Skip to content

Open-source Python library for AI browser agents, plus Browser Use Cloud: stealth browsers and a hosted agent API.

increased threshold for loop detection

Saurav Pandacec4aea3bc05
diff --git a/browser_use/agent/service.py b/browser_use/agent/service.py
index 73d0a4d4a..170cc30e7 100644
--- a/browser_use/agent/service.py
+++ b/browser_use/agent/service.py
@@ -199,7 +199,7 @@ class Agent(Generic[Context, AgentStructuredOutput]):
enable_planning: bool = True,
planning_replan_on_stall: int = 3,
planning_exploration_limit: int = 5,
- loop_detection_window: int = 10,
+ loop_detection_window: int = 20,
loop_detection_enabled: bool = True,
llm_screenshot_size: tuple[int, int] | None = None,
_url_shortening_limit: int = 25,
@@ -1367,9 +1367,14 @@ class Agent(Generic[Context, AgentStructuredOutput]):
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 = {}
diff --git a/browser_use/agent/views.py b/browser_use/agent/views.py
index 4b3967f60..ece2ce6e9 100644
--- a/browser_use/agent/views.py
+++ b/browser_use/agent/views.py
@@ -61,7 +61,7 @@ class AgentSettings(BaseModel):
final_response_after_failure: bool = True # If True, attempt one final recovery call after max_failures
# Loop detection settings
- loop_detection_window: int = 10 # Rolling window size for action similarity tracking
+ loop_detection_window: int = 20 # Rolling window size for action similarity tracking
loop_detection_enabled: bool = True # Whether to enable loop detection nudges
@@ -95,15 +95,15 @@ def _normalize_action_for_hash(action_name: str, params: dict[str, Any]) -> str:
engine = params.get('engine', 'google')
return f'search|{engine}|{"|".join(tokens)}'
- if action_name in ('click_element', 'input_text'):
+ if action_name in ('click', 'input'):
# For element-interaction actions, we only use the index (element identity).
# Two clicks on the same element index are the same action.
index = params.get('index')
- if action_name == 'input_text':
+ if action_name == 'input':
text = str(params.get('text', ''))
# Normalize input text: lowercase, strip whitespace
- return f'input_text|{index}|{text.strip().lower()}'
- return f'click_element|{index}'
+ return f'input|{index}|{text.strip().lower()}'
+ return f'click|{index}'
if action_name == 'navigate':
url = str(params.get('url', ''))
@@ -113,7 +113,8 @@ def _normalize_action_for_hash(action_name: str, params: dict[str, Any]) -> str:
if action_name == 'scroll':
direction = 'down' if params.get('down', True) else 'up'
- return f'scroll|{direction}'
+ index = params.get('index')
+ return f'scroll|{direction}|{index}'
# Default: hash by action name + sorted params (excluding None values)
filtered = {k: v for k, v in sorted(params.items()) if v is not None}
@@ -136,7 +137,7 @@ class ActionLoopDetector(BaseModel):
model_config = ConfigDict(arbitrary_types_allowed=True)
# Rolling window of recent action hashes
- window_size: int = 10
+ window_size: int = 20
recent_action_hashes: list[str] = Field(default_factory=list)
# Page fingerprint tracking for stagnation detection
@@ -184,30 +185,30 @@ class ActionLoopDetector(BaseModel):
"""Return an escalating nudge message based on repetition severity, or None if no loop detected."""
messages: list[str] = []
- # Action repetition nudges (escalating at 3, 5, 8)
- if self.max_repetition_count >= 8:
+ # Action repetition nudges (escalating at 5, 8, 12)
+ if self.max_repetition_count >= 12:
messages.append(
f'CRITICAL LOOP DETECTED: You have repeated a similar action {self.max_repetition_count} times '
f'in the last {len(self.recent_action_hashes)} actions. '
- 'Your remaining action budget is low. Extract what you can and report results using the done action. '
+ 'Try a fundamentally different approach, or if the task is partially complete, report your progress. '
'Continuing to repeat the same action will not yield different results.'
)
- elif self.max_repetition_count >= 5:
+ elif self.max_repetition_count >= 8:
messages.append(
f'STRONG LOOP WARNING: You have repeated a similar action {self.max_repetition_count} times '
f'in the last {len(self.recent_action_hashes)} actions. '
'Strongly consider abandoning this approach. Try a fundamentally different strategy '
'or report that the task cannot be completed with the current approach.'
)
- elif self.max_repetition_count >= 3:
+ elif self.max_repetition_count >= 5:
messages.append(
f'LOOP DETECTED: You have repeated a similar action {self.max_repetition_count} times '
f'in the last {len(self.recent_action_hashes)} actions without meaningful progress. '
- 'Try a fundamentally different approach or report that the task cannot be completed on this site.'
+ 'Consider trying a different approach.'
)
# Page stagnation nudge
- if self.consecutive_stagnant_pages >= 3:
+ if self.consecutive_stagnant_pages >= 5:
messages.append(
f'PAGE STAGNATION: The page content has not changed across {self.consecutive_stagnant_pages} '
'consecutive actions. Your actions are not affecting the page state. '
diff --git a/tests/ci/test_action_loop_detection.py b/tests/ci/test_action_loop_detection.py
index f9a0fe3f3..fe07d2c02 100644
--- a/tests/ci/test_action_loop_detection.py
+++ b/tests/ci/test_action_loop_detection.py
@@ -54,31 +54,31 @@ def test_search_different_queries_produce_different_hashes():
assert h1 != h2
-def test_click_element_same_index_same_hash():
+def test_click_same_index_same_hash():
"""Clicking the same element index produces the same hash."""
- h1 = compute_action_hash('click_element', {'index': 5})
- h2 = compute_action_hash('click_element', {'index': 5})
+ h1 = compute_action_hash('click', {'index': 5})
+ h2 = compute_action_hash('click', {'index': 5})
assert h1 == h2
-def test_click_element_different_index_different_hash():
+def test_click_different_index_different_hash():
"""Clicking different element indices produces different hashes."""
- h1 = compute_action_hash('click_element', {'index': 5})
- h2 = compute_action_hash('click_element', {'index': 12})
+ h1 = compute_action_hash('click', {'index': 5})
+ h2 = compute_action_hash('click', {'index': 12})
assert h1 != h2
-def test_input_text_same_element_same_text():
+def test_input_same_element_same_text():
"""Same element + same text = same hash."""
- h1 = compute_action_hash('input_text', {'index': 3, 'text': 'hello world', 'clear': True})
- h2 = compute_action_hash('input_text', {'index': 3, 'text': 'hello world', 'clear': False})
+ h1 = compute_action_hash('input', {'index': 3, 'text': 'hello world', 'clear': True})
+ h2 = compute_action_hash('input', {'index': 3, 'text': 'hello world', 'clear': False})
assert h1 == h2 # clear flag doesn't affect hash
-def test_input_text_different_text_different_hash():
+def test_input_different_text_different_hash():
"""Same element but different text = different hash."""
- h1 = compute_action_hash('input_text', {'index': 3, 'text': 'hello'})
- h2 = compute_action_hash('input_text', {'index': 3, 'text': 'goodbye'})
+ h1 = compute_action_hash('input', {'index': 3, 'text': 'hello'})
+ h2 = compute_action_hash('input', {'index': 3, 'text': 'goodbye'})
assert h1 != h2
@@ -105,15 +105,29 @@ def test_navigate_different_domain_different_hash():
def test_scroll_direction_matters():
"""Scroll up and scroll down are different actions."""
- h1 = compute_action_hash('scroll', {'down': True})
- h2 = compute_action_hash('scroll', {'down': False})
+ h1 = compute_action_hash('scroll', {'down': True, 'index': None})
+ h2 = compute_action_hash('scroll', {'down': False, 'index': None})
assert h1 != h2
+def test_scroll_different_elements_different_hash():
+ """Scrolling different elements produces different hashes."""
+ h1 = compute_action_hash('scroll', {'down': True, 'index': 5})
+ h2 = compute_action_hash('scroll', {'down': True, 'index': 10})
+ assert h1 != h2
+
+
+def test_scroll_same_element_same_hash():
+ """Scrolling the same element in the same direction produces the same hash."""
+ h1 = compute_action_hash('scroll', {'down': True, 'index': 5})
+ h2 = compute_action_hash('scroll', {'down': True, 'index': 5})
+ assert h1 == h2
+
+
def test_different_action_types_different_hashes():
"""Different action types always produce different hashes."""
- h1 = compute_action_hash('click_element', {'index': 5})
- h2 = compute_action_hash('scroll', {'down': True})
+ h1 = compute_action_hash('click', {'index': 5})
+ h2 = compute_action_hash('scroll', {'down': True, 'index': None})
h3 = compute_action_hash('search', {'query': 'test'})
assert len({h1, h2, h3}) == 3
@@ -123,76 +137,108 @@ def test_different_action_types_different_hashes():
def test_detector_no_nudge_for_diverse_actions():
"""No nudge when actions are all different."""
- detector = ActionLoopDetector(window_size=10)
- detector.record_action('click_element', {'index': 1})
- detector.record_action('scroll', {'down': True})
- detector.record_action('click_element', {'index': 2})
+ detector = ActionLoopDetector(window_size=20)
+ detector.record_action('click', {'index': 1})
+ detector.record_action('scroll', {'down': True, 'index': None})
+ detector.record_action('click', {'index': 2})
detector.record_action('search', {'query': 'something'})
detector.record_action('navigate', {'url': 'https://example.com'})
assert detector.get_nudge_message() is None
-def test_detector_nudge_at_3_repeats():
- """Nudge triggers at 3 repetitions of the same action."""
- detector = ActionLoopDetector(window_size=10)
- for _ in range(3):
+def test_detector_nudge_at_5_repeats():
+ """Nudge triggers at 5 repetitions of the same action."""
+ detector = ActionLoopDetector(window_size=20)
+ for _ in range(5):
detector.record_action('search', {'query': 'site:hinative.com answers votes'})
msg = detector.get_nudge_message()
assert msg is not None
assert 'LOOP DETECTED' in msg
- assert '3 times' in msg
+ assert '5 times' in msg
-def test_detector_nudge_escalates_at_5_repeats():
- """Stronger nudge at 5 repetitions."""
- detector = ActionLoopDetector(window_size=10)
- for _ in range(5):
+def test_detector_no_nudge_at_4_repeats():
+ """No nudge at only 4 repetitions (below threshold)."""
+ detector = ActionLoopDetector(window_size=20)
+ for _ in range(4):
+ detector.record_action('search', {'query': 'site:hinative.com answers votes'})
+ assert detector.get_nudge_message() is None
+
+
+def test_detector_nudge_escalates_at_8_repeats():
+ """Stronger nudge at 8 repetitions."""
+ detector = ActionLoopDetector(window_size=20)
+ for _ in range(8):
detector.record_action('search', {'query': 'site:hinative.com answers votes'})
msg = detector.get_nudge_message()
assert msg is not None
assert 'STRONG LOOP WARNING' in msg
- assert '5 times' in msg
+ assert '8 times' in msg
-def test_detector_nudge_escalates_at_8_repeats():
- """Critical nudge at 8 repetitions."""
- detector = ActionLoopDetector(window_size=10)
- for _ in range(8):
+def test_detector_nudge_escalates_at_12_repeats():
+ """Critical nudge at 12 repetitions."""
+ detector = ActionLoopDetector(window_size=20)
+ for _ in range(12):
detector.record_action('search', {'query': 'site:hinative.com answers votes'})
msg = detector.get_nudge_message()
assert msg is not None
assert 'CRITICAL LOOP DETECTED' in msg
- assert '8 times' in msg
+ assert '12 times' in msg
+
+
+def test_detector_critical_message_no_done_directive():
+ """Critical nudge should NOT tell the agent to call done — just suggest a different approach."""
+ detector = ActionLoopDetector(window_size=20)
+ for _ in range(12):
+ detector.record_action('search', {'query': 'site:hinative.com answers votes'})
+ msg = detector.get_nudge_message()
+ assert msg is not None
+ assert 'done action' not in msg
+ assert 'report your progress' in msg
+
+
+def test_detector_first_nudge_no_cannot_complete():
+ """First nudge should NOT say task 'cannot be completed on this site'."""
+ detector = ActionLoopDetector(window_size=20)
+ for _ in range(5):
+ detector.record_action('search', {'query': 'site:hinative.com answers votes'})
+ msg = detector.get_nudge_message()
+ assert msg is not None
+ assert 'cannot be completed' not in msg
+ assert 'Consider trying a different approach' in msg
def test_detector_window_slides():
"""Old actions fall out of the window."""
- detector = ActionLoopDetector(window_size=5)
+ detector = ActionLoopDetector(window_size=10)
# Fill window with repeated actions
- for _ in range(3):
- detector.record_action('click_element', {'index': 7})
- assert detector.max_repetition_count == 3
+ for _ in range(5):
+ detector.record_action('click', {'index': 7})
+ assert detector.max_repetition_count == 5
# Push them out with diverse actions
- for i in range(5):
- detector.record_action('click_element', {'index': 100 + i})
- # The 3 old repeated actions should have been pushed out
- assert detector.max_repetition_count < 3
+ for i in range(10):
+ detector.record_action('click', {'index': 100 + i})
+ # The 5 old repeated actions should have been pushed out
+ assert detector.max_repetition_count < 5
assert detector.get_nudge_message() is None
def test_detector_search_variations_detected_as_same():
"""Minor variations of the same search (the hinative pattern) are detected as repetition."""
- detector = ActionLoopDetector(window_size=10)
+ detector = ActionLoopDetector(window_size=20)
# These are the kind of variations the agent produces
queries = [
'site:hinative.com answers votes questions',
'site:hinative.com questions answers votes',
'site:hinative.com votes answers questions',
+ 'site:hinative.com questions votes answers',
+ 'site:hinative.com answers questions votes',
]
for q in queries:
detector.record_action('search', {'query': q})
- assert detector.max_repetition_count == 3
+ assert detector.max_repetition_count == 5
assert detector.get_nudge_message() is not None
@@ -201,7 +247,7 @@ def test_detector_search_variations_detected_as_same():
def test_page_stagnation_no_nudge_when_pages_change():
"""No stagnation nudge when page content changes each step."""
- detector = ActionLoopDetector(window_size=10)
+ detector = ActionLoopDetector(window_size=20)
detector.record_page_state('https://example.com', 'page content 1', 50)
detector.record_page_state('https://example.com', 'page content 2', 55)
detector.record_page_state('https://example.com', 'page content 3', 60)
@@ -209,23 +255,31 @@ def test_page_stagnation_no_nudge_when_pages_change():
assert detector.get_nudge_message() is None
-def test_page_stagnation_nudge_at_3_identical_pages():
- """Stagnation nudge fires after 3 consecutive identical page states."""
- detector = ActionLoopDetector(window_size=10)
+def test_page_stagnation_nudge_at_5_identical_pages():
+ """Stagnation nudge fires after 5 consecutive identical page states."""
+ detector = ActionLoopDetector(window_size=20)
# First recording establishes baseline (doesn't count as stagnant)
- detector.record_page_state('https://example.com', 'same content', 50)
- detector.record_page_state('https://example.com', 'same content', 50)
- detector.record_page_state('https://example.com', 'same content', 50)
- detector.record_page_state('https://example.com', 'same content', 50)
- assert detector.consecutive_stagnant_pages >= 3
+ for _ in range(6):
+ detector.record_page_state('https://example.com', 'same content', 50)
+ assert detector.consecutive_stagnant_pages >= 5
msg = detector.get_nudge_message()
assert msg is not None
assert 'PAGE STAGNATION' in msg
+def test_page_stagnation_no_nudge_at_4_identical_pages():
+ """No stagnation nudge at only 4 consecutive identical pages (below threshold)."""
+ detector = ActionLoopDetector(window_size=20)
+ # First recording establishes baseline, then 4 stagnant = 5 total recordings
+ for _ in range(5):
+ detector.record_page_state('https://example.com', 'same content', 50)
+ assert detector.consecutive_stagnant_pages == 4
+ assert detector.get_nudge_message() is None
+
+
def test_page_stagnation_resets_on_change():
"""Stagnation counter resets when page content changes."""
- detector = ActionLoopDetector(window_size=10)
+ detector = ActionLoopDetector(window_size=20)
detector.record_page_state('https://example.com', 'same content', 50)
detector.record_page_state('https://example.com', 'same content', 50)
detector.record_page_state('https://example.com', 'same content', 50)
@@ -237,13 +291,13 @@ def test_page_stagnation_resets_on_change():
def test_combined_loop_and_stagnation():
"""Both action loop and page stagnation messages appear together."""
- detector = ActionLoopDetector(window_size=10)
- # Create action repetition
- for _ in range(5):
- detector.record_action('click_element', {'index': 7})
- # Create page stagnation
+ detector = ActionLoopDetector(window_size=20)
+ # Create action repetition (8 for STRONG LOOP WARNING)
+ for _ in range(8):
+ detector.record_action('click', {'index': 7})
+ # Create page stagnation (need 5 consecutive stagnant)
detector.record_page_state('https://example.com', 'same', 50)
- for _ in range(3):
+ for _ in range(5):
detector.record_page_state('https://example.com', 'same', 50)
msg = detector.get_nudge_message()
assert msg is not None
@@ -290,8 +344,8 @@ async def test_loop_nudge_injected_into_context():
llm = create_mock_llm()
agent = Agent(task='Test task', llm=llm)
- # Simulate 3 repeated actions
- for _ in range(3):
+ # Simulate 5 repeated actions (new threshold)
+ for _ in range(5):
agent.state.loop_detector.record_action('search', {'query': 'site:example.com answers'})
agent._inject_loop_detection_nudge()
@@ -310,8 +364,8 @@ async def test_no_loop_nudge_when_disabled():
loop_detection_enabled=False,
)
- # Simulate 5 repeated actions
- for _ in range(5):
+ # Simulate 8 repeated actions
+ for _ in range(8):
agent.state.loop_detector.record_action('search', {'query': 'site:example.com answers'})
agent._inject_loop_detection_nudge()
@@ -325,9 +379,9 @@ async def test_no_loop_nudge_for_diverse_actions():
llm = create_mock_llm()
agent = Agent(task='Test task', llm=llm)
- agent.state.loop_detector.record_action('click_element', {'index': 1})
- agent.state.loop_detector.record_action('scroll', {'down': True})
- agent.state.loop_detector.record_action('click_element', {'index': 2})
+ agent.state.loop_detector.record_action('click', {'index': 1})
+ agent.state.loop_detector.record_action('scroll', {'down': True, 'index': None})
+ agent.state.loop_detector.record_action('click', {'index': 2})
agent._inject_loop_detection_nudge()
@@ -338,13 +392,13 @@ async def test_no_loop_nudge_for_diverse_actions():
async def test_loop_detector_initialized_from_settings():
"""Loop detector window size is set from agent settings."""
llm = create_mock_llm()
- agent = Agent(task='Test task', llm=llm, loop_detection_window=20)
- assert agent.state.loop_detector.window_size == 20
+ agent = Agent(task='Test task', llm=llm, loop_detection_window=30)
+ assert agent.state.loop_detector.window_size == 30
-async def test_loop_detector_default_enabled():
- """Loop detection is enabled by default."""
+async def test_loop_detector_default_window_size():
+ """Loop detection default window size is 20."""
llm = create_mock_llm()
agent = Agent(task='Test task', llm=llm)
assert agent.settings.loop_detection_enabled is True
- assert agent.state.loop_detector.window_size == 10
+ assert agent.state.loop_detector.window_size == 20