diff --git a/src/google/adk/models/lite_llm.py b/src/google/adk/models/lite_llm.py
index 162db05945..0ca03ce439 100644
--- a/src/google/adk/models/lite_llm.py
+++ b/src/google/adk/models/lite_llm.py
@@ -823,11 +823,15 @@ def _model_response_to_generate_content_response(
 ) -> LlmResponse:
   """Converts a litellm response to LlmResponse. Also adds usage metadata.
 
+  When a response has no message (e.g., turn ends with only tool calls),
+  returns an empty LlmResponse with finish_reason and usage_metadata if
+  available, instead of raising ValueError.
+
   Args:
     response: The model response.
 
   Returns:
-    The LlmResponse.
+    The LlmResponse. May have empty content if message is None.
   """
 
   message = None
@@ -837,17 +841,25 @@ def _model_response_to_generate_content_response(
     message = first_choice.get("message", None)
     finish_reason = first_choice.get("finish_reason", None)
 
-  if not message:
-    raise ValueError("No message in response")
+  # Handle case where message is None or empty (valid when turn ends with
+  # tool calls only). Create empty LlmResponse instead of raising error.
+  if message:
+    thought_parts = _convert_reasoning_value_to_parts(
+        _extract_reasoning_value(message)
+    )
+    llm_response = _message_to_generate_content_response(
+        message,
+        model_version=response.model,
+        thought_parts=thought_parts or None,
+    )
+  else:
+    # Create empty LlmResponse when message is None or empty
+    llm_response = LlmResponse(
+        content=types.Content(role="model", parts=[]),
+        model_version=response.model,
+    )
 
-  thought_parts = _convert_reasoning_value_to_parts(
-      _extract_reasoning_value(message)
-  )
-  llm_response = _message_to_generate_content_response(
-      message,
-      model_version=response.model,
-      thought_parts=thought_parts or None,
-  )
+  # Common logic for finish_reason and usage_metadata
   if finish_reason:
     # If LiteLLM already provides a FinishReason enum (e.g., for Gemini), use
     # it directly. Otherwise, map the finish_reason string to the enum.
diff --git a/tests/unittests/models/test_litellm.py b/tests/unittests/models/test_litellm.py
index f65fc77a61..82e7ac5b5b 100644
--- a/tests/unittests/models/test_litellm.py
+++ b/tests/unittests/models/test_litellm.py
@@ -2610,6 +2610,84 @@ async def test_finish_reason_propagation(
   mock_acompletion.assert_called_once()
 
 
+def test_model_response_to_generate_content_response_no_message_with_finish_reason():
+  """Test response with no message but finish_reason returns empty LlmResponse.
+
+  This test covers issue #3618: when a turn ends with tool calls and no final
+  message, we should return an empty LlmResponse instead of raising ValueError.
+  """
+  response = ModelResponse(
+      model="test_model",
+      choices=[{
+          "finish_reason": "tool_calls",
+          # message is missing/None
+      }],
+      usage={
+          "prompt_tokens": 10,
+          "completion_tokens": 5,
+          "total_tokens": 15,
+      },
+  )
+
+  llm_response = _model_response_to_generate_content_response(response)
+
+  # Should return empty LlmResponse, not raise ValueError
+  assert llm_response.content is not None
+  assert llm_response.content.role == "model"
+  assert len(llm_response.content.parts) == 0
+  # tool_calls maps to STOP
+  assert llm_response.finish_reason == types.FinishReason.STOP
+  assert llm_response.usage_metadata is not None
+  assert llm_response.usage_metadata.prompt_token_count == 10
+  assert llm_response.usage_metadata.candidates_token_count == 5
+  assert llm_response.model_version == "test_model"
+
+
+def test_model_response_to_generate_content_response_no_message_no_finish_reason():
+  """Test response with no message and no finish_reason returns empty LlmResponse."""
+  response = ModelResponse(
+      model="test_model",
+      choices=[{
+          # Both message and finish_reason are missing
+      }],
+  )
+
+  llm_response = _model_response_to_generate_content_response(response)
+
+  # Should return empty LlmResponse, not raise ValueError
+  assert llm_response.content is not None
+  assert llm_response.content.role == "model"
+  assert len(llm_response.content.parts) == 0
+  # finish_reason may be None or have a default value - the important thing
+  # is that we don't raise ValueError
+  assert llm_response.model_version == "test_model"
+
+
+def test_model_response_to_generate_content_response_empty_message_dict():
+  """Test response with empty message dict returns empty LlmResponse."""
+  response = ModelResponse(
+      model="test_model",
+      choices=[{
+          "message": {},  # Empty dict is falsy
+          "finish_reason": "stop",
+      }],
+      usage={
+          "prompt_tokens": 5,
+          "completion_tokens": 3,
+          "total_tokens": 8,
+      },
+  )
+
+  llm_response = _model_response_to_generate_content_response(response)
+
+  # Should return empty LlmResponse, not raise ValueError
+  assert llm_response.content is not None
+  assert llm_response.content.role == "model"
+  assert len(llm_response.content.parts) == 0
+  assert llm_response.finish_reason == types.FinishReason.STOP
+  assert llm_response.usage_metadata is not None
+
+
 @pytest.mark.asyncio
 async def test_finish_reason_unknown_maps_to_other(
     mock_acompletion, lite_llm_instance