Fix Gemini URL retrieval defaults and empty response errors

docs/ai_configuration.md

@@ -92,8 +92,8 @@ Endpoint overrides remain available in APIs that already expose them.
 
 All packaged operations default to one additional retry after the first attempt.
 Set `retries: 0` to disable retries. Temperature is model-specific: modern OpenAI
-models leave it unset, legacy GPT-4o Chat Completions keeps `0.2`, and the configured
-Gemini URL-retrieval model keeps `0.1`. Unset temperature uses the provider's default.
+models and Gemini URL retrieval leave it unset, while legacy GPT-4o Chat Completions
+keeps `0.2`. Unset temperature uses the provider's default (`1.0` for Gemini 3).
 
 For extraction, a saved definition's model retains its existing precedence over
 the caller's `model`. Both tuning and runtime defaults are resolved for that

tests/test_ai_caller_config.py

@@ -449,6 +449,8 @@ def google_transport(monkeypatch):
 
     def generate_content(**kwargs):
         calls["requests"].append(kwargs)
+        if "response" in calls:
+            return calls["response"]
         return SimpleNamespace(candidates=[SimpleNamespace(
             url_context_metadata=None,
             content=SimpleNamespace(parts=[SimpleNamespace(text='{"name":"Synthetic"}')]),
@@ -507,6 +509,65 @@ def test_google_retrieval_uses_config_at_sdk_boundary(configured_ai, google_tran
     assert len(google_transport["closed"]) == 1
 
 
+def test_google_packaged_defaults_omit_temperature(google_transport, monkeypatch):
+    monkeypatch.delenv("WRANGLES_AI_CONFIG", raising=False)
+    ai_config.clear_cache()
+    try:
+        result = gemini.GeminiURLContextClient(api_key="fake-key").retrieve(
+            "https://product.example/one", output_format="json",
+        )
+        assert "temperature" not in google_transport["requests"][0]["config"]
+        assert result["error"] is None
+        assert result["extracted_content"] == {"name": "Synthetic"}
+        assert google_transport["closed"] == [True]
+    finally:
+        ai_config.clear_cache()
+
+
+@pytest.mark.parametrize("content,finish_message", [
+    (None, "Model generated function call(s)."),
+    ({}, "Model generated function call(s)."),
+    ({"parts": []}, "Model generated function call(s)."),
+    ({}, None),
+], ids=["missing", "empty-object", "empty-parts", "no-finish-message"])
+@pytest.mark.parametrize("url_retrieved", [False, True], ids=["no-url-metadata", "url-retrieved"])
+@pytest.mark.filterwarnings("ignore:TOO_MANY_TOOL_CALLS is not a valid FinishReason:UserWarning")
+def test_google_empty_content_preserves_provider_failure(
+    configured_ai, google_transport, monkeypatch, content, finish_message, url_retrieved,
+):
+    from google.genai import types
+
+    candidate = {
+        "finishReason": "TOO_MANY_TOOL_CALLS",
+    }
+    if finish_message is not None:
+        candidate["finishMessage"] = finish_message
+    if content is not None:
+        candidate["content"] = content
+    if url_retrieved:
+        candidate["urlContextMetadata"] = {"urlMetadata": [{
+            "retrievedUrl": "https://product.example/one",
+            "urlRetrievalStatus": "URL_RETRIEVAL_STATUS_SUCCESS",
+        }]}
+    google_transport["response"] = types.GenerateContentResponse.model_validate({
+        "candidates": [candidate],
+    })
+    genai, _, errors = gemini._get_genai()
+    monkeypatch.setattr(gemini, "_get_genai", lambda: (genai, types, errors))
+
+    result = gemini.GeminiURLContextClient(api_key="fake-key").retrieve(
+        "https://product.example/one", output_format="json",
+    )
+
+    assert result["status"] == "Failure"
+    assert "TOO_MANY_TOOL_CALLS" in result["error"]
+    if finish_message is not None:
+        assert finish_message in result["error"]
+    assert result["extracted_content"] is None
+    assert len(google_transport["requests"]) == 1
+    assert google_transport["closed"] == [True]
+
+
 def test_google_direct_client_reads_changed_config_each_call(configured_ai, google_transport):
     config, save = configured_ai
     client = gemini.GeminiURLContextClient(api_key="fake-key")

wrangles/ai_defaults.yml

@@ -86,8 +86,7 @@ providers:
         status: active
         applications: [url_retrieval]
         default_for: [search.retrieve_link_content]
-        defaults:
-          temperature: 0.1
+        defaults: {}
   jina:
     endpoints:
       embeddings: https://api.jina.ai/v1/embeddings

wrangles/clients/gemini.py

@@ -147,11 +147,12 @@ def _retrieve(self, url, prompt, output_format, policy):
             # Process Standard Text Response
             if not response.candidates:
                 result["error"] = "API returned zero candidates."
-            elif not response.candidates[0].content:
-                reason = response.candidates[0].finish_reason
+            elif not cand.content or not cand.content.parts:
+                reason = cand.finish_reason
                 result["error"] = f"Blocked/Empty. Finish Reason: {reason}"
-                if result["status"] != "Success":
-                    result["status"] = "Failure"
+                if cand.finish_message:
+                    result["error"] += f". {cand.finish_message}"
+                result["status"] = "Failure"
             else:
                 full_text = "\n".join([part.text for part in response.candidates[0].content.parts])