Merge branch 'speed-ups3' of https://github.com/browser-use/browser-use into speed-ups3

browser_use/agent/cloud_events.py

@@ -188,7 +188,7 @@ class CreateAgentTaskEvent(BaseEvent):
 	user_id: str = Field(max_length=255)  # Added for authorization checks
 	device_id: str | None = Field(None, max_length=255)  # Device ID for auth lookup
 	agent_session_id: str
-	llm_model: str = Field(max_length=100)  # LLMModel enum value as string
+	llm_model: str = Field(max_length=200)  # LLMModel enum value as string
 	stopped: bool = False
 	paused: bool = False
 	task: str = Field(max_length=MAX_TASK_LENGTH)

browser_use/agent/prompts.py

@@ -28,7 +28,7 @@ class SystemPrompt:
 		self.use_thinking = use_thinking
 		self.flash_mode = flash_mode
 		prompt = ''
-		if override_system_message:
+		if override_system_message is not None:
 			prompt = override_system_message
 		else:
 			self._load_prompt_template()
@@ -265,7 +265,8 @@ class AgentMessagePrompt:
 		# Check if current page is a PDF viewer and add appropriate message
 		pdf_message = ''
 		if self.browser_state.is_pdf_viewer:
-			pdf_message = 'PDF viewer cannot be rendered. In this page, DO NOT use the extract_structured_data action as PDF content cannot be rendered. Use the read_file action on the downloaded PDF in available_file_paths to read the full content.\n\n'
+			pdf_message = 'PDF viewer cannot be rendered. In this page, DO NOT use the extract_structured_data action as PDF content cannot be rendered. '
+			pdf_message += 'Use the read_file action on the downloaded PDF in available_file_paths to read the full text content or scroll in the page to see images/figures if needed.\n\n'
 
 		# Add recent events if available and requested
 		recent_events_text = ''

browser_use/agent/service.py

@@ -624,7 +624,10 @@ class Agent(Generic[Context, AgentStructuredOutput]):
 		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
-		self.eventbus = EventBus(name=f'Agent_{str(self.id)[-self.state.n_steps :]}')
+		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}')
 
 		# Re-register cloud sync handler if it exists (if not disabled)
 		if hasattr(self, 'cloud_sync') and self.cloud_sync and self.enable_cloud_sync:
@@ -681,7 +684,6 @@ class Agent(Generic[Context, AgentStructuredOutput]):
 		# Always take screenshots for all steps
 		self.logger.debug('📸 Requesting browser state with include_screenshot=True')
 		browser_state_summary = await self.browser_session.get_browser_state_summary(
-			cache_clickable_elements_hashes=True,
 			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,
 		)
@@ -1662,7 +1664,6 @@ class Agent(Generic[Context, AgentStructuredOutput]):
 			# This prevents stale element detection but doesn't refresh before execution
 			if action.get_index() is not None and i != 0:
 				new_browser_state_summary = await self.browser_session.get_browser_state_summary(
-					cache_clickable_elements_hashes=False,
 					include_screenshot=False,
 				)
 				new_selector_map = new_browser_state_summary.dom_state.selector_map
@@ -1888,9 +1889,7 @@ class Agent(Generic[Context, AgentStructuredOutput]):
 	async def _execute_history_step(self, history_item: AgentHistory, delay: float) -> list[ActionResult]:
 		"""Execute a single step from history with element validation"""
 		assert self.browser_session is not None, 'BrowserSession is not set up'
-		state = await self.browser_session.get_browser_state_summary(
-			cache_clickable_elements_hashes=False, include_screenshot=False
-		)
+		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')
 		updated_actions = []

browser_use/browser/events.py

@@ -190,7 +190,6 @@ class BrowserStateRequestEvent(BaseEvent[BrowserStateSummary]):
 
 	include_dom: bool = True
 	include_screenshot: bool = True
-	cache_clickable_elements_hashes: bool = True
 	include_recent_events: bool = False
 
 	event_timeout: float | None = _get_timeout('TIMEOUT_BrowserStateRequestEvent', 30.0)  # seconds

browser_use/browser/session.py

@@ -837,6 +837,17 @@ class BrowserSession(BaseModel):
 				self.agent_focus = await self.get_or_create_cdp_session(target_id=last_target_id, focus=True)
 				raise
 
+		# Dispatch NavigationCompleteEvent when tab focus changes
+		# This ensures PDF detection and downloads work when switching tabs
+		if event.target_id and event.url:
+			self.logger.debug(f'🔄 Dispatching NavigationCompleteEvent for tab switch to {event.url[:50]}...')
+			await self.event_bus.dispatch(
+				NavigationCompleteEvent(
+					target_id=event.target_id,
+					url=event.url,
+				)
+			)
+
 		# self.logger.debug('🔄 AgentFocusChangedEvent handler completed successfully')
 
 	async def on_FileDownloadedEvent(self, event: FileDownloadedEvent) -> None:
@@ -1061,7 +1072,6 @@ class BrowserSession(BaseModel):
 	@observe_debug(ignore_input=True, ignore_output=True, name='get_browser_state_summary')
 	async def get_browser_state_summary(
 		self,
-		cache_clickable_elements_hashes: bool = True,
 		include_screenshot: bool = True,
 		cached: bool = False,
 		include_recent_events: bool = False,
@@ -1088,7 +1098,6 @@ class BrowserSession(BaseModel):
 				BrowserStateRequestEvent(
 					include_dom=True,
 					include_screenshot=include_screenshot,
-					cache_clickable_elements_hashes=cache_clickable_elements_hashes,
 					include_recent_events=include_recent_events,
 				)
 			),

browser_use/browser/watchdogs/downloads_watchdog.py

@@ -56,6 +56,7 @@ class DownloadsWatchdog(BaseWatchdog):
 	_cdp_event_tasks: set[asyncio.Task] = PrivateAttr(default_factory=set)  # Track CDP event handler tasks
 	_cdp_downloads_info: dict[str, dict[str, Any]] = PrivateAttr(default_factory=dict)  # Map guid -> info
 	_use_js_fetch_for_local: bool = PrivateAttr(default=False)  # Guard JS fetch path for local regular downloads
+	_session_pdf_urls: dict[str, str] = PrivateAttr(default_factory=dict)  # URL -> path for PDFs downloaded this session
 
 	async def on_BrowserLaunchEvent(self, event: BrowserLaunchEvent) -> None:
 		self.logger.debug(f'[DownloadsWatchdog] Received BrowserLaunchEvent, EventBus ID: {id(self.event_bus)}')
@@ -123,6 +124,7 @@ class DownloadsWatchdog(BaseWatchdog):
 		self._sessions_with_listeners.clear()
 		self._active_downloads.clear()
 		self._pdf_viewer_cache.clear()
+		self._session_pdf_urls.clear()
 
 	async def on_NavigationCompleteEvent(self, event: NavigationCompleteEvent) -> None:
 		"""Check for PDFs after navigation completes."""
@@ -801,13 +803,26 @@ class DownloadsWatchdog(BaseWatchdog):
 
 			self.logger.debug(f'[DownloadsWatchdog] Generated filename: {pdf_filename}')
 
-			# Check if already downloaded by looking in the downloads directory
+			# Check if already downloaded in this session
+			self.logger.debug(f'[DownloadsWatchdog] PDF_URL: {pdf_url}, session_pdf_urls: {self._session_pdf_urls}')
+			if pdf_url in self._session_pdf_urls:
+				existing_path = self._session_pdf_urls[pdf_url]
+				self.logger.debug(f'[DownloadsWatchdog] PDF already downloaded in session: {existing_path}')
+				return existing_path
+
+			# Generate unique filename if file exists from previous run
 			downloads_dir = str(self.browser_session.browser_profile.downloads_path)
-			if os.path.exists(downloads_dir):
-				existing_files = os.listdir(downloads_dir)
-				if pdf_filename in existing_files:
-					self.logger.debug(f'[DownloadsWatchdog] PDF already downloaded: {pdf_filename}')
-					return None
+			os.makedirs(downloads_dir, exist_ok=True)
+			final_filename = pdf_filename
+			existing_files = os.listdir(downloads_dir)
+			if pdf_filename in existing_files:
+				# Generate unique name with (1), (2), etc.
+				base, ext = os.path.splitext(pdf_filename)
+				counter = 1
+				while f'{base} ({counter}){ext}' in existing_files:
+					counter += 1
+				final_filename = f'{base} ({counter}){ext}'
+				self.logger.debug(f'[DownloadsWatchdog] File exists, using: {final_filename}')
 
 			self.logger.debug(f'[DownloadsWatchdog] Starting PDF download from: {pdf_url[:100]}...')
 
@@ -858,12 +873,10 @@ class DownloadsWatchdog(BaseWatchdog):
 				download_result = result.get('result', {}).get('value', {})
 
 				if download_result and download_result.get('data') and len(download_result['data']) > 0:
-					# Ensure unique filename
-					downloads_dir = str(self.browser_session.browser_profile.downloads_path)
 					# Ensure downloads directory exists
+					downloads_dir = str(self.browser_session.browser_profile.downloads_path)
 					os.makedirs(downloads_dir, exist_ok=True)
-					unique_filename = await self._get_unique_filename(downloads_dir, pdf_filename)
-					download_path = os.path.join(downloads_dir, unique_filename)
+					download_path = os.path.join(downloads_dir, final_filename)
 
 					# Save the PDF asynchronously
 					async with await anyio.open_file(download_path, 'wb') as f:
@@ -886,13 +899,16 @@ class DownloadsWatchdog(BaseWatchdog):
 						f'[DownloadsWatchdog] ✅ Auto-downloaded PDF ({cache_status}, {response_size:,} bytes): {download_path}'
 					)
 
+					# Store URL->path mapping for this session
+					self._session_pdf_urls[pdf_url] = download_path
+
 					# Emit file downloaded event
-					self.logger.debug(f'[DownloadsWatchdog] Dispatching FileDownloadedEvent for {unique_filename}')
+					self.logger.debug(f'[DownloadsWatchdog] Dispatching FileDownloadedEvent for {final_filename}')
 					self.event_bus.dispatch(
 						FileDownloadedEvent(
 							url=pdf_url,
 							path=download_path,
-							file_name=unique_filename,
+							file_name=final_filename,
 							file_size=response_size,
 							file_type='pdf',
 							mime_type='application/pdf',

browser_use/filesystem/file_system.py

@@ -272,7 +272,7 @@ class FileSystem:
 
 					reader = pypdf.PdfReader(full_filename)
 					num_pages = len(reader.pages)
-					MAX_PDF_PAGES = 10
+					MAX_PDF_PAGES = 20
 					extra_pages = num_pages - MAX_PDF_PAGES
 					extracted_text = ''
 					for page in reader.pages[:MAX_PDF_PAGES]:

browser_use/mcp/server.py

@@ -768,7 +768,7 @@ class BrowserUseServer:
 		if not self.browser_session:
 			return 'Error: No browser session active'
 
-		state = await self.browser_session.get_browser_state_summary(cache_clickable_elements_hashes=False)
+		state = await self.browser_session.get_browser_state_summary()
 
 		result = {
 			'url': state.url,
@@ -819,10 +819,15 @@ class BrowserUseServer:
 		ExtractAction = create_model(
 			'ExtractAction',
 			__base__=ActionModel,
-			extract_structured_data=(dict[str, Any], {'query': query, 'extract_links': extract_links}),
+			extract_structured_data=dict[str, Any],
 		)
 
-		action = ExtractAction()
+		# Use model_validate because Pyright does not understand the dynamic model
+		action = ExtractAction.model_validate(
+			{
+				'extract_structured_data': {'query': query, 'extract_links': extract_links},
+			}
+		)
 		action_result = await self.tools.act(
 			action=action,
 			browser_session=self.browser_session,

docs/customize/agent/supported-models.mdx

@@ -243,6 +243,33 @@ Required environment variables:
 ALIBABA_CLOUD=
 ```
 
+## ModelScope [example](https://github.com/browser-use/browser-use/blob/main/examples/models/modelscope_example.py)
+
+```python
+from browser_use import Agent, ChatOpenAI
+from dotenv import load_dotenv
+import os
+
+load_dotenv()
+
+# Get API key from https://www.modelscope.cn/docs/model-service/API-Inference/intro
+api_key = os.getenv('MODELSCOPE_API_KEY')
+base_url = 'https://api-inference.modelscope.cn/v1/'
+
+llm = ChatOpenAI(model='Qwen/Qwen2.5-VL-72B-Instruct', api_key=api_key, base_url=base_url)
+
+agent = Agent(
+    task="Your task here",
+    llm=llm,
+    use_vision=True
+)
+```
+
+Required environment variables:
+
+```bash .env
+MODELSCOPE_API_KEY=
+```
 
 ## Other models (DeepSeek, Novita, X...)
 

examples/custom-functions/parallel_agents.py

@@ -0,0 +1,312 @@
+"""
+Simple parallel multi-agent example.
+
+This launches multiple agents in parallel to work on different tasks simultaneously.
+No complex orchestrator - just direct parallel execution.
+
+@file purpose: Demonstrates parallel multi-agent execution using asyncio
+"""
+
+import asyncio
+import os
+import sys
+from typing import List
+
+sys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))))
+
+from dotenv import load_dotenv
+load_dotenv()
+
+from browser_use import Agent
+from browser_use.llm.google import ChatGoogle
+
+# ============================================================================
+# 🔧 SIMPLE CONFIGURATION - CHANGE THIS TO YOUR DESIRED TASK
+# ============================================================================
+
+MAIN_TASK = "find age of ronaldo and messi"
+
+# Simple test - let's start with just one person to see what happens
+# MAIN_TASK = "find age of elon musk"
+
+# ============================================================================
+
+
+async def create_subtasks(main_task: str, llm) -> list[str]:
+    """
+    Use LLM to break down main task into logical subtasks
+    
+    Real examples of how this works:
+    
+    Input: "what is the revenue of nvidia, microsoft, tesla"
+    Output: [
+        "Find Nvidia's current revenue and financial data",
+        "Find Microsoft's current revenue and financial data", 
+        "Find Tesla's current revenue and financial data"
+    ]
+    
+    Input: "what are ages of musk, altman, bezos, gates"
+    Output: [
+        "Find Elon Musk's age and birth date",
+        "Find Sam Altman's age and birth date",
+        "Find Jeff Bezos's age and birth date",
+        "Find Bill Gates's age and birth date"
+    ]
+    
+    Input: "what is the population of tokyo, new york, london, paris"
+    Output: [
+        "Find Tokyo's current population",
+        "Find New York's current population",
+        "Find London's current population", 
+        "Find Paris's current population"
+    ]
+    
+    Input: "name top 10 yc companies by revenue"
+    Output: [
+        "Research Y Combinator's top companies by revenue",
+        "Find revenue data for top YC companies",
+        "Compile list of top 10 YC companies by revenue"
+    ]
+    """
+    
+    prompt = f"""
+    Break down this main task into individual, separate subtasks where each subtask focuses on ONLY ONE specific person, company, or item:
+    
+    Main task: {main_task}
+    
+    RULES:
+    - Each subtask must focus on ONLY ONE person/company/item
+    - Do NOT combine multiple people/companies/items in one subtask
+    - Each subtask should be completely independent
+    - If the main task mentions multiple items, create one subtask per item
+    
+    Return only the subtasks, one per line, without numbering or bullets.
+    Each line should focus on exactly ONE person/company/item.
+    """
+    
+    try:
+        # Use the correct method for ChatGoogle
+        response = await llm.ainvoke(prompt)
+        
+        # Debug: Print the response type and content
+        print(f"DEBUG: Response type: {type(response)}")
+        print(f"DEBUG: Response content: {response}")
+        
+        # Handle different response types - ChatGoogle returns string content
+        if hasattr(response, 'content'):
+            content = response.content
+        elif isinstance(response, str):
+            content = response
+        elif hasattr(response, 'text'):
+            content = response.text
+        else:
+            # Convert to string if it's some other type
+            content = str(response)
+        
+        # Split by newlines and clean up
+        subtasks = [task.strip() for task in content.strip().split('\n') if task.strip()]
+        
+        # Remove any numbering or bullets that the LLM might add
+        cleaned_subtasks = []
+        for task in subtasks:
+            # Remove common prefixes like "1. ", "- ", "* ", etc.
+            cleaned = task.lstrip('0123456789.-* ')
+            if cleaned:
+                cleaned_subtasks.append(cleaned)
+        
+        return cleaned_subtasks if cleaned_subtasks else simple_split_task(main_task)
+    except Exception as e:
+        print(f"Error creating subtasks: {e}")
+        # Fallback to simple split
+        return simple_split_task(main_task)
+
+
+def simple_split_task(main_task: str) -> list[str]:
+    """Simple fallback: split task by common separators"""
+    task_lower = main_task.lower()
+    
+    # Try to split by common separators
+    if " and " in task_lower:
+        parts = main_task.split(" and ")
+        return [part.strip() for part in parts if part.strip()]
+    elif ", " in main_task:
+        parts = main_task.split(", ")
+        return [part.strip() for part in parts if part.strip()]
+    elif "," in main_task:
+        parts = main_task.split(",")
+        return [part.strip() for part in parts if part.strip()]
+    
+    # If no separators found, return the original task
+    return [main_task]
+
+
+async def run_single_agent(task: str, llm, agent_id: int) -> tuple[int, str]:
+    """Run a single agent and return its result"""
+    print(f"🚀 Agent {agent_id} starting: {task}")
+    print(f"   📝 This agent will focus ONLY on: {task}")
+    print(f"   🌐 Creating isolated browser instance for agent {agent_id}")
+    
+    try:
+        # Create agent with its own browser session (separate browser instance)
+        from browser_use.browser import BrowserSession
+        from browser_use.browser.profile import BrowserProfile
+        import tempfile
+        
+        # Create a unique temp directory for this agent's browser data
+        temp_dir = tempfile.mkdtemp(prefix=f"browser_agent_{agent_id}_")
+        
+        # Create browser profile with custom user data directory and single tab focus
+        profile = BrowserProfile()
+        profile.user_data_dir = temp_dir
+        profile.headless = False  # Set to True if you want headless mode
+        profile.keep_alive = False  # Don't keep browser alive after task
+        
+        # Add custom args to prevent new tabs and popups
+        profile.args = [
+            '--disable-popup-blocking',
+            '--disable-extensions',
+            '--disable-plugins',
+            '--disable-images',  # Faster loading
+            '--no-first-run',
+            '--disable-default-apps',
+            '--disable-background-timer-throttling',
+            '--disable-backgrounding-occluded-windows',
+            '--disable-renderer-backgrounding',
+        ]
+        
+        # Create a new browser session for each agent with the custom profile
+        browser_session = BrowserSession(browser_profile=profile)
+        
+        # Debug: Check initial tab count
+        try:
+            await browser_session.start()
+            initial_tabs = await browser_session._cdp_get_all_pages()
+            print(f"   📊 Agent {agent_id} initial tab count: {len(initial_tabs)}")
+        except Exception as e:
+            print(f"   ⚠️ Could not check initial tabs for agent {agent_id}: {e}")
+        
+        # Create agent with the dedicated browser session and disable auto URL detection
+        agent = Agent(task=task, llm=llm, browser_session=browser_session, preload=False)
+        
+        # Run the agent with timeout to prevent hanging
+        try:
+            result = await asyncio.wait_for(agent.run(), timeout=300)  # 5 minute timeout
+        except asyncio.TimeoutError:
+            print(f"⏰ Agent {agent_id} timed out after 5 minutes")
+            result = "Task timed out"
+        
+        # Debug: Check final tab count
+        try:
+            final_tabs = await browser_session._cdp_get_all_pages()
+            print(f"   📊 Agent {agent_id} final tab count: {len(final_tabs)}")
+            for i, tab in enumerate(final_tabs):
+                print(f"      Tab {i+1}: {tab.get('url', 'unknown')[:50]}...")
+        except Exception as e:
+            print(f"   ⚠️ Could not check final tabs for agent {agent_id}: {e}")
+        
+        # Extract clean result from the agent history
+        clean_result = extract_clean_result(result)
+        
+        # Close the browser session for this agent
+        try:
+            await browser_session.kill()
+        except Exception as e:
+            print(f"⚠️ Warning: Error closing browser for agent {agent_id}: {e}")
+        
+        print(f"✅ Agent {agent_id} completed and browser closed: {task}")
+        
+        return agent_id, clean_result
+        
+    except Exception as e:
+        error_msg = f"Agent {agent_id} failed: {str(e)}"
+        print(f"❌ {error_msg}")
+        return agent_id, error_msg
+
+
+def extract_clean_result(agent_result) -> str:
+    """Extract clean result from agent history"""
+    try:
+        # Get the last result from the agent history
+        if hasattr(agent_result, 'all_results') and agent_result.all_results:
+            last_result = agent_result.all_results[-1]
+            if hasattr(last_result, 'extracted_content') and last_result.extracted_content:
+                return last_result.extracted_content
+        
+        # Fallback to string representation
+        return str(agent_result)
+    except Exception:
+        return "Result extraction failed"
+
+
+async def run_parallel_agents():
+    """Run multiple agents in parallel on different tasks"""
+    
+    # Use Gemini 1.5 Flash
+    llm = ChatGoogle(model="gemini-1.5-flash")
+    
+    # Main task to break down - use the simple configuration
+    main_task = MAIN_TASK
+    
+    print(f"🎯 Main task: {main_task}")
+    print("🧠 Creating subtasks using LLM...")
+    
+    # Create subtasks using LLM
+    subtasks = await create_subtasks(main_task, llm)
+    
+    print(f"📋 Created {len(subtasks)} subtasks:")
+    for i, task in enumerate(subtasks, 1):
+        print(f"  {i}. {task}")
+    
+    print(f"\n🔥 Starting {len(subtasks)} agents in parallel...")
+    print(f"🔍 Each agent will get its own browser instance with exactly ONE tab")
+    print(f"📊 Expected: {len(subtasks)} browser instances, {len(subtasks)} tabs total")
+    
+    # Create tasks for parallel execution
+    agent_tasks = [
+        run_single_agent(task, llm, i+1) 
+        for i, task in enumerate(subtasks)
+    ]
+    
+    # Run all agents in parallel using asyncio.gather
+    results = await asyncio.gather(*agent_tasks)
+    
+    # Print results
+    print("\n" + "="*60)
+    print("📊 PARALLEL EXECUTION RESULTS")
+    print("="*60)
+    
+    for agent_id, result in results:
+        print(f"\n🤖 Agent {agent_id} result:")
+        print(f"Task: {subtasks[agent_id-1]}")
+        print(f"Result: {result}")
+        print("-" * 50)
+    
+    print(f"\n🎉 All {len(subtasks)} parallel agents completed!")
+
+
+def main():
+    """Main function to run parallel agents"""
+    # Check if Google API key is available
+    api_key = os.getenv('GOOGLE_API_KEY')
+    if not api_key:
+        print('❌ Error: GOOGLE_API_KEY environment variable not set')
+        print('Please set your Google API key to use parallel agents')
+        print('You can set it with: export GOOGLE_API_KEY="your-key-here"')
+        sys.exit(1)
+    
+    # Check if API key looks valid (Google API keys are typically 39 characters)
+    if len(api_key) < 20:
+        print(f'⚠️  Warning: GOOGLE_API_KEY seems too short ({len(api_key)} characters)')
+        print('Google API keys are typically 39 characters long')
+        print('Continuing anyway, but this might cause authentication issues...')
+    
+    print('🚀 Starting parallel multi-agent example...')
+    print(f'📝 Task: {MAIN_TASK}')
+    print('This will dynamically create agents based on task complexity')
+    print('-' * 60)
+    
+    asyncio.run(run_parallel_agents())
+
+
+if __name__ == "__main__":
+    main()

examples/models/modelscope_example.py

@@ -0,0 +1,34 @@
+"""
+Simple try of the agent.
+
+@dev You need to add MODELSCOPE_API_KEY to your environment variables.
+"""
+
+import asyncio
+import os
+
+from dotenv import load_dotenv
+
+from browser_use import Agent, ChatOpenAI
+
+# dotenv
+load_dotenv()
+
+api_key = os.getenv('MODELSCOPE_API_KEY', '')
+if not api_key:
+	raise ValueError('MODELSCOPE_API_KEY is not set')
+
+
+async def run_search():
+	agent = Agent(
+		# task=('go to amazon.com, search for laptop'),
+		task=('go to google, search for modelscope'),
+		llm=ChatOpenAI(base_url='https://api-inference.modelscope.cn/v1/', model='Qwen/Qwen2.5-VL-72B-Instruct', api_key=api_key),
+		use_vision=False,
+	)
+
+	await agent.run()
+
+
+if __name__ == '__main__':
+	asyncio.run(run_search())

tests/ci/test_browser_event_ClickElementEvent.py

@@ -143,7 +143,7 @@ class TestClickElementEvent:
 		await asyncio.sleep(0.5)  # Give page time to load
 
 		# Initialize the DOM state to populate the selector map
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map
 		selector_map = await browser_session.get_selector_map()
@@ -406,7 +406,7 @@ class TestClickElementEvent:
 		await asyncio.sleep(0.5)
 
 		# Get the clickable elements
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 		selector_map = await browser_session.get_selector_map()
 
 		# Find the inline element
@@ -488,7 +488,7 @@ class TestClickElementEvent:
 		await asyncio.sleep(0.5)
 
 		# Get the clickable elements
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 		selector_map = await browser_session.get_selector_map()
 
 		# Find the block element inside inline
@@ -576,7 +576,7 @@ class TestClickElementEvent:
 		await asyncio.sleep(0.5)
 
 		# Get the clickable elements
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 		selector_map = await browser_session.get_selector_map()
 
 		# Find the target element
@@ -636,7 +636,7 @@ class TestClickElementEvent:
 		await asyncio.sleep(0.5)
 
 		# Get the clickable elements
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 		selector_map = await browser_session.get_selector_map()
 
 		# Find the file input
@@ -699,7 +699,7 @@ class TestClickElementEvent:
 		await asyncio.sleep(0.5)
 
 		# Get the clickable elements
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 		selector_map = await browser_session.get_selector_map()
 
 		# Find the select element
@@ -1098,7 +1098,7 @@ class TestClickElementEvent:
 			await asyncio.sleep(0.5)
 
 			# Initialize the DOM state to populate the selector map
-			await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+			await browser_session.get_browser_state_summary()
 
 			# Get the selector map
 			selector_map = await browser_session.get_selector_map()

tests/ci/test_browser_event_GetDropdownOptionsEvent.py

@@ -286,7 +286,7 @@ class TestGetDropdownOptionsEvent:
 		await tools.act(GoToUrlActionModel(**goto_action), browser_session)
 
 		# Initialize the DOM state to populate the selector map
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map and find the select element
 		selector_map = await browser_session.get_selector_map()
@@ -344,7 +344,7 @@ class TestGetDropdownOptionsEvent:
 		await tools.act(GoToUrlActionModel(**goto_action), browser_session)
 
 		# Initialize the DOM state
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map and find the ARIA menu
 		selector_map = await browser_session.get_selector_map()
@@ -406,7 +406,7 @@ class TestGetDropdownOptionsEvent:
 		await tools.act(GoToUrlActionModel(**goto_action), browser_session)
 
 		# Initialize the DOM state
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map and find the custom dropdown
 		selector_map = await browser_session.get_selector_map()
@@ -495,7 +495,7 @@ class TestSelectDropdownOptionEvent:
 		await browser_session.event_bus.expect(NavigationCompleteEvent, timeout=10.0)
 
 		# Initialize the DOM state
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map and find the select element
 		selector_map = await browser_session.get_selector_map()
@@ -543,7 +543,7 @@ class TestSelectDropdownOptionEvent:
 		await browser_session.event_bus.expect(NavigationCompleteEvent, timeout=10.0)
 
 		# Initialize the DOM state
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map and find the ARIA menu
 		selector_map = await browser_session.get_selector_map()
@@ -595,7 +595,7 @@ class TestSelectDropdownOptionEvent:
 		await browser_session.event_bus.expect(NavigationCompleteEvent, timeout=10.0)
 
 		# Initialize the DOM state
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map and find the custom dropdown
 		selector_map = await browser_session.get_selector_map()
@@ -643,7 +643,7 @@ class TestSelectDropdownOptionEvent:
 		await browser_session.event_bus.expect(NavigationCompleteEvent, timeout=10.0)
 
 		# Initialize the DOM state
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map and find the select element
 		selector_map = await browser_session.get_selector_map()

tests/ci/test_browser_event_GetDropdownOptionsEvent_aria_menus.py

@@ -165,7 +165,7 @@ class TestARIAMenuDropdown:
 		await browser_session.event_bus.expect(NavigationCompleteEvent, timeout=10.0)
 
 		# Initialize the DOM state to populate the selector map
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map
 		selector_map = await browser_session.get_selector_map()
@@ -232,7 +232,7 @@ class TestARIAMenuDropdown:
 		await browser_session.event_bus.expect(NavigationCompleteEvent, timeout=10.0)
 
 		# Initialize the DOM state to populate the selector map
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map
 		selector_map = await browser_session.get_selector_map()
@@ -302,7 +302,7 @@ class TestARIAMenuDropdown:
 		await browser_session.event_bus.expect(NavigationCompleteEvent, timeout=10.0)
 
 		# Initialize the DOM state to populate the selector map
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map
 		selector_map = await browser_session.get_selector_map()

tests/ci/test_browser_event_NavigateToUrlEvent.py

@@ -97,7 +97,7 @@ class TestNavigateToUrlEvent:
 
 		# Test that get_state_summary works
 		try:
-			await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+			await browser_session.get_browser_state_summary()
 			assert False, 'Expected throw error when navigating to non-existent page'
 		except Exception as e:
 			pass

tests/ci/test_browser_session_element_cache.py

@@ -88,7 +88,7 @@ async def test_assumption_1_dom_processing_works(browser_session, httpserver):
 	await event.event_result(raise_if_any=True, raise_if_none=False)
 
 	# Trigger DOM processing
-	state = await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=False)
+	state = await browser_session.get_browser_state_summary()
 
 	print('DOM processing result:')
 	print(f'  - Elements found: {len(state.dom_state.selector_map)}')
@@ -109,7 +109,7 @@ async def test_assumption_2_cached_selector_map_persists(browser_session, httpse
 	await event.event_result(raise_if_any=True, raise_if_none=False)
 
 	# Trigger DOM processing and cache
-	state = await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=False)
+	state = await browser_session.get_browser_state_summary()
 	initial_selector_map = dict(state.dom_state.selector_map)
 
 	# Check if cached selector map is still available
@@ -136,7 +136,7 @@ async def test_assumption_3_action_gets_same_selector_map(browser_session, tools
 	await event.event_result(raise_if_any=True, raise_if_none=False)
 
 	# Trigger DOM processing and cache
-	await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=False)
+	await browser_session.get_browser_state_summary()
 	cached_selector_map = await browser_session.get_selector_map()
 
 	print('Pre-action state:')
@@ -174,7 +174,7 @@ async def test_assumption_4_click_action_specific_issue(browser_session, tools,
 	await event.event_result(raise_if_any=True, raise_if_none=False)
 
 	# Trigger DOM processing and cache
-	await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=False)
+	await browser_session.get_browser_state_summary()
 	cached_selector_map = await browser_session.get_selector_map()
 
 	print('Pre-click state:')
@@ -224,7 +224,7 @@ async def test_assumption_5_multiple_get_selector_map_calls(browser_session, htt
 	await event.event_result(raise_if_any=True, raise_if_none=False)
 
 	# Trigger DOM processing and cache
-	await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=False)
+	await browser_session.get_browser_state_summary()
 
 	# Call get_selector_map multiple times
 	map1 = await browser_session.get_selector_map()

tests/ci/test_browser_watchdog_screenshots.py

@@ -104,7 +104,7 @@ class TestHeadlessScreenshots:
 			await event.event_result(raise_if_any=True, raise_if_none=False)
 
 			# Get state summary
-			state = await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=False)
+			state = await browser_session.get_browser_state_summary()
 
 			# Verify screenshot is included
 			assert state.screenshot is not None
@@ -143,7 +143,7 @@ class TestHeadlessScreenshots:
 
 			# Browser should auto-create a new page on about:blank with animation
 			# With AboutBlankWatchdog, about:blank pages now have animated content, so they should have screenshots
-			state = await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=False)
+			state = await browser_session.get_browser_state_summary()
 			assert state.screenshot is not None, 'Screenshot should not be None for animated about:blank pages'
 			assert state.url == 'about:blank' or state.url.startswith('chrome://'), f'Expected empty page but got {state.url}'
 
@@ -153,7 +153,7 @@ class TestHeadlessScreenshots:
 			await event.event_result(raise_if_any=True, raise_if_none=False)
 
 			# Get state with screenshot
-			state = await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=False)
+			state = await browser_session.get_browser_state_summary()
 			# Should have a screenshot now
 			assert state.screenshot is not None, 'Screenshot should not be None for real pages'
 			assert isinstance(state.screenshot, str)

tests/ci/test_tools.py

@@ -413,7 +413,7 @@ class TestToolsIntegration:
 			await asyncio.sleep(1.0)
 
 		# Initialize the DOM state to populate the selector map
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Get the selector map
 		selector_map = await browser_session.get_selector_map()
@@ -540,7 +540,7 @@ class TestToolsIntegration:
 			await asyncio.sleep(1.0)
 
 		# populate the selector map with highlight indices
-		await browser_session.get_browser_state_summary(cache_clickable_elements_hashes=True)
+		await browser_session.get_browser_state_summary()
 
 		# Now get the selector map which should contain our dropdown
 		selector_map = await browser_session.get_selector_map()

tests/scripts/debug_iframe_scrolling.py

@@ -131,9 +131,7 @@ async def debug_iframe_scrolling():
 			"""Capture DOM state and return analysis"""
 			print(f'\n📸 Capturing DOM state: {label}')
 			state_event = browser_session.event_bus.dispatch(
-				BrowserStateRequestEvent(
-					include_dom=True, include_screenshot=False, cache_clickable_elements_hashes=True, include_recent_events=False
-				)
+				BrowserStateRequestEvent(include_dom=True, include_screenshot=False, include_recent_events=False)
 			)
 			browser_state = await state_event.event_result()