provided if Kokoro is unavailable. Additionally, a voiceId hash mismatch issue was fixed on the Node side, and the Fly.io VM was upgraded to 4GB to prevent auto-suspension.