Fix llama-stack function tool serialization for Gemini compatibility

llama-stack's Responses API to Chat Completions conversion in
streaming.py leaks a "type" field into the function tool dict via
model_dump(). The OpenAI Chat Completions FunctionDefinition schema
only allows name, description, parameters, and strict — not type.
OpenAI silently ignores the extra field, but Gemini's OpenAI-compatible
endpoint strictly validates and rejects it with:
"Unknown name 'type' at 'tools[N].function': Cannot find field."

Fix: exclude type from serialization in streaming.py:
 function=input_tool.model_dump(exclude={"type"})

Signed-off-by: Dan Prince <dprince@redhat.com>

Author: dprince

src/llama_stack/providers/inline/agents/meta_reference/responses/streaming.py

@@ -1397,8 +1397,12 @@ def make_openai_tool(tool_name: str, tool: ToolDef) -> ChatCompletionToolParam:
 
         for input_tool in tools:
             if input_tool.type == "function":
+                # Exclude "type" from the function dict — it belongs on the outer
+                # ChatCompletionToolParam, not inside the FunctionDefinition.
+                # OpenAI silently ignores the extra field, but stricter endpoints
+                # (e.g. Gemini) reject it.
                 self.ctx.chat_tools.append(
-                    ChatCompletionToolParam(type="function", function=input_tool.model_dump(exclude_none=True))
+                    ChatCompletionToolParam(type="function", function=input_tool.model_dump(exclude={"type"}, exclude_none=True))
                 )  # type: ignore[typeddict-item,arg-type]  # Dict compatible with FunctionDefinition
             elif input_tool.type in WebSearchToolTypes:
                 tool_name = "web_search"

tests/unit/providers/agents/meta_reference/test_openai_responses.py

@@ -2540,6 +2540,47 @@ async def test_function_tool_strict_false_included(openai_responses_impl, mock_i
     assert tool_function["strict"] is False, "strict field should be False"
 
 
+async def test_function_tool_type_field_excluded_from_function_dict(openai_responses_impl, mock_inference_api):
+    """Test that the 'type' field from the input tool is excluded from the function dict.
+
+    The function dict inside ChatCompletionToolParam should only contain function-specific
+    fields (name, description, parameters, strict) and not the top-level 'type' field,
+    which belongs on the outer tool param.
+    """
+    input_text = "What is the weather?"
+    model = "meta-llama/Llama-3.1-8B-Instruct"
+
+    mock_inference_api.openai_chat_completion.return_value = fake_stream()
+
+    await openai_responses_impl.create_openai_response(
+        input=input_text,
+        model=model,
+        stream=False,
+        tools=[
+            OpenAIResponseInputToolFunction(
+                type="function",
+                name="get_weather",
+                description="Get weather information",
+                parameters={"type": "object", "properties": {"location": {"type": "string"}}, "required": ["location"]},
+            )
+        ],
+    )
+
+    assert mock_inference_api.openai_chat_completion.call_count == 1
+    params = mock_inference_api.openai_chat_completion.call_args[0][0]
+
+    assert len(params.tools) == 1
+    # The outer tool param should have type="function"
+    assert params.tools[0]["type"] == "function"
+    # The inner function dict should NOT contain 'type'
+    tool_function = params.tools[0]["function"]
+    assert "type" not in tool_function, (
+        "The input tool 'type' field should be excluded from the function dict"
+    )
+    assert tool_function["name"] == "get_weather"
+    assert tool_function["description"] == "Get weather information"
+
+
 async def test_create_openai_response_with_truncation_disabled_streaming(
     openai_responses_impl, mock_inference_api, mock_responses_store
 ):