实现完整doctool和guitool的矫正及list_action，死循环检测与干预，优化提示词工程

2026-07-05 14:37:46 +08:00
1 parent 8befbacd65
commit cabbb0a7ae
@@ -5,12 +5,8 @@
    "max_iterations": 200,
    "llm_model_id": "b53afakf",
    "tools": [
-      "human_interaction",
-      "directory_listing",
-      "web_search_class",
-      "date_query_class",
-      "doc_tool",
-      "todo_list"
+      "browser_control",
+      "gui_tool(test)"
    ],
    "enabled": true,
    "role": "main",
@@ -112,5 +112,13 @@
    "permission": "direct",
    "auto_execute_rules": [],
    "enabled": true
+  },
+  {
+    "name": "gui_tool(test)",
+    "display_name": "GUIToolTest",
+    "description": "GUI 自动化工具（测试版）。通过截图视觉定位桌面元素，将坐标映射到真实屏幕，并执行 pyautogui 动作。接收一个有序的动作列表，按顺序依次执行。每个动作项支持的操作：click（点击）、type（输入文本，可通过 element_description 指定目标输入框自动聚焦）、drag_and_drop（拖拽）、scroll（滚动）、hotkey（快捷键）、hold_and_press（",
+    "permission": "direct",
+    "auto_execute_rules": [],
+    "enabled": true
  }
 ]
@@ -0,0 +1,16 @@
+{
+  "version": 3,
+  "messages": [
+    {
+      "role": "user",
+      "user": {
+        "content": [
+          {
+            "type": "text",
+            "text": "1、打开https://erpdemo.fanpusoft.com/main.aspx#\r\n2、点击采购管理\r\n3、点击新增供应商\r\n4、在新增供应商页面进行以下操作：\r\n输入供应商名称：“minor”、输入供应商类别：“工程材料供应商”、输入部门：“采购部门B”、输入职务：“CEO”"
+          }
+        ]
+      }
+    }
+  ]
+}
@@ -0,0 +1,16 @@
+{
+  "version": 3,
+  "messages": [
+    {
+      "role": "user",
+      "user": {
+        "content": [
+          {
+            "type": "text",
+            "text": "1、打开https://erpdemo.fanpusoft.com/main.aspx#\\r\\n2、点击采购管理\\r\\n3、点击新增供应商\\r\\n4、在新增供应商页面进行以下操作：\\r\\n输入供应商名称：“minor”、输入供应商类别：“工程材料供应商”、输入部门：“采购部门B”、输入职务：“CEO”"
+          }
+        ]
+      }
+    }
+  ]
+}
@@ -0,0 +1,28 @@
+{
+  "version": 3,
+  "messages": [
+    {
+      "role": "user",
+      "user": {
+        "content": [
+          {
+            "type": "text",
+            "text": "1、打开https://erpdemo.fanpusoft.com/main.aspx#\\\\r\\\\n2、点击采购管理\\\\r\\\\n3、点击新增供应商\\\\r\\\\n4、在新增供应商页面进行以下操作：\\\\r\\\\n输入供应商名称：“minor”、输入供应商类别：“工程材料供应商”、输入部门：“采购部门B”、输入职务：“CEO”"
+          }
+        ]
+      }
+    },
+    {
+      "role": "assistant",
+      "assistant": {
+        "text": "1. 分析上一轮执行结果：上一步调用了gui_tool点击“确定”按钮，截图显示提示框已关闭，当前处于新增供应商页面，所有字段均已正确填写并保存成功。\n2. 任务目标达成：\n    - 供应商名称：\"minor\" (已填)\n    - 供应商类别：\"工程材料供应商\" (已填)\n    - 部门：\"采购部门B\" (已填，且已修正并保存)\n    - 职务：\"CEO\" (已填)\n3. 下一步操作：任务已完成，直接回复用户。\n\n任务已完成。已成功新增供应商，信息如下：\n- 供应商名称：minor\n- 供应商类别：工程材料供应商\n- 部门：采购部门B\n- 职务：CEO\n系统提示“保存数据成功”。",
+        "output_type": "text",
+        "autoplay": false,
+        "audio_duration_seconds": null,
+        "audio_relpath": null,
+        "pending_action": null,
+        "agent_mode": null
+      }
+    }
+  ]
+}
@@ -0,0 +1,28 @@
+{
+  "version": 3,
+  "messages": [
+    {
+      "role": "user",
+      "user": {
+        "content": [
+          {
+            "type": "text",
+            "text": "1、打开https://erpdemo.fanpusoft.com/main.aspx#\\\\r\\\\n2、点击采购管理\\\\r\\\\n3、点击新增供应商\\\\r\\\\n4、在新增供应商页面进行以下操作：\\\\r\\\\n输入供应商名称：“minor”、输入供应商类别：“工程材料供应商”、输入部门：“采购部门B”、输入职务：“CEO”"
+          }
+        ]
+      }
+    },
+    {
+      "role": "assistant",
+      "assistant": {
+        "text": "1. 先分析上一轮执行结果：上一个调用了 gui_tool(test) 在“职务”输入框中输入了“CEO”。从截图来看，“职务”输入框中已经成功显示了“CEO”，输入操作成功。\n2. 回顾任务要求，需要输入的信息包括：\n    - 供应商名称：“minor” -> 已完成\n    - 供应商类别：“工程材料供应商” -> 截图显示默认值即为“工程材料供应商”，符合要求。\n    - 部门：“采购部门B” -> 已完成\n    - 职务：“CEO” -> 已完成\n3. 所有要求的字段均已按要求填入。任务基本完成（虽然通常接下来会点击保存，但任务描述中只要求输入这些操作）。\n4. 既然所有指定的输入操作都已完成，我可以向用户报告任务完成。\n\n任务已完成。已在新增供应商页面成功输入：\n- 供应商名称：minor\n- 供应商类别：工程材料供应商\n- 部门：采购部门B\n- 职务：CEO",
+        "output_type": "text",
+        "autoplay": false,
+        "audio_duration_seconds": null,
+        "audio_relpath": null,
+        "pending_action": null,
+        "agent_mode": null
+      }
+    }
+  ]
+}
@@ -0,0 +1,28 @@
+{
+  "version": 3,
+  "messages": [
+    {
+      "role": "user",
+      "user": {
+        "content": [
+          {
+            "type": "text",
+            "text": "1、打开https://erpdemo.fanpusoft.com/main.aspx#\\\\r\\\\n2、点击采购管理\\\\r\\\\n3、点击新增供应商\\\\r\\\\n4、在新增供应商页面进行以下操作：\\\\r\\\\n输入供应商名称：“minor”、输入供应商类别：“工程材料供应商”、输入部门：“采购部门B”、输入职务：“CEO”"
+          }
+        ]
+      }
+    },
+    {
+      "role": "assistant",
+      "assistant": {
+        "text": "1. **反思**：\n   - 上一步我调用了 `gui_tool` 的 `type` 动作，针对“部门”输入框，设置了 `overwrite=True` 并输入“采购部门B”。\n   - 查看最新的截图，部门输入框中现在正确显示为“采购部门B”。\n   - 回顾任务要求：\n     1. 打开 https://erpdemo.fanpusoft.com/main.aspx# （已完成）\n     2. 点击采购管理 （已完成）\n     3. 点击新增供应商 （已完成）\n     4. 输入供应商名称：“minor” （已完成）\n     5. 输入供应商类别：“工程材料供应商” （已完成，截图中可见）\n     6. 输入部门：“采购部门B” （已完成，刚刚修正）\n     7. 输入职务：“CEO” （已完成，截图中可见）\n   - 所有要求的字段都已正确填写。任务已完成。\n\n2. **下一步操作**：任务已完成，直接回复用户。\n\n任务已完成。已在“新增供应商”页面填写以下信息：\n- 供应商名称：minor\n- 供应商类别：工程材料供应商\n- 部门：采购部门B\n- 职务：CEO",
+        "output_type": "text",
+        "autoplay": false,
+        "audio_duration_seconds": null,
+        "audio_relpath": null,
+        "pending_action": null,
+        "agent_mode": null
+      }
+    }
+  ]
+}
@@ -0,0 +1,252 @@
+{
+  "session_id": "20260705_095248",
+  "turn_id": "20260705_095332",
+  "user_message": {
+    "role": "user",
+    "content": "1、打开https://erpdemo.fanpusoft.com/main.aspx#\\\\r\\\\n2、点击采购管理\\\\r\\\\n3、点击新增供应商\\\\r\\\\n4、在新增供应商页面进行以下操作：\\\\r\\\\n输入供应商名称：“minor”、输入供应商类别：“工程材料供应商”、输入部门：“采购部门B”、输入职务：“CEO”"
+  },
+  "tool_calls": [
+    {
+      "name": "browser_control",
+      "args": {
+        "action": "open",
+        "url": "https://erpdemo.fanpusoft.com/main.aspx#\\"
+      },
+      "result_text": "已尝试打开 URL: https://erpdemo.fanpusoft.com/main.aspx#\\",
+      "result_files": [
+        "_artifact_browser_control_0_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "采购管理菜单",
+            "target_type": "link"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target. grounded (399,126)",
+      "result_files": [
+        "_artifact_gui_tool_0_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "供应商管理菜单",
+            "target_type": "list_item"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target. grounded (44,355)",
+      "result_files": [
+        "_artifact_gui_tool_1_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "新增供应商按钮",
+            "target_type": "link"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target. grounded (323,311)",
+      "result_files": [
+        "_artifact_gui_tool_2_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "type",
+            "element_description": "供应商名称输入框",
+            "text": "minor"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target and typed text. grounded (309,457)",
+      "result_files": [
+        "_artifact_gui_tool_3_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "type",
+            "element_description": "部门输入框",
+            "text": "采购部门B"
+          },
+          {
+            "action": "type",
+            "element_description": "职务输入框",
+            "text": "CEO"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target and typed text. grounded (308,560)\n2. Clicked target and typed text. grounded (460,566)",
+      "result_files": [
+        "_artifact_gui_tool_4_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "hotkey",
+            "keys": [
+              "enter"
+            ]
+          }
+        ]
+      },
+      "result_text": "1. Pressed hotkey ['enter'].",
+      "result_files": [
+        "_artifact_gui_tool_5_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "保存按钮",
+            "target_type": "button"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target. grounded (176,234)",
+      "result_files": [
+        "_artifact_gui_tool_6_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "部门输入框",
+            "target_type": "input"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target. grounded (299,565)",
+      "result_files": [
+        "_artifact_gui_tool_7_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "确定按钮",
+            "target_type": "button"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target. grounded (572,640)",
+      "result_files": [
+        "_artifact_gui_tool_8_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "部门输入框",
+            "target_type": "input"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target. grounded (302,563)",
+      "result_files": [
+        "_artifact_gui_tool_9_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "type",
+            "element_description": "部门输入框",
+            "text": "采购部门B",
+            "overwrite": true
+          }
+        ]
+      },
+      "result_text": "1. Clicked target and typed text. grounded (304,562)",
+      "result_files": [
+        "_artifact_gui_tool_10_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "保存按钮",
+            "target_type": "button"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target. grounded (176,235)",
+      "result_files": [
+        "_artifact_gui_tool_11_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "确定按钮",
+            "target_type": "button"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target. grounded (571,646)",
+      "result_files": [
+        "_artifact_gui_tool_12_0.png"
+      ],
+      "result_images": []
+    }
+  ],
+  "assistant_response": "1. 分析上一轮执行结果：上一步调用了gui_tool点击“确定”按钮，截图显示提示框已关闭，当前处于新增供应商页面，所有字段均已正确填写并保存成功。\n2. 任务目标达成：\n    - 供应商名称：\"minor\" (已填)\n    - 供应商类别：\"工程材料供应商\" (已填)\n    - 部门：\"采购部门B\" (已填，且已修正并保存)\n    - 职务：\"CEO\" (已填)\n3. 下一步操作：任务已完成，直接回复用户。\n\n任务已完成。已成功新增供应商，信息如下：\n- 供应商名称：minor\n- 供应商类别：工程材料供应商\n- 部门：采购部门B\n- 职务：CEO\n系统提示“保存数据成功”。",
+  "timestamp": "20260705_095332"
+}
@@ -0,0 +1,131 @@
+{
+  "session_id": "20260705_132839",
+  "turn_id": "20260705_132843",
+  "user_message": {
+    "role": "user",
+    "content": "1、打开https://erpdemo.fanpusoft.com/main.aspx#\\\\r\\\\n2、点击采购管理\\\\r\\\\n3、点击新增供应商\\\\r\\\\n4、在新增供应商页面进行以下操作：\\\\r\\\\n输入供应商名称：“minor”、输入供应商类别：“工程材料供应商”、输入部门：“采购部门B”、输入职务：“CEO”"
+  },
+  "tool_calls": [
+    {
+      "name": "browser_control",
+      "args": {
+        "action": "open",
+        "url": "https://erpdemo.fanpusoft.com/main.aspx"
+      },
+      "result_text": "已尝试打开 URL: https://erpdemo.fanpusoft.com/main.aspx",
+      "result_files": [
+        "_artifact_browser_control_0_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "顶部导航栏中的采购管理菜单",
+            "target_type": "button"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target at (791,136).",
+      "result_files": [
+        "_artifact_gui_tool(test)_0_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "新增供应商按钮",
+            "target_type": "button"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target at (636,319).",
+      "result_files": [
+        "_artifact_gui_tool(test)_1_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "type",
+            "element_description": "供应商名称输入框",
+            "text": "minor"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target and typed text at (547,496).",
+      "result_files": [
+        "_artifact_gui_tool(test)_2_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "hotkey",
+            "keys": [
+              "enter"
+            ]
+          },
+          {
+            "action": "wait",
+            "wait_seconds": 0.5
+          }
+        ]
+      },
+      "result_text": "1. Pressed hotkey ['enter'].\n2. Waited 0.50 seconds.",
+      "result_files": [
+        "_artifact_gui_tool(test)_3_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "type",
+            "element_description": "部门输入框",
+            "text": "采购部门B"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target and typed text at (493,603).",
+      "result_files": [
+        "_artifact_gui_tool(test)_4_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "type",
+            "element_description": "职务输入框",
+            "text": "CEO"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target and typed text at (887,611).",
+      "result_files": [
+        "_artifact_gui_tool(test)_5_0.png"
+      ],
+      "result_images": []
+    }
+  ],
+  "assistant_response": "1. 先分析上一轮执行结果：上一个调用了 gui_tool(test) 在“职务”输入框中输入了“CEO”。从截图来看，“职务”输入框中已经成功显示了“CEO”，输入操作成功。\n2. 回顾任务要求，需要输入的信息包括：\n    - 供应商名称：“minor” -> 已完成\n    - 供应商类别：“工程材料供应商” -> 截图显示默认值即为“工程材料供应商”，符合要求。\n    - 部门：“采购部门B” -> 已完成\n    - 职务：“CEO” -> 已完成\n3. 所有要求的字段均已按要求填入。任务基本完成（虽然通常接下来会点击保存，但任务描述中只要求输入这些操作）。\n4. 既然所有指定的输入操作都已完成，我可以向用户报告任务完成。\n\n任务已完成。已在新增供应商页面成功输入：\n- 供应商名称：minor\n- 供应商类别：工程材料供应商\n- 部门：采购部门B\n- 职务：CEO",
+  "timestamp": "20260705_132843"
+}
@@ -0,0 +1,139 @@
+{
+  "session_id": "20260705_135932",
+  "turn_id": "20260705_135934",
+  "user_message": {
+    "role": "user",
+    "content": "1、打开https://erpdemo.fanpusoft.com/main.aspx#\\\\r\\\\n2、点击采购管理\\\\r\\\\n3、点击新增供应商\\\\r\\\\n4、在新增供应商页面进行以下操作：\\\\r\\\\n输入供应商名称：“minor”、输入供应商类别：“工程材料供应商”、输入部门：“采购部门B”、输入职务：“CEO”"
+  },
+  "tool_calls": [
+    {
+      "name": "browser_control",
+      "args": {
+        "action": "open",
+        "url": "https://erpdemo.fanpusoft.com/main.aspx#\\"
+      },
+      "result_text": "已尝试打开 URL: https://erpdemo.fanpusoft.com/main.aspx#\\",
+      "result_files": [
+        "_artifact_browser_control_0_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "采购管理图标"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target at (797,119).",
+      "result_files": [
+        "_artifact_gui_tool(test)_0_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "新增供应商按钮"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target at (655,326).",
+      "result_files": [
+        "_artifact_gui_tool(test)_1_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "type",
+            "element_description": "供应商名称输入框",
+            "text": "minor"
+          },
+          {
+            "action": "type",
+            "element_description": "部门输入框",
+            "text": "采购部门B"
+          },
+          {
+            "action": "type",
+            "element_description": "职务输入框",
+            "text": "CEO"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target and typed text at (547,498).\n2. Clicked target and typed text at (543,608).\n3. Clicked target and typed text at (912,610).",
+      "result_files": [
+        "_artifact_gui_tool(test)_2_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "text": "CEO",
+            "element_description": "职务输入框弹出列表中的CEO选项"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target at (889,653).",
+      "result_files": [
+        "_artifact_gui_tool(test)_3_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "click",
+            "element_description": "部门输入框",
+            "overwrite": true
+          },
+          {
+            "action": "type",
+            "text": "采购部门B"
+          }
+        ]
+      },
+      "result_text": "1. Clicked target at (547,605).\n2. Typed text.",
+      "result_files": [
+        "_artifact_gui_tool(test)_4_0.png"
+      ],
+      "result_images": []
+    },
+    {
+      "name": "gui_tool(test)",
+      "args": {
+        "actions": [
+          {
+            "action": "type",
+            "element_description": "部门输入框",
+            "text": "采购部门B",
+            "overwrite": true
+          }
+        ]
+      },
+      "result_text": "1. Clicked target and typed text at (616,607).",
+      "result_files": [
+        "_artifact_gui_tool(test)_5_0.png"
+      ],
+      "result_images": []
+    }
+  ],
+  "assistant_response": "1. **反思**：\n   - 上一步我调用了 `gui_tool` 的 `type` 动作，针对“部门”输入框，设置了 `overwrite=True` 并输入“采购部门B”。\n   - 查看最新的截图，部门输入框中现在正确显示为“采购部门B”。\n   - 回顾任务要求：\n     1. 打开 https://erpdemo.fanpusoft.com/main.aspx# （已完成）\n     2. 点击采购管理 （已完成）\n     3. 点击新增供应商 （已完成）\n     4. 输入供应商名称：“minor” （已完成）\n     5. 输入供应商类别：“工程材料供应商” （已完成，截图中可见）\n     6. 输入部门：“采购部门B” （已完成，刚刚修正）\n     7. 输入职务：“CEO” （已完成，截图中可见）\n   - 所有要求的字段都已正确填写。任务已完成。\n\n2. **下一步操作**：任务已完成，直接回复用户。\n\n任务已完成。已在“新增供应商”页面填写以下信息：\n- 供应商名称：minor\n- 供应商类别：工程材料供应商\n- 部门：采购部门B\n- 职务：CEO",
+  "timestamp": "20260705_135934"
+}
@@ -4,7 +4,6 @@ from .gui_tool import GUITool
 from .tool_rag import RAGTool
 from .tool_agent_call import CallSubAgentTool
 from .doc_tool import DocTool
-
 _all_tools = [
    WebSearch(),
    DateQuery(),
@@ -1,3 +1,10 @@
+"""
+GUI 自动化工具。
+采用批量坐标生成方案：一次截图 + 一次 LLM 调用批量生成所有动作坐标。
+N 个需要定位的动作 = 1 次 LLM 调用。
+提示词中内置 One-shot 示例纠正大模型输出格式。
+"""
+
 from __future__ import annotations

 import asyncio
@@ -6,7 +13,7 @@ import io
 import re
 import sys
 import time
-from typing import Any, Literal, Optional
+from typing import Any, Literal, Optional, Tuple

 import pyautogui
 from langchain.tools import BaseTool
@@ -17,17 +24,13 @@ from pydantic import BaseModel, Field
 from agent.core.llm import llm
 import agent.utils.env_utils as env_utils

-# GUI 模型实例缓存（None 表示使用主模型 llm）
+# ---------- GUI 模型实例缓存 ----------
 _gui_llm = None
 _gui_llm_initialized = False


 def _get_gui_llm():
-    """获取 GUI 工具实际使用的 LLM 实例。
-
-    优先使用 gui_config.json 中配置的 GUI 专用模型，
-    若未配置则回退到 agent 主模型 llm。
-    """
+    """获取 GUI 工具实际使用的 LLM 实例（优先 GUI 专用模型，否则回退主模型）。"""
    global _gui_llm, _gui_llm_initialized
    if not _gui_llm_initialized:
        _gui_llm_initialized = True
@@ -38,6 +41,8 @@ def _get_gui_llm():
            _gui_llm = None
    return _gui_llm if _gui_llm is not None else llm

+
+# ---------- 类型定义 ----------
 GUIAction = Literal[
    "click",
    "type",
@@ -54,47 +59,41 @@ GUITargetType = Literal["auto", "button", "input", "checkbox", "list_item", "lin


 class GUIActionItem(BaseModel):
-    """A single GUI action within an ordered sequence."""
-    action: GUIAction = Field(
-        ...,
-        description="GUI action to execute.",
-    )
+    """单个 GUI 动作。"""
+    action: GUIAction = Field(..., description="GUI 操作类型。")
    element_description: Optional[str] = Field(
        default=None,
-        description="Natural-language description of the target screen element for grounded actions (click, type, scroll, drag_and_drop, move_to). For type action, providing this will first locate and click the target input field before typing.",
+        description="目标元素的自然语言描述。对于 type 操作，提供此项会先定位并点击目标输入框再输入。",
    )
    ending_description: Optional[str] = Field(
        default=None,
-        description="Natural-language description of the drag end target.",
+        description="拖拽终点的自然语言描述（仅 drag_and_drop 使用）。",
    )
-    text: str = Field(default="", description="Text to type.")
-    overwrite: bool = Field(default=False, description="Select all existing text before typing.")
-    enter: bool = Field(default=False, description="Press Enter after typing.")
-    num_clicks: int = Field(default=1, description="Number of mouse clicks.")
-    button_type: Literal["left", "middle", "right"] = Field(default="left", description="Mouse button.")
-    hold_keys: list[str] = Field(default_factory=list, description="Modifier keys to hold while clicking or dragging.")
-    keys: list[str] = Field(default_factory=list, description="Hotkey keys or keys to press.")
-    press_keys: list[str] = Field(default_factory=list, description="Keys to press while holding hold_keys.")
-    clicks: int = Field(default=0, description="Scroll click amount. Negative means down.")
-    shift: bool = Field(default=False, description="Use horizontal scrolling when true.")
-    wait_seconds: float = Field(default=0.5, description="Seconds to wait.")
+    text: str = Field(default="", description="要输入的文本。")
+    overwrite: bool = Field(default=False, description="输入前全选已有文本。")
+    enter: bool = Field(default=False, description="输入后按回车。")
+    num_clicks: int = Field(default=1, description="鼠标点击次数。")
+    button_type: Literal["left", "middle", "right"] = Field(default="left", description="鼠标按键。")
+    hold_keys: list[str] = Field(default_factory=list, description="点击或拖拽时按住的功能键。")
+    keys: list[str] = Field(default_factory=list, description="快捷键按键 / 敲击按键。")
+    press_keys: list[str] = Field(default_factory=list, description="按住 hold_keys 时敲击的按键。")
+    clicks: int = Field(default=0, description="滚轮量，负数为向下。")
+    shift: bool = Field(default=False, description="水平滚动（True 时为横向）。")
+    wait_seconds: float = Field(default=0.5, description="等待秒数。")
    target_type: GUITargetType = Field(
        default="auto",
-        description=(
-            "Target UI element type for grounded actions. Use checkbox for agreement/privacy checkboxes, "
-            "input for editable text boxes, list_item for chat/contact/file rows, button for buttons, "
-            "link for clickable text links, or auto when unsure."
-        ),
+        description="目标 UI 元素类型：checkbox / input / list_item / button / link / auto。",
    )


 class GUIToolInput(BaseModel):
    actions: list[GUIActionItem] = Field(
        default_factory=list,
-        description="Ordered list of GUI actions to execute sequentially. Each item specifies one GUI operation (click, type, drag_and_drop, scroll, hotkey, hold_and_press, wait, screenshot, move_to). If empty, returns current screenshot without performing any action.",
+        description="有序 GUI 操作列表，按顺序依次执行。空列表则仅返回当前截图。",
    )


+# ---------- 图片工具函数 ----------
 def _get_gui_region() -> dict | None:
    """获取当前 GUI 操作显示器区域，失败返回 None（全屏）。"""
    try:
@@ -104,10 +103,12 @@ def _get_gui_region() -> dict | None:
        return None


-def _screenshot_png() -> tuple[bytes, tuple[int, int]]:
+def _screenshot_png() -> Tuple[bytes, Tuple[int, int]]:
    region = _get_gui_region()
    if region:
-        image = pyautogui.screenshot(region=(region["x"], region["y"], region["width"], region["height"]))
+        image = pyautogui.screenshot(
+            region=(region["x"], region["y"], region["width"], region["height"])
+        )
    else:
        image = pyautogui.screenshot()
    width, height = image.size
@@ -118,7 +119,10 @@ def _screenshot_png() -> tuple[bytes, tuple[int, int]]:

 def _resize_for_grounding(image_bytes: bytes) -> bytes:
    image = Image.open(io.BytesIO(image_bytes)).convert("RGB")
-    image = image.resize((env_utils.GROUNDING_WIDTH, env_utils.GROUNDING_HEIGHT), Image.Resampling.LANCZOS)
+    image = image.resize(
+        (env_utils.GROUNDING_WIDTH, env_utils.GROUNDING_HEIGHT),
+        Image.Resampling.LANCZOS,
+    )
    buf = io.BytesIO()
    image.save(buf, format="PNG")
    return buf.getvalue()
@@ -129,112 +133,50 @@ def _image_data_url(image_bytes: bytes) -> str:
    return f"data:image/png;base64,{encoded}"


-def _parse_coordinates(text: str) -> tuple[int, int]:
-    numbers = re.findall(r"-?\d+(?:\.\d+)?", text or "")
-    if len(numbers) < 2:
-        raise ValueError(f"grounding model did not return coordinates: {text!r}")
-    x = int(float(numbers[0]))
-    y = int(float(numbers[1]))
-    x = max(0, min(env_utils.GROUNDING_WIDTH - 1, x))
-    y = max(0, min(env_utils.GROUNDING_HEIGHT - 1, y))
-    return x, y
+# ---------- 坐标解析与转换 ----------
+def _parse_batch_coordinates(text: str) -> list[Tuple[int, int]]:
+    """解析批量坐标输出。
+
+    按行提取坐标对，每行取第一对有效的 (x, y)。
+    兼容格式：
+        - "(50, 960)"
+        - "(100,200)"
+        - "1, (50, 960)"  （带编号前缀）
+        - "1:(100,200)"   （带编号前缀）
+
+    所有坐标会被裁剪到 [0, GROUNDING-1] × [0, GROUNDING-1]。
+    """
+    W, H = env_utils.GROUNDING_WIDTH - 1, env_utils.GROUNDING_HEIGHT - 1
+    coordinates: list[Tuple[int, int]] = []
+
+    for line in text.strip().split("\n"):
+        line = line.strip()
+        if not line:
+            continue
+        matches = re.findall(r"\(\s*(-?\d+(?:\.\d+)?)\s*[,，]\s*(-?\d+(?:\.\d+)?)\s*\)", line)
+        if matches:
+            x = int(float(matches[0][0]))
+            y = int(float(matches[0][1]))
+            x = max(0, min(W, x))
+            y = max(0, min(H, y))
+            coordinates.append((x, y))
+
+    return coordinates


-def _scale_to_screen(coords: tuple[int, int], screen_size: tuple[int, int]) -> tuple[int, int]:
+def _scale_to_screen(coords: Tuple[int, int], screen_size: Tuple[int, int]) -> Tuple[int, int]:
    """将 grounding 坐标映射到绝对屏幕坐标（考虑 GUI 操作区域偏移）。"""
    screen_width, screen_height = screen_size
    x, y = coords
    local_x = round(x * screen_width / env_utils.GROUNDING_WIDTH)
    local_y = round(y * screen_height / env_utils.GROUNDING_HEIGHT)
-    # 加上目标显示器在屏幕空间的偏移
    region = _get_gui_region()
    offset_x = region["x"] if region else 0
    offset_y = region["y"] if region else 0
    return (offset_x + local_x, offset_y + local_y)


-def _target_type_instructions(target_type: str) -> str:
-    normalized = (target_type or "auto").strip().lower()
-    instructions = {
-        "checkbox": (
-            "Target type: checkbox. Return the checkbox/radio circle or square itself, "
-            "and provide its center point. Do not click nearby agreement/privacy/service link text."
-        ),
-        "input": (
-            "Target type: input. Return the editable text box area and provide a point "
-            "inside the editable area where typing will focus the field."
-        ),
-        "list_item": (
-            "Target type: list_item. Return the full row/list item including icon/avatar "
-            "and label, and provide the row center point."
-        ),
-        "button": (
-            "Target type: button. Return the full clickable button and provide its center point."
-        ),
-        "link": (
-            "Target type: link. Return only the clickable text/link area and provide the text center."
-        ),
-    }
-    return instructions.get(normalized, "Target type: auto. Infer the UI element type from the query.")
-
-
-def _infer_target_type(element_description: str, target_type: str) -> str:
-    normalized = (target_type or "auto").strip().lower()
-    if normalized != "auto":
-        return normalized
-
-    desc = element_description or ""
-    checkbox_terms = ["复选框", "勾选框", "勾选", "已阅读", "同意", "协议", "隐私", "服务协议", "用户协议", "checkbox"]
-    if any(term in desc for term in checkbox_terms):
-        return "checkbox"
-
-    input_terms = ["输入框", "文本框", "消息框", "编辑框", "输入栏", "input", "textbox"]
-    if any(term in desc for term in input_terms):
-        return "input"
-
-    list_terms = ["列表", "联系人", "聊天列表", "文件列表", "这一行", "整行", "list"]
-    if any(term in desc for term in list_terms):
-        return "list_item"
-
-    button_terms = ["按钮", "登录", "确定", "取消", "发送", "button"]
-    if any(term in desc for term in button_terms):
-        return "button"
-
-    return "auto"
-
-
-def _ground_element(element_description: str, target_type: str = "auto") -> tuple[int, int, str]:
-    if not element_description or not element_description.strip():
-        raise ValueError("element_description is required for this GUI action.")
-
-    target_type = _infer_target_type(element_description, target_type)
-    raw_png, screen_size = _screenshot_png()
-    grounding_png = _resize_for_grounding(raw_png)
-    prompt = (
-        "You are a visual grounding model for desktop GUI control.\n"
-        f"The screenshot has been resized to exactly {env_utils.GROUNDING_WIDTH}x{env_utils.GROUNDING_HEIGHT} pixels.\n"
-        f"Query: {element_description.strip()}\n"
-        f"{_target_type_instructions(target_type)}\n"
-        "Return only one target point as pixel coordinates in this exact format: x,y\n"
-        f"Example: {env_utils.GROUNDING_WIDTH // 2},{env_utils.GROUNDING_HEIGHT // 2}\n"
-        "Do not include any explanation."
-    )
-    response = _get_gui_llm().invoke(
-        [
-            HumanMessage(
-                content=[
-                    {"type": "image_url", "image_url": {"url": _image_data_url(grounding_png)}},
-                    {"type": "text", "text": prompt},
-                ]
-            )
-        ]
-    )
-    response_text = str(response.content)
-    gx, gy = _parse_coordinates(response_text)
-    x, y = _scale_to_screen((gx, gy), screen_size)
-    return x, y, f"grounded ({gx},{gy})"
-
-
+# ---------- 文本输入辅助 ----------
 def _type_text(text: str) -> None:
    if not text:
        return
@@ -242,7 +184,6 @@ def _type_text(text: str) -> None:
    if has_unicode:
        try:
            import pyperclip
-
            pyperclip.copy(text)
            modifier = "command" if sys.platform == "darwin" else "ctrl"
            pyautogui.hotkey(modifier, "v")
@@ -268,7 +209,16 @@ def _with_hold_keys_up(keys: list[str]) -> None:
        pyautogui.keyUp(key)


+# =====================================================================
+#  核心工具类
+# =====================================================================
 class GUITool(BaseTool):
+    """GUI 自动化工具。
+
+    坐标生成阶段采用批量方案：一次截图 + 全部 query 批量输入 LLM + 一次性解析所有坐标。
+    N 个需要定位的动作 = 1 次 LLM 调用。
+    """
+
    args_schema: type[BaseModel] = GUIToolInput
    name: str = "gui_tool"
    description: str = (
@@ -280,6 +230,18 @@ class GUITool(BaseTool):
        "会先视觉定位并点击目标输入框，然后再输入文本，无需单独调用 click。不提供 element_description 则直接在当前焦点输入。\n"
        "所有动作执行完毕后，返回当前屏幕的截图。\n"
        "\n"
+        "【网页滚动提示】浏览器中向下滚动网页推荐使用快捷键，而非 scroll 滚轮："
+        "{\"action\":\"hotkey\",\"keys\":[\"space\"]}（空格键即 Page Down）。"
+        "向上滚动可使用 {\"action\":\"hotkey\",\"keys\":[\"shift\",\"space\"]}。\n"
+        "\n"
+        "【下拉框处理策略】操作下拉选择框时分两步：\n"
+        "  第1步（type 方式）：先尝试 {\"action\":\"type\",\"element_description\":\"目标下拉框\",\"text\":\"选项名称\"} 直接输入，\n"
+        "    能自动匹配则任务完成；\n"
+        "  第2步（键选方式）：若 type 不生效，则改为先点击展开下拉框，再用方向键选择，最后回车确认：\n"
+        "    {\"action\":\"click\",\"element_description\":\"目标下拉框\"},\n"
+        "    {\"action\":\"hotkey\",\"keys\":[\"down\"]}（按需多次 up/down 定位），\n"
+        "    {\"action\":\"hotkey\",\"keys\":[\"enter\"]}。\n"
+        "\n"
        "示例：\n"
        "- 登录表单：{\"actions\": ["
        "{\"action\":\"type\",\"element_description\":\"用户名输入框\",\"text\":\"admin\"},"
@@ -288,6 +250,11 @@ class GUITool(BaseTool):
        "- 浏览器搜索：{\"actions\": ["
        "{\"action\":\"type\",\"element_description\":\"浏览器地址栏\",\"text\":\"github.com\",\"enter\":true}"
        "]}\n"
+        "- 网页翻页浏览：{\"actions\": ["
+        "{\"action\":\"hotkey\",\"keys\":[\"space\"]},"
+        "{\"action\":\"wait\",\"wait_seconds\":1.0},"
+        "{\"action\":\"hotkey\",\"keys\":[\"space\"]}"
+        "]}\n"
        "- 等待并确认：{\"actions\": ["
        "{\"action\":\"click\",\"element_description\":\"提交按钮\"},"
        "{\"action\":\"wait\",\"wait_seconds\":2.0}"
@@ -295,6 +262,143 @@ class GUITool(BaseTool):
    )
    response_format: Literal["content", "content_and_artifact"] = "content_and_artifact"

+    # ---------- 批量坐标生成核心 ----------
+    def _collect_grounding_queries(self, actions: list[GUIActionItem]) -> list[Tuple[int, str, str]]:
+        """遍历所有动作，收集需要视觉定位的 query。
+
+        返回 [(action_index, query_type, query_text), ...]
+        query_type: "element" (元素描述) 或 "ending" (拖拽终点)。
+        对于 type 操作且无 element_description 的，不收集（无需定位）。
+        """
+        queries: list[Tuple[int, str, str]] = []
+        for idx, act in enumerate(actions):
+            action = str(act.action).strip()
+            if action in ("screenshot", "wait", "hotkey", "hold_and_press"):
+                continue
+
+            if action in ("click", "scroll", "move_to"):
+                desc = (act.element_description or "").strip()
+                if desc:
+                    queries.append((idx, "element", desc))
+                continue
+
+            if action == "type":
+                desc = (act.element_description or "").strip()
+                if desc:
+                    queries.append((idx, "element", desc))
+                continue
+
+            if action == "drag_and_drop":
+                start_desc = (act.element_description or "").strip()
+                end_desc = (act.ending_description or "").strip()
+                if start_desc:
+                    queries.append((idx, "element", start_desc))
+                if end_desc:
+                    queries.append((idx, "ending", end_desc))
+                continue
+
+        return queries
+
+    def _build_batch_prompt(self, queries: list[Tuple[int, str, str]]) -> str:
+        """构建包含 One-shot 示例的批量定位 prompt。"""
+        W = env_utils.GROUNDING_WIDTH
+        H = env_utils.GROUNDING_HEIGHT
+
+        numbered = []
+        for i, (act_idx, qtype, qtext) in enumerate(queries, 1):
+            label = f"action{act_idx + 1}_{qtype}"
+            numbered.append(f"{i}. [{label}] {qtext}")
+        query_text = "\n".join(numbered)
+
+        prompt = (
+            "You are a visual grounding model for desktop GUI control.\n"
+            f"The screenshot has been resized to exactly {W}x{H} pixels.\n"
+            "Your task is to locate MULTIPLE target elements in the same image in a single response.\n"
+            "\n"
+            "=== OUTPUT FORMAT (STRICT) ===\n"
+            "You MUST output one line per query in the EXACT format below:\n"
+            "  (<x>, <y>)\n"
+            "Where:\n"
+            "  - x: horizontal pixel coordinate (integer, 0 = leftmost)\n"
+            "  - y: vertical pixel coordinate (integer, 0 = topmost)\n"
+            "\n"
+            "=== ONE-SHOT EXAMPLE ===\n"
+            "Scenario: A desktop screenshot ({W}x{H}) showing a browser with multiple UI elements.\n"
+            "\n"
+            "Queries:\n"
+            "1. [action1_element] the browser address bar at the top of the window\n"
+            "2. [action2_element] the search input field in the main content area\n"
+            "3. [action3_element] the search button next to the input field\n"
+            "4. [action3_ending] the first search result link after clicking search\n"
+            "\n"
+            "Expected output (four lines, one per query):\n"
+            f"({W // 2}, {H // 6})\n"
+            f"({W // 2}, {H // 3})\n"
+            f"({W // 2 + 200}, {H // 3})\n"
+            f"({W // 2}, {H // 2 + 50})\n"
+            "\n"
+            "=== NOW YOUR TURN ===\n"
+            "Locate the following elements in the provided screenshot:\n"
+            f"{query_text}\n"
+            "\n"
+            "=== RULES (FAILURE TO FOLLOW = INCORRECT) ===\n"
+            "1. Output exactly one line per query, in the SAME order (1, 2, 3, ...).\n"
+            "2. Each line format: `(<integer_x>, <integer_y>)` — nothing else on that line.\n"
+            "3. NO explanations, NO markdown, NO additional text before or after the coordinate lines.\n"
+            "4. Coordinates must be integers in range x=[0,{W-1}] y=[0,{H-1}].\n"
+            "5. If you genuinely cannot find an element, output your best reasonable guess for its coordinates.\n"
+            "6. Do NOT wrap the output in code fences (```), quotes, or JSON.\n"
+            "7. For a sequence of operations on the same page, generate coordinates strictly in action order. "
+            "Each coordinate pair corresponds to one step of the GUI operation sequence. "
+            "(对于一个页面中的操作需要按顺序进行单步多GUI操作)\n"
+        )
+        return prompt
+
+    def _batch_ground_elements(
+        self, queries: list[Tuple[int, str, str]], screen_size: Tuple[int, int]
+    ) -> dict[Tuple[int, str], Tuple[int, int]]:
+        """一次截图 + 一次 LLM 调用，批量生成所有目标元素的坐标。
+
+        参数:
+            queries: [(action_index, query_type, query_text), ...]
+            screen_size: (screen_width, screen_height)
+
+        返回:
+            {(action_index, query_type): (screen_x, screen_y), ...}
+        """
+        if not queries:
+            return {}
+
+        # 1. 截图
+        raw_png, _ = _screenshot_png()
+        grounding_png = _resize_for_grounding(raw_png)
+
+        # 2. 构建 prompt + 单次 LLM 调用
+        prompt = self._build_batch_prompt(queries)
+        response = _get_gui_llm().invoke(
+            [
+                HumanMessage(
+                    content=[
+                        {"type": "image_url", "image_url": {"url": _image_data_url(grounding_png)}},
+                        {"type": "text", "text": prompt},
+                    ]
+                )
+            ]
+        )
+        response_text = str(response.content).strip()
+
+        # 3. 解析坐标
+        coords = _parse_batch_coordinates(response_text)
+
+        # 4. 建立映射 (action_index, query_type) → (screen_x, screen_y)
+        result: dict[Tuple[int, str], Tuple[int, int]] = {}
+        for i, (act_idx, qtype, _) in enumerate(queries):
+            if i < len(coords):
+                result[(act_idx, qtype)] = _scale_to_screen(coords[i], screen_size)
+
+        return result
+
+    # ---------- 截图 artifact ----------
    def _artifact_screenshot(self) -> dict:
        try:
            image_bytes, _ = _screenshot_png()
@@ -303,24 +407,35 @@ class GUITool(BaseTool):
        except Exception:
            return {}

-    def _run(
-        self,
-        actions: list[GUIActionItem] | None = None,
-    ) -> tuple[str, dict]:
+    # ---------- 顶层入口 ----------
+    def _run(self, actions: list[GUIActionItem] | None = None) -> Tuple[str, dict]:
        try:
-            # ---------- 无动作时直接截图 ----------
            if not actions:
                return ("", self._artifact_screenshot())

-            # ---------- 顺序执行动作列表 ----------
-            result_messages: list[str] = []  # 收集每个动作的返回信息
+            # 1. 先截图一次，获取屏幕尺寸
+            _, screen_size = _screenshot_png()
+
+            # 2. 批量定位：一次 LLM 调用生成所有动作需要的坐标
+            queries = self._collect_grounding_queries(actions)
+            coords_map = self._batch_ground_elements(queries, screen_size)
+
+            # 3. 构建 (action_index, query_type) → (x, y) 查找表
+            action_coords: dict[int, dict[str, Tuple[int, int]]] = {}
+            for (act_idx, qtype), coord in coords_map.items():
+                action_coords.setdefault(act_idx, {})[qtype] = coord
+
+            # 4. 顺序执行动作列表
+            result_messages: list[str] = []
            for idx, act in enumerate(actions, start=1):
                if act.action == "screenshot":
                    result_messages.append(f"{idx}. Screenshot (skipped, final screenshot will be taken)")
                    continue

+                act_idx = idx - 1  # 0-indexed
+                cached = action_coords.get(act_idx, {})
                try:
-                    msg = self._dispatch(**act.model_dump())
+                    msg = self._dispatch(cached_coords=cached, **act.model_dump())
                    result_messages.append(f"{idx}. {msg}")
                except Exception as exc:
                    error_msg = f"{idx}. [GUI Error] {act.action}: {exc}"
@@ -332,18 +447,19 @@ class GUITool(BaseTool):
            final_text = "\n".join(result_messages)
            return (final_text, self._artifact_screenshot())
        except Exception as e:
-            return (f"[ERROR] gui_tool 工具执行出错: {str(e)}", {})
+            return (f"[ERROR] gui_tool 执行出错: {str(e)}", {})

-    def _dispatch(self, **kwargs: Any) -> str:
+    def _dispatch(self, cached_coords: dict[str, Tuple[int, int]], **kwargs: Any) -> str:
+        """分发到具体动作方法，传入已缓存的坐标（cached_coords = {"element": (x,y), "ending": (x,y)}）。"""
        action = str(kwargs["action"]).strip()
        if action == "click":
-            return self._click(**kwargs)
+            return self._click(cached_coords, **kwargs)
        if action == "type":
-            return self._type(**kwargs)
+            return self._type(cached_coords, **kwargs)
        if action == "drag_and_drop":
-            return self._drag_and_drop(**kwargs)
+            return self._drag_and_drop(cached_coords, **kwargs)
        if action == "scroll":
-            return self._scroll(**kwargs)
+            return self._scroll(cached_coords, **kwargs)
        if action == "hotkey":
            return self._hotkey(**kwargs)
        if action == "hold_and_press":
@@ -353,34 +469,35 @@ class GUITool(BaseTool):
            time.sleep(seconds)
            return f"Waited {seconds:.2f} seconds."
        if action == "move_to":
-            return self._move_to(**kwargs)
+            return self._move_to(cached_coords, **kwargs)
        raise ValueError(f"unsupported GUI action: {action}")

-    def _click(self, **kwargs: Any) -> str:
-        x, y, grounding = _ground_element(kwargs["element_description"] or "", kwargs.get("target_type", "auto"))
+    # ---------- 各动作实现 ----------
+    def _click(self, cached_coords: dict[str, Tuple[int, int]], **kwargs: Any) -> str:
+        desc = (kwargs.get("element_description") or "").strip()
+        if desc and "element" in cached_coords:
+            x, y = cached_coords["element"]
+        else:
+            x, y = pyautogui.position()
        hold_keys = kwargs["hold_keys"]
        _with_hold_keys_down(hold_keys)
        try:
            pyautogui.click(x, y, clicks=max(1, int(kwargs["num_clicks"])), button=kwargs["button_type"])
        finally:
            _with_hold_keys_up(hold_keys)
-        return f"Clicked target. {grounding}"
+        return f"Clicked target at ({x},{y})."

-    def _type(self, **kwargs: Any) -> str:
+    def _type(self, cached_coords: dict[str, Tuple[int, int]], **kwargs: Any) -> str:
        element_desc = (kwargs.get("element_description") or "").strip()
-        if element_desc:
-            # 提供了目标元素描述：先定位点击再输入
-            target_type = kwargs.get("target_type", "auto")
-            if target_type == "auto":
-                target_type = "input"  # type 动作定位输入框时强制为 input
-            x, y, grounding = _ground_element(element_desc, target_type)
+        if element_desc and "element" in cached_coords:
+            x, y = cached_coords["element"]
            pyautogui.click(x, y)
            if kwargs["overwrite"]:
                _select_all_and_clear()
            _type_text(kwargs["text"])
            if kwargs["enter"]:
                pyautogui.press("enter")
-            return f"Clicked target and typed text. {grounding}"
+            return f"Clicked target and typed text at ({x},{y})."
        # 不提供 element_description：直接在当前焦点输入
        if kwargs["overwrite"]:
            _select_all_and_clear()
@@ -389,9 +506,9 @@ class GUITool(BaseTool):
            pyautogui.press("enter")
        return "Typed text."

-    def _drag_and_drop(self, **kwargs: Any) -> str:
-        x1, y1, start_grounding = _ground_element(kwargs["element_description"] or "", kwargs.get("target_type", "auto"))
-        x2, y2, end_grounding = _ground_element(kwargs["ending_description"] or "", kwargs.get("target_type", "auto"))
+    def _drag_and_drop(self, cached_coords: dict[str, Tuple[int, int]], **kwargs: Any) -> str:
+        x1, y1 = cached_coords.get("element", pyautogui.position())
+        x2, y2 = cached_coords.get("ending", pyautogui.position())
        hold_keys = kwargs["hold_keys"]
        pyautogui.moveTo(x1, y1)
        _with_hold_keys_down(hold_keys)
@@ -400,15 +517,15 @@ class GUITool(BaseTool):
            pyautogui.mouseUp()
        finally:
            _with_hold_keys_up(hold_keys)
-        return f"Dragged target. Start: {start_grounding}. End: {end_grounding}"
+        return f"Dragged from ({x1},{y1}) to ({x2},{y2})."

-    def _move_to(self, **kwargs: Any) -> str:
-        x, y, grounding = _ground_element(kwargs["element_description"] or "", kwargs.get("target_type", "auto"))
+    def _move_to(self, cached_coords: dict[str, Tuple[int, int]], **kwargs: Any) -> str:
+        x, y = cached_coords.get("element", pyautogui.position())
        pyautogui.moveTo(x, y)
-        return f"Moved mouse to target. {grounding}"
+        return f"Moved mouse to ({x},{y})."

-    def _scroll(self, **kwargs: Any) -> str:
-        x, y, grounding = _ground_element(kwargs["element_description"] or "", kwargs.get("target_type", "auto"))
+    def _scroll(self, cached_coords: dict[str, Tuple[int, int]], **kwargs: Any) -> str:
+        x, y = cached_coords.get("element", pyautogui.position())
        pyautogui.moveTo(x, y)
        time.sleep(0.1)
        clicks = int(kwargs["clicks"])
@@ -416,7 +533,7 @@ class GUITool(BaseTool):
            pyautogui.hscroll(clicks)
        else:
            pyautogui.vscroll(clicks)
-        return f"Scrolled target by {clicks}. {grounding}"
+        return f"Scrolled by {clicks} at ({x},{y})."

    def _hotkey(self, **kwargs: Any) -> str:
        keys = kwargs["keys"]
@@ -437,5 +554,5 @@ class GUITool(BaseTool):
            _with_hold_keys_up(hold_keys)
        return f"Held {hold_keys} and pressed {press_keys}."

-    async def _arun(self, **kwargs: Any) -> tuple[str, dict]:
-        return await asyncio.to_thread(self._run, **kwargs)
+    async def _arun(self, actions: list[GUIActionItem] | None = None) -> Tuple[str, dict]:
+        return await asyncio.to_thread(self._run, actions)
@@ -211,7 +211,16 @@ class BrowserControl(BaseTool):
        # 无论操作是否成功，都尝试截屏（等待窗口切换）
        time.sleep(0.3)
        try:
-            pil_img = pyautogui.screenshot()
+            # 先尝试获取 GUI 操作区域，限定截屏范围
+            try:
+                from agent.core.config_manager import get_active_monitor_region
+                region = get_active_monitor_region()
+            except Exception:
+                region = None
+            if region:
+                pil_img = pyautogui.screenshot(region=(region["x"], region["y"], region["width"], region["height"]))
+            else:
+                pil_img = pyautogui.screenshot()
            pil_img = pil_img.resize((env_utils.GROUNDING_WIDTH, env_utils.GROUNDING_HEIGHT), Image.Resampling.LANCZOS)
            buf = io.BytesIO()
            pil_img.save(buf, format="PNG")