GUI-Plus: A model for user interface interaction

import os
import re
import json
import math
import time
import uuid
import pyautogui
import pyperclip
import dashscope
from PIL import Image

# ===================== Step 1: System Prompt =====================

system_prompt = """# Tools

You may call one or more functions to assist with the user query.

You are provided with function signatures within <tools></tools> XML tags:
<tools>
{"type": "function", "function": {"name": "computer_use", "description": "Use a mouse and keyboard to interact with a computer, and take screenshots.\\n* This is an interface to a desktop GUI. You do not have access to a terminal or applications menu. You must click on desktop icons to start applications.\\n* Some applications may take time to start or process actions, so you may need to wait and take successive screenshots to see the results of your actions. E.g. if you click on Firefox and a window doesn't open, try wait and taking another screenshot.\\n* The screen's resolution is 1000x1000.\\n* Make sure to click any buttons, links, icons, etc with the cursor tip in the center of the element. Don't click boxes on their edges unless asked.", "parameters": {"properties": {"action": {"description": "The action to perform. The available actions are:\\n* `key`: Performs key down presses on the arguments passed in order, then performs key releases in reverse order.\\n* `type`: Type a string of text on the keyboard.\\n* `mouse_move`: Move the cursor to a specified (x, y) pixel coordinate on the screen.\\n* `left_click`: Click the left mouse button at a specified (x, y) pixel coordinate on the screen.\\n* `left_click_drag`: Click and drag the cursor to a specified (x, y) pixel coordinate on the screen.\\n* `right_click`: Click the right mouse button at a specified (x, y) pixel coordinate on the screen.\\n* `middle_click`: Click the middle mouse button at a specified (x, y) pixel coordinate on the screen.\\n* `double_click`: Double-click the left mouse button at a specified (x, y) pixel coordinate on the screen.\\n* `triple_click`: Triple-click the left mouse button at a specified (x, y) pixel coordinate on the screen (simulated as double-click since it's the closest action).\\n* `scroll`: Performs a scroll of the mouse scroll wheel.\\n* `hscroll`: Performs a horizontal scroll (mapped to regular scroll).\\n* `wait`: Wait specified seconds for the change to happen.\\n* `terminate`: Terminate the current task and report its completion status.\\n* `answer`: Answer a question.\\n* `interact`: Resolve the blocking window by interacting with the user.", "enum": ["key", "type", "mouse_move", "left_click", "left_click_drag", "right_click", "middle_click", "double_click", "triple_click", "scroll", "hscroll", "wait", "terminate", "answer", "interact"], "type": "string"}, "keys": {"description": "Required only by `action=key`.", "type": "array"}, "text": {"description": "Required only by `action=type`, `action=answer` and `action=interact`.", "type": "string"}, "coordinate": {"description": "(x, y): The x (pixels from the left edge) and y (pixels from the top edge) coordinates to move the mouse to. Required only by `action=mouse_move` and `action=left_click_drag`.", "type": "array"}, "pixels": {"description": "The amount of scrolling to perform. Positive values scroll up, negative values scroll down. Required only by `action=scroll` and `action=hscroll`.", "type": "number"}, "time": {"description": "The seconds to wait. Required only by `action=wait`.", "type": "number"}, "status": {"description": "The status of the task. Required only by `action=terminate`.", "type": "string", "enum": ["success", "failure"]}}, "required": ["action"], "type": "object"}}}
</tools>

For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
<tool_call>
{"name": <function-name>, "arguments": <args-json-object>}
</tool_call>

# Response format

Response format for every step:
1) Action: a short imperative describing what to do in the UI.
2) A single <tool_call>...</tool_call> block containing only the JSON: {"name": <function-name>, "arguments": <args-json-object>}.

Rules:
- Output exactly in the order: Action, <tool_call>.
- Be brief: one for Action.
- Do not output anything else outside those two parts.
- If finishing, use action=terminate in the tool call."""


# ===================== Step 2: Construct multi-turn conversation messages =====================

def get_messages(image, instruction, history_output, system_prompt):
    history_n = 4
    current_step = len(history_output)

    history_start_idx = max(0, current_step - history_n)
    previous_actions = []
    for i in range(history_start_idx):
        if i < len(history_output):
            history_output_str = history_output[i]['output']
            if 'Action:' in history_output_str and '<tool_call>':
                history_output_str = history_output_str.split('Action:')[1].split('<tool_call>')[0].strip()
            previous_actions.append(f"Step {i + 1}: {history_output_str}")

    previous_actions_str = "\\n".join(previous_actions) if previous_actions else "None"

    instruction_prompt = f"""
      Please generate the next move according to the UI screenshot, instruction and previous actions.

      Instruction: {instruction}

      Previous actions:
      {previous_actions_str}"""

    messages = [{"role": "system", "content": [{"text": system_prompt}]}]

    history_len = min(history_n, len(history_output))
    if history_len > 0:
        for history_id, history_item in enumerate(history_output[-history_n:], 0):
            if history_id == 0:
                messages.append({
                    "role": "user",
                    "content": [
                        {"text": instruction_prompt},
                        {"image": "file://" + history_item['image']}
                    ]
                })
            else:
                messages.append({
                    "role": "user",
                    "content": [{"image": "file://" + history_item['image']}]
                })
            messages.append({
                "role": "assistant",
                "content": [{"text": history_item['output']}],
            })
        messages.append({
            "role": "user",
            "content": [{"image": "file://" + image}]
        })
    else:
        messages.append({
            "role": "user",
            "content": [
                {"text": instruction_prompt},
                {"image": "file://" + image}
            ]
        })
    return messages


# ===================== Step 3: Parse model output and map coordinates =====================

def extract_tool_calls(text):
    pattern = re.compile(r'<tool_call>(.*?)</tool_call>', re.DOTALL | re.IGNORECASE)
    blocks = pattern.findall(text)
    actions = []
    for blk in blocks:
        blk = blk.strip()
        try:
            actions.append(json.loads(blk))
        except json.JSONDecodeError as e:
            print(f'Parsing failed: {e} | Snippet: {blk[:80]}...')
    return actions

def smart_resize(height, width, factor=32, min_pixels=32*32*4, max_pixels=32*32*1280, max_long_side=8192):
    def round_by_factor(number, factor):
        return round(number / factor) * factor
    def ceil_by_factor(number, factor):
        return math.ceil(number / factor) * factor
    def floor_by_factor(number, factor):
        return math.floor(number / factor) * factor

    if height < 2 or width < 2:
        raise ValueError(f"height:{height} or width:{width} must be larger than factor:{factor}")
    elif max(height, width) / min(height, width) > 200:
        raise ValueError(f"absolute aspect ratio must be smaller than 200, got {height} / {width}")

    if max(height, width) > max_long_side:
        beta = max(height, width) / max_long_side
        height, width = int(height / beta), int(width / beta)

    h_bar = round_by_factor(height, factor)
    w_bar = round_by_factor(width, factor)

    if h_bar * w_bar > max_pixels:
        beta = math.sqrt((height * width) / max_pixels)
        h_bar = floor_by_factor(height / beta, factor)
        w_bar = floor_by_factor(width / beta, factor)
    elif h_bar * w_bar < min_pixels:
        beta = math.sqrt(min_pixels / (height * width))
        h_bar = ceil_by_factor(height * beta, factor)
        w_bar = ceil_by_factor(width * beta, factor)
    return h_bar, w_bar


# ===================== Step 4: GUI operation utility class =====================

class ComputerTools:
    def __init__(self):
        self.image_info = None

    def load_image_info(self, path):
        width, height = Image.open(path).size
        self.image_info = (width, height)

    def get_screenshot(self, image_path, retry_times=3):
        if os.path.exists(image_path):
            os.remove(image_path)
        for i in range(retry_times):
            screenshot = pyautogui.screenshot()
            screenshot.save(image_path)
            if os.path.exists(image_path):
                self.load_image_info(image_path)
                return True
            else:
                time.sleep(0.1)
        return False

    def reset(self):
        pyautogui.hotkey('win', 'd')

    def press_key(self, keys):
        if isinstance(keys, list):
            cleaned_keys = []
            for key in keys:
                if isinstance(key, str):
                    if key.startswith("keys=["): key = key[6:]
                    if key.endswith("]"): key = key[:-1]
                    if key.startswith("['") or key.startswith('["'): key = key[2:] if len(key) > 2 else key
                    if key.endswith("']") or key.endswith('"]'): key = key[:-2] if len(key) > 2 else key
                    key = key.strip()
                    key_map = {"arrowleft": "left", "arrowright": "right", "arrowup": "up", "arrowdown": "down"}
                    key = key_map.get(key, key)
                    cleaned_keys.append(key)
                else:
                    cleaned_keys.append(key)
            keys = cleaned_keys
        else:
            keys = [keys]
        if len(keys) > 1:
            pyautogui.hotkey(*keys)
        else:
            pyautogui.press(keys[0])

    def type(self, text):
        pyperclip.copy(text)
        pyautogui.keyDown('ctrl')
        pyautogui.keyDown('v')
        pyautogui.keyUp('v')
        pyautogui.keyUp('ctrl')

    def mouse_move(self, x, y):
        pyautogui.moveTo(x, y)
        time.sleep(0.1)
        pyautogui.moveTo(x, y)

    def left_click(self, x, y):
        pyautogui.moveTo(x, y)
        time.sleep(0.1)
        pyautogui.click()

    def left_click_drag(self, x, y):
        pyautogui.dragTo(x, y, duration=0.5)
        pyautogui.moveTo(x, y)

    def right_click(self, x, y):
        pyautogui.moveTo(x, y)
        time.sleep(0.1)
        pyautogui.rightClick()

    def middle_click(self, x, y):
        pyautogui.moveTo(x, y)
        time.sleep(0.1)
        pyautogui.middleClick()

    def double_click(self, x, y):
        pyautogui.moveTo(x, y)
        time.sleep(0.1)
        pyautogui.doubleClick()

    def triple_click(self, x, y):
        pyautogui.moveTo(x, y)
        time.sleep(0.1)
        pyautogui.tripleClick()

    def scroll(self, pixels):
        pyautogui.scroll(pixels)


# ===================== Step 5: Complete automation flow =====================

def run_gui_automation(instruction, max_step=30):
    dashscope.api_key = os.getenv("DASHSCOPE_API_KEY")
    dashscope.base_http_api_url = 'https://dashscope.aliyuncs.com/api/v1'
    model_name = 'gui-plus-2026-02-26'

    computer_tools = ComputerTools()
    computer_tools.reset()

    output_dir = os.path.join(os.path.expanduser("~"), "Desktop", "gui_automation")
    os.makedirs(output_dir, exist_ok=True)

    history = []
    stop_flag = False
    session_id = str(uuid.uuid4())
    print('session_id ', session_id)

    print(f"[Task] {instruction}")
    print("=" * 60)

    for step_id in range(max_step):
        if stop_flag:
            break

        print(f"\n[Step {step_id + 1}]")

        screen_shot = os.path.join(output_dir, f'screenshot_{step_id}.png')
        computer_tools.get_screenshot(screen_shot)

        messages = get_messages(screen_shot, instruction, history, system_prompt)

        retry_time = 3
        for _ in range(retry_time):
            response = dashscope.MultiModalConversation.call(
                model=model_name,
                messages=messages,
                vl_high_resolution_images=True,
                headers={"x-dashscope-gui-session-id": session_id},
                stream=False
            )
            print(response['request_id'])
            try:
                output_text = response.output.choices[0].message.content[0]['text']
                break
            except Exception as e:
                print(response)
                print(e)
        else:
            raise Exception('retry_time out')
        print(f"[Model output]\n{output_text}\n")

        action_list = extract_tool_calls(output_text)
        if not action_list:
            print("No valid operation extracted")
            break

        for action_id, action in enumerate(action_list):
            action_parameter = action['arguments']
            action_type = action_parameter['action']

            dummy_image = Image.open(screen_shot)
            resized_height, resized_width = smart_resize(
                dummy_image.height, dummy_image.width,
                factor=16, min_pixels=3136, max_pixels=1003520 * 200
            )

            for key in ['coordinate', 'coordinate1', 'coordinate2']:
                if key in action_parameter:
                    action_parameter[key][0] = int(action_parameter[key][0] / 1000 * resized_width)
                    action_parameter[key][1] = int(action_parameter[key][1] / 1000 * resized_height)

            if action_type in ['click', 'left_click']:
                computer_tools.left_click(action_parameter['coordinate'][0], action_parameter['coordinate'][1])
                print(f"✓ Left-click ({action_parameter['coordinate'][0]}, {action_parameter['coordinate'][1]})")
            elif action_type == 'mouse_move':
                computer_tools.mouse_move(action_parameter['coordinate'][0], action_parameter['coordinate'][1])
                print(f"✓ Move mouse")
            elif action_type == 'middle_click':
                computer_tools.middle_click(action_parameter['coordinate'][0], action_parameter['coordinate'][1])
                print(f"✓ Middle-click")
            elif action_type in ['right click', 'right_click']:
                computer_tools.right_click(action_parameter['coordinate'][0], action_parameter['coordinate'][1])
                print(f"✓ Right-click")
            elif action_type in ['key', 'hotkey']:
                computer_tools.press_key(action_parameter['keys'])
                print(f"✓ Key press {action_parameter['keys']}")
            elif action_type == 'type':
                computer_tools.type(action_parameter['text'])
                print(f"✓ Type text: {action_parameter['text']}")
            elif action_type == 'drag':
                computer_tools.left_click_drag(action_parameter['coordinate'][0], action_parameter['coordinate'][1])
                print(f"✓ Drag")
            elif action_type == 'scroll':
                if 'coordinate' in action_parameter:
                    computer_tools.mouse_move(action_parameter['coordinate'][0], action_parameter['coordinate'][1])
                computer_tools.scroll(action_parameter.get("pixels", 1))
                print(f"✓ Scroll {action_parameter.get('pixels', 1)} pixels")
            elif action_type in ['computer_double_click', 'double_click']:
                computer_tools.double_click(action_parameter['coordinate'][0], action_parameter['coordinate'][1])
                print(f"✓ Double-click")
            elif action_type == 'wait':
                time.sleep(action_parameter.get('time', 2))
                print(f"✓ Wait {action_parameter.get('time', 2)} seconds")
            elif action_type == 'answer':
                print(f"✓ Task complete: {action_parameter.get('text', '')}")
                stop_flag = True
                break
            elif action_type in ['stop', 'terminate', 'done']:
                print(f"✓ Task terminated: {action_parameter.get('status', 'success')}")
                stop_flag = True
                break
            else:
                print(f"Unknown action type: {action_type}")

        history.append({'output': output_text, 'image': screen_shot})
        time.sleep(2)

    print("\n" + "=" * 60)
    print(f"[Complete] Total executed {len(history)} steps")


if __name__ == '__main__':
    run_gui_automation(
        instruction='Open Chrome for me and search for Alibaba on Baidu',
        max_step=30
    )

1. 1. Construct the mobile System Prompt

   import json, os, subprocess
   import dashscope, time, math
   from PIL import Image, ImageDraw
   import shutil, requests
   from datetime import datetime
   
   mobile_system_prompt = '''# Tools
           You may call one or more functions to assist with the user query.
           
           You are provided with function signatures within <tools></tools> XML tags:
           <tools>
           {"type": "function", "function": {"name_for_human": "mobile_use", "name": "mobile_use", "description": "Use a touchscreen to interact with a mobile device, and take screenshots.
           * This is an interface to a mobile device with touchscreen. You can perform actions like clicking, typing, swiping, etc.
           * Some applications may take time to start or process actions, so you may need to wait and take successive screenshots to see the results of your actions.
           * The screen's resolution is 1000x1000.
           * Make sure to click any buttons, links, icons, etc with the cursor tip in the center of the element. Don't click boxes on their edges unless asked.", "parameters": {"properties": {"action": {"description": "The action to perform. The available actions are:
           * `key`: Perform a key event on the mobile device.
               - This supports adb's `keyevent` syntax.
               - Examples: "volume_up", "volume_down", "power", "camera", "clear".
           * `click`: Click the point on the screen with coordinate (x, y).
           * `long_press`: Press the point on the screen with coordinate (x, y) for specified seconds.
           * `swipe`: Swipe from the starting point with coordinate (x, y) to the end point with coordinates2 (x2, y2).
           * `type`: Input the specified text into the activated input box.
           * `system_button`: Press the system button.
           * `open`: Open an app on the device.
           * `wait`: Wait specified seconds for the change to happen.
           * `answer`: Terminate the current task and output the answer.
           * `interact`: Resolve the blocking window by interacting with the user.
           * `terminate`: Terminate the current task and report its completion status.", "enum": ["key", "click", "long_press", "swipe", "type", "system_button", "open", "wait", "answer", "interact", "terminate"], "type": "string"}, "coordinate": {"description": "(x, y): The x (pixels from the left edge) and y (pixels from the top edge) coordinates to move the mouse to. Required only by `action=click`, `action=long_press`, and `action=swipe`.", "type": "array"}, "coordinate2": {"description": "(x, y): The x (pixels from the left edge) and y (pixels from the top edge) coordinates to move the mouse to. Required only by `action=swipe`.", "type": "array"}, "text": {"description": "Required only by `action=key`, `action=type`, `action=open`, `action=answer`,and `action=interact`.", "type": "string"}, "time": {"description": "The seconds to wait. Required only by `action=long_press` and `action=wait`.", "type": "number"}, "button": {"description": "Back means returning to the previous interface, Home means returning to the desktop, Menu means opening the application background menu, and Enter means pressing the enter. Required only by `action=system_button`", "enum": ["Back", "Home", "Menu", "Enter"], "type": "string"}, "status": {"description": "The status of the task. Required only by `action=terminate`.", "type": "string", "enum": ["success", "failure"]}}, "required": ["action"], "type": "object"}, "args_format": "Format the arguments as a JSON object."}}
           </tools>
           
           For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
           <tool_call>
           {"name": <function-name>, "arguments": <args-json-object>}
           </tool_call>
           
           # Response format
           
           Response format for every step:
           1) Action: a short imperative describing what to do in the UI.
           2) A single <tool_call>...</tool_call> block containing only the JSON: {"name": <function-name>, "arguments": <args-json-object>}.
           
           Rules:
           - Output exactly in the order: Action, <tool_call>.
           - Be brief: one for Action.
           - Do not output anything else outside those two parts.
           - If finishing, use action=terminate in the tool call.'''

2. Constructing multi-turn conversation messages

   from datetime import datetime
   
   def get_messages(image, instruction, history_output, system_prompt):
       history_n = 4
       current_step = len(history_output)
   
       history_start_idx = max(0, current_step - history_n)
   
       previous_actions = []
       for i in range(history_start_idx):
           if i < len(history_output):
               history_output_str = history_output[i]['output']
               if 'Action:' in history_output_str and '<tool_call>':
                   history_output_str = history_output_str.split('Action:')[1].split('<tool_call>')[0].strip()
               previous_actions.append(f"Step {i + 1}: {history_output_str}")
   
       previous_actions_str = (
           "\n".join(previous_actions) if previous_actions else "None"
       )
       # Add background information
       today = datetime.today()
       weekday_names = ["Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday", "Sunday"]
       weekday = weekday_names[today.weekday()]
       formatted_date = today.strftime("%Y-%m-%d") + " " + weekday
       ground_info = f'''Today's date is: {formatted_date}.'''
   
   
       instruction_prompt = f"""
           Please generate the next move according to the UI screenshot, instruction and previous actions.
           
           Instruction: {ground_info}{instruction}
           
           Previous actions:
           {previous_actions_str}"""
   
       ## Call the model
       messages = [
           {
               "role": "system",
               "content": [
                   {"text": system_prompt}
               ],
           }
       ]
       history_len = min(history_n, len(history_output))
       if history_len > 0:
           for history_id, history_item in enumerate(history_output[-history_n:], 0):
               if history_id == 0:
                   messages.append({
                       "role": "user",
                       "content": [
                           {"text": instruction_prompt},
                           {"image": "file://" +history_item['image']}
                       ]
                   })
               else:
                   messages.append({
                       "role": "user",
                       "content": [
                           {"image": "file://" +history_item['image']}
                       ]
                   })
               messages.append({
                   "role": "assistant",
                   "content": [
                       {"text": history_item['output']},
                   ]
               })
           messages.append({
               "role": "user",
               "content": [
                   {"image": "file://" +image},
               ]
           })
       else:
           messages.append(
               {
                   "role": "user",
                   "content": [
                       {
                           "text": instruction_prompt
                       },
                       {
                           "image": "file://" +image,
                       },
                   ],
               }
           )
   
       return messages

3. Calculating the size of the scaled image

   The mobile and computer versions share the same smart_resize function. For more information, see Coordinate mapping function.

4. Perform operations in the GUI

   Use ADB commands to perform the actual phone operations.

   import subprocess
   import os
   import time
   from PIL import Image
   
   class AdbTools:
       def __init__(self, adb_path, device=None):
           self.adb_path = adb_path
           self.device = device
           self.__device_str__ = f" -s {device} " if device is not None else ' '
           self.image_info = None
   
       def adb_shell(self, command):
           command = self.adb_path + self.__device_str__ + command
           subprocess.run(command, capture_output=True, text=True, shell=True)
   
       ## Load phone size
       def load_image_info(self, path):
           width, height = Image.open(path).size
           self.image_info = (width, height)
   
       ## Get screenshot
       def get_screenshot(self, image_path, retry_times=3):
           command = self.adb_path + (f" -s {self.device}" if self.device is not None else '') + f" exec-out screencap -p > {image_path}"
   
           for i in range(retry_times):
               subprocess.run(command, capture_output=True, text=True, shell=True)
               if os.path.exists(image_path):
                   self.load_image_info(image_path)
                   return True
               else:
                   time.sleep(0.1)
           else:
               return False
   
       ## Click (x,y)
       ## coordinate_size: The dimensions of the input image. Default is None, which uses the current phone's dimensions. Pass as {'x': int, 'y': int}.
       def click(self, x, y, coordinate_size=None):
           command = self.adb_path + self.__device_str__ + f" shell input tap {x} {y}"
           subprocess.run(command, capture_output=True, text=True, shell=True)
   
       def long_press(self, x, y, time=800):
           command = self.adb_path + self.__device_str__ + f" shell input swipe {x} {y} {x} {y} {time}"
           subprocess.run(command, capture_output=True, text=True, shell=True)
   
       ## Swipe from (x1,y1) to (x2,y2)
       ## coordinate_size: The dimensions of the input image. Default is None, which uses the current phone's dimensions. Pass as {'x': int, 'y': int}.
       def slide(self, x1, y1, x2, y2, coordinate_size=None, slide_time=800):
           command = self.adb_path + self.__device_str__ + f" shell input swipe {x1} {y1} {x2} {y2} {slide_time}"
           subprocess.run(command, capture_output=True, text=True, shell=True)
   
       ## Back
       def back(self):
           command = self.adb_path + self.__device_str__ + f"  shell input keyevent 4"
           subprocess.run(command, capture_output=True, text=True, shell=True)
   
       # Press the Home key
       def home(self):
           command = self.adb_path + self.__device_str__ + f" shell am start -a android.intent.action.MAIN -c android.intent.category.HOME"
           subprocess.run(command, capture_output=True, text=True, shell=True)
   
       ## Type (supports both Chinese and English, other languages are not guaranteed). Note: ADB Keyboard must be installed on the phone first.
       def type(self, text):
           escaped_text = text.replace('"', '\\"').replace("'", "\\'")
           command_list = [
               f"shell ime enable com.android.adbkeyboard/.AdbIME ",
               f"shell ime set com.android.adbkeyboard/.AdbIME ",
               0.1,
               f'shell am broadcast -a ADB_INPUT_TEXT --es msg "{escaped_text}" ',
               0.1,
               f"shell ime disable com.android.adbkeyboard/.AdbIME"
           ]
   
           for command in command_list:
               if isinstance(command, float):
                   time.sleep(command)
               elif isinstance(command, str):
                   subprocess.run(self.adb_path + self.__device_str__ + command.strip(), capture_output=True, text=True, shell=True)
   
       def get_package_name(self, all_packages=False):
           try:
               if all_packages:
                   command = self.adb_path + self.__device_str__ + " shell pm list packages"
               else:
                   command = self.adb_path + self.__device_str__ + " shell pm list packages -3"
               res = subprocess.run(command, capture_output=True, text=True, shell=True)
               pkgs = []
               for line in res.stdout.splitlines():
                   s = line.strip()
                   if not s:
                       continue
                   # Remove the "package:" prefix
                   if s.startswith("package:"):
                       s = s[len("package:"):]
                   # If it contains "=", the package name is on the right
                   if "=" in s:
                       _, s = s.split("=", 1)
                   if s:
                       pkgs.append(s)
               return sorted(set(pkgs))
           except Exception as e:
               print(e)
               return []
   
       def open_app(self, package_name):
           command = self.adb_path + self.__device_str__ + f" shell monkey -p {package_name} -c android.intent.category.LAUNCHER 1"
           subprocess.run(command, capture_output=True, text=True, shell=True)                             

5. 5. Application package name mapping

   The package name is the unique identifier for an Android application (format: com.companyname.productname). For example, com.tencent.mm is for WeChat by Tencent (mm stands for Mobile Messenger).

   To support opening applications by name (action=open), you need to maintain a mapping from application names to package names.

   # Common application package name mapping (example, can be extended as needed)
   package_str_list = '''com.tencent.mm	WeChat	wechat			
   com.tencent.mobileqq	qq	Tencent QQ			
   com.sina.weibo	Weibo				
   com.taobao.taobao	Taobao				
   com.jingdong.app.mall	JD.com	JD Now			
   com.xunmeng.pinduoduo	Pinduoduo				
   com.xingin.xhs	Xiaohongshu				
   com.douban.frodo	Douban				
   com.zhihu.android	Zhihu				
   com.autonavi.minimap	AMAP	Gaode			
   com.baidu.BaiduMap	Baidu Maps				
   com.sankuai.meituan.takeoutnew	Meituan Waimai				
   com.sankuai.meituan	Meituan	Meituan Waimai			
   com.dianping.v1	Dazhong Dianping	Dianping			
   me.ele	Ele.me	Taobao Flash Purchase			
   com.yek.android.kfc.activitys	KFC				
   ctrip.android.view	Ctrip	Ctrip Travel			
   com.MobileTicket	China Railway 12306	12306			
   com.Qunar	Qunar Travel	Qunar.com	Qunar		
   com.sdu.didi.psnger	DiDi Chuxing	DiDi			
   tv.danmaku.bili	bilibili	B Site	Bilibili	B Site	bili
   com.ss.android.ugc.aweme	Douyin				
   com.smile.gifmaker	Kuaishou				
   com.tencent.qqlive	Tencent Video				
   com.qiyi.video	iQIYI				
   com.youku.phone	Youku	Youku Video			
   com.hunantv.imgo.activity	Mango TV	Mango			
   com.phoenix.read	Hongguo Short Video	Hongguo			
   com.netease.cloudmusic	NetEase Cloud Music	NetEase Cloud			
   com.tencent.qqmusic	QQ Music				
   com.luna.music	Qishui Music				
   com.ximalaya.ting.android	Ximalaya				
   com.dragon.read	Fanqie Free Novels	Fanqie Novels			
   com.kmxs.reader	Qimao Free Novels				
   com.ss.android.lark	Lark				
   com.tencent.androidqqmail	QQ Mailbox				
   com.larus.nova	Doubao	Doubao			
   com.gotokeep.keep	Keep				
   com.lingan.seeyou	Meiyou				
   com.tencent.news	Tencent News				
   com.ss.android.article.news	Toutiao				
   com.lianjia.beike	Beike Zhaofang				
   com.anjuke.android.app	Anjuke				
   com.hexin.plat.android	Tonghuashun				
   com.miHoYo.hkrpg	Honkai: Star Rail	Honkai			
   com.papegames.lysk.cn	Love and Deepspace				
   com.android.settings	settings	androidsystemsettings			
   com.android.soundrecorder	audiorecorder				
   com.rammigsoftware.bluecoins	bluecoins				
   com.flauschcode.broccoli	broccoli				
   com.booking	booking				
   com.android.chrome	Google Chrome	googlechrome	chrome		
   com.android.deskclock	Clock	Alarm Clock	clock		
   com.android.contacts	contacts				
   com.duolingo	duolingo	Duolingo			
   com.expedia.bookings	expedia				
   com.android.fileexplorer	files	filemanager			
   com.google.android.gm	gmail	googlemail			
   com.google.android.apps.nbu.files	googlefiles	filesbygoogle			
   com.google.android.calendar	googlecalendar				
   com.google.android.apps.dynamite	googlechat				
   com.google.android.deskclock	googleclock				
   com.google.android.contacts	googlecontacts				
   com.google.android.apps.docs.editors.docs	googledocs				
   com.google.android.apps.docs	googledrive				
   com.google.android.apps.fitness	googlefit				
   com.google.android.keep	googlekeep				
   com.google.android.apps.maps	googlemaps				
   com.google.android.apps.books	googleplaybooks				
   com.android.vending	googleplaystore				
   com.google.android.apps.docs.editors.slides	googleslides				
   com.google.android.apps.tasks	googletasks				
   net.cozic.joplin	joplin				
   com.mcdonalds.app	McDonald's	mcdonald			
   net.osmand	osmand				
   com.Project100Pi.themusicplayer	pimusicplayer				
   com.quora.android	quora				
   com.reddit.frontpage	reddit				
   code.name.monkey.retromusic	retromusic				
   com.scientificcalculatorplus.simplecalculator.basiccalculator.mathcalc	simplecalendarpro				
   com.simplemobiletools.smsmessenger	simplesmsmessenger				
   org.telegram.messenger	telegram				
   com.einnovation.temu	temu				
   com.zhiliaoapp.musically	tiktok				
   com.twitter.android	twitter	x			
   org.videolan.vlc	vlc				
   com.whatsapp	whatsapp				
   com.taobao.movie.android	Taopiaopiao				
   com.tongcheng.android	Tongcheng Travel	Tongcheng			
   com.sankuai.movie	Maoyan				
   com.wuba.zhuanzhuan	Zhuanzhuan				
   com.tencent.weread	WeChat Read				
   com.taobao.idlefish	Xianyu				
   com.wudaokou.hippo	Hema				
   com.eg.android.AlipayGphone	Alipay				
   com.jd.jrapp	JD Finance				
   com.achievo.vipshop	Vipshop				
   com.smzdm.client.android	SMZDM				
   cn.kuwo.player	Kuwo Music				
   com.taobao.trip	Fliggy	Fliggy Travel			
   com.jingdong.pdj	JD Daojia				
   com.tencent.map	Tencent Maps				
   com.shizhuang.duapp	Dewu				
   cn.damai	Damai	Damai.cn			
   com.ss.android.auto	Dongchedi				
   com.cubic.autohome	Autohome				
   com.wuba	58.com	58 Tongcheng			
   com.android.calendar	Calendar				
   com.alibaba.android.rimet	DingTalk				
   com.meituan.retail.v.android	Xiaoxiang Supermarket				
   com.aliyun.tongyi	Tongyi	Qwen	Tongyi Qwen		
   com.hupu.games	Hupu	Hupu Sports			
   com.quark.browser	Quark	Quark Browser			
   com.yuantiku.tutor	Yuanfudao				
   com.tencent.mtt	QQ Browser				
   com.umetrip.android.msky.app	Umetrip				
   com.UCMobile	UC Browser				
   com.ss.android.ugc.aweme.lite	Douyin Lite	Douyin			
   air.tv.douyu.android	Douyu				
   com.tencent.hunyuan.app.chat	Yuanbao				
   com.baidu.searchbox	Baidu				
   com.lemon.lv	Jianying				
   cn.soulapp.android	Soul				
   com.baidu.netdisk	Baidu Netdisk				
   com.tmri.app.main	Jiaoguan 12123	12123			
   com.kugou.android	Kugou	Kugou Music			
   com.ss.android.lark	Lark				
   com.tencent.android.qqdownloader	App宝				
   com.mt.mtxx.mtxx	Meitu	Meitu XiuXiu			
   com.tencent.karaoke	WeSing				
   com.intsig.camscanner	CamScanner				
   com.android.bankabc	Agricultural Bank of China	ABC			
   cmb.pb	China Merchants Bank	CMB			
   com.ganji.android.haoche_c	Guazi	Guazi			
   com.sf.activity	SF Express	SF Express	SF Express		
   com.ziroom.ziroomcustomer	Ziroom				
   com.yumc.phsuperapp	Pizza Hut				
   cn.dominos.pizza	Domino's Pizza	Domino's			
   cn.wps.moffice_eng	WPS Office	WPS			
   com.mfw.roadbook	Mafengwo				
   com.moonshot.kimichat	Kimi				
   com.tencent.wemeet.app	Tencent Meeting				
   com.deepseek.chat	DeepSeek				
   com.spdbccc.app	SPD Bank				
   cn.samsclub.app	Sam's Club	Sam's	Sam's Club	Sam's Club	
   com.tencent.qqsports	Tencent Sports				
   com.hanweb.android.zhejiang.activity	Zheli Ban				
   com.ss.android.article.video	Xigua Video				
   com.taou.maimai	Maimai	'''
   
   PACKAGES_NAME_DICT = {}
   NAME_PACKAGE_DICT = {}
   
   def normalize_package_name(name):
       name = name.lower().strip().replace(" ", "").replace("-", "")
       return name
   
   for package_str in package_str_list.split("\n"):
       package_name = package_str.strip().split("\t")
       PACKAGES_NAME_DICT[package_name[0]] = [normalize_package_name(i) for i in package_name[1:]]
       for name in package_name[1:]:
           name = normalize_package_name(name)
           if name not in NAME_PACKAGE_DICT:
               NAME_PACKAGE_DICT[name] = [package_name[0]]
           else:
               NAME_PACKAGE_DICT[name].append(package_name[0])

6. Fully automated flow

   import os
   import uuid
   import dashscope
   import time
   import shutil
   import json
   from PIL import Image
   
   if __name__ == '__main__':
   
       add_info = ''
       ## Specify an app that is in the app list and installed locally
       # instruction = 'Book a high-speed train ticket from Shanghai to Beijing for the day after tomorrow on Ctrip'
       # instruction = 'Comment on today's NBA game in Hupu'
       # instruction = 'Check tomorrow's flights for me on Umetrip'
       # instruction = 'Look at past exam papers in Yuanfudao'
   
       ## Specify an app that is in the app list but not installed locally
       # instruction = 'Book a ticket for Jay Chou's concert on Maoyan'
       # instruction = 'Buy a pomelo in Xiaoxiang Supermarket'
       # instruction = 'Watch a live stream on Huya Live'
       # instruction = 'Play a song by Vae Xu on QQ Music'
   
       ## Do not specify an app, but it is installed locally
       # instruction = 'Navigate to the village'
       # instruction = 'Order a coffee for delivery'
       # instruction = 'Play a song by Vae Xu'
       instruction = 'Help me book a train ticket'
   
       ## Do not specify an app, and it is not installed locally
       # instruction = 'Check today's Shanghai Composite Index in a stock trading app'
       # instruction = 'Help me send a message to my wife: I won't be home for dinner tomorrow night.'
   
       history = []
       session_id = str(uuid.uuid4())
       print('session_id ', session_id)
       max_step = 50
   
       model_name = 'gui-plus-2026-02-26'
       dashscope.api_key = os.getenv("DASHSCOPE_API_KEY", None)
       print("DashScope API Key: ", dashscope.api_key)
       dashscope.base_http_api_url = 'https://dashscope.aliyuncs.com/api/v1'
       dashscope.base_websocket_api_url = 'https://dashscope.aliyuncs.com/api-ws/v1/inference'
   
       ## Note: You need to enter your own adb path
       adb_tools = AdbTools(adb_path="xxx/adb")
       # package_name_list = adb_tools.get_package_name()
       # adb_tools.home()
       # time.sleep(1)
       task_dir = instruction
       anno_dir = f"{instruction}_anno"
   
       if os.path.exists(task_dir):
           shutil.rmtree(task_dir)
       os.mkdir(task_dir)
   
       if os.path.exists(anno_dir):
           shutil.rmtree(anno_dir)
       os.mkdir(anno_dir)
   
       ## {"image": image, "output": model output}
       history = []
       open_app_retry = False
       # max_step = 1
       for step_id in range(max_step):
           print(f'\nSTEP {step_id}:\n------------------------------------')
           screen_shot = os.path.join(task_dir, f'screen_shot_{step_id}.png')
           adb_tools.get_screenshot(screen_shot)
   
           width, height = Image.open(screen_shot).size
           messages = get_messages(screen_shot, instruction, history, model_name)
   
           retry_time = 3
           for _ in range(retry_time):
               response = dashscope.MultiModalConversation.call(
                                                   model=model_name,
                                                   messages=messages,
                                                   vl_high_resolution_images=True,
                                                   headers={"x-dashscope-gui-session-id": session_id},
                                                   enable_thinking=False,
                                                   stream=False)
               print(response['request_id'])
               try:
                   output_text = response.output.choices[0].message.content[0]['text']
                   break
               except Exception as e:
                   print(response)
                   print(e)
           else:
               raise Exception('retry_time out')
   
           thought = response.output.choices[0].message.reasoning_content
           if thought != '':
               output_text = f"<think>\n{thought}\n</think>{output_text}"
           action = json.loads(output_text.split('<tool_call>\n')[1].split('}}\n')[0] + '}}\n')
           conclusion = output_text.split('<tool_call>')[0].strip()
   
           action_parameter = action['arguments']
           dummy_image = Image.open(screen_shot)
           resized_height, resized_width = smart_resize(dummy_image.height,
                                                        dummy_image.width,
                                                        factor=16,
                                                        min_pixels=3136,
                                                        max_pixels=1003520*200,
                                                        )
           for key in ['coordinate', 'coordinate1', 'coordinate2']:
               if key in action_parameter:
                   action_parameter[key][0] = int(action_parameter[key][0]/1000 * resized_width)
                   action_parameter[key][1] = int(action_parameter[key][1]/1000 * resized_height)
   
           print(output_text)
           action_type = action_parameter['action']
           if action_type == 'click':
               adb_tools.click(action_parameter['coordinate'][0],
                               action_parameter['coordinate'][1])
           elif action_type == 'long_press':
               adb_tools.long_press(action_parameter['coordinate'][0],
                                    action_parameter['coordinate'][1])
           elif action_type == 'type':
               adb_tools.type(action_parameter['text'])
           elif action_type in ['scroll', 'swipe']:
               adb_tools.slide(action_parameter['coordinate'][0],
                               action_parameter['coordinate'][1],
                               action_parameter['coordinate2'][0],
                               action_parameter['coordinate2'][1])
           elif action_type == 'system_button':
               system = action_parameter['button']
               if system == 'Back':
                   adb_tools.back()
               elif system == 'Home':
                   adb_tools.home()
           elif action_type == 'wait':
               time.sleep(2)
           elif action_type == 'terminate':
               print(f'Action completed')
               break
           elif action_type == 'open':
               app_name = action_parameter['text']
               package_name = NAME_PACKAGE_DICT.get(app_name, [])
               package_name_list = adb_tools.get_package_name()
               output_app_name = ''
               if app_name != '':
                   output_app_name = app_name
               for sub_package_name in package_name:
                   if sub_package_name in package_name_list:
                       adb_tools.open_app(sub_package_name)
                       break
               else:
                   input(f"Please install the relevant APP {output_app_name}")
                   continue
   
           elif action_type == 'answer':
               print(f'Answer: {conclusion}\nAction completed')
               break
           elif action_type in ['call_user', 'calluser', 'interact']:
               text = action_parameter['text']
               input(f"Please complete the related action for {text}")
               print("Action completed, continuing execution")
               pass
           else:
               raise Exception(f"mobile-e2e action_type not supported {action}")
   
           history.append({'output': output_text,
                           'image': screen_shot})
           show_screenshot(screen_shot, action_parameter, f"{anno_dir}/screenshot_anno_{step_id}.png")
           time.sleep(2)

import os
import re
import json
import math
import time
import uuid
import shutil
import subprocess
import dashscope
from PIL import Image
from datetime import datetime


# ===================== Step 1: System Prompt =====================

mobile_system_prompt = '''# Tools

You may call one or more functions to assist with the user query.

You are provided with function signatures within <tools></tools> XML tags:
<tools>
{"type": "function", "function": {"name_for_human": "mobile_use", "name": "mobile_use", "description": "Use a touchscreen to interact with a mobile device, and take screenshots.\\n* This is an interface to a mobile device with touchscreen. You can perform actions like clicking, typing, swiping, etc.\\n* Some applications may take time to start or process actions, so you may need to wait and take successive screenshots to see the results of your actions.\\n* The screen's resolution is 1000x1000.\\n* Make sure to click any buttons, links, icons, etc with the cursor tip in the center of the element. Don't click boxes on their edges unless asked.", "parameters": {"properties": {"action": {"description": "The action to perform. The available actions are:\\n* `key`: Perform a key event on the mobile device.\\n    - This supports adb's `keyevent` syntax.\\n    - Examples: \\"volume_up\\", \\"volume_down\\", \\"power\\", \\"camera\\", \\"clear\\".\\n* `click`: Click the point on the screen with coordinate (x, y).\\n* `long_press`: Press the point on the screen with coordinate (x, y) for specified seconds.\\n* `swipe`: Swipe from the starting point with coordinate (x, y) to the end point with coordinates2 (x2, y2).\\n* `type`: Input the specified text into the activated input box.\\n* `system_button`: Press the system button.\\n* `open`: Open an app on the device.\\n* `wait`: Wait specified seconds for the change to happen.\\n* `answer`: Terminate the current task and output the answer.\\n* `interact`: Resolve the blocking window by interacting with the user.\\n* `terminate`: Terminate the current task and report its completion status.", "enum": ["key", "click", "long_press", "swipe", "type", "system_button", "open", "wait", "answer", "interact", "terminate"], "type": "string"}, "coordinate": {"description": "(x, y): The x (pixels from the left edge) and y (pixels from the top edge) coordinates to move the mouse to. Required only by `action=click`, `action=long_press`, and `action=swipe`.", "type": "array"}, "coordinate2": {"description": "(x, y): The x (pixels from the left edge) and y (pixels from the top edge) coordinates to move the mouse to. Required only by `action=swipe`.", "type": "array"}, "text": {"description": "Required only by `action=key`, `action=type`, `action=open`, `action=answer`,and `action=interact`.", "type": "string"}, "time": {"description": "The seconds to wait. Required only by `action=long_press` and `action=wait`.", "type": "number"}, "button": {"description": "Back means returning to the previous interface, Home means returning to the desktop, Menu means opening the application background menu, and Enter means pressing the enter. Required only by `action=system_button`", "enum": ["Back", "Home", "Menu", "Enter"], "type": "string"}, "status": {"description": "The status of the task. Required only by `action=terminate`.", "type": "string", "enum": ["success", "failure"]}}, "required": ["action"], "type": "object"}, "args_format": "Format the arguments as a JSON object."}}
</tools>

For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
<tool_call>
{"name": <function-name>, "arguments": <args-json-object>}
</tool_call>

# Response format

Response format for every step:
1) Action: a short imperative describing what to do in the UI.
2) A single <tool_call>...</tool_call> block containing only the JSON: {"name": <function-name>, "arguments": <args-json-object>}.

Rules:
- Output exactly in the order: Action, <tool_call>.
- Be brief: one for Action.
- Do not output anything else outside those two parts.
- If finishing, use action=terminate in the tool call.'''


# ===================== Step 2: Construct multi-turn conversation messages =====================

def get_messages(image, instruction, history_output, system_prompt):
    history_n = 4
    current_step = len(history_output)
    history_start_idx = max(0, current_step - history_n)

    previous_actions = []
    for i in range(history_start_idx):
        if i < len(history_output):
            history_output_str = history_output[i]['output']
            if 'Action:' in history_output_str and '<tool_call>':
                history_output_str = history_output_str.split('Action:')[1].split('<tool_call>')[0].strip()
            previous_actions.append(f"Step {i + 1}: {history_output_str}")

    previous_actions_str = "\\n".join(previous_actions) if previous_actions else "None"

    # Add background information
    today = datetime.today()
    weekday_names = ["Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday", "Sunday"]
    weekday = weekday_names[today.weekday()]
    formatted_date = today.strftime("%Y-%m-%d") + " " + weekday
    ground_info = f'''Today's date is: {formatted_date}.'''

    instruction_prompt = f"""
Please generate the next move according to the UI screenshot, instruction and previous actions.

Instruction: {ground_info}{instruction}

Previous actions:
{previous_actions_str}"""

    messages = [{"role": "system", "content": [{"text": system_prompt}]}]

    history_len = min(history_n, len(history_output))
    if history_len > 0:
        for history_id, history_item in enumerate(history_output[-history_n:], 0):
            if history_id == 0:
                messages.append({
                    "role": "user",
                    "content": [
                        {"text": instruction_prompt},
                        {"image": "file://" + history_item['image']}
                    ]
                })
            else:
                messages.append({
                    "role": "user",
                    "content": [{"image": "file://" + history_item['image']}]
                })
            messages.append({
                "role": "assistant",
                "content": [{"text": history_item['output']}],
            })
        messages.append({
            "role": "user",
            "content": [{"image": "file://" + image}]
        })
    else:
        messages.append({
            "role": "user",
            "content": [
                {"text": instruction_prompt},
                {"image": "file://" + image}
            ]
        })
    return messages


# ===================== Step 3: Parse model output and map coordinates =====================

def extract_tool_calls(text):
    pattern = re.compile(r'<tool_call>(.*?)</tool_call>', re.DOTALL | re.IGNORECASE)
    blocks = pattern.findall(text)
    actions = []
    for blk in blocks:
        blk = blk.strip()
        try:
            actions.append(json.loads(blk))
        except json.JSONDecodeError as e:
            print(f'Parsing failed: {e} | Snippet: {blk[:80]}...')
    return actions


def smart_resize(height, width, factor=32, min_pixels=32*32*4, max_pixels=32*32*1280, max_long_side=8192):
    def round_by_factor(number, factor):
        return round(number / factor) * factor
    def ceil_by_factor(number, factor):
        return math.ceil(number / factor) * factor
    def floor_by_factor(number, factor):
        return math.floor(number / factor) * factor

    if height < 2 or width < 2:
        raise ValueError(f"height:{height} or width:{width} must be larger than factor:{factor}")
    if max(height, width) > max_long_side:
        beta = max(height, width) / max_long_side
        height, width = int(height / beta), int(width / beta)

    h_bar = round_by_factor(height, factor)
    w_bar = round_by_factor(width, factor)

    if h_bar * w_bar > max_pixels:
        beta = math.sqrt((height * width) / max_pixels)
        h_bar = floor_by_factor(height / beta, factor)
        w_bar = floor_by_factor(width / beta, factor)
    elif h_bar * w_bar < min_pixels:
        beta = math.sqrt(min_pixels / (height * width))
        h_bar = ceil_by_factor(height * beta, factor)
        w_bar = ceil_by_factor(width * beta, factor)
    return h_bar, w_bar


# ===================== Step 4: ADB operation utility class =====================

class AdbTools:
    def __init__(self, adb_path, device=None):
        self.adb_path = adb_path
        self.device = device
        self.__device_str__ = f" -s {device} " if device is not None else ' '
        self.image_info = None

    def adb_shell(self, command):
        command = self.adb_path + self.__device_str__ + command
        subprocess.run(command, capture_output=True, text=True, shell=True)

    def load_image_info(self, path):
        width, height = Image.open(path).size
        self.image_info = (width, height)

    def get_screenshot(self, image_path, retry_times=3):
        command = self.adb_path + (f" -s {self.device}" if self.device is not None else '') + f" exec-out screencap -p > {image_path}"
        for i in range(retry_times):
            subprocess.run(command, capture_output=True, text=True, shell=True)
            if os.path.exists(image_path):
                self.load_image_info(image_path)
                return True
            else:
                time.sleep(0.1)
        return False

    def click(self, x, y, coordinate_size=None):
        command = self.adb_path + self.__device_str__ + f" shell input tap {x} {y}"
        subprocess.run(command, capture_output=True, text=True, shell=True)

    def long_press(self, x, y, time=800):
        command = self.adb_path + self.__device_str__ + f" shell input swipe {x} {y} {x} {y} {time}"
        subprocess.run(command, capture_output=True, text=True, shell=True)

    def slide(self, x1, y1, x2, y2, coordinate_size=None, slide_time=800):
        command = self.adb_path + self.__device_str__ + f" shell input swipe {x1} {y1} {x2} {y2} {slide_time}"
        subprocess.run(command, capture_output=True, text=True, shell=True)

    def back(self):
        command = self.adb_path + self.__device_str__ + f"  shell input keyevent 4"
        subprocess.run(command, capture_output=True, text=True, shell=True)

    def home(self):
        command = self.adb_path + self.__device_str__ + f" shell am start -a android.intent.action.MAIN -c android.intent.category.HOME"
        subprocess.run(command, capture_output=True, text=True, shell=True)

    def type(self, text):
        escaped_text = text.replace('"', '\\"').replace("'", "\\'")
        command_list = [
            f"shell ime enable com.android.adbkeyboard/.AdbIME ",
            f"shell ime set com.android.adbkeyboard/.AdbIME ",
            0.1,
            f'shell am broadcast -a ADB_INPUT_TEXT --es msg "{escaped_text}" ',
            0.1,
            f"shell ime disable com.android.adbkeyboard/.AdbIME"
        ]
        for command in command_list:
            if isinstance(command, float):
                time.sleep(command)
            elif isinstance(command, str):
                subprocess.run(self.adb_path + self.__device_str__ + command.strip(), capture_output=True, text=True, shell=True)

    def get_package_name(self, all_packages=False):
        try:
            if all_packages:
                command = self.adb_path + self.__device_str__ + " shell pm list packages"
            else:
                command = self.adb_path + self.__device_str__ + " shell pm list packages -3"
            res = subprocess.run(command, capture_output=True, text=True, shell=True)
            pkgs = []
            for line in res.stdout.splitlines():
                s = line.strip()
                if not s:
                    continue
                if s.startswith("package:"):
                    s = s[len("package:"):]
                if "=" in s:
                    _, s = s.split("=", 1)
                if s:
                    pkgs.append(s)
            return sorted(set(pkgs))
        except Exception as e:
            print(e)
            return []

    def open_app(self, package_name):
        command = self.adb_path + self.__device_str__ + f" shell monkey -p {package_name} -c android.intent.category.LAUNCHER 1"
        subprocess.run(command, capture_output=True, text=True, shell=True)


# ===================== Step 5: Application package name mapping =====================

package_str_list = '''com.tencent.mm\tWeChat\twechat\t\t\t
com.tencent.mobileqq\tqq\tTencent QQ\t\t\t
com.sina.weibo\tWeibo\t\t\t\t
com.taobao.taobao\tTaobao\t\t\t\t
com.jingdong.app.mall\tJD.com\tJD Now\t\t\t
com.xunmeng.pinduoduo\tPinduoduo\t\t\t\t
com.xingin.xhs\tXiaohongshu\t\t\t\t
com.douban.frodo\tDouban\t\t\t\t
com.zhihu.android\tZhihu\t\t\t\t
com.autonavi.minimap\tAMAP\tGaode\t\t\t
com.baidu.BaiduMap\tBaidu Maps\t\t\t\t
com.sankuai.meituan.takeoutnew\tMeituan Waimai\t\t\t\t
com.sankuai.meituan\tMeituan\tMeituan Waimai\t\t\t
com.dianping.v1\tDazhong Dianping\tDianping\t\t\t
me.ele\tEle.me\tTaobao Flash Purchase\t\t\t
com.yek.android.kfc.activitys\tKFC\t\t\t\t
ctrip.android.view\tCtrip\tCtrip Travel\t\t\t
com.MobileTicket\tChina Railway 12306\t12306\t\t\t
com.Qunar\tQunar Travel\tQunar.com\tQunar\t\t
com.sdu.didi.psnger\tDiDi Chuxing\tDiDi\t\t\t
tv.danmaku.bili\tbilibili\tB Site\tBilibili\tB Site\tbili
com.ss.android.ugc.aweme\tDouyin\t\t\t\t
com.smile.gifmaker\tKuaishou\t\t\t\t
com.tencent.qqlive\tTencent Video\t\t\t\t
com.qiyi.video\tiQIYI\t\t\t\t
com.youku.phone\tYouku\tYouku Video\t\t\t
com.hunantv.imgo.activity\tMango TV\tMango\t\t\t
com.phoenix.read\tHongguo Short Video\tHongguo\t\t\t
com.netease.cloudmusic\tNetEase Cloud Music\tNetEase Cloud\t\t\t
com.tencent.qqmusic\tQQ Music\t\t\t\t
com.luna.music\tQishui Music\t\t\t\t
com.ximalaya.ting.android\tXimalaya\t\t\t\t
com.dragon.read\tFanqie Free Novels\tFanqie Novels\t\t\t
com.kmxs.reader\tQimao Free Novels\t\t\t\t
com.ss.android.lark\tLark\t\t\t\t
com.tencent.androidqqmail\tQQ Mailbox\t\t\t\t
com.larus.nova\tDoubao\tDoubao\t\t\t
com.gotokeep.keep\tKeep\t\t\t\t
com.lingan.seeyou\tMeiyou\t\t\t\t
com.tencent.news\tTencent News\t\t\t\t
com.ss.android.article.news\tToutiao\t\t\t\t
com.lianjia.beike\tBeike Zhaofang\t\t\t\t
com.anjuke.android.app\tAnjuke\t\t\t\t
com.hexin.plat.android\tTonghuashun\t\t\t\t
com.miHoYo.hkrpg\tHonkai: Star Rail\tHonkai\t\t\t
com.papegames.lysk.cn\tLove and Deepspace\t\t\t\t'''

PACKAGES_NAME_DICT = {}
NAME_PACKAGE_DICT = {}

def normalize_package_name(name):
    name = name.lower().strip().replace(" ", "").replace("-", "")
    return name

for package_str in package_str_list.split("\\n"):
    package_name = package_str.strip().split("\\t")
    if len(package_name) < 2:
        continue
    package, *app_names = package_name
    PACKAGES_NAME_DICT[package] = [normalize_package_name(n) for n in app_names if n.strip()]
    for app_name in app_names:
        if not app_name.strip():
            continue
        normalized_name = normalize_package_name(app_name)
        if normalized_name not in NAME_PACKAGE_DICT:
            NAME_PACKAGE_DICT[normalized_name] = []
        NAME_PACKAGE_DICT[normalized_name].append(package)


# ===================== Step 6: Complete automation flow =====================

if __name__ == '__main__':
    instruction = 'Help me book a train ticket'  # Can be changed to other tasks
    history = []
    session_id = str(uuid.uuid4())
    print('session_id ', session_id)
    max_step = 50

    model_name = 'gui-plus-2026-02-26'
    dashscope.api_key = os.getenv("DASHSCOPE_API_KEY")
    dashscope.base_http_api_url = 'https://dashscope.aliyuncs.com/api/v1'

    # Note: You need to enter your own adb path
    adb_tools = AdbTools(adb_path="/path/to/adb")  # Change to your actual path
    # adb_tools.home()
    # time.sleep(1)

    task_dir = instruction
    if os.path.exists(task_dir):
        shutil.rmtree(task_dir)
    os.mkdir(task_dir)

    print(f"[Task] {instruction}")
    print("=" * 60)

    for step_id in range(max_step):
        print(f'\\n[Step {step_id + 1}]')

        screen_shot = os.path.join(task_dir, f'screen_shot_{step_id}.png')
        adb_tools.get_screenshot(screen_shot)

        messages = get_messages(screen_shot, instruction, history, mobile_system_prompt)

        retry_time = 3
        for _ in range(retry_time):
            response = dashscope.MultiModalConversation.call(
                model=model_name,
                messages=messages,
                vl_high_resolution_images=True,
                headers={"x-dashscope-gui-session-id": session_id},
                stream=False
            )
            print(response['request_id'])
            try:
                output_text = response.output.choices[0].message.content[0]['text']
                break
            except Exception as e:
                print(response)
                print(e)
        else:
            raise Exception('retry_time out')
        print(f"[Model output]\\n{output_text}\\n")

        actions = extract_tool_calls(output_text)
        if not actions:
            print("No valid operation extracted")
            break

        action = actions[0]
        action_parameter = action['arguments']

        dummy_image = Image.open(screen_shot)
        resized_height, resized_width = smart_resize(
            dummy_image.height, dummy_image.width,
            factor=16, min_pixels=3136, max_pixels=1003520*200
        )

        for key in ['coordinate', 'coordinate1', 'coordinate2']:
            if key in action_parameter:
                action_parameter[key][0] = int(action_parameter[key][0] / 1000 * resized_width)
                action_parameter[key][1] = int(action_parameter[key][1] / 1000 * resized_height)

        action_type = action_parameter['action']
        if action_type == 'click':
            adb_tools.click(action_parameter['coordinate'][0], action_parameter['coordinate'][1])
            print(f"✓ Click ({action_parameter['coordinate'][0]}, {action_parameter['coordinate'][1]})")
        elif action_type == 'long_press':
            adb_tools.long_press(action_parameter['coordinate'][0], action_parameter['coordinate'][1])
            print(f"✓ Long press")
        elif action_type == 'type':
            adb_tools.type(action_parameter['text'])
            print(f"✓ Type text: {action_parameter['text']}")
        elif action_type in ['scroll', 'swipe']:
            adb_tools.slide(action_parameter['coordinate'][0], action_parameter['coordinate'][1],
                           action_parameter['coordinate2'][0], action_parameter['coordinate2'][1])
            print(f"✓ Swipe")
        elif action_type == 'system_button':
            system = action_parameter['button']
            if system == 'Back':
                adb_tools.back()
                print(f"✓ Back")
            elif system == 'Home':
                adb_tools.home()
                print(f"✓ Home")
        elif action_type == 'wait':
            time.sleep(2)
            print(f"✓ Wait")
        elif action_type == 'terminate':
            print(f'✓ Task complete')
            break
        elif action_type == 'open':
            app_name = action_parameter['text']
            normalized_name = normalize_package_name(app_name)
            package_name = NAME_PACKAGE_DICT.get(normalized_name, [])
            package_name_list = adb_tools.get_package_name()
            for sub_package_name in package_name:
                if sub_package_name in package_name_list:
                    adb_tools.open_app(sub_package_name)
                    print(f"✓ Open app: {app_name}")
                    break
            else:
                print(f"⚠ App not found: {app_name}. Please install it manually.")
                continue
        elif action_type == 'answer':
            print(f'✓ Task complete: {action_parameter.get("text", "")}')
            break
        elif action_type in ['call_user', 'calluser', 'interact']:
            print(f"⚠ User interaction required: {action_parameter.get('text', '')}")
            input("Press Enter to continue after completing the action...")
        else:
            print(f"✗ Unknown action type: {action_type}")

        history.append({'output': output_text, 'image': screen_shot})
        time.sleep(2)

    print("\\n" + "=" * 60)
    print(f"[Complete] Total executed {len(history)} steps")

1. 1. Construct the System Prompt

   import json, os, base64
   import dashscope, time, math
   import uuid
   from PIL import Image, ImageDraw, ImageTk
   import asyncio
   from playwright.async_api import async_playwright, Page, Browser, BrowserContext
   from datetime import datetime
   
   # Define the action display function
   def show_screenshot(image_path, action_parameter, save_path="screenshot_anno.png"):
       image = Image.open(image_path)
       draw = ImageDraw.Draw(image)
       if 'coordinate' in action_parameter:
           radius = 15
           center_x, center_y = action_parameter['coordinate'][0], action_parameter['coordinate'][1]
           draw.ellipse((center_x - radius, center_y - radius, center_x + radius, center_y + radius), 
                        fill="red", outline="red")
       elif 'coordinate1' in action_parameter and 'coordinate2' in action_parameter:
           x1, y1 = action_parameter['coordinate1'][0], action_parameter['coordinate1'][1]
           x2, y2 = action_parameter['coordinate2'][0], action_parameter['coordinate2'][1]
           arrow_size = 10
           color = 'red'
           draw.line((x1, y1, x2, y2), fill=color, width=2)
           angle = math.atan2(y2 - y1, x2 - x1)
           arrow_x1 = x2 - arrow_size * math.cos(angle - math.pi / 6)
           arrow_y1 = y2 - arrow_size * math.sin(angle - math.pi / 6)
           arrow_x2 = x2 - arrow_size * math.cos(angle + math.pi / 6)
           arrow_y2 = y2 - arrow_size * math.sin(angle + math.pi / 6)
           draw.polygon([(x2, y2), (arrow_x1, arrow_y1), (arrow_x2, arrow_y2)], fill=color)
       else:
           return None
       image.save(save_path)
       return save_path
   
   # Prompt assembly function
   def get_messages(image, env_state, instruction, history_output):
       history_n = 1
       current_step = len(history_output)
       history_start_idx = max(0, current_step - history_n)
       SoM_format_ele_text = env_state["SoM"]["format_ele_text"]
   
       system_prompt = '''# Tools
   
   You may call one or more functions to assist with the user query.
   
   You are provided with function signatures within <tools></tools> XML tags:
   <tools>
   {"type": "function", "function": {"name_for_human": "browser_use", "name": "browser_use", "description": "Use a browser to interact with web pages and take labeled screenshots.\\n* This is an interface to a web browser. You can click elements, type into inputs, scroll, wait for loading, go back, etc.\\n* Each Observation screenshot contains Numerical Labels placed at the TOP LEFT of each Web Element. Use these labels to target elements.\\n* Some pages may take time to load; you may need to wait and take successive screenshots.\\n* Avoid clicking near element edges; target the center of the element.\\n* Execute exactly ONE interaction action per step; do not chain multiple interactions in one call.", "parameters": {"properties": {"action": {"description": "The action to perform. The available actions are:\\n* `click`: Click a web element by numerical label.\\n* `type`: Clear existing content in a textbox/input and type content. The system will automatically press ENTER after typing.\\n* `scroll`: Scroll within WINDOW or within a specific scrollable element/area (by label).\\n* `select`: Selects a specific option from a menu or dropdown. Use the option text provided in the textual information.\\n* `wait`: Wait for page processes to finish (default 5 seconds unless specified).\\n* `go_back`: Go back to the previous page.\\n* `wikipedia`: Directly jump to the Wikipedia homepage to search for information.\\n* `answer`: Terminate the current task and output the final answer.", "enum": ["click", "type", "scroll", "select", "wait", "go_back", "wikipedia", "answer"], "type": "string"}, "label": {"description": "Numerical label of the target web element. Required only by `action=click`, `action=type`, `action=scroll`, and `action=select` when scrolling within a specific area. Use string value `WINDOW` to scroll the whole page.", "type": ["integer", "string"]}, "direction": {"description": "Scroll direction. Required only by `action=scroll`.", "enum": ["up", "down"], "type": "string"}, "text": {"description": "Required only by `action=type` and `action=answer`.", "type": "string"}, "option": {"description": "The option to select. Required only by `action=select`", "type": "string"}, "time": {"description": "The seconds to wait. Required only by `action=wait` when overriding the default.", "type": "integer"}}, "required": ["action"], "type": "object"}, "args_format": "Format the arguments as a JSON object."}}
   </tools>
   
   For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
   <tool_call>
   {"name": <function-name>, "arguments": <args-json-object>}
   </tool_call>
   
   # Response format
   
   Response format for every step:
   1) Action: a short imperative describing what to do in the UI.
   2) A single <tool_call>...</tool_call> block containing only the JSON: {"name": <function-name>, "arguments": <args-json-object>}.
   
   Rules:
   - Output exactly in the order: Action, <tool_call>.
   - Be brief: one line for Action.
   - Do not output anything else outside those two parts.
   - Execute ONLY ONE interaction per iteration (one tool call).
   - If finishing, use action=answer in the tool call.'''
   
       today = datetime.today()
       weekday_names = ["Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday", "Sunday"]
       weekday = weekday_names[today.weekday()]
       formatted_date = today.strftime("%Y-%m-%d") + " " + weekday
       day_info = f'''Today's date is: {formatted_date}.'''
   
       previous_actions = []
       for i in range(history_start_idx):
           if i < len(history_output):
               history_output_str = history_output[i]['output']
               if 'Action:' in history_output_str and '<tool_call>':
                   history_output_str = history_output_str.split('Action:')[1].split('<tool_call>')[0].strip()
               previous_actions.append(f"Step {i + 1}: {history_output_str}")
   
       previous_actions_str = (
           "\\n".join(previous_actions) if previous_actions else "None"
       )
   
       instruction_prompt = f'''\nPlease generate the next move according to the UI screenshot, instruction and previous actions.
   
   Instruction: {day_info}{instruction}
   
   Previous actions:
   {previous_actions_str}'''
   
       messages = [
           {
               "role": "system",
               "content": [
                   {"text": system_prompt}
               ],
           }
       ]
       history_len = min(history_n, len(history_output))
       if history_len > 0:
           for history_id, history_item in enumerate(history_output[-history_n:], 0):
               if history_id == 0:
                   messages.append({
                       "role": "user",
                       "content": [
                           {"text": instruction_prompt},
                           {"text": "Current screenshot:"},
                           {"image": "file://" + history_item['image']},
                       ]
                   })
               else:
                   messages.append({
                       "role": "user",
                       "content": [
                           {"text": "Current screenshot:"},
                           {"image": "file://" + history_item['image']}
                       ]
                   })
               messages.append({
                   "role": "assistant",
                   "content": [
                       {"text": history_item['output']},
                   ]
               })
           messages.append({
               "role": "user",
               "content": [
                   {"image": "file://" +image},
                   {"text": SoM_format_ele_text}
               ]
           })
       else:
           messages.append(
               {
                   "role": "user",
                   "content": [
                       {
                           "text": instruction_prompt
                       },
                       {
                           "image": "file://" +image,
                       },
                       {"text": SoM_format_ele_text}
                   ],
               }
           )
   
       return messages
   
   
   def extract_tool_calls(text: str):
       import re, ast
       """Extract all JSON strings between <tool_call> ... </tool_call> from the text"""
       pattern = re.compile(r'<tool_call>(.*?)</tool_call>', re.DOTALL | re.IGNORECASE)
       blocks = pattern.findall(text)
       
       actions = []
       for blk in blocks:
           blk = blk.strip()
           try:
               actions.append(ast.literal_eval(blk))
           except json.JSONDecodeError as e:
               print(f'⚠️ Parsing failed: {e} | Snippet: {blk[:80]}...')
       return actions

2. 2. Browser control

   PLAYWRIGHT_KEY_MAP = {
       "backspace": "Backspace", "tab": "Tab", "return": "Enter", "enter": "Enter",
       "shift": "Shift", "control": "ControlOrMeta", "alt": "Alt", "escape": "Escape",
       "space": "Space", "pageup": "PageUp", "pagedown": "PageDown", "end": "End",
       "home": "Home", "left": "ArrowLeft", "up": "ArrowUp", "right": "ArrowRight",
       "down": "ArrowDown", "insert": "Insert", "delete": "Delete", "semicolon": ";",
       "equals": "=", "multiply": "Multiply", "add": "Add", "separator": "Separator",
       "subtract": "Subtract", "decimal": "Decimal", "divide": "Divide",
       "f1": "F1", "f2": "F2", "f3": "F3", "f4": "F4", "f5": "F5", "f6": "F6",
       "f7": "F7", "f8": "F8", "f9": "F9", "f10": "F10", "f11": "F11", "f12": "F12",
       "command": "Meta",
   }
   
   class PlaywrightComputer:
       """Async Playwright wrapper for web agent"""
   
       def __init__(self, initial_url="https://baidu.com/", task_dir='data', highlight_mouse=False):
           self._initial_url = initial_url
           self._search_engine_url = "https://baike.baidu.com/"
           self.task_dir = task_dir
           self._highlight_mouse = highlight_mouse
           
           self._playwright = None
           self._browser: Browser | None = None
           self._context: BrowserContext | None = None
           self._page: Page | None = None
           self._storage_state_path = "storage_state.json"
   
       async def _handle_new_page(self, new_page: Page):
           """Only keep one tab: redirect new tab url into current page"""
           new_url = new_page.url
           await new_page.close()
           try:
               await self._page.goto(new_url)
           except Exception as e:
               if "interrupted by another navigation" in str(e):
                   pass
               else:
                   raise
   
       async def reset(self):
           print("Creating session...")
           self._playwright = await async_playwright().start()
           self._browser = await self._playwright.chromium.launch(
               args=["--start-maximized", "--disable-blink-features=AutomationControlled"],
               headless=False,
           )
   
           storage_state = None
           if os.path.exists(self._storage_state_path):
               print("Loading storage state")
               storage_state = self._storage_state_path
   
           if sys.platform == "darwin":
               user_agent = (
                   "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) "
                   "AppleWebKit/537.36 (KHTML, like Gecko) Chrome/143.0.0.0 "
                   "Chrome/143.0.0.0 Safari/537.36 Edg/143.0.0.0 "
               )
               import tkinter as tk
               root = tk.Tk()
               root.withdraw()
               width = root.winfo_screenwidth()
               height = root.winfo_screenheight()
               print(f"Screen size: {width}x{height}", flush=True)
               root.destroy()
               self._context = await self._browser.new_context(
                   no_viewport=True,
                   user_agent=user_agent,
                   locale="en-US",
                   timezone_id="America/New_York",
                   storage_state=storage_state,
               )
           else:
               user_agent = (
                   "Mozilla/5.0 (Windows NT 10.0; Win64; x64) "
                   "AppleWebKit/537.36 (KHTML, like Gecko) "
                   "Chrome/122.0.0.0 Safari/537.36"
               )
               self._context = await self._browser.new_context(
                   no_viewport=True,
                   user_agent=user_agent,
                   locale="en-US",
                   timezone_id="America/New_York",
                   storage_state=storage_state,
               )
   
           self._page = await self._context.new_page()
           await self._page.goto(self._initial_url, timeout=60000, wait_until="domcontentloaded")
           self._context.on("page", lambda p: asyncio.create_task(self._handle_new_page(p)))
           print("Started local playwright (async).")
           return self
   
       async def close(self):
           if self._context:
               try:
                   await self._context.storage_state(path=self._storage_state_path)
               except Exception as e:
                   print(f"Failed to save storage state: {e}")
   
           if self._context:
               await self._context.close()
           try:
               if self._browser:
                   await self._browser.close()
           except Exception as e:
               if "Browser.close: Connection closed while reading from the driver" in str(e):
                   pass
               else:
                   raise
           if self._playwright:
               await self._playwright.stop()
   
       async def click_at(self, x: int, y: int):
           await self._page.mouse.click(x, y)
           await self._page.mouse.move(x, y)
           await self._page.mouse.down()
           await self._page.wait_for_timeout(100)
           await self._page.mouse.up()
           await self._page.wait_for_load_state()
   
       async def type_text_at(self, x: int, y: int, text: str, press_enter=True, clear_before_typing=True):
           await self.click_at(x, y)
           await asyncio.sleep(0.1)
           await self._page.wait_for_load_state()
   
           if clear_before_typing:
               if sys.platform == "darwin":
                   await self.key_combination(["Command", "A"])
               else:
                   await self.key_combination(["Control", "A"])
               await asyncio.sleep(0.1)
               await self.key_combination(["Delete"])
               await self._page.wait_for_load_state()
   
           await self.click_at(x, y)
           await asyncio.sleep(0.1)
           await self._page.keyboard.type(text)
           await self._page.wait_for_load_state()
   
           if press_enter:
               await self.key_combination(["Enter"])
           await self._page.wait_for_load_state()
   
       async def scroll_at(self, x: int, y: int, direction: str, magnitude: int = 400):
           await self._page.mouse.move(x, y)
           await asyncio.sleep(0.1)
   
           dx = dy = 0
           if direction == "up":
               dy = -magnitude
           elif direction == "down":
               dy = magnitude
           elif direction == "left":
               dx = -magnitude
           elif direction == "right":
               dx = magnitude
           else:
               raise ValueError("Unsupported direction: ", direction)
   
           await self._page.mouse.wheel(dx, dy)
           await self._page.wait_for_load_state()
   
       async def go_back(self):
           await self._page.go_back()
           await self._page.wait_for_load_state()
   
       async def navigate(self, url: str, normalize=True):
           normalized_url = url
           if normalize and not normalized_url.startswith(("http://", "https://")):
               normalized_url = "https://" + normalized_url
           await self._page.goto(normalized_url)
           await self._page.wait_for_load_state()
   
       async def key_combination(self, keys: list[str]):
           keys = [PLAYWRIGHT_KEY_MAP.get(k.lower(), k) for k in keys]
           for key in keys[:-1]:
               await self._page.keyboard.down(key)
           await self._page.keyboard.press(keys[-1])
           for key in reversed(keys[:-1]):
               await self._page.keyboard.up(key)
   
       async def current_state(self, it):
           try:
               await self._page.wait_for_load_state("networkidle", timeout=10000)
           except Exception:
               try:
                   await self._page.wait_for_load_state("load", timeout=5000)
               except:
                   pass
   
           await asyncio.sleep(1)
   
           os.makedirs(os.path.join(self.task_dir, "trajectory_som"), exist_ok=True)
           os.makedirs(os.path.join(self.task_dir, "trajectory"), exist_ok=True)
           img_path = os.path.join(self.task_dir, f"trajectory_som/screenshot{it}.png")
           img_path_no_box = os.path.join(self.task_dir, f"trajectory/{it}_full_screenshot.png")
   
           SoM_list, format_ele_text = await get_som(self._page, img_path, img_path_no_box)
           width, height = Image.open(img_path).size
           return {
               "img_path": img_path,
               "img_path_no_box": img_path_no_box,
               "SoM": {
                   "SoM_list": SoM_list,
                   "format_ele_text": format_ele_text
               },
               "current_url": self._page.url,
               "width": width,
               "height": height
           }
   
       async def _select(self, x, y, text):
           await self._page.mouse.click(x, y)
           target_text = text
           handle = await self._page.evaluate_handle("""
           ([x,y]) => document.elementFromPoint(x,y)
           """, [x, y])
           
           tag = await handle.evaluate("el => el && el.tagName")
           if tag == "SELECT":
               await handle.evaluate("""(sel, label) => {
                   const opt = [...sel.options].find(o => o.label.trim() === label.trim());
                   if (!opt) throw new Error("Option not found: " + label);
                   sel.value = opt.value;
                   sel.dispatchEvent(new Event('input', {bubbles:true}));
                   sel.dispatchEvent(new Event('change', {bubbles:true}));
               }""", target_text)
           else:
               raise RuntimeError(f"The element at the coordinates is not a SELECT, but a {tag}")

3. SoM Annotation Module

   import asyncio
   import base64
   import io
   import json
   from typing import Any, Dict, List
   from PIL import Image, ImageDraw, ImageFont
   import random
   
   # Some structural/decorative tags, usually not directly interactive
   STRUCTURAL_TAGS = {
       "html", "head", "meta", "link", "style", "script", "base",
       "title", "body",
       "header", "footer", "main", "nav", "section", "article",
       "aside", "summary", "details",
   }
   
   # Obviously non-interactive ARIA roles
   NON_INTERACTIVE_ROLES = {
       "presentation", "none",
       "img", "banner", "main", "contentinfo",
       "navigation", "region",
   }
   
   def looks_interactive(el: Dict[str, Any]) -> bool:
       tag = (el.get("tag") or "").lower()
       role = (el.get("role") or "").lower()
       typ = (el.get("type") or "").lower()
       text = (el.get("text") or "").strip()
       cls = (el.get("cls") or "").lower()
       id_ = (el.get("id") or "").lower()
       aria_label = (el.get("ariaLabel") or "").strip()
   
       # Native interactive controls
       if tag in {"a", "button", "select", "textarea"}:
           return True
       if tag == "input" and typ not in {"hidden"}:
           return True
   
       # Has a typical interactive role
       if role in {"button", "link", "tab", "menuitem", "option", "switch", "checkbox", "radio", "textbox"}:
           return True
   
       # Has common events or can receive focus
       if any(x in el for x in ("onclick", "onmousedown", "onmouseup")):
           return True
   
       # Has obvious interactive keywords in class / id
       if any(k in cls or k in id_ for k in ("btn", "button", "link", "click", "nav", "tab", "menu")):
           return True
   
       # Has text/aria-label and is a small area
       if (text or aria_label) and tag in {"div", "span", "li"}:
           return True
   
       return False
   
   def is_obviously_non_interactive(el: Dict[str, Any]) -> bool:
       tag = (el.get("tag") or "").lower()
       role = (el.get("role") or "").lower()
       typ = (el.get("type") or "").lower()
       bbox = el.get("bbox") or {}
       w = bbox.get("width") or 0
       h = bbox.get("height") or 0
   
       if tag in STRUCTURAL_TAGS:
           if not (tag == "body" and (el.get("isContentEditable") or "ke-content" in (el.get("cls") or ""))):
               return True
   
       if w * h < 9:
           return True
       if tag in STRUCTURAL_TAGS or tag in {"video", "audio", "source"}:
           return True
       if tag == "input" and typ == "hidden":
           return True
       if role in NON_INTERACTIVE_ROLES:
           return True
       if not looks_interactive(el):
           return True
   
       return False
   
   JS_COLLECT_ALL_RECURSIVE = r"""(args) => {
     const selector = args[0];
     const maxDepth = args[1];
     const maxPerDoc = args[2];
     const minArea = args[3];
     const includeIframes = args[4];
   
     function oneLine(s){ return String(s||"").replace(/\\s+/g," ").trim(); }
   
     function describe(el){
       const tagName = (el.tagName || "").toLowerCase();
       const hrefAttr = el.getAttribute("href") || "";
       const typeAttr = (el.getAttribute("type") || "").toLowerCase();
       const roleAttr = (el.getAttribute("role") || "").toLowerCase();
       const ariaAttr = el.getAttribute("aria-label") || "";
   
       let clickable = true;
       if (tagName == "div") clickable = false;
   
       if (tagName === "a" && hrefAttr) clickable = true;
       if (tagName === "button") clickable = true;
       if (tagName === "input" && typeAttr !== "hidden") clickable = true;
   
       if (["button","link","tab","menuitem","option","switch","checkbox","radio"].includes(roleAttr)) {
           clickable = true;
       }
   
       const isContentEditable = el.isContentEditable || el.getAttribute('contenteditable') === 'true';
       if (isContentEditable) clickable = true;
   
       if (el.onclick === "function") clickable = true;
       try {
           if (!clickable && typeof el.onclick === "function") {
               clickable = true;
           }
       } catch(e) {}
   
       const cls = (el.getAttribute("class") || "").toLowerCase();
       if (cls.includes("ke-content") || cls.includes("ke-edit-textarea")) {
           clickable = true;
       }
   
       return {
           tag: tagName,
           id: oneLine(el.getAttribute("id")),
           cls: oneLine(el.getAttribute("class")),
           role: roleAttr,
           name: oneLine(el.getAttribute("name")),
           type: typeAttr,
           href: oneLine(hrefAttr),
           src: oneLine(el.getAttribute("src")),
           ariaLabel: oneLine(ariaAttr),
           text: oneLine(el.textContent).slice(0, 40),
           clickable: clickable,
           isContentEditable: el.isContentEditable || el.getAttribute('contenteditable') === 'true',
       };
     }
   
     const out = [];
     let counter = 0;
   
     function walk(doc, depth, ox, oy, path){
       if (!doc || depth > maxDepth) return;
       const win = doc.defaultView;
       if (!win) return;
   
       const sx = 0;
       const sy = 0;
       const dpr = win.devicePixelRatio || 1;
   
       const nodes = Array.from(doc.querySelectorAll(selector)).slice(0, maxPerDoc);
       for (const el of nodes){
         try {
           const r = el.getBoundingClientRect();
           if (!r) continue;
           const area = r.width * r.height;
           if (area < minArea) continue;
   
           const vw = win.innerWidth || doc.documentElement.clientWidth || 0;
           const vh = win.innerHeight || doc.documentElement.clientHeight || 0;
           if (r.right <= 0 || r.bottom <= 0 || r.left >= vw || r.top >= vh) {
             continue;
           }
   
           const desc = describe(el);
           if (!desc.clickable) continue;
   
           if (!el.checkVisibility({ checkOpacity: true, checkVisibilityCSS: true })) continue;
   
           out.push({
             id: `e_${counter++}`,
             path,
             depth,
             dpr,
             ...describe(el),
             bbox: {
               x: (ox + r.x - sx) * dpr,
               y: (oy + r.y - sy) * dpr,
               width: r.width * dpr,
               height: r.height * dpr
             }
           });
         } catch(e) {}
       }
   
       if (!includeIframes) return;
   
       const iframes = Array.from(doc.querySelectorAll("iframe")).slice(0, maxPerDoc);
       for (let i = 0; i < iframes.length; i++){
         const fr = iframes[i];
         try {
           const rr = fr.getBoundingClientRect();
           const iframeOx = ox + rr.x - sx;
           const iframeOy = oy + rr.y - sy;
           const nextPath = path + `/iframe[${i}]`;
   
           try {
             const childDoc = fr.contentDocument;
             if (!childDoc) {
               out.push({
                 id: `iframe_${counter++}`,
                 path: nextPath,
                 depth,
                 dpr,
                 tag: "iframe",
                 src: oneLine(fr.getAttribute("src")),
                 error: "iframe not ready (contentDocument is null)",
                 bbox: {x: iframeOx * dpr, y: iframeOy * dpr, width: rr.width * dpr, height: rr.height * dpr},
               });
               continue;
             }
             walk(childDoc, depth + 1, iframeOx, iframeOy, nextPath);
           } catch(e) {
             out.push({
               id: `iframe_${counter++}`,
               path: nextPath,
               depth,
               dpr,
               tag: "iframe",
               src: oneLine(fr.getAttribute("src")),
               error: String(e),
               bbox: {x: iframeOx * dpr, y: iframeOy * dpr, width: rr.width * dpr, height: rr.height * dpr},
             });
           }
         } catch(e) {}
       }
     }
   
     walk(document, 0, 0, 0, "root");
     return JSON.stringify(out);
   }"""
   
   def mark_containing_items_for_removal(items):
       """Mark containing items for removal"""
       def bbox_contains(a, b):
           ax1, ay1 = a["x"], a["y"]
           ax2, ay2 = ax1 + a["width"], ay1 + a["height"]
           bx1, by1 = b["x"], b["y"]
           bx2, by2 = bx1 + b["width"], by1 + b["height"]
           return (bx1 >= ax1 and by1 >= ay1 and bx2 <= ax2 and by2 <= ay2)
   
       for item in items:
           item["to_remove"] = False
   
       n = len(items)
       for i in range(n):
           a = items[i]
           for j in range(n):
               if i == j:
                   continue
               b = items[j]
               if (a.get("text") or "").strip() == (b.get("text") or "").strip():
                   if bbox_contains(a["bbox"], b["bbox"]):
                       a["to_remove"] = True
                       break
   
       return [item for item in items if not item["to_remove"]]
   
   def draw_dashed_line(draw, xy, dash_len=6, gap_len=4, fill=(255, 0, 0, 200), width=2):
       """Draw a dashed line"""
       x1, y1, x2, y2 = xy
       if y1 == y2:  # Horizontal line
           total_len = abs(x2 - x1)
           step = dash_len + gap_len
           n = max(1, int(total_len // step) + 1)
           direction = 1 if x2 >= x1 else -1
           for i in range(n):
               start = x1 + direction * i * step
               end = start + direction * dash_len
               if direction == 1:
                   if start > x2: break
                   end = min(end, x2)
               else:
                   if start < x2: break
                   end = max(end, x2)
               draw.line((start, y1, end, y2), fill=fill, width=width)
       elif x1 == x2:  # Vertical line
           total_len = abs(y2 - y1)
           step = dash_len + gap_len
           n = max(1, int(total_len // step) + 1)
           direction = 1 if y2 >= y1 else -1
           for i in range(n):
               start = y1 + direction * i * step
               end = start + direction * dash_len
               if direction == 1:
                   if start > y2: break
                   end = min(end, y2)
               else:
                   if start < y2: break
                   end = max(end, y2)
               draw.line((x1, start, x2, end), fill=fill, width=width)
       else:
           draw.line((x1, y1, x2, y2), fill=fill, width=width)
   
   def draw_dashed_rect(draw, x1, y1, x2, y2, dash_len=6, gap_len=4, fill=(255, 0, 0, 200), width=2):
       """Draw a dashed rectangle"""
       draw_dashed_line(draw, (x1, y1, x2, y1), dash_len, gap_len, fill, width)
       draw_dashed_line(draw, (x1, y2, x2, y2), dash_len, gap_len, fill, width)
       draw_dashed_line(draw, (x1, y1, x1, y2), dash_len, gap_len, fill, width)
       draw_dashed_line(draw, (x2, y1, x2, y2), dash_len, gap_len, fill, width)
   
   def screenshot_to_png_bytes(s: Any) -> bytes:
       if isinstance(s, (bytes, bytearray)):
           return bytes(s)
       if isinstance(s, str):
           ss = s.strip()
           if ss.startswith("data:image"):
               ss = ss.split(",", 1)[-1].strip()
           try:
               return base64.b64decode(ss, validate=True)
           except Exception:
               with open(s, "rb") as f:
                   return f.read()
       raise TypeError(f"Unsupported screenshot type: {type(s)}")
   
   def items_to_text(items_raw):
       format_ele_text = []
       for web_ele_id in range(len(items_raw)):
           item = items_raw[web_ele_id]
           is_menu = item.get('isMenu', False)
           menu_options = item.get('menuOptions', [])
           label_text = item.get('text', "")
           ele_tag_name = item.get("tag", "button")
           ele_type = item.get("type", "")
           ele_aria_label = item.get("ariaLabel", "")
           input_attr_types = ['text', 'search', 'password', 'email', 'tel']
   
           if is_menu and menu_options:
               trigger_text = label_text.split('\\n')[0].strip()
               options_str = ', '.join([f'"{opt}"' for opt in menu_options])
               base_text = f"[{web_ele_id}]: <{ele_tag_name}>"
               if trigger_text:
                   base_text += f' "{trigger_text}"'
               elif ele_aria_label:
                   base_text += f' "{ele_aria_label}"'
               format_ele_text.append(f"{base_text} is a menu with options: [{options_str}];")
               continue
   
           if not label_text:
               if (ele_tag_name.lower() == 'input' and ele_type in input_attr_types) or \\
                  ele_tag_name.lower() == 'textarea' or \\
                  (ele_tag_name.lower() == 'button' and ele_type in ['submit', 'button']):
                   if ele_aria_label:
                       format_ele_text.append(f'[{web_ele_id}]: <{ele_tag_name}> "{ele_aria_label}";')
                   else:
                       format_ele_text.append(f'[{web_ele_id}]: <{ele_tag_name}> "{label_text}";')
           elif label_text and len(label_text) < 200:
               if not ("<img" in label_text and "src=" in label_text):
                   if ele_tag_name in ["button", "input", "textarea"]:
                       if ele_aria_label and (ele_aria_label != label_text):
                           format_ele_text.append(f'[{web_ele_id}]: <{ele_tag_name}> "{label_text}", "{ele_aria_label}";')
                       else:
                           format_ele_text.append(f'[{web_ele_id}]: <{ele_tag_name}> "{label_text}";')
                   else:
                       if ele_aria_label and (ele_aria_label != label_text):
                           format_ele_text.append(f'[{web_ele_id}]: "{label_text}", "{ele_aria_label}";')
                       else:
                           format_ele_text.append(f'[{web_ele_id}]: "{label_text}";')
   
       return '\\t'.join(format_ele_text)
   
   def draw_som(items, overlay, max_draw):
       try:
           font = ImageFont.truetype(ImageFont.load_default().path, size=20)
       except Exception:
           font = ImageFont.load_default()
   
       placed_label_boxes = []
       draw = ImageDraw.Draw(overlay)
       for idx, it in enumerate(items[:max_draw]):
           b = it.get("bbox") or {}
           x, y, w, h = b.get("x"), b.get("y"), b.get("width"), b.get("height")
           if None in (x, y, w, h) or w <= 0 or h <= 0:
               continue
   
           r, g, b_color = random.randint(0, 255), random.randint(0, 255), random.randint(0, 255)
           color = (r, g, b_color, 255)
   
           x1, y1, x2, y2 = x, y, x + w, y + h
           draw_dashed_rect(draw, x1, y1, x2, y2, dash_len=6, gap_len=4, fill=color, width=2)
   
           if idx < max_draw:
               label = f'{idx}'
               try:
                   tb = draw.textbbox((0, 0), label, font=font)
                   tw, th = tb[2] - tb[0], tb[3] - tb[1]
               except Exception:
                   tw, th = len(label) * 6, 12
   
               padding_x, padding_y = 6, 4
               label_w, label_h = tw + padding_x * 2, th + padding_y * 2
   
               candidates = [
                   ("top_left", x1, y1 - label_h - 2),
                   ("top_right", x2 - label_w, y1 - label_h - 2),
                   ("bottom_left", x1, y2 + 2),
                   ("bottom_right", x2 - label_w, y2 + 2),
               ]
   
               img_w, img_h = overlay.size
               normalized_candidates = [
                   (name, max(0, min(lx, img_w - label_w)), max(0, min(ly, img_h - label_h)))
                   for name, lx, ly in candidates
               ]
   
               best_pos = None
               best_overlap = None
               for name, lx_c, ly_c in normalized_candidates:
                   candidate_rect = (lx_c, ly_c, lx_c + label_w, ly_c + label_h)
                   total_overlap = sum(
                       rect_intersection_area(candidate_rect,placed)
                       for placed in placed_label_boxes
                   )
                   if total_overlap == 0:
                       best_pos = (lx_c, ly_c)
                       best_overlap = 0
                       break
                   if best_overlap is None or total_overlap < best_overlap:
                       best_overlap = total_overlap
                       best_pos = (lx_c, ly_c)
   
               lx, ly = best_pos
               label_box = (lx, ly, lx + label_w, ly + label_h)
               placed_label_boxes.append(label_box)
   
               color = list(color)
               color[-1] = 150
               draw.rectangle([lx, ly, lx + label_w, ly + label_h], fill=tuple(color))
               draw.text((lx + padding_x, ly + padding_y), label, fill=(255, 255, 255, 255), font=font)
   
   def rect_intersection_area(a, b):
       ax1, ay1, ax2, ay2 = a
       bx1, by1, bx2, by2 = b
       ix1, iy1 = max(ax1, bx1), max(ay1, by1)
       ix2, iy2 = min(ax2, bx2), min(ay2, by2)
       if ix2 <= ix1 or iy2 <= iy1:
           return 0
       return (ix2 - ix1) * (iy2 - iy1)
   
   def is_inside_strict(box_inner: dict, box_outer: dict) -> bool:
       x1, y1, w1, h1 = box_inner.get("x"), box_inner.get("y"), box_inner.get("width"), box_inner.get("height")
       x2, y2, w2, h2 = box_outer.get("x"), box_outer.get("y"), box_outer.get("width"), box_outer.get("height")
       i_left, i_top = x1, y1
       i_right, i_bottom = x1 + w1, y1 + h1
       o_left, o_top = x2, y2
       o_right, o_bottom = x2 + w2, y2 + h2
       return (i_left > o_left and i_top > o_top and i_right < o_right and i_bottom < o_bottom)
   
   def remove_outer_boxes(A: list) -> list:
       n = len(A)
       remove_flags = [False] * n
       for i in range(n):
           if remove_flags[i]:
               continue
           box_i = A[i]["bbox"]
           for j in range(n):
               if i == j:
                   continue
               box_j = A[j]["bbox"]
               if is_inside_strict(box_j, box_i):
                   remove_flags[i] = True
                   break
       return [box for k, box in enumerate(A) if not remove_flags[k]]
   
   async def get_css_som(page, selector="*", max_depth=16, max_per_doc=3000, min_area=-1.0, max_retry=3):
       items = []
       for _ in range(max_retry):
           try:
               items_json = await page.evaluate(
                   JS_COLLECT_ALL_RECURSIVE,
                   [selector, max_depth, max_per_doc, float(min_area), True],
               )
               items = json.loads(items_json)
               break
           except Exception:
               try:
                   await page.wait_for_load_state()
               except Exception:
                   pass
               await asyncio.sleep(1)
               continue
   
       if items:
           items = mark_containing_items_for_removal(items)
       return items
   
   def remove_neg_boxes(A: list) -> list:
       n = len(A)
       remove_flags = [False] * n
       for i in range(n):
           box_i = A[i]["bbox"]
           x, y, w, h = box_i.get("x"), box_i.get("y"), box_i.get("width"), box_i.get("height")
           if None in (x, y, w, h) or w <= 0 or h <= 0:
               remove_flags[i] = True
           if box_i.get("x", 0) < 0 or box_i.get("y", 0) < 0:
               remove_flags[i] = True
       return [box for k, box in enumerate(A) if not remove_flags[k]]
   
   async def get_som(page, img_path, img_path_no_box, selector="*", max_depth=16, 
                     max_per_doc=3000, min_area=-1.0, max_draw=2000, max_retry=3):
       """Get SoM annotations for the page"""
       items = await get_css_som(page, selector=selector, max_depth=max_depth, 
                                  max_per_doc=max_per_doc, min_area=min_area, max_retry=max_retry)
       
       shot = await page.screenshot()
       with open(img_path_no_box, "wb") as f:
           f.write(shot)
       
       items = remove_neg_boxes(items)
       
       png_bytes = screenshot_to_png_bytes(shot)
       img = Image.open(io.BytesIO(png_bytes)).convert("RGBA")
       overlay = Image.new("RGBA", img.size, (0, 0, 0, 0))
       draw_som(items, overlay, max_draw)
       img = Image.alpha_composite(img, overlay)
       img.save(img_path)
       
       return items, items_to_text(items)

4. Primary Link Operation

   async def main():
       instruction = 'Search for the results of today's NBA Rockets game'
       history = []
       max_step = 30
       model_name = 'gui-plus-2026-02-26'
       task_dir = f'data_{time.strftime("%Y%m%d_%H%M%S")}'
       web_tools = PlaywrightComputer(initial_url='https://www.baidu.com/', task_dir=task_dir)
   
       await web_tools.reset()
       stop_flag = False
       dashscope.api_key = os.getenv("DASHSCOPE_API_KEY", None)
       session_id = str(uuid.uuid4())
       print('session_id ', session_id)
   
       for step_id in range(max_step):
           if stop_flag:
               break
           safe_instruction = "".join(c if c.isalnum() or c in (" ", "_", "-") else "_" for c in instruction).strip()
           env_state = await web_tools.current_state(it=step_id)
           screen_shot = env_state['img_path']
           messages = get_messages(screen_shot, env_state, instruction, history)
   
           retry_time = 3
           for _ in range(retry_time):
               response = dashscope.MultiModalConversation.call(
                   model=model_name,
                   messages=messages,
                   vl_high_resolution_images=True,
                   headers={"x-dashscope-gui-session-id": session_id},
                   stream=False
               )
               print(response['request_id'])
               try:
                   output_text = response.output.choices[0].message.content[0]['text']
                   break
               except Exception as e:
                   print(response)
                   print(e)
           else:
               raise Exception('retry_time out')
           print(output_text)
   
           thought = response.output.choices[0].message.reasoning_content
           if thought != '':
               output_text = f"<think>\\n{thought}\\n</think>{output_text}"
   
           action_list = extract_tool_calls(output_text)
           conclusion = output_text.split('<tool_call>')[0].strip()
   
           for action_id, action in enumerate(action_list):
               action_parameter = action['arguments']
               action_type = action_parameter['action']
               label = action_parameter.get('label', None)
               weki_url = "https://baike.baidu.com/"
               coordicate = []
               
               if label is not None:
                   if label == "WINDOW":
                       coordicate = [500, 500]
                   else:
                       ele = env_state["SoM"]["SoM_list"][label]
                       box = ele["bbox"]
                       x, y, w, h = box.get("x"), box.get("y"), box.get("width"), box.get("height")
                       nx = x + w / 2
                       ny = y + h / 2
                       coordicate = [nx, ny]
                       action_parameter['coordinate'] = coordicate
   
               if action_type == 'wait':
                   await asyncio.sleep(action_parameter.get('time', 2))
               elif action_type == 'scroll':
                   direction = action_parameter['direction']
                   await web_tools.scroll_at(coordicate[0], coordicate[1], direction=direction, magnitude=300)
               elif action_type == 'select':
                   text = action_parameter['option']
                   await web_tools._select(coordicate[0], coordicate[1], text=text)
               elif action_type == 'goback':
                   await web_tools.go_back()
               elif action_type == 'goto':
                   url = action_parameter.get('url', '')
                   await web_tools.navigate(url, normalize=True)
               elif action_type == 'click':
                   await web_tools.click_at(coordicate[0], coordicate[1])
               elif action_type == 'type':
                   text = action_parameter['text']
                   await web_tools.type_text_at(coordicate[0], coordicate[1], text)
               elif action_type == 'wikipedia':
                   await web_tools.navigate(weki_url, normalize=True)
               elif action_type == 'answer':
                   text = action_parameter['text']
                   print(f'Answer: {text}\\nAction completed')
                   stop_flag = True
                   break
   
               anno_dir = os.path.join(task_dir, 'anno')
               if not os.path.exists(anno_dir):
                   os.makedirs(anno_dir)
               anno_path = show_screenshot(
                   env_state['img_path_no_box'],
                   action_parameter,
                   f'{anno_dir}/anno_{step_id}_{action_id}.png'
               )
   
           history.append({'output': output_text, 'image': screen_shot})
           await asyncio.sleep(2)
   
   if __name__ == '__main__':
       asyncio.run(main())

"""
Complete example code for browser GUI automation
Combines browser_local.py and som.py into a single file that can be run directly

Install dependencies:
pip install playwright dashscope pillow termcolor playwright-stealth
playwright install chromium

Set environment variables before running:
export DASHSCOPE_API_KEY=your_api_key
"""

import json, os, sys, base64
import dashscope, time, math
import uuid
from PIL import Image, ImageDraw
import asyncio
from playwright.async_api import async_playwright, Page, Browser, BrowserContext
from typing import Literal, Any, List, Dict
import random
from datetime import datetime
import io
import re

# ===================== SoM annotation module =====================

# Structural/decorative tags, usually not directly interactive
STRUCTURAL_TAGS = {
    "html", "head", "meta", "link", "style", "script", "base",
    "title", "body", "header", "footer", "main", "nav", "section", "article",
    "aside", "summary", "details",
}

# Obviously non-interactive ARIA roles
NON_INTERACTIVE_ROLES = {
    "presentation", "none", "img", "banner", "main", "contentinfo", "navigation", "region",
}

def looks_interactive(el: Dict[str, Any]) -> bool:
    """Determine if an element looks interactive"""
    tag = (el.get("tag") or "").lower()
    role = (el.get("role") or "").lower()
    typ = (el.get("type") or "").lower()
    text = (el.get("text") or "").strip()
    cls = (el.get("cls") or "").lower()
    id_ = (el.get("id") or "").lower()
    aria_label = (el.get("ariaLabel") or "").strip()

    if tag in {"a", "button", "select", "textarea"}:
        return True
    if tag == "input" and typ not in {"hidden"}:
        return True
    if role in {"button", "link", "tab", "menuitem", "option", "switch", "checkbox", "radio", "textbox"}:
        return True
    if any(x in el for x in ("onclick", "onmousedown", "onmouseup")):
        return True
    if any(k in cls or k in id_ for k in ("btn", "button", "link", "click", "nav", "tab", "menu")):
        return True
    if (text or aria_label) and tag in {"div", "span", "li"}:
        return True
    return False

def is_obviously_non_interactive(el: Dict[str, Any]) -> bool:
    """Determine if an element is obviously non-interactive"""
    tag = (el.get("tag") or "").lower()
    role = (el.get("role") or "").lower()
    typ = (el.get("type") or "").lower()
    bbox = el.get("bbox") or {}
    w = bbox.get("width") or 0
    h = bbox.get("height") or 0

    if tag in STRUCTURAL_TAGS:
        if not (tag == "body" and (el.get("isContentEditable") or "ke-content" in (el.get("cls") or ""))):
            return True
    if w * h < 9:
        return True
    if tag in STRUCTURAL_TAGS or tag in {"video", "audio", "source"}:
        return True
    if tag == "input" and typ == "hidden":
        return True
    if role in NON_INTERACTIVE_ROLES:
        return True
    if not looks_interactive(el):
        return True
    return False

JS_COLLECT_ALL_RECURSIVE = r"""(args) => {
  const selector = args[0];
  const maxDepth = args[1];
  const maxPerDoc = args[2];
  const minArea = args[3];
  const includeIframes = args[4];

  function oneLine(s){ return String(s||"").replace(/\s+/g," ").trim(); }

  function describe(el){
    const tagName = (el.tagName || "").toLowerCase();
    const hrefAttr = el.getAttribute("href") || "";
    const typeAttr = (el.getAttribute("type") || "").toLowerCase();
    const roleAttr = (el.getAttribute("role") || "").toLowerCase();
    const ariaAttr = el.getAttribute("aria-label") || "";

    let clickable = true;
    if (tagName == "div") clickable = false;
    if (tagName === "a" && hrefAttr) clickable = true;
    if (tagName === "button") clickable = true;
    if (tagName === "input" && typeAttr !== "hidden") clickable = true;
    if (["button","link","tab","menuitem","option","switch","checkbox","radio"].includes(roleAttr)) {
        clickable = true;
    }
    const isContentEditable = el.isContentEditable || el.getAttribute('contenteditable') === 'true';
    if (isContentEditable) clickable = true;
    if (el.onclick === "function") clickable = true;
    try {
        if (!clickable && typeof el.onclick === "function") clickable = true;
    } catch(e) {}
    const cls = (el.getAttribute("class") || "").toLowerCase();
    if (cls.includes("ke-content") || cls.includes("ke-edit-textarea")) clickable = true;

    return {
        tag: tagName, id: oneLine(el.getAttribute("id")), cls: oneLine(el.getAttribute("class")),
        role: roleAttr, name: oneLine(el.getAttribute("name")), type: typeAttr,
        href: oneLine(hrefAttr), src: oneLine(el.getAttribute("src")), ariaLabel: oneLine(ariaAttr),
        text: oneLine(el.textContent).slice(0, 40), clickable: clickable,
        isContentEditable: el.isContentEditable || el.getAttribute('contenteditable') === 'true',
    };
  }

  const out = [];
  let counter = 0;

  function walk(doc, depth, ox, oy, path){
    if (!doc || depth > maxDepth) return;
    const win = doc.defaultView;
    if (!win) return;
    const sx = 0, sy = 0;
    const dpr = win.devicePixelRatio || 1;

    const nodes = Array.from(doc.querySelectorAll(selector)).slice(0, maxPerDoc);
    for (const el of nodes){
      try {
        const r = el.getBoundingClientRect();
        if (!r) continue;
        const area = r.width * r.height;
        if (area < minArea) continue;
        const vw = win.innerWidth || doc.documentElement.clientWidth || 0;
        const vh = win.innerHeight || doc.documentElement.clientHeight || 0;
        if (r.right <= 0 || r.bottom <= 0 || r.left >= vw || r.top >= vh) continue;
        const desc = describe(el);
        if (!desc.clickable) continue;
        if (!el.checkVisibility({ checkOpacity: true, checkVisibilityCSS: true })) continue;

        out.push({
          id: `e_${counter++}`, path, depth, dpr, ...describe(el),
          bbox: { x: (ox + r.x - sx) * dpr, y: (oy + r.y - sy) * dpr, width: r.width * dpr, height: r.height * dpr }
        });
      } catch(e) {}
    }
    if (!includeIframes) return;
    const iframes = Array.from(doc.querySelectorAll("iframe")).slice(0, maxPerDoc);
    for (let i = 0; i < iframes.length; i++){
      const fr = iframes[i];
      try {
        const rr = fr.getBoundingClientRect();
        const iframeOx = ox + rr.x - sx;
        const iframeOy = oy + rr.y - sy;
        const nextPath = path + `/iframe[${i}]`;
        try {
          const childDoc = fr.contentDocument;
          if (!childDoc) {
            out.push({ id: `iframe_${counter++}`, path: nextPath, depth, dpr, tag: "iframe",
              src: oneLine(fr.getAttribute("src")), error: "iframe not ready",
              bbox: {x: iframeOx * dpr, y: iframeOy * dpr, width: rr.width * dpr, height: rr.height * dpr} });
            continue;
          }
          walk(childDoc, depth + 1, iframeOx, iframeOy, nextPath);
        } catch(e) {
          out.push({ id: `iframe_${counter++}`, path: nextPath, depth, dpr, tag: "iframe",
            src: oneLine(fr.getAttribute("src")), error: String(e),
            bbox: {x: iframeOx * dpr, y: iframeOy * dpr, width: rr.width * dpr, height: rr.height * dpr} });
        }
      } catch(e) {}
    }
  }
  walk(document, 0, 0, 0, "root");
  return JSON.stringify(out);
}"""

def mark_containing_items_for_removal(items):
    """Remove duplicate elements that are contained"""
    def bbox_contains(a, b):
        ax1, ay1, ax2, ay2 = a["x"], a["y"], a["x"] + a["width"], a["y"] + a["height"]
        bx1, by1, bx2, by2 = b["x"], b["y"], b["x"] + b["width"], b["y"] + b["height"]
        return (bx1 >= ax1 and by1 >= ay1 and bx2 <= ax2 and by2 <= ay2)

    for item in items:
        item["to_remove"] = False
    n = len(items)
    for i in range(n):
        a = items[i]
        for j in range(n):
            if i == j: continue
            b = items[j]
            if (a.get("text") or "").strip() == (b.get("text") or "").strip():
                if bbox_contains(a["bbox"], b["bbox"]):
                    a["to_remove"] = True
                    break
    return [item for item in items if not item["to_remove"]]

def screenshot_to_png_bytes(s: Any) -> bytes:
    if isinstance(s, (bytes, bytearray)):
        return bytes(s)
    if isinstance(s, str):
        ss = s.strip()
        if ss.startswith("data:image"):
            ss = ss.split(",", 1)[-1].strip()
        try:
            return base64.b64decode(ss, validate=True)
        except Exception:
            with open(s, "rb") as f:
                return f.read()
    raise TypeError(f"Unsupported screenshot type: {type(s)}")

def items_to_text(items_raw):
    """Convert a list of elements to a text description"""
    format_ele_text = []
    for web_ele_id in range(len(items_raw)):
        item = items_raw[web_ele_id]
        label_text = item.get('text', "")
        ele_tag_name = item.get("tag", "button")
        ele_type = item.get("type", "")
        ele_aria_label = item.get("ariaLabel", "")
        input_attr_types = ['text', 'search', 'password', 'email', 'tel']

        if not label_text:
            if (ele_tag_name.lower() == 'input' and ele_type in input_attr_types) or \
               ele_tag_name.lower() == 'textarea' or \
               (ele_tag_name.lower() == 'button' and ele_type in ['submit', 'button']):
                format_ele_text.append(f"[{web_ele_id}]: <{ele_tag_name}> \"{ele_aria_label or label_text}\";")
        elif label_text and len(label_text) < 200:
            if not ("<img" in label_text and "src=" in label_text):
                if ele_tag_name in ["button", "input", "textarea"]:
                    format_ele_text.append(f"[{web_ele_id}]: <{ele_tag_name}> \"{label_text}\";")
                else:
                    format_ele_text.append(f"[{web_ele_id}]: \"{label_text}\";")
    return '\t'.join(format_ele_text)

def draw_dashed_rect(draw, x1, y1, x2, y2, dash_len=6, gap_len=4, fill=(255, 0, 0, 200), width=2):
    """Draw a dashed rectangle"""
    def draw_dashed_line(xy):
        px1, py1, px2, py2 = xy
        if py1 == py2:  # Horizontal line
            total_len = abs(px2 - px1)
            step = dash_len + gap_len
            direction = 1 if px2 >= px1 else -1
            for i in range(max(1, int(total_len // step) + 1)):
                start = px1 + direction * i * step
                end = start + direction * dash_len
                if direction == 1:
                    if start > px2: break
                    end = min(end, px2)
                else:
                    if start < px2: break
                    end = max(end, px2)
                draw.line((start, py1, end, py2), fill=fill, width=width)
        elif px1 == px2:  # Vertical line
            total_len = abs(py2 - py1)
            step = dash_len + gap_len
            direction = 1 if py2 >= py1 else -1
            for i in range(max(1, int(total_len // step) + 1)):
                start = py1 + direction * i * step
                end = start + direction * dash_len
                if direction == 1:
                    if start > py2: break
                    end = min(end, py2)
                else:
                    if start < py2: break
                    end = max(end, py2)
                draw.line((px1, start, px2, end), fill=fill, width=width)
    draw_dashed_line((x1, y1, x2, y1))
    draw_dashed_line((x1, y2, x2, y2))
    draw_dashed_line((x1, y1, x1, y2))
    draw_dashed_line((x2, y1, x2, y2))

def draw_som(items, overlay, max_draw):
    """Draw SoM annotations on the screenshot"""
    try:
        font = Image.ImageFont.load_default()
    except Exception:
        font = None
    draw = ImageDraw.Draw(overlay)
    placed_label_boxes = []

    for idx, it in enumerate(items[:max_draw]):
        b = it.get("bbox") or {}
        x, y, w, h = b.get("x"), b.get("y"), b.get("width"), b.get("height")
        if None in (x, y, w, h) or w <= 0 or h <= 0:
            continue

        color = (random.randint(0, 255), random.randint(0, 255), random.randint(0, 255), 255)
        x1, y1, x2, y2 = x, y, x + w, y + h
        draw_dashed_rect(draw, x1, y1, x2, y2, fill=color, width=2)

        if idx < max_draw:
            label = f'{idx}'
            try:
                tb = draw.textbbox((0, 0), label, font=font) if font else (0, 0, len(label) * 6, 12)
                tw, th = tb[2] - tb[0], tb[3] - tb[1]
            except Exception:
                tw, th = len(label) * 6, 12

            padding_x, padding_y = 6, 4
            label_w, label_h = tw + padding_x * 2, th + padding_y * 2

            # Prioritize top-left corner
            lx = max(0, min(x1, overlay.size[0] - label_w))
            ly = max(0, y1 - label_h - 2)

            color_list = list(color)
            color_list[-1] = 150
            draw.rectangle([lx, ly, lx + label_w, ly + label_h], fill=tuple(color_list))
            draw.text((lx + padding_x, ly + padding_y), label, fill=(255, 255, 255, 255), font=font)

def remove_neg_boxes(A: list) -> list:
    """Remove boxes with negative coordinates"""
    return [box for box in A if not (box["bbox"].get("x", 0) < 0 or box["bbox"].get("y", 0) < 0)]

async def get_css_som(page, selector="*", max_depth=16, max_per_doc=3000, min_area=-1.0, max_retry=3):
    """Get CSS elements from the page"""
    items = []
    for _ in range(max_retry):
        try:
            items_json = await page.evaluate(JS_COLLECT_ALL_RECURSIVE, [selector, max_depth, max_per_doc, float(min_area), True])
            items = json.loads(items_json)
            break
        except Exception:
            try:
                await page.wait_for_load_state()
            except Exception:
                pass
            await asyncio.sleep(1)
    if items:
        items = mark_containing_items_for_removal(items)
    return items

async def get_som(page, img_path, img_path_no_box, selector="*", max_depth=16, max_per_doc=3000, min_area=-1.0, max_draw=2000, max_retry=3):
    """Get SoM annotations for the page"""
    items = await get_css_som(page, selector=selector, max_depth=max_depth, max_per_doc=max_per_doc, min_area=min_area, max_retry=max_retry)
    shot = await page.screenshot()
    with open(img_path_no_box, "wb") as f:
        f.write(shot)
    items = remove_neg_boxes(items)
    png_bytes = screenshot_to_png_bytes(shot)
    img = Image.open(io.BytesIO(png_bytes)).convert("RGBA")
    overlay = Image.new("RGBA", img.size, (0, 0, 0, 0))
    draw_som(items, overlay, max_draw)
    img = Image.alpha_composite(img, overlay)
    img.save(img_path)
    return items, items_to_text(items)


# ===================== PlaywrightComputer class =====================

PLAYWRIGHT_KEY_MAP = {
    "backspace": "Backspace", "tab": "Tab", "return": "Enter", "enter": "Enter",
    "shift": "Shift", "control": "ControlOrMeta", "alt": "Alt", "escape": "Escape",
    "space": "Space", "pageup": "PageUp", "pagedown": "PageDown", "end": "End", "home": "Home",
    "left": "ArrowLeft", "up": "ArrowUp", "right": "ArrowRight", "down": "ArrowDown",
    "insert": "Insert", "delete": "Delete", "f1": "F1", "f2": "F2", "f3": "F3", "f4": "F4",
    "f5": "F5", "f6": "F6", "f7": "F7", "f8": "F8", "f9": "F9", "f10": "F10", "f11": "F11", "f12": "F12",
    "command": "Meta",
}

class PlaywrightComputer:
    """Playwright browser automation wrapper"""

    def __init__(self, initial_url="https://www.baidu.com/", task_dir='data', highlight_mouse: bool = False):
        self._initial_url = initial_url
        self.task_dir = task_dir
        self._highlight_mouse = highlight_mouse
        self._playwright = None
        self._browser: Browser | None = None
        self._context: BrowserContext | None = None
        self._page: Page | None = None

    async def reset(self):
        """Initialize the browser"""
        self._playwright = await async_playwright().start()
        self._browser = await self._playwright.chromium.launch(
            args=["--start-maximized", "--disable-blink-features=AutomationControlled"],
            headless=False,
        )
        user_agent = "Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/143.0.0.0 Safari/537.36"
        self._context = await self._browser.new_context(no_viewport=True, user_agent=user_agent, locale="en-US")
        self._page = await self._context.new_page()
        await self._page.goto(self._initial_url, timeout=60000, wait_until="domcontentloaded")
        print("Browser started successfully")

    async def close(self):
        """Close the browser"""
        if self._context:
            await self._context.close()
        if self._browser:
            await self._browser.close()
        if self._playwright:
            await self._playwright.stop()

    async def click_at(self, x: int, y: int):
        await self._page.mouse.click(x, y)
        await self._page.wait_for_load_state()

    async def type_text_at(self, x: int, y: int, text: str, press_enter: bool = True, clear_before_typing: bool = True):
        await self._page.mouse.click(x, y)
        await asyncio.sleep(0.1)
        if clear_before_typing:
            await self.key_combination(["Control", "A"])
            await asyncio.sleep(0.1)
            await self.key_combination(["Delete"])
        await self._page.keyboard.type(text)
        if press_enter:
            await self.key_combination(["Enter"])
        await self._page.wait_for_load_state()

    async def scroll_at(self, x: int, y: int, direction: Literal["up", "down"], magnitude: int = 400):
        await self._page.mouse.move(x, y)
        await asyncio.sleep(0.1)
        dy = -magnitude if direction == "up" else magnitude
        await self._page.mouse.wheel(0, dy)
        await self._page.wait_for_load_state()

    async def go_back(self):
        await self._page.go_back()
        await self._page.wait_for_load_state()

    async def navigate(self, url: str, normalize=True):
        if normalize and not url.startswith(("http://", "https://")):
            url = "https://" + url
        await self._page.goto(url)
        await self._page.wait_for_load_state()

    async def key_combination(self, keys: list[str]):
        keys = [PLAYWRIGHT_KEY_MAP.get(k.lower(), k) for k in keys]
        for key in keys[:-1]:
            await self._page.keyboard.down(key)
        await self._page.keyboard.press(keys[-1])
        for key in reversed(keys[:-1]):
            await self._page.keyboard.up(key)

    async def current_state(self, it):
        """Get the current page state and SoM annotations"""
        try:
            await self._page.wait_for_load_state("networkidle", timeout=10000)
        except Exception:
            pass
        await asyncio.sleep(1)

        os.makedirs(os.path.join(self.task_dir, "trajectory_som"), exist_ok=True)
        os.makedirs(os.path.join(self.task_dir, "trajectory"), exist_ok=True)
        img_path = os.path.join(self.task_dir, f"trajectory_som/screenshot{it}.png")
        img_path_no_box = os.path.join(self.task_dir, f"trajectory/{it}_full_screenshot.png")

        SoM_list, format_ele_text = await get_som(self._page, img_path, img_path_no_box)
        width, height = Image.open(img_path).size
        return {
            "img_path": img_path, "img_path_no_box": img_path_no_box,
            "SoM": {"SoM_list": SoM_list, "format_ele_text": format_ele_text},
            "current_url": self._page.url, "width": width, "height": height
        }

    async def _select(self, x, y, text):
        await self._page.mouse.click(x, y)
        handle = await self._page.evaluate_handle("([x,y]) => document.elementFromPoint(x,y)", [x, y])
        tag = await handle.evaluate("el => el && el.tagName")
        if tag == "SELECT":
            await handle.evaluate("""(sel, label) => {
                const opt = [...sel.options].find(o => o.label.trim() === label.trim());
                if (!opt) throw new Error("Option not found: " + label);
                sel.value = opt.value;
                sel.dispatchEvent(new Event('input', {bubbles:true}));
                sel.dispatchEvent(new Event('change', {bubbles:true}));
            }""", text)
        else:
            raise RuntimeError(f"The element at the coordinates is not a SELECT, but a {tag}")


# ===================== Prompt construction =====================

def get_messages(image, env_state, instruction, history_output):
    """Construct multi-turn conversation messages"""
    history_n = 1
    current_step = len(history_output)
    history_start_idx = max(0, current_step - history_n)
    SoM_format_ele_text = env_state["SoM"]["format_ele_text"]

    system_prompt = '''# Tools

You may call one or more functions to assist with the user query.

You are provided with function signatures within <tools></tools> XML tags:
<tools>
{"type": "function", "function": {"name_for_human": "browser_use", "name": "browser_use", "description": "Use a browser to interact with web pages and take labeled screenshots.\\n* This is an interface to a web browser. You can click elements, type into inputs, scroll, wait for loading, go back, etc.\\n* Each Observation screenshot contains Numerical Labels placed at the TOP LEFT of each Web Element. Use these labels to target elements.\\n* Some pages may take time to load; you may need to wait and take successive screenshots.\\n* Avoid clicking near element edges; target the center of the element.\\n* Execute exactly ONE interaction action per step; do not chain multiple interactions in one call.", "parameters": {"properties": {"action": {"description": "The action to perform. The available actions are:\\n* `click`: Click a web element by numerical label.\\n* `type`: Clear existing content in a textbox/input and type content. The system will automatically press ENTER after typing.\\n* `scroll`: Scroll within WINDOW or within a specific scrollable element/area (by label).\\n* `select`: Selects a specific option from a menu or dropdown. Use the option text provided in the textual information.\\n* `wait`: Wait for page processes to finish (default 5 seconds unless specified).\\n* `go_back`: Go back to the previous page.\\n* `wikipedia`: Directly jump to the Wikipedia homepage to search for information.\\n* `answer`: Terminate the current task and output the final answer.", "enum": ["click", "type", "scroll", "select", "wait", "go_back", "wikipedia", "answer"], "type": "string"}, "label": {"description": "Numerical label of the target web element. Required only by `action=click`, `action=type`, `action=scroll`, and `action=select` when scrolling within a specific area. Use string value `WINDOW` to scroll the whole page.", "type": ["integer", "string"]}, "direction": {"description": "Scroll direction. Required only by `action=scroll`.", "enum": ["up", "down"], "type": "string"}, "text": {"description": "Required only by `action=type` and `action=answer`.", "type": "string"}, "option": {"description": "The option to select. Required only by `action=select`", "type": "string"}, "time": {"description": "The seconds to wait. Required only by `action=wait` when overriding the default.", "type": "integer"}}, "required": ["action"], "type": "object"}, "args_format": "Format the arguments as a JSON object."}}
</tools>

For each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:
<tool_call>
{"name": <function-name>, "arguments": <args-json-object>}
</tool_call>

# Response format

Response format for every step:
1) Action: a short imperative describing what to do in the UI.
2) A single <tool_call>...</tool_call> block containing only the JSON: {"name": <function-name>, "arguments": <args-json-object>}.

Rules:
- Output exactly in the order: Action, <tool_call>.
- Be brief: one line for Action.
- Do not output anything else outside those two parts.
- Execute ONLY ONE interaction per iteration (one tool call).
- If finishing, use action=answer in the tool call.'''

    today = datetime.today()
    weekday_names = ["Monday", "Tuesday", "Wednesday", "Thursday", "Friday", "Saturday", "Sunday"]
    weekday = weekday_names[today.weekday()]
    formatted_date = today.strftime("%Y-%m-%d") + " " + weekday
    day_info = f'''Today's date is: {formatted_date}.'''

    previous_actions = []
    for i in range(history_start_idx):
        if i < len(history_output):
            history_output_str = history_output[i]['output']
            if 'Action:' in history_output_str and '<tool_call>':
                history_output_str = history_output_str.split('Action:')[1].split('<tool_call>')[0].strip()
            previous_actions.append(f"Step {i + 1}: {history_output_str}")
    previous_actions_str = "\n".join(previous_actions) if previous_actions else "None"

    instruction_prompt = f'''\nPlease generate the next move according to the UI screenshot, instruction and previous actions.

Instruction: {day_info}{instruction}

Previous actions:
{previous_actions_str}'''

    messages = [{"role": "system", "content": [{"text": system_prompt}]}]
    history_len = min(history_n, len(history_output))

    if history_len > 0:
        for history_id, history_item in enumerate(history_output[-history_n:], 0):
            if history_id == 0:
                messages.append({"role": "user", "content": [
                    {"text": instruction_prompt}, {"text": "Current screenshot:"},
                    {"image": "file://" + history_item['image']}
                ]})
            else:
                messages.append({"role": "user", "content": [
                    {"text": "Current screenshot:"}, {"image": "file://" + history_item['image']}
                ]})
            messages.append({"role": "assistant", "content": [{"text": history_item['output']}]})
        messages.append({"role": "user", "content": [
            {"image": "file://" + image}, {"text": SoM_format_ele_text}
        ]})
    else:
        messages.append({"role": "user", "content": [
            {"text": instruction_prompt}, {"image": "file://" + image}, {"text": SoM_format_ele_text}
        ]})
    return messages


def extract_tool_calls(text: str):
    """Extract tool_call from text"""
    pattern = re.compile(r'<tool_call>(.*?)</tool_call>', re.DOTALL | re.IGNORECASE)
    blocks = pattern.findall(text)
    actions = []
    for blk in blocks:
        blk = blk.strip()
        try:
            actions.append(json.loads(blk))
        except json.JSONDecodeError as e:
            print(f'Parsing failed: {e}')
    return actions


# ===================== Main program =====================

def show_screenshot(image_path, action_parameter, save_path="screenshot_anno.png"):
    """Annotate the action point on the screenshot"""
    image = Image.open(image_path)
    draw = ImageDraw.Draw(image)
    if 'coordinate' in action_parameter:
        radius = 15
        center_x, center_y = action_parameter['coordinate'][0], action_parameter['coordinate'][1]
        draw.ellipse((center_x - radius, center_y - radius, center_x + radius, center_y + radius), fill="red", outline="red")
    else:
        return None
    image.save(save_path)
    return save_path

async def main():
    instruction = 'Search for the results of today's NBA Rockets game'
    history = []
    max_step = 30
    model_name = 'gui-plus-2026-02-26'
    task_dir = f'data_{time.strftime("%Y%m%d_%H%M%S")}'
    web_tools = PlaywrightComputer(initial_url='https://www.baidu.com/', task_dir=task_dir)

    await web_tools.reset()
    stop_flag = False
    dashscope.api_key = os.getenv("DASHSCOPE_API_KEY", None)
    session_id = str(uuid.uuid4())
    print('session_id ', session_id)

    for step_id in range(max_step):
        if stop_flag:
            break
        env_state = await web_tools.current_state(it=step_id)
        screen_shot = env_state['img_path']
        messages = get_messages(screen_shot, env_state, instruction, history)

        retry_time = 3
        for _ in range(retry_time):
            response = dashscope.MultiModalConversation.call(
                model=model_name, messages=messages, vl_high_resolution_images=True,
                headers={"x-dashscope-gui-session-id": session_id},
                stream=False
            )
            print(response['request_id'])
            try:
                output_text = response.output.choices[0].message.content[0]['text']
                break
            except Exception as e:
                print(response)
                print(e)
        else:
            raise Exception('retry_time out')
        print(output_text)

        thought = response.output.choices[0].message.reasoning_content
        if thought != '':
            output_text = f"<think>\n{thought}\n</think>{output_text}"

        action_list = extract_tool_calls(output_text)

        for action_id, action in enumerate(action_list):
            action_parameter = action['arguments']
            action_type = action_parameter['action']
            label = action_parameter.get('label', None)
            weki_url = "https://baike.baidu.com/"
            coordicate = []

            if label is not None:
                if label == "WINDOW":
                    coordicate = [500, 500]
                else:
                    ele = env_state["SoM"]["SoM_list"][label]
                    box = ele["bbox"]
                    x, y, w, h = box.get("x"), box.get("y"), box.get("width"), box.get("height")
                    nx = x + w / 2
                    ny = y + h / 2
                    coordicate = [nx, ny]
                    action_parameter['coordinate'] = coordicate

            if action_type == 'wait':
                await asyncio.sleep(action_parameter.get('time', 2))
            elif action_type == 'scroll':
                direction = action_parameter['direction']
                await web_tools.scroll_at(coordicate[0], coordicate[1], direction=direction, magnitude=300)
            elif action_type == 'select':
                text = action_parameter['option']
                await web_tools._select(coordicate[0], coordicate[1], text=text)
            elif action_type == 'goback':
                await web_tools.go_back()
            elif action_type == 'goto':
                url = action_parameter.get('url', '')
                await web_tools.navigate(url, normalize=True)
            elif action_type == 'click':
                await web_tools.click_at(coordicate[0], coordicate[1])
            elif action_type == 'type':
                text = action_parameter['text']
                await web_tools.type_text_at(coordicate[0], coordicate[1], text)
            elif action_type == 'wikipedia':
                await web_tools.navigate(weki_url, normalize=True)
            elif action_type == 'answer':
                text = action_parameter['text']
                print(f'Answer: {text}\nTask completed')
                stop_flag = True
                break

            anno_dir = os.path.join(task_dir, 'anno')
            if not os.path.exists(anno_dir):
                os.makedirs(anno_dir)
            show_screenshot(env_state['img_path_no_box'], action_parameter, f'{anno_dir}/anno_{step_id}_{action_id}.png')

        history.append({'output': output_text, 'image': screen_shot})
        await asyncio.sleep(2)

if __name__ == '__main__':
    asyncio.run(main())