chore: import upstream snapshot with attribution
Test Browser Use CLI Install / uv pip install (ubuntu-latest) (push) Failing after 1s
Test Browser Use CLI Install / uvx browser-use from local wheel (push) Failing after 1s
Test Browser Use CLI Install / uvx browser-use[cli] from PyPI (push) Failing after 1s
package / pip-install-on-macos-latest-py-3.11 (push) Has been skipped
package / pip-install-on-macos-latest-py-3.13 (push) Has been skipped
package / pip-install-on-ubuntu-latest-py-3.11 (push) Has been skipped
package / pip-install-on-windows-latest-py-3.13 (push) Has been skipped
cloud_evals / trigger_cloud_eval_image_build (push) Failing after 1s
docker / build_publish_image (push) Failing after 1s
Test Browser Use CLI Install / browser-use skill sync (push) Failing after 1s
lint / code-style (push) Failing after 0s
lint / type-checker (push) Failing after 1s
package / pip-build (push) Failing after 1s
lint / syntax-errors (push) Failing after 3s
package / pip-install-on-ubuntu-latest-py-3.13 (push) Has been skipped
package / pip-install-on-windows-latest-py-3.11 (push) Has been skipped
test / ${{ matrix.test_filename }} (push) Has been skipped
test / evaluate-tasks (push) Has been skipped
test / setup-chromium (push) Failing after 2s
test / find_tests (push) Failing after 2s
Test Browser Use CLI Install / uv pip install (windows-latest) (push) Has been cancelled
Test Browser Use CLI Install / uv pip install (macos-latest) (push) Has been cancelled
Test Browser Use CLI Install / uv pip install (ubuntu-latest) (push) Failing after 1s
Test Browser Use CLI Install / uvx browser-use from local wheel (push) Failing after 1s
Test Browser Use CLI Install / uvx browser-use[cli] from PyPI (push) Failing after 1s
package / pip-install-on-macos-latest-py-3.11 (push) Has been skipped
package / pip-install-on-macos-latest-py-3.13 (push) Has been skipped
package / pip-install-on-ubuntu-latest-py-3.11 (push) Has been skipped
package / pip-install-on-windows-latest-py-3.13 (push) Has been skipped
cloud_evals / trigger_cloud_eval_image_build (push) Failing after 1s
docker / build_publish_image (push) Failing after 1s
Test Browser Use CLI Install / browser-use skill sync (push) Failing after 1s
lint / code-style (push) Failing after 0s
lint / type-checker (push) Failing after 1s
package / pip-build (push) Failing after 1s
lint / syntax-errors (push) Failing after 3s
package / pip-install-on-ubuntu-latest-py-3.13 (push) Has been skipped
package / pip-install-on-windows-latest-py-3.11 (push) Has been skipped
test / ${{ matrix.test_filename }} (push) Has been skipped
test / evaluate-tasks (push) Has been skipped
test / setup-chromium (push) Failing after 2s
test / find_tests (push) Failing after 2s
Test Browser Use CLI Install / uv pip install (windows-latest) (push) Has been cancelled
Test Browser Use CLI Install / uv pip install (macos-latest) (push) Has been cancelled
This commit is contained in:
@@ -0,0 +1,918 @@
|
||||
"""Event-driven CDP session management.
|
||||
|
||||
Manages CDP sessions by listening to Target.attachedToTarget and Target.detachedFromTarget
|
||||
events, ensuring the session pool always reflects the current browser state.
|
||||
"""
|
||||
|
||||
import asyncio
|
||||
from collections import deque
|
||||
from typing import TYPE_CHECKING, Any
|
||||
|
||||
from cdp_use.cdp.target import AttachedToTargetEvent, DetachedFromTargetEvent, SessionID, TargetID
|
||||
|
||||
from browser_use.utils import create_task_with_error_handling
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from browser_use.browser.session import BrowserSession, CDPSession, Target
|
||||
|
||||
|
||||
class SessionManager:
|
||||
"""Event-driven CDP session manager.
|
||||
|
||||
Automatically synchronizes the CDP session pool with browser state via CDP events.
|
||||
|
||||
Key features:
|
||||
- Sessions added/removed automatically via Target attach/detach events
|
||||
- Multiple sessions can attach to the same target
|
||||
- Targets only removed when ALL sessions detach
|
||||
- No stale sessions - pool always reflects browser reality
|
||||
|
||||
SessionManager is the SINGLE SOURCE OF TRUTH for all targets and sessions.
|
||||
"""
|
||||
|
||||
def __init__(self, browser_session: 'BrowserSession'):
|
||||
self.browser_session = browser_session
|
||||
self.logger = browser_session.logger
|
||||
|
||||
# All targets (entities: pages, iframes, workers)
|
||||
self._targets: dict[TargetID, 'Target'] = {}
|
||||
|
||||
# All sessions (communication channels)
|
||||
self._sessions: dict[SessionID, 'CDPSession'] = {}
|
||||
|
||||
# Mapping: target -> sessions attached to it
|
||||
self._target_sessions: dict[TargetID, set[SessionID]] = {}
|
||||
|
||||
# Reverse mapping: session -> target it belongs to
|
||||
self._session_to_target: dict[SessionID, TargetID] = {}
|
||||
|
||||
# Page lifecycle events per target, fed by ONE global Page.lifecycleEvent handler
|
||||
# registered in start_monitoring(). cdp-use's event registry is single-slot per
|
||||
# CDP method, so per-session handler registrations would replace each other and
|
||||
# leave every tab but the most recently attached one without lifecycle events.
|
||||
self._lifecycle_events: dict[TargetID, deque[dict[str, Any]]] = {}
|
||||
|
||||
self._lock = asyncio.Lock()
|
||||
self._recovery_lock = asyncio.Lock()
|
||||
|
||||
# Focus recovery coordination - event-driven instead of polling
|
||||
self._recovery_in_progress: bool = False
|
||||
self._recovery_complete_event: asyncio.Event | None = None
|
||||
self._recovery_task: asyncio.Task | None = None
|
||||
|
||||
async def start_monitoring(self) -> None:
|
||||
"""Start monitoring Target attach/detach events.
|
||||
|
||||
Registers CDP event handlers to keep the session pool synchronized with browser state.
|
||||
Also discovers and initializes all existing targets on startup.
|
||||
"""
|
||||
if not self.browser_session._cdp_client_root:
|
||||
raise RuntimeError('CDP client not initialized')
|
||||
|
||||
# Capture cdp_client_root in closure to avoid type errors
|
||||
cdp_client = self.browser_session._cdp_client_root
|
||||
|
||||
# Enable target discovery to receive targetInfoChanged events automatically
|
||||
# This eliminates the need for getTargetInfo() polling calls
|
||||
await cdp_client.send.Target.setDiscoverTargets(
|
||||
params={'discover': True, 'filter': [{'type': 'page'}, {'type': 'iframe'}]}
|
||||
)
|
||||
|
||||
# Register synchronous event handlers (CDP requirement)
|
||||
def on_attached(event: AttachedToTargetEvent, session_id: SessionID | None = None):
|
||||
# _handle_target_attached() handles:
|
||||
# - setAutoAttach for children
|
||||
# - Create CDPSession
|
||||
# - Enable monitoring (for pages/tabs)
|
||||
# - Add to pool
|
||||
create_task_with_error_handling(
|
||||
self._handle_target_attached(event),
|
||||
name='handle_target_attached',
|
||||
logger_instance=self.logger,
|
||||
suppress_exceptions=True,
|
||||
)
|
||||
|
||||
def on_detached(event: DetachedFromTargetEvent, session_id: SessionID | None = None):
|
||||
create_task_with_error_handling(
|
||||
self._handle_target_detached(event),
|
||||
name='handle_target_detached',
|
||||
logger_instance=self.logger,
|
||||
suppress_exceptions=True,
|
||||
)
|
||||
|
||||
def on_target_info_changed(event, session_id: SessionID | None = None):
|
||||
# Update session info from targetInfoChanged events (no polling needed!)
|
||||
create_task_with_error_handling(
|
||||
self._handle_target_info_changed(event),
|
||||
name='handle_target_info_changed',
|
||||
logger_instance=self.logger,
|
||||
suppress_exceptions=True,
|
||||
)
|
||||
|
||||
def on_lifecycle_event(event, session_id: SessionID | None = None):
|
||||
# ONE global handler for all targets: route by session_id -> target_id.
|
||||
# Registering per-session closures instead would clobber each other in
|
||||
# cdp-use's single-slot registry (one handler per CDP method).
|
||||
if not session_id:
|
||||
return
|
||||
target_id = self.get_target_id_from_session_id(session_id)
|
||||
if not target_id:
|
||||
return
|
||||
self.get_lifecycle_events(target_id).append(
|
||||
{
|
||||
'name': event.get('name', 'unknown'),
|
||||
'loaderId': event.get('loaderId'),
|
||||
'timestamp': asyncio.get_event_loop().time(),
|
||||
}
|
||||
)
|
||||
|
||||
cdp_client.register.Target.attachedToTarget(on_attached)
|
||||
cdp_client.register.Target.detachedFromTarget(on_detached)
|
||||
cdp_client.register.Target.targetInfoChanged(on_target_info_changed)
|
||||
cdp_client.register.Page.lifecycleEvent(on_lifecycle_event)
|
||||
|
||||
self.logger.debug('[SessionManager] Event monitoring started')
|
||||
|
||||
# Discover and initialize ALL existing targets
|
||||
await self._initialize_existing_targets()
|
||||
|
||||
def get_lifecycle_events(self, target_id: TargetID) -> 'deque[dict[str, Any]]':
|
||||
"""Get (creating if needed) the lifecycle event buffer for a target."""
|
||||
events = self._lifecycle_events.get(target_id)
|
||||
if events is None:
|
||||
events = deque(maxlen=50)
|
||||
self._lifecycle_events[target_id] = events
|
||||
return events
|
||||
|
||||
def _get_session_for_target(self, target_id: TargetID) -> 'CDPSession | None':
|
||||
"""Internal: Get ANY valid session for a target (picks first available).
|
||||
|
||||
⚠️ INTERNAL API - Use browser_session.get_or_create_cdp_session() instead!
|
||||
This method has no validation, no focus management, no recovery.
|
||||
|
||||
Args:
|
||||
target_id: Target ID to get session for
|
||||
|
||||
Returns:
|
||||
CDPSession if exists, None if target has detached
|
||||
"""
|
||||
session_ids = self._target_sessions.get(target_id, set())
|
||||
if not session_ids:
|
||||
# Check if this is the focused target - indicates stale focus that needs cleanup
|
||||
if self.browser_session.agent_focus_target_id == target_id:
|
||||
self.logger.warning(
|
||||
f'[SessionManager] ⚠️ Attempted to get session for stale focused target {target_id[:8]}... '
|
||||
f'Clearing stale focus and triggering recovery.'
|
||||
)
|
||||
|
||||
# Clear stale focus immediately (defense in depth)
|
||||
self.browser_session.agent_focus_target_id = None
|
||||
|
||||
# Trigger recovery if not already in progress
|
||||
if not self._recovery_in_progress:
|
||||
self.logger.warning('[SessionManager] Recovery was not in progress! Triggering now.')
|
||||
self._recovery_task = create_task_with_error_handling(
|
||||
self._recover_agent_focus(target_id),
|
||||
name='recover_agent_focus_from_stale_get',
|
||||
logger_instance=self.logger,
|
||||
suppress_exceptions=False,
|
||||
)
|
||||
return None
|
||||
return self._sessions.get(next(iter(session_ids)))
|
||||
|
||||
def get_all_page_targets(self) -> list:
|
||||
"""Get all page/tab targets using owned data.
|
||||
|
||||
Returns:
|
||||
List of Target objects for all page/tab targets
|
||||
"""
|
||||
page_targets = []
|
||||
for target in self._targets.values():
|
||||
if target.target_type in ('page', 'tab'):
|
||||
page_targets.append(target)
|
||||
return page_targets
|
||||
|
||||
async def validate_session(self, target_id: TargetID) -> bool:
|
||||
"""Check if a target still has active sessions.
|
||||
|
||||
Args:
|
||||
target_id: Target ID to validate
|
||||
|
||||
Returns:
|
||||
True if target has active sessions, False if it should be removed
|
||||
"""
|
||||
if target_id not in self._target_sessions:
|
||||
return False
|
||||
return len(self._target_sessions[target_id]) > 0
|
||||
|
||||
async def clear(self) -> None:
|
||||
"""Clear all owned data structures for cleanup."""
|
||||
async with self._lock:
|
||||
# Clear owned data (single source of truth)
|
||||
self._targets.clear()
|
||||
self._sessions.clear()
|
||||
self._target_sessions.clear()
|
||||
self._session_to_target.clear()
|
||||
|
||||
self.logger.info('[SessionManager] Cleared all owned data (targets, sessions, mappings)')
|
||||
|
||||
async def is_target_valid(self, target_id: TargetID) -> bool:
|
||||
"""Check if a target is still valid and has active sessions.
|
||||
|
||||
Args:
|
||||
target_id: Target ID to validate
|
||||
|
||||
Returns:
|
||||
True if target is valid and has active sessions, False otherwise
|
||||
"""
|
||||
if target_id not in self._target_sessions:
|
||||
return False
|
||||
return len(self._target_sessions[target_id]) > 0
|
||||
|
||||
def get_target_id_from_session_id(self, session_id: SessionID) -> TargetID | None:
|
||||
"""Look up which target a session belongs to.
|
||||
|
||||
Args:
|
||||
session_id: The session ID to look up
|
||||
|
||||
Returns:
|
||||
Target ID if found, None otherwise
|
||||
"""
|
||||
return self._session_to_target.get(session_id)
|
||||
|
||||
def get_target(self, target_id: TargetID) -> 'Target | None':
|
||||
"""Get target from owned data.
|
||||
|
||||
Args:
|
||||
target_id: Target ID to get
|
||||
|
||||
Returns:
|
||||
Target object if found, None otherwise
|
||||
"""
|
||||
return self._targets.get(target_id)
|
||||
|
||||
def get_all_targets(self) -> dict[TargetID, 'Target']:
|
||||
"""Get all targets (read-only access to owned data).
|
||||
|
||||
Returns:
|
||||
Dict mapping target_id to Target objects
|
||||
"""
|
||||
return self._targets
|
||||
|
||||
def get_all_target_ids(self) -> list[TargetID]:
|
||||
"""Get all target IDs from owned data.
|
||||
|
||||
Returns:
|
||||
List of all target IDs
|
||||
"""
|
||||
return list(self._targets.keys())
|
||||
|
||||
def get_all_sessions(self) -> dict[SessionID, 'CDPSession']:
|
||||
"""Get all sessions (read-only access to owned data).
|
||||
|
||||
Returns:
|
||||
Dict mapping session_id to CDPSession objects
|
||||
"""
|
||||
return self._sessions
|
||||
|
||||
def get_session(self, session_id: SessionID) -> 'CDPSession | None':
|
||||
"""Get session from owned data.
|
||||
|
||||
Args:
|
||||
session_id: Session ID to get
|
||||
|
||||
Returns:
|
||||
CDPSession object if found, None otherwise
|
||||
"""
|
||||
return self._sessions.get(session_id)
|
||||
|
||||
def get_all_sessions_for_target(self, target_id: TargetID) -> list['CDPSession']:
|
||||
"""Get ALL sessions attached to a target from owned data.
|
||||
|
||||
Args:
|
||||
target_id: Target ID to get sessions for
|
||||
|
||||
Returns:
|
||||
List of all CDPSession objects for this target
|
||||
"""
|
||||
session_ids = self._target_sessions.get(target_id, set())
|
||||
return [self._sessions[sid] for sid in session_ids if sid in self._sessions]
|
||||
|
||||
def get_target_sessions_mapping(self) -> dict[TargetID, set[SessionID]]:
|
||||
"""Get target->sessions mapping (read-only access).
|
||||
|
||||
Returns:
|
||||
Dict mapping target_id to set of session_ids
|
||||
"""
|
||||
return self._target_sessions
|
||||
|
||||
def get_focused_target(self) -> 'Target | None':
|
||||
"""Get the target that currently has agent focus.
|
||||
|
||||
Convenience method that uses browser_session.agent_focus_target_id.
|
||||
|
||||
Returns:
|
||||
Target object if agent has focus, None otherwise
|
||||
"""
|
||||
if not self.browser_session.agent_focus_target_id:
|
||||
return None
|
||||
return self.get_target(self.browser_session.agent_focus_target_id)
|
||||
|
||||
async def ensure_valid_focus(self, timeout: float = 3.0) -> bool:
|
||||
"""Ensure agent_focus_target_id points to a valid, attached CDP session.
|
||||
|
||||
If the focus target is stale (detached), this method waits for automatic recovery.
|
||||
Uses event-driven coordination instead of polling for efficiency.
|
||||
|
||||
Args:
|
||||
timeout: Maximum time to wait for recovery in seconds (default: 3.0)
|
||||
|
||||
Returns:
|
||||
True if focus is valid or successfully recovered, False if no focus or recovery failed
|
||||
"""
|
||||
if not self.browser_session.agent_focus_target_id:
|
||||
# No focus at all - might be initial state or complete failure
|
||||
if self._recovery_in_progress and self._recovery_complete_event:
|
||||
# Recovery is happening, wait for it
|
||||
try:
|
||||
await asyncio.wait_for(self._recovery_complete_event.wait(), timeout=timeout)
|
||||
# Check again after recovery - simple existence check
|
||||
focus_id = self.browser_session.agent_focus_target_id
|
||||
return bool(focus_id and self._get_session_for_target(focus_id))
|
||||
except TimeoutError:
|
||||
self.logger.error(f'[SessionManager] ❌ Timed out waiting for recovery after {timeout}s')
|
||||
return False
|
||||
return False
|
||||
|
||||
# Simple existence check - does the focused target have a session?
|
||||
cdp_session = self._get_session_for_target(self.browser_session.agent_focus_target_id)
|
||||
if cdp_session:
|
||||
# Session exists - validate it's still active
|
||||
is_valid = await self.validate_session(self.browser_session.agent_focus_target_id)
|
||||
if is_valid:
|
||||
return True
|
||||
|
||||
# Focus is stale - wait for recovery using event instead of polling
|
||||
stale_target_id = self.browser_session.agent_focus_target_id
|
||||
self.logger.warning(
|
||||
f'[SessionManager] ⚠️ Stale agent_focus detected (target {stale_target_id[:8] if stale_target_id else "None"}... detached), '
|
||||
f'waiting for recovery...'
|
||||
)
|
||||
|
||||
# Check if recovery is already in progress
|
||||
if not self._recovery_in_progress:
|
||||
self.logger.warning(
|
||||
'[SessionManager] ⚠️ Recovery not in progress for stale focus! '
|
||||
'This indicates a bug - recovery should have been triggered.'
|
||||
)
|
||||
return False
|
||||
|
||||
# Wait for recovery complete event (event-driven, not polling!)
|
||||
if self._recovery_complete_event:
|
||||
try:
|
||||
start_time = asyncio.get_event_loop().time()
|
||||
await asyncio.wait_for(self._recovery_complete_event.wait(), timeout=timeout)
|
||||
elapsed = asyncio.get_event_loop().time() - start_time
|
||||
|
||||
# Verify recovery succeeded - simple existence check
|
||||
focus_id = self.browser_session.agent_focus_target_id
|
||||
if focus_id and self._get_session_for_target(focus_id):
|
||||
self.logger.info(
|
||||
f'[SessionManager] ✅ Agent focus recovered to {self.browser_session.agent_focus_target_id[:8]}... '
|
||||
f'after {elapsed * 1000:.0f}ms'
|
||||
)
|
||||
return True
|
||||
else:
|
||||
self.logger.error(
|
||||
f'[SessionManager] ❌ Recovery completed but focus still invalid after {elapsed * 1000:.0f}ms'
|
||||
)
|
||||
return False
|
||||
|
||||
except TimeoutError:
|
||||
self.logger.error(
|
||||
f'[SessionManager] ❌ Recovery timed out after {timeout}s '
|
||||
f'(was: {stale_target_id[:8] if stale_target_id else "None"}..., '
|
||||
f'now: {self.browser_session.agent_focus_target_id[:8] if self.browser_session.agent_focus_target_id else "None"})'
|
||||
)
|
||||
return False
|
||||
else:
|
||||
self.logger.error('[SessionManager] ❌ Recovery event not initialized')
|
||||
return False
|
||||
|
||||
async def _handle_target_attached(self, event: AttachedToTargetEvent) -> None:
|
||||
"""Handle Target.attachedToTarget event.
|
||||
|
||||
Called automatically by Chrome when a new target/session is created.
|
||||
This is the ONLY place where sessions are added to the pool.
|
||||
"""
|
||||
target_id = event['targetInfo']['targetId']
|
||||
session_id = event['sessionId']
|
||||
target_type = event['targetInfo']['type']
|
||||
target_info = event['targetInfo']
|
||||
waiting_for_debugger = event.get('waitingForDebugger', False)
|
||||
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Target attached: {target_id[:8]}... (session={session_id[:8]}..., '
|
||||
f'type={target_type}, waitingForDebugger={waiting_for_debugger})'
|
||||
)
|
||||
|
||||
# Defensive check: browser may be shutting down and _cdp_client_root could be None
|
||||
if self.browser_session._cdp_client_root is None:
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Skipping target attach for {target_id[:8]}... - browser shutting down (no CDP client)'
|
||||
)
|
||||
return
|
||||
|
||||
# Enable auto-attach for this session's children (do this FIRST, outside lock)
|
||||
try:
|
||||
await self.browser_session._cdp_client_root.send.Target.setAutoAttach(
|
||||
params={'autoAttach': True, 'waitForDebuggerOnStart': False, 'flatten': True}, session_id=session_id
|
||||
)
|
||||
except Exception as e:
|
||||
error_str = str(e)
|
||||
# Expected for short-lived targets (workers, temp iframes) that detach before this executes
|
||||
if '-32001' not in error_str and 'Session with given id not found' not in error_str:
|
||||
self.logger.debug(f'[SessionManager] Auto-attach failed for {target_type}: {e}')
|
||||
|
||||
from browser_use.browser.session import Target
|
||||
|
||||
async with self._lock:
|
||||
# Track this session for the target
|
||||
if target_id not in self._target_sessions:
|
||||
self._target_sessions[target_id] = set()
|
||||
|
||||
self._target_sessions[target_id].add(session_id)
|
||||
self._session_to_target[session_id] = target_id
|
||||
|
||||
# Create or update Target inside the same lock so that get_target() is never
|
||||
# called in the window between _target_sessions being set and _targets being set.
|
||||
if target_id not in self._targets:
|
||||
target = Target(
|
||||
target_id=target_id,
|
||||
target_type=target_type,
|
||||
url=target_info.get('url', 'about:blank'),
|
||||
title=target_info.get('title', 'Unknown title'),
|
||||
)
|
||||
self._targets[target_id] = target
|
||||
self.logger.debug(f'[SessionManager] Created target {target_id[:8]}... (type={target_type})')
|
||||
else:
|
||||
# Update existing target info
|
||||
existing_target = self._targets[target_id]
|
||||
existing_target.url = target_info.get('url', existing_target.url)
|
||||
existing_target.title = target_info.get('title', existing_target.title)
|
||||
|
||||
# Create CDPSession (communication channel)
|
||||
from browser_use.browser.session import CDPSession
|
||||
|
||||
assert self.browser_session._cdp_client_root is not None, 'Root CDP client required'
|
||||
|
||||
cdp_session = CDPSession(
|
||||
cdp_client=self.browser_session._cdp_client_root,
|
||||
target_id=target_id,
|
||||
session_id=session_id,
|
||||
)
|
||||
|
||||
# Add to sessions dict
|
||||
self._sessions[session_id] = cdp_session
|
||||
|
||||
# If proxy auth is configured, enable Fetch auth handling on this session
|
||||
# Avoids overwriting Target.attachedToTarget handlers elsewhere
|
||||
try:
|
||||
proxy_cfg = self.browser_session.browser_profile.proxy
|
||||
username = proxy_cfg.username if proxy_cfg else None
|
||||
password = proxy_cfg.password if proxy_cfg else None
|
||||
if username and password:
|
||||
await cdp_session.cdp_client.send.Fetch.enable(
|
||||
params={'handleAuthRequests': True},
|
||||
session_id=cdp_session.session_id,
|
||||
)
|
||||
self.logger.debug(f'[SessionManager] Fetch.enable(handleAuthRequests=True) on session {session_id[:8]}...')
|
||||
except Exception as e:
|
||||
self.logger.debug(f'[SessionManager] Fetch.enable on attached session failed: {type(e).__name__}: {e}')
|
||||
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Created session {session_id[:8]}... for target {target_id[:8]}... '
|
||||
f'(total sessions: {len(self._sessions)})'
|
||||
)
|
||||
|
||||
# Enable lifecycle events and network monitoring for page targets
|
||||
if target_type in ('page', 'tab'):
|
||||
await self._enable_page_monitoring(cdp_session)
|
||||
|
||||
# Resume execution if waiting for debugger
|
||||
if waiting_for_debugger:
|
||||
try:
|
||||
assert self.browser_session._cdp_client_root is not None
|
||||
await self.browser_session._cdp_client_root.send.Runtime.runIfWaitingForDebugger(session_id=session_id)
|
||||
except Exception as e:
|
||||
self.logger.warning(f'[SessionManager] Failed to resume execution: {e}')
|
||||
|
||||
async def _handle_target_info_changed(self, event: dict) -> None:
|
||||
"""Handle Target.targetInfoChanged event.
|
||||
|
||||
Updates target title/URL without polling getTargetInfo().
|
||||
Chrome fires this automatically when title or URL changes.
|
||||
"""
|
||||
target_info = event.get('targetInfo', {})
|
||||
target_id = target_info.get('targetId')
|
||||
|
||||
if not target_id:
|
||||
return
|
||||
|
||||
async with self._lock:
|
||||
# Update target if it exists (source of truth for url/title)
|
||||
if target_id in self._targets:
|
||||
target = self._targets[target_id]
|
||||
|
||||
target.title = target_info.get('title', target.title)
|
||||
target.url = target_info.get('url', target.url)
|
||||
|
||||
async def _handle_target_detached(self, event: DetachedFromTargetEvent) -> None:
|
||||
"""Handle Target.detachedFromTarget event.
|
||||
|
||||
Called automatically by Chrome when a target/session is destroyed.
|
||||
This is the ONLY place where sessions are removed from the pool.
|
||||
"""
|
||||
session_id = event['sessionId']
|
||||
target_id = event.get('targetId') # May be empty
|
||||
|
||||
# If targetId not in event, look it up via session mapping
|
||||
if not target_id:
|
||||
async with self._lock:
|
||||
target_id = self._session_to_target.get(session_id)
|
||||
|
||||
if not target_id:
|
||||
self.logger.warning(f'[SessionManager] Session detached but target unknown (session={session_id[:8]}...)')
|
||||
return
|
||||
|
||||
agent_focus_lost = False
|
||||
target_fully_removed = False
|
||||
target_type = None
|
||||
|
||||
async with self._lock:
|
||||
# Remove this session from target's session set
|
||||
if target_id in self._target_sessions:
|
||||
self._target_sessions[target_id].discard(session_id)
|
||||
|
||||
remaining_sessions = len(self._target_sessions[target_id])
|
||||
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Session detached: target={target_id[:8]}... '
|
||||
f'session={session_id[:8]}... (remaining={remaining_sessions})'
|
||||
)
|
||||
|
||||
# Only remove target when NO sessions remain
|
||||
if remaining_sessions == 0:
|
||||
self.logger.debug(f'[SessionManager] No sessions remain for target {target_id[:8]}..., removing target')
|
||||
|
||||
target_fully_removed = True
|
||||
|
||||
# Check if agent_focus points to this target
|
||||
agent_focus_lost = self.browser_session.agent_focus_target_id == target_id
|
||||
|
||||
# Immediately clear stale focus to prevent operations on detached target
|
||||
if agent_focus_lost:
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Clearing stale agent_focus_target_id {target_id[:8]}... '
|
||||
f'to prevent operations on detached target'
|
||||
)
|
||||
self.browser_session.agent_focus_target_id = None
|
||||
|
||||
# Get target type before removing (needed for TabClosedEvent dispatch)
|
||||
target = self._targets.get(target_id)
|
||||
target_type = target.target_type if target else None
|
||||
|
||||
# Remove target (entity) from owned data
|
||||
if target_id in self._targets:
|
||||
self._targets.pop(target_id)
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Removed target {target_id[:8]}... (remaining targets: {len(self._targets)})'
|
||||
)
|
||||
|
||||
# Clean up tracking
|
||||
del self._target_sessions[target_id]
|
||||
self._lifecycle_events.pop(target_id, None)
|
||||
else:
|
||||
# Target not tracked - already removed or never attached
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Session detached from untracked target: target={target_id[:8]}... '
|
||||
f'session={session_id[:8]}... (target was already removed or attach event was missed)'
|
||||
)
|
||||
|
||||
# Remove session from owned sessions dict
|
||||
if session_id in self._sessions:
|
||||
self._sessions.pop(session_id)
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Removed session {session_id[:8]}... (remaining sessions: {len(self._sessions)})'
|
||||
)
|
||||
|
||||
# Remove from reverse mapping
|
||||
if session_id in self._session_to_target:
|
||||
del self._session_to_target[session_id]
|
||||
|
||||
# Dispatch TabClosedEvent only for page/tab targets that are fully removed (not iframes/workers or partial detaches)
|
||||
if target_fully_removed:
|
||||
if target_type in ('page', 'tab'):
|
||||
from browser_use.browser.events import TabClosedEvent
|
||||
|
||||
self.browser_session.event_bus.dispatch(TabClosedEvent(target_id=target_id))
|
||||
self.logger.debug(f'[SessionManager] Dispatched TabClosedEvent for page target {target_id[:8]}...')
|
||||
elif target_type:
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Target {target_id[:8]}... fully removed (type={target_type}) - not dispatching TabClosedEvent'
|
||||
)
|
||||
|
||||
# Auto-recover agent_focus outside the lock to avoid blocking other operations
|
||||
if agent_focus_lost:
|
||||
# Create recovery task instead of awaiting directly - allows concurrent operations to wait on same recovery
|
||||
if not self._recovery_in_progress:
|
||||
self._recovery_task = create_task_with_error_handling(
|
||||
self._recover_agent_focus(target_id),
|
||||
name='recover_agent_focus',
|
||||
logger_instance=self.logger,
|
||||
suppress_exceptions=False,
|
||||
)
|
||||
|
||||
async def _recover_agent_focus(self, crashed_target_id: TargetID) -> None:
|
||||
"""Auto-recover agent_focus when the focused target crashes/detaches.
|
||||
|
||||
Uses recovery lock to prevent concurrent recovery attempts from creating multiple emergency tabs.
|
||||
Coordinates with ensure_valid_focus() via events for efficient waiting.
|
||||
|
||||
Args:
|
||||
crashed_target_id: The target ID that was lost
|
||||
"""
|
||||
try:
|
||||
# Prevent concurrent recovery attempts
|
||||
async with self._recovery_lock:
|
||||
# Set recovery state INSIDE lock to prevent race conditions
|
||||
if self._recovery_in_progress:
|
||||
self.logger.debug('[SessionManager] Recovery already in progress, waiting for it to complete')
|
||||
# Wait for ongoing recovery instead of starting a new one
|
||||
if self._recovery_complete_event:
|
||||
try:
|
||||
await asyncio.wait_for(self._recovery_complete_event.wait(), timeout=5.0)
|
||||
except TimeoutError:
|
||||
self.logger.error('[SessionManager] Timed out waiting for ongoing recovery')
|
||||
return
|
||||
|
||||
# Set recovery state
|
||||
self._recovery_in_progress = True
|
||||
self._recovery_complete_event = asyncio.Event()
|
||||
|
||||
if self.browser_session._cdp_client_root is None:
|
||||
self.logger.debug('[SessionManager] Skipping focus recovery - browser shutting down (no CDP client)')
|
||||
return
|
||||
|
||||
# Check if another recovery already fixed agent_focus
|
||||
if self.browser_session.agent_focus_target_id and self.browser_session.agent_focus_target_id != crashed_target_id:
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Agent focus already recovered by concurrent operation '
|
||||
f'(now: {self.browser_session.agent_focus_target_id[:8]}...), skipping recovery'
|
||||
)
|
||||
return
|
||||
|
||||
# Note: agent_focus_target_id may already be None (cleared in _handle_target_detached)
|
||||
current_focus_desc = (
|
||||
f'{self.browser_session.agent_focus_target_id[:8]}...'
|
||||
if self.browser_session.agent_focus_target_id
|
||||
else 'None (already cleared)'
|
||||
)
|
||||
|
||||
self.logger.warning(
|
||||
f'[SessionManager] Agent focus target {crashed_target_id[:8]}... detached! '
|
||||
f'Current focus: {current_focus_desc}. Auto-recovering by switching to another target...'
|
||||
)
|
||||
|
||||
# Perform recovery (outside lock to allow concurrent operations)
|
||||
# Try to find another valid page target
|
||||
page_targets = self.get_all_page_targets()
|
||||
|
||||
new_target_id = None
|
||||
is_existing_tab = False
|
||||
|
||||
if page_targets:
|
||||
# Switch to most recent page that's not the crashed one
|
||||
new_target_id = page_targets[-1].target_id
|
||||
is_existing_tab = True
|
||||
self.logger.info(f'[SessionManager] Switching agent_focus to existing tab {new_target_id[:8]}...')
|
||||
else:
|
||||
# No pages exist - create a new one
|
||||
self.logger.warning('[SessionManager] No tabs remain! Creating new tab for agent...')
|
||||
new_target_id = await self.browser_session._cdp_create_new_page('about:blank')
|
||||
self.logger.info(f'[SessionManager] Created new tab {new_target_id[:8]}... for agent')
|
||||
|
||||
# Dispatch TabCreatedEvent so watchdogs can initialize
|
||||
from browser_use.browser.events import TabCreatedEvent
|
||||
|
||||
self.browser_session.event_bus.dispatch(TabCreatedEvent(url='about:blank', target_id=new_target_id))
|
||||
|
||||
# Wait for CDP attach event to create session
|
||||
# Note: This polling is necessary - waiting for external Chrome CDP event
|
||||
# _handle_target_attached will add session to pool when Chrome fires attachedToTarget
|
||||
new_session = None
|
||||
for attempt in range(20): # Wait up to 2 seconds
|
||||
await asyncio.sleep(0.1)
|
||||
new_session = self._get_session_for_target(new_target_id)
|
||||
if new_session:
|
||||
break
|
||||
|
||||
if new_session:
|
||||
self.browser_session.agent_focus_target_id = new_target_id
|
||||
self.logger.info(f'[SessionManager] ✅ Agent focus recovered: {new_target_id[:8]}...')
|
||||
|
||||
# Visually activate the tab in browser (only for existing tabs)
|
||||
if is_existing_tab:
|
||||
try:
|
||||
assert self.browser_session._cdp_client_root is not None
|
||||
await self.browser_session._cdp_client_root.send.Target.activateTarget(params={'targetId': new_target_id})
|
||||
self.logger.debug(f'[SessionManager] Activated tab {new_target_id[:8]}... in browser UI')
|
||||
except Exception as e:
|
||||
self.logger.debug(f'[SessionManager] Failed to activate tab visually: {e}')
|
||||
|
||||
# Get target to access url (from owned data)
|
||||
target = self.get_target(new_target_id)
|
||||
target_url = target.url if target else 'about:blank'
|
||||
|
||||
# Dispatch focus changed event
|
||||
from browser_use.browser.events import AgentFocusChangedEvent
|
||||
|
||||
self.browser_session.event_bus.dispatch(AgentFocusChangedEvent(target_id=new_target_id, url=target_url))
|
||||
return
|
||||
|
||||
# Recovery failed - create emergency fallback tab
|
||||
self.logger.error(
|
||||
f'[SessionManager] ❌ Failed to get session for {new_target_id[:8]}... after 2s, creating emergency fallback tab'
|
||||
)
|
||||
|
||||
fallback_target_id = await self.browser_session._cdp_create_new_page('about:blank')
|
||||
self.logger.warning(f'[SessionManager] Created emergency fallback tab {fallback_target_id[:8]}...')
|
||||
|
||||
# Try one more time with fallback
|
||||
# Note: This polling is necessary - waiting for external Chrome CDP event
|
||||
for _ in range(20):
|
||||
await asyncio.sleep(0.1)
|
||||
fallback_session = self._get_session_for_target(fallback_target_id)
|
||||
if fallback_session:
|
||||
self.browser_session.agent_focus_target_id = fallback_target_id
|
||||
self.logger.warning(f'[SessionManager] ⚠️ Agent focus set to emergency fallback: {fallback_target_id[:8]}...')
|
||||
|
||||
from browser_use.browser.events import AgentFocusChangedEvent, TabCreatedEvent
|
||||
|
||||
self.browser_session.event_bus.dispatch(TabCreatedEvent(url='about:blank', target_id=fallback_target_id))
|
||||
self.browser_session.event_bus.dispatch(
|
||||
AgentFocusChangedEvent(target_id=fallback_target_id, url='about:blank')
|
||||
)
|
||||
return
|
||||
|
||||
# Complete failure - this should never happen
|
||||
self.logger.critical(
|
||||
'[SessionManager] 🚨 CRITICAL: Failed to recover agent_focus even with fallback! Agent may be in broken state.'
|
||||
)
|
||||
|
||||
except Exception as e:
|
||||
self.logger.error(f'[SessionManager] ❌ Error during agent_focus recovery: {type(e).__name__}: {e}')
|
||||
finally:
|
||||
# Always signal completion and reset recovery state
|
||||
# This allows all waiting operations to proceed (success or failure)
|
||||
if self._recovery_complete_event:
|
||||
self._recovery_complete_event.set()
|
||||
self._recovery_in_progress = False
|
||||
self._recovery_task = None
|
||||
self.logger.debug('[SessionManager] Recovery state reset')
|
||||
|
||||
async def _initialize_existing_targets(self) -> None:
|
||||
"""Discover and initialize all existing targets at startup.
|
||||
|
||||
Attaches to each target and initializes it SYNCHRONOUSLY.
|
||||
Chrome will also fire attachedToTarget events, but _handle_target_attached() is
|
||||
idempotent (checks if target already in pool), so duplicate handling is safe.
|
||||
|
||||
This eliminates race conditions - monitoring is guaranteed ready before navigation.
|
||||
"""
|
||||
cdp_client = self.browser_session._cdp_client_root
|
||||
assert cdp_client is not None
|
||||
|
||||
# Get all existing targets
|
||||
targets_result = await cdp_client.send.Target.getTargets()
|
||||
existing_targets = targets_result.get('targetInfos', [])
|
||||
|
||||
self.logger.debug(f'[SessionManager] Discovered {len(existing_targets)} existing targets')
|
||||
|
||||
# Track target IDs for verification
|
||||
target_ids_to_wait_for = []
|
||||
|
||||
# Just attach to ALL existing targets - Chrome fires attachedToTarget events
|
||||
# The on_attached handler (via create_task) does ALL the work
|
||||
for target in existing_targets:
|
||||
target_id = target['targetId']
|
||||
target_type = target.get('type', 'unknown')
|
||||
|
||||
try:
|
||||
# Just attach - event handler does everything
|
||||
await cdp_client.send.Target.attachToTarget(params={'targetId': target_id, 'flatten': True})
|
||||
target_ids_to_wait_for.append(target_id)
|
||||
except Exception as e:
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Failed to attach to existing target {target_id[:8]}... (type={target_type}): {e}'
|
||||
)
|
||||
|
||||
# Wait for event handlers to complete their work (they run via create_task)
|
||||
# Use event-driven approach instead of polling for better performance
|
||||
ready_event = asyncio.Event()
|
||||
|
||||
async def check_all_ready():
|
||||
"""Check if all sessions are ready and signal completion."""
|
||||
while True:
|
||||
ready_count = 0
|
||||
for tid in target_ids_to_wait_for:
|
||||
session = self._get_session_for_target(tid)
|
||||
if session:
|
||||
target = self._targets.get(tid)
|
||||
target_type = target.target_type if target else 'unknown'
|
||||
# For pages, verify monitoring is enabled
|
||||
if target_type in ('page', 'tab'):
|
||||
if hasattr(session, '_lifecycle_events') and session._lifecycle_events is not None:
|
||||
ready_count += 1
|
||||
else:
|
||||
# Non-page targets don't need monitoring
|
||||
ready_count += 1
|
||||
|
||||
if ready_count == len(target_ids_to_wait_for):
|
||||
ready_event.set()
|
||||
return
|
||||
|
||||
await asyncio.sleep(0.05)
|
||||
|
||||
# Start checking in background
|
||||
check_task = create_task_with_error_handling(
|
||||
check_all_ready(), name='check_all_targets_ready', logger_instance=self.logger
|
||||
)
|
||||
|
||||
try:
|
||||
# Wait for completion with timeout
|
||||
await asyncio.wait_for(ready_event.wait(), timeout=2.0)
|
||||
except TimeoutError:
|
||||
# Timeout - count what's ready
|
||||
ready_count = 0
|
||||
for tid in target_ids_to_wait_for:
|
||||
session = self._get_session_for_target(tid)
|
||||
if session:
|
||||
target = self._targets.get(tid)
|
||||
target_type = target.target_type if target else 'unknown'
|
||||
# For pages, verify monitoring is enabled
|
||||
if target_type in ('page', 'tab'):
|
||||
if hasattr(session, '_lifecycle_events') and session._lifecycle_events is not None:
|
||||
ready_count += 1
|
||||
else:
|
||||
# Non-page targets don't need monitoring
|
||||
ready_count += 1
|
||||
self.logger.warning(
|
||||
f'[SessionManager] Initialization timeout after 2.0s: {ready_count}/{len(target_ids_to_wait_for)} sessions ready'
|
||||
)
|
||||
finally:
|
||||
check_task.cancel()
|
||||
try:
|
||||
await check_task
|
||||
except asyncio.CancelledError:
|
||||
pass
|
||||
|
||||
async def _enable_page_monitoring(self, cdp_session: 'CDPSession') -> None:
|
||||
"""Enable lifecycle events and network monitoring for a page target.
|
||||
|
||||
This is called once per page when it's created, avoiding handler accumulation.
|
||||
Registers a SINGLE lifecycle handler per session that stores events for navigations to consume.
|
||||
|
||||
Args:
|
||||
cdp_session: The CDP session to enable monitoring on
|
||||
"""
|
||||
try:
|
||||
# Enable Page domain first (required for lifecycle events)
|
||||
await cdp_session.cdp_client.send.Page.enable(session_id=cdp_session.session_id)
|
||||
|
||||
# Enable lifecycle events (load, DOMContentLoaded, networkIdle, etc.)
|
||||
await cdp_session.cdp_client.send.Page.setLifecycleEventsEnabled(
|
||||
params={'enabled': True}, session_id=cdp_session.session_id
|
||||
)
|
||||
|
||||
# Enable network monitoring for networkIdle detection
|
||||
await cdp_session.cdp_client.send.Network.enable(session_id=cdp_session.session_id)
|
||||
|
||||
# Event storage and the Page.lifecycleEvent handler live in SessionManager
|
||||
# (one global handler registered in start_monitoring, routed by session_id):
|
||||
# cdp-use's registry is single-slot per method, so a per-session registration
|
||||
# here would replace the previous tab's handler and freeze its event buffer.
|
||||
# Expose the shared per-target buffer on the session for readiness checks.
|
||||
cdp_session._lifecycle_events = self.get_lifecycle_events(cdp_session.target_id)
|
||||
|
||||
except Exception as e:
|
||||
# Don't fail - target might be short-lived or already detached
|
||||
error_str = str(e)
|
||||
if '-32001' in error_str or 'Session with given id not found' in error_str:
|
||||
self.logger.debug(
|
||||
f'[SessionManager] Target {cdp_session.target_id[:8]}... detached before monitoring could be enabled (normal for short-lived targets)'
|
||||
)
|
||||
else:
|
||||
self.logger.warning(
|
||||
f'[SessionManager] Failed to enable monitoring for target {cdp_session.target_id[:8]}...: {e}'
|
||||
)
|
||||
Reference in New Issue
Block a user