{"repo_name": "web-ui", "file_name": "/web-ui/src/agent/deep_research/deep_research_agent.py", "inference_info": {"prefix_code": "import asyncio\nimport json\nimport logging\nimport os\nimport threading\nimport uuid\nfrom pathlib import Path\nfrom typing import Any, Dict, List, Optional, TypedDict\n\nfrom browser_use.browser.browser import BrowserConfig\nfrom langchain_community.tools.file_management import (\n ListDirectoryTool,\n ReadFileTool,\n WriteFileTool,\n)\n\n# Langchain imports\nfrom langchain_core.messages import (\n AIMessage,\n BaseMessage,\n HumanMessage,\n SystemMessage,\n ToolMessage,\n)\nfrom langchain_core.prompts import ChatPromptTemplate\nfrom langchain_core.tools import StructuredTool, Tool\n\n# Langgraph imports\nfrom langgraph.graph import StateGraph\nfrom pydantic import BaseModel, Field\n\nfrom browser_use.browser.context import BrowserContextConfig\n\nfrom src.agent.browser_use.browser_use_agent import BrowserUseAgent\nfrom src.browser.custom_browser import CustomBrowser\nfrom src.controller.custom_controller import CustomController\nfrom src.utils.mcp_client import setup_mcp_client_and_tools\n\nlogger = logging.getLogger(__name__)\n\n# Constants\nREPORT_FILENAME = \"report.md\"\nPLAN_FILENAME = \"research_plan.md\"\nSEARCH_INFO_FILENAME = \"search_info.json\"\n\n_AGENT_STOP_FLAGS = {}\n_BROWSER_AGENT_INSTANCES = {}\n\n\nasync def run_single_browser_task(\n task_query: str,\n task_id: str,\n llm: Any, # Pass the main LLM\n browser_config: Dict[str, Any],\n stop_event: threading.Event,\n use_vision: bool = False,\n) -> Dict[str, Any]:\n \"\"\"\n Runs a single BrowserUseAgent task.\n Manages browser creation and closing for this specific task.\n \"\"\"\n if not BrowserUseAgent:\n return {\n \"query\": task_query,\n \"error\": \"BrowserUseAgent components not available.\",\n }\n\n # --- Browser Setup ---\n # These should ideally come from the main agent's config\n headless = browser_config.get(\"headless\", False)\n window_w = browser_config.get(\"window_width\", 1280)\n window_h = browser_config.get(\"window_height\", 1100)\n browser_user_data_dir = browser_config.get(\"user_data_dir\", None)\n use_own_browser = browser_config.get(\"use_own_browser\", False)\n browser_binary_path = browser_config.get(\"browser_binary_path\", None)\n wss_url = browser_config.get(\"wss_url\", None)\n cdp_url = browser_config.get(\"cdp_url\", None)\n disable_security = browser_config.get(\"disable_security\", False)\n\n bu_browser = None\n bu_browser_context = None\n try:\n logger.info(f\"Starting browser task for query: {task_query}\")\n extra_args = []\n if use_own_browser:\n browser_binary_path = os.getenv(\"BROWSER_PATH\", None) or browser_binary_path\n if browser_binary_path == \"\":\n browser_binary_path = None\n browser_user_data = browser_user_data_dir or os.getenv(\"BROWSER_USER_DATA\", None)\n if browser_user_data:\n extra_args += [f\"--user-data-dir={browser_user_data}\"]\n else:\n browser_binary_path = None\n\n bu_browser = CustomBrowser(\n config=BrowserConfig(\n headless=headless,\n browser_binary_path=browser_binary_path,\n extra_browser_args=extra_args,\n wss_url=wss_url,\n cdp_url=cdp_url,\n new_context_config=BrowserContextConfig(\n window_width=window_w,\n window_height=window_h,\n )\n )\n )\n\n context_config = BrowserContextConfig(\n save_downloads_path=\"./tmp/downloads\",\n window_height=window_h,\n window_width=window_w,\n force_new_context=True,\n )\n bu_browser_context = await bu_browser.new_context(config=context_config)\n\n # Simple controller example, replace with your actual implementation if needed\n bu_controller = CustomController()\n\n # Construct the task prompt for BrowserUseAgent\n # Instruct it to find specific info and return title/URL\n bu_task_prompt = f\"\"\"\n Research Task: {task_query}\n Objective: Find relevant information answering the query.\n Output Requirements: For each relevant piece of information found, please provide:\n 1. A concise summary of the information.\n 2. The title of the source page or document.\n 3. The URL of the source.\n Focus on accuracy and relevance. Avoid irrelevant details.\n PDF cannot directly extract _content, please try to download first, then using read_file, if you can't save or read, please try other methods.\n \"\"\"\n\n bu_agent_instance = BrowserUseAgent(\n task=bu_task_prompt,\n llm=llm, # Use the passed LLM\n browser=bu_browser,\n browser_context=bu_browser_context,\n controller=bu_controller,\n use_vision=use_vision,\n source=\"webui\",\n )\n\n # Store instance for potential stop() call\n task_key = f\"{task_id}_{uuid.uuid4()}\"\n _BROWSER_AGENT_INSTANCES[task_key] = bu_agent_instance\n\n # --- Run with Stop Check ---\n # BrowserUseAgent needs to internally check a stop signal or have a stop method.\n # We simulate checking before starting and assume `run` might be interruptible\n # or have its own stop mechanism we can trigger via bu_agent_instance.stop().\n if stop_event.is_set():\n logger.info(f\"Browser task for '{task_query}' cancelled before start.\")\n return {\"query\": task_query, \"result\": None, \"status\": \"cancelled\"}\n\n # The run needs to be awaitable and ideally accept a stop signal or have a .stop() method\n # result = await bu_agent_instance.run(max_steps=max_steps) # Add max_steps if applicable\n # Let's assume a simplified run for now\n logger.info(f\"Running BrowserUseAgent for: {task_query}\")\n result = await bu_agent_instance.run() # Assuming run is the main method\n logger.info(f\"BrowserUseAgent finished for: {task_query}\")\n\n final_data = result.final_result()\n\n if stop_event.is_set():\n logger.info(f\"Browser task for '{task_query}' stopped during execution.\")\n return {\"query\": task_query, \"result\": final_data, \"status\": \"stopped\"}\n else:\n logger.info(f\"Browser result for '{task_query}': {final_data}\")\n return {\"query\": task_query, \"result\": final_data, \"status\": \"completed\"}\n\n except Exception as e:\n logger.error(\n f\"Error during browser task for query '{task_query}': {e}\", exc_info=True\n )\n return {\"query\": task_query, \"error\": str(e), \"status\": \"failed\"}\n finally:\n if bu_browser_context:\n try:\n await bu_browser_context.close()\n bu_browser_context = None\n logger.info(\"Closed browser context.\")\n except Exception as e:\n logger.error(f\"Error closing browser context: {e}\")\n if bu_browser:\n try:\n await bu_browser.close()\n bu_browser = None\n logger.info(\"Closed browser.\")\n except Exception as e:\n logger.error(f\"Error closing browser: {e}\")\n\n if task_key in _BROWSER_AGENT_INSTANCES:\n del _BROWSER_AGENT_INSTANCES[task_key]\n\n\nclass BrowserSearchInput(BaseModel):\n queries: List[str] = Field(\n description=\"List of distinct search queries to find information relevant to the research task.\"\n )\n\n\nasync def _run_browser_search_tool(\n queries: List[str],\n task_id: str, # Injected dependency\n llm: Any, # Injected dependency\n browser_config: Dict[str, Any],\n stop_event: threading.Event,\n max_parallel_browsers: int = 1,\n) -> List[Dict[str, Any]]:\n \"\"\"\n Internal function to execute parallel browser searches based on LLM-provided queries.\n Handles concurrency and stop signals.\n \"\"\"\n\n # Limit queries just in case LLM ignores the description\n queries = queries[:max_parallel_browsers]\n logger.info(\n f\"[Browser Tool {task_id}] Running search for {len(queries)} queries: {queries}\"\n )\n\n results = []\n semaphore = asyncio.Semaphore(max_parallel_browsers)\n\n async def task_wrapper(query):\n async with semaphore:\n if stop_event.is_set():\n logger.info(\n f\"[Browser Tool {task_id}] Skipping task due to stop signal: {query}\"\n )\n return {\"query\": query, \"result\": None, \"status\": \"cancelled\"}\n # Pass necessary injected configs and the stop event\n return await run_single_browser_task(\n query,\n task_id,\n llm, # Pass the main LLM (or a dedicated one if needed)\n browser_config,\n stop_event,\n # use_vision could be added here if needed\n )\n\n tasks = [task_wrapper(query) for query in queries]\n search_results = await asyncio.gather(*tasks, return_exceptions=True)\n\n processed_results = []\n for i, res in enumerate(search_results):\n query = queries[i] # Get corresponding query\n if isinstance(res, Exception):\n logger.error(\n f\"[Browser Tool {task_id}] Gather caught exception for query '{query}': {res}\",\n exc_info=True,\n )\n processed_results.append(\n {\"query\": query, \"error\": str(res), \"status\": \"failed\"}\n )\n elif isinstance(res, dict):\n processed_results.append(res)\n else:\n logger.error(\n f\"[Browser Tool {task_id}] Unexpected result type for query '{query}': {type(res)}\"\n )\n processed_results.append(\n {\"query\": query, \"error\": \"Unexpected result type\", \"status\": \"failed\"}\n )\n\n logger.info(\n f\"[Browser Tool {task_id}] Finished search. Results count: {len(processed_results)}\"\n )\n return processed_results\n\n\ndef create_browser_search_tool(\n llm: Any,\n browser_config: Dict[str, Any],\n task_id: str,\n stop_event: threading.Event,\n max_parallel_browsers: int = 1,\n) -> StructuredTool:\n \"\"\"Factory function to create the browser search tool with necessary dependencies.\"\"\"\n # Use partial to bind the dependencies that aren't part of the LLM call arguments\n from functools import partial\n\n bound_tool_func = partial(\n _run_browser_search_tool,\n task_id=task_id,\n llm=llm,\n browser_config=browser_config,\n stop_event=stop_event,\n max_parallel_browsers=max_parallel_browsers,\n )\n\n return StructuredTool.from_function(\n coroutine=bound_tool_func,\n name=\"parallel_browser_search\",\n description=f\"\"\"Use this tool to actively search the web for information related to a specific research task or question.\nIt runs up to {max_parallel_browsers} searches in parallel using a browser agent for better results than simple scraping.\nProvide a list of distinct search queries(up to {max_parallel_browsers}) that are likely to yield relevant information.\"\"\",\n args_schema=BrowserSearchInput,\n )\n\n\n# --- Langgraph State Definition ---\n\n\nclass ResearchTaskItem(TypedDict):\n # step: int # Maybe step within category, or just implicit by order\n task_description: str\n status: str # \"pending\", \"completed\", \"failed\"\n queries: Optional[List[str]]\n result_summary: Optional[str]\n\n\nclass ResearchCategoryItem(TypedDict):\n category_name: str\n tasks: List[ResearchTaskItem]\n # Optional: category_status: str # Could be \"pending\", \"in_progress\", \"completed\"\n\n\nclass DeepResearchState(TypedDict):\n task_id: str\n topic: str\n research_plan: List[ResearchCategoryItem] # CHANGED\n search_results: List[Dict[str, Any]]\n llm: Any\n tools: List[Tool]\n output_dir: Path\n browser_config: Dict[str, Any]\n final_report: Optional[str]\n current_category_index: int\n current_task_index_in_category: int\n stop_requested: bool\n error_message: Optional[str]\n messages: List[BaseMessage]\n\n\n# --- Langgraph Nodes ---\n\n\ndef _load_previous_state(task_id: str, output_dir: str) -> Dict[str, Any]:\n state_updates = {}\n plan_file = os.path.join(output_dir, PLAN_FILENAME)\n search_file = os.path.join(output_dir, SEARCH_INFO_FILENAME)\n\n loaded_plan: List[ResearchCategoryItem] = []\n next_cat_idx, next_task_idx = 0, 0\n found_pending = False\n\n ", "suffix_code": "\n\n if os.path.exists(search_file):\n try:\n with open(search_file, \"r\", encoding=\"utf-8\") as f:\n state_updates[\"search_results\"] = json.load(f)\n logger.info(f\"Loaded search results from {search_file}\")\n except Exception as e:\n logger.error(f\"Failed to load search results {search_file}: {e}\")\n state_updates[\"error_message\"] = (\n state_updates.get(\"error_message\", \"\") + f\" Failed to load search results: {e}\").strip()\n\n return state_updates\n\n\ndef _save_plan_to_md(plan: List[ResearchCategoryItem], output_dir: str):\n plan_file = os.path.join(output_dir, PLAN_FILENAME)\n try:\n with open(plan_file, \"w\", encoding=\"utf-8\") as f:\n f.write(f\"# Research Plan\\n\\n\")\n for cat_idx, category in enumerate(plan):\n f.write(f\"## {cat_idx + 1}. {category['category_name']}\\n\\n\")\n for task_idx, task in enumerate(category['tasks']):\n marker = \"- [x]\" if task[\"status\"] == \"completed\" else \"- [ ]\" if task[\n \"status\"] == \"pending\" else \"- [-]\" # [-] for failed\n f.write(f\" {marker} {task['task_description']}\\n\")\n f.write(\"\\n\")\n logger.info(f\"Hierarchical research plan saved to {plan_file}\")\n except Exception as e:\n logger.error(f\"Failed to save research plan to {plan_file}: {e}\")\n\n\ndef _save_search_results_to_json(results: List[Dict[str, Any]], output_dir: str):\n \"\"\"Appends or overwrites search results to a JSON file.\"\"\"\n search_file = os.path.join(output_dir, SEARCH_INFO_FILENAME)\n try:\n # Simple overwrite for now, could be append\n with open(search_file, \"w\", encoding=\"utf-8\") as f:\n json.dump(results, f, indent=2, ensure_ascii=False)\n logger.info(f\"Search results saved to {search_file}\")\n except Exception as e:\n logger.error(f\"Failed to save search results to {search_file}: {e}\")\n\n\ndef _save_report_to_md(report: str, output_dir: Path):\n \"\"\"Saves the final report to a markdown file.\"\"\"\n report_file = os.path.join(output_dir, REPORT_FILENAME)\n try:\n with open(report_file, \"w\", encoding=\"utf-8\") as f:\n f.write(report)\n logger.info(f\"Final report saved to {report_file}\")\n except Exception as e:\n logger.error(f\"Failed to save final report to {report_file}: {e}\")\n\n\nasync def planning_node(state: DeepResearchState) -> Dict[str, Any]:\n logger.info(\"--- Entering Planning Node ---\")\n if state.get(\"stop_requested\"):\n logger.info(\"Stop requested, skipping planning.\")\n return {\"stop_requested\": True}\n\n llm = state[\"llm\"]\n topic = state[\"topic\"]\n existing_plan = state.get(\"research_plan\")\n output_dir = state[\"output_dir\"]\n\n if existing_plan and (\n state.get(\"current_category_index\", 0) > 0 or state.get(\"current_task_index_in_category\", 0) > 0):\n logger.info(\"Resuming with existing plan.\")\n _save_plan_to_md(existing_plan, output_dir) # Ensure it's saved initially\n # current_category_index and current_task_index_in_category should be set by _load_previous_state\n return {\"research_plan\": existing_plan}\n\n logger.info(f\"Generating new research plan for topic: {topic}\")\n\n prompt_text = f\"\"\"You are a meticulous research assistant. Your goal is to create a hierarchical research plan to thoroughly investigate the topic: \"{topic}\".\nThe plan should be structured into several main research categories. Each category should contain a list of specific, actionable research tasks or questions.\nFormat the output as a JSON list of objects. Each object represents a research category and should have:\n1. \"category_name\": A string for the name of the research category.\n2. \"tasks\": A list of strings, where each string is a specific research task for that category.\n\nExample JSON Output:\n[\n {{\n \"category_name\": \"Understanding Core Concepts and Definitions\",\n \"tasks\": [\n \"Define the primary terminology associated with '{topic}'.\",\n \"Identify the fundamental principles and theories underpinning '{topic}'.\"\n ]\n }},\n {{\n \"category_name\": \"Historical Development and Key Milestones\",\n \"tasks\": [\n \"Trace the historical evolution of '{topic}'.\",\n \"Identify key figures, events, or breakthroughs in the development of '{topic}'.\"\n ]\n }},\n {{\n \"category_name\": \"Current State-of-the-Art and Applications\",\n \"tasks\": [\n \"Analyze the current advancements and prominent applications of '{topic}'.\",\n \"Investigate ongoing research and active areas of development related to '{topic}'.\"\n ]\n }},\n {{\n \"category_name\": \"Challenges, Limitations, and Future Outlook\",\n \"tasks\": [\n \"Identify the major challenges and limitations currently facing '{topic}'.\",\n \"Explore potential future trends, ethical considerations, and societal impacts of '{topic}'.\"\n ]\n }}\n]\n\nGenerate a plan with 3-10 categories, and 2-6 tasks per category for the topic: \"{topic}\" according to the complexity of the topic.\nEnsure the output is a valid JSON array.\n\"\"\"\n messages = [\n SystemMessage(content=\"You are a research planning assistant outputting JSON.\"),\n HumanMessage(content=prompt_text)\n ]\n\n try:\n response = await llm.ainvoke(messages)\n raw_content = response.content\n # The LLM might wrap the JSON in backticks\n if raw_content.strip().startswith(\"```json\"):\n raw_content = raw_content.strip()[7:-3].strip()\n elif raw_content.strip().startswith(\"```\"):\n raw_content = raw_content.strip()[3:-3].strip()\n\n logger.debug(f\"LLM response for plan: {raw_content}\")\n parsed_plan_from_llm = json.loads(raw_content)\n\n new_plan: List[ResearchCategoryItem] = []\n for cat_idx, category_data in enumerate(parsed_plan_from_llm):\n if not isinstance(category_data,\n dict) or \"category_name\" not in category_data or \"tasks\" not in category_data:\n logger.warning(f\"Skipping invalid category data: {category_data}\")\n continue\n\n tasks: List[ResearchTaskItem] = []\n for task_idx, task_desc in enumerate(category_data[\"tasks\"]):\n if isinstance(task_desc, str):\n tasks.append(\n ResearchTaskItem(\n task_description=task_desc,\n status=\"pending\",\n queries=None,\n result_summary=None,\n )\n )\n else: # Sometimes LLM puts tasks as {\"task\": \"description\"}\n if isinstance(task_desc, dict) and \"task_description\" in task_desc:\n tasks.append(\n ResearchTaskItem(\n task_description=task_desc[\"task_description\"],\n status=\"pending\",\n queries=None,\n result_summary=None,\n )\n )\n elif isinstance(task_desc, dict) and \"task\" in task_desc: # common LLM mistake\n tasks.append(\n ResearchTaskItem(\n task_description=task_desc[\"task\"],\n status=\"pending\",\n queries=None,\n result_summary=None,\n )\n )\n else:\n logger.warning(\n f\"Skipping invalid task data: {task_desc} in category {category_data['category_name']}\")\n\n new_plan.append(\n ResearchCategoryItem(\n category_name=category_data[\"category_name\"],\n tasks=tasks,\n )\n )\n\n if not new_plan:\n logger.error(\"LLM failed to generate a valid plan structure from JSON.\")\n return {\"error_message\": \"Failed to generate research plan structure.\"}\n\n logger.info(f\"Generated research plan with {len(new_plan)} categories.\")\n _save_plan_to_md(new_plan, output_dir) # Save the hierarchical plan\n\n return {\n \"research_plan\": new_plan,\n \"current_category_index\": 0,\n \"current_task_index_in_category\": 0,\n \"search_results\": [],\n }\n\n except json.JSONDecodeError as e:\n logger.error(f\"Failed to parse JSON from LLM for plan: {e}. Response was: {raw_content}\", exc_info=True)\n return {\"error_message\": f\"LLM generated invalid JSON for research plan: {e}\"}\n except Exception as e:\n logger.error(f\"Error during planning: {e}\", exc_info=True)\n return {\"error_message\": f\"LLM Error during planning: {e}\"}\n\n\nasync def research_execution_node(state: DeepResearchState) -> Dict[str, Any]:\n logger.info(\"--- Entering Research Execution Node ---\")\n if state.get(\"stop_requested\"):\n logger.info(\"Stop requested, skipping research execution.\")\n return {\n \"stop_requested\": True,\n \"current_category_index\": state[\"current_category_index\"],\n \"current_task_index_in_category\": state[\"current_task_index_in_category\"],\n }\n\n plan = state[\"research_plan\"]\n cat_idx = state[\"current_category_index\"]\n task_idx = state[\"current_task_index_in_category\"]\n llm = state[\"llm\"]\n tools = state[\"tools\"]\n output_dir = str(state[\"output_dir\"])\n task_id = state[\"task_id\"] # For _AGENT_STOP_FLAGS\n\n # This check should ideally be handled by `should_continue`\n if not plan or cat_idx >= len(plan):\n logger.info(\"Research plan complete or categories exhausted.\")\n return {} # should route to synthesis\n\n current_category = plan[cat_idx]\n if task_idx >= len(current_category[\"tasks\"]):\n logger.info(f\"All tasks in category '{current_category['category_name']}' completed. Moving to next category.\")\n # This logic is now effectively handled by should_continue and the index updates below\n # The next iteration will be caught by should_continue or this node with updated indices\n return {\n \"current_category_index\": cat_idx + 1,\n \"current_task_index_in_category\": 0,\n \"messages\": state[\"messages\"] # Pass messages along\n }\n\n current_task = current_category[\"tasks\"][task_idx]\n\n if current_task[\"status\"] == \"completed\":\n logger.info(\n f\"Task '{current_task['task_description']}' in category '{current_category['category_name']}' already completed. Skipping.\")\n # Logic to find next task\n next_task_idx = task_idx + 1\n next_cat_idx = cat_idx\n if next_task_idx >= len(current_category[\"tasks\"]):\n next_cat_idx += 1\n next_task_idx = 0\n return {\n \"current_category_index\": next_cat_idx,\n \"current_task_index_in_category\": next_task_idx,\n \"messages\": state[\"messages\"] # Pass messages along\n }\n\n logger.info(\n f\"Executing research task: '{current_task['task_description']}' (Category: '{current_category['category_name']}')\"\n )\n\n llm_with_tools = llm.bind_tools(tools)\n\n # Construct messages for LLM invocation\n task_prompt_content = (\n f\"Current Research Category: {current_category['category_name']}\\n\"\n f\"Specific Task: {current_task['task_description']}\\n\\n\"\n \"Please use the available tools, especially 'parallel_browser_search', to gather information for this specific task. \"\n \"Provide focused search queries relevant ONLY to this task. \"\n \"If you believe you have sufficient information from previous steps for this specific task, you can indicate that you are ready to summarize or that no further search is needed.\"\n )\n current_task_message_history = [\n HumanMessage(content=task_prompt_content)\n ]\n if not state[\"messages\"]: # First actual execution message\n invocation_messages = [\n SystemMessage(\n content=\"You are a research assistant executing one task of a research plan. Focus on the current task only.\"),\n ] + current_task_message_history\n else:\n invocation_messages = state[\"messages\"] + current_task_message_history\n\n try:\n logger.info(f\"Invoking LLM with tools for task: {current_task['task_description']}\")\n ai_response: BaseMessage = await llm_with_tools.ainvoke(invocation_messages)\n logger.info(\"LLM invocation complete.\")\n\n tool_results = []\n executed_tool_names = []\n current_search_results = state.get(\"search_results\", []) # Get existing search results\n\n if not isinstance(ai_response, AIMessage) or not ai_response.tool_calls:\n logger.warning(\n f\"LLM did not call any tool for task '{current_task['task_description']}'. Response: {ai_response.content[:100]}...\"\n )\n current_task[\"status\"] = \"pending\" # Or \"completed_no_tool\" if LLM explains it's done\n current_task[\"result_summary\"] = f\"LLM did not use a tool. Response: {ai_response.content}\"\n current_task[\"current_category_index\"] = cat_idx\n current_task[\"current_task_index_in_category\"] = task_idx\n return current_task\n # We still save the plan and advance.\n else:\n # Process tool calls\n for tool_call in ai_response.tool_calls:\n tool_name = tool_call.get(\"name\")\n tool_args = tool_call.get(\"args\", {})\n tool_call_id = tool_call.get(\"id\")\n\n logger.info(f\"LLM requested tool call: {tool_name} with args: {tool_args}\")\n executed_tool_names.append(tool_name)\n selected_tool = next((t for t in tools if t.name == tool_name), None)\n\n if not selected_tool:\n logger.error(f\"LLM called tool '{tool_name}' which is not available.\")\n tool_results.append(\n ToolMessage(content=f\"Error: Tool '{tool_name}' not found.\", tool_call_id=tool_call_id))\n continue\n\n try:\n stop_event = _AGENT_STOP_FLAGS.get(task_id)\n if stop_event and stop_event.is_set():\n logger.info(f\"Stop requested before executing tool: {tool_name}\")\n current_task[\"status\"] = \"pending\" # Or a new \"stopped\" status\n _save_plan_to_md(plan, output_dir)\n return {\"stop_requested\": True, \"research_plan\": plan, \"current_category_index\": cat_idx,\n \"current_task_index_in_category\": task_idx}\n\n logger.info(f\"Executing tool: {tool_name}\")\n tool_output = await selected_tool.ainvoke(tool_args)\n logger.info(f\"Tool '{tool_name}' executed successfully.\")\n\n if tool_name == \"parallel_browser_search\":\n current_search_results.extend(tool_output) # tool_output is List[Dict]\n else: # For other tools, we might need specific handling or just log\n logger.info(f\"Result from tool '{tool_name}': {str(tool_output)[:200]}...\")\n # Storing non-browser results might need a different structure or key in search_results\n current_search_results.append(\n {\"tool_name\": tool_name, \"args\": tool_args, \"output\": str(tool_output),\n \"status\": \"completed\"})\n\n tool_results.append(ToolMessage(content=json.dumps(tool_output), tool_call_id=tool_call_id))\n\n except Exception as e:\n logger.error(f\"Error executing tool '{tool_name}': {e}\", exc_info=True)\n tool_results.append(\n ToolMessage(content=f\"Error executing tool {tool_name}: {e}\", tool_call_id=tool_call_id))\n current_search_results.append(\n {\"tool_name\": tool_name, \"args\": tool_args, \"status\": \"failed\", \"error\": str(e)})\n\n # After processing all tool calls for this task\n step_failed_tool_execution = any(\"Error:\" in str(tr.content) for tr in tool_results)\n # Consider a task successful if a browser search was attempted and didn't immediately error out during call\n # The browser search itself returns status for each query.\n browser_tool_attempted_successfully = \"parallel_browser_search\" in executed_tool_names and not step_failed_tool_execution\n\n if step_failed_tool_execution:\n current_task[\"status\"] = \"failed\"\n current_task[\n \"result_summary\"] = f\"Tool execution failed. Errors: {[tr.content for tr in tool_results if 'Error' in str(tr.content)]}\"\n elif executed_tool_names: # If any tool was called\n current_task[\"status\"] = \"completed\"\n current_task[\"result_summary\"] = f\"Executed tool(s): {', '.join(executed_tool_names)}.\"\n # TODO: Could ask LLM to summarize the tool_results for this task if needed, rather than just listing tools.\n else: # No tool calls but AI response had .tool_calls structure (empty)\n current_task[\"status\"] = \"failed\" # Or a more specific status\n current_task[\"result_summary\"] = \"LLM prepared for tool call but provided no tools.\"\n\n # Save progress\n _save_plan_to_md(plan, output_dir)\n _save_search_results_to_json(current_search_results, output_dir)\n\n # Determine next indices\n next_task_idx = task_idx + 1\n next_cat_idx = cat_idx\n if next_task_idx >= len(current_category[\"tasks\"]):\n next_cat_idx += 1\n next_task_idx = 0\n\n updated_messages = state[\"messages\"] + current_task_message_history + [ai_response] + tool_results\n\n return {\n \"research_plan\": plan,\n \"search_results\": current_search_results,\n \"current_category_index\": next_cat_idx,\n \"current_task_index_in_category\": next_task_idx,\n \"messages\": updated_messages,\n }\n\n except Exception as e:\n logger.error(f\"Unhandled error during research execution for task '{current_task['task_description']}': {e}\",\n exc_info=True)\n current_task[\"status\"] = \"failed\"\n _save_plan_to_md(plan, output_dir)\n # Determine next indices even on error to attempt to move on\n next_task_idx = task_idx + 1\n next_cat_idx = cat_idx\n if next_task_idx >= len(current_category[\"tasks\"]):\n next_cat_idx += 1\n next_task_idx = 0\n return {\n \"research_plan\": plan,\n \"current_category_index\": next_cat_idx,\n \"current_task_index_in_category\": next_task_idx,\n \"error_message\": f\"Core Execution Error on task '{current_task['task_description']}': {e}\",\n \"messages\": state[\"messages\"] + current_task_message_history # Preserve messages up to error\n }\n\n\nasync def synthesis_node(state: DeepResearchState) -> Dict[str, Any]:\n \"\"\"Synthesizes the final report from the collected search results.\"\"\"\n logger.info(\"--- Entering Synthesis Node ---\")\n if state.get(\"stop_requested\"):\n logger.info(\"Stop requested, skipping synthesis.\")\n return {\"stop_requested\": True}\n\n llm = state[\"llm\"]\n topic = state[\"topic\"]\n search_results = state.get(\"search_results\", [])\n output_dir = state[\"output_dir\"]\n plan = state[\"research_plan\"] # Include plan for context\n\n if not search_results:\n logger.warning(\"No search results found to synthesize report.\")\n report = f\"# Research Report: {topic}\\n\\nNo information was gathered during the research process.\"\n _save_report_to_md(report, output_dir)\n return {\"final_report\": report}\n\n logger.info(\n f\"Synthesizing report from {len(search_results)} collected search result entries.\"\n )\n\n # Prepare context for the LLM\n # Format search results nicely, maybe group by query or original plan step\n formatted_results = \"\"\n references = {}\n ref_count = 1\n for i, result_entry in enumerate(search_results):\n query = result_entry.get(\"query\", \"Unknown Query\") # From parallel_browser_search\n tool_name = result_entry.get(\"tool_name\") # From other tools\n status = result_entry.get(\"status\", \"unknown\")\n result_data = result_entry.get(\"result\") # From BrowserUseAgent's final_result\n tool_output_str = result_entry.get(\"output\") # From other tools\n\n if tool_name == \"parallel_browser_search\" and status == \"completed\" and result_data:\n # result_data is the summary from BrowserUseAgent\n formatted_results += f'### Finding from Web Search Query: \"{query}\"\\n'\n formatted_results += f\"- **Summary:**\\n{result_data}\\n\" # result_data is already a summary string here\n # If result_data contained title/URL, you'd format them here.\n # The current BrowserUseAgent returns a string summary directly as 'final_data' in run_single_browser_task\n formatted_results += \"---\\n\"\n elif tool_name != \"parallel_browser_search\" and status == \"completed\" and tool_output_str:\n formatted_results += f'### Finding from Tool: \"{tool_name}\" (Args: {result_entry.get(\"args\")})\\n'\n formatted_results += f\"- **Output:**\\n{tool_output_str}\\n\"\n formatted_results += \"---\\n\"\n elif status == \"failed\":\n error = result_entry.get(\"error\")\n q_or_t = f\"Query: \\\"{query}\\\"\" if query != \"Unknown Query\" else f\"Tool: \\\"{tool_name}\\\"\"\n formatted_results += f'### Failed {q_or_t}\\n'\n formatted_results += f\"- **Error:** {error}\\n\"\n formatted_results += \"---\\n\"\n\n # Prepare the research plan context\n plan_summary = \"\\nResearch Plan Followed:\\n\"\n for cat_idx, category in enumerate(plan):\n plan_summary += f\"\\n#### Category {cat_idx + 1}: {category['category_name']}\\n\"\n for task_idx, task in enumerate(category['tasks']):\n marker = \"[x]\" if task[\"status\"] == \"completed\" else \"[ ]\" if task[\"status\"] == \"pending\" else \"[-]\"\n plan_summary += f\" - {marker} {task['task_description']}\\n\"\n\n synthesis_prompt = ChatPromptTemplate.from_messages(\n [\n (\n \"system\",\n \"\"\"You are a professional researcher tasked with writing a comprehensive and well-structured report based on collected findings.\n The report should address the research topic thoroughly, synthesizing the information gathered from various sources.\n Structure the report logically:\n 1. Briefly introduce the topic and the report's scope (mentioning the research plan followed, including categories and tasks, is good).\n 2. Discuss the key findings, organizing them thematically, possibly aligning with the research categories. Analyze, compare, and contrast information.\n 3. Summarize the main points and offer concluding thoughts.\n\n Ensure the tone is objective and professional.\n If findings are contradictory or incomplete, acknowledge this.\n \"\"\", # Removed citation part for simplicity for now, as browser agent returns summaries.\n ),\n (\n \"human\",\n f\"\"\"\n **Research Topic:** {topic}\n\n {plan_summary}\n\n **Collected Findings:**\n ```\n {formatted_results}\n ```\n\n Please generate the final research report in Markdown format based **only** on the information above.\n \"\"\",\n ),\n ]\n )\n\n try:\n response = await llm.ainvoke(\n synthesis_prompt.format_prompt(\n topic=topic,\n plan_summary=plan_summary,\n formatted_results=formatted_results,\n ).to_messages()\n )\n final_report_md = response.content\n\n # Append the reference list automatically to the end of the generated markdown\n if references:\n report_references_section = \"\\n\\n## References\\n\\n\"\n # Sort refs by ID for consistent output\n sorted_refs = sorted(references.values(), key=lambda x: x[\"id\"])\n for ref in sorted_refs:\n report_references_section += (\n f\"[{ref['id']}] {ref['title']} - {ref['url']}\\n\"\n )\n final_report_md += report_references_section\n\n logger.info(\"Successfully synthesized the final report.\")\n _save_report_to_md(final_report_md, output_dir)\n return {\"final_report\": final_report_md}\n\n except Exception as e:\n logger.error(f\"Error during report synthesis: {e}\", exc_info=True)\n return {\"error_message\": f\"LLM Error during synthesis: {e}\"}\n\n\n# --- Langgraph Edges and Conditional Logic ---\n\n\ndef should_continue(state: DeepResearchState) -> str:\n logger.info(\"--- Evaluating Condition: Should Continue? ---\")\n if state.get(\"stop_requested\"):\n logger.info(\"Stop requested, routing to END.\")\n return \"end_run\"\n if state.get(\"error_message\") and \"Core Execution Error\" in state[\"error_message\"]: # Critical error in node\n logger.warning(f\"Critical error detected: {state['error_message']}. Routing to END.\")\n return \"end_run\"\n\n plan = state.get(\"research_plan\")\n cat_idx = state.get(\"current_category_index\", 0)\n task_idx = state.get(\"current_task_index_in_category\", 0) # This is the *next* task to check\n\n if not plan:\n logger.warning(\"No research plan found. Routing to END.\")\n return \"end_run\"\n\n # Check if the current indices point to a valid pending task\n if cat_idx < len(plan):\n current_category = plan[cat_idx]\n if task_idx < len(current_category[\"tasks\"]):\n # We are trying to execute the task at plan[cat_idx][\"tasks\"][task_idx]\n # The research_execution_node will handle if it's already completed.\n logger.info(\n f\"Plan has potential pending tasks (next up: Category {cat_idx}, Task {task_idx}). Routing to Research Execution.\"\n )\n return \"execute_research\"\n else: # task_idx is out of bounds for current category, means we need to check next category\n if cat_idx + 1 < len(plan): # If there is a next category\n logger.info(\n f\"Finished tasks in category {cat_idx}. Moving to category {cat_idx + 1}. Routing to Research Execution.\"\n )\n # research_execution_node will update state to {current_category_index: cat_idx + 1, current_task_index_in_category: 0}\n # Or rather, the previous execution node already set these indices to the start of the next category.\n return \"execute_research\"\n\n # If we've gone through all categories and tasks (cat_idx >= len(plan))\n logger.info(\"All plan categories and tasks processed or current indices are out of bounds. Routing to Synthesis.\")\n return \"synthesize_report\"\n\n\n# --- DeepSearchAgent Class ---\n\n\nclass DeepResearchAgent:\n def __init__(\n self,\n llm: Any,\n browser_config: Dict[str, Any],\n mcp_server_config: Optional[Dict[str, Any]] = None,\n ):\n \"\"\"\n Initializes the DeepSearchAgent.\n\n Args:\n llm: The Langchain compatible language model instance.\n browser_config: Configuration dictionary for the BrowserUseAgent tool.\n Example: {\"headless\": True, \"window_width\": 1280, ...}\n mcp_server_config: Optional configuration for the MCP client.\n \"\"\"\n self.llm = llm\n self.browser_config = browser_config\n self.mcp_server_config = mcp_server_config\n self.mcp_client = None\n self.stopped = False\n self.graph = self._compile_graph()\n self.current_task_id: Optional[str] = None\n self.stop_event: Optional[threading.Event] = None\n self.runner: Optional[asyncio.Task] = None # To hold the asyncio task for run\n\n async def _setup_tools(\n self, task_id: str, stop_event: threading.Event, max_parallel_browsers: int = 1\n ) -> List[Tool]:\n \"\"\"Sets up the basic tools (File I/O) and optional MCP tools.\"\"\"\n tools = [\n WriteFileTool(),\n ReadFileTool(),\n ListDirectoryTool(),\n ] # Basic file operations\n browser_use_tool = create_browser_search_tool(\n llm=self.llm,\n browser_config=self.browser_config,\n task_id=task_id,\n stop_event=stop_event,\n max_parallel_browsers=max_parallel_browsers,\n )\n tools += [browser_use_tool]\n # Add MCP tools if config is provided\n if self.mcp_server_config:\n try:\n logger.info(\"Setting up MCP client and tools...\")\n if not self.mcp_client:\n self.mcp_client = await setup_mcp_client_and_tools(\n self.mcp_server_config\n )\n mcp_tools = self.mcp_client.get_tools()\n logger.info(f\"Loaded {len(mcp_tools)} MCP tools.\")\n tools.extend(mcp_tools)\n except Exception as e:\n logger.error(f\"Failed to set up MCP tools: {e}\", exc_info=True)\n elif self.mcp_server_config:\n logger.warning(\n \"MCP server config provided, but setup function unavailable.\"\n )\n tools_map = {tool.name: tool for tool in tools}\n return tools_map.values()\n\n async def close_mcp_client(self):\n if self.mcp_client:\n await self.mcp_client.__aexit__(None, None, None)\n self.mcp_client = None\n\n def _compile_graph(self) -> StateGraph:\n \"\"\"Compiles the Langgraph state machine.\"\"\"\n workflow = StateGraph(DeepResearchState)\n\n # Add nodes\n workflow.add_node(\"plan_research\", planning_node)\n workflow.add_node(\"execute_research\", research_execution_node)\n workflow.add_node(\"synthesize_report\", synthesis_node)\n workflow.add_node(\n \"end_run\", lambda state: logger.info(\"--- Reached End Run Node ---\") or {}\n ) # Simple end node\n\n # Define edges\n workflow.set_entry_point(\"plan_research\")\n\n workflow.add_edge(\n \"plan_research\", \"execute_research\"\n ) # Always execute after planning\n\n # Conditional edge after execution\n workflow.add_conditional_edges(\n \"execute_research\",\n should_continue,\n {\n \"execute_research\": \"execute_research\", # Loop back if more steps\n \"synthesize_report\": \"synthesize_report\", # Move to synthesis if done\n \"end_run\": \"end_run\", # End if stop requested or error\n },\n )\n\n workflow.add_edge(\"synthesize_report\", \"end_run\") # End after synthesis\n\n app = workflow.compile()\n return app\n\n async def run(\n self,\n topic: str,\n task_id: Optional[str] = None,\n save_dir: str = \"./tmp/deep_research\",\n max_parallel_browsers: int = 1,\n ) -> Dict[str, Any]:\n \"\"\"\n Starts the deep research process (Async Generator Version).\n\n Args:\n topic: The research topic.\n task_id: Optional existing task ID to resume. If None, a new ID is generated.\n\n Yields:\n Intermediate state updates or messages during execution.\n \"\"\"\n if self.runner and not self.runner.done():\n logger.warning(\n \"Agent is already running. Please stop the current task first.\"\n )\n # Return an error status instead of yielding\n return {\n \"status\": \"error\",\n \"message\": \"Agent already running.\",\n \"task_id\": self.current_task_id,\n }\n\n self.current_task_id = task_id if task_id else str(uuid.uuid4())\n safe_root_dir = \"./tmp/deep_research\"\n normalized_save_dir = os.path.normpath(save_dir)\n if not normalized_save_dir.startswith(os.path.abspath(safe_root_dir)):\n logger.warning(f\"Unsafe save_dir detected: {save_dir}. Using default directory.\")\n normalized_save_dir = os.path.abspath(safe_root_dir)\n output_dir = os.path.join(normalized_save_dir, self.current_task_id)\n os.makedirs(output_dir, exist_ok=True)\n\n logger.info(\n f\"[AsyncGen] Starting research task ID: {self.current_task_id} for topic: '{topic}'\"\n )\n logger.info(f\"[AsyncGen] Output directory: {output_dir}\")\n\n self.stop_event = threading.Event()\n _AGENT_STOP_FLAGS[self.current_task_id] = self.stop_event\n agent_tools = await self._setup_tools(\n self.current_task_id, self.stop_event, max_parallel_browsers\n )\n initial_state: DeepResearchState = {\n \"task_id\": self.current_task_id,\n \"topic\": topic,\n \"research_plan\": [],\n \"search_results\": [],\n \"messages\": [],\n \"llm\": self.llm,\n \"tools\": agent_tools,\n \"output_dir\": Path(output_dir),\n \"browser_config\": self.browser_config,\n \"final_report\": None,\n \"current_category_index\": 0,\n \"current_task_index_in_category\": 0,\n \"stop_requested\": False,\n \"error_message\": None,\n }\n\n if task_id:\n logger.info(f\"Attempting to resume task {task_id}...\")\n loaded_state = _load_previous_state(task_id, output_dir)\n initial_state.update(loaded_state)\n if loaded_state.get(\"research_plan\"):\n logger.info(\n f\"Resuming with {len(loaded_state['research_plan'])} plan categories \"\n f\"and {len(loaded_state.get('search_results', []))} existing results. \"\n f\"Next task: Cat {initial_state['current_category_index']}, Task {initial_state['current_task_index_in_category']}\"\n )\n initial_state[\"topic\"] = (\n topic # Allow overriding topic even when resuming? Or use stored topic? Let's use new one.\n )\n else:\n logger.warning(\n f\"Resume requested for {task_id}, but no previous plan found. Starting fresh.\"\n )\n\n # --- Execute Graph using ainvoke ---\n final_state = None\n status = \"unknown\"\n message = None\n try:\n logger.info(f\"Invoking graph execution for task {self.current_task_id}...\")\n self.runner = asyncio.create_task(self.graph.ainvoke(initial_state))\n final_state = await self.runner\n logger.info(f\"Graph execution finished for task {self.current_task_id}.\")\n\n # Determine status based on final state\n if self.stop_event and self.stop_event.is_set():\n status = \"stopped\"\n message = \"Research process was stopped by request.\"\n logger.info(message)\n elif final_state and final_state.get(\"error_message\"):\n status = \"error\"\n message = final_state[\"error_message\"]\n logger.error(f\"Graph execution completed with error: {message}\")\n elif final_state and final_state.get(\"final_report\"):\n status = \"completed\"\n message = \"Research process completed successfully.\"\n logger.info(message)\n else:\n # If it ends without error/report (e.g., empty plan, stopped before synthesis)\n status = \"finished_incomplete\"\n message = \"Research process finished, but may be incomplete (no final report generated).\"\n logger.warning(message)\n\n except asyncio.CancelledError:\n status = \"cancelled\"\n message = f\"Agent run task cancelled for {self.current_task_id}.\"\n logger.info(message)\n # final_state will remain None or the state before cancellation if checkpointing was used\n except Exception as e:\n status = \"error\"\n message = f\"Unhandled error during graph execution for {self.current_task_id}: {e}\"\n logger.error(message, exc_info=True)\n # final_state will remain None or the state before the error\n finally:\n logger.info(f\"Cleaning up resources for task {self.current_task_id}\")\n task_id_to_clean = self.current_task_id\n\n self.stop_event = None\n self.current_task_id = None\n self.runner = None # Mark runner as finished\n if self.mcp_client:\n await self.mcp_client.__aexit__(None, None, None)\n\n # Return a result dictionary including the status and the final state if available\n return {\n \"status\": status,\n \"message\": message,\n \"task_id\": task_id_to_clean, # Use the stored task_id\n \"final_state\": final_state\n if final_state\n else {}, # Return the final state dict\n }\n\n async def _stop_lingering_browsers(self, task_id):\n \"\"\"Attempts to stop any BrowserUseAgent instances associated with the task_id.\"\"\"\n keys_to_stop = [\n key for key in _BROWSER_AGENT_INSTANCES if key.startswith(f\"{task_id}_\")\n ]\n if not keys_to_stop:\n return\n\n logger.warning(\n f\"Found {len(keys_to_stop)} potentially lingering browser agents for task {task_id}. Attempting stop...\"\n )\n for key in keys_to_stop:\n agent_instance = _BROWSER_AGENT_INSTANCES.get(key)\n try:\n if agent_instance:\n # Assuming BU agent has an async stop method\n await agent_instance.stop()\n logger.info(f\"Called stop() on browser agent instance {key}\")\n except Exception as e:\n logger.error(\n f\"Error calling stop() on browser agent instance {key}: {e}\"\n )\n\n async def stop(self):\n \"\"\"Signals the currently running agent task to stop.\"\"\"\n if not self.current_task_id or not self.stop_event:\n logger.info(\"No agent task is currently running.\")\n return\n\n logger.info(f\"Stop requested for task ID: {self.current_task_id}\")\n self.stop_event.set() # Signal the stop event\n self.stopped = True\n await self._stop_lingering_browsers(self.current_task_id)\n\n def close(self):\n self.stopped = False\n", "middle_code": "if os.path.exists(plan_file):\n try:\n with open(plan_file, \"r\", encoding=\"utf-8\") as f:\n current_category: Optional[ResearchCategoryItem] = None\n lines = f.readlines()\n cat_counter = 0\n task_counter_in_cat = 0\n for line_num, line_content in enumerate(lines):\n line = line_content.strip()\n if line.startswith(\"\n if current_category: \n loaded_plan.append(current_category)\n if not found_pending: \n cat_counter += 1\n task_counter_in_cat = 0\n category_name = line[line.find(\" \"):].strip() \n current_category = ResearchCategoryItem(category_name=category_name, tasks=[])\n elif (line.startswith(\"- [ ]\") or line.startswith(\"- [x]\") or line.startswith(\n \"- [-]\")) and current_category: \n status = \"pending\"\n if line.startswith(\"- [x]\"):\n status = \"completed\"\n elif line.startswith(\"- [-]\"):\n status = \"failed\"\n task_desc = line[5:].strip()\n current_category[\"tasks\"].append(\n ResearchTaskItem(task_description=task_desc, status=status, queries=None,\n result_summary=None)\n )\n if status == \"pending\" and not found_pending:\n next_cat_idx = cat_counter\n next_task_idx = task_counter_in_cat\n found_pending = True\n if not found_pending: \n task_counter_in_cat += 1\n if current_category: \n loaded_plan.append(current_category)\n if loaded_plan:\n state_updates[\"research_plan\"] = loaded_plan\n if not found_pending and loaded_plan: \n next_cat_idx = len(loaded_plan) \n next_task_idx = 0\n state_updates[\"current_category_index\"] = next_cat_idx\n state_updates[\"current_task_index_in_category\"] = next_task_idx\n logger.info(\n f\"Loaded hierarchical research plan from {plan_file}. \"\n f\"Next task: Category {next_cat_idx}, Task {next_task_idx} in category.\"\n )\n else:\n logger.warning(f\"Plan file {plan_file} was empty or malformed.\")\n except Exception as e:\n logger.error(f\"Failed to load or parse research plan {plan_file}: {e}\", exc_info=True)\n state_updates[\"error_message\"] = f\"Failed to load research plan: {e}\"\n else:\n logger.info(f\"Plan file {plan_file} not found. Will start fresh.\")", "code_description": null, "fill_type": "BLOCK_TYPE", "language_type": "python", "sub_task_type": "if_statement"}, "context_code": [["/web-ui/src/webui/components/deep_research_agent_tab.py", "import gradio as gr\nfrom gradio.components import Component\nfrom functools import partial\n\nfrom src.webui.webui_manager import WebuiManager\nfrom src.utils import config\nimport logging\nimport os\nfrom typing import Any, Dict, AsyncGenerator, Optional, Tuple, Union\nimport asyncio\nimport json\nfrom src.agent.deep_research.deep_research_agent import DeepResearchAgent\nfrom src.utils import llm_provider\n\nlogger = logging.getLogger(__name__)\n\n\nasync def _initialize_llm(provider: Optional[str], model_name: Optional[str], temperature: float,\n base_url: Optional[str], api_key: Optional[str], num_ctx: Optional[int] = None):\n \"\"\"Initializes the LLM based on settings. Returns None if provider/model is missing.\"\"\"\n if not provider or not model_name:\n logger.info(\"LLM Provider or Model Name not specified, LLM will be None.\")\n return None\n try:\n logger.info(f\"Initializing LLM: Provider={provider}, Model={model_name}, Temp={temperature}\")\n # Use your actual LLM provider logic here\n llm = llm_provider.get_llm_model(\n provider=provider,\n model_name=model_name,\n temperature=temperature,\n base_url=base_url or None,\n api_key=api_key or None,\n num_ctx=num_ctx if provider == \"ollama\" else None\n )\n return llm\n except Exception as e:\n logger.error(f\"Failed to initialize LLM: {e}\", exc_info=True)\n gr.Warning(\n f\"Failed to initialize LLM '{model_name}' for provider '{provider}'. Please check settings. Error: {e}\")\n return None\n\n\ndef _read_file_safe(file_path: str) -> Optional[str]:\n \"\"\"Safely read a file, returning None if it doesn't exist or on error.\"\"\"\n if not os.path.exists(file_path):\n return None\n try:\n with open(file_path, 'r', encoding='utf-8') as f:\n return f.read()\n except Exception as e:\n logger.error(f\"Error reading file {file_path}: {e}\")\n return None\n\n\n# --- Deep Research Agent Specific Logic ---\n\nasync def run_deep_research(webui_manager: WebuiManager, components: Dict[Component, Any]) -> AsyncGenerator[\n Dict[Component, Any], None]:\n \"\"\"Handles initializing and running the DeepResearchAgent.\"\"\"\n\n # --- Get Components ---\n research_task_comp = webui_manager.get_component_by_id(\"deep_research_agent.research_task\")\n resume_task_id_comp = webui_manager.get_component_by_id(\"deep_research_agent.resume_task_id\")\n parallel_num_comp = webui_manager.get_component_by_id(\"deep_research_agent.parallel_num\")\n save_dir_comp = webui_manager.get_component_by_id(\n \"deep_research_agent.max_query\") # Note: component ID seems misnamed in original code\n start_button_comp = webui_manager.get_component_by_id(\"deep_research_agent.start_button\")\n stop_button_comp = webui_manager.get_component_by_id(\"deep_research_agent.stop_button\")\n markdown_display_comp = webui_manager.get_component_by_id(\"deep_research_agent.markdown_display\")\n markdown_download_comp = webui_manager.get_component_by_id(\"deep_research_agent.markdown_download\")\n mcp_server_config_comp = webui_manager.get_component_by_id(\"deep_research_agent.mcp_server_config\")\n\n # --- 1. Get Task and Settings ---\n task_topic = components.get(research_task_comp, \"\").strip()\n task_id_to_resume = components.get(resume_task_id_comp, \"\").strip() or None\n max_parallel_agents = int(components.get(parallel_num_comp, 1))\n base_save_dir = components.get(save_dir_comp, \"./tmp/deep_research\").strip()\n safe_root_dir = \"./tmp/deep_research\"\n normalized_base_save_dir = os.path.abspath(os.path.normpath(base_save_dir))\n if os.path.commonpath([normalized_base_save_dir, os.path.abspath(safe_root_dir)]) != os.path.abspath(safe_root_dir):\n logger.warning(f\"Unsafe base_save_dir detected: {base_save_dir}. Using default directory.\")\n normalized_base_save_dir = os.path.abspath(safe_root_dir)\n base_save_dir = normalized_base_save_dir\n mcp_server_config_str = components.get(mcp_server_config_comp)\n mcp_config = json.loads(mcp_server_config_str) if mcp_server_config_str else None\n\n if not task_topic:\n gr.Warning(\"Please enter a research task.\")\n yield {start_button_comp: gr.update(interactive=True)} # Re-enable start button\n return\n\n # Store base save dir for stop handler\n webui_manager.dr_save_dir = base_save_dir\n os.makedirs(base_save_dir, exist_ok=True)\n\n # --- 2. Initial UI Update ---\n yield {\n start_button_comp: gr.update(value=\"⏳ Running...\", interactive=False),\n stop_button_comp: gr.update(interactive=True),\n research_task_comp: gr.update(interactive=False),\n resume_task_id_comp: gr.update(interactive=False),\n parallel_num_comp: gr.update(interactive=False),\n save_dir_comp: gr.update(interactive=False),\n markdown_display_comp: gr.update(value=\"Starting research...\"),\n markdown_download_comp: gr.update(value=None, interactive=False)\n }\n\n agent_task = None\n running_task_id = None\n plan_file_path = None\n report_file_path = None\n last_plan_content = None\n last_plan_mtime = 0\n\n try:\n # --- 3. Get LLM and Browser Config from other tabs ---\n # Access settings values via components dict, getting IDs from webui_manager\n def get_setting(tab: str, key: str, default: Any = None):\n comp = webui_manager.id_to_component.get(f\"{tab}.{key}\")\n return components.get(comp, default) if comp else default\n\n # LLM Config (from agent_settings tab)\n llm_provider_name = get_setting(\"agent_settings\", \"llm_provider\")\n llm_model_name = get_setting(\"agent_settings\", \"llm_model_name\")\n llm_temperature = max(get_setting(\"agent_settings\", \"llm_temperature\", 0.5), 0.5)\n llm_base_url = get_setting(\"agent_settings\", \"llm_base_url\")\n llm_api_key = get_setting(\"agent_settings\", \"llm_api_key\")\n ollama_num_ctx = get_setting(\"agent_settings\", \"ollama_num_ctx\")\n\n llm = await _initialize_llm(\n llm_provider_name, llm_model_name, llm_temperature, llm_base_url, llm_api_key,\n ollama_num_ctx if llm_provider_name == \"ollama\" else None\n )\n if not llm:\n raise ValueError(\"LLM Initialization failed. Please check Agent Settings.\")\n\n # Browser Config (from browser_settings tab)\n # Note: DeepResearchAgent constructor takes a dict, not full Browser/Context objects\n browser_config_dict = {\n \"headless\": get_setting(\"browser_settings\", \"headless\", False),\n \"disable_security\": get_setting(\"browser_settings\", \"disable_security\", False),\n \"browser_binary_path\": get_setting(\"browser_settings\", \"browser_binary_path\"),\n \"user_data_dir\": get_setting(\"browser_settings\", \"browser_user_data_dir\"),\n \"window_width\": int(get_setting(\"browser_settings\", \"window_w\", 1280)),\n \"window_height\": int(get_setting(\"browser_settings\", \"window_h\", 1100)),\n # Add other relevant fields if DeepResearchAgent accepts them\n }\n\n # --- 4. Initialize or Get Agent ---\n if not webui_manager.dr_agent:\n webui_manager.dr_agent = DeepResearchAgent(\n llm=llm,\n browser_config=browser_config_dict,\n mcp_server_config=mcp_config\n )\n logger.info(\"DeepResearchAgent initialized.\")\n\n # --- 5. Start Agent Run ---\n agent_run_coro = webui_manager.dr_agent.run(\n topic=task_topic,\n task_id=task_id_to_resume,\n save_dir=base_save_dir,\n max_parallel_browsers=max_parallel_agents\n )\n agent_task = asyncio.create_task(agent_run_coro)\n webui_manager.dr_current_task = agent_task\n\n # Wait briefly for the agent to start and potentially create the task ID/folder\n await asyncio.sleep(1.0)\n\n # Determine the actual task ID being used (agent sets this)\n running_task_id = webui_manager.dr_agent.current_task_id\n if not running_task_id:\n # Agent might not have set it yet, try to get from result later? Risky.\n # Or derive from resume_task_id if provided?\n running_task_id = task_id_to_resume\n if not running_task_id:\n logger.warning(\"Could not determine running task ID immediately.\")\n # We can still monitor, but might miss initial plan if ID needed for path\n else:\n logger.info(f\"Assuming task ID based on resume ID: {running_task_id}\")\n else:\n logger.info(f\"Agent started with Task ID: {running_task_id}\")\n\n webui_manager.dr_task_id = running_task_id # Store for stop handler\n\n # --- 6. Monitor Progress via research_plan.md ---\n if running_task_id:\n task_specific_dir = os.path.join(base_save_dir, str(running_task_id))\n plan_file_path = os.path.join(task_specific_dir, \"research_plan.md\")\n report_file_path = os.path.join(task_specific_dir, \"report.md\")\n logger.info(f\"Monitoring plan file: {plan_file_path}\")\n else:\n logger.warning(\"Cannot monitor plan file: Task ID unknown.\")\n plan_file_path = None\n last_plan_content = None\n while not agent_task.done():\n update_dict = {}\n update_dict[resume_task_id_comp] = gr.update(value=running_task_id)\n agent_stopped = getattr(webui_manager.dr_agent, 'stopped', False)\n if agent_stopped:\n logger.info(\"Stop signal detected from agent state.\")\n break # Exit monitoring loop\n\n # Check and update research plan display\n if plan_file_path:\n try:\n current_mtime = os.path.getmtime(plan_file_path) if os.path.exists(plan_file_path) else 0\n if current_mtime > last_plan_mtime:\n logger.info(f\"Detected change in {plan_file_path}\")\n plan_content = _read_file_safe(plan_file_path)\n if last_plan_content is None or (\n plan_content is not None and plan_content != last_plan_content):\n update_dict[markdown_display_comp] = gr.update(value=plan_content)\n last_plan_content = plan_content\n last_plan_mtime = current_mtime\n elif plan_content is None:\n # File might have been deleted or became unreadable\n last_plan_mtime = 0 # Reset to force re-read attempt later\n except Exception as e:\n logger.warning(f\"Error checking/reading plan file {plan_file_path}: {e}\")\n # Avoid continuous logging for the same error\n await asyncio.sleep(2.0)\n\n # Yield updates if any\n if update_dict:\n yield update_dict\n\n await asyncio.sleep(1.0) # Check file changes every second\n\n # --- 7. Task Finalization ---\n logger.info(\"Agent task processing finished. Awaiting final result...\")\n final_result_dict = await agent_task # Get result or raise exception\n logger.info(f\"Agent run completed. Result keys: {final_result_dict.keys() if final_result_dict else 'None'}\")\n\n # Try to get task ID from result if not known before\n if not running_task_id and final_result_dict and 'task_id' in final_result_dict:\n running_task_id = final_result_dict['task_id']\n webui_manager.dr_task_id = running_task_id\n task_specific_dir = os.path.join(base_save_dir, str(running_task_id))\n report_file_path = os.path.join(task_specific_dir, \"report.md\")\n logger.info(f\"Task ID confirmed from result: {running_task_id}\")\n\n final_ui_update = {}\n if report_file_path and os.path.exists(report_file_path):\n logger.info(f\"Loading final report from: {report_file_path}\")\n report_content = _read_file_safe(report_file_path)\n if report_content:\n final_ui_update[markdown_display_comp] = gr.update(value=report_content)\n final_ui_update[markdown_download_comp] = gr.File(value=report_file_path,\n label=f\"Report ({running_task_id}.md)\",\n interactive=True)\n else:\n final_ui_update[markdown_display_comp] = gr.update(\n value=\"# Research Complete\\n\\n*Error reading final report file.*\")\n elif final_result_dict and 'report' in final_result_dict:\n logger.info(\"Using report content directly from agent result.\")\n # If agent directly returns report content\n final_ui_update[markdown_display_comp] = gr.update(value=final_result_dict['report'])\n # Cannot offer download if only content is available\n final_ui_update[markdown_download_comp] = gr.update(value=None, label=\"Download Research Report\",\n interactive=False)\n else:\n logger.warning(\"Final report file not found and not in result dict.\")\n final_ui_update[markdown_display_comp] = gr.update(value=\"# Research Complete\\n\\n*Final report not found.*\")\n\n yield final_ui_update\n\n\n except Exception as e:\n logger.error(f\"Error during Deep Research Agent execution: {e}\", exc_info=True)\n gr.Error(f\"Research failed: {e}\")\n yield {markdown_display_comp: gr.update(value=f\"# Research Failed\\n\\n**Error:**\\n```\\n{e}\\n```\")}\n\n finally:\n # --- 8. Final UI Reset ---\n webui_manager.dr_current_task = None # Clear task reference\n webui_manager.dr_task_id = None # Clear running task ID\n\n yield {\n start_button_comp: gr.update(value=\"▶️ Run\", interactive=True),\n stop_button_comp: gr.update(interactive=False),\n research_task_comp: gr.update(interactive=True),\n resume_task_id_comp: gr.update(value=\"\", interactive=True),\n parallel_num_comp: gr.update(interactive=True),\n save_dir_comp: gr.update(interactive=True),\n # Keep download button enabled if file exists\n markdown_download_comp: gr.update() if report_file_path and os.path.exists(report_file_path) else gr.update(\n interactive=False)\n }\n\n\nasync def stop_deep_research(webui_manager: WebuiManager) -> Dict[Component, Any]:\n \"\"\"Handles the Stop button click.\"\"\"\n logger.info(\"Stop button clicked for Deep Research.\")\n agent = webui_manager.dr_agent\n task = webui_manager.dr_current_task\n task_id = webui_manager.dr_task_id\n base_save_dir = webui_manager.dr_save_dir\n\n stop_button_comp = webui_manager.get_component_by_id(\"deep_research_agent.stop_button\")\n start_button_comp = webui_manager.get_component_by_id(\"deep_research_agent.start_button\")\n markdown_display_comp = webui_manager.get_component_by_id(\"deep_research_agent.markdown_display\")\n markdown_download_comp = webui_manager.get_component_by_id(\"deep_research_agent.markdown_download\")\n\n final_update = {\n stop_button_comp: gr.update(interactive=False, value=\"⏹️ Stopping...\")\n }\n\n if agent and task and not task.done():\n logger.info(\"Signalling DeepResearchAgent to stop.\")\n try:\n # Assuming stop is synchronous or sets a flag quickly\n await agent.stop()\n except Exception as e:\n logger.error(f\"Error calling agent.stop(): {e}\")\n\n # The run_deep_research loop should detect the stop and exit.\n # We yield an intermediate \"Stopping...\" state. The final reset is done by run_deep_research.\n\n # Try to show the final report if available after stopping\n await asyncio.sleep(1.5) # Give agent a moment to write final files potentially\n report_file_path = None\n if task_id and base_save_dir:\n report_file_path = os.path.join(base_save_dir, str(task_id), \"report.md\")\n\n if report_file_path and os.path.exists(report_file_path):\n report_content = _read_file_safe(report_file_path)\n if report_content:\n final_update[markdown_display_comp] = gr.update(\n value=report_content + \"\\n\\n---\\n*Research stopped by user.*\")\n final_update[markdown_download_comp] = gr.File(value=report_file_path, label=f\"Report ({task_id}.md)\",\n interactive=True)\n else:\n final_update[markdown_display_comp] = gr.update(\n value=\"# Research Stopped\\n\\n*Error reading final report file after stop.*\")\n else:\n final_update[markdown_display_comp] = gr.update(value=\"# Research Stopped by User\")\n\n # Keep start button disabled, run_deep_research finally block will re-enable it.\n final_update[start_button_comp] = gr.update(interactive=False)\n\n else:\n logger.warning(\"Stop clicked but no active research task found.\")\n # Reset UI state just in case\n final_update = {\n start_button_comp: gr.update(interactive=True),\n stop_button_comp: gr.update(interactive=False),\n webui_manager.get_component_by_id(\"deep_research_agent.research_task\"): gr.update(interactive=True),\n webui_manager.get_component_by_id(\"deep_research_agent.resume_task_id\"): gr.update(interactive=True),\n webui_manager.get_component_by_id(\"deep_research_agent.max_iteration\"): gr.update(interactive=True),\n webui_manager.get_component_by_id(\"deep_research_agent.max_query\"): gr.update(interactive=True),\n }\n\n return final_update\n\n\nasync def update_mcp_server(mcp_file: str, webui_manager: WebuiManager):\n \"\"\"\n Update the MCP server.\n \"\"\"\n if hasattr(webui_manager, \"dr_agent\") and webui_manager.dr_agent:\n logger.warning(\"⚠️ Close controller because mcp file has changed!\")\n await webui_manager.dr_agent.close_mcp_client()\n\n if not mcp_file or not os.path.exists(mcp_file) or not mcp_file.endswith('.json'):\n logger.warning(f\"{mcp_file} is not a valid MCP file.\")\n return None, gr.update(visible=False)\n\n with open(mcp_file, 'r') as f:\n mcp_server = json.load(f)\n\n return json.dumps(mcp_server, indent=2), gr.update(visible=True)\n\n\ndef create_deep_research_agent_tab(webui_manager: WebuiManager):\n \"\"\"\n Creates a deep research agent tab\n \"\"\"\n input_components = set(webui_manager.get_components())\n tab_components = {}\n\n with gr.Group():\n with gr.Row():\n mcp_json_file = gr.File(label=\"MCP server json\", interactive=True, file_types=[\".json\"])\n mcp_server_config = gr.Textbox(label=\"MCP server\", lines=6, interactive=True, visible=False)\n\n with gr.Group():\n research_task = gr.Textbox(label=\"Research Task\", lines=5,\n value=\"Give me a detailed travel plan to Switzerland from June 1st to 10th.\",\n interactive=True)\n with gr.Row():\n resume_task_id = gr.Textbox(label=\"Resume Task ID\", value=\"\",\n interactive=True)\n parallel_num = gr.Number(label=\"Parallel Agent Num\", value=1,\n precision=0,\n interactive=True)\n max_query = gr.Textbox(label=\"Research Save Dir\", value=\"./tmp/deep_research\",\n interactive=True)\n with gr.Row():\n stop_button = gr.Button(\"⏹️ Stop\", variant=\"stop\", scale=2)\n start_button = gr.Button(\"▶️ Run\", variant=\"primary\", scale=3)\n with gr.Group():\n markdown_display = gr.Markdown(label=\"Research Report\")\n markdown_download = gr.File(label=\"Download Research Report\", interactive=False)\n tab_components.update(\n dict(\n research_task=research_task,\n parallel_num=parallel_num,\n max_query=max_query,\n start_button=start_button,\n stop_button=stop_button,\n markdown_display=markdown_display,\n markdown_download=markdown_download,\n resume_task_id=resume_task_id,\n mcp_json_file=mcp_json_file,\n mcp_server_config=mcp_server_config,\n )\n )\n webui_manager.add_components(\"deep_research_agent\", tab_components)\n webui_manager.init_deep_research_agent()\n\n async def update_wrapper(mcp_file):\n \"\"\"Wrapper for handle_pause_resume.\"\"\"\n update_dict = await update_mcp_server(mcp_file, webui_manager)\n yield update_dict\n\n mcp_json_file.change(\n update_wrapper,\n inputs=[mcp_json_file],\n outputs=[mcp_server_config, mcp_server_config]\n )\n\n dr_tab_outputs = list(tab_components.values())\n all_managed_inputs = set(webui_manager.get_components())\n\n # --- Define Event Handler Wrappers ---\n async def start_wrapper(comps: Dict[Component, Any]) -> AsyncGenerator[Dict[Component, Any], None]:\n async for update in run_deep_research(webui_manager, comps):\n yield update\n\n async def stop_wrapper() -> AsyncGenerator[Dict[Component, Any], None]:\n update_dict = await stop_deep_research(webui_manager)\n yield update_dict\n\n # --- Connect Handlers ---\n start_button.click(\n fn=start_wrapper,\n inputs=all_managed_inputs,\n outputs=dr_tab_outputs\n )\n\n stop_button.click(\n fn=stop_wrapper,\n inputs=None,\n outputs=dr_tab_outputs\n )\n"], ["/web-ui/src/webui/components/browser_use_agent_tab.py", "import asyncio\nimport json\nimport logging\nimport os\nimport uuid\nfrom typing import Any, AsyncGenerator, Dict, Optional\n\nimport gradio as gr\n\n# from browser_use.agent.service import Agent\nfrom browser_use.agent.views import (\n AgentHistoryList,\n AgentOutput,\n)\nfrom browser_use.browser.browser import BrowserConfig\nfrom browser_use.browser.context import BrowserContext, BrowserContextConfig\nfrom browser_use.browser.views import BrowserState\nfrom gradio.components import Component\nfrom langchain_core.language_models.chat_models import BaseChatModel\n\nfrom src.agent.browser_use.browser_use_agent import BrowserUseAgent\nfrom src.browser.custom_browser import CustomBrowser\nfrom src.controller.custom_controller import CustomController\nfrom src.utils import llm_provider\nfrom src.webui.webui_manager import WebuiManager\n\nlogger = logging.getLogger(__name__)\n\n\n# --- Helper Functions --- (Defined at module level)\n\n\nasync def _initialize_llm(\n provider: Optional[str],\n model_name: Optional[str],\n temperature: float,\n base_url: Optional[str],\n api_key: Optional[str],\n num_ctx: Optional[int] = None,\n) -> Optional[BaseChatModel]:\n \"\"\"Initializes the LLM based on settings. Returns None if provider/model is missing.\"\"\"\n if not provider or not model_name:\n logger.info(\"LLM Provider or Model Name not specified, LLM will be None.\")\n return None\n try:\n # Use your actual LLM provider logic here\n logger.info(\n f\"Initializing LLM: Provider={provider}, Model={model_name}, Temp={temperature}\"\n )\n # Example using a placeholder function\n llm = llm_provider.get_llm_model(\n provider=provider,\n model_name=model_name,\n temperature=temperature,\n base_url=base_url or None,\n api_key=api_key or None,\n # Add other relevant params like num_ctx for ollama\n num_ctx=num_ctx if provider == \"ollama\" else None,\n )\n return llm\n except Exception as e:\n logger.error(f\"Failed to initialize LLM: {e}\", exc_info=True)\n gr.Warning(\n f\"Failed to initialize LLM '{model_name}' for provider '{provider}'. Please check settings. Error: {e}\"\n )\n return None\n\n\ndef _get_config_value(\n webui_manager: WebuiManager,\n comp_dict: Dict[gr.components.Component, Any],\n comp_id_suffix: str,\n default: Any = None,\n) -> Any:\n \"\"\"Safely get value from component dictionary using its ID suffix relative to the tab.\"\"\"\n # Assumes component ID format is \"tab_name.comp_name\"\n tab_name = \"browser_use_agent\" # Hardcode or derive if needed\n comp_id = f\"{tab_name}.{comp_id_suffix}\"\n # Need to find the component object first using the ID from the manager\n try:\n comp = webui_manager.get_component_by_id(comp_id)\n return comp_dict.get(comp, default)\n except KeyError:\n # Try accessing settings tabs as well\n for prefix in [\"agent_settings\", \"browser_settings\"]:\n try:\n comp_id = f\"{prefix}.{comp_id_suffix}\"\n comp = webui_manager.get_component_by_id(comp_id)\n return comp_dict.get(comp, default)\n except KeyError:\n continue\n logger.warning(\n f\"Component with suffix '{comp_id_suffix}' not found in manager for value lookup.\"\n )\n return default\n\n\ndef _format_agent_output(model_output: AgentOutput) -> str:\n \"\"\"Formats AgentOutput for display in the chatbot using JSON.\"\"\"\n content = \"\"\n if model_output:\n try:\n # Directly use model_dump if actions and current_state are Pydantic models\n action_dump = [\n action.model_dump(exclude_none=True) for action in model_output.action\n ]\n\n state_dump = model_output.current_state.model_dump(exclude_none=True)\n model_output_dump = {\n \"current_state\": state_dump,\n \"action\": action_dump,\n }\n # Dump to JSON string with indentation\n json_string = json.dumps(model_output_dump, indent=4, ensure_ascii=False)\n # Wrap in 
 for proper display in HTML\n            content = f\"
{json_string}
\"\n\n        except AttributeError as ae:\n            logger.error(\n                f\"AttributeError during model dump: {ae}. Check if 'action' or 'current_state' or their items support 'model_dump'.\"\n            )\n            content = f\"
Error: Could not format agent output (AttributeError: {ae}).\\nRaw output: {str(model_output)}
\"\n        except Exception as e:\n            logger.error(f\"Error formatting agent output: {e}\", exc_info=True)\n            # Fallback to simple string representation on error\n            content = f\"
Error formatting agent output.\\nRaw output:\\n{str(model_output)}
\"\n\n    return content.strip()\n\n\n# --- Updated Callback Implementation ---\n\n\nasync def _handle_new_step(\n        webui_manager: WebuiManager, state: BrowserState, output: AgentOutput, step_num: int\n):\n    \"\"\"Callback for each step taken by the agent, including screenshot display.\"\"\"\n\n    # Use the correct chat history attribute name from the user's code\n    if not hasattr(webui_manager, \"bu_chat_history\"):\n        logger.error(\n            \"Attribute 'bu_chat_history' not found in webui_manager! Cannot add chat message.\"\n        )\n        # Initialize it maybe? Or raise an error? For now, log and potentially skip chat update.\n        webui_manager.bu_chat_history = []  # Initialize if missing (consider if this is the right place)\n        # return # Or stop if this is critical\n    step_num -= 1\n    logger.info(f\"Step {step_num} completed.\")\n\n    # --- Screenshot Handling ---\n    screenshot_html = \"\"\n    # Ensure state.screenshot exists and is not empty before proceeding\n    # Use getattr for safer access\n    screenshot_data = getattr(state, \"screenshot\", None)\n    if screenshot_data:\n        try:\n            # Basic validation: check if it looks like base64\n            if (\n                    isinstance(screenshot_data, str) and len(screenshot_data) > 100\n            ):  # Arbitrary length check\n                # *** UPDATED STYLE: Removed centering, adjusted width ***\n                img_tag = f'\"Step'\n                screenshot_html = (\n                        img_tag + \"
\"\n                )  # Use 
 for line break after inline-block image\n            else:\n                logger.warning(\n                    f\"Screenshot for step {step_num} seems invalid (type: {type(screenshot_data)}, len: {len(screenshot_data) if isinstance(screenshot_data, str) else 'N/A'}).\"\n                )\n                screenshot_html = \"**[Invalid screenshot data]**
\"\n\n        except Exception as e:\n            logger.error(\n                f\"Error processing or formatting screenshot for step {step_num}: {e}\",\n                exc_info=True,\n            )\n            screenshot_html = \"**[Error displaying screenshot]**
\"\n    else:\n        logger.debug(f\"No screenshot available for step {step_num}.\")\n\n    # --- Format Agent Output ---\n    formatted_output = _format_agent_output(output)  # Use the updated function\n\n    # --- Combine and Append to Chat ---\n    step_header = f\"--- **Step {step_num}** ---\"\n    # Combine header, image (with line break), and JSON block\n    final_content = step_header + \"
\" + screenshot_html + formatted_output\n\n    chat_message = {\n        \"role\": \"assistant\",\n        \"content\": final_content.strip(),  # Remove leading/trailing whitespace\n    }\n\n    # Append to the correct chat history list\n    webui_manager.bu_chat_history.append(chat_message)\n\n    await asyncio.sleep(0.05)\n\n\ndef _handle_done(webui_manager: WebuiManager, history: AgentHistoryList):\n    \"\"\"Callback when the agent finishes the task (success or failure).\"\"\"\n    logger.info(\n        f\"Agent task finished. Duration: {history.total_duration_seconds():.2f}s, Tokens: {history.total_input_tokens()}\"\n    )\n    final_summary = \"**Task Completed**\\n\"\n    final_summary += f\"- Duration: {history.total_duration_seconds():.2f} seconds\\n\"\n    final_summary += f\"- Total Input Tokens: {history.total_input_tokens()}\\n\"  # Or total tokens if available\n\n    final_result = history.final_result()\n    if final_result:\n        final_summary += f\"- Final Result: {final_result}\\n\"\n\n    errors = history.errors()\n    if errors and any(errors):\n        final_summary += f\"- **Errors:**\\n```\\n{errors}\\n```\\n\"\n    else:\n        final_summary += \"- Status: Success\\n\"\n\n    webui_manager.bu_chat_history.append(\n        {\"role\": \"assistant\", \"content\": final_summary}\n    )\n\n\nasync def _ask_assistant_callback(\n        webui_manager: WebuiManager, query: str, browser_context: BrowserContext\n) -> Dict[str, Any]:\n    \"\"\"Callback triggered by the agent's ask_for_assistant action.\"\"\"\n    logger.info(\"Agent requires assistance. Waiting for user input.\")\n\n    if not hasattr(webui_manager, \"_chat_history\"):\n        logger.error(\"Chat history not found in webui_manager during ask_assistant!\")\n        return {\"response\": \"Internal Error: Cannot display help request.\"}\n\n    webui_manager.bu_chat_history.append(\n        {\n            \"role\": \"assistant\",\n            \"content\": f\"**Need Help:** {query}\\nPlease provide information or perform the required action in the browser, then type your response/confirmation below and click 'Submit Response'.\",\n        }\n    )\n\n    # Use state stored in webui_manager\n    webui_manager.bu_response_event = asyncio.Event()\n    webui_manager.bu_user_help_response = None  # Reset previous response\n\n    try:\n        logger.info(\"Waiting for user response event...\")\n        await asyncio.wait_for(\n            webui_manager.bu_response_event.wait(), timeout=3600.0\n        )  # Long timeout\n        logger.info(\"User response event received.\")\n    except asyncio.TimeoutError:\n        logger.warning(\"Timeout waiting for user assistance.\")\n        webui_manager.bu_chat_history.append(\n            {\n                \"role\": \"assistant\",\n                \"content\": \"**Timeout:** No response received. Trying to proceed.\",\n            }\n        )\n        webui_manager.bu_response_event = None  # Clear the event\n        return {\"response\": \"Timeout: User did not respond.\"}  # Inform the agent\n\n    response = webui_manager.bu_user_help_response\n    webui_manager.bu_chat_history.append(\n        {\"role\": \"user\", \"content\": response}\n    )  # Show user response in chat\n    webui_manager.bu_response_event = (\n        None  # Clear the event for the next potential request\n    )\n    return {\"response\": response}\n\n\n# --- Core Agent Execution Logic --- (Needs access to webui_manager)\n\n\nasync def run_agent_task(\n        webui_manager: WebuiManager, components: Dict[gr.components.Component, Any]\n) -> AsyncGenerator[Dict[gr.components.Component, Any], None]:\n    \"\"\"Handles the entire lifecycle of initializing and running the agent.\"\"\"\n\n    # --- Get Components ---\n    # Need handles to specific UI components to update them\n    user_input_comp = webui_manager.get_component_by_id(\"browser_use_agent.user_input\")\n    run_button_comp = webui_manager.get_component_by_id(\"browser_use_agent.run_button\")\n    stop_button_comp = webui_manager.get_component_by_id(\n        \"browser_use_agent.stop_button\"\n    )\n    pause_resume_button_comp = webui_manager.get_component_by_id(\n        \"browser_use_agent.pause_resume_button\"\n    )\n    clear_button_comp = webui_manager.get_component_by_id(\n        \"browser_use_agent.clear_button\"\n    )\n    chatbot_comp = webui_manager.get_component_by_id(\"browser_use_agent.chatbot\")\n    history_file_comp = webui_manager.get_component_by_id(\n        \"browser_use_agent.agent_history_file\"\n    )\n    gif_comp = webui_manager.get_component_by_id(\"browser_use_agent.recording_gif\")\n    browser_view_comp = webui_manager.get_component_by_id(\n        \"browser_use_agent.browser_view\"\n    )\n\n    # --- 1. Get Task and Initial UI Update ---\n    task = components.get(user_input_comp, \"\").strip()\n    if not task:\n        gr.Warning(\"Please enter a task.\")\n        yield {run_button_comp: gr.update(interactive=True)}\n        return\n\n    # Set running state indirectly via _current_task\n    webui_manager.bu_chat_history.append({\"role\": \"user\", \"content\": task})\n\n    yield {\n        user_input_comp: gr.Textbox(\n            value=\"\", interactive=False, placeholder=\"Agent is running...\"\n        ),\n        run_button_comp: gr.Button(value=\"⏳ Running...\", interactive=False),\n        stop_button_comp: gr.Button(interactive=True),\n        pause_resume_button_comp: gr.Button(value=\"⏸️ Pause\", interactive=True),\n        clear_button_comp: gr.Button(interactive=False),\n        chatbot_comp: gr.update(value=webui_manager.bu_chat_history),\n        history_file_comp: gr.update(value=None),\n        gif_comp: gr.update(value=None),\n    }\n\n    # --- Agent Settings ---\n    # Access settings values via components dict, getting IDs from webui_manager\n    def get_setting(key, default=None):\n        comp = webui_manager.id_to_component.get(f\"agent_settings.{key}\")\n        return components.get(comp, default) if comp else default\n\n    override_system_prompt = get_setting(\"override_system_prompt\") or None\n    extend_system_prompt = get_setting(\"extend_system_prompt\") or None\n    llm_provider_name = get_setting(\n        \"llm_provider\", None\n    )  # Default to None if not found\n    llm_model_name = get_setting(\"llm_model_name\", None)\n    llm_temperature = get_setting(\"llm_temperature\", 0.6)\n    use_vision = get_setting(\"use_vision\", True)\n    ollama_num_ctx = get_setting(\"ollama_num_ctx\", 16000)\n    llm_base_url = get_setting(\"llm_base_url\") or None\n    llm_api_key = get_setting(\"llm_api_key\") or None\n    max_steps = get_setting(\"max_steps\", 100)\n    max_actions = get_setting(\"max_actions\", 10)\n    max_input_tokens = get_setting(\"max_input_tokens\", 128000)\n    tool_calling_str = get_setting(\"tool_calling_method\", \"auto\")\n    tool_calling_method = tool_calling_str if tool_calling_str != \"None\" else None\n    mcp_server_config_comp = webui_manager.id_to_component.get(\n        \"agent_settings.mcp_server_config\"\n    )\n    mcp_server_config_str = (\n        components.get(mcp_server_config_comp) if mcp_server_config_comp else None\n    )\n    mcp_server_config = (\n        json.loads(mcp_server_config_str) if mcp_server_config_str else None\n    )\n\n    # Planner LLM Settings (Optional)\n    planner_llm_provider_name = get_setting(\"planner_llm_provider\") or None\n    planner_llm = None\n    planner_use_vision = False\n    if planner_llm_provider_name:\n        planner_llm_model_name = get_setting(\"planner_llm_model_name\")\n        planner_llm_temperature = get_setting(\"planner_llm_temperature\", 0.6)\n        planner_ollama_num_ctx = get_setting(\"planner_ollama_num_ctx\", 16000)\n        planner_llm_base_url = get_setting(\"planner_llm_base_url\") or None\n        planner_llm_api_key = get_setting(\"planner_llm_api_key\") or None\n        planner_use_vision = get_setting(\"planner_use_vision\", False)\n\n        planner_llm = await _initialize_llm(\n            planner_llm_provider_name,\n            planner_llm_model_name,\n            planner_llm_temperature,\n            planner_llm_base_url,\n            planner_llm_api_key,\n            planner_ollama_num_ctx if planner_llm_provider_name == \"ollama\" else None,\n        )\n\n    # --- Browser Settings ---\n    def get_browser_setting(key, default=None):\n        comp = webui_manager.id_to_component.get(f\"browser_settings.{key}\")\n        return components.get(comp, default) if comp else default\n\n    browser_binary_path = get_browser_setting(\"browser_binary_path\") or None\n    browser_user_data_dir = get_browser_setting(\"browser_user_data_dir\") or None\n    use_own_browser = get_browser_setting(\n        \"use_own_browser\", False\n    )  # Logic handled by CDP/WSS presence\n    keep_browser_open = get_browser_setting(\"keep_browser_open\", False)\n    headless = get_browser_setting(\"headless\", False)\n    disable_security = get_browser_setting(\"disable_security\", False)\n    window_w = int(get_browser_setting(\"window_w\", 1280))\n    window_h = int(get_browser_setting(\"window_h\", 1100))\n    cdp_url = get_browser_setting(\"cdp_url\") or None\n    wss_url = get_browser_setting(\"wss_url\") or None\n    save_recording_path = get_browser_setting(\"save_recording_path\") or None\n    save_trace_path = get_browser_setting(\"save_trace_path\") or None\n    save_agent_history_path = get_browser_setting(\n        \"save_agent_history_path\", \"./tmp/agent_history\"\n    )\n    save_download_path = get_browser_setting(\"save_download_path\", \"./tmp/downloads\")\n\n    stream_vw = 70\n    stream_vh = int(70 * window_h // window_w)\n\n    os.makedirs(save_agent_history_path, exist_ok=True)\n    if save_recording_path:\n        os.makedirs(save_recording_path, exist_ok=True)\n    if save_trace_path:\n        os.makedirs(save_trace_path, exist_ok=True)\n    if save_download_path:\n        os.makedirs(save_download_path, exist_ok=True)\n\n    # --- 2. Initialize LLM ---\n    main_llm = await _initialize_llm(\n        llm_provider_name,\n        llm_model_name,\n        llm_temperature,\n        llm_base_url,\n        llm_api_key,\n        ollama_num_ctx if llm_provider_name == \"ollama\" else None,\n    )\n\n    # Pass the webui_manager instance to the callback when wrapping it\n    async def ask_callback_wrapper(\n            query: str, browser_context: BrowserContext\n    ) -> Dict[str, Any]:\n        return await _ask_assistant_callback(webui_manager, query, browser_context)\n\n    if not webui_manager.bu_controller:\n        webui_manager.bu_controller = CustomController(\n            ask_assistant_callback=ask_callback_wrapper\n        )\n        await webui_manager.bu_controller.setup_mcp_client(mcp_server_config)\n\n    # --- 4. Initialize Browser and Context ---\n    should_close_browser_on_finish = not keep_browser_open\n\n    try:\n        # Close existing resources if not keeping open\n        if not keep_browser_open:\n            if webui_manager.bu_browser_context:\n                logger.info(\"Closing previous browser context.\")\n                await webui_manager.bu_browser_context.close()\n                webui_manager.bu_browser_context = None\n            if webui_manager.bu_browser:\n                logger.info(\"Closing previous browser.\")\n                await webui_manager.bu_browser.close()\n                webui_manager.bu_browser = None\n\n        # Create Browser if needed\n        if not webui_manager.bu_browser:\n            logger.info(\"Launching new browser instance.\")\n            extra_args = []\n            if use_own_browser:\n                browser_binary_path = os.getenv(\"BROWSER_PATH\", None) or browser_binary_path\n                if browser_binary_path == \"\":\n                    browser_binary_path = None\n                browser_user_data = browser_user_data_dir or os.getenv(\"BROWSER_USER_DATA\", None)\n                if browser_user_data:\n                    extra_args += [f\"--user-data-dir={browser_user_data}\"]\n            else:\n                browser_binary_path = None\n\n            webui_manager.bu_browser = CustomBrowser(\n                config=BrowserConfig(\n                    headless=headless,\n                    disable_security=disable_security,\n                    browser_binary_path=browser_binary_path,\n                    extra_browser_args=extra_args,\n                    wss_url=wss_url,\n                    cdp_url=cdp_url,\n                    new_context_config=BrowserContextConfig(\n                        window_width=window_w,\n                        window_height=window_h,\n                    )\n                )\n            )\n\n        # Create Context if needed\n        if not webui_manager.bu_browser_context:\n            logger.info(\"Creating new browser context.\")\n            context_config = BrowserContextConfig(\n                trace_path=save_trace_path if save_trace_path else None,\n                save_recording_path=save_recording_path\n                if save_recording_path\n                else None,\n                save_downloads_path=save_download_path if save_download_path else None,\n                window_height=window_h,\n                window_width=window_w,\n            )\n            if not webui_manager.bu_browser:\n                raise ValueError(\"Browser not initialized, cannot create context.\")\n            webui_manager.bu_browser_context = (\n                await webui_manager.bu_browser.new_context(config=context_config)\n            )\n\n        # --- 5. Initialize or Update Agent ---\n        webui_manager.bu_agent_task_id = str(uuid.uuid4())  # New ID for this task run\n        os.makedirs(\n            os.path.join(save_agent_history_path, webui_manager.bu_agent_task_id),\n            exist_ok=True,\n        )\n        history_file = os.path.join(\n            save_agent_history_path,\n            webui_manager.bu_agent_task_id,\n            f\"{webui_manager.bu_agent_task_id}.json\",\n        )\n        gif_path = os.path.join(\n            save_agent_history_path,\n            webui_manager.bu_agent_task_id,\n            f\"{webui_manager.bu_agent_task_id}.gif\",\n        )\n\n        # Pass the webui_manager to callbacks when wrapping them\n        async def step_callback_wrapper(\n                state: BrowserState, output: AgentOutput, step_num: int\n        ):\n            await _handle_new_step(webui_manager, state, output, step_num)\n\n        def done_callback_wrapper(history: AgentHistoryList):\n            _handle_done(webui_manager, history)\n\n        if not webui_manager.bu_agent:\n            logger.info(f\"Initializing new agent for task: {task}\")\n            if not webui_manager.bu_browser or not webui_manager.bu_browser_context:\n                raise ValueError(\n                    \"Browser or Context not initialized, cannot create agent.\"\n                )\n            webui_manager.bu_agent = BrowserUseAgent(\n                task=task,\n                llm=main_llm,\n                browser=webui_manager.bu_browser,\n                browser_context=webui_manager.bu_browser_context,\n                controller=webui_manager.bu_controller,\n                register_new_step_callback=step_callback_wrapper,\n                register_done_callback=done_callback_wrapper,\n                use_vision=use_vision,\n                override_system_message=override_system_prompt,\n                extend_system_message=extend_system_prompt,\n                max_input_tokens=max_input_tokens,\n                max_actions_per_step=max_actions,\n                tool_calling_method=tool_calling_method,\n                planner_llm=planner_llm,\n                use_vision_for_planner=planner_use_vision if planner_llm else False,\n                source=\"webui\",\n            )\n            webui_manager.bu_agent.state.agent_id = webui_manager.bu_agent_task_id\n            webui_manager.bu_agent.settings.generate_gif = gif_path\n        else:\n            webui_manager.bu_agent.state.agent_id = webui_manager.bu_agent_task_id\n            webui_manager.bu_agent.add_new_task(task)\n            webui_manager.bu_agent.settings.generate_gif = gif_path\n            webui_manager.bu_agent.browser = webui_manager.bu_browser\n            webui_manager.bu_agent.browser_context = webui_manager.bu_browser_context\n            webui_manager.bu_agent.controller = webui_manager.bu_controller\n\n        # --- 6. Run Agent Task and Stream Updates ---\n        agent_run_coro = webui_manager.bu_agent.run(max_steps=max_steps)\n        agent_task = asyncio.create_task(agent_run_coro)\n        webui_manager.bu_current_task = agent_task  # Store the task\n\n        last_chat_len = len(webui_manager.bu_chat_history)\n        while not agent_task.done():\n            is_paused = webui_manager.bu_agent.state.paused\n            is_stopped = webui_manager.bu_agent.state.stopped\n\n            # Check for pause state\n            if is_paused:\n                yield {\n                    pause_resume_button_comp: gr.update(\n                        value=\"▶️ Resume\", interactive=True\n                    ),\n                    stop_button_comp: gr.update(interactive=True),\n                }\n                # Wait until pause is released or task is stopped/done\n                while is_paused and not agent_task.done():\n                    # Re-check agent state in loop\n                    is_paused = webui_manager.bu_agent.state.paused\n                    is_stopped = webui_manager.bu_agent.state.stopped\n                    if is_stopped:  # Stop signal received while paused\n                        break\n                    await asyncio.sleep(0.2)\n\n                if (\n                        agent_task.done() or is_stopped\n                ):  # If stopped or task finished while paused\n                    break\n\n                # If resumed, yield UI update\n                yield {\n                    pause_resume_button_comp: gr.update(\n                        value=\"⏸️ Pause\", interactive=True\n                    ),\n                    run_button_comp: gr.update(\n                        value=\"⏳ Running...\", interactive=False\n                    ),\n                }\n\n            # Check if agent stopped itself or stop button was pressed (which sets agent.state.stopped)\n            if is_stopped:\n                logger.info(\"Agent has stopped (internally or via stop button).\")\n                if not agent_task.done():\n                    # Ensure the task coroutine finishes if agent just set flag\n                    try:\n                        await asyncio.wait_for(\n                            agent_task, timeout=1.0\n                        )  # Give it a moment to exit run()\n                    except asyncio.TimeoutError:\n                        logger.warning(\n                            \"Agent task did not finish quickly after stop signal, cancelling.\"\n                        )\n                        agent_task.cancel()\n                    except Exception:  # Catch task exceptions if it errors on stop\n                        pass\n                break  # Exit the streaming loop\n\n            # Check if agent is asking for help (via response_event)\n            update_dict = {}\n            if webui_manager.bu_response_event is not None:\n                update_dict = {\n                    user_input_comp: gr.update(\n                        placeholder=\"Agent needs help. Enter response and submit.\",\n                        interactive=True,\n                    ),\n                    run_button_comp: gr.update(\n                        value=\"✔️ Submit Response\", interactive=True\n                    ),\n                    pause_resume_button_comp: gr.update(interactive=False),\n                    stop_button_comp: gr.update(interactive=False),\n                    chatbot_comp: gr.update(value=webui_manager.bu_chat_history),\n                }\n                last_chat_len = len(webui_manager.bu_chat_history)\n                yield update_dict\n                # Wait until response is submitted or task finishes\n                while (\n                        webui_manager.bu_response_event is not None\n                        and not agent_task.done()\n                ):\n                    await asyncio.sleep(0.2)\n                # Restore UI after response submitted or if task ended unexpectedly\n                if not agent_task.done():\n                    yield {\n                        user_input_comp: gr.update(\n                            placeholder=\"Agent is running...\", interactive=False\n                        ),\n                        run_button_comp: gr.update(\n                            value=\"⏳ Running...\", interactive=False\n                        ),\n                        pause_resume_button_comp: gr.update(interactive=True),\n                        stop_button_comp: gr.update(interactive=True),\n                    }\n                else:\n                    break  # Task finished while waiting for response\n\n            # Update Chatbot if new messages arrived via callbacks\n            if len(webui_manager.bu_chat_history) > last_chat_len:\n                update_dict[chatbot_comp] = gr.update(\n                    value=webui_manager.bu_chat_history\n                )\n                last_chat_len = len(webui_manager.bu_chat_history)\n\n            # Update Browser View\n            if headless and webui_manager.bu_browser_context:\n                try:\n                    screenshot_b64 = (\n                        await webui_manager.bu_browser_context.take_screenshot()\n                    )\n                    if screenshot_b64:\n                        html_content = f''\n                        update_dict[browser_view_comp] = gr.update(\n                            value=html_content, visible=True\n                        )\n                    else:\n                        html_content = f\"
Waiting for browser session...
\"\n                        update_dict[browser_view_comp] = gr.update(\n                            value=html_content, visible=True\n                        )\n                except Exception as e:\n                    logger.debug(f\"Failed to capture screenshot: {e}\")\n                    update_dict[browser_view_comp] = gr.update(\n                        value=\"
Error loading view...
\",\n                        visible=True,\n                    )\n            else:\n                update_dict[browser_view_comp] = gr.update(visible=False)\n\n            # Yield accumulated updates\n            if update_dict:\n                yield update_dict\n\n            await asyncio.sleep(0.1)  # Polling interval\n\n        # --- 7. Task Finalization ---\n        webui_manager.bu_agent.state.paused = False\n        webui_manager.bu_agent.state.stopped = False\n        final_update = {}\n        try:\n            logger.info(\"Agent task completing...\")\n            # Await the task ensure completion and catch exceptions if not already caught\n            if not agent_task.done():\n                await agent_task  # Retrieve result/exception\n            elif agent_task.exception():  # Check if task finished with exception\n                agent_task.result()  # Raise the exception to be caught below\n            logger.info(\"Agent task completed processing.\")\n\n            logger.info(f\"Explicitly saving agent history to: {history_file}\")\n            webui_manager.bu_agent.save_history(history_file)\n\n            if os.path.exists(history_file):\n                final_update[history_file_comp] = gr.File(value=history_file)\n\n            if gif_path and os.path.exists(gif_path):\n                logger.info(f\"GIF found at: {gif_path}\")\n                final_update[gif_comp] = gr.Image(value=gif_path)\n\n        except asyncio.CancelledError:\n            logger.info(\"Agent task was cancelled.\")\n            if not any(\n                    \"Cancelled\" in msg.get(\"content\", \"\")\n                    for msg in webui_manager.bu_chat_history\n                    if msg.get(\"role\") == \"assistant\"\n            ):\n                webui_manager.bu_chat_history.append(\n                    {\"role\": \"assistant\", \"content\": \"**Task Cancelled**.\"}\n                )\n            final_update[chatbot_comp] = gr.update(value=webui_manager.bu_chat_history)\n        except Exception as e:\n            logger.error(f\"Error during agent execution: {e}\", exc_info=True)\n            error_message = (\n                f\"**Agent Execution Error:**\\n```\\n{type(e).__name__}: {e}\\n```\"\n            )\n            if not any(\n                    error_message in msg.get(\"content\", \"\")\n                    for msg in webui_manager.bu_chat_history\n                    if msg.get(\"role\") == \"assistant\"\n            ):\n                webui_manager.bu_chat_history.append(\n                    {\"role\": \"assistant\", \"content\": error_message}\n                )\n            final_update[chatbot_comp] = gr.update(value=webui_manager.bu_chat_history)\n            gr.Error(f\"Agent execution failed: {e}\")\n\n        finally:\n            webui_manager.bu_current_task = None  # Clear the task reference\n\n            # Close browser/context if requested\n            if should_close_browser_on_finish:\n                if webui_manager.bu_browser_context:\n                    logger.info(\"Closing browser context after task.\")\n                    await webui_manager.bu_browser_context.close()\n                    webui_manager.bu_browser_context = None\n                if webui_manager.bu_browser:\n                    logger.info(\"Closing browser after task.\")\n                    await webui_manager.bu_browser.close()\n                    webui_manager.bu_browser = None\n\n            # --- 8. Final UI Update ---\n            final_update.update(\n                {\n                    user_input_comp: gr.update(\n                        value=\"\",\n                        interactive=True,\n                        placeholder=\"Enter your next task...\",\n                    ),\n                    run_button_comp: gr.update(value=\"▶️ Submit Task\", interactive=True),\n                    stop_button_comp: gr.update(value=\"⏹️ Stop\", interactive=False),\n                    pause_resume_button_comp: gr.update(\n                        value=\"⏸️ Pause\", interactive=False\n                    ),\n                    clear_button_comp: gr.update(interactive=True),\n                    # Ensure final chat history is shown\n                    chatbot_comp: gr.update(value=webui_manager.bu_chat_history),\n                }\n            )\n            yield final_update\n\n    except Exception as e:\n        # Catch errors during setup (before agent run starts)\n        logger.error(f\"Error setting up agent task: {e}\", exc_info=True)\n        webui_manager.bu_current_task = None  # Ensure state is reset\n        yield {\n            user_input_comp: gr.update(\n                interactive=True, placeholder=\"Error during setup. Enter task...\"\n            ),\n            run_button_comp: gr.update(value=\"▶️ Submit Task\", interactive=True),\n            stop_button_comp: gr.update(value=\"⏹️ Stop\", interactive=False),\n            pause_resume_button_comp: gr.update(value=\"⏸️ Pause\", interactive=False),\n            clear_button_comp: gr.update(interactive=True),\n            chatbot_comp: gr.update(\n                value=webui_manager.bu_chat_history\n                      + [{\"role\": \"assistant\", \"content\": f\"**Setup Error:** {e}\"}]\n            ),\n        }\n\n\n# --- Button Click Handlers --- (Need access to webui_manager)\n\n\nasync def handle_submit(\n        webui_manager: WebuiManager, components: Dict[gr.components.Component, Any]\n):\n    \"\"\"Handles clicks on the main 'Submit' button.\"\"\"\n    user_input_comp = webui_manager.get_component_by_id(\"browser_use_agent.user_input\")\n    user_input_value = components.get(user_input_comp, \"\").strip()\n\n    # Check if waiting for user assistance\n    if webui_manager.bu_response_event and not webui_manager.bu_response_event.is_set():\n        logger.info(f\"User submitted assistance: {user_input_value}\")\n        webui_manager.bu_user_help_response = (\n            user_input_value if user_input_value else \"User provided no text response.\"\n        )\n        webui_manager.bu_response_event.set()\n        # UI updates handled by the main loop reacting to the event being set\n        yield {\n            user_input_comp: gr.update(\n                value=\"\",\n                interactive=False,\n                placeholder=\"Waiting for agent to continue...\",\n            ),\n            webui_manager.get_component_by_id(\n                \"browser_use_agent.run_button\"\n            ): gr.update(value=\"⏳ Running...\", interactive=False),\n        }\n    # Check if a task is currently running (using _current_task)\n    elif webui_manager.bu_current_task and not webui_manager.bu_current_task.done():\n        logger.warning(\n            \"Submit button clicked while agent is already running and not asking for help.\"\n        )\n        gr.Info(\"Agent is currently running. Please wait or use Stop/Pause.\")\n        yield {}  # No change\n    else:\n        # Handle submission for a new task\n        logger.info(\"Submit button clicked for new task.\")\n        # Use async generator to stream updates from run_agent_task\n        async for update in run_agent_task(webui_manager, components):\n            yield update\n\n\nasync def handle_stop(webui_manager: WebuiManager):\n    \"\"\"Handles clicks on the 'Stop' button.\"\"\"\n    logger.info(\"Stop button clicked.\")\n    agent = webui_manager.bu_agent\n    task = webui_manager.bu_current_task\n\n    if agent and task and not task.done():\n        # Signal the agent to stop by setting its internal flag\n        agent.state.stopped = True\n        agent.state.paused = False  # Ensure not paused if stopped\n        return {\n            webui_manager.get_component_by_id(\n                \"browser_use_agent.stop_button\"\n            ): gr.update(interactive=False, value=\"⏹️ Stopping...\"),\n            webui_manager.get_component_by_id(\n                \"browser_use_agent.pause_resume_button\"\n            ): gr.update(interactive=False),\n            webui_manager.get_component_by_id(\n                \"browser_use_agent.run_button\"\n            ): gr.update(interactive=False),\n        }\n    else:\n        logger.warning(\"Stop clicked but agent is not running or task is already done.\")\n        # Reset UI just in case it's stuck\n        return {\n            webui_manager.get_component_by_id(\n                \"browser_use_agent.run_button\"\n            ): gr.update(interactive=True),\n            webui_manager.get_component_by_id(\n                \"browser_use_agent.stop_button\"\n            ): gr.update(interactive=False),\n            webui_manager.get_component_by_id(\n                \"browser_use_agent.pause_resume_button\"\n            ): gr.update(interactive=False),\n            webui_manager.get_component_by_id(\n                \"browser_use_agent.clear_button\"\n            ): gr.update(interactive=True),\n        }\n\n\nasync def handle_pause_resume(webui_manager: WebuiManager):\n    \"\"\"Handles clicks on the 'Pause/Resume' button.\"\"\"\n    agent = webui_manager.bu_agent\n    task = webui_manager.bu_current_task\n\n    if agent and task and not task.done():\n        if agent.state.paused:\n            logger.info(\"Resume button clicked.\")\n            agent.resume()\n            # UI update happens in main loop\n            return {\n                webui_manager.get_component_by_id(\n                    \"browser_use_agent.pause_resume_button\"\n                ): gr.update(value=\"⏸️ Pause\", interactive=True)\n            }  # Optimistic update\n        else:\n            logger.info(\"Pause button clicked.\")\n            agent.pause()\n            return {\n                webui_manager.get_component_by_id(\n                    \"browser_use_agent.pause_resume_button\"\n                ): gr.update(value=\"▶️ Resume\", interactive=True)\n            }  # Optimistic update\n    else:\n        logger.warning(\n            \"Pause/Resume clicked but agent is not running or doesn't support state.\"\n        )\n        return {}  # No change\n\n\nasync def handle_clear(webui_manager: WebuiManager):\n    \"\"\"Handles clicks on the 'Clear' button.\"\"\"\n    logger.info(\"Clear button clicked.\")\n\n    # Stop any running task first\n    task = webui_manager.bu_current_task\n    if task and not task.done():\n        logger.info(\"Clearing requires stopping the current task.\")\n        webui_manager.bu_agent.stop()\n        task.cancel()\n        try:\n            await asyncio.wait_for(task, timeout=2.0)  # Wait briefly\n        except (asyncio.CancelledError, asyncio.TimeoutError):\n            pass\n        except Exception as e:\n            logger.warning(f\"Error stopping task on clear: {e}\")\n    webui_manager.bu_current_task = None\n\n    if webui_manager.bu_controller:\n        await webui_manager.bu_controller.close_mcp_client()\n        webui_manager.bu_controller = None\n    webui_manager.bu_agent = None\n\n    # Reset state stored in manager\n    webui_manager.bu_chat_history = []\n    webui_manager.bu_response_event = None\n    webui_manager.bu_user_help_response = None\n    webui_manager.bu_agent_task_id = None\n\n    logger.info(\"Agent state and browser resources cleared.\")\n\n    # Reset UI components\n    return {\n        webui_manager.get_component_by_id(\"browser_use_agent.chatbot\"): gr.update(\n            value=[]\n        ),\n        webui_manager.get_component_by_id(\"browser_use_agent.user_input\"): gr.update(\n            value=\"\", placeholder=\"Enter your task here...\"\n        ),\n        webui_manager.get_component_by_id(\n            \"browser_use_agent.agent_history_file\"\n        ): gr.update(value=None),\n        webui_manager.get_component_by_id(\"browser_use_agent.recording_gif\"): gr.update(\n            value=None\n        ),\n        webui_manager.get_component_by_id(\"browser_use_agent.browser_view\"): gr.update(\n            value=\"
Browser Cleared
\"\n        ),\n        webui_manager.get_component_by_id(\"browser_use_agent.run_button\"): gr.update(\n            value=\"▶️ Submit Task\", interactive=True\n        ),\n        webui_manager.get_component_by_id(\"browser_use_agent.stop_button\"): gr.update(\n            interactive=False\n        ),\n        webui_manager.get_component_by_id(\n            \"browser_use_agent.pause_resume_button\"\n        ): gr.update(value=\"⏸️ Pause\", interactive=False),\n        webui_manager.get_component_by_id(\"browser_use_agent.clear_button\"): gr.update(\n            interactive=True\n        ),\n    }\n\n\n# --- Tab Creation Function ---\n\n\ndef create_browser_use_agent_tab(webui_manager: WebuiManager):\n    \"\"\"\n    Create the run agent tab, defining UI, state, and handlers.\n    \"\"\"\n    webui_manager.init_browser_use_agent()\n\n    # --- Define UI Components ---\n    tab_components = {}\n    with gr.Column():\n        chatbot = gr.Chatbot(\n            lambda: webui_manager.bu_chat_history,  # Load history dynamically\n            elem_id=\"browser_use_chatbot\",\n            label=\"Agent Interaction\",\n            type=\"messages\",\n            height=600,\n            show_copy_button=True,\n        )\n        user_input = gr.Textbox(\n            label=\"Your Task or Response\",\n            placeholder=\"Enter your task here or provide assistance when asked.\",\n            lines=3,\n            interactive=True,\n            elem_id=\"user_input\",\n        )\n        with gr.Row():\n            stop_button = gr.Button(\n                \"⏹️ Stop\", interactive=False, variant=\"stop\", scale=2\n            )\n            pause_resume_button = gr.Button(\n                \"⏸️ Pause\", interactive=False, variant=\"secondary\", scale=2, visible=True\n            )\n            clear_button = gr.Button(\n                \"🗑️ Clear\", interactive=True, variant=\"secondary\", scale=2\n            )\n            run_button = gr.Button(\"▶️ Submit Task\", variant=\"primary\", scale=3)\n\n        browser_view = gr.HTML(\n            value=\"
Browser View (Requires Headless=True)
\",\n            label=\"Browser Live View\",\n            elem_id=\"browser_view\",\n            visible=False,\n        )\n        with gr.Column():\n            gr.Markdown(\"### Task Outputs\")\n            agent_history_file = gr.File(label=\"Agent History JSON\", interactive=False)\n            recording_gif = gr.Image(\n                label=\"Task Recording GIF\",\n                format=\"gif\",\n                interactive=False,\n                type=\"filepath\",\n            )\n\n    # --- Store Components in Manager ---\n    tab_components.update(\n        dict(\n            chatbot=chatbot,\n            user_input=user_input,\n            clear_button=clear_button,\n            run_button=run_button,\n            stop_button=stop_button,\n            pause_resume_button=pause_resume_button,\n            agent_history_file=agent_history_file,\n            recording_gif=recording_gif,\n            browser_view=browser_view,\n        )\n    )\n    webui_manager.add_components(\n        \"browser_use_agent\", tab_components\n    )  # Use \"browser_use_agent\" as tab_name prefix\n\n    all_managed_components = set(\n        webui_manager.get_components()\n    )  # Get all components known to manager\n    run_tab_outputs = list(tab_components.values())\n\n    async def submit_wrapper(\n            components_dict: Dict[Component, Any],\n    ) -> AsyncGenerator[Dict[Component, Any], None]:\n        \"\"\"Wrapper for handle_submit that yields its results.\"\"\"\n        async for update in handle_submit(webui_manager, components_dict):\n            yield update\n\n    async def stop_wrapper() -> AsyncGenerator[Dict[Component, Any], None]:\n        \"\"\"Wrapper for handle_stop.\"\"\"\n        update_dict = await handle_stop(webui_manager)\n        yield update_dict\n\n    async def pause_resume_wrapper() -> AsyncGenerator[Dict[Component, Any], None]:\n        \"\"\"Wrapper for handle_pause_resume.\"\"\"\n        update_dict = await handle_pause_resume(webui_manager)\n        yield update_dict\n\n    async def clear_wrapper() -> AsyncGenerator[Dict[Component, Any], None]:\n        \"\"\"Wrapper for handle_clear.\"\"\"\n        update_dict = await handle_clear(webui_manager)\n        yield update_dict\n\n    # --- Connect Event Handlers using the Wrappers --\n    run_button.click(\n        fn=submit_wrapper, inputs=all_managed_components, outputs=run_tab_outputs\n    )\n    user_input.submit(\n        fn=submit_wrapper, inputs=all_managed_components, outputs=run_tab_outputs\n    )\n    stop_button.click(fn=stop_wrapper, inputs=None, outputs=run_tab_outputs)\n    pause_resume_button.click(\n        fn=pause_resume_wrapper, inputs=None, outputs=run_tab_outputs\n    )\n    clear_button.click(fn=clear_wrapper, inputs=None, outputs=run_tab_outputs)\n"], ["/web-ui/src/agent/browser_use/browser_use_agent.py", "from __future__ import annotations\n\nimport asyncio\nimport logging\nimport os\n\n# from lmnr.sdk.decorators import observe\nfrom browser_use.agent.gif import create_history_gif\nfrom browser_use.agent.service import Agent, AgentHookFunc\nfrom browser_use.agent.views import (\n    ActionResult,\n    AgentHistory,\n    AgentHistoryList,\n    AgentStepInfo,\n    ToolCallingMethod,\n)\nfrom browser_use.browser.views import BrowserStateHistory\nfrom browser_use.utils import time_execution_async\nfrom dotenv import load_dotenv\nfrom browser_use.agent.message_manager.utils import is_model_without_tool_support\n\nload_dotenv()\nlogger = logging.getLogger(__name__)\n\nSKIP_LLM_API_KEY_VERIFICATION = (\n        os.environ.get(\"SKIP_LLM_API_KEY_VERIFICATION\", \"false\").lower()[0] in \"ty1\"\n)\n\n\nclass BrowserUseAgent(Agent):\n    def _set_tool_calling_method(self) -> ToolCallingMethod | None:\n        tool_calling_method = self.settings.tool_calling_method\n        if tool_calling_method == 'auto':\n            if is_model_without_tool_support(self.model_name):\n                return 'raw'\n            elif self.chat_model_library == 'ChatGoogleGenerativeAI':\n                return None\n            elif self.chat_model_library == 'ChatOpenAI':\n                return 'function_calling'\n            elif self.chat_model_library == 'AzureChatOpenAI':\n                return 'function_calling'\n            else:\n                return None\n        else:\n            return tool_calling_method\n\n    @time_execution_async(\"--run (agent)\")\n    async def run(\n            self, max_steps: int = 100, on_step_start: AgentHookFunc | None = None,\n            on_step_end: AgentHookFunc | None = None\n    ) -> AgentHistoryList:\n        \"\"\"Execute the task with maximum number of steps\"\"\"\n\n        loop = asyncio.get_event_loop()\n\n        # Set up the Ctrl+C signal handler with callbacks specific to this agent\n        from browser_use.utils import SignalHandler\n\n        signal_handler = SignalHandler(\n            loop=loop,\n            pause_callback=self.pause,\n            resume_callback=self.resume,\n            custom_exit_callback=None,  # No special cleanup needed on forced exit\n            exit_on_second_int=True,\n        )\n        signal_handler.register()\n\n        try:\n            self._log_agent_run()\n\n            # Execute initial actions if provided\n            if self.initial_actions:\n                result = await self.multi_act(self.initial_actions, check_for_new_elements=False)\n                self.state.last_result = result\n\n            for step in range(max_steps):\n                # Check if waiting for user input after Ctrl+C\n                if self.state.paused:\n                    signal_handler.wait_for_resume()\n                    signal_handler.reset()\n\n                # Check if we should stop due to too many failures\n                if self.state.consecutive_failures >= self.settings.max_failures:\n                    logger.error(f'❌ Stopping due to {self.settings.max_failures} consecutive failures')\n                    break\n\n                # Check control flags before each step\n                if self.state.stopped:\n                    logger.info('Agent stopped')\n                    break\n\n                while self.state.paused:\n                    await asyncio.sleep(0.2)  # Small delay to prevent CPU spinning\n                    if self.state.stopped:  # Allow stopping while paused\n                        break\n\n                if on_step_start is not None:\n                    await on_step_start(self)\n\n                step_info = AgentStepInfo(step_number=step, max_steps=max_steps)\n                await self.step(step_info)\n\n                if on_step_end is not None:\n                    await on_step_end(self)\n\n                if self.state.history.is_done():\n                    if self.settings.validate_output and step < max_steps - 1:\n                        if not await self._validate_output():\n                            continue\n\n                    await self.log_completion()\n                    break\n            else:\n                error_message = 'Failed to complete task in maximum steps'\n\n                self.state.history.history.append(\n                    AgentHistory(\n                        model_output=None,\n                        result=[ActionResult(error=error_message, include_in_memory=True)],\n                        state=BrowserStateHistory(\n                            url='',\n                            title='',\n                            tabs=[],\n                            interacted_element=[],\n                            screenshot=None,\n                        ),\n                        metadata=None,\n                    )\n                )\n\n                logger.info(f'❌ {error_message}')\n\n            return self.state.history\n\n        except KeyboardInterrupt:\n            # Already handled by our signal handler, but catch any direct KeyboardInterrupt as well\n            logger.info('Got KeyboardInterrupt during execution, returning current history')\n            return self.state.history\n\n        finally:\n            # Unregister signal handlers before cleanup\n            signal_handler.unregister()\n\n            if self.settings.save_playwright_script_path:\n                logger.info(\n                    f'Agent run finished. Attempting to save Playwright script to: {self.settings.save_playwright_script_path}'\n                )\n                try:\n                    # Extract sensitive data keys if sensitive_data is provided\n                    keys = list(self.sensitive_data.keys()) if self.sensitive_data else None\n                    # Pass browser and context config to the saving method\n                    self.state.history.save_as_playwright_script(\n                        self.settings.save_playwright_script_path,\n                        sensitive_data_keys=keys,\n                        browser_config=self.browser.config,\n                        context_config=self.browser_context.config,\n                    )\n                except Exception as script_gen_err:\n                    # Log any error during script generation/saving\n                    logger.error(f'Failed to save Playwright script: {script_gen_err}', exc_info=True)\n\n            await self.close()\n\n            if self.settings.generate_gif:\n                output_path: str = 'agent_history.gif'\n                if isinstance(self.settings.generate_gif, str):\n                    output_path = self.settings.generate_gif\n\n                create_history_gif(task=self.task, history=self.state.history, output_path=output_path)\n"], ["/web-ui/src/controller/custom_controller.py", "import pdb\n\nimport pyperclip\nfrom typing import Optional, Type, Callable, Dict, Any, Union, Awaitable, TypeVar\nfrom pydantic import BaseModel\nfrom browser_use.agent.views import ActionResult\nfrom browser_use.browser.context import BrowserContext\nfrom browser_use.controller.service import Controller, DoneAction\nfrom browser_use.controller.registry.service import Registry, RegisteredAction\nfrom main_content_extractor import MainContentExtractor\nfrom browser_use.controller.views import (\n    ClickElementAction,\n    DoneAction,\n    ExtractPageContentAction,\n    GoToUrlAction,\n    InputTextAction,\n    OpenTabAction,\n    ScrollAction,\n    SearchGoogleAction,\n    SendKeysAction,\n    SwitchTabAction,\n)\nimport logging\nimport inspect\nimport asyncio\nimport os\nfrom langchain_core.language_models.chat_models import BaseChatModel\nfrom browser_use.agent.views import ActionModel, ActionResult\n\nfrom src.utils.mcp_client import create_tool_param_model, setup_mcp_client_and_tools\n\nfrom browser_use.utils import time_execution_sync\n\nlogger = logging.getLogger(__name__)\n\nContext = TypeVar('Context')\n\n\nclass CustomController(Controller):\n    def __init__(self, exclude_actions: list[str] = [],\n                 output_model: Optional[Type[BaseModel]] = None,\n                 ask_assistant_callback: Optional[Union[Callable[[str, BrowserContext], Dict[str, Any]], Callable[\n                     [str, BrowserContext], Awaitable[Dict[str, Any]]]]] = None,\n                 ):\n        super().__init__(exclude_actions=exclude_actions, output_model=output_model)\n        self._register_custom_actions()\n        self.ask_assistant_callback = ask_assistant_callback\n        self.mcp_client = None\n        self.mcp_server_config = None\n\n    def _register_custom_actions(self):\n        \"\"\"Register all custom browser actions\"\"\"\n\n        @self.registry.action(\n            \"When executing tasks, prioritize autonomous completion. However, if you encounter a definitive blocker \"\n            \"that prevents you from proceeding independently – such as needing credentials you don't possess, \"\n            \"requiring subjective human judgment, needing a physical action performed, encountering complex CAPTCHAs, \"\n            \"or facing limitations in your capabilities – you must request human assistance.\"\n        )\n        async def ask_for_assistant(query: str, browser: BrowserContext):\n            if self.ask_assistant_callback:\n                if inspect.iscoroutinefunction(self.ask_assistant_callback):\n                    user_response = await self.ask_assistant_callback(query, browser)\n                else:\n                    user_response = self.ask_assistant_callback(query, browser)\n                msg = f\"AI ask: {query}. User response: {user_response['response']}\"\n                logger.info(msg)\n                return ActionResult(extracted_content=msg, include_in_memory=True)\n            else:\n                return ActionResult(extracted_content=\"Human cannot help you. Please try another way.\",\n                                    include_in_memory=True)\n\n        @self.registry.action(\n            'Upload file to interactive element with file path ',\n        )\n        async def upload_file(index: int, path: str, browser: BrowserContext, available_file_paths: list[str]):\n            if path not in available_file_paths:\n                return ActionResult(error=f'File path {path} is not available')\n\n            if not os.path.exists(path):\n                return ActionResult(error=f'File {path} does not exist')\n\n            dom_el = await browser.get_dom_element_by_index(index)\n\n            file_upload_dom_el = dom_el.get_file_upload_element()\n\n            if file_upload_dom_el is None:\n                msg = f'No file upload element found at index {index}'\n                logger.info(msg)\n                return ActionResult(error=msg)\n\n            file_upload_el = await browser.get_locate_element(file_upload_dom_el)\n\n            if file_upload_el is None:\n                msg = f'No file upload element found at index {index}'\n                logger.info(msg)\n                return ActionResult(error=msg)\n\n            try:\n                await file_upload_el.set_input_files(path)\n                msg = f'Successfully uploaded file to index {index}'\n                logger.info(msg)\n                return ActionResult(extracted_content=msg, include_in_memory=True)\n            except Exception as e:\n                msg = f'Failed to upload file to index {index}: {str(e)}'\n                logger.info(msg)\n                return ActionResult(error=msg)\n\n    @time_execution_sync('--act')\n    async def act(\n            self,\n            action: ActionModel,\n            browser_context: Optional[BrowserContext] = None,\n            #\n            page_extraction_llm: Optional[BaseChatModel] = None,\n            sensitive_data: Optional[Dict[str, str]] = None,\n            available_file_paths: Optional[list[str]] = None,\n            #\n            context: Context | None = None,\n    ) -> ActionResult:\n        \"\"\"Execute an action\"\"\"\n\n        try:\n            for action_name, params in action.model_dump(exclude_unset=True).items():\n                if params is not None:\n                    if action_name.startswith(\"mcp\"):\n                        # this is a mcp tool\n                        logger.debug(f\"Invoke MCP tool: {action_name}\")\n                        mcp_tool = self.registry.registry.actions.get(action_name).function\n                        result = await mcp_tool.ainvoke(params)\n                    else:\n                        result = await self.registry.execute_action(\n                            action_name,\n                            params,\n                            browser=browser_context,\n                            page_extraction_llm=page_extraction_llm,\n                            sensitive_data=sensitive_data,\n                            available_file_paths=available_file_paths,\n                            context=context,\n                        )\n\n                    if isinstance(result, str):\n                        return ActionResult(extracted_content=result)\n                    elif isinstance(result, ActionResult):\n                        return result\n                    elif result is None:\n                        return ActionResult()\n                    else:\n                        raise ValueError(f'Invalid action result type: {type(result)} of {result}')\n            return ActionResult()\n        except Exception as e:\n            raise e\n\n    async def setup_mcp_client(self, mcp_server_config: Optional[Dict[str, Any]] = None):\n        self.mcp_server_config = mcp_server_config\n        if self.mcp_server_config:\n            self.mcp_client = await setup_mcp_client_and_tools(self.mcp_server_config)\n            self.register_mcp_tools()\n\n    def register_mcp_tools(self):\n        \"\"\"\n        Register the MCP tools used by this controller.\n        \"\"\"\n        if self.mcp_client:\n            for server_name in self.mcp_client.server_name_to_tools:\n                for tool in self.mcp_client.server_name_to_tools[server_name]:\n                    tool_name = f\"mcp.{server_name}.{tool.name}\"\n                    self.registry.registry.actions[tool_name] = RegisteredAction(\n                        name=tool_name,\n                        description=tool.description,\n                        function=tool,\n                        param_model=create_tool_param_model(tool),\n                    )\n                    logger.info(f\"Add mcp tool: {tool_name}\")\n                logger.debug(\n                    f\"Registered {len(self.mcp_client.server_name_to_tools[server_name])} mcp tools for {server_name}\")\n        else:\n            logger.warning(f\"MCP client not started.\")\n\n    async def close_mcp_client(self):\n        if self.mcp_client:\n            await self.mcp_client.__aexit__(None, None, None)\n"], ["/web-ui/src/utils/mcp_client.py", "import inspect\nimport logging\nimport uuid\nfrom datetime import date, datetime, time\nfrom enum import Enum\nfrom typing import Any, Dict, List, Optional, Set, Type, Union, get_type_hints\n\nfrom browser_use.controller.registry.views import ActionModel\nfrom langchain.tools import BaseTool\nfrom langchain_mcp_adapters.client import MultiServerMCPClient\nfrom pydantic import BaseModel, Field, create_model\nfrom pydantic.v1 import BaseModel, Field\n\nlogger = logging.getLogger(__name__)\n\n\nasync def setup_mcp_client_and_tools(mcp_server_config: Dict[str, Any]) -> Optional[MultiServerMCPClient]:\n    \"\"\"\n    Initializes the MultiServerMCPClient, connects to servers, fetches tools,\n    filters them, and returns a flat list of usable tools and the client instance.\n\n    Returns:\n        A tuple containing:\n        - list[BaseTool]: The filtered list of usable LangChain tools.\n        - MultiServerMCPClient | None: The initialized and started client instance, or None on failure.\n    \"\"\"\n\n    logger.info(\"Initializing MultiServerMCPClient...\")\n\n    if not mcp_server_config:\n        logger.error(\"No MCP server configuration provided.\")\n        return None\n\n    try:\n        if \"mcpServers\" in mcp_server_config:\n            mcp_server_config = mcp_server_config[\"mcpServers\"]\n        client = MultiServerMCPClient(mcp_server_config)\n        await client.__aenter__()\n        return client\n\n    except Exception as e:\n        logger.error(f\"Failed to setup MCP client or fetch tools: {e}\", exc_info=True)\n        return None\n\n\ndef create_tool_param_model(tool: BaseTool) -> Type[BaseModel]:\n    \"\"\"Creates a Pydantic model from a LangChain tool's schema\"\"\"\n\n    # Get tool schema information\n    json_schema = tool.args_schema\n    tool_name = tool.name\n\n    # If the tool already has a schema defined, convert it to a new param_model\n    if json_schema is not None:\n\n        # Create new parameter model\n        params = {}\n\n        # Process properties if they exist\n        if 'properties' in json_schema:\n            # Find required fields\n            required_fields: Set[str] = set(json_schema.get('required', []))\n\n            for prop_name, prop_details in json_schema['properties'].items():\n                field_type = resolve_type(prop_details, f\"{tool_name}_{prop_name}\")\n\n                # Check if parameter is required\n                is_required = prop_name in required_fields\n\n                # Get default value and description\n                default_value = prop_details.get('default', ... if is_required else None)\n                description = prop_details.get('description', '')\n\n                # Add field constraints\n                field_kwargs = {'default': default_value}\n                if description:\n                    field_kwargs['description'] = description\n\n                # Add additional constraints if present\n                if 'minimum' in prop_details:\n                    field_kwargs['ge'] = prop_details['minimum']\n                if 'maximum' in prop_details:\n                    field_kwargs['le'] = prop_details['maximum']\n                if 'minLength' in prop_details:\n                    field_kwargs['min_length'] = prop_details['minLength']\n                if 'maxLength' in prop_details:\n                    field_kwargs['max_length'] = prop_details['maxLength']\n                if 'pattern' in prop_details:\n                    field_kwargs['pattern'] = prop_details['pattern']\n\n                # Add to parameters dictionary\n                params[prop_name] = (field_type, Field(**field_kwargs))\n\n        return create_model(\n            f'{tool_name}_parameters',\n            __base__=ActionModel,\n            **params,  # type: ignore\n        )\n\n    # If no schema is defined, extract parameters from the _run method\n    run_method = tool._run\n    sig = inspect.signature(run_method)\n\n    # Get type hints for better type information\n    try:\n        type_hints = get_type_hints(run_method)\n    except Exception:\n        type_hints = {}\n\n    params = {}\n    for name, param in sig.parameters.items():\n        # Skip 'self' parameter and any other parameters you want to exclude\n        if name == 'self':\n            continue\n\n        # Get annotation from type hints if available, otherwise from signature\n        annotation = type_hints.get(name, param.annotation)\n        if annotation == inspect.Parameter.empty:\n            annotation = Any\n\n        # Use default value if available, otherwise make it required\n        if param.default != param.empty:\n            params[name] = (annotation, param.default)\n        else:\n            params[name] = (annotation, ...)\n\n    return create_model(\n        f'{tool_name}_parameters',\n        __base__=ActionModel,\n        **params,  # type: ignore\n    )\n\n\ndef resolve_type(prop_details: Dict[str, Any], prefix: str = \"\") -> Any:\n    \"\"\"Recursively resolves JSON schema type to Python/Pydantic type\"\"\"\n\n    # Handle reference types\n    if '$ref' in prop_details:\n        # In a real application, reference resolution would be needed\n        return Any\n\n    # Basic type mapping\n    type_mapping = {\n        'string': str,\n        'integer': int,\n        'number': float,\n        'boolean': bool,\n        'array': List,\n        'object': Dict,\n        'null': type(None),\n    }\n\n    # Handle formatted strings\n    if prop_details.get('type') == 'string' and 'format' in prop_details:\n        format_mapping = {\n            'date-time': datetime,\n            'date': date,\n            'time': time,\n            'email': str,\n            'uri': str,\n            'url': str,\n            'uuid': uuid.UUID,\n            'binary': bytes,\n        }\n        return format_mapping.get(prop_details['format'], str)\n\n    # Handle enum types\n    if 'enum' in prop_details:\n        enum_values = prop_details['enum']\n        # Create dynamic enum class with safe names\n        enum_dict = {}\n        for i, v in enumerate(enum_values):\n            # Ensure enum names are valid Python identifiers\n            if isinstance(v, str):\n                key = v.upper().replace(' ', '_').replace('-', '_')\n                if not key.isidentifier():\n                    key = f\"VALUE_{i}\"\n            else:\n                key = f\"VALUE_{i}\"\n            enum_dict[key] = v\n\n        # Only create enum if we have values\n        if enum_dict:\n            return Enum(f\"{prefix}_Enum\", enum_dict)\n        return str  # Fallback\n\n    # Handle array types\n    if prop_details.get('type') == 'array' and 'items' in prop_details:\n        item_type = resolve_type(prop_details['items'], f\"{prefix}_item\")\n        return List[item_type]  # type: ignore\n\n    # Handle object types with properties\n    if prop_details.get('type') == 'object' and 'properties' in prop_details:\n        nested_params = {}\n        for nested_name, nested_details in prop_details['properties'].items():\n            nested_type = resolve_type(nested_details, f\"{prefix}_{nested_name}\")\n            # Get required field info\n            required_fields = prop_details.get('required', [])\n            is_required = nested_name in required_fields\n            default_value = nested_details.get('default', ... if is_required else None)\n            description = nested_details.get('description', '')\n\n            field_kwargs = {'default': default_value}\n            if description:\n                field_kwargs['description'] = description\n\n            nested_params[nested_name] = (nested_type, Field(**field_kwargs))\n\n        # Create nested model\n        nested_model = create_model(f\"{prefix}_Model\", **nested_params)\n        return nested_model\n\n    # Handle union types (oneOf, anyOf)\n    if 'oneOf' in prop_details or 'anyOf' in prop_details:\n        union_schema = prop_details.get('oneOf') or prop_details.get('anyOf')\n        union_types = []\n        for i, t in enumerate(union_schema):\n            union_types.append(resolve_type(t, f\"{prefix}_{i}\"))\n\n        if union_types:\n            return Union.__getitem__(tuple(union_types))  # type: ignore\n        return Any\n\n    # Handle allOf (intersection types)\n    if 'allOf' in prop_details:\n        nested_params = {}\n        for i, schema_part in enumerate(prop_details['allOf']):\n            if 'properties' in schema_part:\n                for nested_name, nested_details in schema_part['properties'].items():\n                    nested_type = resolve_type(nested_details, f\"{prefix}_allOf_{i}_{nested_name}\")\n                    # Check if required\n                    required_fields = schema_part.get('required', [])\n                    is_required = nested_name in required_fields\n                    nested_params[nested_name] = (nested_type, ... if is_required else None)\n\n        # Create composite model\n        if nested_params:\n            composite_model = create_model(f\"{prefix}_CompositeModel\", **nested_params)\n            return composite_model\n        return Dict\n\n    # Default to basic types\n    schema_type = prop_details.get('type', 'string')\n    if isinstance(schema_type, list):\n        # Handle multiple types (e.g., [\"string\", \"null\"])\n        non_null_types = [t for t in schema_type if t != 'null']\n        if non_null_types:\n            primary_type = type_mapping.get(non_null_types[0], Any)\n            if 'null' in schema_type:\n                return Optional[primary_type]  # type: ignore\n            return primary_type\n        return Any\n\n    return type_mapping.get(schema_type, Any)\n"], ["/web-ui/src/webui/webui_manager.py", "import json\nfrom collections.abc import Generator\nfrom typing import TYPE_CHECKING\nimport os\nimport gradio as gr\nfrom datetime import datetime\nfrom typing import Optional, Dict, List\nimport uuid\nimport asyncio\nimport time\n\nfrom gradio.components import Component\nfrom browser_use.browser.browser import Browser\nfrom browser_use.browser.context import BrowserContext\nfrom browser_use.agent.service import Agent\nfrom src.browser.custom_browser import CustomBrowser\nfrom src.browser.custom_context import CustomBrowserContext\nfrom src.controller.custom_controller import CustomController\nfrom src.agent.deep_research.deep_research_agent import DeepResearchAgent\n\n\nclass WebuiManager:\n    def __init__(self, settings_save_dir: str = \"./tmp/webui_settings\"):\n        self.id_to_component: dict[str, Component] = {}\n        self.component_to_id: dict[Component, str] = {}\n\n        self.settings_save_dir = settings_save_dir\n        os.makedirs(self.settings_save_dir, exist_ok=True)\n\n    def init_browser_use_agent(self) -> None:\n        \"\"\"\n        init browser use agent\n        \"\"\"\n        self.bu_agent: Optional[Agent] = None\n        self.bu_browser: Optional[CustomBrowser] = None\n        self.bu_browser_context: Optional[CustomBrowserContext] = None\n        self.bu_controller: Optional[CustomController] = None\n        self.bu_chat_history: List[Dict[str, Optional[str]]] = []\n        self.bu_response_event: Optional[asyncio.Event] = None\n        self.bu_user_help_response: Optional[str] = None\n        self.bu_current_task: Optional[asyncio.Task] = None\n        self.bu_agent_task_id: Optional[str] = None\n\n    def init_deep_research_agent(self) -> None:\n        \"\"\"\n        init deep research agent\n        \"\"\"\n        self.dr_agent: Optional[DeepResearchAgent] = None\n        self.dr_current_task = None\n        self.dr_agent_task_id: Optional[str] = None\n        self.dr_save_dir: Optional[str] = None\n\n    def add_components(self, tab_name: str, components_dict: dict[str, \"Component\"]) -> None:\n        \"\"\"\n        Add tab components\n        \"\"\"\n        for comp_name, component in components_dict.items():\n            comp_id = f\"{tab_name}.{comp_name}\"\n            self.id_to_component[comp_id] = component\n            self.component_to_id[component] = comp_id\n\n    def get_components(self) -> list[\"Component\"]:\n        \"\"\"\n        Get all components\n        \"\"\"\n        return list(self.id_to_component.values())\n\n    def get_component_by_id(self, comp_id: str) -> \"Component\":\n        \"\"\"\n        Get component by id\n        \"\"\"\n        return self.id_to_component[comp_id]\n\n    def get_id_by_component(self, comp: \"Component\") -> str:\n        \"\"\"\n        Get id by component\n        \"\"\"\n        return self.component_to_id[comp]\n\n    def save_config(self, components: Dict[\"Component\", str]) -> None:\n        \"\"\"\n        Save config\n        \"\"\"\n        cur_settings = {}\n        for comp in components:\n            if not isinstance(comp, gr.Button) and not isinstance(comp, gr.File) and str(\n                    getattr(comp, \"interactive\", True)).lower() != \"false\":\n                comp_id = self.get_id_by_component(comp)\n                cur_settings[comp_id] = components[comp]\n\n        config_name = datetime.now().strftime(\"%Y%m%d-%H%M%S\")\n        with open(os.path.join(self.settings_save_dir, f\"{config_name}.json\"), \"w\") as fw:\n            json.dump(cur_settings, fw, indent=4)\n\n        return os.path.join(self.settings_save_dir, f\"{config_name}.json\")\n\n    def load_config(self, config_path: str):\n        \"\"\"\n        Load config\n        \"\"\"\n        with open(config_path, \"r\") as fr:\n            ui_settings = json.load(fr)\n\n        update_components = {}\n        for comp_id, comp_val in ui_settings.items():\n            if comp_id in self.id_to_component:\n                comp = self.id_to_component[comp_id]\n                if comp.__class__.__name__ == \"Chatbot\":\n                    update_components[comp] = comp.__class__(value=comp_val, type=\"messages\")\n                else:\n                    update_components[comp] = comp.__class__(value=comp_val)\n                    if comp_id == \"agent_settings.planner_llm_provider\":\n                        yield update_components  # yield provider, let callback run\n                        time.sleep(0.1)  # wait for Gradio UI callback\n\n        config_status = self.id_to_component[\"load_save_config.config_status\"]\n        update_components.update(\n            {\n                config_status: config_status.__class__(value=f\"Successfully loaded config: {config_path}\")\n            }\n        )\n        yield update_components\n"], ["/web-ui/src/utils/llm_provider.py", "from openai import OpenAI\nimport pdb\nfrom langchain_openai import ChatOpenAI\nfrom langchain_core.globals import get_llm_cache\nfrom langchain_core.language_models.base import (\n    BaseLanguageModel,\n    LangSmithParams,\n    LanguageModelInput,\n)\nimport os\nfrom langchain_core.load import dumpd, dumps\nfrom langchain_core.messages import (\n    AIMessage,\n    SystemMessage,\n    AnyMessage,\n    BaseMessage,\n    BaseMessageChunk,\n    HumanMessage,\n    convert_to_messages,\n    message_chunk_to_message,\n)\nfrom langchain_core.outputs import (\n    ChatGeneration,\n    ChatGenerationChunk,\n    ChatResult,\n    LLMResult,\n    RunInfo,\n)\nfrom langchain_ollama import ChatOllama\nfrom langchain_core.output_parsers.base import OutputParserLike\nfrom langchain_core.runnables import Runnable, RunnableConfig\nfrom langchain_core.tools import BaseTool\n\nfrom typing import (\n    TYPE_CHECKING,\n    Any,\n    Callable,\n    Literal,\n    Optional,\n    Union,\n    cast, List,\n)\nfrom langchain_anthropic import ChatAnthropic\nfrom langchain_mistralai import ChatMistralAI\nfrom langchain_google_genai import ChatGoogleGenerativeAI\nfrom langchain_ollama import ChatOllama\nfrom langchain_openai import AzureChatOpenAI, ChatOpenAI\nfrom langchain_ibm import ChatWatsonx\nfrom langchain_aws import ChatBedrock\nfrom pydantic import SecretStr\n\nfrom src.utils import config\n\n\nclass DeepSeekR1ChatOpenAI(ChatOpenAI):\n\n    def __init__(self, *args: Any, **kwargs: Any) -> None:\n        super().__init__(*args, **kwargs)\n        self.client = OpenAI(\n            base_url=kwargs.get(\"base_url\"),\n            api_key=kwargs.get(\"api_key\")\n        )\n\n    async def ainvoke(\n            self,\n            input: LanguageModelInput,\n            config: Optional[RunnableConfig] = None,\n            *,\n            stop: Optional[list[str]] = None,\n            **kwargs: Any,\n    ) -> AIMessage:\n        message_history = []\n        for input_ in input:\n            if isinstance(input_, SystemMessage):\n                message_history.append({\"role\": \"system\", \"content\": input_.content})\n            elif isinstance(input_, AIMessage):\n                message_history.append({\"role\": \"assistant\", \"content\": input_.content})\n            else:\n                message_history.append({\"role\": \"user\", \"content\": input_.content})\n\n        response = self.client.chat.completions.create(\n            model=self.model_name,\n            messages=message_history\n        )\n\n        reasoning_content = response.choices[0].message.reasoning_content\n        content = response.choices[0].message.content\n        return AIMessage(content=content, reasoning_content=reasoning_content)\n\n    def invoke(\n            self,\n            input: LanguageModelInput,\n            config: Optional[RunnableConfig] = None,\n            *,\n            stop: Optional[list[str]] = None,\n            **kwargs: Any,\n    ) -> AIMessage:\n        message_history = []\n        for input_ in input:\n            if isinstance(input_, SystemMessage):\n                message_history.append({\"role\": \"system\", \"content\": input_.content})\n            elif isinstance(input_, AIMessage):\n                message_history.append({\"role\": \"assistant\", \"content\": input_.content})\n            else:\n                message_history.append({\"role\": \"user\", \"content\": input_.content})\n\n        response = self.client.chat.completions.create(\n            model=self.model_name,\n            messages=message_history\n        )\n\n        reasoning_content = response.choices[0].message.reasoning_content\n        content = response.choices[0].message.content\n        return AIMessage(content=content, reasoning_content=reasoning_content)\n\n\nclass DeepSeekR1ChatOllama(ChatOllama):\n\n    async def ainvoke(\n            self,\n            input: LanguageModelInput,\n            config: Optional[RunnableConfig] = None,\n            *,\n            stop: Optional[list[str]] = None,\n            **kwargs: Any,\n    ) -> AIMessage:\n        org_ai_message = await super().ainvoke(input=input)\n        org_content = org_ai_message.content\n        reasoning_content = org_content.split(\"\")[0].replace(\"\", \"\")\n        content = org_content.split(\"\")[1]\n        if \"**JSON Response:**\" in content:\n            content = content.split(\"**JSON Response:**\")[-1]\n        return AIMessage(content=content, reasoning_content=reasoning_content)\n\n    def invoke(\n            self,\n            input: LanguageModelInput,\n            config: Optional[RunnableConfig] = None,\n            *,\n            stop: Optional[list[str]] = None,\n            **kwargs: Any,\n    ) -> AIMessage:\n        org_ai_message = super().invoke(input=input)\n        org_content = org_ai_message.content\n        reasoning_content = org_content.split(\"\")[0].replace(\"\", \"\")\n        content = org_content.split(\"\")[1]\n        if \"**JSON Response:**\" in content:\n            content = content.split(\"**JSON Response:**\")[-1]\n        return AIMessage(content=content, reasoning_content=reasoning_content)\n\n\ndef get_llm_model(provider: str, **kwargs):\n    \"\"\"\n    Get LLM model\n    :param provider: LLM provider\n    :param kwargs:\n    :return:\n    \"\"\"\n    if provider not in [\"ollama\", \"bedrock\"]:\n        env_var = f\"{provider.upper()}_API_KEY\"\n        api_key = kwargs.get(\"api_key\", \"\") or os.getenv(env_var, \"\")\n        if not api_key:\n            provider_display = config.PROVIDER_DISPLAY_NAMES.get(provider, provider.upper())\n            error_msg = f\"💥 {provider_display} API key not found! 🔑 Please set the `{env_var}` environment variable or provide it in the UI.\"\n            raise ValueError(error_msg)\n        kwargs[\"api_key\"] = api_key\n\n    if provider == \"anthropic\":\n        if not kwargs.get(\"base_url\", \"\"):\n            base_url = \"https://api.anthropic.com\"\n        else:\n            base_url = kwargs.get(\"base_url\")\n\n        return ChatAnthropic(\n            model=kwargs.get(\"model_name\", \"claude-3-5-sonnet-20241022\"),\n            temperature=kwargs.get(\"temperature\", 0.0),\n            base_url=base_url,\n            api_key=api_key,\n        )\n    elif provider == 'mistral':\n        if not kwargs.get(\"base_url\", \"\"):\n            base_url = os.getenv(\"MISTRAL_ENDPOINT\", \"https://api.mistral.ai/v1\")\n        else:\n            base_url = kwargs.get(\"base_url\")\n        if not kwargs.get(\"api_key\", \"\"):\n            api_key = os.getenv(\"MISTRAL_API_KEY\", \"\")\n        else:\n            api_key = kwargs.get(\"api_key\")\n\n        return ChatMistralAI(\n            model=kwargs.get(\"model_name\", \"mistral-large-latest\"),\n            temperature=kwargs.get(\"temperature\", 0.0),\n            base_url=base_url,\n            api_key=api_key,\n        )\n    elif provider == \"openai\":\n        if not kwargs.get(\"base_url\", \"\"):\n            base_url = os.getenv(\"OPENAI_ENDPOINT\", \"https://api.openai.com/v1\")\n        else:\n            base_url = kwargs.get(\"base_url\")\n\n        return ChatOpenAI(\n            model=kwargs.get(\"model_name\", \"gpt-4o\"),\n            temperature=kwargs.get(\"temperature\", 0.0),\n            base_url=base_url,\n            api_key=api_key,\n        )\n    elif provider == \"grok\":\n        if not kwargs.get(\"base_url\", \"\"):\n            base_url = os.getenv(\"GROK_ENDPOINT\", \"https://api.x.ai/v1\")\n        else:\n            base_url = kwargs.get(\"base_url\")\n\n        return ChatOpenAI(\n            model=kwargs.get(\"model_name\", \"grok-3\"),\n            temperature=kwargs.get(\"temperature\", 0.0),\n            base_url=base_url,\n            api_key=api_key,\n        )\n    elif provider == \"deepseek\":\n        if not kwargs.get(\"base_url\", \"\"):\n            base_url = os.getenv(\"DEEPSEEK_ENDPOINT\", \"\")\n        else:\n            base_url = kwargs.get(\"base_url\")\n\n        if kwargs.get(\"model_name\", \"deepseek-chat\") == \"deepseek-reasoner\":\n            return DeepSeekR1ChatOpenAI(\n                model=kwargs.get(\"model_name\", \"deepseek-reasoner\"),\n                temperature=kwargs.get(\"temperature\", 0.0),\n                base_url=base_url,\n                api_key=api_key,\n            )\n        else:\n            return ChatOpenAI(\n                model=kwargs.get(\"model_name\", \"deepseek-chat\"),\n                temperature=kwargs.get(\"temperature\", 0.0),\n                base_url=base_url,\n                api_key=api_key,\n            )\n    elif provider == \"google\":\n        return ChatGoogleGenerativeAI(\n            model=kwargs.get(\"model_name\", \"gemini-2.0-flash-exp\"),\n            temperature=kwargs.get(\"temperature\", 0.0),\n            api_key=api_key,\n        )\n    elif provider == \"ollama\":\n        if not kwargs.get(\"base_url\", \"\"):\n            base_url = os.getenv(\"OLLAMA_ENDPOINT\", \"http://localhost:11434\")\n        else:\n            base_url = kwargs.get(\"base_url\")\n\n        if \"deepseek-r1\" in kwargs.get(\"model_name\", \"qwen2.5:7b\"):\n            return DeepSeekR1ChatOllama(\n                model=kwargs.get(\"model_name\", \"deepseek-r1:14b\"),\n                temperature=kwargs.get(\"temperature\", 0.0),\n                num_ctx=kwargs.get(\"num_ctx\", 32000),\n                base_url=base_url,\n            )\n        else:\n            return ChatOllama(\n                model=kwargs.get(\"model_name\", \"qwen2.5:7b\"),\n                temperature=kwargs.get(\"temperature\", 0.0),\n                num_ctx=kwargs.get(\"num_ctx\", 32000),\n                num_predict=kwargs.get(\"num_predict\", 1024),\n                base_url=base_url,\n            )\n    elif provider == \"azure_openai\":\n        if not kwargs.get(\"base_url\", \"\"):\n            base_url = os.getenv(\"AZURE_OPENAI_ENDPOINT\", \"\")\n        else:\n            base_url = kwargs.get(\"base_url\")\n        api_version = kwargs.get(\"api_version\", \"\") or os.getenv(\"AZURE_OPENAI_API_VERSION\", \"2025-01-01-preview\")\n        return AzureChatOpenAI(\n            model=kwargs.get(\"model_name\", \"gpt-4o\"),\n            temperature=kwargs.get(\"temperature\", 0.0),\n            api_version=api_version,\n            azure_endpoint=base_url,\n            api_key=api_key,\n        )\n    elif provider == \"alibaba\":\n        if not kwargs.get(\"base_url\", \"\"):\n            base_url = os.getenv(\"ALIBABA_ENDPOINT\", \"https://dashscope.aliyuncs.com/compatible-mode/v1\")\n        else:\n            base_url = kwargs.get(\"base_url\")\n\n        return ChatOpenAI(\n            model=kwargs.get(\"model_name\", \"qwen-plus\"),\n            temperature=kwargs.get(\"temperature\", 0.0),\n            base_url=base_url,\n            api_key=api_key,\n        )\n    elif provider == \"ibm\":\n        parameters = {\n            \"temperature\": kwargs.get(\"temperature\", 0.0),\n            \"max_tokens\": kwargs.get(\"num_ctx\", 32000)\n        }\n        if not kwargs.get(\"base_url\", \"\"):\n            base_url = os.getenv(\"IBM_ENDPOINT\", \"https://us-south.ml.cloud.ibm.com\")\n        else:\n            base_url = kwargs.get(\"base_url\")\n\n        return ChatWatsonx(\n            model_id=kwargs.get(\"model_name\", \"ibm/granite-vision-3.1-2b-preview\"),\n            url=base_url,\n            project_id=os.getenv(\"IBM_PROJECT_ID\"),\n            apikey=os.getenv(\"IBM_API_KEY\"),\n            params=parameters\n        )\n    elif provider == \"moonshot\":\n        return ChatOpenAI(\n            model=kwargs.get(\"model_name\", \"moonshot-v1-32k-vision-preview\"),\n            temperature=kwargs.get(\"temperature\", 0.0),\n            base_url=os.getenv(\"MOONSHOT_ENDPOINT\"),\n            api_key=os.getenv(\"MOONSHOT_API_KEY\"),\n        )\n    elif provider == \"unbound\":\n        return ChatOpenAI(\n            model=kwargs.get(\"model_name\", \"gpt-4o-mini\"),\n            temperature=kwargs.get(\"temperature\", 0.0),\n            base_url=os.getenv(\"UNBOUND_ENDPOINT\", \"https://api.getunbound.ai\"),\n            api_key=api_key,\n        )\n    elif provider == \"siliconflow\":\n        if not kwargs.get(\"api_key\", \"\"):\n            api_key = os.getenv(\"SiliconFLOW_API_KEY\", \"\")\n        else:\n            api_key = kwargs.get(\"api_key\")\n        if not kwargs.get(\"base_url\", \"\"):\n            base_url = os.getenv(\"SiliconFLOW_ENDPOINT\", \"\")\n        else:\n            base_url = kwargs.get(\"base_url\")\n        return ChatOpenAI(\n            api_key=api_key,\n            base_url=base_url,\n            model_name=kwargs.get(\"model_name\", \"Qwen/QwQ-32B\"),\n            temperature=kwargs.get(\"temperature\", 0.0),\n        )\n    elif provider == \"modelscope\":\n        if not kwargs.get(\"api_key\", \"\"):\n            api_key = os.getenv(\"MODELSCOPE_API_KEY\", \"\")\n        else:\n            api_key = kwargs.get(\"api_key\")\n        if not kwargs.get(\"base_url\", \"\"):\n            base_url = os.getenv(\"MODELSCOPE_ENDPOINT\", \"\")\n        else:\n            base_url = kwargs.get(\"base_url\")\n        return ChatOpenAI(\n            api_key=api_key,\n            base_url=base_url,\n            model_name=kwargs.get(\"model_name\", \"Qwen/QwQ-32B\"),\n            temperature=kwargs.get(\"temperature\", 0.0),\n        )\n    else:\n        raise ValueError(f\"Unsupported provider: {provider}\")\n"], ["/web-ui/src/webui/components/agent_settings_tab.py", "import json\nimport os\n\nimport gradio as gr\nfrom gradio.components import Component\nfrom typing import Any, Dict, Optional\nfrom src.webui.webui_manager import WebuiManager\nfrom src.utils import config\nimport logging\nfrom functools import partial\n\nlogger = logging.getLogger(__name__)\n\n\ndef update_model_dropdown(llm_provider):\n    \"\"\"\n    Update the model name dropdown with predefined models for the selected provider.\n    \"\"\"\n    # Use predefined models for the selected provider\n    if llm_provider in config.model_names:\n        return gr.Dropdown(choices=config.model_names[llm_provider], value=config.model_names[llm_provider][0],\n                           interactive=True)\n    else:\n        return gr.Dropdown(choices=[], value=\"\", interactive=True, allow_custom_value=True)\n\n\nasync def update_mcp_server(mcp_file: str, webui_manager: WebuiManager):\n    \"\"\"\n    Update the MCP server.\n    \"\"\"\n    if hasattr(webui_manager, \"bu_controller\") and webui_manager.bu_controller:\n        logger.warning(\"⚠️ Close controller because mcp file has changed!\")\n        await webui_manager.bu_controller.close_mcp_client()\n        webui_manager.bu_controller = None\n\n    if not mcp_file or not os.path.exists(mcp_file) or not mcp_file.endswith('.json'):\n        logger.warning(f\"{mcp_file} is not a valid MCP file.\")\n        return None, gr.update(visible=False)\n\n    with open(mcp_file, 'r') as f:\n        mcp_server = json.load(f)\n\n    return json.dumps(mcp_server, indent=2), gr.update(visible=True)\n\n\ndef create_agent_settings_tab(webui_manager: WebuiManager):\n    \"\"\"\n    Creates an agent settings tab.\n    \"\"\"\n    input_components = set(webui_manager.get_components())\n    tab_components = {}\n\n    with gr.Group():\n        with gr.Column():\n            override_system_prompt = gr.Textbox(label=\"Override system prompt\", lines=4, interactive=True)\n            extend_system_prompt = gr.Textbox(label=\"Extend system prompt\", lines=4, interactive=True)\n\n    with gr.Group():\n        mcp_json_file = gr.File(label=\"MCP server json\", interactive=True, file_types=[\".json\"])\n        mcp_server_config = gr.Textbox(label=\"MCP server\", lines=6, interactive=True, visible=False)\n\n    with gr.Group():\n        with gr.Row():\n            llm_provider = gr.Dropdown(\n                choices=[provider for provider, model in config.model_names.items()],\n                label=\"LLM Provider\",\n                value=os.getenv(\"DEFAULT_LLM\", \"openai\"),\n                info=\"Select LLM provider for LLM\",\n                interactive=True\n            )\n            llm_model_name = gr.Dropdown(\n                label=\"LLM Model Name\",\n                choices=config.model_names[os.getenv(\"DEFAULT_LLM\", \"openai\")],\n                value=config.model_names[os.getenv(\"DEFAULT_LLM\", \"openai\")][0],\n                interactive=True,\n                allow_custom_value=True,\n                info=\"Select a model in the dropdown options or directly type a custom model name\"\n            )\n        with gr.Row():\n            llm_temperature = gr.Slider(\n                minimum=0.0,\n                maximum=2.0,\n                value=0.6,\n                step=0.1,\n                label=\"LLM Temperature\",\n                info=\"Controls randomness in model outputs\",\n                interactive=True\n            )\n\n            use_vision = gr.Checkbox(\n                label=\"Use Vision\",\n                value=True,\n                info=\"Enable Vision(Input highlighted screenshot into LLM)\",\n                interactive=True\n            )\n\n            ollama_num_ctx = gr.Slider(\n                minimum=2 ** 8,\n                maximum=2 ** 16,\n                value=16000,\n                step=1,\n                label=\"Ollama Context Length\",\n                info=\"Controls max context length model needs to handle (less = faster)\",\n                visible=False,\n                interactive=True\n            )\n\n        with gr.Row():\n            llm_base_url = gr.Textbox(\n                label=\"Base URL\",\n                value=\"\",\n                info=\"API endpoint URL (if required)\"\n            )\n            llm_api_key = gr.Textbox(\n                label=\"API Key\",\n                type=\"password\",\n                value=\"\",\n                info=\"Your API key (leave blank to use .env)\"\n            )\n\n    with gr.Group():\n        with gr.Row():\n            planner_llm_provider = gr.Dropdown(\n                choices=[provider for provider, model in config.model_names.items()],\n                label=\"Planner LLM Provider\",\n                info=\"Select LLM provider for LLM\",\n                value=None,\n                interactive=True\n            )\n            planner_llm_model_name = gr.Dropdown(\n                label=\"Planner LLM Model Name\",\n                interactive=True,\n                allow_custom_value=True,\n                info=\"Select a model in the dropdown options or directly type a custom model name\"\n            )\n        with gr.Row():\n            planner_llm_temperature = gr.Slider(\n                minimum=0.0,\n                maximum=2.0,\n                value=0.6,\n                step=0.1,\n                label=\"Planner LLM Temperature\",\n                info=\"Controls randomness in model outputs\",\n                interactive=True\n            )\n\n            planner_use_vision = gr.Checkbox(\n                label=\"Use Vision(Planner LLM)\",\n                value=False,\n                info=\"Enable Vision(Input highlighted screenshot into LLM)\",\n                interactive=True\n            )\n\n            planner_ollama_num_ctx = gr.Slider(\n                minimum=2 ** 8,\n                maximum=2 ** 16,\n                value=16000,\n                step=1,\n                label=\"Ollama Context Length\",\n                info=\"Controls max context length model needs to handle (less = faster)\",\n                visible=False,\n                interactive=True\n            )\n\n        with gr.Row():\n            planner_llm_base_url = gr.Textbox(\n                label=\"Base URL\",\n                value=\"\",\n                info=\"API endpoint URL (if required)\"\n            )\n            planner_llm_api_key = gr.Textbox(\n                label=\"API Key\",\n                type=\"password\",\n                value=\"\",\n                info=\"Your API key (leave blank to use .env)\"\n            )\n\n    with gr.Row():\n        max_steps = gr.Slider(\n            minimum=1,\n            maximum=1000,\n            value=100,\n            step=1,\n            label=\"Max Run Steps\",\n            info=\"Maximum number of steps the agent will take\",\n            interactive=True\n        )\n        max_actions = gr.Slider(\n            minimum=1,\n            maximum=100,\n            value=10,\n            step=1,\n            label=\"Max Number of Actions\",\n            info=\"Maximum number of actions the agent will take per step\",\n            interactive=True\n        )\n\n    with gr.Row():\n        max_input_tokens = gr.Number(\n            label=\"Max Input Tokens\",\n            value=128000,\n            precision=0,\n            interactive=True\n        )\n        tool_calling_method = gr.Dropdown(\n            label=\"Tool Calling Method\",\n            value=\"auto\",\n            interactive=True,\n            allow_custom_value=True,\n            choices=['function_calling', 'json_mode', 'raw', 'auto', 'tools', \"None\"],\n            visible=True\n        )\n    tab_components.update(dict(\n        override_system_prompt=override_system_prompt,\n        extend_system_prompt=extend_system_prompt,\n        llm_provider=llm_provider,\n        llm_model_name=llm_model_name,\n        llm_temperature=llm_temperature,\n        use_vision=use_vision,\n        ollama_num_ctx=ollama_num_ctx,\n        llm_base_url=llm_base_url,\n        llm_api_key=llm_api_key,\n        planner_llm_provider=planner_llm_provider,\n        planner_llm_model_name=planner_llm_model_name,\n        planner_llm_temperature=planner_llm_temperature,\n        planner_use_vision=planner_use_vision,\n        planner_ollama_num_ctx=planner_ollama_num_ctx,\n        planner_llm_base_url=planner_llm_base_url,\n        planner_llm_api_key=planner_llm_api_key,\n        max_steps=max_steps,\n        max_actions=max_actions,\n        max_input_tokens=max_input_tokens,\n        tool_calling_method=tool_calling_method,\n        mcp_json_file=mcp_json_file,\n        mcp_server_config=mcp_server_config,\n    ))\n    webui_manager.add_components(\"agent_settings\", tab_components)\n\n    llm_provider.change(\n        fn=lambda x: gr.update(visible=x == \"ollama\"),\n        inputs=llm_provider,\n        outputs=ollama_num_ctx\n    )\n    llm_provider.change(\n        lambda provider: update_model_dropdown(provider),\n        inputs=[llm_provider],\n        outputs=[llm_model_name]\n    )\n    planner_llm_provider.change(\n        fn=lambda x: gr.update(visible=x == \"ollama\"),\n        inputs=[planner_llm_provider],\n        outputs=[planner_ollama_num_ctx]\n    )\n    planner_llm_provider.change(\n        lambda provider: update_model_dropdown(provider),\n        inputs=[planner_llm_provider],\n        outputs=[planner_llm_model_name]\n    )\n\n    async def update_wrapper(mcp_file):\n        \"\"\"Wrapper for handle_pause_resume.\"\"\"\n        update_dict = await update_mcp_server(mcp_file, webui_manager)\n        yield update_dict\n\n    mcp_json_file.change(\n        update_wrapper,\n        inputs=[mcp_json_file],\n        outputs=[mcp_server_config, mcp_server_config]\n    )\n"], ["/web-ui/src/browser/custom_browser.py", "import asyncio\nimport pdb\n\nfrom playwright.async_api import Browser as PlaywrightBrowser\nfrom playwright.async_api import (\n    BrowserContext as PlaywrightBrowserContext,\n)\nfrom playwright.async_api import (\n    Playwright,\n    async_playwright,\n)\nfrom browser_use.browser.browser import Browser, IN_DOCKER\nfrom browser_use.browser.context import BrowserContext, BrowserContextConfig\nfrom playwright.async_api import BrowserContext as PlaywrightBrowserContext\nimport logging\n\nfrom browser_use.browser.chrome import (\n    CHROME_ARGS,\n    CHROME_DETERMINISTIC_RENDERING_ARGS,\n    CHROME_DISABLE_SECURITY_ARGS,\n    CHROME_DOCKER_ARGS,\n    CHROME_HEADLESS_ARGS,\n)\nfrom browser_use.browser.context import BrowserContext, BrowserContextConfig\nfrom browser_use.browser.utils.screen_resolution import get_screen_resolution, get_window_adjustments\nfrom browser_use.utils import time_execution_async\nimport socket\n\nfrom .custom_context import CustomBrowserContext\n\nlogger = logging.getLogger(__name__)\n\n\nclass CustomBrowser(Browser):\n\n    async def new_context(self, config: BrowserContextConfig | None = None) -> CustomBrowserContext:\n        \"\"\"Create a browser context\"\"\"\n        browser_config = self.config.model_dump() if self.config else {}\n        context_config = config.model_dump() if config else {}\n        merged_config = {**browser_config, **context_config}\n        return CustomBrowserContext(config=BrowserContextConfig(**merged_config), browser=self)\n\n    async def _setup_builtin_browser(self, playwright: Playwright) -> PlaywrightBrowser:\n        \"\"\"Sets up and returns a Playwright Browser instance with anti-detection measures.\"\"\"\n        assert self.config.browser_binary_path is None, 'browser_binary_path should be None if trying to use the builtin browsers'\n\n        # Use the configured window size from new_context_config if available\n        if (\n                not self.config.headless\n                and hasattr(self.config, 'new_context_config')\n                and hasattr(self.config.new_context_config, 'window_width')\n                and hasattr(self.config.new_context_config, 'window_height')\n        ):\n            screen_size = {\n                'width': self.config.new_context_config.window_width,\n                'height': self.config.new_context_config.window_height,\n            }\n            offset_x, offset_y = get_window_adjustments()\n        elif self.config.headless:\n            screen_size = {'width': 1920, 'height': 1080}\n            offset_x, offset_y = 0, 0\n        else:\n            screen_size = get_screen_resolution()\n            offset_x, offset_y = get_window_adjustments()\n\n        chrome_args = {\n            f'--remote-debugging-port={self.config.chrome_remote_debugging_port}',\n            *CHROME_ARGS,\n            *(CHROME_DOCKER_ARGS if IN_DOCKER else []),\n            *(CHROME_HEADLESS_ARGS if self.config.headless else []),\n            *(CHROME_DISABLE_SECURITY_ARGS if self.config.disable_security else []),\n            *(CHROME_DETERMINISTIC_RENDERING_ARGS if self.config.deterministic_rendering else []),\n            f'--window-position={offset_x},{offset_y}',\n            f'--window-size={screen_size[\"width\"]},{screen_size[\"height\"]}',\n            *self.config.extra_browser_args,\n        }\n\n        # check if chrome remote debugging port is already taken,\n        # if so remove the remote-debugging-port arg to prevent conflicts\n        with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:\n            if s.connect_ex(('localhost', self.config.chrome_remote_debugging_port)) == 0:\n                chrome_args.remove(f'--remote-debugging-port={self.config.chrome_remote_debugging_port}')\n\n        browser_class = getattr(playwright, self.config.browser_class)\n        args = {\n            'chromium': list(chrome_args),\n            'firefox': [\n                *{\n                    '-no-remote',\n                    *self.config.extra_browser_args,\n                }\n            ],\n            'webkit': [\n                *{\n                    '--no-startup-window',\n                    *self.config.extra_browser_args,\n                }\n            ],\n        }\n\n        browser = await browser_class.launch(\n            channel='chromium',  # https://github.com/microsoft/playwright/issues/33566\n            headless=self.config.headless,\n            args=args[self.config.browser_class],\n            proxy=self.config.proxy.model_dump() if self.config.proxy else None,\n            handle_sigterm=False,\n            handle_sigint=False,\n        )\n        return browser\n"], ["/web-ui/src/webui/components/browser_settings_tab.py", "import os\nfrom distutils.util import strtobool\nimport gradio as gr\nimport logging\nfrom gradio.components import Component\n\nfrom src.webui.webui_manager import WebuiManager\nfrom src.utils import config\n\nlogger = logging.getLogger(__name__)\n\nasync def close_browser(webui_manager: WebuiManager):\n    \"\"\"\n    Close browser\n    \"\"\"\n    if webui_manager.bu_current_task and not webui_manager.bu_current_task.done():\n        webui_manager.bu_current_task.cancel()\n        webui_manager.bu_current_task = None\n\n    if webui_manager.bu_browser_context:\n        logger.info(\"⚠️ Closing browser context when changing browser config.\")\n        await webui_manager.bu_browser_context.close()\n        webui_manager.bu_browser_context = None\n\n    if webui_manager.bu_browser:\n        logger.info(\"⚠️ Closing browser when changing browser config.\")\n        await webui_manager.bu_browser.close()\n        webui_manager.bu_browser = None\n\ndef create_browser_settings_tab(webui_manager: WebuiManager):\n    \"\"\"\n    Creates a browser settings tab.\n    \"\"\"\n    input_components = set(webui_manager.get_components())\n    tab_components = {}\n\n    with gr.Group():\n        with gr.Row():\n            browser_binary_path = gr.Textbox(\n                label=\"Browser Binary Path\",\n                lines=1,\n                interactive=True,\n                placeholder=\"e.g. '/Applications/Google\\\\ Chrome.app/Contents/MacOS/Google\\\\ Chrome'\"\n            )\n            browser_user_data_dir = gr.Textbox(\n                label=\"Browser User Data Dir\",\n                lines=1,\n                interactive=True,\n                placeholder=\"Leave it empty if you use your default user data\",\n            )\n    with gr.Group():\n        with gr.Row():\n            use_own_browser = gr.Checkbox(\n                label=\"Use Own Browser\",\n                value=bool(strtobool(os.getenv(\"USE_OWN_BROWSER\", \"false\"))),\n                info=\"Use your existing browser instance\",\n                interactive=True\n            )\n            keep_browser_open = gr.Checkbox(\n                label=\"Keep Browser Open\",\n                value=bool(strtobool(os.getenv(\"KEEP_BROWSER_OPEN\", \"true\"))),\n                info=\"Keep Browser Open between Tasks\",\n                interactive=True\n            )\n            headless = gr.Checkbox(\n                label=\"Headless Mode\",\n                value=False,\n                info=\"Run browser without GUI\",\n                interactive=True\n            )\n            disable_security = gr.Checkbox(\n                label=\"Disable Security\",\n                value=False,\n                info=\"Disable browser security\",\n                interactive=True\n            )\n\n    with gr.Group():\n        with gr.Row():\n            window_w = gr.Number(\n                label=\"Window Width\",\n                value=1280,\n                info=\"Browser window width\",\n                interactive=True\n            )\n            window_h = gr.Number(\n                label=\"Window Height\",\n                value=1100,\n                info=\"Browser window height\",\n                interactive=True\n            )\n    with gr.Group():\n        with gr.Row():\n            cdp_url = gr.Textbox(\n                label=\"CDP URL\",\n                value=os.getenv(\"BROWSER_CDP\", None),\n                info=\"CDP URL for browser remote debugging\",\n                interactive=True,\n            )\n            wss_url = gr.Textbox(\n                label=\"WSS URL\",\n                info=\"WSS URL for browser remote debugging\",\n                interactive=True,\n            )\n    with gr.Group():\n        with gr.Row():\n            save_recording_path = gr.Textbox(\n                label=\"Recording Path\",\n                placeholder=\"e.g. ./tmp/record_videos\",\n                info=\"Path to save browser recordings\",\n                interactive=True,\n            )\n\n            save_trace_path = gr.Textbox(\n                label=\"Trace Path\",\n                placeholder=\"e.g. ./tmp/traces\",\n                info=\"Path to save Agent traces\",\n                interactive=True,\n            )\n\n        with gr.Row():\n            save_agent_history_path = gr.Textbox(\n                label=\"Agent History Save Path\",\n                value=\"./tmp/agent_history\",\n                info=\"Specify the directory where agent history should be saved.\",\n                interactive=True,\n            )\n            save_download_path = gr.Textbox(\n                label=\"Save Directory for browser downloads\",\n                value=\"./tmp/downloads\",\n                info=\"Specify the directory where downloaded files should be saved.\",\n                interactive=True,\n            )\n    tab_components.update(\n        dict(\n            browser_binary_path=browser_binary_path,\n            browser_user_data_dir=browser_user_data_dir,\n            use_own_browser=use_own_browser,\n            keep_browser_open=keep_browser_open,\n            headless=headless,\n            disable_security=disable_security,\n            save_recording_path=save_recording_path,\n            save_trace_path=save_trace_path,\n            save_agent_history_path=save_agent_history_path,\n            save_download_path=save_download_path,\n            cdp_url=cdp_url,\n            wss_url=wss_url,\n            window_h=window_h,\n            window_w=window_w,\n        )\n    )\n    webui_manager.add_components(\"browser_settings\", tab_components)\n\n    async def close_wrapper():\n        \"\"\"Wrapper for handle_clear.\"\"\"\n        await close_browser(webui_manager)\n\n    headless.change(close_wrapper)\n    keep_browser_open.change(close_wrapper)\n    disable_security.change(close_wrapper)\n    use_own_browser.change(close_wrapper)\n"], ["/web-ui/src/utils/utils.py", "import base64\nimport os\nimport time\nfrom pathlib import Path\nfrom typing import Dict, Optional\nimport requests\nimport json\nimport gradio as gr\nimport uuid\n\n\ndef encode_image(img_path):\n    if not img_path:\n        return None\n    with open(img_path, \"rb\") as fin:\n        image_data = base64.b64encode(fin.read()).decode(\"utf-8\")\n    return image_data\n\n\ndef get_latest_files(directory: str, file_types: list = ['.webm', '.zip']) -> Dict[str, Optional[str]]:\n    \"\"\"Get the latest recording and trace files\"\"\"\n    latest_files: Dict[str, Optional[str]] = {ext: None for ext in file_types}\n\n    if not os.path.exists(directory):\n        os.makedirs(directory, exist_ok=True)\n        return latest_files\n\n    for file_type in file_types:\n        try:\n            matches = list(Path(directory).rglob(f\"*{file_type}\"))\n            if matches:\n                latest = max(matches, key=lambda p: p.stat().st_mtime)\n                # Only return files that are complete (not being written)\n                if time.time() - latest.stat().st_mtime > 1.0:\n                    latest_files[file_type] = str(latest)\n        except Exception as e:\n            print(f\"Error getting latest {file_type} file: {e}\")\n\n    return latest_files\n"], ["/web-ui/src/webui/interface.py", "import gradio as gr\n\nfrom src.webui.webui_manager import WebuiManager\nfrom src.webui.components.agent_settings_tab import create_agent_settings_tab\nfrom src.webui.components.browser_settings_tab import create_browser_settings_tab\nfrom src.webui.components.browser_use_agent_tab import create_browser_use_agent_tab\nfrom src.webui.components.deep_research_agent_tab import create_deep_research_agent_tab\nfrom src.webui.components.load_save_config_tab import create_load_save_config_tab\n\ntheme_map = {\n    \"Default\": gr.themes.Default(),\n    \"Soft\": gr.themes.Soft(),\n    \"Monochrome\": gr.themes.Monochrome(),\n    \"Glass\": gr.themes.Glass(),\n    \"Origin\": gr.themes.Origin(),\n    \"Citrus\": gr.themes.Citrus(),\n    \"Ocean\": gr.themes.Ocean(),\n    \"Base\": gr.themes.Base()\n}\n\n\ndef create_ui(theme_name=\"Ocean\"):\n    css = \"\"\"\n    .gradio-container {\n        width: 70vw !important; \n        max-width: 70% !important; \n        margin-left: auto !important;\n        margin-right: auto !important;\n        padding-top: 10px !important;\n    }\n    .header-text {\n        text-align: center;\n        margin-bottom: 20px;\n    }\n    .tab-header-text {\n        text-align: center;\n    }\n    .theme-section {\n        margin-bottom: 10px;\n        padding: 15px;\n        border-radius: 10px;\n    }\n    \"\"\"\n\n    # dark mode in default\n    js_func = \"\"\"\n    function refresh() {\n        const url = new URL(window.location);\n\n        if (url.searchParams.get('__theme') !== 'dark') {\n            url.searchParams.set('__theme', 'dark');\n            window.location.href = url.href;\n        }\n    }\n    \"\"\"\n\n    ui_manager = WebuiManager()\n\n    with gr.Blocks(\n            title=\"Browser Use WebUI\", theme=theme_map[theme_name], css=css, js=js_func,\n    ) as demo:\n        with gr.Row():\n            gr.Markdown(\n                \"\"\"\n                # 🌐 Browser Use WebUI\n                ### Control your browser with AI assistance\n                \"\"\",\n                elem_classes=[\"header-text\"],\n            )\n\n        with gr.Tabs() as tabs:\n            with gr.TabItem(\"⚙️ Agent Settings\"):\n                create_agent_settings_tab(ui_manager)\n\n            with gr.TabItem(\"🌐 Browser Settings\"):\n                create_browser_settings_tab(ui_manager)\n\n            with gr.TabItem(\"🤖 Run Agent\"):\n                create_browser_use_agent_tab(ui_manager)\n\n            with gr.TabItem(\"🎁 Agent Marketplace\"):\n                gr.Markdown(\n                    \"\"\"\n                    ### Agents built on Browser-Use\n                    \"\"\",\n                    elem_classes=[\"tab-header-text\"],\n                )\n                with gr.Tabs():\n                    with gr.TabItem(\"Deep Research\"):\n                        create_deep_research_agent_tab(ui_manager)\n\n            with gr.TabItem(\"📁 Load & Save Config\"):\n                create_load_save_config_tab(ui_manager)\n\n    return demo\n"], ["/web-ui/src/browser/custom_context.py", "import json\nimport logging\nimport os\n\nfrom browser_use.browser.browser import Browser, IN_DOCKER\nfrom browser_use.browser.context import BrowserContext, BrowserContextConfig\nfrom playwright.async_api import Browser as PlaywrightBrowser\nfrom playwright.async_api import BrowserContext as PlaywrightBrowserContext\nfrom typing import Optional\nfrom browser_use.browser.context import BrowserContextState\n\nlogger = logging.getLogger(__name__)\n\n\nclass CustomBrowserContext(BrowserContext):\n    def __init__(\n            self,\n            browser: 'Browser',\n            config: BrowserContextConfig | None = None,\n            state: Optional[BrowserContextState] = None,\n    ):\n        super(CustomBrowserContext, self).__init__(browser=browser, config=config, state=state)\n"], ["/web-ui/src/webui/components/load_save_config_tab.py", "import gradio as gr\nfrom gradio.components import Component\n\nfrom src.webui.webui_manager import WebuiManager\nfrom src.utils import config\n\n\ndef create_load_save_config_tab(webui_manager: WebuiManager):\n    \"\"\"\n    Creates a load and save config tab.\n    \"\"\"\n    input_components = set(webui_manager.get_components())\n    tab_components = {}\n\n    config_file = gr.File(\n        label=\"Load UI Settings from json\",\n        file_types=[\".json\"],\n        interactive=True\n    )\n    with gr.Row():\n        load_config_button = gr.Button(\"Load Config\", variant=\"primary\")\n        save_config_button = gr.Button(\"Save UI Settings\", variant=\"primary\")\n\n    config_status = gr.Textbox(\n        label=\"Status\",\n        lines=2,\n        interactive=False\n    )\n\n    tab_components.update(dict(\n        load_config_button=load_config_button,\n        save_config_button=save_config_button,\n        config_status=config_status,\n        config_file=config_file,\n    ))\n\n    webui_manager.add_components(\"load_save_config\", tab_components)\n\n    save_config_button.click(\n        fn=webui_manager.save_config,\n        inputs=set(webui_manager.get_components()),\n        outputs=[config_status]\n    )\n\n    load_config_button.click(\n        fn=webui_manager.load_config,\n        inputs=[config_file],\n        outputs=webui_manager.get_components(),\n    )\n\n"], ["/web-ui/src/utils/config.py", "PROVIDER_DISPLAY_NAMES = {\n    \"openai\": \"OpenAI\",\n    \"azure_openai\": \"Azure OpenAI\",\n    \"anthropic\": \"Anthropic\",\n    \"deepseek\": \"DeepSeek\",\n    \"google\": \"Google\",\n    \"alibaba\": \"Alibaba\",\n    \"moonshot\": \"MoonShot\",\n    \"unbound\": \"Unbound AI\",\n    \"ibm\": \"IBM\",\n    \"grok\": \"Grok\",\n}\n\n# Predefined model names for common providers\nmodel_names = {\n    \"anthropic\": [\"claude-3-5-sonnet-20241022\", \"claude-3-5-sonnet-20240620\", \"claude-3-opus-20240229\"],\n    \"openai\": [\"gpt-4o\", \"gpt-4\", \"gpt-3.5-turbo\", \"o3-mini\"],\n    \"deepseek\": [\"deepseek-chat\", \"deepseek-reasoner\"],\n    \"google\": [\"gemini-2.0-flash\", \"gemini-2.0-flash-thinking-exp\", \"gemini-1.5-flash-latest\",\n               \"gemini-1.5-flash-8b-latest\", \"gemini-2.0-flash-thinking-exp-01-21\", \"gemini-2.0-pro-exp-02-05\",\n               \"gemini-2.5-pro-preview-03-25\", \"gemini-2.5-flash-preview-04-17\"],\n    \"ollama\": [\"qwen2.5:7b\", \"qwen2.5:14b\", \"qwen2.5:32b\", \"qwen2.5-coder:14b\", \"qwen2.5-coder:32b\", \"llama2:7b\",\n               \"deepseek-r1:14b\", \"deepseek-r1:32b\"],\n    \"azure_openai\": [\"gpt-4o\", \"gpt-4\", \"gpt-3.5-turbo\"],\n    \"mistral\": [\"pixtral-large-latest\", \"mistral-large-latest\", \"mistral-small-latest\", \"ministral-8b-latest\"],\n    \"alibaba\": [\"qwen-plus\", \"qwen-max\", \"qwen-vl-max\", \"qwen-vl-plus\", \"qwen-turbo\", \"qwen-long\"],\n    \"moonshot\": [\"moonshot-v1-32k-vision-preview\", \"moonshot-v1-8k-vision-preview\"],\n    \"unbound\": [\"gemini-2.0-flash\", \"gpt-4o-mini\", \"gpt-4o\", \"gpt-4.5-preview\"],\n    \"grok\": [\n        \"grok-3\",\n        \"grok-3-fast\",\n        \"grok-3-mini\",\n        \"grok-3-mini-fast\",\n        \"grok-2-vision\",\n        \"grok-2-image\",\n        \"grok-2\",\n    ],\n    \"siliconflow\": [\n        \"deepseek-ai/DeepSeek-R1\",\n        \"deepseek-ai/DeepSeek-V3\",\n        \"deepseek-ai/DeepSeek-R1-Distill-Qwen-32B\",\n        \"deepseek-ai/DeepSeek-R1-Distill-Qwen-14B\",\n        \"deepseek-ai/DeepSeek-R1-Distill-Qwen-7B\",\n        \"deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B\",\n        \"deepseek-ai/DeepSeek-V2.5\",\n        \"deepseek-ai/deepseek-vl2\",\n        \"Qwen/Qwen2.5-72B-Instruct-128K\",\n        \"Qwen/Qwen2.5-72B-Instruct\",\n        \"Qwen/Qwen2.5-32B-Instruct\",\n        \"Qwen/Qwen2.5-14B-Instruct\",\n        \"Qwen/Qwen2.5-7B-Instruct\",\n        \"Qwen/Qwen2.5-Coder-32B-Instruct\",\n        \"Qwen/Qwen2.5-Coder-7B-Instruct\",\n        \"Qwen/Qwen2-7B-Instruct\",\n        \"Qwen/Qwen2-1.5B-Instruct\",\n        \"Qwen/QwQ-32B-Preview\",\n        \"Qwen/Qwen2-VL-72B-Instruct\",\n        \"Qwen/Qwen2.5-VL-32B-Instruct\",\n        \"Qwen/Qwen2.5-VL-72B-Instruct\",\n        \"TeleAI/TeleChat2\",\n        \"THUDM/glm-4-9b-chat\",\n        \"Vendor-A/Qwen/Qwen2.5-72B-Instruct\",\n        \"internlm/internlm2_5-7b-chat\",\n        \"internlm/internlm2_5-20b-chat\",\n        \"Pro/Qwen/Qwen2.5-7B-Instruct\",\n        \"Pro/Qwen/Qwen2-7B-Instruct\",\n        \"Pro/Qwen/Qwen2-1.5B-Instruct\",\n        \"Pro/THUDM/chatglm3-6b\",\n        \"Pro/THUDM/glm-4-9b-chat\",\n    ],\n    \"ibm\": [\"ibm/granite-vision-3.1-2b-preview\", \"meta-llama/llama-4-maverick-17b-128e-instruct-fp8\",\n            \"meta-llama/llama-3-2-90b-vision-instruct\"],\n    \"modelscope\":[\n        \"Qwen/Qwen2.5-Coder-32B-Instruct\",\n        \"Qwen/Qwen2.5-Coder-14B-Instruct\",\n        \"Qwen/Qwen2.5-Coder-7B-Instruct\",\n        \"Qwen/Qwen2.5-72B-Instruct\",\n        \"Qwen/Qwen2.5-32B-Instruct\",\n        \"Qwen/Qwen2.5-14B-Instruct\",\n        \"Qwen/Qwen2.5-7B-Instruct\",\n        \"Qwen/QwQ-32B-Preview\",\n        \"Qwen/Qwen2.5-VL-3B-Instruct\",\n        \"Qwen/Qwen2.5-VL-7B-Instruct\",\n        \"Qwen/Qwen2.5-VL-32B-Instruct\",\n        \"Qwen/Qwen2.5-VL-72B-Instruct\",\n        \"deepseek-ai/DeepSeek-R1-Distill-Qwen-32B\",\n        \"deepseek-ai/DeepSeek-R1-Distill-Qwen-14B\",\n        \"deepseek-ai/DeepSeek-R1-Distill-Qwen-7B\",\n        \"deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B\",\n        \"deepseek-ai/DeepSeek-R1\",\n        \"deepseek-ai/DeepSeek-V3\",\n        \"Qwen/Qwen3-1.7B\",\n        \"Qwen/Qwen3-4B\",\n        \"Qwen/Qwen3-8B\",\n        \"Qwen/Qwen3-14B\",\n        \"Qwen/Qwen3-30B-A3B\",\n        \"Qwen/Qwen3-32B\",\n        \"Qwen/Qwen3-235B-A22B\",\n    ],\n}\n"], ["/web-ui/webui.py", "from dotenv import load_dotenv\nload_dotenv()\nimport argparse\nfrom src.webui.interface import theme_map, create_ui\n\n\ndef main():\n    parser = argparse.ArgumentParser(description=\"Gradio WebUI for Browser Agent\")\n    parser.add_argument(\"--ip\", type=str, default=\"127.0.0.1\", help=\"IP address to bind to\")\n    parser.add_argument(\"--port\", type=int, default=7788, help=\"Port to listen on\")\n    parser.add_argument(\"--theme\", type=str, default=\"Ocean\", choices=theme_map.keys(), help=\"Theme to use for the UI\")\n    args = parser.parse_args()\n\n    demo = create_ui(theme_name=args.theme)\n    demo.queue().launch(server_name=args.ip, server_port=args.port)\n\n\nif __name__ == '__main__':\n    main()\n"], ["/web-ui/src/browser/__init__.py", ""], ["/web-ui/src/controller/__init__.py", ""], ["/web-ui/src/webui/components/__init__.py", ""], ["/web-ui/src/webui/__init__.py", ""], ["/web-ui/src/utils/__init__.py", ""], ["/web-ui/src/agent/__init__.py", ""], ["/web-ui/src/__init__.py", ""]], "task_instance_info": {"created_time": "2025-08-20 15:37:48", "created_task_model": "DeepSeek-R1", "code_description": "This code segment functions to: Parse a hierarchical research plan from a markdown document and track execution progress by identifying pending tasks while maintaining category-task structure integrity."}, "inference_content": {"inference_model": "DeepSeek-R1", "inference_result": "```python\n[TASK_BEGIN]\n                current_category = None\n                for line in lines:\n                    line = line.strip()\n                    if line.startswith('## '):\n                        # Handle category line\n                        cat_name = line[3:].strip().split('. ', 1)[-1]  # Remove leading numbers if present\n                        current_category = ResearchCategoryItem(\n                            category_name=cat_name,\n                            tasks=[]\n                        )\n                        loaded_plan.append(current_category)\n                    elif line.startswith('- [') and current_category is not None:\n                        # Handle task line with status marker\n                        marker = line[2:5]\n                        status = \"pending\"\n                        if marker == '[x]':\n                            status = \"completed\"\n                        elif marker == '[-]':\n                            status = \"failed\"\n                        \n                        task_desc = line[5:].strip().lstrip('- ').strip()\n                        task_item = ResearchTaskItem(\n                            task_description=task_desc,\n                            status=status,\n                            queries=None,\n                            result_summary=None\n                        )\n                        current_category['tasks'].append(task_item)\n                        \n                        # Track first pending task\n                        if not found_pending and status == \"pending\":\n                            next_cat_idx = len(loaded_plan) - 1\n                            next_task_idx = len(current_category['tasks']) - 1\n                            found_pending = True\n\n                # Update state with loaded plan and indices\n                if loaded_plan:\n                    state_updates[\"research_plan\"] = loaded_plan\n                    state_updates[\"current_category_index\"] = next_cat_idx\n                    state_updates[\"current_task_index_in_category\"] = next_task_idx\n                    logger.info(f\"Loaded research plan with {len(loaded_plan)} categories\")\n[TASK_END]\n```", "inference_time": "2025-08-20 15-38-57"}, "editdistance_info": {"edit_distance": 25.5725, "calculate_time": "2025-08-20 15:38:57", "true_code_clean": "if os.path.exists(plan_file):\n        try:\n            with open(plan_file, \"r\", encoding=\"utf-8\") as f:\n                current_category: Optional[ResearchCategoryItem] = None\n                lines = f.readlines()\n                cat_counter = 0\n                task_counter_in_cat = 0\n                for line_num, line_content in enumerate(lines):\n                    line = line_content.strip()\n                    if line.startswith(\"\n                        if current_category:  \n                            loaded_plan.append(current_category)\n                            if not found_pending:  \n                                cat_counter += 1\n                                task_counter_in_cat = 0\n                        category_name = line[line.find(\" \"):].strip()  \n                        current_category = ResearchCategoryItem(category_name=category_name, tasks=[])\n                    elif (line.startswith(\"- [ ]\") or line.startswith(\"- [x]\") or line.startswith(\n                            \"- [-]\")) and current_category:  \n                        status = \"pending\"\n                        if line.startswith(\"- [x]\"):\n                            status = \"completed\"\n                        elif line.startswith(\"- [-]\"):\n                            status = \"failed\"\n                        task_desc = line[5:].strip()\n                        current_category[\"tasks\"].append(\n                            ResearchTaskItem(task_description=task_desc, status=status, queries=None,\n                                             result_summary=None)\n                        )\n                        if status == \"pending\" and not found_pending:\n                            next_cat_idx = cat_counter\n                            next_task_idx = task_counter_in_cat\n                            found_pending = True\n                        if not found_pending:  \n                            task_counter_in_cat += 1\n                if current_category:  \n                    loaded_plan.append(current_category)\n            if loaded_plan:\n                state_updates[\"research_plan\"] = loaded_plan\n                if not found_pending and loaded_plan:  \n                    next_cat_idx = len(loaded_plan)  \n                    next_task_idx = 0\n                state_updates[\"current_category_index\"] = next_cat_idx\n                state_updates[\"current_task_index_in_category\"] = next_task_idx\n                logger.info(\n                    f\"Loaded hierarchical research plan from {plan_file}. \"\n                    f\"Next task: Category {next_cat_idx}, Task {next_task_idx} in category.\"\n                )\n            else:\n                logger.warning(f\"Plan file {plan_file} was empty or malformed.\")\n        except Exception as e:\n            logger.error(f\"Failed to load or parse research plan {plan_file}: {e}\", exc_info=True)\n            state_updates[\"error_message\"] = f\"Failed to load research plan: {e}\"\n    else:\n        logger.info(f\"Plan file {plan_file} not found. Will start fresh.\")", "predict_code_clean": "current_category = None\n                for line in lines:\n                    line = line.strip()\n                    if line.startswith('\n                        cat_name = line[3:].strip().split('. ', 1)[-1]  \n                        current_category = ResearchCategoryItem(\n                            category_name=cat_name,\n                            tasks=[]\n                        )\n                        loaded_plan.append(current_category)\n                    elif line.startswith('- [') and current_category is not None:\n                        marker = line[2:5]\n                        status = \"pending\"\n                        if marker == '[x]':\n                            status = \"completed\"\n                        elif marker == '[-]':\n                            status = \"failed\"\n                        task_desc = line[5:].strip().lstrip('- ').strip()\n                        task_item = ResearchTaskItem(\n                            task_description=task_desc,\n                            status=status,\n                            queries=None,\n                            result_summary=None\n                        )\n                        current_category['tasks'].append(task_item)\n                        if not found_pending and status == \"pending\":\n                            next_cat_idx = len(loaded_plan) - 1\n                            next_task_idx = len(current_category['tasks']) - 1\n                            found_pending = True\n                if loaded_plan:\n                    state_updates[\"research_plan\"] = loaded_plan\n                    state_updates[\"current_category_index\"] = next_cat_idx\n                    state_updates[\"current_task_index_in_category\"] = next_task_idx\n                    logger.info(f\"Loaded research plan with {len(loaded_plan)} categories\")"}}