fix: handle malformed JSON in LiteLLM tool call arguments gracefully

src/google/adk/models/lite_llm.py

@@ -2178,9 +2178,18 @@ def _message_to_generate_content_response(
     for tool_call in tool_calls:
       if tool_call.type == "function":
         thought_signature = _extract_thought_signature_from_tool_call(tool_call)
+        try:
+          args = _parse_tool_call_arguments(tool_call.function.arguments)
+        except json.JSONDecodeError:
+          logger.warning(
+              "Skipping tool call %s (id=REDACTED) with malformed JSON"
+              " arguments.",
+              tool_call.function.name,
+          )
+          continue
         part = types.Part.from_function_call(
             name=tool_call.function.name,
-            args=_parse_tool_call_arguments(tool_call.function.arguments),
+            args=args,
         )
         part.function_call.id = tool_call.id
         if thought_signature:

tests/unittests/models/test_litellm.py

@@ -6760,3 +6760,117 @@ async def test_generate_content_async_omits_tool_choice_when_functions_override(
   _, kwargs = mock_acompletion.call_args
   assert kwargs.get("tools") is None
   assert "tool_choice" not in kwargs
+
+
+def test_message_to_generate_content_response_malformed_tool_call_json():
+  """Malformed tool call arguments should be skipped, not crash."""
+  message = ChatCompletionAssistantMessage(
+      role="assistant",
+      content=None,
+      tool_calls=[
+          ChatCompletionMessageToolCall(
+              type="function",
+              id="call_bad",
+              function=Function(
+                  name="broken_tool",
+                  arguments='{"a":"unterminated',
+              ),
+          ),
+          ChatCompletionMessageToolCall(
+              type="function",
+              id="call_good",
+              function=Function(
+                  name="valid_tool",
+                  arguments='{"key": "value"}',
+              ),
+          ),
+      ],
+  )
+
+  response = _message_to_generate_content_response(message)
+  assert response.content.role == "model"
+  # The malformed tool call should be skipped; only the valid one remains
+  function_parts = [
+      p for p in response.content.parts if p.function_call is not None
+  ]
+  assert len(function_parts) == 1
+  assert function_parts[0].function_call.name == "valid_tool"
+  assert function_parts[0].function_call.id == "call_good"
+
+
+def test_message_to_generate_content_response_all_malformed_tool_calls():
+  """When all tool calls have malformed JSON, response should have no parts."""
+  message = ChatCompletionAssistantMessage(
+      role="assistant",
+      content=None,
+      tool_calls=[
+          ChatCompletionMessageToolCall(
+              type="function",
+              id="call_1",
+              function=Function(
+                  name="broken",
+                  arguments="not json at all",
+              ),
+          ),
+      ],
+  )
+
+  response = _message_to_generate_content_response(message)
+  assert response.content.role == "model"
+  assert len(response.content.parts) == 0
+
+
+def test_message_to_generate_content_response_tool_call_none_arguments():
+  """Tool call with arguments=None should produce a function call with empty args."""
+  message = ChatCompletionAssistantMessage(
+      role="assistant",
+      content=None,
+      tool_calls=[
+          ChatCompletionMessageToolCall(
+              type="function",
+              id="call_none_args",
+              function=Function(
+                  name="no_args_tool",
+                  arguments=None,
+              ),
+          ),
+      ],
+  )
+
+  response = _message_to_generate_content_response(message)
+  assert response.content.role == "model"
+  function_parts = [
+      p for p in response.content.parts if p.function_call is not None
+  ]
+  assert len(function_parts) == 1
+  assert function_parts[0].function_call.name == "no_args_tool"
+  assert function_parts[0].function_call.id == "call_none_args"
+  assert dict(function_parts[0].function_call.args) == {}
+
+
+def test_message_to_generate_content_response_tool_call_empty_string_arguments():
+  """Tool call with arguments='' should produce a function call with empty args."""
+  message = ChatCompletionAssistantMessage(
+      role="assistant",
+      content=None,
+      tool_calls=[
+          ChatCompletionMessageToolCall(
+              type="function",
+              id="call_empty_args",
+              function=Function(
+                  name="empty_args_tool",
+                  arguments="",
+              ),
+          ),
+      ],
+  )
+
+  response = _message_to_generate_content_response(message)
+  assert response.content.role == "model"
+  function_parts = [
+      p for p in response.content.parts if p.function_call is not None
+  ]
+  assert len(function_parts) == 1
+  assert function_parts[0].function_call.name == "empty_args_tool"
+  assert function_parts[0].function_call.id == "call_empty_args"
+  assert dict(function_parts[0].function_call.args) == {}