beswarm 0.2.34__py3-none-any.whl → 0.2.36__py3-none-any.whl

beswarm/tools/worker.py

@@ -4,9 +4,186 @@ import sys
 import copy
 import json
 import difflib
+import asyncio
 import platform
 from pathlib import Path
 from datetime import datetime
+from typing import List, Dict, Union
+
+from ..broker import MessageBroker
+from ..aient.src.aient.models import chatgpt
+from ..aient.src.aient.plugins import register_tool, get_function_call_list, registry
+from ..prompt import worker_system_prompt, instruction_system_prompt
+from ..utils import extract_xml_content, get_current_screen_image_message, replace_xml_content, register_mcp_tools
+from ..bemcp.bemcp import MCPManager
+
+class BaseAgent:
+    """Base class for agents, handling common initialization and disposal."""
+    def __init__(self, goal: str, tools_json: List, agent_config: Dict, work_dir: str, cache_messages: Union[bool, List[Dict]], broker: MessageBroker, listen_topic: str, publish_topic: str, status_topic: str):
+        self.goal = goal
+        self.tools_json = tools_json
+        self.work_dir = work_dir
+        self.cache_file = Path(work_dir) / ".beswarm" / "work_agent_conversation_history.json"
+        self.config = agent_config
+        self.cache_messages = cache_messages
+        if cache_messages and isinstance(cache_messages, bool) and cache_messages == True:
+            self.cache_messages = json.loads(self.cache_file.read_text(encoding="utf-8"))
+        self.broker = broker
+        self.listen_topic = listen_topic
+        self.error_topic = listen_topic + ".error"
+        self.publish_topic = publish_topic
+        self.status_topic = status_topic
+        self._subscription = self.broker.subscribe(self.handle_message, [self.listen_topic, self.error_topic])
+
+    async def handle_message(self, message: Dict):
+        """Process incoming messages. Must be implemented by subclasses."""
+        raise NotImplementedError
+
+    def dispose(self):
+        """Cancels the subscription and cleans up resources."""
+        if self._subscription:
+            self._subscription.dispose()
+
+
+class InstructionAgent(BaseAgent):
+    """Generates instructions and publishes them to a message broker."""
+    def __init__(self, goal: str, tools_json: List, agent_config: Dict, work_dir: str, cache_messages: Union[bool, List[Dict]], broker: MessageBroker, listen_topic: str, publish_topic: str, status_topic: str):
+        super().__init__(goal, tools_json, agent_config, work_dir, cache_messages, broker, listen_topic, publish_topic, status_topic)
+
+        self.last_instruction = None
+        self.agent = chatgpt(**self.config)
+
+        self.goal_diff = None
+
+        if self.cache_messages and isinstance(self.cache_messages, list) and len(self.cache_messages) > 1:
+            old_goal = extract_xml_content(self.cache_messages[1]["content"], "goal")
+            if old_goal.strip() != goal.strip():
+                diff_generator = difflib.ndiff(old_goal.splitlines(), goal.splitlines())
+                changed_lines = []
+                for line in diff_generator:
+                    if (line.startswith('+ ') or line.startswith('- ')) and line[2:].strip():
+                        changed_lines.append(line)
+                self.goal_diff = '\n'.join(changed_lines).strip()
+
+    def get_conversation_history(self, conversation_history: List[Dict]):
+        conversation_history = copy.deepcopy(conversation_history)
+
+        self.cache_file.write_text(json.dumps(conversation_history, ensure_ascii=False, indent=4), encoding="utf-8")
+
+        work_agent_system_prompt = conversation_history.pop(0)
+        if conversation_history:
+            original_content = work_agent_system_prompt["content"]
+            regex = r"<latest_file_content>(.*?)</latest_file_content>"
+            match = re.search(regex, original_content, re.DOTALL)
+            if match:
+                extracted_content = f"<latest_file_content>{match.group(1)}</latest_file_content>\n\n"
+            else:
+                extracted_content = ""
+            if isinstance(conversation_history[0]["content"], str):
+                conversation_history[0]["content"] = extracted_content + conversation_history[0]["content"]
+            elif isinstance(conversation_history[0]["content"], list) and extracted_content:
+                conversation_history[0]["content"].append({"type": "text", "text": extracted_content})
+
+        return conversation_history
+
+    async def handle_message(self, message: Dict):
+        """Receives a worker response, generates the next instruction, and publishes it."""
+
+        if len(message["conversation"]) > 1 and message["conversation"][-2]["role"] == "user" \
+        and "<task_complete_message>" in message["conversation"][-2]["content"]:
+            task_complete_message = extract_xml_content(message["conversation"][-2]["content"], "task_complete_message")
+            self.broker.publish({"status": "finished", "result": task_complete_message}, self.status_topic)
+            return
+
+        instruction_prompt = "".join([
+                "</work_agent_conversation_end>\n\n",
+                f"任务目标: {self.goal}\n\n",
+                f"任务目标新变化：\n{self.goal_diff}\n\n" if self.goal_diff else "",
+                "在 tag <work_agent_conversation_start>...</work_agent_conversation_end> 之前的对话历史都是工作智能体的对话历史。\n\n",
+                "根据以上对话历史和目标，请生成下一步指令。如果任务已完成，指示工作智能体调用task_complete工具。\n\n",
+            ])
+        if self.last_instruction and 'fetch_gpt_response_stream HTTP Error' not in self.last_instruction:
+            instruction_prompt = (
+                f"{instruction_prompt}\n\n"
+                "你生成的指令格式错误，必须把给assistant的指令放在<instructions>...</instructions>标签内。请重新生成格式正确的指令。"
+                f"这是你上次给assistant的错误格式的指令：\n{self.last_instruction}"
+            )
+
+        self.agent.conversation["default"][1:] = self.get_conversation_history(message["conversation"])
+
+        if "find_and_click_element" in json.dumps(self.tools_json):
+            instruction_prompt = await get_current_screen_image_message(instruction_prompt)
+
+        raw_response = await self.agent.ask_async(instruction_prompt)
+
+        if "fetch_gpt_response_stream HTTP Error', 'status_code': 404" in raw_response:
+            raise Exception(f"Model: {self.config['engine']} not found!")
+        if "'status_code': 413" in raw_response or \
+        "'status_code': 400" in raw_response:
+            self.broker.publish({"status": "error", "result": raw_response}, self.status_topic)
+            return
+
+        self.broker.publish({"status": "new_message", "result": "\n🤖 指令智能体:\n" + raw_response}, self.status_topic)
+
+        self.last_instruction = raw_response
+        instruction = extract_xml_content(raw_response, "instructions")
+        if instruction:
+            if len(message["conversation"]) == 1:
+                instruction = (
+                    "任务描述：\n"
+                    f"<goal>{self.goal}</goal>\n\n"
+                    "你作为指令的**执行者**，而非任务的**规划师**，你必须严格遵循以下单步工作流程：\n"
+                    "**执行指令**\n"
+                    "   - **严格遵从：** 只执行我当前下达的明确指令。在我明确给出下一步指令前，绝不擅自行动或推测、执行任何未明确要求的后续步骤。\n"
+                    "   - **严禁越权：** 禁止执行任何我未指定的步骤。`<goal>` 标签中的内容仅为背景信息，不得据此进行任务规划或推测。\n"
+                    "**汇报结果**\n"
+                    "   - **聚焦单步：** 指令完成后，仅汇报该步骤的执行结果与产出。\n"
+                    "**暂停等待**\n"
+                    "   - **原地待命：** 汇报后，任务暂停。在收到我新的指令前，严禁发起任何新的工具调用或操作。\n"
+                    "   - **请求指令：** 回复的最后必须明确请求我提供下一步指令。\n"
+                    "**注意：** 禁止完成超出下面我未规定的步骤，`<goal>` 标签中的内容仅为背景信息。"
+                    "现在开始执行第一步：\n"
+                    f"{instruction}"
+                )
+            self.broker.publish({"instruction": instruction, "conversation": message["conversation"]}, self.publish_topic)
+        else:
+            print("\n❌ 指令智能体生成的指令不符合要求，正在重新生成。")
+            self.broker.publish(message, self.error_topic)
+
+
+class WorkerAgent(BaseAgent):
+    """Executes instructions and publishes results to a message broker."""
+    def __init__(self, goal: str, tools_json: List, agent_config: Dict, work_dir: str, cache_messages: Union[bool, List[Dict]], broker: MessageBroker, listen_topic: str, publish_topic: str, status_topic: str):
+        super().__init__(goal, tools_json, agent_config, work_dir, cache_messages, broker, listen_topic, publish_topic, status_topic)
+
+        if self.cache_messages and isinstance(self.cache_messages, list) and len(self.cache_messages) > 1:
+            first_user_message = replace_xml_content(self.cache_messages[1]["content"], "goal", goal)
+            self.config["cache_messages"] = self.cache_messages[0:1] + [{"role": "user", "content": first_user_message}] + self.cache_messages[2:]
+
+        self.agent = chatgpt(**self.config)
+
+    async def handle_message(self, message: Dict):
+        """Receives an instruction, executes it, and publishes the response."""
+
+        if message.get("instruction") == "Initial kickoff":
+            self.broker.publish({
+                "conversation": self.agent.conversation["default"]
+            }, self.publish_topic)
+            return
+
+        instruction = message["instruction"]
+        if "find_and_click_element" in json.dumps(self.tools_json):
+            instruction = await get_current_screen_image_message(instruction)
+        response = await self.agent.ask_async(instruction)
+
+        if response.strip() == '':
+            print("\n❌ 工作智能体回复为空，请重新生成指令。")
+            self.broker.publish(message, self.error_topic)
+        else:
+            self.broker.publish({"status": "new_message", "result": "\n✅ 工作智能体:\n" + response}, self.status_topic)
+            self.broker.publish({
+                "conversation": self.agent.conversation["default"]
+            }, self.publish_topic)
 
 class Tee:
     def __init__(self, *files):
@@ -21,435 +198,170 @@ class Tee:
         for f in self.files:
             f.flush()
 
-from ..aient.src.aient.models import chatgpt
-from ..aient.src.aient.plugins import register_tool, get_function_call_list, registry
-from ..prompt import worker_system_prompt, instruction_system_prompt
-from ..utils import extract_xml_content, get_current_screen_image_message, replace_xml_content, register_mcp_tools
-from ..bemcp.bemcp import MCPClient, convert_tool_format, MCPManager
+broker = MessageBroker()
+mcp_manager = MCPManager()
+
+class BrokerWorker:
+    """The 'glue' class that orchestrates agents via a MessageBroker."""
+    def __init__(self, goal: str, tools: List[Union[str, Dict]], work_dir: str, cache_messages: Union[bool, List[Dict]] = None, broker: MessageBroker = None, mcp_manager: MCPManager = None):
+        self.goal = goal
+        self.tools = tools
+        self.work_dir = Path(work_dir)
+        self.cache_messages = cache_messages
+
+        self.broker = broker
+        self.mcp_manager = mcp_manager
+        self.task_completion_event = asyncio.Event()
+        self.final_result = None
+        self._status_subscription = None
+        self.setup()
+
+        self.channel = self.broker.request_channel()
+        self.INSTRUCTION_TOPIC = self.channel + ".instructions"
+        self.WORKER_RESPONSE_TOPIC = self.channel + ".worker_responses"
+        self.TASK_STATUS_TOPIC =self.channel +  ".task_status"
+
+    def setup(self):
+        cache_dir = self.work_dir / ".beswarm"
+        cache_dir.mkdir(parents=True, exist_ok=True)
+        task_manager.set_root_path(self.work_dir)
+        self.cache_file = cache_dir / "work_agent_conversation_history.json"
+        if not self.cache_file.exists():
+            self.cache_file.write_text("[]", encoding="utf-8")
+
+        DEBUG = os.getenv("DEBUG", "false").lower() in ("true", "1", "t", "yes")
+        if DEBUG:
+            log_file = open(cache_dir / "history.log", "a", encoding="utf-8")
+            log_file.write(f"========== {datetime.now().strftime('%Y-%m-%d %H:%M:%S')} ==========\n")
+            original_stdout = sys.stdout
+            original_stderr = sys.stderr
+            sys.stdout = Tee(original_stdout, log_file)
+            sys.stderr = Tee(original_stderr, log_file)
+
+    async def _configure_tools(self):
+        mcp_list = [item for item in self.tools if isinstance(item, dict)]
+        if mcp_list:
+            for mcp_item in mcp_list:
+                mcp_name, mcp_config = list(mcp_item.items())[0]
+                await self.mcp_manager.add_server(mcp_name, mcp_config)
+                client = self.mcp_manager.clients.get(mcp_name)
+                await register_mcp_tools(client, registry)
+            all_mcp_tools = await self.mcp_manager.get_all_tools()
+            self.tools.extend([tool.name for tool in sum(all_mcp_tools.values(), [])])
+        self.tools = [item for item in self.tools if not isinstance(item, dict)]
+        if "task_complete" not in self.tools: self.tools.append("task_complete")
+        self.tools_json = [value for _, value in get_function_call_list(self.tools).items()]
+
+    def _task_status_subscriber(self, message: Dict):
+        """Subscriber for task status changes."""
+        if message.get("status") == "finished":
+            self.final_result = message.get("result")
+            self.task_completion_event.set()
+
+        if message.get("status") == "error":
+            raise Exception(message.get("result"))
+
+        if message.get("status") == "new_message":
+            print(message.get("result"))
+
+    def _setup_agents(self):
+        instruction_agent_config = {
+            "api_key": os.getenv("API_KEY"), "api_url": os.getenv("BASE_URL"),
+            "engine": os.getenv("MODEL"),
+            "system_prompt": instruction_system_prompt.format(
+                os_version=platform.platform(), tools_list=self.tools_json,
+                workspace_path=self.work_dir, current_time=datetime.now().strftime("%Y-%m-%d %H:%M:%S")
+            ),
+            "print_log": os.getenv("DEBUG", "false").lower() in ("true", "1", "t", "yes"),
+            "temperature": 0.7, "use_plugins": False
+        }
+
+        worker_agent_config = {
+            "api_key": os.getenv("API_KEY"), "api_url": os.getenv("BASE_URL"),
+            "engine": os.getenv("FAST_MODEL") or os.getenv("MODEL"),
+            "system_prompt": worker_system_prompt.format(
+                os_version=platform.platform(), workspace_path=self.work_dir,
+                shell=os.getenv('SHELL', 'Unknown'), current_time=datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
+                tools_list=self.tools_json
+            ),
+            "print_log": True, "temperature": 0.5, "function_call_max_loop": 100
+        }
+
+        instruction_agent = InstructionAgent(
+            goal=self.goal, tools_json=self.tools_json, agent_config=instruction_agent_config, work_dir=self.work_dir, cache_messages=self.cache_messages,
+            broker=self.broker, listen_topic=self.WORKER_RESPONSE_TOPIC,
+            publish_topic=self.INSTRUCTION_TOPIC, status_topic=self.TASK_STATUS_TOPIC
+        )
+
+        worker_agent = WorkerAgent(
+            goal=self.goal, tools_json=self.tools_json, agent_config=worker_agent_config, work_dir=self.work_dir, cache_messages=self.cache_messages,
+            broker=self.broker, listen_topic=self.INSTRUCTION_TOPIC,
+            publish_topic=self.WORKER_RESPONSE_TOPIC, status_topic=self.TASK_STATUS_TOPIC
+        )
+        return instruction_agent, worker_agent
+
+    async def run(self):
+        """Sets up subscriptions and starts the workflow."""
+        os.chdir(self.work_dir.absolute())
+        await self._configure_tools()
+
+        instruction_agent, worker_agent = self._setup_agents()
+
+        self.broker.publish({"instruction": "Initial kickoff"}, self.INSTRUCTION_TOPIC)
+
+        self._status_subscription = self.broker.subscribe(self._task_status_subscriber, self.TASK_STATUS_TOPIC)
+        await self.task_completion_event.wait()
+
+        instruction_agent.dispose()
+        worker_agent.dispose()
+        self._status_subscription.dispose()
+        await self.mcp_manager.cleanup()
+        return self.final_result
+
+    async def stream_run(self):
+        """Runs the workflow and yields status messages."""
+        os.chdir(self.work_dir.absolute())
+        await self._configure_tools()
+
+        instruction_agent, worker_agent = self._setup_agents()
+
+        self.broker.publish({"instruction": "Initial kickoff"}, self.INSTRUCTION_TOPIC)
+
+        try:
+            async for message in self.broker.iter_topic(self.TASK_STATUS_TOPIC):
+                if message.get("status") == "new_message":
+                    yield message.get("result")
+                elif message.get("status") == "finished":
+                    yield message.get("result")
+                    break
+                elif message.get("status") == "error":
+                    raise Exception(message.get("result"))
+        finally:
+            instruction_agent.dispose()
+            worker_agent.dispose()
+            await self.mcp_manager.cleanup()
 
-manager = MCPManager()
 
 @register_tool()
-async def worker(goal, tools, work_dir, cache_messages=None):
-    cache_dir = Path(work_dir) / ".beswarm"
-    cache_dir.mkdir(parents=True, exist_ok=True)
-    task_manager.set_root_path(work_dir)
-    cache_file = cache_dir / "work_agent_conversation_history.json"
-    if not cache_file.exists():
-        cache_file.write_text("[]", encoding="utf-8")
-
-    DEBUG = os.getenv("DEBUG", "false").lower() in ("true", "1", "t", "yes")
-    if DEBUG:
-        log_file = open(cache_dir / "history.log", "a", encoding="utf-8")
-        log_file.write(f"========== {datetime.now().strftime('%Y-%m-%d %H:%M:%S')} ==========\n")
-        original_stdout = sys.stdout
-        original_stderr = sys.stderr
-        sys.stdout = Tee(original_stdout, log_file)
-        sys.stderr = Tee(original_stderr, log_file)
-
+async def worker(goal: str, tools: List[Union[str, Dict]], work_dir: str, cache_messages: Union[bool, List[Dict]] = None):
     start_time = datetime.now()
-    os.chdir(Path(work_dir).absolute())
-    finish_flag = 0
-    goal_diff = None
-
-    mcp_list = [item for item in tools if isinstance(item, dict)]
-    if mcp_list:
-        for mcp_item in mcp_list:
-            mcp_name, mcp_config = list(mcp_item.items())[0]
-            await manager.add_server(mcp_name, mcp_config)
-            client = manager.clients.get(mcp_name)
-            await register_mcp_tools(client, registry)
-        all_tools = await manager.get_all_tools()
-        mcp_tools_name = [tool.name for tool in sum(all_tools.values(), [])]
-        tools += mcp_tools_name
-
-    tools = [item for item in tools if not isinstance(item, dict)]
-    if "task_complete" not in tools:
-        tools.append("task_complete")
-
-    tools_json = [value for _, value in get_function_call_list(tools).items()]
-    work_agent_system_prompt = worker_system_prompt.format(
-        os_version=platform.platform(),
-        workspace_path=work_dir,
-        shell=os.getenv('SHELL', 'Unknown'),
-        current_time=datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
-        tools_list=tools_json
-    )
-
-    work_agent_config = {
-        "api_key": os.getenv("API_KEY"),
-        "api_url": os.getenv("BASE_URL"),
-        "engine": os.getenv("FAST_MODEL") or os.getenv("MODEL"),
-        "system_prompt": work_agent_system_prompt,
-        "print_log": True,
-        # "max_tokens": 8000,
-        "temperature": 0.5,
-        "function_call_max_loop": 100,
-    }
-    if cache_messages:
-        if isinstance(cache_messages, bool) and cache_messages == True:
-            cache_messages = json.loads(cache_file.read_text(encoding="utf-8"))
-        if cache_messages and isinstance(cache_messages, list) and len(cache_messages) > 1:
-            old_goal = extract_xml_content(cache_messages[1]["content"], "goal")
-            if old_goal.strip() != goal.strip():
-                diff_generator = difflib.ndiff(old_goal.splitlines(), goal.splitlines())
-                changed_lines = []
-                for line in diff_generator:
-                    if (line.startswith('+ ') or line.startswith('- ')) and line[2:].strip():
-                        changed_lines.append(line)
-                goal_diff = '\n'.join(changed_lines).strip()
-            first_user_message = replace_xml_content(cache_messages[1]["content"], "goal", goal)
-            work_agent_config["cache_messages"] = cache_messages[0:1] + [{"role": "user", "content": first_user_message}] + cache_messages[2:]
-
-    instruction_agent_config = {
-        "api_key": os.getenv("API_KEY"),
-        "api_url": os.getenv("BASE_URL"),
-        "engine": os.getenv("MODEL"),
-        "system_prompt": instruction_system_prompt.format(os_version=platform.platform(), tools_list=tools_json, workspace_path=work_dir, current_time=datetime.now().strftime("%Y-%m-%d %H:%M:%S")),
-        "print_log": DEBUG,
-        # "max_tokens": 4000,
-        "temperature": 0.7,
-        "use_plugins": False,
-    }
-
-    # 工作agent初始化
-    work_agent = chatgpt(**work_agent_config)
-    async def instruction_agent_task():
-        last_instruction = None
-        while True:
-            instruction_prompt = "".join([
-                    "</work_agent_conversation_end>\n\n",
-                    f"任务目标: {goal}\n\n",
-                    f"任务目标新变化：\n{goal_diff}\n\n" if goal_diff else "",
-                    "在 tag <work_agent_conversation_start>...</work_agent_conversation_end> 之前的对话历史都是工作智能体的对话历史。\n\n",
-                    "根据以上对话历史和目标，请生成下一步指令。如果任务已完成，指示工作智能体调用task_complete工具。\n\n",
-                ])
-            if last_instruction and 'fetch_gpt_response_stream HTTP Error' not in last_instruction:
-                instruction_prompt = (
-                    f"{instruction_prompt}\n\n"
-                    "你生成的指令格式错误，必须把给assistant的指令放在<instructions>...</instructions>标签内。请重新生成格式正确的指令。"
-                    f"这是你上次给assistant的错误格式的指令：\n{last_instruction}"
-                )
-            # 让指令agent分析对话历史并生成新指令
-            instruction_agent = chatgpt(**instruction_agent_config)
-            conversation_history = copy.deepcopy(work_agent.conversation["default"])
-            if len(conversation_history) > 1 and conversation_history[-2]["role"] == "user" \
-            and "<task_complete_message>" in conversation_history[-2]["content"]:
-                task_complete_message = extract_xml_content(conversation_history[-2]["content"], "task_complete_message")
-                # del work_agent.conversation["default"][-4:]
-                return "<task_complete_message>" + task_complete_message + "</task_complete_message>"
-
-            cache_file.write_text(json.dumps(conversation_history, ensure_ascii=False, indent=4), encoding="utf-8")
-
-            work_agent_system_prompt = conversation_history.pop(0)
-            if conversation_history:
-                # 获取原始内容
-                original_content = work_agent_system_prompt["content"]
-
-                # 定义正则表达式
-                regex = r"<latest_file_content>(.*?)</latest_file_content>"
-
-                # 进行匹配
-                match = re.search(regex, original_content, re.DOTALL)
-
-                # 提取内容或设置为空字符串
-                if match:
-                    extracted_content = f"<latest_file_content>{match.group(1)}</latest_file_content>\n\n"
-                else:
-                    extracted_content = ""
-                if isinstance(conversation_history[0]["content"], str):
-                    conversation_history[0]["content"] = extracted_content + conversation_history[0]["content"]
-                elif isinstance(conversation_history[0]["content"], list) and extracted_content:
-                    conversation_history[0]["content"].append({"type": "text", "text": extracted_content})
-
-            instruction_agent.conversation["default"][1:] = conversation_history
-            if "find_and_click_element" in str(tools_json):
-                instruction_prompt = await get_current_screen_image_message(instruction_prompt)
-            next_instruction = await instruction_agent.ask_async(instruction_prompt)
-            print("\n🤖 指令智能体生成的下一步指令:", next_instruction)
-            if "fetch_gpt_response_stream HTTP Error', 'status_code': 404" in next_instruction:
-                raise Exception(f"Model: {instruction_agent_config['engine']} not found!")
-            if "'status_code': 413" in next_instruction or \
-            "'status_code': 400" in next_instruction:
-                end_time = datetime.now()
-                total_time = end_time - start_time
-                print(f"\n任务开始时间: {start_time.strftime('%Y-%m-%d %H:%M:%S')}")
-                print(f"任务结束时间: {end_time.strftime('%Y-%m-%d %H:%M:%S')}")
-                print(f"总用时: {total_time}")
-                raise Exception(f"The request body is too long, please try again.")
-
-            last_instruction = next_instruction
-            next_instruction = extract_xml_content(next_instruction, "instructions")
-            if not next_instruction:
-                print("\n❌ 指令智能体生成的指令不符合要求，请重新生成。")
-                continue
-            else:
-                if conversation_history == []:
-                    next_instruction = (
-                        "任务描述：\n"
-                        f"<goal>{goal}</goal>\n\n"
-                        "你作为指令的**执行者**，而非任务的**规划师**，你必须严格遵循以下单步工作流程：\n"
-                        "**执行指令**\n"
-                        "   - **严格遵从：** 只执行我当前下达的明确指令。在我明确给出下一步指令前，绝不擅自行动或推测、执行任何未明确要求的后续步骤。\n"
-                        "   - **严禁越权：** 禁止执行任何我未指定的步骤。`<goal>` 标签中的内容仅为背景信息，不得据此进行任务规划或推测。\n"
-                        "**汇报结果**\n"
-                        "   - **聚焦单步：** 指令完成后，仅汇报该步骤的执行结果与产出。\n"
-                        "**暂停等待**\n"
-                        "   - **原地待命：** 汇报后，任务暂停。在收到我新的指令前，严禁发起任何新的工具调用或操作。\n"
-                        "   - **请求指令：** 回复的最后必须明确请求我提供下一步指令。\n"
-                        "**注意：** 禁止完成超出下面我未规定的步骤，`<goal>` 标签中的内容仅为背景信息。"
-                        "现在开始执行第一步：\n"
-                        f"{next_instruction}"
-                    )
-                break
-        return next_instruction
-
-    need_instruction = True
-    result = None
-    while True:
-        next_instruction = ''
-        if need_instruction:
-            next_instruction = await instruction_agent_task()
-
-            # 检查任务是否完成
-            if "<task_complete_message>" in next_instruction:
-                if finish_flag == 0:
-                    finish_flag = 1
-                    continue
-                elif finish_flag == 1:
-                    result = extract_xml_content(next_instruction, "task_complete_message")
-                    break
-            else:
-                finish_flag = 0
-        if "find_and_click_element" in str(tools_json):
-            next_instruction = await get_current_screen_image_message(next_instruction)
-        result = await work_agent.ask_async(next_instruction)
-        if result.strip() == '' or result.strip() == '</content>\n</write_to_file>':
-            print("\n❌ 工作智能体回复为空，请重新生成指令。")
-            need_instruction = False
-            continue
-        print("✅ 工作智能体回复:", result)
-        need_instruction = True
-
+    worker_instance = BrokerWorker(goal, tools, work_dir, cache_messages, broker, mcp_manager)
+    result = await worker_instance.run()
     end_time = datetime.now()
-    total_time = end_time - start_time
-    print("\n✅ 任务已完成：", result)
     print(f"\n任务开始时间: {start_time.strftime('%Y-%m-%d %H:%M:%S')}")
     print(f"任务结束时间: {end_time.strftime('%Y-%m-%d %H:%M:%S')}")
-    print(f"总用时: {total_time}")
-    await manager.cleanup()
+    print(f"总用时: {end_time - start_time}")
     return result
 
-async def worker_gen(goal, tools, work_dir, cache_messages=None):
-    cache_dir = Path(work_dir) / ".beswarm"
-    cache_dir.mkdir(parents=True, exist_ok=True)
-    task_manager.set_root_path(work_dir)
-    cache_file = cache_dir / "work_agent_conversation_history.json"
-    if not cache_file.exists():
-        cache_file.write_text("[]", encoding="utf-8")
-
-    DEBUG = os.getenv("DEBUG", "false").lower() in ("true", "1", "t", "yes")
-    if DEBUG:
-        log_file = open(cache_dir / "history.log", "a", encoding="utf-8")
-        log_file.write(f"========== {datetime.now().strftime('%Y-%m-%d %H:%M:%S')} ==========\n")
-        original_stdout = sys.stdout
-        original_stderr = sys.stderr
-        sys.stdout = Tee(original_stdout, log_file)
-        sys.stderr = Tee(original_stderr, log_file)
-
+@register_tool()
+async def worker_gen(goal: str, tools: List[Union[str, Dict]], work_dir: str, cache_messages: Union[bool, List[Dict]] = None):
     start_time = datetime.now()
-    os.chdir(Path(work_dir).absolute())
-    finish_flag = 0
-    goal_diff = None
-
-    mcp_list = [item for item in tools if isinstance(item, dict)]
-    if mcp_list:
-        for mcp_item in mcp_list:
-            mcp_name, mcp_config = list(mcp_item.items())[0]
-            await manager.add_server(mcp_name, mcp_config)
-            client = manager.clients.get(mcp_name)
-            await register_mcp_tools(client, registry)
-        all_tools = await manager.get_all_tools()
-        mcp_tools_name = [tool.name for tool in sum(all_tools.values(), [])]
-        tools += mcp_tools_name
-
-    tools = [item for item in tools if not isinstance(item, dict)]
-    if "task_complete" not in tools:
-        tools.append("task_complete")
-
-    tools_json = [value for _, value in get_function_call_list(tools).items()]
-    work_agent_system_prompt = worker_system_prompt.format(
-        os_version=platform.platform(),
-        workspace_path=work_dir,
-        shell=os.getenv('SHELL', 'Unknown'),
-        current_time=datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
-        tools_list=tools_json
-    )
-
-    work_agent_config = {
-        "api_key": os.getenv("API_KEY"),
-        "api_url": os.getenv("BASE_URL"),
-        "engine": os.getenv("FAST_MODEL") or os.getenv("MODEL"),
-        "system_prompt": work_agent_system_prompt,
-        "print_log": True,
-        # "max_tokens": 8000,
-        "temperature": 0.5,
-        "function_call_max_loop": 100,
-    }
-    if cache_messages:
-        if isinstance(cache_messages, bool) and cache_messages == True:
-            cache_messages = json.loads(cache_file.read_text(encoding="utf-8"))
-        if cache_messages and isinstance(cache_messages, list) and len(cache_messages) > 1:
-            old_goal = extract_xml_content(cache_messages[1]["content"], "goal")
-            if old_goal.strip() != goal.strip():
-                diff_generator = difflib.ndiff(old_goal.splitlines(), goal.splitlines())
-                changed_lines = []
-                for line in diff_generator:
-                    if (line.startswith('+ ') or line.startswith('- ')) and line[2:].strip():
-                        changed_lines.append(line)
-                goal_diff = '\n'.join(changed_lines).strip()
-            first_user_message = replace_xml_content(cache_messages[1]["content"], "goal", goal)
-            work_agent_config["cache_messages"] = cache_messages[0:1] + [{"role": "user", "content": first_user_message}] + cache_messages[2:]
-
-    instruction_agent_config = {
-        "api_key": os.getenv("API_KEY"),
-        "api_url": os.getenv("BASE_URL"),
-        "engine": os.getenv("MODEL"),
-        "system_prompt": instruction_system_prompt.format(os_version=platform.platform(), tools_list=tools_json, workspace_path=work_dir, current_time=datetime.now().strftime("%Y-%m-%d %H:%M:%S")),
-        "print_log": DEBUG,
-        # "max_tokens": 4000,
-        "temperature": 0.7,
-        "use_plugins": False,
-    }
-
-    # 工作agent初始化
-    work_agent = chatgpt(**work_agent_config)
-    async def instruction_agent_task():
-        last_instruction = None
-        while True:
-            instruction_prompt = "".join([
-                    "</work_agent_conversation_end>\n\n",
-                    f"任务目标: {goal}\n\n",
-                    f"任务目标新变化：\n{goal_diff}\n\n" if goal_diff else "",
-                    "在 tag <work_agent_conversation_start>...</work_agent_conversation_end> 之前的对话历史都是工作智能体的对话历史。\n\n",
-                    "根据以上对话历史和目标，请生成下一步指令。如果任务已完成，指示工作智能体调用task_complete工具。\n\n",
-                ])
-            if last_instruction and 'fetch_gpt_response_stream HTTP Error' not in last_instruction:
-                instruction_prompt = (
-                    f"{instruction_prompt}\n\n"
-                    "你生成的指令格式错误，必须把给assistant的指令放在<instructions>...</instructions>标签内。请重新生成格式正确的指令。"
-                    f"这是你上次给assistant的错误格式的指令：\n{last_instruction}"
-                )
-            # 让指令agent分析对话历史并生成新指令
-            instruction_agent = chatgpt(**instruction_agent_config)
-            conversation_history = copy.deepcopy(work_agent.conversation["default"])
-            if len(conversation_history) > 1 and conversation_history[-2]["role"] == "user" \
-            and "<task_complete_message>" in conversation_history[-2]["content"]:
-                task_complete_message = extract_xml_content(conversation_history[-2]["content"], "task_complete_message")
-                # del work_agent.conversation["default"][-4:]
-                return "<task_complete_message>" + task_complete_message + "</task_complete_message>"
-
-            cache_file.write_text(json.dumps(conversation_history, ensure_ascii=False, indent=4), encoding="utf-8")
-
-            work_agent_system_prompt = conversation_history.pop(0)
-            if conversation_history:
-                # 获取原始内容
-                original_content = work_agent_system_prompt["content"]
-
-                # 定义正则表达式
-                regex = r"<latest_file_content>(.*?)</latest_file_content>"
-
-                # 进行匹配
-                match = re.search(regex, original_content, re.DOTALL)
-
-                # 提取内容或设置为空字符串
-                if match:
-                    extracted_content = f"<latest_file_content>{match.group(1)}</latest_file_content>\n\n"
-                else:
-                    extracted_content = ""
-                if isinstance(conversation_history[0]["content"], str):
-                    conversation_history[0]["content"] = extracted_content + conversation_history[0]["content"]
-                elif isinstance(conversation_history[0]["content"], list) and extracted_content:
-                    conversation_history[0]["content"].append({"type": "text", "text": extracted_content})
-
-            instruction_agent.conversation["default"][1:] = conversation_history
-            if "find_and_click_element" in str(tools_json):
-                instruction_prompt = await get_current_screen_image_message(instruction_prompt)
-            next_instruction = await instruction_agent.ask_async(instruction_prompt)
-            print("\n🤖 指令智能体生成的下一步指令:", next_instruction)
-            if "fetch_gpt_response_stream HTTP Error', 'status_code': 404" in next_instruction:
-                raise Exception(f"Model: {instruction_agent_config['engine']} not found!")
-            if "'status_code': 413" in next_instruction or \
-            "'status_code': 400" in next_instruction:
-                end_time = datetime.now()
-                total_time = end_time - start_time
-                print(f"\n任务开始时间: {start_time.strftime('%Y-%m-%d %H:%M:%S')}")
-                print(f"任务结束时间: {end_time.strftime('%Y-%m-%d %H:%M:%S')}")
-                print(f"总用时: {total_time}")
-                raise Exception(f"The request body is too long, please try again.")
-
-            last_instruction = next_instruction
-            next_instruction = extract_xml_content(next_instruction, "instructions")
-            if not next_instruction:
-                print("\n❌ 指令智能体生成的指令不符合要求，请重新生成。")
-                continue
-            else:
-                if conversation_history == []:
-                    next_instruction = (
-                        "任务描述：\n"
-                        f"<goal>{goal}</goal>\n\n"
-                        "你作为指令的**执行者**，而非任务的**规划师**，你必须严格遵循以下单步工作流程：\n"
-                        "**执行指令**\n"
-                        "   - **严格遵从：** 只执行我当前下达的明确指令。在我明确给出下一步指令前，绝不擅自行动或推测、执行任何未明确要求的后续步骤。\n"
-                        "   - **严禁越权：** 禁止执行任何我未指定的步骤。`<goal>` 标签中的内容仅为背景信息，不得据此进行任务规划或推测。\n"
-                        "**汇报结果**\n"
-                        "   - **聚焦单步：** 指令完成后，仅汇报该步骤的执行结果与产出。\n"
-                        "**暂停等待**\n"
-                        "   - **原地待命：** 汇报后，任务暂停。在收到我新的指令前，严禁发起任何新的工具调用或操作。\n"
-                        "   - **请求指令：** 回复的最后必须明确请求我提供下一步指令。\n"
-                        "**注意：** 禁止完成超出下面我未规定的步骤，`<goal>` 标签中的内容仅为背景信息。"
-                        "现在开始执行第一步：\n"
-                        f"{next_instruction}"
-                    )
-                break
-        return next_instruction
-
-    need_instruction = True
-    result = None
-    while True:
-        next_instruction = ''
-        if need_instruction:
-            next_instruction = await instruction_agent_task()
-
-            yield {"user": next_instruction}
-
-            # 检查任务是否完成
-            if "<task_complete_message>" in next_instruction:
-                if finish_flag == 0:
-                    finish_flag = 1
-                    continue
-                elif finish_flag == 1:
-                    result = extract_xml_content(next_instruction, "task_complete_message")
-                    break
-            else:
-                finish_flag = 0
-        if "find_and_click_element" in str(tools_json):
-            next_instruction = await get_current_screen_image_message(next_instruction)
-        result = await work_agent.ask_async(next_instruction)
-        if result.strip() == '' or result.strip() == '</content>\n</write_to_file>':
-            print("\n❌ 工作智能体回复为空，请重新生成指令。")
-            need_instruction = False
-            continue
-        yield {"assistant": result}
-        print("✅ 工作智能体回复:", result)
-        need_instruction = True
-
+    worker_instance = BrokerWorker(goal, tools, work_dir, cache_messages, broker, mcp_manager)
+    async for result in worker_instance.stream_run():
+        yield result
     end_time = datetime.now()
-    total_time = end_time - start_time
-    print("\n✅ 任务已完成：", result)
     print(f"\n任务开始时间: {start_time.strftime('%Y-%m-%d %H:%M:%S')}")
     print(f"任务结束时间: {end_time.strftime('%Y-%m-%d %H:%M:%S')}")
-    print(f"总用时: {total_time}")
-    await manager.cleanup()
+    print(f"总用时: {end_time - start_time}")
 
-from .taskmanager import task_manager
+from .taskmanager import task_manager