From cdc0734347454cb77e58a8393d4aadf5a526fd68 Mon Sep 17 00:00:00 2001 From: pewdiepie-archdaemon Date: Thu, 17 Sep 2026 18:51:56 +0000 Subject: [PATCH] require breadth for broad current research --- src/clean_agent_preview.py | 37 ++++++++++++++++++--- tests/test_clean_agent_preview.py | 53 +++++++++++++++++++++++-------- 2 files changed, 72 insertions(+), 18 deletions(-) diff --git a/src/clean_agent_preview.py b/src/clean_agent_preview.py index e733fe37b..733df9a99 100644 --- a/src/clean_agent_preview.py +++ b/src/clean_agent_preview.py @@ -888,19 +888,24 @@ def contentless_final_response(content): )) -def incomplete_broad_web_answer(content, user_text): - """Reject a fragmentary answer to a broad current-information request.""" +def broad_current_web_request(user_text): + """Whether the user requested a broad current-information briefing.""" request = str(user_text or '') - if not ( + return bool( re.search(r'\b(?:latest|recent|current|today(?:\'s)?)\b', request, re.I) and re.search(r'\b(?:info(?:rmation)?|news|nees|updates?)\b', request, re.I) - ): + ) + + +def incomplete_broad_web_answer(content, user_text): + """Reject a shallow answer to a broad current-information request.""" + if not broad_current_web_request(user_text): return False answer = re.sub(r'https?://\S+', ' ', str(content or '')).strip() words = re.findall(r"[A-Za-z0-9][A-Za-z0-9'’-]*", answer) # A broad briefing cannot be fulfilled by one headline fragment. This is # intentionally inapplicable to narrow quick-fact searches. - return len(words) < 25 + return len(words) < 80 def progressive_thinking_for_turn(model, offered_schemas): @@ -4061,6 +4066,28 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac }) yield event({'type': 'completion_recovery', 'reason': 'contentless_answer'}) continue + if ( + broad_current_web_request(direct_user_text) + and successful_web_searches == 1 + and round_number < round_limit + ): + force_web_search_next_round = True + replace_streamed_draft_on_finish = True + history.pop() + history.append({ + 'role': 'user', '_harness_control': True, + 'content': ( + 'Research breadth check: one search is insufficient for this broad ' + 'current-information request. Run one materially different follow-up ' + 'search that fills gaps or corroborates the strongest findings. Then ' + 'inspect the best source evidence before synthesizing the answer.' + ), + }) + yield event({ + 'type': 'completion_recovery', + 'reason': 'insufficient_research_breadth', + }) + continue if ( successful_web_searches and incomplete_broad_web_answer(content, direct_user_text) diff --git a/tests/test_clean_agent_preview.py b/tests/test_clean_agent_preview.py index 242773db2..45c05ccff 100644 --- a/tests/test_clean_agent_preview.py +++ b/tests/test_clean_agent_preview.py @@ -1154,7 +1154,7 @@ async def test_stream_bounds_research_to_two_searches_fetch_then_synthesis(monke @pytest.mark.asyncio async def test_stream_retries_an_obviously_truncated_broad_web_answer(monkeypatch): - """A successful search must not end in a fragmentary headline stub.""" + """Broad current research expands, retrieves evidence, then synthesizes.""" import src.clean_agent_preview as module packets = iter([ @@ -1165,10 +1165,30 @@ async def test_stream_retries_an_obviously_truncated_broad_web_answer(monkeypatc }, }]}}]}, {'choices': [{'delta': {'content': 'Current AI news includes reports about U.'}}]}, + {'choices': [{'delta': {'tool_calls': [{ + 'index': 0, 'id': 'search-2', 'function': { + 'name': 'web_search', + 'arguments': json.dumps({'query': 'AI policy and model releases today'}), + }, + }]}}]}, + {'choices': [{'delta': {'tool_calls': [{ + 'index': 0, 'id': 'fetch-1', 'function': { + 'name': 'web_fetch', + 'arguments': json.dumps({'url': 'https://example.org/ai-news'}), + }, + }]}}]}, {'choices': [{'delta': {'content': ( - 'Here is a fuller evidence-based briefing covering the major current AI ' - 'developments, what each source actually reports, and the limits of the ' - 'available evidence. Source: https://example.org/ai-news' + 'Here is a fuller evidence-based briefing. Recent developments include ' + 'new model releases, updated deployment commitments, and policy proposals ' + 'from several governments. The first source explains what changed in the ' + 'models and how developers can access them. A second independent report ' + 'adds context about evaluation, safety, and likely industry effects. The ' + 'policy coverage distinguishes proposals from rules already in force and ' + 'identifies the dates involved. Taken together, the evidence suggests ' + 'continued rapid deployment alongside stronger demands for transparency, ' + 'although several announced measures remain preliminary. Readers should ' + 'check the linked primary material because this is a changing story. ' + 'Sources: https://example.org/ai-news and https://example.org/ai-policy' )}}]}, ]) requests = [] @@ -1191,7 +1211,7 @@ async def test_stream_retries_an_obviously_truncated_broad_web_answer(monkeypatc return Response(next(packets)) async def execute(block, **kwargs): - return 'web_search', { + return block.tool_type, { 'output': '[1] AI News\n https://example.org/ai-news', 'exit_code': 0, 'evidence_status': 'available', @@ -1199,23 +1219,30 @@ async def test_stream_retries_an_obviously_truncated_broad_web_answer(monkeypatc monkeypatch.setattr(module.httpx, 'AsyncClient', Client) monkeypatch.setattr(module, 'execute_tool_block', execute) - schema = next( - s for s in FUNCTION_TOOL_SCHEMAS if s['function']['name'] == 'web_search' - ) - contract = resolve_full_inventory_contract(schemas=[schema], policy=ToolPolicy()) + schemas = [ + s for s in FUNCTION_TOOL_SCHEMAS + if s['function']['name'] in {'web_search', 'web_fetch'} + ] + contract = resolve_full_inventory_contract(schemas=schemas, policy=ToolPolicy()) raw = [chunk async for chunk in stream_preview( endpoint_url='http://test', model='test', messages=[{'role': 'user', 'content': 'Latest news in AI?'}], headers={}, turn_contract=contract, session_id='test', owner='test', - disabled_tools=set(), tool_policy=ToolPolicy(), max_rounds=3, + disabled_tools=set(), tool_policy=ToolPolicy(), max_rounds=5, )] events = [json.loads(chunk[6:]) for chunk in raw if '[DONE]' not in chunk] - assert len(requests) == 3 - assert 'tools' not in requests[2] + assert len(requests) == 5 + assert requests[2]['tool_choice'] == { + 'type': 'function', 'function': {'name': 'web_search'}, + } + assert requests[3]['tool_choice'] == { + 'type': 'function', 'function': {'name': 'web_fetch'}, + } + assert 'tools' not in requests[4] assert any( event.get('type') == 'completion_recovery' - and event.get('reason') == 'incomplete_research_answer' + and event.get('reason') == 'insufficient_research_breadth' for event in events ) assert any(