Schedule broad research before synthesis and clarify briefing layout

scripts/verify_clean_v3_search_quality.mjs

@@ -88,6 +88,9 @@ async function send(page, prompt) {
     prompt, seconds: (performance.now() - started) / 1000,
     streamed_text_chunks: events.filter(event => typeof event.delta === 'string' && event.delta.length).length,
     final_replacement_count: events.filter(event => event.type === 'final_response').length,
+    event_order: events.filter(event => event.type === 'tool_start' || event.type === 'final_response' || event.delta)
+      .map(event => event.type === 'tool_start' ? `tool:${event.tool}` : event.type === 'final_response' ? 'final' : 'text')
+      .filter((value, index, array) => index === 0 || value !== array[index - 1]),
     rendered_answers: renderedAnswers,
     rounds: metrics?.agent_rounds ?? null,
     tool_execution_timings: metrics?.tool_execution_timings || [],
@@ -121,6 +124,7 @@ try {
     const cases = [
       ['news-typo', ['latset ai neews?', 'more about the second story, with sources'], true],
       ['country-casual', ['whats new in japan rn'], true],
+      ['country-sweden', ['Latest news in sweden'], true],
       ['software-typo', ['latest pythno verison? official source pls'], true],
       ['manual', ['find official english manual for Sony WH-1000XM5'], true],
       ['comparison', ['compare current firefox and chrome privacy features with sources'], true],

src/clean_agent_preview.py

@@ -4089,6 +4089,26 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
             for round_number in range(1, round_limit + 1):
                 rounds_used = round_number
                 yield event({'type': 'agent_step', 'round': round_number})
+                # Enforce a known research prerequisite before asking the model
+                # for another response, not after streaming a premature answer.
+                if (
+                    broad_current_web_request(direct_user_text)
+                    and successful_web_searches == 1 and web_search_attempts < 2
+                    and not breadth_recovery_attempted and not search_completion_attempted
+                    and not force_no_tools_next_round and not required_artifacts
+                    and round_number < round_limit
+                    and any(canonical(s['function']['name']) == 'web_search' for s in offered)
+                    and 'web_search' not in permanently_suppressed_tools
+                ):
+                    breadth_recovery_attempted = True
+                    force_web_search_next_round = True
+                    history.append({'role': 'user', '_harness_control': True, 'content': (
+                        'Continue research before drafting the answer. Use the findings from the '
+                        'first search to choose one materially different follow-up query that '
+                        'fills a gap or corroborates the strongest findings. Then assess the '
+                        'source evidence and prepare the briefing. Do not repeat the same query.'
+                    )})
+                    yield event({'type': 'completion_recovery', 'reason': 'research_before_synthesis'})
                 if (
                     required_artifacts
                     and not successful_write
@@ -4260,6 +4280,7 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                             if (
                                 not prior_summary_answer
                                 and not progressive_thinking
+                                and request.get('tool_choice') in (None, 'auto', 'none')
                             ):
                                 text_event = {'delta': text}
                                 if replace_streamed_draft_on_finish and not streamed_round_text:
@@ -4385,9 +4406,12 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                                 'Completion check: the draft is still too shallow and does not '
                                 'answer the broad current-information request. Using the Web '
                                 'evidence already gathered, provide a complete useful briefing '
-                                'of at least several substantive paragraphs or equivalent bullets, '
+                                'with a short opening and clearly separated **bold topic labels** '
+                                'or Markdown headings, followed by substantive paragraphs or bullets, '
                                 'with the main findings, context, source links, and any evidence '
-                                'limitations. Do not call another tool or return another one-sentence summary.'
+                                'limitations. Use descriptive Markdown links next to the supported '
+                                'findings rather than raw URLs. Do not call another tool or return '
+                                'another one-sentence summary.'
                             ),
                         })
                         yield event({
@@ -5632,7 +5656,9 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                         recovery = (
                             'Research budget reached: no more tools will be offered. Using only '
                             'the search evidence already returned, provide the complete final '
-                            'answer now with useful detail and source URLs. Do not emit a tool call.'
+                            'answer now with useful detail. For a multi-topic briefing, use bold '
+                            'topic labels or headings, separated paragraphs or bullets, and '
+                            'descriptive source links alongside supported findings. Do not emit a tool call.'
                         )
                         if history and history[-1].get('_harness_control'):
                             history[-1]['content'] = (

tests/test_clean_agent_preview.py

@@ -1200,7 +1200,7 @@ async def test_stream_bounds_research_to_two_searches_fetch_then_synthesis(monke
 @pytest.mark.parametrize('embedded_article', [False, True])
 @pytest.mark.parametrize('empty_second_search', [False, True])
 @pytest.mark.parametrize('sources_requested', [False, True])
-async def test_stream_retries_an_obviously_truncated_broad_web_answer(monkeypatch, embedded_article, empty_second_search, sources_requested):
+async def test_stream_research_prerequisite_precedes_broad_web_answer(monkeypatch, embedded_article, empty_second_search, sources_requested):
     """Broad current research expands, retrieves evidence, then synthesizes."""
     import src.clean_agent_preview as module
 
@@ -1211,7 +1211,6 @@ async def test_stream_retries_an_obviously_truncated_broad_web_answer(monkeypatc
                 'arguments': json.dumps({'query': 'latest AI news'}),
             },
         }]}}]},
-        {'choices': [{'delta': {'content': 'Current AI news includes reports about U.'}}]},
         {'choices': [{'delta': {'tool_calls': [{
             'index': 0, 'id': 'search-2', 'function': {
                 'name': 'web_search',
@@ -1240,7 +1239,7 @@ async def test_stream_retries_an_obviously_truncated_broad_web_answer(monkeypatc
     ]
     article = packets[-1]['choices'][0]['delta']['content']
     if embedded_article or empty_second_search:
-        packets.pop(3)
+        packets.pop(2)
     packets = iter(packets)
     requests = []
 
@@ -1293,27 +1292,27 @@ async def test_stream_retries_an_obviously_truncated_broad_web_answer(monkeypatc
     )]
     events = [json.loads(chunk[6:]) for chunk in raw if '[DONE]' not in chunk]
 
-    assert len(requests) == (4 if embedded_article or empty_second_search else 5)
+    assert len(requests) == (3 if embedded_article or empty_second_search else 4)
-    assert requests[2]['tool_choice'] == 'required'
+    assert requests[1]['tool_choice'] == 'required'
-    assert [s['function']['name'] for s in requests[2]['tools']] == ['web_search']
+    assert [s['function']['name'] for s in requests[1]['tools']] == ['web_search']
     if not embedded_article and not empty_second_search:
-        assert requests[3]['tool_choice'] == {
+        assert requests[2]['tool_choice'] == {
             'type': 'function', 'function': {'name': 'web_fetch'},
         }
     assert any(s['function']['name'] == 'web_fetch' for s in requests[-1]['tools'])
     assert requests[-1].get('tool_choice') != 'none'
     assert any(
         event.get('type') == 'completion_recovery'
-        and event.get('reason') == 'insufficient_research_breadth'
+        and event.get('reason') == 'research_before_synthesis'
         for event in events
     )
-    assert sum(event.get('reason') == 'insufficient_research_breadth' for event in events) == 1
+    assert sum(event.get('reason') == 'research_before_synthesis' for event in events) == 1
     assert any(
         event.get('type') == 'final_response'
         and 'fuller evidence-based briefing' in event.get('content', '')
         for event in events
     )
-    assert any(
+    assert not any(
         event.get('delta') == 'Current AI news includes reports about U.'
         for event in events
     )
@@ -1324,7 +1323,7 @@ async def test_stream_retries_an_obviously_truncated_broad_web_answer(monkeypatc
     replacement = next(event for event in events if event.get('type') == 'final_response')
     assert replacement['replacement_scope'] == 'turn'
     assert replacement['render_owner'] == 'streamed'
-    assert any(event.get('delta') and event.get('replacement_scope') == 'turn' for event in events)
+    assert any(event.get('delta') for event in events)
 
 
 def test_task_renderer_honors_few_and_filters_confirmed_morning_schedule():