litellm - 💡(How to fix) Fix bug: Guardrail pre_call_hook crashes with KeyError: 'name' on /v1/messages endpoint when tools are present [1 comments, 2 participants]

File "/app/litellm/proxy/anthropic_endpoints/endpoints.py", line 53, in anthropic_response result = await base_llm_response_processor.base_process_llm_request(...)

File "/app/litellm/proxy/common_request_processing.py", line 837, in base_process_llm_request self.data, logging_obj = await self.common_processing_pre_call_logic(...)

File "/app/litellm/proxy/common_request_processing.py", line 695, in common_processing_pre_call_logic self.data = await proxy_logging_obj.pre_call_hook( user_api_key_dict=user_api_key_dict, data=self.data, call_type=route_type )

File "/app/litellm/proxy/utils.py", line 1351, in pre_call_hook result = await self._process_guardrail_callback(...)

File "/app/litellm/proxy/utils.py", line 1019, in _process_guardrail_callback response = await self._execute_guardrail_hook(...)

File "/app/litellm/proxy/utils.py", line 894, in _execute_guardrail_hook return await target.async_pre_call_hook(...)

File "/app/litellm/proxy/guardrails/guardrail_hooks/unified_guardrail/unified_guardrail.py", line 124, in async_pre_call_hook data = await endpoint_translation.process_input_messages(...)

File "/app/litellm/llms/anthropic/chat/guardrail_translation/handler.py", line 79, in process_input_messages LiteLLMAnthropicMessagesAdapter().translate_anthropic_to_openai( anthropic_message_request=cast(AnthropicMessagesRequest, data.copy()) )

File "/app/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py", line 910, in translate_anthropic_to_openai new_kwargs["tools"], tool_name_mapping = self.translate_anthropic_tools_to_openai( tools=cast(List[AllAnthropicToolsValues], regular_tools), model=new_kwargs.get("model"), )

File "/app/litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py", line 736, in translate_anthropic_tools_to_openai original_name = tool["name"] ~~~~^^^^^^^^ KeyError: 'name'

Fix Plan

To resolve the KeyError: 'name' issue in the unified guardrail's async_pre_call_hook, we need to ensure that only the original Anthropic request fields are passed to translate_anthropic_to_openai(). We can achieve this by filtering out proxy-internal fields before creating the AnthropicMessagesRequest object.

Here are the concrete steps:

• Modify the process_input_messages() function in guardrail_translation/handler.py to filter out proxy-internal fields:

def process_input_messages(self, data):
    # Filter out proxy-internal fields
    original_request_fields = {
        "model": data.get("model"),
        "messages": data.get("messages"),
        "tools": data.get("tools"),
        # Add other original Anthropic request fields as needed
    }
    return LiteLLMAnthropicMessagesAdapter().translate_anthropic_to_openai(
        anthropic_message_request=cast(AnthropicMessagesRequest, original_request_fields)
    )

• Alternatively, modify the translate_anthropic_to_openai() function in adapters/transformation.py to ignore proxy-internal fields when creating the AnthropicMessagesRequest object:

def translate_anthropic_to_openai(self, anthropic_message_request):
    # Ignore proxy-internal fields
    original_request_fields = {
        key: value for key, value in anthropic_message_request.items()
        if key in ["model", "messages", "tools"]  # Add other original Anthropic request fields as needed
    }
    request_body = AnthropicMessagesRequest(**original_request_fields)
    # ...

Verification

To verify that the fix worked, send a /v1/messages request with tools and check that the response does not contain a 500: 'name' error. You can use the same curl command provided in the issue description:

curl -s -X POST 'http://localhost:4000/v1/messages' \
  -H 'Content-Type: application/json' \
  -H 'x-litellm-api-key: sk-...' \
  -H 'anthropic-version: 2023-06-01' \
  -d '{
    "model": "claude-sonnet-4-6",
    "max_tokens": 100,
    "tools": [{"name": "get_weather", "description": "Get weather", "input_schema": {"type": "object", "properties": {"city": {"type": "string"}}, "required": ["city"]}}],
    "messages": [{"role": "user", "content": "What is the weather in London?"}]
  }'

Extra Tips

To prevent similar issues in the future, consider adding input validation and filtering to ensure that only expected fields are passed to sensitive functions. Additionally, review the