LLM 只需处理极短的 "Questions sent to user." 而不是完整的问题 JSON，减少了 token 消耗和延迟

routes/chat.py

@@ -144,12 +144,11 @@ async def enhanced_generate_stream_response(
                         is_ui_resource = (
                             msg.text
                             and msg.text.lstrip().startswith('{"')
-                            and (
-                                ('"ui://' in msg.text and ('"text/html' in msg.text or '"text/uri-list' in msg.text))
-                                or '"__ask_user__"' in msg.text
-                            )
+                            and '"ui://' in msg.text
+                            and ('"text/html' in msg.text or '"text/uri-list' in msg.text)
                         )
-                        if config.tool_response or is_ui_resource:
+                        is_ask_user = msg.name == 'ask_user'
+                        if config.tool_response or is_ui_resource or is_ask_user:
                             new_content = f"[{message_tag}] {msg.name}\n{msg.text}\n"
 
                     # Collect full content

skills/common/mcp-ui/ui_render_server.py

@@ -21,32 +21,20 @@ from mcp_common import (
 )
 
 
-ASK_USER_MARKER = "__ask_user__"
+ASK_USER_RESPONSE = "Questions sent to user."
 
 
-def ask_user(questions: list) -> Dict[str, Any]:
-    """Create an ask_user response.
+def ask_user() -> Dict[str, Any]:
+    """Return a minimal fixed response for ask_user tool.
 
-    Args:
-        questions: List of dicts, each with "question", "options", and "multi_select".
-
-    Returns a JSON structure with a marker so the backend can detect it
-    and emit it as a special delta.ask_user event at the end of the stream.
+    The actual questions/options are already in the TOOL_CALL arguments,
+    so the frontend parses them directly from there. This response only
+    serves to acknowledge the tool call and minimize token usage in the
+    subsequent LLM inference round.
     """
-    normalized = []
-    for q in questions:
-        normalized.append({
-            "question": q.get("question", ""),
-            "options": q.get("options", []),
-            "multi_select": q.get("multi_select", False),
-        })
-    payload = {
-        "__type__": ASK_USER_MARKER,
-        "questions": normalized,
-    }
     return {
         "content": [
-            {"type": "text", "text": json.dumps(payload, ensure_ascii=False)}
+            {"type": "text", "text": ASK_USER_RESPONSE}
         ]
     }
 
@@ -168,7 +156,7 @@ async def handle_request(request: Dict[str, Any]) -> Dict[str, Any]:
                         request_id, -32602, "Missing required parameter: questions"
                     )
 
-                result = ask_user(questions)
+                result = ask_user()
                 return {"jsonrpc": "2.0", "id": request_id, "result": result}
 
             else: