update: 视觉模块直接识别输出，不经过二次LLM

main/xiaozhi-server/core/api/vision_handler.py

@@ -8,6 +8,7 @@ from config.config_loader import get_private_config_from_api
 from core.utils.auth import AuthToken
 import base64
 from typing import Tuple, Optional
+from plugins_func.register import Action
 
 TAG = __name__
 
@@ -122,7 +123,8 @@ class VisionHandler:
 
             return_json = {
                 "success": True,
-                "result": result,
+                "action": Action.RESPONSE.name,
+                "response": result,
             }
 
             response = web.Response(

main/xiaozhi-server/core/connection.py

@@ -751,9 +751,24 @@ class ConnectionHandler:
                             self.loop,
                         ).result()
                         self.logger.bind(tag=TAG).debug(f"MCP工具调用结果: {result}")
-                        result = ActionResponse(
-                            action=Action.REQLLM, result=result, response=""
-                        )
+
+                        if isinstance(result, str):
+                            try:
+                                result = json.loads(result)
+                            except Exception as e:
+                                self.logger.bind(tag=TAG).error(f"解析MCP工具返回结果失败: {e}")
+
+                        # 视觉大模型不经过二次LLM处理
+                        if isinstance(result, dict) and "action" in result:
+                            result = ActionResponse(
+                                action=Action[result["action"]],
+                                result=None,
+                                response=result.get("response", "")
+                            )
+                        else:
+                            result = ActionResponse(
+                                action=Action.REQLLM, result=result, response=""
+                            )
                     except Exception as e:
                         self.logger.bind(tag=TAG).error(f"MCP工具调用失败: {e}")
                         result = ActionResponse(