Fix AI output showing UTF-8 bytes as hex escapes instead of Chinese.

Decode <0xE5><0xA7><0xA4> style model output to proper characters; add prompt rule to use normal Chinese text.

Co-authored-by: Cursor <cursoragent@cursor.com>

lib/ai/client-stream.ts

@@ -1,4 +1,9 @@
 import type { AiRequestBody } from "@/lib/ai/types";
+import { decodeHexByteEscapes } from "@/lib/ai/decode-text";
+
+function emitText(text: string, onUpdate: (text: string) => void) {
+  onUpdate(decodeHexByteEscapes(text));
+}
 
 function parseApiError(text: string, status: number): string {
   const trimmed = text.trim();
@@ -41,11 +46,11 @@ export async function streamAiCompletion(
       break;
     }
     text += decoder.decode(value, { stream: true });
-    onUpdate(text);
+    emitText(text, onUpdate);
   }
 
   text += decoder.decode();
-  onUpdate(text);
+  emitText(text, onUpdate);
 
   if (!text.trim()) {
     throw new Error("AI 返回内容为空，请检查模型配置或稍后重试");