import asyncio import gc import inspect import json import logging import re import tempfile import time from collections.abc import Awaitable, Callable from pathlib import Path from typing import TYPE_CHECKING, Any, Generic, Literal, TypeVar, cast from urllib.parse import urlparse if TYPE_CHECKING: from browser_use.skills.views import Skill from dotenv import load_dotenv from browser_use.agent.cloud_events import ( CreateAgentOutputFileEvent, CreateAgentSessionEvent, CreateAgentStepEvent, CreateAgentTaskEvent, UpdateAgentTaskEvent, ) from browser_use.agent.message_manager.utils import save_conversation from browser_use.llm.base import BaseChatModel from browser_use.llm.exceptions import ModelOutputTruncatedError, ModelProviderError, ModelRateLimitError from browser_use.llm.messages import BaseMessage, ContentPartImageParam, ContentPartTextParam, UserMessage from browser_use.tokens.service import TokenCost load_dotenv() from bubus import EventBus from pydantic import BaseModel, ValidationError from uuid_extensions import uuid7str from browser_use import Browser, BrowserProfile, BrowserSession from browser_use.agent.judge import construct_judge_messages # Lazy import for gif to avoid heavy agent.views import at startup # from browser_use.agent.gif import create_history_gif from browser_use.agent.message_manager.service import ( MessageManager, ) from browser_use.agent.prompts import SystemPrompt from browser_use.agent.views import ( ActionResult, AgentError, AgentHistory, AgentHistoryList, AgentOutput, AgentSettings, AgentState, AgentStepInfo, AgentStructuredOutput, BrowserStateHistory, DetectedVariable, JudgementResult, MessageCompactionSettings, PlanItem, StepMetadata, ) from browser_use.browser.events import _get_timeout from browser_use.browser.session import DEFAULT_BROWSER_PROFILE from browser_use.browser.views import BrowserStateSummary from browser_use.config import CONFIG from browser_use.dom.views import DOMInteractedElement, MatchLevel from browser_use.filesystem.file_system import FileSystem from browser_use.observability import observe, observe_debug from browser_use.telemetry.service import ProductTelemetry from browser_use.telemetry.views import AgentTelemetryEvent from browser_use.tools.registry.views import ActionModel from browser_use.tools.service import Tools from browser_use.utils import ( URL_PATTERN, _log_pretty_path, check_latest_browser_use_version, get_browser_use_version, is_placeholder_url, sanitize_url_candidate, time_execution_async, time_execution_sync, ) logger = logging.getLogger(__name__) def log_response(response: AgentOutput, registry=None, logger=None) -> None: """Utility function to log the model's response.""" # Use module logger if no logger provided if logger is None: logger = logging.getLogger(__name__) # Only log thinking if it's present if response.current_state.thinking: logger.debug(f'πŸ’‘ Thinking:\n{response.current_state.thinking}') # Only log evaluation if it's not empty eval_goal = response.current_state.evaluation_previous_goal if eval_goal: if 'success' in eval_goal.lower(): emoji = 'πŸ‘' # Green color for success logger.info(f' \033[32m{emoji} Eval: {eval_goal}\033[0m') elif 'failure' in eval_goal.lower(): emoji = '⚠️' # Red color for failure logger.info(f' \033[31m{emoji} Eval: {eval_goal}\033[0m') else: emoji = '❔' # No color for unknown/neutral logger.info(f' {emoji} Eval: {eval_goal}') # Always log memory if present if response.current_state.memory: logger.info(f' 🧠 Memory: {response.current_state.memory}') # Only log next goal if it's not empty next_goal = response.current_state.next_goal if next_goal: # Blue color for next goal logger.info(f' \033[34m🎯 Next goal: {next_goal}\033[0m') Context = TypeVar('Context') AgentHookFunc = Callable[['Agent'], Awaitable[None]] class Agent(Generic[Context, AgentStructuredOutput]): @time_execution_sync('--init') def __init__( self, task: str, llm: BaseChatModel | None = None, # Optional parameters browser_profile: BrowserProfile | None = None, browser_session: BrowserSession | None = None, browser: Browser | None = None, # Alias for browser_session tools: Tools[Context] | None = None, controller: Tools[Context] | None = None, # Alias for tools # Skills integration skill_ids: list[str | Literal['*']] | None = None, skills: list[str | Literal['*']] | None = None, # Alias for skill_ids skill_service: Any | None = None, # Initial agent run parameters sensitive_data: dict[str, str | dict[str, str]] | None = None, initial_actions: list[dict[str, dict[str, Any]]] | None = None, # Cloud Callbacks register_new_step_callback: ( Callable[['BrowserStateSummary', 'AgentOutput', int], None] # Sync callback | Callable[['BrowserStateSummary', 'AgentOutput', int], Awaitable[None]] # Async callback | None ) = None, register_done_callback: ( Callable[['AgentHistoryList'], Awaitable[None]] # Async Callback | Callable[['AgentHistoryList'], None] # Sync Callback | None ) = None, register_external_agent_status_raise_error_callback: Callable[[], Awaitable[bool]] | None = None, register_should_stop_callback: Callable[[], Awaitable[bool]] | None = None, # Agent settings output_model_schema: type[AgentStructuredOutput] | None = None, extraction_schema: dict | None = None, use_vision: bool | Literal['auto'] = True, save_conversation_path: str | Path | None = None, save_conversation_path_encoding: str | None = 'utf-8', max_failures: int = 5, override_system_message: str | None = None, extend_system_message: str | None = None, generate_gif: bool | str = False, available_file_paths: list[str] | None = None, include_attributes: list[str] | None = None, max_actions_per_step: int = 5, use_thinking: bool = True, flash_mode: bool = False, demo_mode: bool | None = None, max_history_items: int | None = None, page_extraction_llm: BaseChatModel | None = None, fallback_llm: BaseChatModel | None = None, use_judge: bool = True, ground_truth: str | None = None, judge_llm: BaseChatModel | None = None, injected_agent_state: AgentState | None = None, source: str | None = None, file_system_path: str | None = None, task_id: str | None = None, calculate_cost: bool = False, pricing_url: str | None = None, display_files_in_done_text: bool = True, include_tool_call_examples: bool = False, vision_detail_level: Literal['auto', 'low', 'high'] = 'auto', llm_timeout: int | None = None, step_timeout: int = 180, directly_open_url: bool = True, include_recent_events: bool = False, sample_images: list[ContentPartTextParam | ContentPartImageParam] | None = None, final_response_after_failure: bool = True, enable_planning: bool = True, planning_replan_on_stall: int = 3, planning_exploration_limit: int = 5, loop_detection_window: int = 20, loop_detection_enabled: bool = True, llm_screenshot_size: tuple[int, int] | None = None, message_compaction: MessageCompactionSettings | bool | None = True, max_clickable_elements_length: int = 40000, _url_shortening_limit: int = 25, enable_signal_handler: bool = True, **kwargs, ): # Validate llm_screenshot_size if llm_screenshot_size is not None: if not isinstance(llm_screenshot_size, tuple) or len(llm_screenshot_size) != 2: raise ValueError('llm_screenshot_size must be a tuple of (width, height)') width, height = llm_screenshot_size if not isinstance(width, int) or not isinstance(height, int): raise ValueError('llm_screenshot_size dimensions must be integers') if width < 100 or height < 100: raise ValueError('llm_screenshot_size dimensions must be at least 100 pixels') self.logger.info(f'πŸ–ΌοΈ LLM screenshot resizing enabled: {width}x{height}') if llm is None: default_llm_name = CONFIG.DEFAULT_LLM if default_llm_name: from browser_use.llm.models import get_llm_by_name llm = get_llm_by_name(default_llm_name) else: # No default LLM specified, use the original default from browser_use import ChatBrowserUse llm = ChatBrowserUse() # set flashmode = True if llm is ChatBrowserUse if llm.provider == 'browser-use': flash_mode = True # Flash mode strips plan fields from the output schema, so planning is structurally impossible if flash_mode: enable_planning = False # Auto-configure llm_screenshot_size for Claude Sonnet, including gateway ids like # 'anthropic/claude-sonnet-4-6' (rsplit drops the provider prefix before matching). if llm_screenshot_size is None: model_name = getattr(llm, 'model', '') if isinstance(model_name, str) and model_name.rsplit('/', 1)[-1].startswith('claude-sonnet'): llm_screenshot_size = (1400, 850) logger.info('πŸ–ΌοΈ Auto-configured LLM screenshot size for Claude Sonnet: 1400x850') if page_extraction_llm is None: page_extraction_llm = llm if judge_llm is None: judge_llm = llm if available_file_paths is None: available_file_paths = [] # Set timeout based on model name if not explicitly provided if llm_timeout is None: def _get_model_timeout(llm_model: BaseChatModel) -> int: """Determine timeout based on model name""" model_name = getattr(llm_model, 'model', '').lower() if 'gemini' in model_name: if '3-pro' in model_name: return 90 return 75 elif 'groq' in model_name: return 30 elif 'o3' in model_name or 'claude' in model_name or 'sonnet' in model_name or 'deepseek' in model_name: return 90 else: return 75 # Default timeout llm_timeout = _get_model_timeout(llm) self.id = task_id or uuid7str() self.task_id: str = self.id self.session_id: str = uuid7str() base_profile = browser_profile or DEFAULT_BROWSER_PROFILE if base_profile is DEFAULT_BROWSER_PROFILE: base_profile = base_profile.model_copy() if demo_mode is not None and base_profile.demo_mode != demo_mode: base_profile = base_profile.model_copy(update={'demo_mode': demo_mode}) browser_profile = base_profile # Handle browser vs browser_session parameter (browser takes precedence) if browser and browser_session: raise ValueError('Cannot specify both "browser" and "browser_session" parameters. Use "browser" for the cleaner API.') browser_session = browser or browser_session if browser_session is not None and demo_mode is not None and browser_session.browser_profile.demo_mode != demo_mode: browser_session.browser_profile = browser_session.browser_profile.model_copy(update={'demo_mode': demo_mode}) self.browser_session = browser_session or BrowserSession( browser_profile=browser_profile, id=uuid7str()[:-4] + self.id[-4:], # re-use the same 4-char suffix so they show up together in logs ) self._demo_mode_enabled: bool = bool(self.browser_profile.demo_mode) if self.browser_session else False if self._demo_mode_enabled and getattr(self.browser_profile, 'headless', False): self.logger.warning( 'Demo mode is enabled but the browser is headless=True; set headless=False to view the in-browser panel.' ) # Initialize available file paths as direct attribute self.available_file_paths = available_file_paths # Set up tools first (needed to detect output_model_schema) if tools is not None: self.tools = tools elif controller is not None: self.tools = controller else: # Exclude screenshot tool when use_vision is not auto exclude_actions = ['screenshot'] if use_vision != 'auto' else [] self.tools = Tools(exclude_actions=exclude_actions, display_files_in_done_text=display_files_in_done_text) # Enforce screenshot exclusion when use_vision != 'auto', even if user passed custom tools if use_vision != 'auto': self.tools.exclude_action('screenshot') # Enable coordinate clicking for models that support it model_name = getattr(llm, 'model', '').lower() supports_coordinate_clicking = any( pattern in model_name for pattern in ['claude-sonnet-4', 'claude-opus-4', 'claude-fable-5', 'gemini-3-pro', 'browser-use/'] ) if supports_coordinate_clicking: self.tools.set_coordinate_clicking(True) # Handle skills vs skill_ids parameter (skills takes precedence) if skills and skill_ids: raise ValueError('Cannot specify both "skills" and "skill_ids" parameters. Use "skills" for the cleaner API.') skill_ids = skills or skill_ids # Skills integration - use injected service or create from skill_ids self.skill_service = None self._skills_registered = False if skill_service is not None: self.skill_service = skill_service elif skill_ids: from browser_use.skills import SkillService self.skill_service = SkillService(skill_ids=skill_ids) # Structured output - use explicit param or detect from tools tools_output_model = self.tools.get_output_model() if output_model_schema is not None and tools_output_model is not None: # Both provided - warn if they differ if output_model_schema is not tools_output_model: logger.warning( f'output_model_schema ({output_model_schema.__name__}) differs from Tools output_model ' f'({tools_output_model.__name__}). Using Agent output_model_schema.' ) elif output_model_schema is None and tools_output_model is not None: # Only tools has it - use that (cast is safe: both are BaseModel subclasses) output_model_schema = cast(type[AgentStructuredOutput], tools_output_model) self.output_model_schema = output_model_schema if self.output_model_schema is not None: self.tools.use_structured_output_action(self.output_model_schema) # Extraction schema: explicit param takes priority, otherwise auto-bridge from output_model_schema self.extraction_schema = extraction_schema if self.extraction_schema is None and self.output_model_schema is not None: self.extraction_schema = self.output_model_schema.model_json_schema() # Core components - task enhancement now has access to output_model_schema from tools self.task = self._enhance_task_with_schema(task, output_model_schema) self.llm = llm self.judge_llm = judge_llm # Fallback LLM configuration self._fallback_llm: BaseChatModel | None = fallback_llm self._using_fallback_llm: bool = False self._original_llm: BaseChatModel = llm # Store original for reference self.directly_open_url = directly_open_url self.include_recent_events = include_recent_events self._url_shortening_limit = _url_shortening_limit self.sensitive_data = sensitive_data self.sample_images = sample_images if isinstance(message_compaction, bool): message_compaction = MessageCompactionSettings(enabled=message_compaction) self.settings = AgentSettings( use_vision=use_vision, vision_detail_level=vision_detail_level, save_conversation_path=save_conversation_path, save_conversation_path_encoding=save_conversation_path_encoding, max_failures=max_failures, override_system_message=override_system_message, extend_system_message=extend_system_message, generate_gif=generate_gif, include_attributes=include_attributes, max_actions_per_step=max_actions_per_step, use_thinking=use_thinking, flash_mode=flash_mode, max_history_items=max_history_items, page_extraction_llm=page_extraction_llm, calculate_cost=calculate_cost, include_tool_call_examples=include_tool_call_examples, llm_timeout=llm_timeout, step_timeout=step_timeout, final_response_after_failure=final_response_after_failure, use_judge=use_judge, ground_truth=ground_truth, enable_planning=enable_planning, planning_replan_on_stall=planning_replan_on_stall, planning_exploration_limit=planning_exploration_limit, loop_detection_window=loop_detection_window, loop_detection_enabled=loop_detection_enabled, message_compaction=message_compaction, max_clickable_elements_length=max_clickable_elements_length, ) # Token cost service self.token_cost_service = TokenCost(include_cost=calculate_cost, pricing_url=pricing_url) self.token_cost_service.register_llm(llm) self.token_cost_service.register_llm(page_extraction_llm) self.token_cost_service.register_llm(judge_llm) if self.settings.message_compaction and self.settings.message_compaction.compaction_llm: self.token_cost_service.register_llm(self.settings.message_compaction.compaction_llm) # Store signal handler setting (not part of AgentSettings as it's runtime behavior) self.enable_signal_handler = enable_signal_handler # Initialize state self.state = injected_agent_state or AgentState() # Configure loop detector window size from settings self.state.loop_detector.window_size = self.settings.loop_detection_window # Initialize history self.history = AgentHistoryList(history=[], usage=None) # Initialize agent directory import time timestamp = int(time.time()) base_tmp = Path(tempfile.gettempdir()) self.agent_directory = base_tmp / f'browser_use_agent_{self.id}_{timestamp}' # Initialize file system and screenshot service self._set_file_system(file_system_path) self._set_screenshot_service() # Action setup self._setup_action_models() self._set_browser_use_version_and_source(source) initial_url = None # only load url if no initial actions are provided if self.directly_open_url and not self.state.follow_up_task and not initial_actions: initial_url = self._extract_start_url(self.task) if initial_url: self.logger.info(f'πŸ”— Found URL in task: {initial_url}, adding as initial action...') initial_actions = [{'navigate': {'url': initial_url, 'new_tab': False}}] self.initial_url = initial_url self.initial_actions = self._convert_initial_actions(initial_actions) if initial_actions else None # Verify we can connect to the model self._verify_and_setup_llm() # TODO: move this logic to the LLMs # Handle users trying to use use_vision=True with DeepSeek models if 'deepseek' in self.llm.model.lower(): self.logger.warning('⚠️ DeepSeek models do not support use_vision=True yet. Setting use_vision=False for now...') self.settings.use_vision = False # Handle users trying to use use_vision=True with XAI models that don't support it # grok-3 variants and grok-code don't support vision; grok-2 and grok-4 do model_lower = self.llm.model.lower() if 'grok-3' in model_lower or 'grok-code' in model_lower: self.logger.warning('⚠️ This XAI model does not support use_vision=True yet. Setting use_vision=False for now...') self.settings.use_vision = False logger.debug( f'{" +vision" if self.settings.use_vision else ""}' f' extraction_model={self.settings.page_extraction_llm.model if self.settings.page_extraction_llm else "Unknown"}' f'{" +file_system" if self.file_system else ""}' ) # Store llm_screenshot_size in browser_session so tools can access it self.browser_session.llm_screenshot_size = llm_screenshot_size # Check if LLM is ChatAnthropic instance from browser_use.llm.anthropic.chat import ChatAnthropic is_anthropic = isinstance(self.llm, ChatAnthropic) # Check if model is a browser-use fine-tuned model (uses simplified prompts) is_browser_use_model = 'browser-use/' in self.llm.model.lower() # Initialize message manager with state # Initial system prompt with all actions - will be updated during each step self._message_manager = MessageManager( task=self.task, system_message=SystemPrompt( max_actions_per_step=self.settings.max_actions_per_step, override_system_message=override_system_message, extend_system_message=extend_system_message, use_thinking=self.settings.use_thinking, flash_mode=self.settings.flash_mode, is_anthropic=is_anthropic, is_browser_use_model=is_browser_use_model, model_name=self.llm.model, ).get_system_message(), file_system=self.file_system, state=self.state.message_manager_state, use_thinking=self.settings.use_thinking, # Settings that were previously in MessageManagerSettings include_attributes=self.settings.include_attributes, sensitive_data=sensitive_data, max_history_items=self.settings.max_history_items, vision_detail_level=self.settings.vision_detail_level, include_tool_call_examples=self.settings.include_tool_call_examples, include_recent_events=self.include_recent_events, sample_images=self.sample_images, llm_screenshot_size=llm_screenshot_size, max_clickable_elements_length=self.settings.max_clickable_elements_length, ) if self.sensitive_data: # Check if sensitive_data has domain-specific credentials has_domain_specific_credentials = any(isinstance(v, dict) for v in self.sensitive_data.values()) # If no allowed_domains are configured, show a security warning if not self.browser_profile.allowed_domains: self.logger.warning( '⚠️ Agent(sensitive_data=β€’β€’β€’β€’β€’β€’β€’β€’) was provided but Browser(allowed_domains=[...]) is not locked down! ⚠️\n' ' ☠️ If the agent visits a malicious website and encounters a prompt-injection attack, your sensitive_data may be exposed!\n\n' ' \n' ) # If we're using domain-specific credentials, validate domain patterns elif has_domain_specific_credentials: # For domain-specific format, ensure all domain patterns are included in allowed_domains domain_patterns = [k for k, v in self.sensitive_data.items() if isinstance(v, dict)] # Validate each domain pattern against allowed_domains for domain_pattern in domain_patterns: is_allowed = False for allowed_domain in self.browser_profile.allowed_domains: # Special cases that don't require URL matching if domain_pattern == allowed_domain or allowed_domain == '*': is_allowed = True break # Need to create example URLs to compare the patterns # Extract the domain parts, ignoring scheme pattern_domain = domain_pattern.split('://')[-1] if '://' in domain_pattern else domain_pattern allowed_domain_part = allowed_domain.split('://')[-1] if '://' in allowed_domain else allowed_domain # Check if pattern is covered by an allowed domain # Example: "google.com" is covered by "*.google.com" if pattern_domain == allowed_domain_part or ( allowed_domain_part.startswith('*.') and ( pattern_domain == allowed_domain_part[2:] or pattern_domain.endswith('.' + allowed_domain_part[2:]) ) ): is_allowed = True break if not is_allowed: self.logger.warning( f'⚠️ Domain pattern "{domain_pattern}" in sensitive_data is not covered by any pattern in allowed_domains={self.browser_profile.allowed_domains}\n' f' This may be a security risk as credentials could be used on unintended domains.' ) # Callbacks self.register_new_step_callback = register_new_step_callback self.register_done_callback = register_done_callback self.register_should_stop_callback = register_should_stop_callback self.register_external_agent_status_raise_error_callback = register_external_agent_status_raise_error_callback # Telemetry self.telemetry = ProductTelemetry() # Event bus with WAL persistence # Default to ~/.config/browseruse/events/{agent_session_id}.jsonl # wal_path = CONFIG.BROWSER_USE_CONFIG_DIR / 'events' / f'{self.session_id}.jsonl' self.eventbus = EventBus(name=f'Agent_{str(self.id)[-4:]}') if self.settings.save_conversation_path: self.settings.save_conversation_path = Path(self.settings.save_conversation_path).expanduser().resolve() self.logger.info(f'πŸ’¬ Saving conversation to {_log_pretty_path(self.settings.save_conversation_path)}') # Initialize download tracking assert self.browser_session is not None, 'BrowserSession is not set up' self.has_downloads_path = self.browser_session.browser_profile.downloads_path is not None if self.has_downloads_path: self._last_known_downloads: list[str] = [] self.logger.debug('πŸ“ Initialized download tracking for agent') # Event-based pause control (kept out of AgentState for serialization) self._external_pause_event = asyncio.Event() self._external_pause_event.set() def _enhance_task_with_schema(self, task: str, output_model_schema: type[AgentStructuredOutput] | None) -> str: """Enhance task description with output schema information if provided.""" if output_model_schema is None: return task try: schema = output_model_schema.model_json_schema() import json schema_json = json.dumps(schema, indent=2) enhancement = f'\nExpected output format: {output_model_schema.__name__}\n{schema_json}' return task + enhancement except Exception as e: self.logger.debug(f'Could not parse output schema: {e}') return task @property def logger(self) -> logging.Logger: """Get instance-specific logger with task ID in the name""" # logger may be called in __init__ so we don't assume self.* attributes have been initialized _task_id = task_id[-4:] if (task_id := getattr(self, 'task_id', None)) else '----' _browser_session_id = browser_session.id[-4:] if (browser_session := getattr(self, 'browser_session', None)) else '----' _current_target_id = ( browser_session.agent_focus_target_id[-2:] if (browser_session := getattr(self, 'browser_session', None)) and browser_session.agent_focus_target_id else '--' ) return logging.getLogger(f'browser_use.AgentπŸ…° {_task_id} β‡’ πŸ…‘ {_browser_session_id} πŸ…£ {_current_target_id}') @property def browser_profile(self) -> BrowserProfile: assert self.browser_session is not None, 'BrowserSession is not set up' return self.browser_session.browser_profile @property def is_using_fallback_llm(self) -> bool: """Check if the agent is currently using the fallback LLM.""" return self._using_fallback_llm @property def current_llm_model(self) -> str: """Get the model name of the currently active LLM.""" return self.llm.model if hasattr(self.llm, 'model') else 'unknown' async def _check_and_update_downloads(self, context: str = '') -> None: """Check for new downloads and update available file paths.""" if not self.has_downloads_path: return assert self.browser_session is not None, 'BrowserSession is not set up' try: current_downloads = self.browser_session.downloaded_files if current_downloads != self._last_known_downloads: self._update_available_file_paths(current_downloads) self._last_known_downloads = current_downloads if context: self.logger.debug(f'πŸ“ {context}: Updated available files') except Exception as e: error_context = f' {context}' if context else '' self.logger.debug(f'πŸ“ Failed to check for downloads{error_context}: {type(e).__name__}: {e}') def _update_available_file_paths(self, downloads: list[str]) -> None: """Update available_file_paths with downloaded files.""" if not self.has_downloads_path: return current_files = set(self.available_file_paths or []) new_files = set(downloads) - current_files if new_files: self.available_file_paths = list(current_files | new_files) self.logger.info( f'πŸ“ Added {len(new_files)} downloaded files to available_file_paths (total: {len(self.available_file_paths)} files)' ) for file_path in new_files: self.logger.info(f'πŸ“„ New file available: {file_path}') else: self.logger.debug(f'πŸ“ No new downloads detected (tracking {len(current_files)} files)') def _set_file_system(self, file_system_path: str | None = None) -> None: # Check for conflicting parameters if self.state.file_system_state and file_system_path: raise ValueError( 'Cannot provide both file_system_state (from agent state) and file_system_path. ' 'Either restore from existing state or create new file system at specified path, not both.' ) # Check if we should restore from existing state first if self.state.file_system_state: try: # Restore file system from state at the exact same location self.file_system = FileSystem.from_state(self.state.file_system_state) # The parent directory of base_dir is the original file_system_path self.file_system_path = str(self.file_system.base_dir) self.logger.debug(f'πŸ’Ύ File system restored from state to: {self.file_system_path}') return except Exception as e: self.logger.error(f'πŸ’Ύ Failed to restore file system from state: {e}') raise e # Initialize new file system try: if file_system_path: self.file_system = FileSystem(file_system_path) self.file_system_path = file_system_path else: # Use the agent directory for file system self.file_system = FileSystem(self.agent_directory) self.file_system_path = str(self.agent_directory) except Exception as e: self.logger.error(f'πŸ’Ύ Failed to initialize file system: {e}.') raise e # Save file system state to agent state self.state.file_system_state = self.file_system.get_state() self.logger.debug(f'πŸ’Ύ File system path: {self.file_system_path}') def _set_screenshot_service(self) -> None: """Initialize screenshot service using agent directory""" try: from browser_use.screenshots.service import ScreenshotService self.screenshot_service = ScreenshotService(self.agent_directory) self.logger.debug(f'πŸ“Έ Screenshot service initialized in: {self.agent_directory}/screenshots') except Exception as e: self.logger.error(f'πŸ“Έ Failed to initialize screenshot service: {e}.') raise e def save_file_system_state(self) -> None: """Save current file system state to agent state""" if self.file_system: self.state.file_system_state = self.file_system.get_state() else: self.logger.error('πŸ’Ύ File system is not set up. Cannot save state.') raise ValueError('File system is not set up. Cannot save state.') def _set_browser_use_version_and_source(self, source_override: str | None = None) -> None: """Get the version from pyproject.toml and determine the source of the browser-use package""" # Use the helper function for version detection version = get_browser_use_version() # Determine source try: package_root = Path(__file__).parent.parent.parent repo_files = ['.git', 'README.md', 'docs', 'examples'] if all(Path(package_root / file).exists() for file in repo_files): source = 'git' else: source = 'pip' except Exception as e: self.logger.debug(f'Error determining source: {e}') source = 'unknown' if source_override is not None: source = source_override # self.logger.debug(f'Version: {version}, Source: {source}') # moved later to _log_agent_run so that people are more likely to include it in copy-pasted support ticket logs self.version = version self.source = source def _setup_action_models(self) -> None: """Setup dynamic action models from tools registry""" # Initially only include actions with no filters self.ActionModel = self.tools.registry.create_action_model() # Create output model with the dynamic actions if self.settings.flash_mode: self.AgentOutput = AgentOutput.type_with_custom_actions_flash_mode(self.ActionModel) elif self.settings.use_thinking: self.AgentOutput = AgentOutput.type_with_custom_actions(self.ActionModel) else: self.AgentOutput = AgentOutput.type_with_custom_actions_no_thinking(self.ActionModel) # used to force the done action when max_steps is reached self.DoneActionModel = self.tools.registry.create_action_model(include_actions=['done']) if self.settings.flash_mode: self.DoneAgentOutput = AgentOutput.type_with_custom_actions_flash_mode(self.DoneActionModel) elif self.settings.use_thinking: self.DoneAgentOutput = AgentOutput.type_with_custom_actions(self.DoneActionModel) else: self.DoneAgentOutput = AgentOutput.type_with_custom_actions_no_thinking(self.DoneActionModel) def _get_skill_slug(self, skill: 'Skill', all_skills: list['Skill']) -> str: """Generate a clean slug from skill title for action names Converts title to lowercase, removes special characters, replaces spaces with underscores. Adds UUID suffix if there are duplicate slugs. Args: skill: The skill to get slug for all_skills: List of all skills to check for duplicates Returns: Slug like "cloned_github_stars_tracker" or "get_weather_data_a1b2" if duplicate Examples: "[Cloned] Github Stars Tracker" -> "cloned_github_stars_tracker" "Get Weather Data" -> "get_weather_data" """ import re # Remove special characters and convert to lowercase slug = re.sub(r'[^\w\s]', '', skill.title.lower()) # Replace whitespace and hyphens with underscores slug = re.sub(r'[\s\-]+', '_', slug) # Remove leading/trailing underscores slug = slug.strip('_') # Check for duplicates and add UUID suffix if needed same_slug_count = sum( 1 for s in all_skills if re.sub(r'[\s\-]+', '_', re.sub(r'[^\w\s]', '', s.title.lower()).strip('_')) == slug ) if same_slug_count > 1: return f'{slug}_{skill.id[:4]}' else: return slug async def _register_skills_as_actions(self) -> None: """Register each skill as a separate action using slug as action name""" if not self.skill_service or self._skills_registered: return self.logger.info('πŸ”§ Registering skill actions...') # Fetch all skills (auto-initializes if needed) skills = await self.skill_service.get_all_skills() if not skills: self.logger.warning('No skills loaded from SkillService') return # Register each skill as its own action for skill in skills: slug = self._get_skill_slug(skill, skills) param_model = skill.parameters_pydantic(exclude_cookies=True) # Create description with skill title in quotes description = f'{skill.description} (Skill: "{skill.title}")' # Create handler for this specific skill def make_skill_handler(skill_id: str): async def skill_handler(params: BaseModel) -> ActionResult: """Execute a specific skill""" assert self.skill_service is not None, 'SkillService not initialized' # Convert parameters to dict if isinstance(params, BaseModel): skill_params = params.model_dump() elif isinstance(params, dict): skill_params = params else: return ActionResult(extracted_content=None, error=f'Invalid parameters type: {type(params)}') # Get cookies from browser _cookies = await self.browser_session.cookies() try: result = await self.skill_service.execute_skill( skill_id=skill_id, parameters=skill_params, cookies=_cookies ) if result.success: return ActionResult( extracted_content=str(result.result) if result.result else None, error=None, ) else: return ActionResult(extracted_content=None, error=result.error or 'Skill execution failed') except Exception as e: # Check if it's a MissingCookieException if type(e).__name__ == 'MissingCookieException': # Format: "Missing cookies (name): description" cookie_name = getattr(e, 'cookie_name', 'unknown') cookie_description = getattr(e, 'cookie_description', str(e)) error_msg = f'Missing cookies ({cookie_name}): {cookie_description}' return ActionResult(extracted_content=None, error=error_msg) return ActionResult(extracted_content=None, error=f'Skill execution error: {type(e).__name__}: {e}') return skill_handler # Create the handler for this skill handler = make_skill_handler(skill.id) handler.__name__ = slug # Register the action with the slug as the action name self.tools.registry.action(description=description, param_model=param_model)(handler) # Mark as registered self._skills_registered = True # Rebuild action models to include the new skill actions self._setup_action_models() # Reconvert initial actions with the new ActionModel type if they exist if self.initial_actions: # Convert back to dict form first initial_actions_dict = [] for action in self.initial_actions: action_dump = action.model_dump(exclude_unset=True) initial_actions_dict.append(action_dump) # Reconvert using new ActionModel self.initial_actions = self._convert_initial_actions(initial_actions_dict) self.logger.info(f'βœ“ Registered {len(skills)} skill actions') async def _get_unavailable_skills_info(self) -> str: """Get information about skills that are unavailable due to missing cookies Returns: Formatted string describing unavailable skills and how to make them available """ if not self.skill_service: return '' try: # Get all skills skills = await self.skill_service.get_all_skills() if not skills: return '' # Get current cookies current_cookies = await self.browser_session.cookies() cookie_dict = {cookie['name']: cookie['value'] for cookie in current_cookies} # Check each skill for missing required cookies unavailable_skills: list[dict[str, Any]] = [] for skill in skills: # Get cookie parameters for this skill cookie_params = [p for p in skill.parameters if p.type == 'cookie'] if not cookie_params: # No cookies needed, skip continue # Check for missing required cookies missing_cookies: list[dict[str, str]] = [] for cookie_param in cookie_params: is_required = cookie_param.required if cookie_param.required is not None else True if is_required and cookie_param.name not in cookie_dict: missing_cookies.append( {'name': cookie_param.name, 'description': cookie_param.description or 'No description provided'} ) if missing_cookies: unavailable_skills.append( { 'id': skill.id, 'title': skill.title, 'description': skill.description, 'missing_cookies': missing_cookies, } ) if not unavailable_skills: return '' # Format the unavailable skills info with slugs lines = ['Unavailable Skills (missing required cookies):'] for skill_info in unavailable_skills: # Get the full skill object to use the slug helper skill_obj = next((s for s in skills if s.id == skill_info['id']), None) slug = self._get_skill_slug(skill_obj, skills) if skill_obj else skill_info['title'] title = skill_info['title'] lines.append(f'\n β€’ {slug} ("{title}")') lines.append(f' Description: {skill_info["description"]}') lines.append(' Missing cookies:') for cookie in skill_info['missing_cookies']: lines.append(f' - {cookie["name"]}: {cookie["description"]}') return '\n'.join(lines) except Exception as e: self.logger.error(f'Error getting unavailable skills info: {type(e).__name__}: {e}') return '' def add_new_task(self, new_task: str) -> None: """Add a new task to the agent, keeping the same task_id as tasks are continuous""" # Simply delegate to message manager - no need for new task_id or events # The task continues with new instructions, it doesn't end and start a new one self.task = new_task self._message_manager.add_new_task(new_task) # Mark as follow-up task and recreate eventbus (gets shut down after each run) self.state.follow_up_task = True # Reset control flags so agent can continue self.state.stopped = False self.state.paused = False agent_id_suffix = str(self.id)[-4:].replace('-', '_') if agent_id_suffix and agent_id_suffix[0].isdigit(): agent_id_suffix = 'a' + agent_id_suffix self.eventbus = EventBus(name=f'Agent_{agent_id_suffix}') async def _check_stop_or_pause(self) -> None: """Check if the agent should stop or pause, and handle accordingly.""" # Check new should_stop_callback - sets stopped state cleanly without raising if self.register_should_stop_callback: if await self.register_should_stop_callback(): self.logger.info('External callback requested stop') self.state.stopped = True raise InterruptedError if self.register_external_agent_status_raise_error_callback: if await self.register_external_agent_status_raise_error_callback(): raise InterruptedError if self.state.stopped: raise InterruptedError if self.state.paused: raise InterruptedError @observe(name='agent.step', ignore_output=True, ignore_input=True) @time_execution_async('--step') async def step(self, step_info: AgentStepInfo | None = None) -> None: """Execute one step of the task""" # Initialize timing first, before any exceptions can occur self.step_start_time = time.time() browser_state_summary = None try: if self.browser_session: try: captcha_wait = await self.browser_session.wait_if_captcha_solving() if captcha_wait and captcha_wait.waited: # Reset step timing to exclude the captcha wait from step duration metrics self.step_start_time = time.time() duration_s = captcha_wait.duration_ms / 1000 outcome = captcha_wait.result # 'success' | 'failed' | 'timeout' msg = f'Waited {duration_s:.1f}s for {captcha_wait.vendor} CAPTCHA to be solved. Result: {outcome}.' self.logger.info(f'πŸ”’ {msg}') # Inject the outcome so the LLM sees what happened captcha_result = ActionResult(long_term_memory=msg) if self.state.last_result: self.state.last_result.append(captcha_result) else: self.state.last_result = [captcha_result] except Exception as e: self.logger.warning(f'Phase 0 captcha wait failed (non-fatal): {e}') # Phase 1: Prepare context and timing browser_state_summary = await self._prepare_context(step_info) # Clear previous step state after context preparation (which needs # them for the "previous action result" prompt) but before the LLM # call, so a timeout during _get_next_action or _execute_actions # won't leave stale data from the previous step. self.state.last_model_output = None self.state.last_result = None # Phase 2: Get model output and execute actions await self._get_next_action(browser_state_summary) await self._execute_actions() # Phase 3: Post-processing await self._post_process() except Exception as e: # Handle ALL exceptions in one place await self._handle_step_error(e) finally: await self._finalize(browser_state_summary) async def _prepare_context(self, step_info: AgentStepInfo | None = None) -> BrowserStateSummary: """Prepare the context for the step: browser state, action models, page actions""" # step_start_time is now set in step() method assert self.browser_session is not None, 'BrowserSession is not set up' self.logger.debug(f'🌐 Step {self.state.n_steps}: Getting browser state...') # Always take screenshots for all steps self.logger.debug('πŸ“Έ Requesting browser state with include_screenshot=True') browser_state_summary = await self.browser_session.get_browser_state_summary( include_screenshot=True, # always capture even if use_vision=False so that cloud sync is useful (it's fast now anyway) include_recent_events=self.include_recent_events, ) if browser_state_summary.screenshot: self.logger.debug(f'πŸ“Έ Got browser state WITH screenshot, length: {len(browser_state_summary.screenshot)}') else: self.logger.debug('πŸ“Έ Got browser state WITHOUT screenshot') # Check for new downloads after getting browser state (catches PDF auto-downloads and previous step downloads) await self._check_and_update_downloads(f'Step {self.state.n_steps}: after getting browser state') self._log_step_context(browser_state_summary) await self._check_stop_or_pause() # Update action models with page-specific actions self.logger.debug(f'πŸ“ Step {self.state.n_steps}: Updating action models...') await self._update_action_models_for_page(browser_state_summary.url) # Get page-specific filtered actions page_filtered_actions = self.tools.registry.get_prompt_description(browser_state_summary.url) # Page-specific actions will be included directly in the browser_state message self.logger.debug(f'πŸ’¬ Step {self.state.n_steps}: Creating state messages for context...') # Get unavailable skills info if skills service is enabled unavailable_skills_info = None if self.skill_service is not None: unavailable_skills_info = await self._get_unavailable_skills_info() # Render plan description for injection into agent context plan_description = self._render_plan_description() self._message_manager.prepare_step_state( browser_state_summary=browser_state_summary, model_output=self.state.last_model_output, result=self.state.last_result, step_info=step_info, sensitive_data=self.sensitive_data, ) await self._maybe_compact_messages(step_info) self._message_manager.create_state_messages( browser_state_summary=browser_state_summary, model_output=self.state.last_model_output, result=self.state.last_result, step_info=step_info, use_vision=self.settings.use_vision, page_filtered_actions=page_filtered_actions if page_filtered_actions else None, sensitive_data=self.sensitive_data, available_file_paths=self.available_file_paths, # Always pass current available_file_paths unavailable_skills_info=unavailable_skills_info, plan_description=plan_description, skip_state_update=True, ) await self._inject_budget_warning(step_info) self._inject_replan_nudge() self._inject_exploration_nudge() self._update_loop_detector_page_state(browser_state_summary) self._inject_loop_detection_nudge() await self._force_done_after_last_step(step_info) await self._force_done_after_failure() return browser_state_summary async def _maybe_compact_messages(self, step_info: AgentStepInfo | None = None) -> None: """Optionally compact message history to keep prompts small.""" settings = self.settings.message_compaction if not settings or not settings.enabled: return compaction_llm = settings.compaction_llm or self.settings.page_extraction_llm or self.llm await self._message_manager.maybe_compact_messages( llm=compaction_llm, settings=settings, step_info=step_info, ) @observe_debug(ignore_input=True, name='get_next_action') async def _get_next_action(self, browser_state_summary: BrowserStateSummary) -> None: """Execute LLM interaction with retry logic and handle callbacks""" input_messages = self._message_manager.get_messages() self.logger.debug( f'πŸ€– Step {self.state.n_steps}: Calling LLM with {len(input_messages)} messages (model: {self.llm.model})...' ) try: model_output = await asyncio.wait_for( self._get_model_output_with_retry(input_messages), timeout=self.settings.llm_timeout ) except TimeoutError: @observe(name='_llm_call_timed_out_with_input') async def _log_model_input_to_lmnr(input_messages: list[BaseMessage]) -> None: """Log the model input""" pass await _log_model_input_to_lmnr(input_messages) raise TimeoutError( f'LLM call timed out after {self.settings.llm_timeout} seconds. Keep your thinking and output short.' ) self.state.last_model_output = model_output # Check again for paused/stopped state after getting model output await self._check_stop_or_pause() # Handle callbacks and conversation saving await self._handle_post_llm_processing(browser_state_summary, input_messages) # check again if Ctrl+C was pressed before we commit the output to history await self._check_stop_or_pause() async def _execute_actions(self) -> None: """Execute the actions from model output""" if self.state.last_model_output is None: raise ValueError('No model output to execute actions from') result = await self.multi_act(self.state.last_model_output.action) self.state.last_result = result async def _post_process(self) -> None: """Handle post-action processing like download tracking and result logging""" assert self.browser_session is not None, 'BrowserSession is not set up' # Check for new downloads after executing actions await self._check_and_update_downloads('after executing actions') # Update plan state from model output if self.state.last_model_output is not None: self._update_plan_from_model_output(self.state.last_model_output) # Record executed actions for loop detection self._update_loop_detector_actions() # check for action errors - only count single-action steps toward consecutive failures; # multi-action steps with errors are handled by loop detection and replan nudges instead if self.state.last_result and len(self.state.last_result) == 1 and self.state.last_result[-1].error: self.state.consecutive_failures += 1 self.logger.debug(f'πŸ”„ Step {self.state.n_steps}: Consecutive failures: {self.state.consecutive_failures}') return if self.state.consecutive_failures > 0: self.state.consecutive_failures = 0 self.logger.debug(f'πŸ”„ Step {self.state.n_steps}: Consecutive failures reset to: {self.state.consecutive_failures}') # Log completion results if self.state.last_result and len(self.state.last_result) > 0 and self.state.last_result[-1].is_done: success = self.state.last_result[-1].success if success: # Green color for success self.logger.info(f'\nπŸ“„ \033[32m Final Result:\033[0m \n{self.state.last_result[-1].extracted_content}\n\n') else: # Red color for failure self.logger.info(f'\nπŸ“„ \033[31m Final Result:\033[0m \n{self.state.last_result[-1].extracted_content}\n\n') if self.state.last_result[-1].attachments: total_attachments = len(self.state.last_result[-1].attachments) for i, file_path in enumerate(self.state.last_result[-1].attachments): self.logger.info(f'πŸ‘‰ Attachment {i + 1 if total_attachments > 1 else ""}: {file_path}') async def _handle_step_error(self, error: Exception) -> None: """Handle all types of errors that can occur during a step""" # Handle InterruptedError specially if isinstance(error, InterruptedError): error_msg = 'The agent was interrupted mid-step' + (f' - {str(error)}' if str(error) else '') # NOTE: This is not an error, it's a normal part of the execution when the user interrupts the agent self.logger.warning(f'{error_msg}') return # Handle browser closed/disconnected errors if self._is_connection_like_error(error): # If reconnection is in progress, wait for it instead of stopping if self.browser_session.is_reconnecting: wait_timeout = self.browser_session.RECONNECT_WAIT_TIMEOUT self.logger.warning( f'πŸ”„ Connection error during reconnection, waiting up to {wait_timeout}s for reconnect: {error}' ) try: await asyncio.wait_for(self.browser_session._reconnect_event.wait(), timeout=wait_timeout) except TimeoutError: pass # Check if reconnection succeeded if self.browser_session.is_cdp_connected: self.logger.info('πŸ”„ Reconnection succeeded, retrying step...') self.state.last_result = [ActionResult(error=f'Connection lost and recovered: {error}')] return # Not reconnecting or reconnection failed β€” check if truly terminal if self._is_browser_closed_error(error): self.logger.warning(f'πŸ›‘ Browser closed or disconnected: {error}') self.state.stopped = True self._external_pause_event.set() return # Handle all other exceptions include_trace = self.logger.isEnabledFor(logging.DEBUG) error_msg = AgentError.format_error(error, include_trace=include_trace) max_total_failures = self.settings.max_failures + int(self.settings.final_response_after_failure) prefix = f'❌ Result failed {self.state.consecutive_failures + 1}/{max_total_failures} times: ' self.state.consecutive_failures += 1 # Use WARNING for partial failures, ERROR only when max failures reached is_final_failure = self.state.consecutive_failures >= max_total_failures log_level = logging.ERROR if is_final_failure else logging.WARNING if 'Could not parse response' in error_msg or 'tool_use_failed' in error_msg: # give model a hint how output should look like self.logger.log(log_level, f'Model: {self.llm.model} failed') self.logger.log(log_level, f'{prefix}{error_msg}') else: self.logger.log(log_level, f'{prefix}{error_msg}') await self._demo_mode_log(f'Step error: {error_msg}', 'error', {'step': self.state.n_steps}) self.state.last_result = [ActionResult(error=error_msg)] return None def _is_connection_like_error(self, error: Exception) -> bool: """Check if the error looks like a CDP/WebSocket connection failure. Unlike _is_browser_closed_error(), this does NOT check if the CDP client is None or if reconnection is in progress β€” it purely looks at the error signature. """ error_str = str(error).lower() return ( isinstance(error, ConnectionError) or 'websocket connection closed' in error_str or 'connection closed' in error_str or 'browser has been closed' in error_str or 'browser closed' in error_str or 'no browser' in error_str ) def _is_browser_closed_error(self, error: Exception) -> bool: """Check if the browser has been closed or disconnected. Only returns True when the error itself is a CDP/WebSocket connection failure AND the CDP client is gone AND we're not actively reconnecting. Avoids false positives on unrelated errors (element not found, timeouts, parse errors) that happen to coincide with a transient None state during reconnects or resets. """ # During reconnection, don't treat connection errors as terminal if self.browser_session.is_reconnecting: return False error_str = str(error).lower() is_connection_error = ( isinstance(error, ConnectionError) or 'websocket connection closed' in error_str or 'connection closed' in error_str or 'browser has been closed' in error_str or 'browser closed' in error_str or 'no browser' in error_str ) return is_connection_error and self.browser_session._cdp_client_root is None async def _finalize(self, browser_state_summary: BrowserStateSummary | None) -> None: """Finalize the step with history, logging, and events""" step_end_time = time.time() if not self.state.last_result: return if browser_state_summary: step_interval = None if len(self.history.history) > 0: last_history_item = self.history.history[-1] if last_history_item.metadata: previous_end_time = last_history_item.metadata.step_end_time previous_start_time = last_history_item.metadata.step_start_time step_interval = max(0, previous_end_time - previous_start_time) metadata = StepMetadata( step_number=self.state.n_steps, step_start_time=self.step_start_time, step_end_time=step_end_time, step_interval=step_interval, ) # Use _make_history_item like main branch await self._make_history_item( self.state.last_model_output, browser_state_summary, self.state.last_result, metadata, state_message=self._message_manager.last_state_message_text, ) # Log step completion summary summary_message = self._log_step_completion_summary(self.step_start_time, self.state.last_result) if summary_message: await self._demo_mode_log(summary_message, 'info', {'step': self.state.n_steps}) # Save file system state after step completion self.save_file_system_state() # Emit both step created and executed events if browser_state_summary and self.state.last_model_output: # Extract key step data for the event actions_data = [] if self.state.last_model_output.action: for action in self.state.last_model_output.action: action_dict = action.model_dump() if hasattr(action, 'model_dump') else {} actions_data.append(action_dict) # Emit CreateAgentStepEvent step_event = CreateAgentStepEvent.from_agent_step( self, self.state.last_model_output, self.state.last_result, actions_data, browser_state_summary, ) self.eventbus.dispatch(step_event) # Increment step counter after step is fully completed self.state.n_steps += 1 def _update_plan_from_model_output(self, model_output: AgentOutput) -> None: """Update the plan state from model output fields (current_plan_item, plan_update).""" if not self.settings.enable_planning: return # If model provided a new plan via plan_update, replace the current plan if model_output.plan_update is not None: self.state.plan = [PlanItem(text=step_text) for step_text in model_output.plan_update] self.state.current_plan_item_index = 0 self.state.plan_generation_step = self.state.n_steps if self.state.plan: self.state.plan[0].status = 'current' self.logger.info( f'πŸ“‹ Plan {"updated" if self.state.plan_generation_step else "created"} with {len(self.state.plan)} steps' ) return # If model provided a step index update, advance the plan if model_output.current_plan_item is not None and self.state.plan is not None: new_idx = model_output.current_plan_item # Clamp to valid range new_idx = max(0, min(new_idx, len(self.state.plan) - 1)) old_idx = self.state.current_plan_item_index # Mark steps between old and new as done for i in range(old_idx, new_idx): if i < len(self.state.plan) and self.state.plan[i].status in ('current', 'pending'): self.state.plan[i].status = 'done' # Mark the new step as current if new_idx < len(self.state.plan): self.state.plan[new_idx].status = 'current' self.state.current_plan_item_index = new_idx def _render_plan_description(self) -> str | None: """Render the current plan as a text description for injection into agent context.""" if not self.settings.enable_planning or self.state.plan is None: return None markers = {'done': '[x]', 'current': '[>]', 'pending': '[ ]', 'skipped': '[-]'} lines = [] for i, step in enumerate(self.state.plan): marker = markers.get(step.status, '[ ]') lines.append(f'{marker} {i}: {step.text}') return '\n'.join(lines) def _inject_replan_nudge(self) -> None: """Inject a replan nudge when stall detection threshold is met.""" if not self.settings.enable_planning or self.state.plan is None: return if self.settings.planning_replan_on_stall <= 0: return if self.state.consecutive_failures >= self.settings.planning_replan_on_stall: msg = ( 'REPLAN SUGGESTED: You have failed ' f'{self.state.consecutive_failures} consecutive times. ' 'Your current plan may need revision. ' 'Output a new `plan_update` with revised steps to recover.' ) self.logger.info(f'πŸ“‹ Replan nudge injected after {self.state.consecutive_failures} consecutive failures') self._message_manager._add_context_message(UserMessage(content=msg)) def _inject_exploration_nudge(self) -> None: """Nudge the agent to create a plan (or call done) after exploring without one.""" if not self.settings.enable_planning or self.state.plan is not None: return if self.settings.planning_exploration_limit <= 0: return if self.state.n_steps >= self.settings.planning_exploration_limit: msg = ( 'PLANNING NUDGE: You have taken ' f'{self.state.n_steps} steps without creating a plan. ' 'If the task is complex, output a `plan_update` with clear todo items now. ' 'If the task is already done or nearly done, call `done` instead.' ) self.logger.info(f'πŸ“‹ Exploration nudge injected after {self.state.n_steps} steps without a plan') self._message_manager._add_context_message(UserMessage(content=msg)) def _inject_loop_detection_nudge(self) -> None: """Inject an escalating nudge when behavioral loops are detected.""" if not self.settings.loop_detection_enabled: return nudge = self.state.loop_detector.get_nudge_message() if nudge: self.logger.info( f'πŸ” Loop detection nudge injected (repetition={self.state.loop_detector.max_repetition_count}, ' f'stagnation={self.state.loop_detector.consecutive_stagnant_pages})' ) self._message_manager._add_context_message(UserMessage(content=nudge)) def _update_loop_detector_actions(self) -> None: """Record the actions from the latest step into the loop detector.""" if not self.settings.loop_detection_enabled: return if self.state.last_model_output is None: return # Actions to exclude: wait always hashes identically (instant false positive), # done is terminal, go_back is navigation recovery _LOOP_EXEMPT_ACTIONS = {'wait', 'done', 'go_back'} for action in self.state.last_model_output.action: action_data = action.model_dump(exclude_unset=True) action_name = next(iter(action_data.keys()), 'unknown') if action_name in _LOOP_EXEMPT_ACTIONS: continue params = action_data.get(action_name, {}) if not isinstance(params, dict): params = {} self.state.loop_detector.record_action(action_name, params) def _update_loop_detector_page_state(self, browser_state_summary: BrowserStateSummary) -> None: """Record the current page state for stagnation detection.""" if not self.settings.loop_detection_enabled: return url = browser_state_summary.url or '' element_count = len(browser_state_summary.dom_state.selector_map) if browser_state_summary.dom_state else 0 # Use the DOM text representation for fingerprinting dom_text = '' if browser_state_summary.dom_state: try: dom_text = browser_state_summary.dom_state.llm_representation() except Exception: dom_text = '' self.state.loop_detector.record_page_state(url, dom_text, element_count) async def _inject_budget_warning(self, step_info: AgentStepInfo | None = None) -> None: """Inject a prominent budget warning when the agent has used >= 75% of its step budget. This gives the LLM advance notice to wrap up, save partial results, and call done rather than exhausting all steps with nothing saved. """ if step_info is None: return steps_used = step_info.step_number + 1 # Convert 0-indexed to 1-indexed budget_ratio = steps_used / step_info.max_steps if budget_ratio >= 0.75 and not step_info.is_last_step(): steps_remaining = step_info.max_steps - steps_used pct = int(budget_ratio * 100) msg = ( f'BUDGET WARNING: You have used {steps_used}/{step_info.max_steps} steps ' f'({pct}%). {steps_remaining} steps remaining. ' f'If the task cannot be completed in the remaining steps, prioritize: ' f'(1) consolidate your results (save to files if the file system is in use), ' f'(2) call done with what you have. ' f'Partial results are far more valuable than exhausting all steps with nothing saved.' ) self.logger.info(f'Step budget warning: {steps_used}/{step_info.max_steps} ({pct}%)') self._message_manager._add_context_message(UserMessage(content=msg)) async def _force_done_after_last_step(self, step_info: AgentStepInfo | None = None) -> None: """Handle special processing for the last step""" if step_info and step_info.is_last_step(): # Add last step warning if needed msg = 'You reached max_steps - this is your last step. Your only tool available is the "done" tool. No other tool is available. All other tools which you see in history or examples are not available.' msg += '\nIf the task is not yet fully finished as requested by the user, set success in "done" to false! E.g. if not all steps are fully completed. Else success to true.' msg += '\nInclude everything you found out for the ultimate task in the done text.' self.logger.debug('Last step finishing up') self._message_manager._add_context_message(UserMessage(content=msg)) self.AgentOutput = self.DoneAgentOutput async def _force_done_after_failure(self) -> None: """Force done after failure""" # Create recovery message if self.state.consecutive_failures >= self.settings.max_failures and self.settings.final_response_after_failure: msg = f'You failed {self.settings.max_failures} times. Therefore we terminate the agent.' msg += '\nYour only tool available is the "done" tool. No other tool is available. All other tools which you see in history or examples are not available.' msg += '\nIf the task is not yet fully finished as requested by the user, set success in "done" to false! E.g. if not all steps are fully completed. Else success to true.' msg += '\nInclude everything you found out for the ultimate task in the done text.' self.logger.debug('Force done action, because we reached max_failures.') self._message_manager._add_context_message(UserMessage(content=msg)) self.AgentOutput = self.DoneAgentOutput @observe(ignore_input=True, ignore_output=False) async def _judge_trace(self) -> JudgementResult | None: """Judge the trace of the agent""" task = self.task final_result = self.history.final_result() or '' agent_steps = self.history.agent_steps() screenshot_paths = [p for p in self.history.screenshot_paths() if p is not None] # Construct input messages for judge evaluation input_messages = construct_judge_messages( task=task, final_result=final_result, agent_steps=agent_steps, screenshot_paths=screenshot_paths, max_images=10, ground_truth=self.settings.ground_truth, use_vision=self.settings.use_vision, ) # Call LLM with JudgementResult as output format kwargs: dict = {'output_format': JudgementResult} # Only pass request_type for ChatBrowserUse (other providers don't support it) if self.judge_llm.provider == 'browser-use': kwargs['request_type'] = 'judge' kwargs['session_id'] = self.session_id try: response = await self.judge_llm.ainvoke(input_messages, **kwargs) judgement: JudgementResult = response.completion # type: ignore[assignment] return judgement except Exception as e: self.logger.error(f'Judge trace failed: {e}') # Return a default judgement on failure return None async def _judge_and_log(self) -> None: """Run judge evaluation and log the verdict. The judge verdict is attached to the action result but does NOT override last_result.success β€” that stays as the agent's self-report. Telemetry sends both values so the eval platform can compare agent vs judge. """ judgement = await self._judge_trace() # Attach judgement to last action result if self.history.history[-1].result[-1].is_done: last_result = self.history.history[-1].result[-1] last_result.judgement = judgement # Get self-reported success self_reported_success = last_result.success # Log the verdict based on self-reported success and judge verdict if judgement: # If both self-reported and judge agree on success, don't log if self_reported_success is True and judgement.verdict is True: return judge_log = '\n' # If agent reported success but judge thinks it failed, show warning if self_reported_success is True and judgement.verdict is False: judge_log += '⚠️ \033[33mAgent reported success but judge thinks task failed\033[0m\n' # Otherwise, show full judge result verdict_color = '\033[32m' if judgement.verdict else '\033[31m' verdict_text = 'βœ… PASS' if judgement.verdict else '❌ FAIL' judge_log += f'βš–οΈ {verdict_color}Judge Verdict: {verdict_text}\033[0m\n' if judgement.failure_reason: judge_log += f' Failure Reason: {judgement.failure_reason}\n' if judgement.reached_captcha: self.logger.warning( 'Agent was blocked by a captcha. Cloud browsers include stealth fingerprinting and proxy rotation to avoid this.\n' ' Try: Browser(use_cloud=True) | Get an API key: https://cloud.browser-use.com?utm_source=oss&utm_medium=captcha_nudge' ) judge_log += f' {judgement.reasoning}\n' self.logger.info(judge_log) async def _get_model_output_with_retry(self, input_messages: list[BaseMessage]) -> AgentOutput: """Get model output with retry logic for empty actions""" model_output = await self.get_model_output(input_messages) self.logger.debug( f'βœ… Step {self.state.n_steps}: Got LLM response with {len(model_output.action) if model_output.action else 0} actions' ) if ( not model_output.action or not isinstance(model_output.action, list) or all(action.model_dump() == {} for action in model_output.action) ): self.logger.warning('Model returned empty action. Retrying...') clarification_message = UserMessage( content='You forgot to return an action. Please respond with a valid JSON action according to the expected schema with your assessment and next actions.' ) retry_messages = input_messages + [clarification_message] model_output = await self.get_model_output(retry_messages) if not model_output.action or all(action.model_dump() == {} for action in model_output.action): self.logger.warning('Model still returned empty after retry. Inserting safe noop action.') action_instance = self.ActionModel() setattr( action_instance, 'done', { 'success': False, 'text': 'No next action returned by LLM!', }, ) model_output.action = [action_instance] return model_output async def _handle_post_llm_processing( self, browser_state_summary: BrowserStateSummary, input_messages: list[BaseMessage], ) -> None: """Handle callbacks and conversation saving after LLM interaction""" if self.register_new_step_callback and self.state.last_model_output: if inspect.iscoroutinefunction(self.register_new_step_callback): await self.register_new_step_callback( browser_state_summary, self.state.last_model_output, self.state.n_steps, ) else: self.register_new_step_callback( browser_state_summary, self.state.last_model_output, self.state.n_steps, ) if self.settings.save_conversation_path and self.state.last_model_output: # Treat save_conversation_path as a directory (consistent with other recording paths) conversation_dir = Path(self.settings.save_conversation_path) conversation_filename = f'conversation_{self.id}_{self.state.n_steps}.txt' target = conversation_dir / conversation_filename await save_conversation( input_messages, self.state.last_model_output, target, self.settings.save_conversation_path_encoding, ) async def _make_history_item( self, model_output: AgentOutput | None, browser_state_summary: BrowserStateSummary, result: list[ActionResult], metadata: StepMetadata | None = None, state_message: str | None = None, ) -> None: """Create and store history item""" if model_output: interacted_elements = AgentHistory.get_interacted_element(model_output, browser_state_summary.dom_state.selector_map) else: interacted_elements = [None] # Store screenshot and get path screenshot_path = None if browser_state_summary.screenshot: self.logger.debug( f'πŸ“Έ Storing screenshot for step {self.state.n_steps}, screenshot length: {len(browser_state_summary.screenshot)}' ) screenshot_path = await self.screenshot_service.store_screenshot(browser_state_summary.screenshot, self.state.n_steps) self.logger.debug(f'πŸ“Έ Screenshot stored at: {screenshot_path}') else: self.logger.debug(f'πŸ“Έ No screenshot in browser_state_summary for step {self.state.n_steps}') state_history = BrowserStateHistory( url=browser_state_summary.url, title=browser_state_summary.title, tabs=browser_state_summary.tabs, interacted_element=interacted_elements, screenshot_path=screenshot_path, ) history_item = AgentHistory( model_output=model_output, result=result, state=state_history, metadata=metadata, state_message=state_message, ) self.history.add_item(history_item) def _remove_think_tags(self, text: str) -> str: THINK_TAGS = re.compile(r'.*?', re.DOTALL) STRAY_CLOSE_TAG = re.compile(r'.*?', re.DOTALL) # Step 1: Remove well-formed ... text = re.sub(THINK_TAGS, '', text) # Step 2: If there's an unmatched closing tag , # remove everything up to and including that. text = re.sub(STRAY_CLOSE_TAG, '', text) return text.strip() # region - URL replacement def _replace_urls_in_text(self, text: str) -> tuple[str, dict[str, str]]: """Replace URLs in a text string""" replaced_urls: dict[str, str] = {} def replace_url(match: re.Match) -> str: """Url can only have 1 query and 1 fragment""" import hashlib original_url = match.group(0) # Find where the query/fragment starts query_start = original_url.find('?') fragment_start = original_url.find('#') # Find the earliest position of query or fragment after_path_start = len(original_url) # Default: no query/fragment if query_start != -1: after_path_start = min(after_path_start, query_start) if fragment_start != -1: after_path_start = min(after_path_start, fragment_start) # Split URL into base (up to path) and after_path (query + fragment) base_url = original_url[:after_path_start] after_path = original_url[after_path_start:] # If after_path is within the limit, don't shorten if len(after_path) <= self._url_shortening_limit: return original_url # If after_path is too long, truncate and add hash if after_path: truncated_after_path = after_path[: self._url_shortening_limit] # Create a short hash of the full after_path content hash_obj = hashlib.md5(after_path.encode('utf-8')) short_hash = hash_obj.hexdigest()[:7] # Create shortened URL shortened = f'{base_url}{truncated_after_path}...{short_hash}' # Only use shortened URL if it's actually shorter than the original if len(shortened) < len(original_url): replaced_urls[shortened] = original_url return shortened return original_url return URL_PATTERN.sub(replace_url, text), replaced_urls def _process_messsages_and_replace_long_urls_shorter_ones(self, input_messages: list[BaseMessage]) -> dict[str, str]: """Replace long URLs with shorter ones ? @dev edits input_messages in place returns: tuple[filtered_input_messages, urls we replaced {shorter_url: original_url}] """ from browser_use.llm.messages import AssistantMessage, UserMessage urls_replaced: dict[str, str] = {} # Process each message, in place for message in input_messages: # no need to process SystemMessage, we have control over that anyway if isinstance(message, (UserMessage, AssistantMessage)): if isinstance(message.content, str): # Simple string content message.content, replaced_urls = self._replace_urls_in_text(message.content) urls_replaced.update(replaced_urls) elif isinstance(message.content, list): # List of content parts for part in message.content: if isinstance(part, ContentPartTextParam): part.text, replaced_urls = self._replace_urls_in_text(part.text) urls_replaced.update(replaced_urls) return urls_replaced @staticmethod def _recursive_process_all_strings_inside_pydantic_model(model: BaseModel, url_replacements: dict[str, str]) -> None: """Recursively process all strings inside a Pydantic model, replacing shortened URLs with originals in place.""" for field_name, field_value in model.__dict__.items(): if isinstance(field_value, str): # Replace shortened URLs with original URLs in string processed_string = Agent._replace_shortened_urls_in_string(field_value, url_replacements) setattr(model, field_name, processed_string) elif isinstance(field_value, BaseModel): # Recursively process nested Pydantic models Agent._recursive_process_all_strings_inside_pydantic_model(field_value, url_replacements) elif isinstance(field_value, dict): # Process dictionary values in place Agent._recursive_process_dict(field_value, url_replacements) elif isinstance(field_value, (list, tuple)): processed_value = Agent._recursive_process_list_or_tuple(field_value, url_replacements) setattr(model, field_name, processed_value) @staticmethod def _recursive_process_dict(dictionary: dict, url_replacements: dict[str, str]) -> None: """Helper method to process dictionaries.""" for k, v in dictionary.items(): if isinstance(v, str): dictionary[k] = Agent._replace_shortened_urls_in_string(v, url_replacements) elif isinstance(v, BaseModel): Agent._recursive_process_all_strings_inside_pydantic_model(v, url_replacements) elif isinstance(v, dict): Agent._recursive_process_dict(v, url_replacements) elif isinstance(v, (list, tuple)): dictionary[k] = Agent._recursive_process_list_or_tuple(v, url_replacements) @staticmethod def _recursive_process_list_or_tuple(container: list | tuple, url_replacements: dict[str, str]) -> list | tuple: """Helper method to process lists and tuples.""" if isinstance(container, tuple): # For tuples, create a new tuple with processed items processed_items = [] for item in container: if isinstance(item, str): processed_items.append(Agent._replace_shortened_urls_in_string(item, url_replacements)) elif isinstance(item, BaseModel): Agent._recursive_process_all_strings_inside_pydantic_model(item, url_replacements) processed_items.append(item) elif isinstance(item, dict): Agent._recursive_process_dict(item, url_replacements) processed_items.append(item) elif isinstance(item, (list, tuple)): processed_items.append(Agent._recursive_process_list_or_tuple(item, url_replacements)) else: processed_items.append(item) return tuple(processed_items) else: # For lists, modify in place for i, item in enumerate(container): if isinstance(item, str): container[i] = Agent._replace_shortened_urls_in_string(item, url_replacements) elif isinstance(item, BaseModel): Agent._recursive_process_all_strings_inside_pydantic_model(item, url_replacements) elif isinstance(item, dict): Agent._recursive_process_dict(item, url_replacements) elif isinstance(item, (list, tuple)): container[i] = Agent._recursive_process_list_or_tuple(item, url_replacements) return container @staticmethod def _replace_shortened_urls_in_string(text: str, url_replacements: dict[str, str]) -> str: """Replace all shortened URLs in a string with their original URLs.""" result = text for shortened_url, original_url in url_replacements.items(): result = result.replace(shortened_url, original_url) return result # endregion - URL replacement @time_execution_async('--get_next_action') @observe_debug(ignore_input=True, ignore_output=True, name='get_model_output') async def get_model_output(self, input_messages: list[BaseMessage]) -> AgentOutput: """Get next action from LLM based on current state""" urls_replaced = self._process_messsages_and_replace_long_urls_shorter_ones(input_messages) # Build kwargs for ainvoke # Note: ChatBrowserUse will automatically generate action descriptions from output_format schema kwargs: dict = {'output_format': self.AgentOutput, 'session_id': self.session_id} try: response = await self.llm.ainvoke(input_messages, **kwargs) parsed: AgentOutput = response.completion # type: ignore[assignment] # Replace any shortened URLs in the LLM response back to original URLs if urls_replaced: self._recursive_process_all_strings_inside_pydantic_model(parsed, urls_replaced) # cut the number of actions to max_actions_per_step if needed if len(parsed.action) > self.settings.max_actions_per_step: parsed.action = parsed.action[: self.settings.max_actions_per_step] if not (hasattr(self.state, 'paused') and (self.state.paused or self.state.stopped)): log_response(parsed, self.tools.registry.registry, self.logger) await self._broadcast_model_state(parsed) self._log_next_action_summary(parsed) return parsed except ValidationError: # Just re-raise - Pydantic's validation errors are already descriptive raise except (ModelRateLimitError, ModelProviderError) as e: # Check if we can switch to a fallback LLM if not self._try_switch_to_fallback_llm(e): # No fallback available, re-raise the original error raise # Retry with the fallback LLM return await self.get_model_output(input_messages) def _try_switch_to_fallback_llm(self, error: ModelRateLimitError | ModelProviderError) -> bool: """ Attempt to switch to a fallback LLM after a rate limit or provider error. Returns True if successfully switched to a fallback, False if no fallback available. Once switched, the agent will use the fallback LLM for the rest of the run. """ # Already using fallback - can't switch again if self._using_fallback_llm: self.logger.warning( f'⚠️ Fallback LLM also failed ({type(error).__name__}: {error.message}), no more fallbacks available' ) return False # Check if error is retryable (rate limit, auth errors, or server errors) # 401: API key invalid/expired - fallback to different provider # 402: Insufficient credits/payment required - fallback to different provider # 429: Rate limit exceeded # 500, 502, 503, 504: Server errors # ModelOutputTruncatedError: not retryable on the same model, but a fallback may have a higher cap retryable_status_codes = {401, 402, 429, 500, 502, 503, 504} is_retryable = isinstance(error, (ModelRateLimitError, ModelOutputTruncatedError)) or ( hasattr(error, 'status_code') and error.status_code in retryable_status_codes ) if not is_retryable: return False # Check if we have a fallback LLM configured if self._fallback_llm is None: self.logger.warning(f'⚠️ LLM error ({type(error).__name__}: {error.message}) but no fallback_llm configured') return False self._log_fallback_switch(error, self._fallback_llm) # Switch to the fallback LLM self.llm = self._fallback_llm self._using_fallback_llm = True # Register the fallback LLM for token cost tracking self.token_cost_service.register_llm(self._fallback_llm) return True def _log_fallback_switch(self, error: ModelRateLimitError | ModelProviderError, fallback: BaseChatModel) -> None: """Log when switching to a fallback LLM.""" original_model = self._original_llm.model if hasattr(self._original_llm, 'model') else 'unknown' fallback_model = fallback.model if hasattr(fallback, 'model') else 'unknown' error_type = type(error).__name__ status_code = getattr(error, 'status_code', 'N/A') self.logger.warning( f'⚠️ Primary LLM ({original_model}) failed with {error_type} (status={status_code}), ' f'switching to fallback LLM ({fallback_model})' ) async def _log_agent_run(self) -> None: """Log the agent run""" # Blue color for task self.logger.info(f'\033[34m🎯 Task: {self.task}\033[0m') self.logger.debug(f'πŸ€– Browser-Use Library Version {self.version} ({self.source})') # Check for latest version and log upgrade message if needed if CONFIG.BROWSER_USE_VERSION_CHECK: latest_version = await check_latest_browser_use_version() if latest_version and latest_version != self.version: self.logger.info( f'πŸ“¦ Newer version available: {latest_version} (current: {self.version}). Upgrade with: uv add browser-use=={latest_version}' ) def _log_first_step_startup(self) -> None: """Log startup message only on the first step""" if len(self.history.history) == 0: self.logger.info( f'Starting a browser-use agent with version {self.version}, with provider={self.llm.provider} and model={self.llm.model}' ) def _log_step_context(self, browser_state_summary: BrowserStateSummary) -> None: """Log step context information""" url = browser_state_summary.url if browser_state_summary else '' url_short = url[:50] + '...' if len(url) > 50 else url interactive_count = len(browser_state_summary.dom_state.selector_map) if browser_state_summary else 0 self.logger.info('\n') self.logger.info(f'πŸ“ Step {self.state.n_steps}:') self.logger.debug(f'Evaluating page with {interactive_count} interactive elements on: {url_short}') def _log_next_action_summary(self, parsed: 'AgentOutput') -> None: """Log a comprehensive summary of the next action(s)""" if not (self.logger.isEnabledFor(logging.DEBUG) and parsed.action): return # Collect action details action_details = [] for i, action in enumerate(parsed.action): action_data = action.model_dump(exclude_unset=True) action_name = next(iter(action_data.keys())) if action_data else 'unknown' action_params = action_data.get(action_name, {}) if action_data else {} # Format key parameters concisely param_summary = [] if isinstance(action_params, dict): for key, value in action_params.items(): if key == 'index': param_summary.append(f'#{value}') elif key == 'text' and isinstance(value, str): text_preview = value[:30] + '...' if len(value) > 30 else value param_summary.append(f'text="{text_preview}"') elif key == 'url': param_summary.append(f'url="{value}"') elif key == 'success': param_summary.append(f'success={value}') elif isinstance(value, (str, int, bool)): val_str = str(value)[:30] + '...' if len(str(value)) > 30 else str(value) param_summary.append(f'{key}={val_str}') param_str = f'({", ".join(param_summary)})' if param_summary else '' action_details.append(f'{action_name}{param_str}') def _prepare_demo_message(self, message: str, limit: int = 600) -> str: # Previously truncated long entries; keep full text for better context in demo panel return message.strip() async def _demo_mode_log(self, message: str, level: str = 'info', metadata: dict[str, Any] | None = None) -> None: if not self._demo_mode_enabled or not message or self.browser_session is None: return try: await self.browser_session.send_demo_mode_log( message=self._prepare_demo_message(message), level=level, metadata=metadata or {}, ) except Exception as exc: self.logger.debug(f'[DemoMode] Failed to send overlay log: {exc}') async def _broadcast_model_state(self, parsed: 'AgentOutput') -> None: if not self._demo_mode_enabled: return state = parsed.current_state step_meta = {'step': self.state.n_steps} if state.thinking: await self._demo_mode_log(state.thinking, 'thought', step_meta) if state.evaluation_previous_goal: eval_text = state.evaluation_previous_goal level = 'success' if 'success' in eval_text.lower() else 'warning' if 'failure' in eval_text.lower() else 'info' await self._demo_mode_log(eval_text, level, step_meta) if state.memory: await self._demo_mode_log(f'Memory: {state.memory}', 'info', step_meta) if state.next_goal: await self._demo_mode_log(f'Next goal: {state.next_goal}', 'info', step_meta) def _log_step_completion_summary(self, step_start_time: float, result: list[ActionResult]) -> str | None: """Log step completion summary with action count, timing, and success/failure stats""" if not result: return None step_duration = time.time() - step_start_time action_count = len(result) # Count success and failures success_count = sum(1 for r in result if not r.error) failure_count = action_count - success_count # Format success/failure indicators success_indicator = f'βœ… {success_count}' if success_count > 0 else '' failure_indicator = f'❌ {failure_count}' if failure_count > 0 else '' status_parts = [part for part in [success_indicator, failure_indicator] if part] status_str = ' | '.join(status_parts) if status_parts else 'βœ… 0' message = ( f'πŸ“ Step {self.state.n_steps}: Ran {action_count} action{"" if action_count == 1 else "s"} ' f'in {step_duration:.2f}s: {status_str}' ) self.logger.debug(message) return message def _log_final_outcome_messages(self) -> None: """Log helpful messages to user based on agent run outcome""" # Check if agent failed is_successful = self.history.is_successful() if is_successful is False or is_successful is None: # Get final result to check for specific failure reasons final_result = self.history.final_result() final_result_str = str(final_result).lower() if final_result else '' # Check for captcha/cloudflare related failures captcha_keywords = ['captcha', 'cloudflare', 'recaptcha', 'challenge', 'bot detection', 'access denied'] has_captcha_issue = any(keyword in final_result_str for keyword in captcha_keywords) if has_captcha_issue: self.logger.warning( 'Agent was blocked by a captcha. Cloud browsers include stealth fingerprinting and proxy rotation to avoid this.\n' ' Try: Browser(use_cloud=True) | Get an API key: https://cloud.browser-use.com?utm_source=oss&utm_medium=captcha_nudge' ) # General failure message self.logger.info('') self.logger.info('Did the Agent not work as expected? Let us fix this!') self.logger.info(' Open a short issue on GitHub: https://github.com/browser-use/browser-use/issues') def _log_agent_event(self, max_steps: int, agent_run_error: str | None = None) -> None: """Sent the agent event for this run to telemetry""" token_summary = self.token_cost_service.get_usage_tokens_for_model(self.llm.model) # Prepare action_history data correctly action_history_data = [] for item in self.history.history: if item.model_output and item.model_output.action: # Convert each ActionModel in the step to its dictionary representation step_actions = [ action.model_dump(exclude_unset=True) for action in item.model_output.action if action # Ensure action is not None if list allows it ] action_history_data.append(step_actions) else: # Append None or [] if a step had no actions or no model output action_history_data.append(None) final_res = self.history.final_result() final_result_str = json.dumps(final_res) if final_res is not None else None # Extract judgement data if available judgement_data = self.history.judgement() judge_verdict = judgement_data.get('verdict') if judgement_data else None judge_reasoning = judgement_data.get('reasoning') if judgement_data else None judge_failure_reason = judgement_data.get('failure_reason') if judgement_data else None judge_reached_captcha = judgement_data.get('reached_captcha') if judgement_data else None judge_impossible_task = judgement_data.get('impossible_task') if judgement_data else None self.telemetry.capture( AgentTelemetryEvent( task=self.task, model=self.llm.model, model_provider=self.llm.provider, max_steps=max_steps, max_actions_per_step=self.settings.max_actions_per_step, use_vision=self.settings.use_vision, version=self.version, source=self.source, cdp_url=urlparse(self.browser_session.cdp_url).hostname if self.browser_session and self.browser_session.cdp_url else None, agent_type=None, # Regular Agent (not code-use) action_errors=self.history.errors(), action_history=action_history_data, urls_visited=self.history.urls(), steps=self.state.n_steps, total_input_tokens=token_summary.prompt_tokens, total_output_tokens=token_summary.completion_tokens, prompt_cached_tokens=token_summary.prompt_cached_tokens, total_tokens=token_summary.total_tokens, total_duration_seconds=self.history.total_duration_seconds(), success=self.history.is_successful(), final_result_response=final_result_str, error_message=agent_run_error, judge_verdict=judge_verdict, judge_reasoning=judge_reasoning, judge_failure_reason=judge_failure_reason, judge_reached_captcha=judge_reached_captcha, judge_impossible_task=judge_impossible_task, ) ) async def take_step(self, step_info: AgentStepInfo | None = None) -> tuple[bool, bool]: """Take a step Returns: Tuple[bool, bool]: (is_done, is_valid) """ if step_info is not None and step_info.step_number == 0: # First step self._log_first_step_startup() # Normally there was no try catch here but the callback can raise an InterruptedError which we skip try: await self._execute_initial_actions() except InterruptedError: pass except Exception as e: raise e await self.step(step_info) if self.history.is_done(): await self.log_completion() # Run full judge before done callback if enabled if self.settings.use_judge: await self._judge_and_log() if self.register_done_callback: if inspect.iscoroutinefunction(self.register_done_callback): await self.register_done_callback(self.history) else: self.register_done_callback(self.history) return True, True return False, False def _extract_start_url(self, task: str) -> str | None: """Extract URL from task string using naive pattern matching.""" import re # Remove email addresses from task before looking for URLs task_without_emails = re.sub(r'\b[A-Za-z0-9._%+-]+@[A-Za-z0-9.-]+\.[A-Z|a-z]{2,}\b', '', task) # Look for common URL patterns patterns = [ r'https?://[^\s<>"\']+', # Full URLs with http/https r'(?:www\.)?[a-zA-Z0-9-]+(?:\.[a-zA-Z0-9-]+)*\.[a-zA-Z]{2,}(?:/[^\s<>"\']*)?', # Domain names with subdomains and optional paths ] # File extensions that should be excluded from URL detection # These are likely files rather than web pages to navigate to excluded_extensions = { # Documents 'pdf', 'doc', 'docx', 'xls', 'xlsx', 'ppt', 'pptx', 'odt', 'ods', 'odp', # Text files 'txt', 'md', 'csv', 'json', 'xml', 'yaml', 'yml', # Archives 'zip', 'rar', '7z', 'tar', 'gz', 'bz2', 'xz', # Images 'jpg', 'jpeg', 'png', 'gif', 'bmp', 'svg', 'webp', 'ico', # Audio/Video 'mp3', 'mp4', 'avi', 'mkv', 'mov', 'wav', 'flac', 'ogg', # Code/Data 'py', 'js', 'css', 'java', 'cpp', # Academic/Research 'bib', 'bibtex', 'tex', 'latex', 'cls', 'sty', # Other common file types 'exe', 'msi', 'dmg', 'pkg', 'deb', 'rpm', 'iso', # GitHub/Project paths 'polynomial', } excluded_words = { 'never', 'dont', 'not', "don't", } found_urls = [] for pattern in patterns: matches = re.finditer(pattern, task_without_emails) for match in matches: url = match.group(0) original_position = match.start() # Store original position before URL modification # Remove trailing punctuation that's not part of URLs url = sanitize_url_candidate(url) if is_placeholder_url(url): self.logger.debug(f'Excluding placeholder URL from auto-navigation: {url}') continue # Check if URL ends with a file extension that should be excluded url_lower = url.lower() should_exclude = False for ext in excluded_extensions: if f'.{ext}' in url_lower: should_exclude = True break if should_exclude: self.logger.debug(f'Excluding URL with file extension from auto-navigation: {url}') continue # If in the 20 characters before the url position is a word in excluded_words skip to avoid "Never go to this url" context_start = max(0, original_position - 20) context_text = task_without_emails[context_start:original_position] if any(word.lower() in context_text.lower() for word in excluded_words): self.logger.debug( f'Excluding URL with word in excluded words from auto-navigation: {url} (context: "{context_text.strip()}")' ) continue # Add https:// if missing (after excluded words check to avoid position calculation issues) if not url.startswith(('http://', 'https://')): url = 'https://' + url found_urls.append(url) unique_urls = list(set(found_urls)) # If multiple URLs found, skip directly_open_urling if len(unique_urls) > 1: self.logger.debug(f'Multiple URLs found ({len(found_urls)}), skipping directly_open_url to avoid ambiguity') return None # If exactly one URL found, return it if len(unique_urls) == 1: return unique_urls[0] return None async def _execute_step( self, step: int, max_steps: int, step_info: AgentStepInfo, on_step_start: AgentHookFunc | None = None, on_step_end: AgentHookFunc | None = None, ) -> bool: """ Execute a single step with timeout. Returns: bool: True if task is done, False otherwise """ if on_step_start is not None: await on_step_start(self) await self._demo_mode_log( f'Starting step {step + 1}/{max_steps}', 'info', {'step': step + 1, 'total_steps': max_steps}, ) self.logger.debug(f'🚢 Starting step {step + 1}/{max_steps}...') try: await asyncio.wait_for( self.step(step_info), timeout=self.settings.step_timeout, ) self.logger.debug(f'βœ… Completed step {step + 1}/{max_steps}') except TimeoutError: # Handle step timeout gracefully error_msg = f'Step {step + 1} timed out after {self.settings.step_timeout} seconds' self.logger.error(f'⏰ {error_msg}') await self._demo_mode_log(error_msg, 'error', {'step': step + 1}) self.state.consecutive_failures += 1 self.state.last_result = [ActionResult(error=error_msg)] # Ensure step counter advances on timeout β€” _finalize() may have # been skipped or returned early due to the cancellation. if self.state.n_steps == step + 1: self.state.n_steps += 1 if on_step_end is not None: await on_step_end(self) if self.history.is_done(): await self.log_completion() # Run full judge before done callback if enabled if self.settings.use_judge: await self._judge_and_log() if self.register_done_callback: if inspect.iscoroutinefunction(self.register_done_callback): await self.register_done_callback(self.history) else: self.register_done_callback(self.history) return True return False @observe(name='agent.run', ignore_input=True, ignore_output=True) @time_execution_async('--run') async def run( self, max_steps: int = 500, on_step_start: AgentHookFunc | None = None, on_step_end: AgentHookFunc | None = None, ) -> AgentHistoryList[AgentStructuredOutput]: """Execute the task with maximum number of steps""" loop = asyncio.get_event_loop() agent_run_error: str | None = None # Initialize error tracking variable self._force_exit_telemetry_logged = False # ADDED: Flag for custom telemetry on force exit should_delay_close = False # Set up the signal handler with callbacks specific to this agent from browser_use.utils import SignalHandler # Define the custom exit callback function for second CTRL+C def on_force_exit_log_telemetry(): self._log_agent_event(max_steps=max_steps, agent_run_error='SIGINT: Cancelled by user') # NEW: Call the flush method on the telemetry instance if hasattr(self, 'telemetry') and self.telemetry: self.telemetry.flush() self._force_exit_telemetry_logged = True # Set the flag signal_handler = SignalHandler( loop=loop, pause_callback=self.pause, resume_callback=self.resume, custom_exit_callback=on_force_exit_log_telemetry, # Pass the new telemetrycallback exit_on_second_int=True, disabled=not self.enable_signal_handler, ) signal_handler.register() try: await self._log_agent_run() self.logger.debug( f'πŸ”§ Agent setup: Agent Session ID {self.session_id[-4:]}, Task ID {self.task_id[-4:]}, Browser Session ID {self.browser_session.id[-4:] if self.browser_session else "None"} {"(connecting via CDP)" if (self.browser_session and self.browser_session.cdp_url) else "(launching local browser)"}' ) # Initialize timing for session and task self._session_start_time = time.time() self._task_start_time = self._session_start_time # Initialize task start time # Only dispatch session events if this is the first run if not self.state.session_initialized: self.logger.debug('πŸ“‘ Dispatching CreateAgentSessionEvent...') # Emit CreateAgentSessionEvent at the START of run() self.eventbus.dispatch(CreateAgentSessionEvent.from_agent(self)) self.state.session_initialized = True self.logger.debug('πŸ“‘ Dispatching CreateAgentTaskEvent...') # Emit CreateAgentTaskEvent at the START of run() self.eventbus.dispatch(CreateAgentTaskEvent.from_agent(self)) # Log startup message on first step (only if we haven't already done steps) self._log_first_step_startup() # Start browser session and attach watchdogs await self.browser_session.start() if self._demo_mode_enabled: await self._demo_mode_log(f'Started task: {self.task}', 'info', {'tag': 'task'}) await self._demo_mode_log( 'Demo mode active - follow the side panel for live thoughts and actions.', 'info', {'tag': 'status'}, ) # Register skills as actions if SkillService is configured await self._register_skills_as_actions() # Normally there was no try catch here but the callback can raise an InterruptedError. # Wrap with step_timeout so initial actions (usually a single URL navigate) can't # hang indefinitely on a silent CDP WebSocket β€” without this the agent would take # zero steps and return with an empty history while any outer watchdog waits. try: await asyncio.wait_for( self._execute_initial_actions(), timeout=self.settings.step_timeout, ) except InterruptedError: pass except TimeoutError: initial_timeout_msg = ( f'Initial actions timed out after {self.settings.step_timeout}s ' f'(browser may be unresponsive). Proceeding to main execution loop.' ) self.logger.error(f'⏰ {initial_timeout_msg}') self.state.last_result = [ActionResult(error=initial_timeout_msg)] self.state.consecutive_failures += 1 except Exception as e: raise e self.logger.debug( f'πŸ”„ Starting main execution loop with max {max_steps} steps (currently at step {self.state.n_steps})...' ) while self.state.n_steps <= max_steps: current_step = self.state.n_steps - 1 # Convert to 0-indexed for step_info # Use the consolidated pause state management if self.state.paused: self.logger.debug(f'⏸️ Step {self.state.n_steps}: Agent paused, waiting to resume...') await self._external_pause_event.wait() signal_handler.reset() # Check if we should stop due to too many failures, if final_response_after_failure is True, we try one last time if (self.state.consecutive_failures) >= self.settings.max_failures + int( self.settings.final_response_after_failure ): self.logger.error(f'❌ Stopping due to {self.settings.max_failures} consecutive failures') agent_run_error = f'Stopped due to {self.settings.max_failures} consecutive failures' break # Check control flags before each step if self.state.stopped: self.logger.info('πŸ›‘ Agent stopped') agent_run_error = 'Agent stopped programmatically' break step_info = AgentStepInfo(step_number=current_step, max_steps=max_steps) is_done = await self._execute_step(current_step, max_steps, step_info, on_step_start, on_step_end) if is_done: # Agent has marked the task as done if self._demo_mode_enabled and self.history.history: final_result_text = self.history.final_result() or 'Task completed' await self._demo_mode_log(f'Final Result: {final_result_text}', 'success', {'tag': 'task'}) should_delay_close = True break else: agent_run_error = 'Failed to complete task in maximum steps' self.history.add_item( AgentHistory( model_output=None, result=[ActionResult(error=agent_run_error, include_in_memory=True)], state=BrowserStateHistory( url='', title='', tabs=[], interacted_element=[], screenshot_path=None, ), metadata=None, ) ) self.logger.info(f'❌ {agent_run_error}') self.history.usage = await self.token_cost_service.get_usage_summary() # set the model output schema and call it on the fly if self.history._output_model_schema is None and self.output_model_schema is not None: self.history._output_model_schema = self.output_model_schema return self.history except KeyboardInterrupt: # Already handled by our signal handler, but catch any direct KeyboardInterrupt as well self.logger.debug('Got KeyboardInterrupt during execution, returning current history') agent_run_error = 'KeyboardInterrupt' self.history.usage = await self.token_cost_service.get_usage_summary() return self.history except Exception as e: self.logger.error(f'Agent run failed with exception: {e}', exc_info=True) agent_run_error = str(e) raise e finally: if should_delay_close and self._demo_mode_enabled and agent_run_error is None: await asyncio.sleep(30) if agent_run_error: await self._demo_mode_log(f'Agent stopped: {agent_run_error}', 'error', {'tag': 'run'}) # Log token usage summary await self.token_cost_service.log_usage_summary() # Unregister signal handlers before cleanup signal_handler.unregister() if not self._force_exit_telemetry_logged: # MODIFIED: Check the flag try: self._log_agent_event(max_steps=max_steps, agent_run_error=agent_run_error) except Exception as log_e: # Catch potential errors during logging itself self.logger.error(f'Failed to log telemetry event: {log_e}', exc_info=True) else: # ADDED: Info message when custom telemetry for SIGINT was already logged self.logger.debug('Telemetry for force exit (SIGINT) was logged by custom exit callback.') # NOTE: CreateAgentSessionEvent and CreateAgentTaskEvent are now emitted at the START of run() # to match backend requirements for CREATE events to be fired when entities are created, # not when they are completed # Emit UpdateAgentTaskEvent at the END of run() with final task state self.eventbus.dispatch(UpdateAgentTaskEvent.from_agent(self)) # Generate GIF if needed before stopping event bus if self.settings.generate_gif: output_path: str = 'agent_history.gif' if isinstance(self.settings.generate_gif, str): output_path = self.settings.generate_gif # Lazy import gif module to avoid heavy startup cost from browser_use.agent.gif import create_history_gif create_history_gif(task=self.task, history=self.history, output_path=output_path) # Only emit output file event if GIF was actually created if Path(output_path).exists(): output_event = await CreateAgentOutputFileEvent.from_agent_and_file(self, output_path) self.eventbus.dispatch(output_event) # Log final messages to user based on outcome self._log_final_outcome_messages() # Stop the event bus gracefully, waiting for all events to be processed # Configurable via TIMEOUT_AgentEventBusStop env var (default: 3.0s) await self.eventbus.stop(clear=True, timeout=_get_timeout('TIMEOUT_AgentEventBusStop', 3.0)) await self.close() @observe_debug(ignore_input=True, ignore_output=True) @time_execution_async('--multi_act') async def multi_act(self, actions: list[ActionModel]) -> list[ActionResult]: """Execute multiple actions with page-change guards. Two layers of protection prevent executing actions against stale DOM: 1. Static flag: actions tagged with terminates_sequence=True (navigate, search, go_back, switch) automatically abort remaining queued actions. 2. Runtime detection: after every action, the current URL and focused target are compared to pre-action values. Any change aborts the remaining queue. """ results: list[ActionResult] = [] total_actions = len(actions) assert self.browser_session is not None, 'BrowserSession is not set up' try: if ( self.browser_session._cached_browser_state_summary is not None and self.browser_session._cached_browser_state_summary.dom_state is not None ): cached_selector_map = dict(self.browser_session._cached_browser_state_summary.dom_state.selector_map) else: cached_selector_map = {} except Exception as e: self.logger.error(f'Error getting cached selector map: {e}') cached_selector_map = {} for i, action in enumerate(actions): # Get action name from the action model BEFORE try block to ensure it's always available in except action_data = action.model_dump(exclude_unset=True) action_name = next(iter(action_data.keys())) if action_data else 'unknown' if i > 0: # ONLY ALLOW TO CALL `done` IF IT IS A SINGLE ACTION if action_data.get('done') is not None: msg = f'Done action is allowed only as a single action - stopped after action {i} / {total_actions}.' self.logger.debug(msg) break # wait between actions (only after first action) if i > 0: self.logger.debug(f'Waiting {self.browser_profile.wait_between_actions} seconds between actions') await asyncio.sleep(self.browser_profile.wait_between_actions) try: await self._check_stop_or_pause() # Log action before execution await self._log_action(action, action_name, i + 1, total_actions) # Capture pre-action state for runtime page-change detection pre_action_url = await self.browser_session.get_current_page_url() pre_action_focus = self.browser_session.agent_focus_target_id result = await self.tools.act( action=action, browser_session=self.browser_session, file_system=self.file_system, page_extraction_llm=self.settings.page_extraction_llm, sensitive_data=self.sensitive_data, available_file_paths=self.available_file_paths, extraction_schema=self.extraction_schema, ) if result.error: await self._demo_mode_log( f'Action "{action_name}" failed: {result.error}', 'error', {'action': action_name, 'step': self.state.n_steps}, ) elif result.is_done: completion_text = result.long_term_memory or result.extracted_content or 'Task marked as done.' level = 'success' if result.success is not False else 'warning' await self._demo_mode_log( completion_text, level, {'action': action_name, 'step': self.state.n_steps}, ) results.append(result) if results[-1].is_done or results[-1].error or i == total_actions - 1: break # --- Page-change guards (only when more actions remain) --- # Layer 1: Static flag β€” action metadata declares it changes the page registered_action = self.tools.registry.registry.actions.get(action_name) if registered_action and registered_action.terminates_sequence: self.logger.info( f'Action "{action_name}" terminates sequence β€” skipping {total_actions - i - 1} remaining action(s)' ) break # Layer 2: Runtime detection β€” URL or focus target changed post_action_url = await self.browser_session.get_current_page_url() post_action_focus = self.browser_session.agent_focus_target_id if post_action_url != pre_action_url or post_action_focus != pre_action_focus: self.logger.info(f'Page changed after "{action_name}" β€” skipping {total_actions - i - 1} remaining action(s)') break except Exception as e: # Re-raise InterruptedError so _check_stop_or_pause's stop/pause signal still propagates if isinstance(e, InterruptedError): raise # Re-raise browser/connection errors so _handle_step_error can handle reconnect/shutdown if self._is_connection_like_error(e): raise # Handle any exceptions during action execution self.logger.error(f'❌ Executing action {i + 1} failed -> {type(e).__name__}: {e}') await self._demo_mode_log( f'Action "{action_name}" raised {type(e).__name__}: {e}', 'error', {'action': action_name, 'step': self.state.n_steps}, ) # Preserve partial results so the agent knows which actions succeeded before the failure results.append(ActionResult(error=f'{type(e).__name__}: {e}')) return results return results async def _log_action(self, action, action_name: str, action_num: int, total_actions: int) -> None: """Log the action before execution with colored formatting""" # Color definitions blue = '\033[34m' # Action name magenta = '\033[35m' # Parameter names reset = '\033[0m' # Format action number and name if total_actions > 1: action_header = f'▢️ [{action_num}/{total_actions}] {blue}{action_name}{reset}:' plain_header = f'▢️ [{action_num}/{total_actions}] {action_name}:' else: action_header = f'▢️ {blue}{action_name}{reset}:' plain_header = f'▢️ {action_name}:' # Get action parameters action_data = action.model_dump(exclude_unset=True) params = action_data.get(action_name, {}) # Build parameter parts with colored formatting param_parts = [] plain_param_parts = [] if params and isinstance(params, dict): for param_name, value in params.items(): # Truncate long values for readability if isinstance(value, str) and len(value) > 150: display_value = value[:150] + '...' elif isinstance(value, list) and len(str(value)) > 200: display_value = str(value)[:200] + '...' else: display_value = value param_parts.append(f'{magenta}{param_name}{reset}: {display_value}') plain_param_parts.append(f'{param_name}: {display_value}') # Join all parts if param_parts: params_string = ', '.join(param_parts) self.logger.info(f' {action_header} {params_string}') else: self.logger.info(f' {action_header}') if self._demo_mode_enabled: panel_message = plain_header if plain_param_parts: panel_message = f'{panel_message} {", ".join(plain_param_parts)}' await self._demo_mode_log(panel_message.strip(), 'action', {'action': action_name, 'step': self.state.n_steps}) async def log_completion(self) -> None: """Log the completion of the task""" # self._task_end_time = time.time() # self._task_duration = self._task_end_time - self._task_start_time TODO: this is not working when using take_step if self.history.is_successful(): self.logger.info('βœ… Task completed successfully') await self._demo_mode_log('Task completed successfully', 'success', {'tag': 'task'}) async def _generate_rerun_summary( self, original_task: str, results: list[ActionResult], summary_llm: BaseChatModel | None = None ) -> ActionResult: """Generate AI summary of rerun completion using screenshot and last step info""" from browser_use.agent.views import RerunSummaryAction # Get current screenshot screenshot_b64 = None try: screenshot = await self.browser_session.take_screenshot(full_page=False) if screenshot: import base64 screenshot_b64 = base64.b64encode(screenshot).decode('utf-8') except Exception as e: self.logger.warning(f'Failed to capture screenshot for rerun summary: {e}') # Build summary prompt and message error_count = sum(1 for r in results if r.error) success_count = len(results) - error_count from browser_use.agent.prompts import get_rerun_summary_message, get_rerun_summary_prompt prompt = get_rerun_summary_prompt( original_task=original_task, total_steps=len(results), success_count=success_count, error_count=error_count, ) # Use provided LLM, agent's LLM, or fall back to OpenAI with structured output try: # Determine which LLM to use if summary_llm is None: # Try to use the agent's LLM first summary_llm = self.llm self.logger.debug('Using agent LLM for rerun summary') else: self.logger.debug(f'Using provided LLM for rerun summary: {summary_llm.model}') # Build message with prompt and optional screenshot from browser_use.llm.messages import BaseMessage message = get_rerun_summary_message(prompt, screenshot_b64) messages: list[BaseMessage] = [message] # type: ignore[list-item] # Try calling with structured output first self.logger.debug(f'Calling LLM for rerun summary with {len(messages)} message(s)') try: kwargs: dict = {'output_format': RerunSummaryAction} response = await summary_llm.ainvoke(messages, **kwargs) summary: RerunSummaryAction = response.completion # type: ignore[assignment] self.logger.debug(f'LLM response type: {type(summary)}') self.logger.debug(f'LLM response: {summary}') except Exception as structured_error: # If structured output fails (e.g., Browser-Use LLM doesn't support it for this type), # fall back to text response without parsing self.logger.debug(f'Structured output failed: {structured_error}, falling back to text response') response = await summary_llm.ainvoke(messages, None) response_text = response.completion self.logger.debug(f'LLM text response: {response_text}') # Use the text response directly as the summary summary = RerunSummaryAction( summary=response_text if isinstance(response_text, str) else str(response_text), success=error_count == 0, completion_status='complete' if error_count == 0 else ('partial' if success_count > 0 else 'failed'), ) self.logger.info(f'πŸ“Š Rerun Summary: {summary.summary}') self.logger.info(f'πŸ“Š Status: {summary.completion_status} (success={summary.success})') return ActionResult( is_done=True, success=summary.success, extracted_content=summary.summary, long_term_memory=f'Rerun completed with status: {summary.completion_status}. {summary.summary[:100]}', ) except Exception as e: self.logger.warning(f'Failed to generate AI summary: {e.__class__.__name__}: {e}') self.logger.debug('Full error traceback:', exc_info=True) # Fallback to simple summary return ActionResult( is_done=True, success=error_count == 0, extracted_content=f'Rerun completed: {success_count}/{len(results)} steps succeeded', long_term_memory=f'Rerun completed: {success_count} steps succeeded, {error_count} errors', ) async def _execute_ai_step( self, query: str, include_screenshot: bool = False, extract_links: bool = False, ai_step_llm: BaseChatModel | None = None, ) -> ActionResult: """ Execute an AI step during rerun to re-evaluate extract actions. Analyzes full page DOM/markdown + optional screenshot. Args: query: What to analyze or extract from the current page include_screenshot: Whether to include screenshot in analysis extract_links: Whether to include links in markdown extraction ai_step_llm: Optional LLM to use. If not provided, uses agent's LLM Returns: ActionResult with extracted content """ from browser_use.agent.prompts import get_ai_step_system_prompt, get_ai_step_user_prompt, get_rerun_summary_message from browser_use.llm.messages import SystemMessage, UserMessage from browser_use.utils import sanitize_surrogates # Use provided LLM or agent's LLM llm = ai_step_llm or self.llm self.logger.debug(f'Using LLM for AI step: {llm.model}') # Extract clean markdown try: from browser_use.dom.markdown_extractor import extract_clean_markdown content, content_stats = await extract_clean_markdown( browser_session=self.browser_session, extract_links=extract_links ) except Exception as e: return ActionResult(error=f'Could not extract clean markdown: {type(e).__name__}: {e}') # Get screenshot if requested screenshot_b64 = None if include_screenshot: try: screenshot = await self.browser_session.take_screenshot(full_page=False) if screenshot: import base64 screenshot_b64 = base64.b64encode(screenshot).decode('utf-8') except Exception as e: self.logger.warning(f'Failed to capture screenshot for ai_step: {e}') # Build prompt with content stats original_html_length = content_stats['original_html_chars'] initial_markdown_length = content_stats['initial_markdown_chars'] final_filtered_length = content_stats['final_filtered_chars'] chars_filtered = content_stats['filtered_chars_removed'] stats_summary = f"""Content processed: {original_html_length:,} HTML chars β†’ {initial_markdown_length:,} initial markdown β†’ {final_filtered_length:,} filtered markdown""" if chars_filtered > 0: stats_summary += f' (filtered {chars_filtered:,} chars of noise)' # Sanitize content content = sanitize_surrogates(content) query = sanitize_surrogates(query) # Get prompts from prompts.py system_prompt = get_ai_step_system_prompt() prompt_text = get_ai_step_user_prompt(query, stats_summary, content) # Build user message with optional screenshot if screenshot_b64: user_message = get_rerun_summary_message(prompt_text, screenshot_b64) else: user_message = UserMessage(content=prompt_text) try: import asyncio response = await asyncio.wait_for(llm.ainvoke([SystemMessage(content=system_prompt), user_message]), timeout=120.0) current_url = await self.browser_session.get_current_page_url() extracted_content = ( f'\n{current_url}\n\n\n{query}\n\n\n{response.completion}\n' ) # Simple memory handling MAX_MEMORY_LENGTH = 1000 if len(extracted_content) < MAX_MEMORY_LENGTH: memory = extracted_content include_extracted_content_only_once = False else: file_name = await self.file_system.save_extracted_content(extracted_content) memory = f'Query: {query}\nContent in {file_name} and once in .' include_extracted_content_only_once = True self.logger.info(f'πŸ€– AI Step: {memory}') return ActionResult( extracted_content=extracted_content, include_extracted_content_only_once=include_extracted_content_only_once, long_term_memory=memory, ) except Exception as e: self.logger.warning(f'Failed to execute AI step: {e.__class__.__name__}: {e}') self.logger.debug('Full error traceback:', exc_info=True) return ActionResult(error=f'AI step failed: {e}') async def rerun_history( self, history: AgentHistoryList, max_retries: int = 3, skip_failures: bool = False, delay_between_actions: float = 2.0, max_step_interval: float = 45.0, summary_llm: BaseChatModel | None = None, ai_step_llm: BaseChatModel | None = None, wait_for_elements: bool = False, ) -> list[ActionResult]: """ Rerun a saved history of actions with error handling and retry logic. Args: history: The history to replay max_retries: Maximum number of retries per action skip_failures: Whether to skip failed actions or stop execution. When True, also skips steps that had errors in the original run (e.g., modal close buttons that auto-dismissed, or elements that became non-interactable) delay_between_actions: Delay between actions in seconds (used when no saved interval) max_step_interval: Maximum delay from saved step_interval (caps LLM time from original run) summary_llm: Optional LLM to use for generating the final summary. If not provided, uses the agent's LLM ai_step_llm: Optional LLM to use for AI steps (extract actions). If not provided, uses the agent's LLM wait_for_elements: If True, wait for minimum number of elements before attempting element matching. Useful for SPA pages where shadow DOM content loads dynamically. Default is False. Returns: List of action results (including AI summary as the final result) """ # Skip cloud sync session events for rerunning (we're replaying, not starting new) self.state.session_initialized = True # Initialize browser session await self.browser_session.start() results = [] # Track previous step for redundant retry detection previous_item: AgentHistory | None = None previous_step_succeeded: bool = False try: for i, history_item in enumerate(history.history): goal = history_item.model_output.current_state.next_goal if history_item.model_output else '' step_num = history_item.metadata.step_number if history_item.metadata else i step_name = 'Initial actions' if step_num == 0 else f'Step {step_num}' # Determine step delay if history_item.metadata and history_item.metadata.step_interval is not None: # Cap the saved interval to max_step_interval (saved interval includes LLM time) step_delay = min(history_item.metadata.step_interval, max_step_interval) # Format delay nicely - show ms for values < 1s, otherwise show seconds if step_delay < 1.0: delay_str = f'{step_delay * 1000:.0f}ms' else: delay_str = f'{step_delay:.1f}s' if history_item.metadata.step_interval > max_step_interval: delay_source = f'capped to {delay_str} (saved was {history_item.metadata.step_interval:.1f}s)' else: delay_source = f'using saved step_interval={delay_str}' else: step_delay = delay_between_actions if step_delay < 1.0: delay_str = f'{step_delay * 1000:.0f}ms' else: delay_str = f'{step_delay:.1f}s' delay_source = f'using default delay={delay_str}' self.logger.info(f'Replaying {step_name} ({i + 1}/{len(history.history)}) [{delay_source}]: {goal}') if ( not history_item.model_output or not history_item.model_output.action or history_item.model_output.action == [None] ): self.logger.warning(f'{step_name}: No action to replay, skipping') results.append(ActionResult(error='No action to replay')) continue # Check if the original step had errors - skip if skip_failures is enabled original_had_error = any(r.error for r in history_item.result if r.error) if original_had_error and skip_failures: error_msgs = [r.error for r in history_item.result if r.error] self.logger.warning( f'{step_name}: Original step had error(s), skipping (skip_failures=True): {error_msgs[0][:100] if error_msgs else "unknown"}' ) results.append( ActionResult( error=f'Skipped - original step had error: {error_msgs[0][:100] if error_msgs else "unknown"}' ) ) continue # Check if this step is a redundant retry of the previous step # This handles cases where original run needed to click same element multiple times # due to slow page response, but during replay the first click already worked if self._is_redundant_retry_step(history_item, previous_item, previous_step_succeeded): self.logger.info(f'{step_name}: Skipping redundant retry (previous step already succeeded with same element)') results.append( ActionResult( extracted_content='Skipped - redundant retry of previous step', include_in_memory=False, ) ) # Don't update previous_item/previous_step_succeeded - keep tracking the original step continue retry_count = 0 step_succeeded = False menu_reopened = False # Track if we've already tried reopening the menu # Exponential backoff: 5s base, doubling each retry, capped at 30s base_retry_delay = 5.0 max_retry_delay = 30.0 while retry_count < max_retries: try: result = await self._execute_history_step(history_item, step_delay, ai_step_llm, wait_for_elements) results.extend(result) step_succeeded = True break except Exception as e: error_str = str(e) retry_count += 1 # Check if this is a "Could not find matching element" error for a menu item # If so, try to re-open the dropdown from the previous step before retrying if ( not menu_reopened and 'Could not find matching element' in error_str and previous_item is not None and self._is_menu_opener_step(previous_item) ): # Check if current step targets a menu item element curr_elements = history_item.state.interacted_element if history_item.state else [] curr_elem = curr_elements[0] if curr_elements else None if self._is_menu_item_element(curr_elem): self.logger.info( 'πŸ”„ Dropdown may have closed. Attempting to re-open by re-executing previous step...' ) reopened = await self._reexecute_menu_opener(previous_item, ai_step_llm) if reopened: menu_reopened = True # Don't increment retry_count for the menu reopen attempt # Retry immediately with minimal delay retry_count -= 1 step_delay = 0.5 # Use short delay after reopening self.logger.info('πŸ”„ Dropdown re-opened, retrying element match...') continue if retry_count == max_retries: error_msg = f'{step_name} failed after {max_retries} attempts: {error_str}' self.logger.error(error_msg) # Always record the error in results so AI summary counts it correctly results.append(ActionResult(error=error_msg)) if not skip_failures: raise RuntimeError(error_msg) # With skip_failures=True, continue to next step else: # Exponential backoff: 5s, 10s, 20s, ... capped at 30s retry_delay = min(base_retry_delay * (2 ** (retry_count - 1)), max_retry_delay) self.logger.warning( f'{step_name} failed (attempt {retry_count}/{max_retries}), retrying in {retry_delay}s...' ) await asyncio.sleep(retry_delay) # Update tracking for redundant retry detection previous_item = history_item previous_step_succeeded = step_succeeded # Generate AI summary of rerun completion self.logger.info('πŸ€– Generating AI summary of rerun completion...') summary_result = await self._generate_rerun_summary(self.task, results, summary_llm) results.append(summary_result) return results finally: # Always close resources, even on failure await self.close() async def _execute_initial_actions(self) -> None: # Execute initial actions if provided if self.initial_actions and not self.state.follow_up_task: self.logger.debug(f'⚑ Executing {len(self.initial_actions)} initial actions...') result = await self.multi_act(self.initial_actions) # update result 1 to mention that its was automatically loaded if result and self.initial_url and result[0].long_term_memory: result[0].long_term_memory = f'Found initial url and automatically loaded it. {result[0].long_term_memory}' self.state.last_result = result # Save initial actions to history as step 0 for rerun capability # Skip browser state capture for initial actions (usually just URL navigation) if self.settings.flash_mode: model_output = self.AgentOutput( evaluation_previous_goal=None, memory='Initial navigation', next_goal=None, action=self.initial_actions, ) else: model_output = self.AgentOutput( evaluation_previous_goal='Start', memory=None, next_goal='Initial navigation', action=self.initial_actions, ) metadata = StepMetadata(step_number=0, step_start_time=time.time(), step_end_time=time.time(), step_interval=None) # Create minimal browser state history for initial actions state_history = BrowserStateHistory( url=self.initial_url or '', title='Initial Actions', tabs=[], interacted_element=[None] * len(self.initial_actions), # No DOM elements needed screenshot_path=None, ) history_item = AgentHistory( model_output=model_output, result=result, state=state_history, metadata=metadata, ) self.history.add_item(history_item) self.logger.debug('πŸ“ Saved initial actions to history as step 0') self.logger.debug('Initial actions completed') async def _wait_for_minimum_elements( self, min_elements: int, timeout: float = 30.0, poll_interval: float = 1.0, ) -> BrowserStateSummary | None: """Wait for the page to have at least min_elements interactive elements. This helps handle SPA pages where shadow DOM and dynamic content may not be immediately available even when document.readyState is 'complete'. Args: min_elements: Minimum number of interactive elements to wait for timeout: Maximum time to wait in seconds poll_interval: Time between polling attempts in seconds Returns: BrowserStateSummary if minimum elements found, None if timeout """ assert self.browser_session is not None, 'BrowserSession is not set up' start_time = time.time() last_count = 0 while (time.time() - start_time) < timeout: state = await self.browser_session.get_browser_state_summary(include_screenshot=False) if state and state.dom_state.selector_map: current_count = len(state.dom_state.selector_map) if current_count >= min_elements: self.logger.debug(f'βœ… Page has {current_count} elements (needed {min_elements}), proceeding with action') return state if current_count != last_count: self.logger.debug( f'⏳ Waiting for elements: {current_count}/{min_elements} ' f'(timeout in {timeout - (time.time() - start_time):.1f}s)' ) last_count = current_count await asyncio.sleep(poll_interval) # Return last state even if we didn't reach min_elements self.logger.warning(f'⚠️ Timeout waiting for {min_elements} elements, proceeding with {last_count} elements') return await self.browser_session.get_browser_state_summary(include_screenshot=False) def _count_expected_elements_from_history(self, history_item: AgentHistory) -> int: """Estimate the minimum number of elements expected based on history. Uses the action indices from the history to determine the minimum number of elements the page should have. If an action targets index N, the page needs at least N+1 elements in the selector_map. """ if not history_item.model_output or not history_item.model_output.action: return 0 max_index = -1 # Use -1 to indicate no index found yet for action in history_item.model_output.action: # Get the element index this action targets index = action.get_index() if index is not None: max_index = max(max_index, index) # Need at least max_index + 1 elements (indices are 0-based) # Cap at 50 to avoid waiting forever for very high indices # max_index >= 0 means we found at least one action with an index return min(max_index + 1, 50) if max_index >= 0 else 0 async def _execute_history_step( self, history_item: AgentHistory, delay: float, ai_step_llm: BaseChatModel | None = None, wait_for_elements: bool = False, ) -> list[ActionResult]: """Execute a single step from history with element validation. For extract actions, uses AI to re-evaluate the content since page content may have changed. Args: history_item: The history step to execute delay: Delay before executing the step ai_step_llm: Optional LLM to use for AI steps wait_for_elements: If True, wait for minimum elements before element matching """ assert self.browser_session is not None, 'BrowserSession is not set up' await asyncio.sleep(delay) # Optionally wait for minimum elements before element matching (useful for SPAs) if wait_for_elements: # Determine if we need to wait for elements (actions that interact with DOM elements) needs_element_matching = False if history_item.model_output: for i, action in enumerate(history_item.model_output.action): action_data = action.model_dump(exclude_unset=True) action_name = next(iter(action_data.keys()), None) # Actions that need element matching if action_name in ('click', 'input', 'hover', 'select_option', 'drag_and_drop'): historical_elem = ( history_item.state.interacted_element[i] if i < len(history_item.state.interacted_element) else None ) if historical_elem is not None: needs_element_matching = True break # If we need element matching, wait for minimum elements before proceeding if needs_element_matching: min_elements = self._count_expected_elements_from_history(history_item) if min_elements > 0: state = await self._wait_for_minimum_elements(min_elements, timeout=15.0, poll_interval=1.0) else: state = await self.browser_session.get_browser_state_summary(include_screenshot=False) else: state = await self.browser_session.get_browser_state_summary(include_screenshot=False) else: state = await self.browser_session.get_browser_state_summary(include_screenshot=False) if not state or not history_item.model_output: raise ValueError('Invalid state or model output') results = [] pending_actions = [] for i, action in enumerate(history_item.model_output.action): # Check if this is an extract action - use AI step instead action_data = action.model_dump(exclude_unset=True) action_name = next(iter(action_data.keys()), None) if action_name == 'extract': # Execute any pending actions first to maintain correct order # (e.g., if step is [click, extract], click must happen before extract) if pending_actions: batch_results = await self.multi_act(pending_actions) results.extend(batch_results) pending_actions = [] # Now execute AI step for extract action extract_params = action_data['extract'] query = extract_params.get('query', '') extract_links = extract_params.get('extract_links', False) self.logger.info(f'πŸ€– Using AI step for extract action: {query[:50]}...') ai_result = await self._execute_ai_step( query=query, include_screenshot=False, # Match original extract behavior extract_links=extract_links, ai_step_llm=ai_step_llm, ) results.append(ai_result) else: # For non-extract actions, update indices and collect for batch execution historical_elem = history_item.state.interacted_element[i] updated_action = await self._update_action_indices( historical_elem, action, state, ) if updated_action is None: # Build informative error message with diagnostic info elem_info = self._format_element_for_error(historical_elem) selector_map = state.dom_state.selector_map or {} selector_count = len(selector_map) # Find elements with same node_name for diagnostics hist_node = historical_elem.node_name.lower() if historical_elem else '' similar_elements = [] if historical_elem and historical_elem.attributes: for idx, elem in selector_map.items(): if elem.node_name.lower() == hist_node and elem.attributes: elem_aria = elem.attributes.get('aria-label', '') if elem_aria: similar_elements.append(f'{idx}:{elem_aria[:30]}') if len(similar_elements) >= 5: break diagnostic = '' if similar_elements: diagnostic = f'\n Available <{hist_node.upper()}> with aria-label: {similar_elements}' elif hist_node: same_node_count = sum(1 for e in selector_map.values() if e.node_name.lower() == hist_node) diagnostic = ( f'\n Found {same_node_count} <{hist_node.upper()}> elements (none with matching identifiers)' ) raise ValueError( f'Could not find matching element for action {i} in current page.\n' f' Looking for: {elem_info}\n' f' Page has {selector_count} interactive elements.{diagnostic}\n' f' Tried: EXACT hash β†’ STABLE hash β†’ XPATH β†’ AX_NAME β†’ ATTRIBUTE matching' ) pending_actions.append(updated_action) # Execute any remaining pending actions if pending_actions: batch_results = await self.multi_act(pending_actions) results.extend(batch_results) return results async def _update_action_indices( self, historical_element: DOMInteractedElement | None, action: ActionModel, # Type this properly based on your action model browser_state_summary: BrowserStateSummary, ) -> ActionModel | None: """ Update action indices based on current page state. Returns updated action or None if element cannot be found. Cascading matching strategy (tries each level in order): 1. EXACT: Full element_hash match (includes all attributes + ax_name) 2. STABLE: Hash with dynamic CSS classes filtered out (focus, hover, animation, etc.) 3. XPATH: XPath string match (structural position in DOM) 4. AX_NAME: Accessible name match from accessibility tree (robust for dynamic menus) 5. ATTRIBUTE: Unique attribute match (name, id, aria-label) for old history files """ if not historical_element or not browser_state_summary.dom_state.selector_map: return action selector_map = browser_state_summary.dom_state.selector_map highlight_index: int | None = None match_level: MatchLevel | None = None # Debug: log what we're looking for and what's available self.logger.info( f'πŸ” Searching for element: <{historical_element.node_name}> ' f'hash={historical_element.element_hash} stable_hash={historical_element.stable_hash}' ) # Log what elements are in selector_map for debugging if historical_element.node_name: hist_name = historical_element.node_name.lower() matching_nodes = [ (idx, elem.node_name, elem.attributes.get('name') if elem.attributes else None) for idx, elem in selector_map.items() if elem.node_name.lower() == hist_name ] self.logger.info( f'πŸ” Selector map has {len(selector_map)} elements, ' f'{len(matching_nodes)} are <{hist_name.upper()}>: {matching_nodes}' ) # Level 1: EXACT hash match for idx, elem in selector_map.items(): if elem.element_hash == historical_element.element_hash: highlight_index = idx match_level = MatchLevel.EXACT break if highlight_index is None: self.logger.debug(f'EXACT hash match failed (checked {len(selector_map)} elements)') # Level 2: STABLE hash match (dynamic classes filtered) # Use stored stable_hash (computed at save time from EnhancedDOMTreeNode - single source of truth) if highlight_index is None and historical_element.stable_hash is not None: for idx, elem in selector_map.items(): if elem.compute_stable_hash() == historical_element.stable_hash: highlight_index = idx match_level = MatchLevel.STABLE self.logger.info('Element matched at STABLE level (dynamic classes filtered)') break if highlight_index is None: self.logger.debug('STABLE hash match failed') elif highlight_index is None: self.logger.debug('STABLE hash match skipped (no stable_hash in history)') # Level 3: XPATH match if highlight_index is None and historical_element.x_path: for idx, elem in selector_map.items(): if elem.xpath == historical_element.x_path: highlight_index = idx match_level = MatchLevel.XPATH self.logger.info(f'Element matched at XPATH level: {historical_element.x_path}') break if highlight_index is None: self.logger.debug(f'XPATH match failed for: {historical_element.x_path[-60:]}') # Level 4: ax_name (accessible name) match - robust for dynamic SPAs with menus # This uses the accessible name from the accessibility tree which is stable # even when DOM structure changes (e.g., dynamically generated menu items) if highlight_index is None and historical_element.ax_name: hist_name = historical_element.node_name.lower() hist_ax_name = historical_element.ax_name for idx, elem in selector_map.items(): # Match by node type and accessible name elem_ax_name = elem.ax_node.name if elem.ax_node else None if elem.node_name.lower() == hist_name and elem_ax_name == hist_ax_name: highlight_index = idx match_level = MatchLevel.AX_NAME self.logger.info(f'Element matched at AX_NAME level: "{hist_ax_name}"') break if highlight_index is None: # Log available ax_names for debugging same_type_ax_names = [ (idx, elem.ax_node.name if elem.ax_node else None) for idx, elem in selector_map.items() if elem.node_name.lower() == hist_name and elem.ax_node and elem.ax_node.name ] self.logger.debug( f'AX_NAME match failed for <{hist_name.upper()}> ax_name="{hist_ax_name}". ' f'Page has {len(same_type_ax_names)} <{hist_name.upper()}> with ax_names: ' f'{same_type_ax_names[:5]}{"..." if len(same_type_ax_names) > 5 else ""}' ) # Level 5: Unique attribute fallback (for old history files without stable_hash) if highlight_index is None and historical_element.attributes: hist_attrs = historical_element.attributes hist_name = historical_element.node_name.lower() # Try matching by unique identifiers: name, id, or aria-label for attr_key in ['name', 'id', 'aria-label']: if attr_key in hist_attrs and hist_attrs[attr_key]: for idx, elem in selector_map.items(): if ( elem.node_name.lower() == hist_name and elem.attributes and elem.attributes.get(attr_key) == hist_attrs[attr_key] ): highlight_index = idx match_level = MatchLevel.ATTRIBUTE self.logger.info(f'Element matched via {attr_key} attribute: {hist_attrs[attr_key]}') break if highlight_index is not None: break if highlight_index is None: tried_attrs = [k for k in ['name', 'id', 'aria-label'] if k in hist_attrs and hist_attrs[k]] # Log what was tried and what's available on the page for debugging same_node_elements = [ (idx, elem.attributes.get('aria-label') or elem.attributes.get('id') or elem.attributes.get('name')) for idx, elem in selector_map.items() if elem.node_name.lower() == hist_name and elem.attributes ] self.logger.info( f'πŸ” ATTRIBUTE match failed for <{hist_name.upper()}> ' f'(tried: {tried_attrs}, looking for: {[hist_attrs.get(k) for k in tried_attrs]}). ' f'Page has {len(same_node_elements)} <{hist_name.upper()}> elements with identifiers: ' f'{same_node_elements[:5]}{"..." if len(same_node_elements) > 5 else ""}' ) if highlight_index is None: return None old_index = action.get_index() if old_index != highlight_index: action.set_index(highlight_index) level_name = match_level.name if match_level else 'UNKNOWN' self.logger.info(f'Element index updated {old_index} β†’ {highlight_index} (matched at {level_name} level)') return action def _format_element_for_error(self, elem: DOMInteractedElement | None) -> str: """Format element info for error messages during history rerun.""" if elem is None: return '' parts = [f'<{elem.node_name}>'] # Add key identifying attributes if elem.attributes: for key in ['name', 'id', 'aria-label', 'type']: if key in elem.attributes and elem.attributes[key]: parts.append(f'{key}="{elem.attributes[key]}"') # Add hash info parts.append(f'hash={elem.element_hash}') if elem.stable_hash: parts.append(f'stable_hash={elem.stable_hash}') # Add xpath (truncated) if elem.x_path: xpath_short = elem.x_path if len(elem.x_path) <= 60 else f'...{elem.x_path[-57:]}' parts.append(f'xpath="{xpath_short}"') return ' '.join(parts) def _is_redundant_retry_step( self, current_item: AgentHistory, previous_item: AgentHistory | None, previous_step_succeeded: bool, ) -> bool: """ Detect if current step is a redundant retry of the previous step. This handles cases where the original run needed to click the same element multiple times due to slow page response, but during replay the first click already succeeded. When the page has already navigated, subsequent retry clicks on the same element would fail because that element no longer exists. Returns True if: - Previous step succeeded - Both steps target the same element (by element_hash, stable_hash, or xpath) - Both steps perform the same action type (e.g., both are clicks) """ if not previous_item or not previous_step_succeeded: return False # Get interacted elements from both steps (first action in each) curr_elements = current_item.state.interacted_element prev_elements = previous_item.state.interacted_element if not curr_elements or not prev_elements: return False curr_elem = curr_elements[0] if curr_elements else None prev_elem = prev_elements[0] if prev_elements else None if not curr_elem or not prev_elem: return False # Check if same element by various matching strategies same_by_hash = curr_elem.element_hash == prev_elem.element_hash same_by_stable_hash = ( curr_elem.stable_hash is not None and prev_elem.stable_hash is not None and curr_elem.stable_hash == prev_elem.stable_hash ) same_by_xpath = curr_elem.x_path == prev_elem.x_path if not (same_by_hash or same_by_stable_hash or same_by_xpath): return False # Check if same action type curr_actions = current_item.model_output.action if current_item.model_output else [] prev_actions = previous_item.model_output.action if previous_item.model_output else [] if not curr_actions or not prev_actions: return False # Get the action type (first key in the action dict) curr_action_data = curr_actions[0].model_dump(exclude_unset=True) prev_action_data = prev_actions[0].model_dump(exclude_unset=True) curr_action_type = next(iter(curr_action_data.keys()), None) prev_action_type = next(iter(prev_action_data.keys()), None) if curr_action_type != prev_action_type: return False self.logger.debug( f'πŸ”„ Detected redundant retry: both steps target same element ' f'<{curr_elem.node_name}> with action "{curr_action_type}"' ) return True def _is_menu_opener_step(self, history_item: AgentHistory | None) -> bool: """ Detect if a step opens a dropdown/menu. Checks for common patterns indicating a menu opener: - Element has aria-haspopup attribute - Element has data-gw-click="toggleSubMenu" (Guidewire pattern) - Element has expand-button in class name - Element role is "menuitem" with aria-expanded Returns True if the step appears to open a dropdown/submenu. """ if not history_item or not history_item.state or not history_item.state.interacted_element: return False elem = history_item.state.interacted_element[0] if history_item.state.interacted_element else None if not elem: return False attrs = elem.attributes or {} # Check for common menu opener indicators if attrs.get('aria-haspopup') in ('true', 'menu', 'listbox'): return True if attrs.get('data-gw-click') == 'toggleSubMenu': return True if 'expand-button' in attrs.get('class', ''): return True if attrs.get('role') == 'menuitem' and attrs.get('aria-expanded') in ('false', 'true'): return True if attrs.get('role') == 'button' and attrs.get('aria-expanded') in ('false', 'true'): return True return False def _is_menu_item_element(self, elem: 'DOMInteractedElement | None') -> bool: """ Detect if an element is a menu item that appears inside a dropdown/menu. Checks for: - role="menuitem", "option", "menuitemcheckbox", "menuitemradio" - Element is inside a menu structure (has menu-related parent indicators) - ax_name is set (menu items typically have accessible names) Returns True if the element appears to be a menu item. """ if not elem: return False attrs = elem.attributes or {} # Check for menu item roles role = attrs.get('role', '') if role in ('menuitem', 'option', 'menuitemcheckbox', 'menuitemradio', 'treeitem'): return True # Elements in Guidewire menus have these patterns if 'gw-action--inner' in attrs.get('class', ''): return True if 'menuitem' in attrs.get('class', '').lower(): return True # If element has an ax_name and looks like it could be in a menu # This is a softer check - only used if the previous step was a menu opener if elem.ax_name and elem.ax_name not in ('', None): # Common menu container classes elem_class = attrs.get('class', '').lower() if any(x in elem_class for x in ['dropdown', 'popup', 'menu', 'submenu', 'action']): return True return False async def _reexecute_menu_opener( self, opener_item: AgentHistory, ai_step_llm: 'BaseChatModel | None' = None, ) -> bool: """ Re-execute a menu opener step to re-open a closed dropdown. This is used when a menu item can't be found because the dropdown closed during the wait between steps. Returns True if re-execution succeeded, False otherwise. """ try: self.logger.info('πŸ”„ Re-opening dropdown/menu by re-executing previous step...') # Use a minimal delay - we want to quickly re-open the menu await self._execute_history_step(opener_item, delay=0.5, ai_step_llm=ai_step_llm, wait_for_elements=False) # Small delay to let the menu render await asyncio.sleep(0.3) return True except Exception as e: self.logger.warning(f'Failed to re-open dropdown: {e}') return False async def load_and_rerun( self, history_file: str | Path | None = None, variables: dict[str, str] | None = None, **kwargs, ) -> list[ActionResult]: """ Load history from file and rerun it, optionally substituting variables. Args: history_file: Path to the history file variables: Optional dict mapping variable names to new values (e.g. {'email': 'new@example.com'}) **kwargs: Additional arguments passed to rerun_history: - max_retries: Maximum retries per action (default: 3) - skip_failures: Continue on failure (default: True) - delay_between_actions: Delay when no saved interval (default: 2.0s) - max_step_interval: Cap on saved step_interval (default: 45.0s) - summary_llm: Custom LLM for final summary - ai_step_llm: Custom LLM for extract re-evaluation """ if not history_file: history_file = 'AgentHistory.json' history = AgentHistoryList.load_from_file(history_file, self.AgentOutput) # Substitute variables if provided if variables: history = self._substitute_variables_in_history(history, variables) return await self.rerun_history(history, **kwargs) def save_history(self, file_path: str | Path | None = None) -> None: """Save the history to a file with sensitive data filtering""" if not file_path: file_path = 'AgentHistory.json' self.history.save_to_file(file_path, sensitive_data=self.sensitive_data) def pause(self) -> None: """Pause the agent before the next step""" print('\n\n⏸️ Paused the agent and left the browser open.\n\tPress [Enter] to resume or [Ctrl+C] again to quit.') self.state.paused = True self._external_pause_event.clear() def resume(self) -> None: """Resume the agent""" # TODO: Locally the browser got closed print('----------------------------------------------------------------------') print('▢️ Resuming agent execution where it left off...\n') self.state.paused = False self._external_pause_event.set() def stop(self) -> None: """Stop the agent""" self.logger.info('⏹️ Agent stopping') self.state.stopped = True # Signal pause event to unblock any waiting code so it can check the stopped state self._external_pause_event.set() # Task stopped def _convert_initial_actions(self, actions: list[dict[str, dict[str, Any]]]) -> list[ActionModel]: """Convert dictionary-based actions to ActionModel instances""" converted_actions = [] action_model = self.ActionModel for action_dict in actions: # Each action_dict should have a single key-value pair action_name = next(iter(action_dict)) params = action_dict[action_name] # Get the parameter model for this action from registry action_info = self.tools.registry.registry.actions[action_name] param_model = action_info.param_model # Create validated parameters using the appropriate param model validated_params = param_model(**params) # Create ActionModel instance with the validated parameters action_model = self.ActionModel(**{action_name: validated_params}) converted_actions.append(action_model) return converted_actions def _verify_and_setup_llm(self): """ Verify that the LLM API keys are setup and the LLM API is responding properly. Also handles tool calling method detection if in auto mode. """ # Skip verification if already done if getattr(self.llm, '_verified_api_keys', None) is True or CONFIG.SKIP_LLM_API_KEY_VERIFICATION: setattr(self.llm, '_verified_api_keys', True) return True @property def message_manager(self) -> MessageManager: return self._message_manager async def close(self): """Close all resources""" try: # Only close browser if keep_alive is False (or not set) if self.browser_session is not None: if not self.browser_session.browser_profile.keep_alive: # Kill the browser session - this dispatches BrowserStopEvent, # stops the EventBus with clear=True, and recreates a fresh EventBus await self.browser_session.kill() else: # keep_alive=True sessions shouldn't keep the event loop alive after agent.run() await self.browser_session.event_bus.stop( clear=False, timeout=_get_timeout('TIMEOUT_BrowserSessionEventBusStopOnAgentClose', 1.0), ) try: self.browser_session.event_bus.event_queue = None self.browser_session.event_bus._on_idle = None except Exception: pass # Close skill service if configured if self.skill_service is not None: await self.skill_service.close() # Force garbage collection gc.collect() # Debug: Log remaining threads and asyncio tasks import threading threads = threading.enumerate() self.logger.debug(f'🧡 Remaining threads ({len(threads)}): {[t.name for t in threads]}') # Get all asyncio tasks tasks = asyncio.all_tasks(asyncio.get_event_loop()) # Filter out the current task (this close() coroutine) other_tasks = [t for t in tasks if t != asyncio.current_task()] if other_tasks: self.logger.debug(f'⚑ Remaining asyncio tasks ({len(other_tasks)}):') for task in other_tasks[:10]: # Limit to first 10 to avoid spam self.logger.debug(f' - {task.get_name()}: {task}') except Exception as e: self.logger.error(f'Error during cleanup: {e}') async def _update_action_models_for_page(self, page_url: str) -> None: """Update action models with page-specific actions""" # Create new action model with current page's filtered actions self.ActionModel = self.tools.registry.create_action_model(page_url=page_url) # Update output model with the new actions if self.settings.flash_mode: self.AgentOutput = AgentOutput.type_with_custom_actions_flash_mode(self.ActionModel) elif self.settings.use_thinking: self.AgentOutput = AgentOutput.type_with_custom_actions(self.ActionModel) else: self.AgentOutput = AgentOutput.type_with_custom_actions_no_thinking(self.ActionModel) # Update done action model too self.DoneActionModel = self.tools.registry.create_action_model(include_actions=['done'], page_url=page_url) if self.settings.flash_mode: self.DoneAgentOutput = AgentOutput.type_with_custom_actions_flash_mode(self.DoneActionModel) elif self.settings.use_thinking: self.DoneAgentOutput = AgentOutput.type_with_custom_actions(self.DoneActionModel) else: self.DoneAgentOutput = AgentOutput.type_with_custom_actions_no_thinking(self.DoneActionModel) async def authenticate_cloud_sync(self, show_instructions: bool = True) -> bool: """ Authenticate with cloud service for future runs. This is useful when users want to authenticate after a task has completed so that future runs will sync to the cloud. Args: show_instructions: Whether to show authentication instructions to user Returns: bool: True if authentication was successful """ self.logger.warning('Cloud sync has been removed and is no longer available') return False def run_sync( self, max_steps: int = 500, on_step_start: AgentHookFunc | None = None, on_step_end: AgentHookFunc | None = None, ) -> AgentHistoryList[AgentStructuredOutput]: """Synchronous wrapper around the async run method for easier usage without asyncio.""" import asyncio return asyncio.run(self.run(max_steps=max_steps, on_step_start=on_step_start, on_step_end=on_step_end)) def detect_variables(self) -> dict[str, DetectedVariable]: """Detect reusable variables in agent history""" from browser_use.agent.variable_detector import detect_variables_in_history return detect_variables_in_history(self.history) def _substitute_variables_in_history(self, history: AgentHistoryList, variables: dict[str, str]) -> AgentHistoryList: """Substitute variables in history with new values for rerunning with different data""" from browser_use.agent.variable_detector import detect_variables_in_history # Detect variables in the history detected_vars = detect_variables_in_history(history) # Build a mapping of original values to new values value_replacements: dict[str, str] = {} for var_name, new_value in variables.items(): if var_name in detected_vars: old_value = detected_vars[var_name].original_value value_replacements[old_value] = new_value else: self.logger.warning(f'Variable "{var_name}" not found in history, skipping substitution') if not value_replacements: self.logger.info('No variables to substitute') return history # Create a deep copy of history to avoid modifying the original import copy modified_history = copy.deepcopy(history) # Substitute values in all actions substitution_count = 0 for history_item in modified_history.history: if not history_item.model_output or not history_item.model_output.action: continue for action in history_item.model_output.action: # Handle both Pydantic models and dicts if hasattr(action, 'model_dump'): action_dict = action.model_dump() elif isinstance(action, dict): action_dict = action else: action_dict = vars(action) if hasattr(action, '__dict__') else {} # Substitute in all string fields substitution_count += self._substitute_in_dict(action_dict, value_replacements) # Update the action with modified values if hasattr(action, 'model_dump'): # For Pydantic RootModel, we need to recreate from the modified dict if hasattr(action, 'root'): # This is a RootModel - recreate it from the modified dict new_action = type(action).model_validate(action_dict) # Replace the root field in-place using object.__setattr__ to bypass Pydantic's immutability object.__setattr__(action, 'root', getattr(new_action, 'root')) else: # Regular Pydantic model - update fields in-place for key, val in action_dict.items(): if hasattr(action, key): setattr(action, key, val) elif isinstance(action, dict): action.update(action_dict) self.logger.info(f'Substituted {substitution_count} value(s) in {len(value_replacements)} variable type(s) in history') return modified_history def _substitute_in_dict(self, data: dict, replacements: dict[str, str]) -> int: """Recursively substitute values in a dictionary, returns count of substitutions made""" count = 0 for key, value in data.items(): if isinstance(value, str): # Replace if exact match if value in replacements: data[key] = replacements[value] count += 1 elif isinstance(value, dict): # Recurse into nested dicts count += self._substitute_in_dict(value, replacements) elif isinstance(value, list): # Handle lists for i, item in enumerate(value): if isinstance(item, str) and item in replacements: value[i] = replacements[item] count += 1 elif isinstance(item, dict): count += self._substitute_in_dict(item, replacements) return count _PythonAgent = Agent