feat: extend voice API workspace and speech controls

src/server/voiceApi.ts

@@ -31,6 +31,7 @@ const EVENT_OPEN_TIMEOUT_MS = 10_000;
 const TURN_TIMEOUT_MS = 120_000;
 const WORKING_PROGRESS_INTERVAL_MS = 15_000;
 const DEVICE_WORKSPACE_ROOT = "/home/hope/workspaces";
+const VOICE_CONVERSATION_INSTRUCTIONS = "This is a voice conversation. Respond in natural, concise spoken language suitable for Azure DragonHD text-to-speech. Do not use Markdown, headings, tables, bullet lists, code fences, URLs unless explicitly requested, or visual-only references. Speak punctuation naturally and expand ambiguous symbols when useful.";
 interface TokenRecord {
   id: string;
   hash: string;
@@ -608,7 +609,11 @@ async function runTurn(
   speech: VoiceSpeechGateway,
   text: string
 ): Promise<void> {
-  const prompt = conversation.context === undefined ? text : `${conversation.context}\n\n${text}`;
+  const prompt = [
+    VOICE_CONVERSATION_INSTRUCTIONS,
+    ...(conversation.context === undefined ? [] : [conversation.context]),
+    text,
+  ].join("\n\n");
   // The event subscription is fully open before the prompt is submitted, so
   // a fast Pi run cannot emit agent_settled before this device starts waiting.
   const { settled } = await startSettledWaiter(