ground current lookups and recover failed fetches

scripts/verify_clean_v3_search_quality.mjs

@@ -10,6 +10,7 @@ const base = process.env.BASE_URL || 'http://127.0.0.1:7011';
 const owner = 'sft_alex_creator';
 const endpointId = process.env.ENDPOINT_ID || '1d1022ef';
 const endpointUrl = process.env.ENDPOINT_URL || (() => { throw new Error("ENDPOINT_URL is required"); })();
+const model = process.env.MODEL || 'odysseus-qwen3.5-tools-pre-heretic';
 const run = new Date().toISOString().replace(/[:.]/g, '-');
 const reportPath = path.resolve(process.env.REPORT_PATH || path.join(root, `reports/clean-v3-search-quality-${run}.json`));
 if (!reportPath.startsWith(path.join(root, 'reports') + path.sep) || fs.existsSync(reportPath)) throw Error('Report path must be new and under reports/');
@@ -31,7 +32,7 @@ const parseSSE = body => body.replace(/\r\n/g, '\n').split('\n\n').flatMap(frame
 
 async function createSession(context, name) {
   const response = await context.request.post(`${base}/api/session`, { multipart: {
-    name, model: 'odysseus-qwen3.5-tools-pre-heretic', endpoint_id: endpointId,
+    name, model, endpoint_id: endpointId,
     endpoint_url: endpointUrl, skip_validation: 'true', rag: 'false',
   }});
   if (!response.ok()) throw Error(`Session create HTTP ${response.status()}`);

src/clean_agent_preview.py

@@ -4945,24 +4945,19 @@ async def stream_preview(*, endpoint_url, model, messages, headers, turn_contrac
                             canonical(schema['function']['name']) == 'private_browser'
                             for schema in offered
                         )
-                        and re.search(
-                            r'\b(?:HTTP\s+(?:401|403|429)|access\s+(?:denied|blocked)|'
-                            r'captcha|needs?\s+JS|login|no\s+readable\s+text|timed?\s*out)\b',
-                            output,
-                            re.I,
-                        )
+                        and retrieved_source_urls(args)
                     ):
                         # Static fetchers are routinely rejected by publisher
                         # bot protection.  That is a transport failure, not
                         # evidence that the source is unavailable.  Offer one
                         # rendered-browser attempt at the same evidenced URL.
                         suppressed_tool_until_round['web_fetch'] = round_number + 1
-                        fetch_url = str(args.get('url') or '').strip().rstrip('/')
-                        if fetch_url:
-                            static_fetch_failed_urls.add(fetch_url)
+                        for fetch_url in retrieved_source_urls(args):
+                            static_fetch_failed_urls.add(fetch_url.rstrip('/'))
+                        force_private_browser_next_round = True
                         round_recovery_messages.append(
-                            'The static page fetch was blocked or returned no readable content. '
-                            'Use private_browser once to open the same known URL and inspect the '
+                            'The static page fetch failed or returned no readable content. Use '
+                            'private_browser once to open the strongest known URL and inspect the '
                             'rendered page; if that also fails, report the limitation without '
                             'inventing page content.'
                         )

src/turn_contract.py

@@ -671,6 +671,24 @@ def selected_tools_for_request(message: str) -> frozenset[str] | None:
     }
     if explicitly_named_web:
         return frozenset(explicitly_named_web)
+    if (
+        re.search(r"\b(?:look\s*up|search|find)\b", text, re.I)
+        and re.search(
+            r"\b(?:current|latest|today(?:'s)?|right\s+now|this\s+(?:week|month|year))\b",
+            text,
+            re.I,
+        )
+        and not re.search(r"\bhttps?://", text, re.I)
+        and not re.match(
+            r"^\s*" + _REQUEST_PREFIX + r"(?:open|browse|visit|navigate|go\s+to)\b",
+            text,
+            re.I,
+        )
+    ):
+        # Current lookups need discovery before navigation. Letting the model
+        # begin on an arbitrary browser page can ground an answer in stale or
+        # unrelated content without ever establishing a current source set.
+        return frozenset({"web_search"})
     if re.search(
         r"\buse\s+(?:the\s+)?(?:odysseus\s+)?web_search\b",
         raw_text,

tests/test_turn_contract.py

@@ -67,6 +67,15 @@ def test_latest_topic_info_and_common_news_typo_route_to_web_search():
         assert requested_capabilities(prompt) == frozenset({"search_browser"})
 
 
+@pytest.mark.parametrize('prompt', [
+    'Look up the current stock market and summarize the major US indexes.',
+    'Find the latest Nvidia driver for Linux.',
+    'Search for today\'s exchange rate for USD to JPY.',
+])
+def test_explicit_current_lookup_starts_with_native_search(prompt):
+    assert selected_tools_for_request(prompt) == frozenset({'web_search'})
+
+
 @pytest.mark.parametrize("prompt", [
     "I'm using a WIKING fireplace; can I find an English manual online?",
     "Can you look online",