fix(vertex_ai): omit function_call id on Vertex Gemini 3.5+ tool turns (#28324)

* fix(vertex_ai): omit function_call id on Vertex Gemini 3.5+ tool turns

Vertex AI rejects `id` on function_call/function_response parts; only Google AI Studio accepts it for Gemini 3.5+ strict tool matching.

Co-authored-by: Cursor <cursoragent@cursor.com>

* Update litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py

Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>

* fix(vertex_ai): forward custom_llm_provider in context caching

Pass custom_llm_provider through to _gemini_convert_messages_with_history
in the context caching path so Gemini 3.5+ tool-call `id` forwarding
behaves consistently between cached and non-cached completions on Google
AI Studio.

Co-authored-by: Claude <claude@anthropic.com>

---------

Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Claude <claude@anthropic.com>
(cherry picked from commit fecf212d70)

litellm/litellm_core_utils/prompt_templates/factory.py

@@ -1344,6 +1344,7 @@ def _get_dummy_thought_signature() -> str:
 def convert_to_gemini_tool_call_invoke(
     message: ChatCompletionAssistantMessage,
     model: Optional[str] = None,
+    custom_llm_provider: Optional[str] = None,
 ) -> List[VertexPartType]:
     """
     OpenAI tool invokes:
@@ -1394,7 +1395,10 @@ def convert_to_gemini_tool_call_invoke(
         )
 
         forward_tool_call_id = bool(
-            model and VertexGeminiConfig._is_gemini_3_or_newer(model)
+            model
+            and VertexGeminiConfig._forward_gemini_function_call_id(
+                model, custom_llm_provider
+            )
         )
 
         if tool_calls is not None:
@@ -1475,6 +1479,7 @@ def convert_to_gemini_tool_call_result(  # noqa: PLR0915
     message: Union[ChatCompletionToolMessage, ChatCompletionFunctionMessage],
     last_message_with_tool_calls: Optional[dict],
     model: Optional[str] = None,
+    custom_llm_provider: Optional[str] = None,
 ) -> Union[VertexPartType, List[VertexPartType]]:
     """
     OpenAI message with a tool result looks like:
@@ -1616,14 +1621,16 @@ def convert_to_gemini_tool_call_result(  # noqa: PLR0915
                 name = tool.get("function", {}).get("name", "")
 
     # Echo the OpenAI tool_call_id on functionResponse (strip thought-signature suffix).
-    # Only Gemini 3+ accepts (and returns) an `id` on function_response parts;
-    # older Gemini models reject the field with a 400.
+    # Only Google AI Studio Gemini 3+ accepts `id` on function_response parts.
+    # Vertex AI and older Gemini models reject the field with HTTP 400.
     from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
         VertexGeminiConfig,
     )
 
     gemini_call_id: Optional[str] = None
-    if model and VertexGeminiConfig._is_gemini_3_or_newer(model):
+    if model and VertexGeminiConfig._forward_gemini_function_call_id(
+        model, custom_llm_provider
+    ):
         raw_tool_call_id = message.get("tool_call_id")
         if raw_tool_call_id and isinstance(raw_tool_call_id, str):
             stripped_id = raw_tool_call_id.split(THOUGHT_SIGNATURE_SEPARATOR, 1)[0]

litellm/llms/gemini/chat/transformation.py

@@ -151,4 +151,8 @@ class GoogleAIStudioGeminiConfig(VertexGeminiConfig):
                             except Exception:
                                 # If conversion fails, leave as is and let the API handle it
                                 pass
-        return _gemini_convert_messages_with_history(messages=messages, model=model)
+        return _gemini_convert_messages_with_history(
+            messages=messages,
+            model=model,
+            custom_llm_provider="gemini",
+        )

litellm/llms/vertex_ai/context_caching/transformation.py

@@ -174,7 +174,9 @@ def transform_openai_messages_to_gemini_context_caching(
     )
 
     transformed_messages = _gemini_convert_messages_with_history(
-        messages=new_messages, model=model
+        messages=new_messages,
+        model=model,
+        custom_llm_provider=custom_llm_provider,
     )
 
     model_name = "models/{}".format(model)

litellm/llms/vertex_ai/gemini/transformation.py

@@ -311,6 +311,7 @@ def check_if_part_exists_in_parts(
 def _gemini_convert_messages_with_history(  # noqa: PLR0915
     messages: List[AllMessageValues],
     model: Optional[str] = None,
+    custom_llm_provider: Optional[str] = None,
 ) -> List[ContentType]:
     """
     Converts given messages from OpenAI format to Gemini format
@@ -548,7 +549,9 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
                     or assistant_msg.get("function_call") is not None
                 ):  # support assistant tool invoke conversion
                     gemini_tool_call_parts = convert_to_gemini_tool_call_invoke(
-                        assistant_msg, model=model
+                        assistant_msg,
+                        model=model,
+                        custom_llm_provider=custom_llm_provider,
                     )
                     ## check if gemini_tool_call already exists in assistant_content
                     for gemini_tool_call_part in gemini_tool_call_parts:
@@ -610,6 +613,7 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
                     messages[msg_i],  # type: ignore
                     last_message_with_tool_calls,  # type: ignore
                     model=model,
+                    custom_llm_provider=custom_llm_provider,
                 )
                 msg_i += 1
                 # Handle both single part and list of parts (for Computer Use with images)

litellm/llms/vertex_ai/gemini/vertex_and_google_ai_studio_gemini.py

@@ -289,6 +289,20 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
             return True
         return False
 
+    @staticmethod
+    def _forward_gemini_function_call_id(
+        model: str, custom_llm_provider: Optional[str] = None
+    ) -> bool:
+        """
+        Whether to include `id` on function_call / function_response parts.
+
+        Gemini 3+ on Google AI Studio accepts (and returns) `id` for strict
+        tool-call matching. Vertex AI rejects the field with HTTP 400.
+        """
+        if custom_llm_provider != "gemini":
+            return False
+        return VertexGeminiConfig._is_gemini_3_or_newer(model)
+
     def _supports_penalty_parameters(self, model: str) -> bool:
         # Gemini 3 models do not support penalty parameters
         if VertexGeminiConfig._is_gemini_3_or_newer(model):
@@ -2645,7 +2659,11 @@ class VertexGeminiConfig(VertexAIBaseConfig, BaseConfig):
     def _transform_messages(
         self, messages: List[AllMessageValues], model: Optional[str] = None
     ) -> List[ContentType]:
-        return _gemini_convert_messages_with_history(messages=messages, model=model)
+        return _gemini_convert_messages_with_history(
+            messages=messages,
+            model=model,
+            custom_llm_provider="vertex_ai",
+        )
 
     def get_error_class(
         self, error_message: str, status_code: int, headers: Union[Dict, httpx.Headers]

litellm/types/llms/vertex_ai.py

@@ -16,15 +16,15 @@ GeminiEmbeddingInput = Union[EmbeddingInput, List[List[str]]]
 
 class FunctionResponse(TypedDict, total=False):
     # `id` correlates this response with the originating `functionCall` part.
-    # Required by Gemini 3.5+ for strict function-calling response matching.
+    # Supported on Google AI Studio Gemini 3.5+; Vertex AI rejects this field.
     id: str
     name: Required[str]
     response: Optional[dict]
 
 
 class FunctionCall(TypedDict, total=False):
-    # `id` is returned by Gemini 3.5+ to correlate the corresponding
-    # `functionResponse`. Older Gemini models omit this field.
+    # `id` correlates the corresponding `functionResponse` on Google AI Studio
+    # Gemini 3.5+. Vertex AI and older Gemini models omit/reject this field.
     id: str
     name: Required[str]
     args: Optional[dict]
@@ -52,8 +52,8 @@ class PartType(TypedDict, total=False):
 
 
 class HttpxFunctionCall(TypedDict, total=False):
-    # `id` is returned by Gemini 3.5+ to correlate the corresponding
-    # `functionResponse`. Older Gemini models omit this field.
+    # `id` correlates the corresponding `functionResponse` on Google AI Studio
+    # Gemini 3.5+. Vertex AI and older Gemini models omit/reject this field.
     id: str
     name: Required[str]
     args: dict

tests/test_litellm/llms/vertex_ai/gemini/test_vertex_and_google_ai_studio_gemini.py

@@ -2097,6 +2097,125 @@ def test_is_gemini_3_or_newer():
     assert VertexGeminiConfig._is_gemini_3_or_newer("") == False
 
 
+def test_forward_gemini_function_call_id_vertex_vs_google_ai_studio():
+    """Vertex AI rejects `id` on function_call/function_response; Google AI Studio accepts it on Gemini 3.5+."""
+    from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
+        VertexGeminiConfig,
+    )
+
+    model = "gemini-3.5-flash"
+    assert (
+        VertexGeminiConfig._forward_gemini_function_call_id(model, "vertex_ai") is False
+    )
+    assert (
+        VertexGeminiConfig._forward_gemini_function_call_id(model, "vertex_ai_beta")
+        is False
+    )
+    assert VertexGeminiConfig._forward_gemini_function_call_id(model, "gemini") is True
+    assert VertexGeminiConfig._forward_gemini_function_call_id(model, None) is False
+    assert (
+        VertexGeminiConfig._forward_gemini_function_call_id(
+            "gemini-2.5-flash", "gemini"
+        )
+        is False
+    )
+
+
+def test_vertex_ai_gemini_35_tool_calls_omit_function_call_id():
+    """Regression: Vertex must not send OpenAI tool_call id inside Gemini function_call parts."""
+    from litellm.llms.vertex_ai.gemini.transformation import (
+        _gemini_convert_messages_with_history,
+    )
+
+    messages = [
+        {"role": "user", "content": "Explore this directory"},
+        {
+            "role": "assistant",
+            "content": "",
+            "tool_calls": [
+                {
+                    "id": "call_50e7e0fe0989464a89f188eda443",
+                    "type": "function",
+                    "function": {
+                        "name": "read",
+                        "arguments": '{"filePath": "/tmp"}',
+                    },
+                }
+            ],
+        },
+        {
+            "role": "tool",
+            "tool_call_id": "call_50e7e0fe0989464a89f188eda443",
+            "content": "ok",
+        },
+    ]
+
+    contents = _gemini_convert_messages_with_history(
+        messages=messages,
+        model="gemini-3.5-flash",
+        custom_llm_provider="vertex_ai",
+    )
+
+    for content in contents:
+        for part in content.get("parts", []):
+            fc = part.get("function_call")
+            if fc is not None:
+                assert "id" not in fc, f"Vertex payload must not include id: {fc}"
+            fr = part.get("function_response")
+            if fr is not None:
+                assert "id" not in fr, f"Vertex payload must not include id: {fr}"
+
+
+def test_google_ai_studio_gemini_35_tool_calls_include_function_call_id():
+    from litellm.llms.vertex_ai.gemini.transformation import (
+        _gemini_convert_messages_with_history,
+    )
+
+    tool_call_id = "call_50e7e0fe0989464a89f188eda443"
+    messages = [
+        {"role": "user", "content": "hi"},
+        {
+            "role": "assistant",
+            "content": "",
+            "tool_calls": [
+                {
+                    "id": tool_call_id,
+                    "type": "function",
+                    "function": {
+                        "name": "read",
+                        "arguments": '{"filePath": "/tmp"}',
+                    },
+                }
+            ],
+        },
+        {
+            "role": "tool",
+            "tool_call_id": tool_call_id,
+            "content": "ok",
+        },
+    ]
+
+    contents = _gemini_convert_messages_with_history(
+        messages=messages,
+        model="gemini-3.5-flash",
+        custom_llm_provider="gemini",
+    )
+
+    function_call_ids = []
+    function_response_ids = []
+    for content in contents:
+        for part in content.get("parts", []):
+            fc = part.get("function_call")
+            if fc is not None:
+                function_call_ids.append(fc.get("id"))
+            fr = part.get("function_response")
+            if fr is not None:
+                function_response_ids.append(fr.get("id"))
+
+    assert function_call_ids == [tool_call_id]
+    assert function_response_ids == [tool_call_id]
+
+
 def test_reasoning_effort_maps_to_thinking_level_gemini_3():
     """Test that reasoning_effort maps to thinking_level AND includeThoughts for Gemini 3+ models"""
     from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
@@ -3531,7 +3650,12 @@ def test_video_metadata_supported_for_all_gemini_models():
         }
     ]
 
-    for model in ["gemini-1.5-pro", "gemini-2.5-flash", "gemini-2.5-pro", "gemini-3-pro-preview"]:
+    for model in [
+        "gemini-1.5-pro",
+        "gemini-2.5-flash",
+        "gemini-2.5-pro",
+        "gemini-3-pro-preview",
+    ]:
         contents = _gemini_convert_messages_with_history(messages=messages, model=model)
 
         file_part = None
@@ -3541,19 +3665,25 @@ def test_video_metadata_supported_for_all_gemini_models():
                 break
 
         assert file_part is not None, f"{model}: file part should exist"
-        assert "video_metadata" in file_part, f"{model}: video_metadata should be present"
+        assert (
+            "video_metadata" in file_part
+        ), f"{model}: video_metadata should be present"
         assert file_part["video_metadata"]["fps"] == 5, f"{model}: fps should be 5"
 
     # Per-part media_resolution is Gemini 3+ only; 2.x uses generation_config global
     for model in ["gemini-3-pro-preview"]:
         contents = _gemini_convert_messages_with_history(messages=messages, model=model)
         file_part = next(p for p in contents[0]["parts"] if "file_data" in p)
-        assert "media_resolution" in file_part, f"{model}: media_resolution should be present"
+        assert (
+            "media_resolution" in file_part
+        ), f"{model}: media_resolution should be present"
 
     for model in ["gemini-1.5-pro", "gemini-2.5-flash", "gemini-2.5-pro"]:
         contents = _gemini_convert_messages_with_history(messages=messages, model=model)
         file_part = next(p for p in contents[0]["parts"] if "file_data" in p)
-        assert "media_resolution" not in file_part, f"{model}: per-part media_resolution should not be set"
+        assert (
+            "media_resolution" not in file_part
+        ), f"{model}: per-part media_resolution should not be set"
 
 
 def test_chunk_parser_handles_prompt_feedback_block():
@@ -4186,8 +4316,9 @@ def test_vertex_ai_usage_metadata_with_document_tokens_in_prompt():
 
     # DOCUMENT tokens should be included in text_tokens: 8 (TEXT) + 774 (DOCUMENT) = 782
     assert result.prompt_tokens_details is not None
-    assert result.prompt_tokens_details.text_tokens == 782, \
-        "DOCUMENT modality tokens should be added to text_tokens (8 TEXT + 774 DOCUMENT = 782)"
+    assert (
+        result.prompt_tokens_details.text_tokens == 782
+    ), "DOCUMENT modality tokens should be added to text_tokens (8 TEXT + 774 DOCUMENT = 782)"
 
     # Verify completion token details
     assert result.completion_tokens_details is not None
@@ -4222,8 +4353,9 @@ def test_vertex_ai_usage_metadata_with_document_tokens_cached():
 
     # DOCUMENT cached tokens map to cached_text_tokens, so:
     # text_tokens = (8 TEXT + 774 DOCUMENT) - 400 cached = 382
-    assert result.prompt_tokens_details.text_tokens == 382, \
-        "text_tokens should be (8 + 774) - 400 cached = 382"
+    assert (
+        result.prompt_tokens_details.text_tokens == 382
+    ), "text_tokens should be (8 + 774) - 400 cached = 382"
     assert result.prompt_tokens_details.cached_tokens == 400