diff --git a/lib_run_single.py b/lib_run_single.py
index 0d3fefb..4bc37e1 100644
--- a/lib_run_single.py
+++ b/lib_run_single.py
@@ -267,3 +267,63 @@ def run_single_example_autoglm(agent, env, example, max_steps, instruction, args
     with open(os.path.join(example_result_dir, "result.txt"), "w", encoding="utf-8") as f:
         f.write(f"{result}\n")
     env.controller.end_recording(os.path.join(example_result_dir, "recording.mp4"))
+
+def run_single_example_mano(agent, env, example, max_steps, instruction, args, example_result_dir, scores):
+    runtime_logger = setup_logger(example, example_result_dir)
+    agent.reset(runtime_logger)
+    env.reset(task_config=example)
+    time.sleep(60) # Wait for the environment to be ready
+    obs = env._get_obs() # Get the initial observation
+    done = False
+    step_idx = 0
+    env.controller.start_recording()
+    
+    with open(os.path.join(example_result_dir, f"step_0.png"),
+      "wb") as _f:
+        _f.write(obs['screenshot'])
+    while not done and step_idx < max_steps:
+        response, actions = agent.predict(
+            instruction,
+            obs
+        )
+        if len(actions) > 1:
+            if (("pyautogui.hotkey('shift')" in actions[0] or "pyautogui.hotkey('ctrl')" in actions[0]) 
+                and "pyautogui.click" in actions[1]):
+                hotkey_type = 'shift' if "shift" in actions[0] else 'ctrl'
+                action = f"pyautogui.keyDown('{hotkey_type}')\n{actions[1]}\npyautogui.keyUp('{hotkey_type}')"
+                actions = [action]  
+                
+        for action in actions:
+            # Capture the timestamp before executing the action
+            action_timestamp = datetime.datetime.now().strftime("%Y%m%d@%H%M%S")
+            logger.info("Step %d: %s", step_idx + 1, action)
+            obs, reward, done, info = env.step(action, args.sleep_after_execution)
+
+            logger.info("Reward: %.2f", reward)
+            logger.info("Done: %s", done)
+            # Save screenshot and trajectory information
+            with open(os.path.join(example_result_dir, f"step_{step_idx + 1}_{action_timestamp}.png"),
+                      "wb") as _f:
+                _f.write(obs['screenshot'])
+            with open(os.path.join(example_result_dir, "traj.jsonl"), "a") as f:
+                f.write(json.dumps({
+                    "step_num": step_idx + 1,
+                    "action_timestamp": action_timestamp,
+                    "action": action,
+                    "reward": reward,
+                    "done": done,
+                    "info": info,
+                    "screenshot_file": f"step_{step_idx + 1}_{action_timestamp}.png",
+                    "response":response
+                }))
+                f.write("\n")
+            if done:
+                logger.info("The episode is done.")
+                break
+        step_idx += 1
+    result = env.evaluate()
+    logger.info("Result: %.2f", result)
+    scores.append(result)
+    with open(os.path.join(example_result_dir, "result.txt"), "w", encoding="utf-8") as f:
+        f.write(f"{result}\n")
+    env.controller.end_recording(os.path.join(example_result_dir, "recording.mp4"))
\ No newline at end of file
diff --git a/mm_agents/mano_agent.py b/mm_agents/mano_agent.py
new file mode 100644
index 0000000..d85a6a8
--- /dev/null
+++ b/mm_agents/mano_agent.py
@@ -0,0 +1,1068 @@
+import ast
+import base64
+import logging
+import math
+import re
+import os
+import xml.etree.ElementTree as ET
+from io import BytesIO
+from typing import Dict, List
+
+import backoff
+import numpy as np
+from PIL import Image, ImageDraw
+from requests.exceptions import SSLError
+import openai
+from openai import OpenAI
+from google.api_core.exceptions import (
+    BadRequest,
+    InternalServerError,
+    InvalidArgument,
+    ResourceExhausted,
+)
+
+from mm_agents.accessibility_tree_wrap.heuristic_retrieve import (
+    filter_nodes,
+)
+
+# from mm_agents.prompts import MANO_PROMPT_THOUGHT
+
+
+logger = logging.getLogger("desktopenv.agent")
+
+FINISH_WORD = "finished"
+WAIT_WORD = "wait"
+ENV_FAIL_WORD = "error_env"
+CALL_USER = "call_user"
+
+IMAGE_FACTOR = 28
+MIN_PIXELS = 100 * 28 * 28
+MAX_PIXELS = 16384 * 28 * 28
+MAX_RATIO = 200
+MANO_PROMPT_THOUGHT = """You are a GUI agent. You are given a task and your action history, with screenshots. You need to perform the next action to complete the task. 
+
+## Output Format
+```
+Thought: ...
+Action desp: ...
+Action: ...
+```
+
+## Action Space
+click(start_box='<|box_start|>(x1,y1)<|box_end|>')
+left_double(start_box='<|box_start|>(x1,y1)<|box_end|>')
+right_single(start_box='<|box_start|>(x1,y1)<|box_end|>')
+drag(start_box='<|box_start|>(x1,y1)<|box_end|>', end_box='<|box_start|>(x3,y3)<|box_end|>')
+hotkey(key='')
+type(content='') #If you want to submit your input, use "\\n" at the end of `content`.
+scroll(start_box='<|box_start|>(x1,y1)<|box_end|>', direction='down or up or right or left')
+wait() #Sleep for 5s and take a screenshot to check for any changes.
+finished(content='xxx') # Use escape characters \\', \\", and \\n in content part to ensure we can parse the content in normal python string format.
+
+## Note
+- Use {language} in `Thought` part.
+- Write a small plan and finally summarize your next action (with its target element) in one sentence in `Action desp` part.
+
+## User Instruction
+{instruction}
+"""
+
+
+pure_text_settings = ["a11y_tree"]
+
+attributes_ns_ubuntu = "https://accessibility.windows.example.org/ns/attributes"
+attributes_ns_windows = "https://accessibility.windows.example.org/ns/attributes"
+state_ns_ubuntu = "https://accessibility.ubuntu.example.org/ns/state"
+state_ns_windows = "https://accessibility.windows.example.org/ns/state"
+component_ns_ubuntu = "https://accessibility.ubuntu.example.org/ns/component"
+component_ns_windows = "https://accessibility.windows.example.org/ns/component"
+value_ns_ubuntu = "https://accessibility.ubuntu.example.org/ns/value"
+value_ns_windows = "https://accessibility.windows.example.org/ns/value"
+class_ns_windows = "https://accessibility.windows.example.org/ns/class"
+# More namespaces defined in OSWorld, please check desktop_env/server/main.py
+
+# 定义一个函数来解析每个 action
+def parse_action(action_str):
+    try:
+        # 解析字符串为 AST 节点
+        node = ast.parse(action_str, mode='eval')
+
+        # 确保节点是一个表达式
+        if not isinstance(node, ast.Expression):
+            raise ValueError("Not an expression")
+
+        # 获取表达式的主体
+        call = node.body
+
+        # 确保主体是一个函数调用
+        if not isinstance(call, ast.Call):
+            raise ValueError("Not a function call")
+
+        # 获取函数名
+        if isinstance(call.func, ast.Name):
+            func_name = call.func.id
+        elif isinstance(call.func, ast.Attribute):
+            func_name = call.func.attr
+        else:
+            func_name = None
+
+        # 获取关键字参数
+        kwargs = {}
+        for kw in call.keywords:
+            key = kw.arg
+            # 处理不同类型的值，这里假设都是常量
+            if isinstance(kw.value, ast.Constant):
+                value = kw.value.value
+            elif isinstance(kw.value, ast.Str):  # 兼容旧版本 Python
+                value = kw.value.s
+            else:
+                value = None
+            value = str(value)
+            kwargs[key] = value
+
+        return {
+            'function': func_name,
+            'args': kwargs
+        }
+
+    except Exception as e:
+        print(f"Failed to parse action '{action_str}': {e}")
+        return None
+    
+def escape_single_quotes(text):
+    # 匹配未转义的单引号（不匹配 \\'）
+    pattern = r"(?<!\\)'"
+    return re.sub(pattern, r"\\'", text)
+
+def round_by_factor(number: int, factor: int) -> int:
+    """Returns the closest integer to 'number' that is divisible by 'factor'."""
+    return round(number / factor) * factor
+
+
+def ceil_by_factor(number: int, factor: int) -> int:
+    """Returns the smallest integer greater than or equal to 'number' that is divisible by 'factor'."""
+    return math.ceil(number / factor) * factor
+
+
+def floor_by_factor(number: int, factor: int) -> int:
+    """Returns the largest integer less than or equal to 'number' that is divisible by 'factor'."""
+    return math.floor(number / factor) * factor
+
+def linear_resize(
+    height: int, width: int, factor: int = IMAGE_FACTOR, min_pixels: int = MIN_PIXELS, max_pixels: int = MAX_PIXELS
+) -> tuple[int, int]:
+    if width * height > max_pixels:
+        """
+        如果图片超过/低于像素限制，则计算一个缩放因子resize_factor，使图片的像素数缩小到等于或小于max_pixels。这个缩放因子是通过开平方根计算的，确保纵横比保持不变,这样原始的相对坐标可以不经转换直接复用
+        """
+        resize_factor = math.sqrt(max_pixels / (width * height))
+        width, height = int(width * resize_factor), int(height * resize_factor)
+    if width * height < min_pixels:
+        resize_factor = math.sqrt(min_pixels / (width * height))
+        width, height = math.ceil(width * resize_factor), math.ceil(height * resize_factor)
+
+    return height, width 
+
+def smart_resize(
+    height: int, width: int, factor: int = IMAGE_FACTOR, min_pixels: int = MIN_PIXELS, max_pixels: int = MAX_PIXELS
+) -> tuple[int, int]:
+    """
+    Rescales the image so that the following conditions are met:
+
+    1. Both dimensions (height and width) are divisible by 'factor'.
+
+    2. The total number of pixels is within the range ['min_pixels', 'max_pixels'].
+
+    3. The aspect ratio of the image is maintained as closely as possible.
+    """
+    if max(height, width) / min(height, width) > MAX_RATIO:
+        raise ValueError(
+            f"absolute aspect ratio must be smaller than {MAX_RATIO}, got {max(height, width) / min(height, width)}"
+        )
+    h_bar = max(factor, round_by_factor(height, factor))
+    w_bar = max(factor, 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
+
+def parse_action_to_structure_output(text, factor, origin_resized_height, origin_resized_width, model_type, max_pixels=16384*28*28, min_pixels=100*28*28):
+    text = text.strip()
+    if model_type == "qwen25vl":
+        smart_resize_height, smart_resize_width = smart_resize(origin_resized_height, origin_resized_width, factor=IMAGE_FACTOR, min_pixels=min_pixels, max_pixels=max_pixels)
+
+    # 正则表达式匹配 Action 字符串
+    if text.startswith("Thought:") and "Action desp" not in text:
+        thought_pattern = r"Thought: (.+?)(?=\s*Action:|$)"
+        thought_hint = "Thought: "
+    elif text.startswith("Thought:") and "Action desp" in text:
+        thought_pattern = r"Thought: (.+?)(?=\s*Action desp:|$)"
+        thought_hint = "Thought: "
+    elif text.startswith("Reflection:"):
+        thought_pattern = r"Reflection: (.+?)Action_Summary: (.+?)(?=\s*Action:|$)"
+        thought_hint = "Reflection: "
+    elif text.startswith("Action_Summary:"):
+        thought_pattern = r"Action_Summary: (.+?)(?=\s*Action:|$)"
+        thought_hint = "Action_Summary: "
+    else:
+        thought_pattern = r"Thought: (.+?)(?=\s*Action:|$)"
+        thought_hint = "Thought: "
+    reflection, thought = None, None
+    thought_match = re.search(thought_pattern, text, re.DOTALL)
+    if thought_match:
+        if len(thought_match.groups()) == 1:
+            thought = thought_match.group(1).strip()
+        elif len(thought_match.groups()) == 2:
+            thought = thought_match.group(2).strip()
+            reflection = thought_match.group(1).strip()
+    assert "Action:" in text
+    action_str = text.split("Action:")[-1]
+
+    tmp_all_action = action_str.split("\n\n")
+    all_action = []
+    for action_str in tmp_all_action:
+        if "type(content" in action_str:
+            # 正则表达式匹配 content 中的字符串并转义单引号
+            def escape_quotes(match):
+                content = match.group(1)  # 获取 content 的值
+                return content
+
+            # 使用正则表达式进行替换
+            pattern = r"type\(content='(.*?)'\)"  # 匹配 type(content='...')
+            content = re.sub(pattern, escape_quotes, action_str)
+
+            # 处理字符串
+            action_str = escape_single_quotes(content)
+            action_str = "type(content='" + action_str + "')"
+        all_action.append(action_str)
+
+    parsed_actions = [parse_action(action.replace("\n","\\n").lstrip()) for action in all_action]
+    actions = []
+    for action_instance, raw_str in zip(parsed_actions, all_action):
+        if action_instance == None:
+            print(f"Action can't parse: {raw_str}")
+            raise ValueError(f"Action can't parse: {raw_str}") 
+        action_type = action_instance["function"]
+        params = action_instance["args"]
+
+        # import pdb; pdb.set_trace()
+        action_inputs = {}
+        for param_name, param in params.items():
+            if param == "": continue
+            param = param.lstrip()  # 去掉引号和多余的空格
+            # 处理start_box或者end_box参数格式 '<bbox>x1 y1 x2 y2</bbox>'
+            action_inputs[param_name.strip()] = param
+            
+            if "start_box" in param_name or "end_box" in param_name:
+                ori_box = param
+                # Remove parentheses and split the string by commas
+                numbers = ori_box.replace("(", "").replace(")", "").split(",")
+
+                # Convert to float and scale by 1000
+                # Qwen2.5vl output absolute coordinates, qwen2vl output relative coordinates
+                if model_type == "qwen25vl":
+                    float_numbers = []
+                    for num_idx, num in enumerate(numbers):
+                        num = float(num)
+                        if (num_idx + 1) % 2 == 0:
+                            float_numbers.append(float(num/smart_resize_height))
+                        else:
+                            float_numbers.append(float(num/smart_resize_width))
+                else:
+                    float_numbers = [float(num) / factor for num in numbers]
+
+                if len(float_numbers) == 2:
+                    float_numbers = [float_numbers[0], float_numbers[1], float_numbers[0], float_numbers[1]]
+                action_inputs[param_name.strip()] = str(float_numbers)
+
+        # import pdb; pdb.set_trace()
+        actions.append({
+            "reflection": reflection,
+            "thought": thought,
+            "action_type": action_type,
+            "action_inputs": action_inputs,
+            "text": text
+        })
+    return actions
+
+def parsing_response_to_pyautogui_code(responses, image_height: int, image_width:int, input_swap:bool=True) -> str:
+    '''
+    将M模型的输出解析为OSWorld中的action，生成pyautogui代码字符串
+    参数:
+        response: 包含模型输出的字典，结构类似于：
+        {
+            "action_type": "hotkey",
+            "action_inputs": {
+                "hotkey": "v ctrl",
+                "start_box": None,
+                "end_box": None
+            }
+        }
+    返回:
+        生成的pyautogui代码字符串
+    '''
+
+    pyautogui_code = f"import pyautogui\nimport time\n"
+    if isinstance(responses, dict):
+        responses = [responses]
+    for response_id, response in enumerate(responses):
+        if "observation" in response:
+            observation = response["observation"]
+        else:
+            observation = ""
+
+        if "thought" in response:
+            thought = response["thought"]
+        else:
+            thought = ""
+        
+        if response_id == 0:
+            pyautogui_code += f"'''\nObservation:\n{observation}\n\nThought:\n{thought}\n'''\n"
+        else:
+            pyautogui_code += f"\ntime.sleep(1)\n"
+
+        action_dict = response
+        action_type = action_dict.get("action_type")
+        action_inputs = action_dict.get("action_inputs", {})
+        
+        if action_type == "hotkey":
+            # Parsing hotkey action
+            if "key" in action_inputs:
+                hotkey = action_inputs.get("key", "")
+            else:
+                hotkey = action_inputs.get("hotkey", "")
+
+            if hotkey == "arrowleft":
+                hotkey = "left"
+
+            elif hotkey == "arrowright":
+                hotkey = "right"
+            
+            elif hotkey == "arrowup":
+                hotkey = "up"
+            
+            elif hotkey == "arrowdown":
+                hotkey = "down"
+
+            if hotkey:
+                # Handle other hotkeys
+                keys = hotkey.split()  # Split the keys by space
+                convert_keys = []
+                for key in keys:
+                    if key == "space":
+                        key = ' '
+                    convert_keys.append(key)
+                pyautogui_code += f"\npyautogui.hotkey({', '.join([repr(k) for k in convert_keys])})"
+        
+        elif action_type == "press":
+            # Parsing press action
+            if "key" in action_inputs:
+                key_to_press = action_inputs.get("key", "")
+            else:
+                key_to_press = action_inputs.get("press", "")
+
+            if hotkey == "arrowleft":
+                hotkey = "left"
+
+            elif hotkey == "arrowright":
+                hotkey = "right"
+            
+            elif hotkey == "arrowup":
+                hotkey = "up"
+            
+            elif hotkey == "arrowdown":
+                hotkey = "down"
+            
+            elif hotkey == "space":
+                hotkey = " "
+                
+            if key_to_press:
+                # Simulate pressing a single key
+                pyautogui_code += f"\npyautogui.press({repr(key_to_press)})"
+            
+        elif action_type == "keyup":
+            key_to_up = action_inputs.get("key", "")
+            pyautogui_code += f"\npyautogui.keyUp({repr(key_to_up)})"
+        
+        elif action_type == "keydown":
+            key_to_down = action_inputs.get("key", "")
+            pyautogui_code += f"\npyautogui.keyDown({repr(key_to_down)})"
+
+        elif action_type == "type":
+            # Parsing typing action using clipboard
+            content = action_inputs.get("content", "")
+            content = escape_single_quotes(content)
+            stripped_content = content
+            if content.endswith("\n") or content.endswith("\\n"):
+                stripped_content = stripped_content.rstrip("\\n").rstrip("\n")
+            if content:
+                if input_swap:
+                    pyautogui_code += f"\nimport pyperclip"
+                    pyautogui_code += f"\npyperclip.copy('{stripped_content}')"
+                    pyautogui_code += f"\npyautogui.hotkey('ctrl', 'v')"
+                    pyautogui_code += f"\ntime.sleep(0.5)\n"
+                    if content.endswith("\n") or content.endswith("\\n"):
+                        pyautogui_code += f"\npyautogui.press('enter')"
+                else:
+                    pyautogui_code += f"\npyautogui.write('{stripped_content}', interval=0.1)"
+                    pyautogui_code += f"\ntime.sleep(0.5)\n"
+                    if content.endswith("\n") or content.endswith("\\n"):
+                        pyautogui_code += f"\npyautogui.press('enter')"
+
+        
+        elif action_type in ["drag", "select"]:
+            # Parsing drag or select action based on start and end_boxes
+            start_box = action_inputs.get("start_box")
+            end_box = action_inputs.get("end_box")
+            if start_box and end_box:
+                x1, y1, x2, y2 = eval(start_box)  # Assuming box is in [x1, y1, x2, y2]
+                sx = round(float((x1 + x2) / 2) * image_width, 3)
+                sy = round(float((y1 + y2) / 2) * image_height, 3)
+                x1, y1, x2, y2 = eval(end_box)  # Assuming box is in [x1, y1, x2, y2]
+                ex = round(float((x1 + x2) / 2) * image_width, 3)
+                ey = round(float((y1 + y2) / 2) * image_height, 3)
+                pyautogui_code += (
+                    f"\npyautogui.moveTo({sx}, {sy})\n"
+                    f"\npyautogui.dragTo({ex}, {ey}, duration=1.0)\n"
+                )
+
+
+        elif action_type == "scroll":
+            # Parsing scroll action
+            start_box = action_inputs.get("start_box")
+            if start_box:
+                x1, y1, x2, y2 = eval(start_box)  # Assuming box is in [x1, y1, x2, y2]
+                x = round(float((x1 + x2) / 2) * image_width, 3)
+                y = round(float((y1 + y2) / 2) * image_height, 3)
+            else:
+                x = None
+                y = None
+            
+            direction = action_inputs.get("direction", "")
+            
+            # 获取自定义滑动距离，如果没有则使用默认值
+            scroll_amount = action_inputs.get("scroll_amount", None)
+            
+            if scroll_amount is not None:
+                try:
+                    scroll_amount = int(scroll_amount)
+                except (ValueError, TypeError):
+                    scroll_amount = None  # 转换失败使用默认值 
+                    
+            # 确定滚动值
+            if scroll_amount is not None:
+                # 使用自定义滑动距离
+                if "up" in direction.lower():
+                    scroll_value = abs(scroll_amount)   # 向上滚动：正值
+                elif "down" in direction.lower():
+                    scroll_value = -abs(scroll_amount)  # 向下滚动：负值
+                else:
+                    scroll_value = -abs(scroll_amount)  # 默认向下
+            else:
+                # 使用默认滑动距离
+                if "up" in direction.lower():
+                    scroll_value = 5    # 向上滚动：正值
+                elif "down" in direction.lower():
+                    scroll_value = -5   # 向下滚动：负值
+                else:
+                    scroll_value = -5   # 默认向下
+            
+            # 生成滚动代码
+            if x is None:
+                pyautogui_code += f"\npyautogui.scroll({scroll_value})"
+            else:
+                pyautogui_code += f"\npyautogui.scroll({scroll_value}, x={x}, y={y})"
+
+        elif action_type in ["click", "left_single", "left_double", "right_single", "hover"]:
+            # Parsing mouse click actions
+            start_box = action_inputs.get("start_box")
+            start_box = str(start_box)
+            if start_box:
+                start_box = eval(start_box)
+                if len(start_box) == 4:
+                    x1, y1, x2, y2 = start_box  # Assuming box is in [x1, y1, x2, y2]
+                elif len(start_box) == 2:
+                    x1, y1 = start_box
+                    x2 = x1
+                    y2 = y1
+                x = round(float((x1 + x2) / 2) * image_width, 3)
+                y = round(float((y1 + y2) / 2) * image_height, 3)
+                if action_type == "left_single" or action_type == "click":
+                    pyautogui_code += f"\npyautogui.click({x}, {y}, button='left')"
+                elif action_type == "left_double":
+                    pyautogui_code += f"\npyautogui.doubleClick({x}, {y}, button='left')"
+                elif action_type == "right_single":
+                    pyautogui_code += f"\npyautogui.click({x}, {y}, button='right')"
+                elif action_type == "hover":
+                    pyautogui_code += f"\npyautogui.moveTo({x}, {y})"
+        
+        elif action_type in ["finished"]:
+            pyautogui_code = f"DONE"
+        
+        else:
+            pyautogui_code += f"\n# Unrecognized action type: {action_type}"
+
+    return pyautogui_code
+
+def add_box_token(input_string):
+    # Step 1: Split the string into individual actions
+    if "Action: " in input_string and "start_box=" in input_string:
+        suffix = input_string.split("Action: ")[0] + "Action: "
+        actions = input_string.split("Action: ")[1:]
+        processed_actions = []
+        for action in actions:
+            action = action.strip()
+            # Step 2: Extract coordinates (start_box or end_box) using regex
+            coordinates = re.findall(r"(start_box|end_box)='\((\d+),\s*(\d+)\)'", action)
+            
+            updated_action = action  # Start with the original action
+            for coord_type, x, y in coordinates:
+                # Convert x and y to integers
+                updated_action = updated_action.replace(f"{coord_type}='({x},{y})'", f"{coord_type}='<|box_start|>({x},{y})<|box_end|>'")
+            processed_actions.append(updated_action)
+        
+        # Step 5: Reconstruct the final string
+        final_string = suffix + "\n\n".join(processed_actions)
+    else:
+        final_string = input_string
+    return final_string
+
+def pil_to_base64(image):
+    buffer = BytesIO()
+    image.save(buffer, format="PNG")  # 你可以改成 "JPEG" 等格式
+    return base64.b64encode(buffer.getvalue()).decode("utf-8")
+
+def linearize_accessibility_tree(accessibility_tree, platform="ubuntu"):
+
+    if platform == "ubuntu":
+        _attributes_ns = attributes_ns_ubuntu
+        _state_ns = state_ns_ubuntu
+        _component_ns = component_ns_ubuntu
+        _value_ns = value_ns_ubuntu
+    elif platform == "windows":
+        _attributes_ns = attributes_ns_windows
+        _state_ns = state_ns_windows
+        _component_ns = component_ns_windows
+        _value_ns = value_ns_windows
+    else:
+        raise ValueError("Invalid platform, must be 'ubuntu' or 'windows'")
+
+    filtered_nodes = filter_nodes(ET.fromstring(accessibility_tree), platform)
+    linearized_accessibility_tree = [
+        "tag\tname\ttext\tclass\tdescription\tposition (top-left x&y)\tsize (w&h)"
+    ]
+
+    # Linearize the accessibility tree nodes into a table format
+    for node in filtered_nodes:
+        if node.text:
+            text = (
+                node.text
+                if '"' not in node.text
+                else '"{:}"'.format(node.text.replace('"', '""'))
+            )
+
+        elif node.get("{{{:}}}class".format(class_ns_windows), "").endswith(
+            "EditWrapper"
+        ) and node.get("{{{:}}}value".format(_value_ns)):
+            node_text = node.get("{{{:}}}value".format(_value_ns), "")
+            text = (
+                node_text
+                if '"' not in node_text
+                else '"{:}"'.format(node_text.replace('"', '""'))
+            )
+        else:
+            text = '""'
+
+        linearized_accessibility_tree.append(
+            "{:}\t{:}\t{:}\t{:}\t{:}\t{:}\t{:}".format(
+                node.tag,
+                node.get("name", ""),
+                text,
+                (
+                    node.get("{{{:}}}class".format(_attributes_ns), "")
+                    if platform == "ubuntu"
+                    else node.get("{{{:}}}class".format(class_ns_windows), "")
+                ),
+                node.get("{{{:}}}description".format(_attributes_ns), ""),
+                node.get("{{{:}}}screencoord".format(_component_ns), ""),
+                node.get("{{{:}}}size".format(_component_ns), ""),
+            )
+        )
+
+    return "\n".join(linearized_accessibility_tree)
+
+def trim_accessibility_tree(linearized_accessibility_tree, max_tokens):
+    # enc = tiktoken.encoding_for_model("gpt-4")
+    # tokens = enc.encode(linearized_accessibility_tree)
+    # if len(tokens) > max_tokens:
+    #     linearized_accessibility_tree = enc.decode(tokens[:max_tokens])
+    #     linearized_accessibility_tree += "[...]\n"
+    return linearized_accessibility_tree
+
+def extract_action_desp(raw_response: str) -> str:
+    """
+    从原始响应中提取 "Action desp:" 及其描述部分，不包括后续的Action部分。
+    如果未找到开始或结束标记，则抛出 ValueError。
+    """
+    start_marker = "Action desp:"
+    start_index = raw_response.find(start_marker)
+    
+    # 如果没找到开始标记，直接抛出错误
+    if start_index == -1:
+        raise ValueError("未在响应中找到 'Action desp:' 标记。")
+    
+    # 找到下一个"\nAction:"的位置
+    end_marker = "\nAction:"
+    end_index = raw_response.find(end_marker, start_index)
+    
+    # 如果没有找到结束标记，也抛出错误
+    if end_index == -1:
+        raise ValueError("未在响应中找到 'Action desp:' 对应的结束标记 '\\nAction:'。")
+    
+    # 返回从开始标记到结束标记之前的内容
+    return raw_response[start_index:end_index]
+
+def print_content_without_long_base64(content):
+    """Print content while truncating long base64 image data"""
+    import copy
+    printable_content = copy.deepcopy(content)
+    
+    for item in printable_content:
+        if item.get("type") == "image_url" and "image_url" in item:
+            url = item["image_url"]["url"]
+            if url.startswith("data:image"):
+                # Replace base64 data with a short indication
+                prefix = url[:30]  # Keep the beginning
+                item["image_url"]["url"] = f"{prefix}...[BASE64_DATA_TRUNCATED]"
+    
+    print(printable_content)
+
+class ManoAgent:
+    def __init__(
+        self,
+        platform="ubuntu",
+        action_space="pyautogui",
+        observation_type="screenshot",
+        # observation_type can be in ["screenshot", "a11y_tree", "screenshot_a11y_tree", "som"]
+        max_trajectory_length=100,
+        a11y_tree_max_tokens=10000,
+        model="mano",
+        model_type="qwen25vl",
+        runtime_conf: dict = {
+            "infer_mode": "qwen25vl_normal",
+            "prompt_style": "qwen25vl_normal",
+            "input_swap": True,
+            "language": "English",
+            "history_n": 3,
+            "max_pixels": 16384*28*28,
+            "min_pixels": 100*28*28,
+            "callusr_tolerance": 100,
+            "temperature": 0.0,
+            "top_k": -1,
+            "top_p": 0.9,
+            "max_tokens": 500
+
+        }
+    ):
+        self.platform = platform
+        self.action_space = action_space
+        self.observation_type = observation_type
+        self.max_trajectory_length = max_trajectory_length
+        self.a11y_tree_max_tokens = a11y_tree_max_tokens
+        self.model = model
+        self.model_type = model_type
+        self.runtime_conf = runtime_conf
+        self.vlm = OpenAI(
+            base_url=os.environ['MANO_API_URL'],
+            api_key=os.environ['MANO_API_KEY'],
+        )
+        self.temperature = self.runtime_conf["temperature"]
+        self.top_k = self.runtime_conf["top_k"]
+        self.top_p = self.runtime_conf["top_p"]
+        self.max_tokens = self.runtime_conf["max_tokens"]
+        self.infer_mode = self.runtime_conf["infer_mode"]
+        self.prompt_style = self.runtime_conf["prompt_style"]
+        self.input_swap = self.runtime_conf["input_swap"]
+        self.language = self.runtime_conf["language"]
+        self.max_pixels = self.runtime_conf["max_pixels"]
+        self.min_pixels = self.runtime_conf["min_pixels"]
+        self.callusr_tolerance = self.runtime_conf["callusr_tolerance"]
+
+        self.thoughts = []
+        self.actions = []
+        self.observations = []
+        self.history_images = []
+        self.history_responses = []  
+        self.action_parse_res_factor = 1000
+        self.prompt_template = MANO_PROMPT_THOUGHT
+        
+        if "history_n" in self.runtime_conf:
+            self.history_n = self.runtime_conf["history_n"]
+        else:
+            self.history_n = 5
+        
+        self.cur_callusr_count = 0
+
+    def predict(
+        self, instruction: str, obs: Dict, last_action_after_obs: Dict = None
+    ) -> List:
+        """
+        Predict the next action(s) based on the current observation.
+        """
+
+        # Append trajectory
+        # print(len(self.observations), len(self.actions), len(self.actions))
+        assert len(self.observations) == len(self.actions) and len(self.actions) == len(
+            self.thoughts
+        ), "The number of observations and actions should be the same."
+
+        if len(self.observations) > self.max_trajectory_length:
+            if self.max_trajectory_length == 0:
+                _observations = []
+                _actions = []
+                _thoughts = []
+            else:
+                _observations = self.observations[-self.max_trajectory_length :]
+                _actions = self.actions[-self.max_trajectory_length :]
+                _thoughts = self.thoughts[-self.max_trajectory_length :]
+        else:
+            _observations = self.observations
+            _actions = self.actions
+            _thoughts = self.thoughts
+
+        for previous_obs, previous_action, previous_thought in zip(
+            _observations, _actions, _thoughts
+        ):
+            # {{{1
+            if self.observation_type == "screenshot_a11y_tree":
+                _screenshot = previous_obs["screenshot"]
+                _linearized_accessibility_tree = previous_obs["accessibility_tree"]
+            elif self.observation_type == "screenshot":
+                _screenshot = previous_obs["screenshot"]
+            else:
+                raise ValueError(
+                    "Invalid observation_type type: " + self.observation_type
+                )  # 1}}}
+
+        self.history_images.append(obs["screenshot"])
+
+        if self.observation_type in ["screenshot", "screenshot_a11y_tree"]:
+            base64_image = obs["screenshot"]
+            try:
+                linearized_accessibility_tree = (
+                    linearize_accessibility_tree(
+                        accessibility_tree=obs["accessibility_tree"],
+                        platform=self.platform,
+                    )
+                    if self.observation_type == "screenshot_a11y_tree"
+                    else None
+                )
+            except:
+                linearized_accessibility_tree = None
+            # logger.debug("LINEAR AT: %s", linearized_accessibility_tree)
+
+            if linearized_accessibility_tree:
+                linearized_accessibility_tree = trim_accessibility_tree(
+                    linearized_accessibility_tree, self.a11y_tree_max_tokens
+                )
+
+            if self.observation_type == "screenshot_a11y_tree":
+                self.observations.append(
+                    {
+                        "screenshot": base64_image,
+                        "accessibility_tree": linearized_accessibility_tree,
+                    }
+                )
+            else:
+                self.observations.append(
+                    {"screenshot": base64_image, "accessibility_tree": None}
+                )
+
+        else:
+            raise ValueError(
+                "Invalid observation_type type: " + self.observation_type
+            )  # 1}}}
+        
+        if self.infer_mode == "qwen25vl_normal":
+            user_prompt = self.prompt_template.format(
+                instruction=instruction,
+                language=self.language
+            )
+        # print(user_prompt)
+        # exit()
+        
+        if len(self.history_images) > self.history_n:
+            self.history_images = self.history_images[-self.history_n:]
+
+        messages, images = [], []
+        if isinstance(self.history_images, bytes):
+            self.history_images = [self.history_images]
+        elif isinstance(self.history_images, np.ndarray):
+            self.history_images = list(self.history_images)
+        elif isinstance(self.history_images, list):
+            pass
+        else:
+            raise TypeError(f"Unidentified images type: {type(self.history_images)}")
+
+        for turn, image in enumerate(self.history_images):
+            if len(images) >= self.history_n:
+                break
+            try:
+                image = Image.open(BytesIO(image))
+            except Exception as e:
+                raise RuntimeError(f"Error opening image: {e}")
+
+            if image.width * image.height > self.max_pixels:
+                """
+                如果图片超过/低于像素限制，则计算一个缩放因子resize_factor，使图片的像素数缩小到等于或小于max_pixels。这个缩放因子是通过开平方根计算的，确保纵横比保持不变,这样原始的相对坐标可以不经转换直接复用
+                """
+                resize_factor = math.sqrt(self.max_pixels / (image.width * image.height))
+                width, height = int(image.width * resize_factor), int(image.height * resize_factor)
+                image = image.resize((width, height))
+            if image.width * image.height < self.min_pixels:
+                resize_factor = math.sqrt(self.min_pixels / (image.width * image.height))
+                width, height = math.ceil(image.width * resize_factor), math.ceil(image.height * resize_factor)
+                image = image.resize((width, height))
+
+            if image.mode != "RGB":
+                image = image.convert("RGB")
+
+            images.append(image)
+
+        # 修改为单轮对话形式
+        messages = [
+            {
+                "role": "system",
+                "content": [{"type": "text", "text": "You are a helpful assistant."}]
+            }
+        ]
+        
+        # 构建用户消息内容
+        user_message_content = []
+        
+        user_message_content.append({
+            "type": "text", 
+            "text": user_prompt
+        })
+        #print(user_prompt)
+        
+        # 添加历史操作信息 - 包含所有历史步骤，不限制数量
+        if len(self.history_responses) > 0:
+            # 确定哪些步骤需要显示图片
+            steps_with_images = {}
+            
+            # 计算可用的历史图片数量（除去当前步骤的图片）
+            available_images = len(images) - 1
+            
+            # 为最近的步骤分配图片
+            for i in range(min(available_images, len(self.history_responses))):
+                # 最近的步骤优先获得图片
+                step_idx = len(self.history_responses) - 1 - i
+                img_idx = available_images - 1 - i
+                steps_with_images[step_idx] = img_idx
+            
+            # 构建历史步骤文本
+            for i, history_response in enumerate(self.history_responses):
+                step_num = i + 1
+                
+                # 尝试提取Action描述
+                try:
+                    action_desp = extract_action_desp(history_response)
+                except ValueError as e:
+                    print("获取Action desp失败")
+                    raise ValueError(f"Action extraction error: {str(e)}") from e
+                
+                # 添加步骤描述文本
+                if i > 0:  # 添加换行，除了第一步
+                    user_message_content.append({"type": "text", "text": "\n"})
+                
+                user_message_content.append({
+                    "type": "text", 
+                    "text": f"第{step_num}步：{action_desp}"
+                })
+                
+                if i in steps_with_images:
+                    img_idx = steps_with_images[i]
+                    encoded_string = pil_to_base64(images[img_idx])
+                    user_message_content.append({
+                    "type": "text", 
+                    "text": "对应截图为："
+                    })
+                    user_message_content.append({
+                        "type": "image_url", 
+                        "image_url": {"url": f"data:image/png;base64,{encoded_string}"}
+                    })
+        
+        # 添加当前步骤描述
+        user_message_content.append({
+            "type": "text", 
+            "text": "\n当前步骤的截图："
+        })
+        
+        # 添加当前步骤的图片
+        encoded_string = pil_to_base64(images[-1])
+        user_message_content.append({
+            "type": "image_url", 
+            "image_url": {"url": f"data:image/png;base64,{encoded_string}"}
+        })
+        
+        # 构建最终的用户消息
+        messages.append({
+            "role": "user",
+            "content": user_message_content
+        })
+
+        #print_content_without_long_base64(user_message_content)
+        
+        
+        try_times = 3
+        origin_resized_height = images[-1].height
+        origin_resized_width = images[-1].width
+        temperature = self.temperature
+        top_k = self.top_k
+        #print("[DEBUG] 模型调用中")
+        while True:
+            if try_times <= 0:
+                print(f"Reach max retry times to fetch response from client, as error flag.")
+                return "client error", ["DONE"], []
+            try:
+                # print("in try")
+                response = self.vlm.chat.completions.create(
+                    model="mano",
+                    messages=messages,
+                    #frequency_penalty=1,
+                    max_tokens=self.max_tokens,
+                    temperature=temperature
+                )
+                # print("[DEBUG] 模型调用成功，收到 response")
+                #print(response.choices[0].message.content)
+                prediction = response.choices[0].message.content.strip()
+                #prediction = response[0]["prediction"].strip()
+            except Exception as e:
+                print(f"Error when fetching response from client, with response: {response}")
+                prediction = None
+                try_times -= 1
+            
+            try:
+                parsed_responses = parse_action_to_structure_output(
+                    prediction,
+                    self.action_parse_res_factor,
+                    origin_resized_height,
+                    origin_resized_width,
+                    self.model_type,
+                    self.max_pixels,
+                    self.min_pixels
+                )
+                #print(f"[DEBUG] Step prediction: {prediction}")
+                #print(f"[DEBUG] Parsed responses: {parsed_responses}")
+                break
+            except Exception as e:
+                print(f"Error 111 when parsing response from client, with response: {response}")
+                # If fail to parse the model response, we use sampling parameters to avoid it
+                prediction = None
+                try_times -= 1
+                temperature = 1
+                top_k = -1
+                
+        if prediction is None:
+            return "client error", ["DONE"]
+
+        self.history_responses.append(prediction)
+        self.thoughts.append(prediction)
+
+        try:
+            parsed_responses = parse_action_to_structure_output(
+                prediction,
+                self.action_parse_res_factor,
+                origin_resized_height,
+                origin_resized_width,
+                self.model_type,
+                self.max_pixels,
+                self.min_pixels
+            )
+        except Exception as e:
+            print(f"Parsing action error: {prediction}, with error:\n{e}")
+            return f"Parsing action error: {prediction}, with error:\n{e}", ["DONE"]
+
+        actions = []
+        last_image = Image.open(BytesIO(self.history_images[-1]))
+        obs_image_height = last_image.height
+        obs_image_width = last_image.width
+        
+        for parsed_response in parsed_responses:
+            if "action_type" in parsed_response:
+
+                if parsed_response["action_type"] == FINISH_WORD:
+                    self.actions.append(actions)
+
+                    return prediction, ["DONE"]
+                
+                elif parsed_response["action_type"] == WAIT_WORD:
+                    self.actions.append(actions)
+                    return prediction, ["WAIT"]
+                
+                elif parsed_response["action_type"] == ENV_FAIL_WORD:
+                    self.actions.append(actions)
+                    return prediction, ["FAIL"]
+
+                elif parsed_response["action_type"] == CALL_USER:
+                    if self.callusr_tolerance > self.cur_callusr_count:
+                        self.actions.append(actions)
+                        self.cur_callusr_count += 1
+                        return prediction, ["WAIT"]
+                    else:
+                        self.actions.append(actions)
+                        return prediction, ["FAIL"]
+
+            pyautogui_code = parsing_response_to_pyautogui_code(
+                parsed_response,
+                obs_image_height,
+                obs_image_width,
+                self.input_swap
+            )
+            actions.append(pyautogui_code)
+
+        self.actions.append(actions)
+
+        if len(self.history_responses) >= self.max_trajectory_length:
+            # Default to FAIL if exceed max steps
+            actions = ["FAIL"]
+
+        return prediction, actions
+
+
+    @backoff.on_exception(
+        backoff.constant,
+        # here you should add more model exceptions as you want,
+        # but you are forbidden to add "Exception", that is, a common type of exception
+        # because we want to catch this kind of Exception in the outside to ensure each example won't exceed the time limit
+        (
+            # General exceptions
+            SSLError,
+            # OpenAI exceptions
+            openai.RateLimitError,
+            openai.BadRequestError,
+            openai.InternalServerError,
+            # Google exceptions
+            InvalidArgument,
+            ResourceExhausted,
+            InternalServerError,
+            BadRequest,
+            # Groq exceptions
+            # todo: check
+        ),
+        interval=30,
+        max_tries=10,
+    )
+    
+    def reset(self, runtime_logger):
+        self.thoughts = []
+        self.actions = []
+        self.observations = []
+        self.history_images = []
+        self.history_responses = []
diff --git a/run_multienv_mano.py b/run_multienv_mano.py
new file mode 100644
index 0000000..3006efc
--- /dev/null
+++ b/run_multienv_mano.py
@@ -0,0 +1,568 @@
+from __future__ import annotations
+import argparse
+import datetime
+import json
+import logging
+import os
+import sys
+import signal
+import time
+from typing import List
+from multiprocessing import Process, Manager
+from multiprocessing import current_process
+import lib_run_single
+from desktop_env.desktop_env import DesktopEnv
+from mm_agents.mano_agent import ManoAgent
+import os
+
+# Global variables for signal handling
+active_environments = []
+processes = []
+is_terminating = False
+
+# load the environment variables from .env file
+if os.path.exists(".env"):
+    from dotenv import load_dotenv
+    load_dotenv()
+
+#  Logger Configs {{{ #
+def config() -> argparse.Namespace:
+    parser = argparse.ArgumentParser(
+        description="Run end-to-end evaluation on the benchmark"
+    )
+
+    # environment config
+    parser.add_argument("--path_to_vm", type=str, default=None)
+    parser.add_argument(
+        "--headless", action="store_true", help="Run in headless machine"
+    )
+    parser.add_argument(
+        "--action_space", type=str, default="pyautogui", help="Action type"
+    )
+    parser.add_argument(
+        "--observation_type",
+        choices=["screenshot", "a11y_tree", "screenshot_a11y_tree", "som"],
+        default="screenshot",
+        help="Observation type",
+    )
+    parser.add_argument("--sleep_after_execution", type=float, default=3.0)
+    parser.add_argument("--max_steps", type=int, default=15)
+    
+    # evaluation config
+    parser.add_argument(
+        "--test_config_base_dir", type=str, default="evaluation_examples"
+    )
+
+    # lm config
+    parser.add_argument("--model", type=str, default="mano")
+    parser.add_argument("--model_type", type=str, default="qwen25vl")
+    parser.add_argument("--infer_mode", type=str, default="qwen25vl_normal")
+    parser.add_argument("--prompt_style", type=str, default="qwen25vl_normal")
+    parser.add_argument("--input_swap", action="store_true", help="Use copy and paste to type content")
+    parser.add_argument("--language", type=str, default="Chinese")
+    parser.add_argument("--max_pixels", type=float, default=16384*28*28)
+    parser.add_argument("--min_pixels", type=float, default=100*28*28)
+    parser.add_argument("--temperature", type=float, default=1.0)
+    parser.add_argument("--top_p", type=float, default=0.9)
+    parser.add_argument("--top_k", type=int, default=-1)
+    parser.add_argument("--history_n", type=int, default=5)
+    parser.add_argument("--callusr_tolerance", type=int, default=3)
+    parser.add_argument("--max_tokens", type=int, default=1000)
+    parser.add_argument("--stop_token", type=str, default=None)
+
+    parser.add_argument("--max_trajectory_length", type=int, default=None, help="The max number of trajectory steps.")
+    parser.add_argument("--max_image_history_length", type=int, default=5, help="The max number of images in the history.")
+
+    # example config
+    parser.add_argument("--domain", type=str, default="all")
+    parser.add_argument(
+        "--test_all_meta_path", type=str, default="evaluation_examples/test_all.json"
+    )
+
+    # logging related
+    parser.add_argument("--result_dir", type=str, default="./results")
+    parser.add_argument("--num_envs", type=int, default=1, help="Number of environments to run in parallel")  
+    parser.add_argument("--log_level", type=str, choices=['DEBUG', 'INFO', 'WARNING', 'ERROR', 'CRITICAL'], 
+                       default='INFO', help="Set the logging level")
+    # aws config
+    parser.add_argument(
+        "--region", type=str, default="us-east-1", help="AWS region for the VM"
+    )
+    parser.add_argument(
+        "--provider_name", type=str, default="aws", choices=["aws", "virtualbox", "vmware", "docker", "azure"], help="Provider name"
+    )
+    parser.add_argument(
+        "--client_password", type=str, default="", help="Client password"
+    )
+    parser.add_argument(
+        "--screen_width", type=int, default=1920, help="Screen width"
+    )
+    parser.add_argument(
+        "--screen_height", type=int, default=1080, help="Screen height"
+    )
+    args = parser.parse_args()
+    return args
+
+args = config()  # Get command line arguments first
+
+logger = logging.getLogger()
+log_level = getattr(logging, args.log_level.upper())
+logger.setLevel(log_level)
+
+datetime_str: str = datetime.datetime.now().strftime("%Y%m%d@%H%M%S")
+
+file_handler = logging.FileHandler(
+    os.path.join("logs", "normal-{:}.log".format(datetime_str)), encoding="utf-8"
+)
+debug_handler = logging.FileHandler(
+    os.path.join("logs", "debug-{:}.log".format(datetime_str)), encoding="utf-8"
+)
+stdout_handler = logging.StreamHandler(sys.stdout)
+
+file_handler.setLevel(logging.INFO)
+debug_handler.setLevel(logging.DEBUG)
+stdout_handler.setLevel(log_level)
+
+formatter = logging.Formatter(
+    fmt="\x1b[1;33m[%(asctime)s \x1b[31m%(levelname)s \x1b[32m%(module)s/%(lineno)d-%(processName)s\x1b[1;33m] \x1b[0m%(message)s"
+)
+file_handler.setFormatter(formatter)
+debug_handler.setFormatter(formatter)
+stdout_handler.setFormatter(formatter)
+
+stdout_handler.addFilter(logging.Filter("desktopenv"))
+
+logger.addHandler(file_handler)
+logger.addHandler(debug_handler)
+logger.addHandler(stdout_handler)
+#  }}} Logger Configs #
+
+logger = logging.getLogger("desktopenv.experiment")
+
+
+def distribute_tasks(test_all_meta: dict) -> List[tuple]:
+    all_tasks = []
+    for domain, examples in test_all_meta.items():
+        for example_id in examples:
+            all_tasks.append((domain, example_id))
+    return all_tasks
+
+
+def process_signal_handler(signum, frame, env_idx):
+    """Signal handler for child processes to gracefully shut down their environments."""
+    logger.info(f"Process {env_idx + 1} received signal {signum}. Shutting down...")
+    
+    # Get the active_environments from the caller's frame
+    local_vars = frame.f_locals
+    active_environments = local_vars.get('active_environments', [])
+    
+    # Close environment in the current process context
+    for env in active_environments:
+        if env is not None:
+            try:
+                logger.info(f"Process {env_idx + 1} closing environment...")
+                env.close()
+                logger.info(f"Process {env_idx + 1} environment closed successfully")
+            except Exception as e:
+                logger.error(f"Process {env_idx + 1} error closing environment: {e}")
+    
+    logger.info(f"Process {env_idx + 1} shutdown complete. Exiting.")
+    sys.exit(0)
+
+def run_env_tasks(task_queue: Queue, args: argparse.Namespace, shared_scores: list):
+    active_environments = []
+    env = None
+    try:
+        #from desktop_env.providers.aws.manager import IMAGE_ID_MAP
+        #REGION = args.region
+        screen_size = (args.screen_width, args.screen_height)
+        #ami_id = IMAGE_ID_MAP[REGION].get(screen_size, IMAGE_ID_MAP[REGION][(1920, 1080)])
+        env = DesktopEnv(
+            path_to_vm=args.path_to_vm,
+            action_space=args.action_space,
+            provider_name=args.provider_name,
+            #region=REGION,
+            #snapshot_name=ami_id,
+            screen_size=screen_size,
+            headless=args.headless,
+            os_type="Ubuntu",
+            require_a11y_tree=args.observation_type in ["a11y_tree", "screenshot_a11y_tree", "som"],
+            #enable_proxy=True,
+            #client_password=args.client_password
+        )
+
+
+        active_environments.append(env)
+        args.max_trajectory_length = args.max_steps
+        if args.infer_mode == "qwen25vl_normal":
+            runtime_conf: dict = {
+                "infer_mode": "qwen25vl_normal",
+                "prompt_style": "qwen25vl_normal",
+                "input_swap": False, #True,
+                "language": "Chinese",
+                "history_n": 3,
+                "max_pixels": 16384*28*28,
+                "min_pixels": 100*28*28,
+                "callusr_tolerance": 3,
+                "temperature": 0.0,
+                "top_k": -1,
+                "top_p": 0.9,
+                "max_tokens": 1000
+
+            }
+        else:
+            raise ValueError(f"Unknown infer_mode: {args.infer_mode}")
+        
+        agent = ManoAgent(
+            model=args.model,
+            action_space=args.action_space,
+            observation_type=args.observation_type,
+            max_trajectory_length=args.max_trajectory_length,
+            model_type=args.model_type,
+            runtime_conf = runtime_conf
+        )
+
+        logger.info(f"Process {current_process().name} started.")
+        while True:
+            try:
+                item = task_queue.get(timeout=5)
+            except Exception:
+                break
+            domain, example_id = item
+            try:
+                config_file = os.path.join(
+                    args.test_config_base_dir, f"examples/{domain}/{example_id}.json"
+                )
+                with open(config_file, "r", encoding="utf-8") as f:
+                    example = json.load(f)
+                logger.info(f"[{current_process().name}][Domain]: {domain}")
+                logger.info(f"[{current_process().name}][Example ID]: {example_id}")
+                logger.info(f"[{current_process().name}][Instruction]: {example['instruction']}")
+                example_result_dir = os.path.join(
+                    args.result_dir,
+                    args.action_space,
+                    args.observation_type,
+                    args.model,
+                    domain,
+                    example_id,
+                )
+                os.makedirs(example_result_dir, exist_ok=True)
+                try:
+                    lib_run_single.run_single_example_mano(
+                        agent,
+                        env,
+                        example,
+                        args.max_steps,
+                        example["instruction"],
+                        args,
+                        example_result_dir,
+                        shared_scores,
+                    )
+                except Exception as e:
+                    import traceback
+                    logger.error(f"Exception in {current_process().name} {domain}/{example_id}: {e}")
+                    logger.error(traceback.format_exc())
+                    try:
+                        env.controller.end_recording(
+                            os.path.join(example_result_dir, "recording.mp4")
+                        )
+                    except Exception as rec_e:
+                        logger.error(f"Failed to end recording: {rec_e}")
+                    with open(os.path.join(example_result_dir, "traj.jsonl"), "a") as f:
+                        f.write(
+                            json.dumps(
+                                {"Error": f"{domain}/{example_id} - {e}"}
+                            )
+                        )
+                        f.write("\n")
+            except Exception as e:
+                logger.error(f"Task-level error in {current_process().name}: {e}")
+                import traceback
+                logger.error(traceback.format_exc())
+    except Exception as e:
+        logger.error(f"Process-level error in {current_process().name}: {e}")
+        import traceback
+        logger.error(traceback.format_exc())
+    finally:
+        logger.info(f"{current_process().name} cleaning up environment...")
+        try:
+            if env:
+                env.close()
+                logger.info(f"{current_process().name} environment closed successfully")
+        except Exception as e:
+            logger.error(f"{current_process().name} error during environment cleanup: {e}")
+
+
+
+def signal_handler(signum, frame):
+    """Handle termination signals (SIGINT, SIGTERM) to gracefully shutdown environments."""
+    global is_terminating, active_environments, processes
+    
+    # Avoid duplicate handling
+    if is_terminating:
+        return
+    
+    is_terminating = True
+    logger.info(f"Received signal {signum}. Gracefully shutting down...")
+    
+    # Close all registered environments in the main process
+    for env in active_environments:
+        try:
+            logger.info(f"Closing environment...")
+            env.close()
+            logger.info(f"Environment closed successfully")
+        except Exception as e:
+            logger.error(f"Error closing environment: {e}")
+    
+    # Send termination signal to all child processes first
+    for p in processes:
+        if p.is_alive():
+            try:
+                logger.info(f"Sending termination signal to process {p.name}...")
+                p.terminate()
+            except Exception as e:
+                logger.error(f"Error sending termination signal to process: {e}")
+    
+    # Allow a short time for processes to handle their own cleanup
+    time.sleep(1)
+    
+    # Forcefully terminate any processes that didn't exit
+    for p in processes:
+        if p.is_alive():
+            try:
+                logger.info(f"Forcefully terminating process {p.name}...")
+                import signal as sig
+                os.kill(p.pid, sig.SIGKILL)
+            except Exception as e:
+                logger.error(f"Error forcefully terminating process: {e}")
+    
+    logger.info("Shutdown complete. Exiting.")
+    sys.exit(0)
+
+
+def test(args: argparse.Namespace, test_all_meta: dict) -> None:
+    global processes
+    logger.info("Args: %s", args)
+    all_tasks = distribute_tasks(test_all_meta)
+    logger.info(f"Total tasks: {len(all_tasks)}")
+    with Manager() as manager:
+        shared_scores = manager.list()
+        task_queue = manager.Queue()
+        for item in all_tasks:
+            task_queue.put(item)
+        num_envs = args.num_envs
+        processes = []
+        for i in range(num_envs):
+            p = Process(
+                target=run_env_tasks,
+                args=(task_queue, args, shared_scores),
+                name=f"EnvProcess-{i+1}"
+            )
+            p.daemon = True
+            p.start()
+            processes.append(p)
+            logger.info(f"Started process {p.name} with PID {p.pid}")
+        try:
+            while True:
+                alive_count = 0
+                for idx, p in enumerate(processes):
+                    if not p.is_alive():
+                        logger.warning(f"Process {p.name} died, restarting...")
+                        new_p = Process(
+                            target=run_env_tasks,
+                            args=(task_queue, args, shared_scores),
+                            name=f"EnvProcess-Restart-{idx+1}"
+                        )
+                        new_p.daemon = True
+                        new_p.start()
+                        processes[idx] = new_p
+                        logger.info(f"Restarted process {new_p.name} with PID {new_p.pid}")
+                    else:
+                        alive_count += 1
+                if task_queue.empty():
+                    logger.info("All tasks finished.")
+                    break
+                if alive_count == 0:
+                    logger.error("All processes died, exiting.")
+                    break
+                time.sleep(5)
+            for p in processes:
+                p.join()
+        except KeyboardInterrupt:
+            logger.info("Main process received KeyboardInterrupt. Initiating graceful shutdown...")
+            raise
+        except Exception as e:
+            logger.error(f"Unexpected error while waiting for processes: {e}", exc_info=True)
+            for p in processes:
+                if p.is_alive():
+                    try:
+                        logger.info(f"Terminating process {p.name} due to error...")
+                        p.terminate()
+                    except Exception as term_e:
+                        logger.error(f"Error terminating process {p.name}: {term_e}")
+            raise
+        scores = list(shared_scores)
+    logger.info(f"Average score: {sum(scores) / len(scores) if scores else 0}")
+
+
+def get_unfinished(
+    action_space, use_model, observation_type, result_dir, total_file_json
+):
+    target_dir = os.path.join(result_dir, action_space, observation_type, use_model)
+
+    if not os.path.exists(target_dir):
+        return total_file_json
+
+    finished = {}
+    for domain in os.listdir(target_dir):
+        finished[domain] = []
+        domain_path = os.path.join(target_dir, domain)
+        if os.path.isdir(domain_path):
+            for example_id in os.listdir(domain_path):
+                if example_id == "onboard":
+                    continue
+                example_path = os.path.join(domain_path, example_id)
+                if os.path.isdir(example_path):
+                    if "result.txt" not in os.listdir(example_path):
+                        # empty all files under example_id
+                        for file in os.listdir(example_path):
+                            os.remove(os.path.join(example_path, file))
+                    else:
+                        finished[domain].append(example_id)
+
+    if not finished:
+        return total_file_json
+
+    for domain, examples in finished.items():
+        if domain in total_file_json:
+            total_file_json[domain] = [
+                x for x in total_file_json[domain] if x not in examples
+            ]
+
+    return total_file_json
+
+
+def get_result(action_space, use_model, observation_type, result_dir, total_file_json):
+    target_dir = os.path.join(result_dir, action_space, observation_type, use_model)
+    if not os.path.exists(target_dir):
+        print("New experiment, no result yet.")
+        return None
+
+    all_result = []
+
+    for domain in os.listdir(target_dir):
+        domain_path = os.path.join(target_dir, domain)
+        if os.path.isdir(domain_path):
+            for example_id in os.listdir(domain_path):
+                example_path = os.path.join(domain_path, example_id)
+                if os.path.isdir(example_path):
+                    if "result.txt" in os.listdir(example_path):
+                        # empty all files under example_id
+                        try:
+                            all_result.append(
+                                float(
+                                    open(
+                                        os.path.join(example_path, "result.txt"), "r"
+                                    ).read()
+                                )
+                            )
+                        except:
+                            all_result.append(0.0)
+
+    if not all_result:
+        print("New experiment, no result yet.")
+        return None
+    else:
+        print("Current Success Rate:", sum(all_result) / len(all_result) * 100, "%")
+        return all_result
+
+
+if __name__ == "__main__":
+    ####### The complete version of the list of examples #######
+    os.environ["TOKENIZERS_PARALLELISM"] = "false"
+    
+    # Register signal handlers for graceful termination
+    signal.signal(signal.SIGINT, signal_handler)  # Handle Ctrl+C
+    signal.signal(signal.SIGTERM, signal_handler)  # Handle termination signal
+    
+    try:
+        args = config()
+        
+        # save args to json in result_dir/action_space/observation_type/model/args.json
+        path_to_args = os.path.join(
+            args.result_dir,
+            args.action_space,
+            args.observation_type,
+            args.model,
+            "args.json",
+        )
+        os.makedirs(os.path.dirname(path_to_args), exist_ok=True)
+        with open(path_to_args, "w", encoding="utf-8") as f:
+            json.dump(vars(args), f, indent=4)
+
+        with open(args.test_all_meta_path, "r", encoding="utf-8") as f:
+            test_all_meta = json.load(f)
+
+        if args.domain != "all":
+            test_all_meta = {args.domain: test_all_meta[args.domain]}
+
+        test_file_list = get_unfinished(
+            args.action_space,
+            args.model,
+            args.observation_type,
+            args.result_dir,
+            test_all_meta,
+        )
+        left_info = ""
+        for domain in test_file_list:
+            left_info += f"{domain}: {len(test_file_list[domain])}\n"
+        logger.info(f"Left tasks:\n{left_info}")
+
+        get_result(
+            args.action_space,
+            args.model,
+            args.observation_type,
+            args.result_dir,
+            test_all_meta,
+        )
+        test(args, test_file_list)
+    except KeyboardInterrupt:
+        logger.info("Main process received KeyboardInterrupt.")
+        # Signal handler will take care of cleanup
+    except Exception as e:
+        logger.error(f"Unexpected error in main process: {e}", exc_info=True)
+        # Also trigger cleanup for unhandled exceptions
+        signal_handler(signal.SIGTERM, None)
+    finally:
+        # Final cleanup in case any environments or processes remain
+        logger.info("Main process final cleanup...")
+        for env in active_environments:
+            if env is not None:
+                try:
+                    logger.info(f"Closing environment in final cleanup...")
+                    env.close()
+                    logger.info(f"Environment closed successfully in final cleanup")
+                except Exception as e:
+                    logger.error(f"Error during final environment cleanup: {e}")
+        
+        # First try gentle termination
+        for p in processes:
+            if p is not None and p.is_alive():
+                try:
+                    logger.info(f"Terminating process {p.name}...")
+                    p.terminate()
+                except Exception as e:
+                    logger.error(f"Error terminating process: {e}")
+        
+        # Wait a moment for processes to terminate
+        time.sleep(1)
+        
+        # Then force kill if needed
+        for p in processes:
+            if p is not None and p.is_alive():
+                try:
+                    logger.info(f"Force killing process {p.name}...")
+                    os.kill(p.pid, signal.SIGKILL)
+                    logger.info(f"Process {p.name} force killed")
+                except Exception as e:
+                    logger.error(f"Error force killing process: {e}")