[aidan] fix/web-search: nudge model to browser fallback when search backends fail

backend/apps/web/web.py

@@ -74,6 +74,15 @@ def p_join_text(parts: list[dict[str, Any]]) -> str:
 # Local httpx + trafilatura fetch of a real page; the fast path for /fetch (normal pages return in <2s). Set just above WebFetchTool's own 30s httpx ceiling so a valid-but-slow page still completes locally instead of being clipped down to a grounded summary; only a truly hung server gets cut.
 P_LOCAL_FETCH_TIMEOUT = 32.0
 
+# When every search backend fails, point the model at the in-product browser (always-on CreateBrowserAgent tool) instead of telling it to "wait and retry", which it can't do and just relays as a dead end. The real Chromium renders pages and isn't subject to the DDG scrape throttle.
+def p_browser_fallback_nudge(query: str) -> str:
+    return (
+        "Don't stop here: fall back to the in-product browser, which renders real pages and "
+        "isn't subject to this rate limit. Call CreateBrowserAgent with a task like: "
+        f'"Search the web for: {query}. Report the top results with their titles and URLs, '
+        'plus a direct answer if you find one."'
+    )
+
 
 async def p_gemini_grounded_call(api_key: str, prompt: str, *, use_url_context: bool) -> dict:
     """Call Gemini with googleSearch (+ optionally urlContext) grounding.
@@ -470,11 +479,12 @@ async def try_ddg():
     else:
         tail = (
             "DuckDuckGo is rate-limiting this network and every configured provider "
-            "errored (see details below). Wait a moment and retry."
+            "errored (see details below)."
         )
+    nudge = p_browser_fallback_nudge(body.query)
     return {
         "query": body.query,
-        "results": f"No results for: {body.query}\n\n{tail}",
+        "results": f"No results for: {body.query}\n\n{tail}\n\n{nudge}",
         "backend": "none",
         "cascade_errors": errors,
     }

backend/tests/test_web_search_cascade.py

@@ -128,6 +128,22 @@ async def test_everything_fails_is_honest_not_empty(monkeypatch):
     assert "Settings" in res["results"] or "API key" in res["results"]
 
 
+@pytest.mark.asyncio
+async def test_everything_fails_nudges_browser_not_retry(monkeypatch):
+    # All-fail must hand the model the browser as an escape hatch, not a dead-end "wait and retry".
+    p_ddg_throttled(monkeypatch)
+    monkeypatch.setattr(W, "p_resolve_openai_api_key", lambda: "okey")  # configured but errors
+
+    async def p_openai_boom(*a, **k):
+        raise RuntimeError("openai down")
+    monkeypatch.setattr(W, "p_openai_websearch", p_openai_boom)
+
+    res = await search(SearchBody(query="sony zv-e10 price"))
+    assert res["backend"] == "none"
+    assert "CreateBrowserAgent" in res["results"]
+    assert "retry" not in res["results"].lower()
+
+
 # -------------------------------------------------------------------------- /fetch mirrors /search: local httpx + trafilatura is the fast path, grounded fetchers are the fallback for JS/paywalled pages, every attempt is bounded. --------------------------------------------------------------------------
 
 from backend.apps.web.web import fetch, FetchBody