Skip to content

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

fix(agent): match URL negations as whole words

JayLay03108e880168a941
diff --git a/browser_use/agent/service.py b/browser_use/agent/service.py
index 3e0970ddd..304f5765a 100644
--- a/browser_use/agent/service.py
+++ b/browser_use/agent/service.py
@@ -77,6 +77,7 @@ from browser_use.utils import (
_log_pretty_path,
check_latest_browser_use_version,
get_browser_use_version,
+ has_url_negation,
is_placeholder_url,
sanitize_url_candidate,
time_execution_async,
@@ -2367,13 +2368,6 @@ class Agent(Generic[Context, AgentStructuredOutput]):
'polynomial',
}
- excluded_words = {
- 'never',
- 'dont',
- 'not',
- "don't",
- }
-
found_urls = []
matched_spans: list[tuple[int, int]] = []
for pattern in patterns:
@@ -2411,16 +2405,14 @@ class Agent(Generic[Context, AgentStructuredOutput]):
self.logger.debug(f'Excluding URL with file extension from auto-navigation: {url}')
continue
- # If in the 20 characters before the url position is a word in excluded_words skip to avoid "Never go to this url"
+ # Skip URLs explicitly negated by nearby prose, such as "Never go to this URL".
context_start = max(0, original_position - 20)
context_text = task_without_emails[context_start:original_position]
- if any(word.lower() in context_text.lower() for word in excluded_words):
- self.logger.debug(
- f'Excluding URL with word in excluded words from auto-navigation: {url} (context: "{context_text.strip()}")'
- )
+ if has_url_negation(context_text):
+ self.logger.debug(f'Excluding negated URL from auto-navigation: {url} (context: "{context_text.strip()}")')
continue
- # Add https:// if missing (after excluded words check to avoid position calculation issues)
+ # Add https:// after the negation check to preserve source positions.
if not has_scheme:
url = 'https://' + url
diff --git a/browser_use/beta/service.py b/browser_use/beta/service.py
index d8f116688..3a0bb4c45 100644
--- a/browser_use/beta/service.py
+++ b/browser_use/beta/service.py
@@ -76,6 +76,7 @@ from browser_use.utils import (
check_latest_browser_use_version,
get_browser_use_version,
get_git_info,
+ has_url_negation,
is_placeholder_url,
sanitize_url_candidate,
)
@@ -1460,8 +1461,6 @@ def _extract_start_url(task: str) -> str | None:
'rpm',
'iso',
}
- excluded_words = {'never', 'dont', 'not', "don't"}
-
found_urls = []
matched_spans: list[tuple[int, int]] = []
for pattern in patterns:
@@ -1483,7 +1482,7 @@ def _extract_start_url(task: str) -> str | None:
continue
context_start = max(0, match.start() - 20)
context_text = task_without_emails[context_start : match.start()]
- if any(word in context_text.lower() for word in excluded_words):
+ if has_url_negation(context_text):
continue
if not has_scheme:
url = 'https://' + url
diff --git a/browser_use/utils.py b/browser_use/utils.py
index 03e651d38..5cbfee760 100644
--- a/browser_use/utils.py
+++ b/browser_use/utils.py
@@ -20,6 +20,7 @@ load_dotenv()
# Pre-compiled regex for URL detection - used in URL shortening
URL_PATTERN = re.compile(r'https?://[^\s<>"\']+|www\.[^\s<>"\']+|[^\s<>"\']+\.[a-z]{2,}(?:/[^\s<>"\']*)?', re.IGNORECASE)
+URL_NEGATION_PATTERN = re.compile(r"\b(?:never|not|don'?t)\b", re.IGNORECASE)
logger = logging.getLogger(__name__)
@@ -49,6 +50,11 @@ def sanitize_url_candidate(url: str) -> str:
return re.sub(r'[.,;:!?()\[\]]+$', '', candidate)
+def has_url_negation(context: str) -> bool:
+ """Return whether nearby prose explicitly negates navigation to a URL."""
+ return URL_NEGATION_PATTERN.search(context) is not None
+
+
# Lazy import for error types
# Use sentinel to avoid retrying import when package is not installed
_IMPORT_NOT_FOUND: type = type('_ImportNotFound', (), {})
diff --git a/tests/ci/test_beta_agent.py b/tests/ci/test_beta_agent.py
index 82aef6f9c..359f731e9 100644
--- a/tests/ci/test_beta_agent.py
+++ b/tests/ci/test_beta_agent.py
@@ -4297,6 +4297,10 @@ def test_beta_agent_exposes_task_helper_methods():
assert 'Expected output format: Answer' in enhanced
assert '"answer"' in enhanced
assert agent._extract_start_url('Open example.com and report the title.') == 'https://example.com'
+ assert agent._extract_start_url('Open this notable site: https://example.com') == 'https://example.com'
+ assert browser_use_agent._extract_start_url('Open this notable site: https://example.com') == 'https://example.com'
+ assert agent._extract_start_url('Do not open https://example.com.') is None
+ assert browser_use_agent._extract_start_url('Do not open https://example.com.') is None
assert agent._extract_start_url('Email [email protected] only.') is None
assert agent._extract_start_url('Open https://example.com/report.pdf and summarize it.') is None
assert agent._extract_start_url('Use https://XXX.XX as a placeholder in the table.') is None