Voice-Md

A markdown editor where you edit by talking, and the agent answers in interface instead of text. Hover a talk zone, speak (“summarize this section”), hover out — the rewrite lands and a length dial materializes beneath it, pre-generated variants cached so scrubbing is instant. The model decides which affordance fits: dial, alternative cards, follow-up chip, or nothing. A2UI v0.9 is the generative UI layer (model emits JSONL into a React component catalog, progressive streaming). OpenAI Realtime runs the voice loop with hover replacing VAD. Milkdown is the editor. Six dependencies, working end-to-end.