fix: resolve Gemini tool results by call id

litellm/llms/vertex_ai/gemini/transformation.py

@@ -307,6 +307,7 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
     contents: List[ContentType] = []
 
     last_message_with_tool_calls = None
+    message_by_tool_call_id: Dict[str, ChatCompletionAssistantMessage] = {}
 
     msg_i = 0
     tool_call_responses = []
@@ -527,9 +528,10 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
                                     assistant_content.append(_part)
 
                 ## HANDLE ASSISTANT FUNCTION CALL
+                tool_calls = assistant_msg.get("tool_calls")
+                function_call = assistant_msg.get("function_call")
                 if (
-                    assistant_msg.get("tool_calls", []) is not None
-                    or assistant_msg.get("function_call") is not None
+                    tool_calls or function_call is not None
                 ):  # support assistant tool invoke conversion
                     gemini_tool_call_parts = convert_to_gemini_tool_call_invoke(
                         assistant_msg, model=model
@@ -542,6 +544,11 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
                             excluded_keys=["thoughtSignature"],
                         ):
                             assistant_content.append(gemini_tool_call_part)
+                    if tool_calls:
+                        for tool_call in tool_calls:
+                            tool_call_id = tool_call.get("id")
+                            if tool_call_id:
+                                message_by_tool_call_id[tool_call_id] = assistant_msg
                     last_message_with_tool_calls = assistant_msg
 
                 ## HANDLE SERVER-SIDE TOOL INVOCATIONS (context circulation)
@@ -590,8 +597,17 @@ def _gemini_convert_messages_with_history(  # noqa: PLR0915
                 msg_i < len(messages)
                 and messages[msg_i]["role"] in tool_call_message_roles
             ):
+                tool_result_message = messages[msg_i]
+                tool_call_id = tool_result_message.get("tool_call_id")
+                message_with_tool_call = (
+                    message_by_tool_call_id.get(
+                        tool_call_id, last_message_with_tool_calls
+                    )
+                    if tool_call_id is not None
+                    else last_message_with_tool_calls
+                )
                 _part = convert_to_gemini_tool_call_result(
-                    messages[msg_i], last_message_with_tool_calls  # type: ignore
+                    tool_result_message, message_with_tool_call  # type: ignore
                 )
                 msg_i += 1
                 # Handle both single part and list of parts (for Computer Use with images)

tests/litellm/llms/vertex_ai/gemini/test_transformation.py

@@ -10,8 +10,6 @@ from litellm.llms.vertex_ai.gemini import transformation
 from litellm.llms.vertex_ai.gemini.vertex_and_google_ai_studio_gemini import (
     VertexGeminiConfig,
 )
-from litellm.types.llms import openai
-from litellm.types import completion
 from litellm.types.llms.vertex_ai import RequestBody
 
 
@@ -95,6 +93,65 @@ async def test__transform_request_body_metadata():
     }
 
 
+def test_gemini_tool_result_resolves_after_text_only_assistant_message():
+    messages = [
+        {"role": "user", "content": "Read the file before answering."},
+        {
+            "role": "assistant",
+            "content": None,
+            "tool_calls": [
+                {
+                    "id": "call_read_file",
+                    "type": "function",
+                    "function": {
+                        "name": "read_file",
+                        "arguments": '{"path": "/tmp/example.txt"}',
+                    },
+                }
+            ],
+        },
+        {
+            "role": "assistant",
+            "content": "I will inspect the file and then summarize it.",
+            "tool_calls": [],
+        },
+        {
+            "role": "tool",
+            "tool_call_id": "call_read_file",
+            "content": '{"contents": "hello from the file"}',
+        },
+    ]
+
+    contents = transformation._gemini_convert_messages_with_history(messages)
+
+    assert contents == [
+        {"parts": [{"text": "Read the file before answering."}], "role": "user"},
+        {
+            "parts": [
+                {
+                    "function_call": {
+                        "name": "read_file",
+                        "args": {"path": "/tmp/example.txt"},
+                    }
+                },
+                {"text": "I will inspect the file and then summarize it."},
+            ],
+            "role": "model",
+        },
+        {
+            "parts": [
+                {
+                    "function_response": {
+                        "name": "read_file",
+                        "response": {"contents": "hello from the file"},
+                    }
+                }
+            ],
+            "role": "user",
+        },
+    ]
+
+
 @pytest.mark.asyncio
 async def test__transform_request_body_labels_and_metadata():
     """