Separate blocked browser evidence from successful navigation and allow alternate sources

src/clean_agent_preview.py

@@ -3573,6 +3573,22 @@ def browser_observation_page_missing(raw):
 
 def browser_observation_access_blocked(raw):
     """Identify browser observations containing only an access gate."""
+    try:
+        decoded = json.loads(raw) if isinstance(raw, str) else raw
+    except (TypeError, ValueError):
+        decoded = None
+    rows = decoded if isinstance(decoded, list) else [decoded]
+    page_title = ''
+    for row in rows:
+        if not isinstance(row, dict):
+            continue
+        payload = row.get('result') if isinstance(row.get('result'), dict) else row
+        # A new navigation supersedes the previous page title in a batch.
+        if 'title' in payload:
+            page_title = str(payload['title']).strip().casefold().rstrip('.!')
+        if (page_title in {'client challenge', 'just a moment', 'security verification', 'verify you are human'}
+                and str(payload.get('snapshot', '')).strip() == '(empty page)'):
+            return True
     return bool(re.search(
         r"\b(?:captcha|access\s+(?:is\s+)?temporarily\s+restricted|access\s+denied|"
         r"verify\s+(?:that\s+)?you(?:\s+are|'re)\s+human|checking\s+your\s+browser|"
@@ -5099,11 +5115,11 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                             if browser_url and browser_url in static_fetch_failed_urls:
                                 # Both independent transports have now failed for
                                 # this exact source.  Do not bounce between them.
-                                force_no_tools_next_round = True
                                 round_recovery_messages.append(
                                     'Both static fetch and rendered browser access failed for this '
-                                    'same URL. Do not retry either path; briefly report the source '
-                                    'access limitation without inventing article content.'
+                                    'same URL. Do not retry either path for this source. Use another '
+                                    'relevant source already discovered if available; otherwise '
+                                    'report the access limitation without inventing article content.'
                                 )
                             else:
                                 # A CAPTCHA is transport output, not article

tests/test_clean_agent_preview.py

@@ -1900,6 +1900,21 @@ def test_browser_access_gate_is_not_treated_as_page_evidence():
     )
 
 
+@pytest.mark.parametrize('title,snapshot,blocked', [
+    ('Client Challenge', '(empty page)', True),
+    ('Just a moment...', '(empty page)', True),
+    ('Example documentation', '(empty page)', False),
+    ('Client Challenge', 'An article discussing client challenge design.', False),
+])
+def test_browser_challenge_title_with_empty_snapshot(title, snapshot, blocked):
+    from src.clean_agent_preview import browser_observation_access_blocked
+    raw = json.dumps([
+        {'command': ['open', 'https://example.org'], 'result': {'title': title}, 'success': True},
+        {'command': ['snapshot'], 'result': {'snapshot': snapshot}, 'success': True},
+    ])
+    assert browser_observation_access_blocked(raw) is blocked
+
+
 def test_rendered_missing_page_is_not_article_evidence():
     from src.clean_agent_preview import browser_observation_page_missing
 
@@ -4185,7 +4200,11 @@ async def test_failed_static_fetch_recovers_once_through_rendered_browser(monkey
             'name': 'private_browser',
             'arguments': json.dumps({'action': 'open', 'url': 'https://www.reuters.com/example'}),
         }}]}}]},
-        {'choices': [{'delta': {'content': 'Reuters blocked both access methods.'}}]},
+        {'choices': [{'delta': {'tool_calls': [{'index': 0, 'id': 'alternative', 'function': {
+            'name': 'web_fetch',
+            'arguments': json.dumps({'url': 'https://example.org/report'}),
+        }}]}}]},
+        {'choices': [{'delta': {'content': 'Reuters blocked both access methods. The alternative report was readable.'}}]},
     ])
     requests = []
 
@@ -4211,6 +4230,8 @@ async def test_failed_static_fetch_recovers_once_through_rendered_browser(monkey
     async def execute(block, **kwargs):
         executions.append(block.tool_type)
         if block.tool_type == 'web_fetch':
+            if json.loads(block.content).get('url') == 'https://example.org/report':
+                return 'web_fetch', {'output': 'Independent report with readable evidence.', 'exit_code': 0}
             return 'web_fetch', {'error': 'web_fetch: HTTP 401', 'exit_code': 1}
         return 'private_browser', {
             'output': 'Iframe "DataDome CAPTCHA" Access is temporarily restricted',
@@ -4234,12 +4255,13 @@ async def test_failed_static_fetch_recovers_once_through_rendered_browser(monkey
         disabled_tools=set(), tool_policy=ToolPolicy(), max_rounds=4,
     )]
 
-    assert executions == ['web_fetch', 'private_browser']
+    assert executions == ['web_fetch', 'private_browser', 'web_fetch']
     assert all(
         schema['function']['name'] != 'web_fetch'
         for schema in requests[1].get('tools', [])
     )
-    assert 'tools' not in requests[2]
+    assert any(schema['function']['name'] == 'web_fetch' for schema in requests[2]['tools'])
+    assert any('Use another relevant source' in str(msg.get('content')) for msg in requests[2]['messages'])
     events = [json.loads(chunk[6:]) for chunk in raw if '[DONE]' not in chunk]
     assert any(event.get('type') == 'tool_loop_recovery' for event in events)
     assert any('blocked both access methods' in event.get('delta', '') for event in events)