[eric] browser: fast path refuses multi-page sweeps (3+ URLs deterministic + classifier rule); a grabbed sweep answered INSUFFICIENT from one read (ENG-355)

backend/apps/agents/browser/browser_fast_path.py

@@ -54,7 +54,10 @@ P_CLASSIFIER_SYSTEM = (
     "spreadsheets, SMS to a phone number, or other desktop apps. Also NO for "
     "plain conversation or questions answerable without visiting any site. "
     "Also NO when the task names an app the user message says is connected as "
-    "a native integration: the main agent's own tools beat a browser there.\n"
+    "a native integration: the main agent's own tools beat a browser there. "
+    "Also NO when the task needs MORE THAN TWO distinct pages or sites visited "
+    "(comparing several articles, per-item facts across a list): you route one "
+    "focused visit, not a sweep.\n"
     "The browser is the LAST resort: prefer READ over ACT whenever the goal is "
     "information rather than an action on a page, and prefer NO over READ when "
     "a plain web search could answer it without visiting a specific page.\n"
@@ -66,6 +69,8 @@ P_CLASSIFIER_SYSTEM = (
     "'count the messages in my linkedin thread with bob' -> ACT\n"
     "'find the report on stripe.com and save it to my desktop' -> NO\n"
     "'text 555-0102 that I'm late' -> NO\n"
+    "'fetch the wikipedia articles for six scientists and give a table' -> NO (a "
+    "multi-page sweep, not one visit)\n"
     "If line 1 is NO, reply with exactly the word NO and nothing else.\n"
     "If line 1 is READ or ACT, follow it with a short browsing brief:\n"
     "ENTRY: the best starting URL; use a direct deep/search URL when the site's "
@@ -93,6 +98,9 @@ def fast_path_eligible(
         return False
     if not prompt or not prompt.strip():
         return False
+    # A sweep across many pages is main-loop work: the fast path does ONE focused visit, and a grabbed sweep answers "INSUFFICIENT" from a single read (ENG-355).
+    if len(set(re.findall(r'https?://[^\s<>"\')\]]+', prompt))) >= 3:
+        return False
     return bool(P_BROWSY_RE.search(prompt))
 
 

backend/tests/test_browser_fast_path.py

@@ -189,3 +189,19 @@ def test_connected_integration_names_route_native_not_browser():
     assert fp.browsy_beyond_connected("go to slack.com and read the pricing page", ["slack"]) is True
     assert fp.browsy_beyond_connected("add a row to my Notion tracker", ["notion"]) is False
     assert fp.browsy_beyond_connected("send a linkedin message to my recruiter", ["slack", "notion"]) is True
+
+
+def test_multi_url_sweep_not_eligible():
+    """ENG-355: 3+ distinct URLs is a sweep; the fast path does one visit and a grabbed sweep answered INSUFFICIENT."""
+    from backend.apps.agents.browser.browser_fast_path import fast_path_eligible
+    sweep = ("fetch https://en.wikipedia.org/wiki/A and https://en.wikipedia.org/wiki/B "
+             "and https://en.wikipedia.org/wiki/C and report each")
+    assert fast_path_eligible(sweep, "agent", "d1", True, False) is False
+
+
+def test_single_and_double_url_still_eligible():
+    from backend.apps.agents.browser.browser_fast_path import fast_path_eligible
+    one = "open https://news.ycombinator.com and tell me the top story"
+    two = "compare https://a.com/x with https://b.com/y and tell me which loads"
+    assert fast_path_eligible(one, "agent", "d1", True, False) is True
+    assert fast_path_eligible(two, "agent", "d1", True, False) is True