Skip to content

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

fix: keep balanced trailing brackets in URLs extracted from task text

Aniket Waghd33cc68f06d5
diff --git a/browser_use/utils.py b/browser_use/utils.py
index 03e651d38..eb1814123 100644
--- a/browser_use/utils.py
+++ b/browser_use/utils.py
@@ -39,6 +39,10 @@ def is_placeholder_url(url: str) -> bool:
return len(labels) >= 2 and all(re.fullmatch(r'x+', label) for label in labels)
+_TRAILING_PROSE_PUNCTUATION = frozenset('.,;:!?([')
+_CLOSING_TO_OPENING_BRACKET = {')': '(', ']': '['}
+
+
def sanitize_url_candidate(url: str) -> str:
"""Normalize a URL candidate captured from prose before auto-navigation."""
candidate = url.strip()
@@ -46,7 +50,22 @@ def sanitize_url_candidate(url: str) -> str:
# "https://example.com/search.\\n2. Next step". Those are task text,
# not part of the URL.
candidate = re.split(r'\\[nrt]', candidate, maxsplit=1)[0]
- return re.sub(r'[.,;:!?()\[\]]+$', '', candidate)
+
+ # Strip trailing prose punctuation, but keep a closing bracket the URL opened
+ # itself, e.g. /wiki/Python_(programming_language). A closing bracket is only
+ # prose when it has no opener inside the candidate, as in "(see https://x.com/a)".
+ while candidate:
+ last_char = candidate[-1]
+ if last_char in _TRAILING_PROSE_PUNCTUATION:
+ candidate = candidate[:-1]
+ continue
+ opening_bracket = _CLOSING_TO_OPENING_BRACKET.get(last_char)
+ if opening_bracket is not None and candidate.count(last_char) > candidate.count(opening_bracket):
+ candidate = candidate[:-1]
+ continue
+ break
+
+ return candidate
# Lazy import for error types
diff --git a/tests/ci/test_beta_agent.py b/tests/ci/test_beta_agent.py
index 82aef6f9c..29f638856 100644
--- a/tests/ci/test_beta_agent.py
+++ b/tests/ci/test_beta_agent.py
@@ -4305,6 +4305,14 @@ def test_beta_agent_exposes_task_helper_methods():
assert agent._extract_start_url(numbered_task) == 'https://elibrary.ferc.gov/eLibrary/search'
assert browser_use_agent._extract_start_url(numbered_task) == 'https://elibrary.ferc.gov/eLibrary/search'
+ # A closing bracket the URL opened itself is part of the path, not prose.
+ wikipedia_task = 'Summarize https://en.wikipedia.org/wiki/Python_(programming_language)'
+ assert agent._extract_start_url(wikipedia_task) == 'https://en.wikipedia.org/wiki/Python_(programming_language)'
+ assert browser_use_agent._extract_start_url(wikipedia_task) == 'https://en.wikipedia.org/wiki/Python_(programming_language)'
+ assert agent._extract_start_url('Check https://example.com/a[1] please') == 'https://example.com/a[1]'
+ # ...but one the prose opened is still dropped.
+ assert agent._extract_start_url('See the docs (https://example.com/guide) for details.') == 'https://example.com/guide'
+
def test_beta_agent_exposes_url_text_helper_methods():
from browser_use.beta import Agent