Critical2

High2

Medium0

#1high+4−3

Add request_data to A2A output handler

📄 litellm/llms/a2a/chat/guardrail_translation/handler.py

Adds the request_data parameter to process_output_response and merges it into local_request_data via dict unpacking, preserving any input-phase context.

111111         guardrail_to_apply: "CustomGuardrail",

112112         litellm_logging_obj: Optional["LiteLLMLoggingObj"] = None,

113113         user_api_key_dict: Optional["UserAPIKeyAuth"] = None,

114+        request_data: Optional[dict] = None,

114115     ) -> Any:

115116         """

116117         Process A2A output response by applying guardrails to text content.

⋯ 50 lines hidden

167168         # Step 2: Apply guardrail to all texts in batch

168169         # Create a local_request_data dict with response info

169−        request_data: dict = {"response": response_dict}

170+        local_request_data: dict = {**(request_data or {}), "response": response_dict}

💬Conversation · 2 messages▼

Y

You2 min ago

Why use dict unpacking instead of dict.update()?

AI

AI Assistant1 min ago

The `{(request_data or {}), "response": response_dict}` pattern creates a new dict rather than mutating `request_data`. This matters because `request_data` is shared across multiple handler invocations — mutating it would cause cross-contamination between guardrail handlers.

Y

171172         # Add user API key metadata with prefixed keys

172173         user_metadata = self.transform_user_api_key_dict_to_metadata(user_api_key_dict)

173174         if user_metadata:

174175             local_request_data["litellm_metadata"] = user_metadata

Hover a line and click + to ask about specific lines

#5critical+8−3

Update base class abstract signatures

📄 litellm/llms/base_llm/guardrail_translation/base_translation.py

Updates the abstract BaseTranslation to declare request_data on both process_output_response and process_output_streaming_response, defining the interface contract all subclasses must follow.

7373         guardrail_to_apply: "CustomGuardrail",

7474         litellm_logging_obj: Optional["LiteLLMLoggingObj"] = None,

7575         user_api_key_dict: Optional["UserAPIKeyAuth"] = None,

76+        request_data: Optional[dict] = None,

7677     ) -> Any:

7778         """

7879         Process output response with guardrails.

8283             litellm_logging_obj: Optional logging object

8384             user_api_key_dict: User API key metadata

8485             request_data: Optional original request data dict from the proxy

86+        """

8587         pass

Hover a line and click + to ask about specific lines

AI

Ask about Chunk #5

AI Summary

Updates the abstract BaseTranslation to declare request_data on both process_output_response and process_output_streaming_response, defining the interface contract all subclasses must follow.

RELATED →#1#2#3#8#20

Y

You3 min ago

Why is this chunk marked as critical?

AI

AI Assistant2 min ago

This chunk modifies the abstract base class `BaseTranslation`, which defines the interface contract for all guardrail handlers. Every concrete handler must conform to this signature. Changing the abstract interface is critical because:

Any subclass missing `request_data` breaks Liskov substitution
It affects 10+ handler implementations
A missing parameter means silent loss of input context

Y

#8high+6−4

Update OpenAI handler + use local_request_data

📄 litellm/llms/openai/chat/guardrail_translation/handler.py

Adds request_data to the OpenAI chat handler and fixes the streaming path to pass local_request_data instead of the raw parameter, ensuring enriched context reaches the guardrail.

260260         guardrail_to_apply: "CustomGuardrail",

261261         litellm_logging_obj: Optional[Any] = None,

262262         user_api_key_dict: Optional[Any] = None,

263+        request_data: Optional[dict] = None,

263264     ) -> Any:

264265         """

265266         Process output response by applying guardrails to text content.

⋯ 43 lines hidden

308309         # Step 2: Apply guardrail to all texts and tool calls in batch

309310         if texts_to_check or tool_calls_to_check:

310311             # Create a local_request_data dict

311−            request_data: dict = {"response": response}

312+            local_request_data: dict = {**(request_data or {}), "response": response}

313314             # Add user API key metadata with prefixed keys

314315             user_metadata = self.transform_user_api_key_dict_to_metadata(

315316                 user_api_key_dict

316317             )

Hover a line and click + to ask about specific lines

#20critical+4−0

Forward request_data from unified guardrail

📄 litellm/proxy/guardrails/guardrail_hooks/unified_guardrail/unified_guardrail.py

The entry point: passes the full request `data` dict into process_output_response and streaming hooks, making it available to all downstream handlers.

247247                 guardrail_to_apply=guardrail_to_apply,

248248                 litellm_logging_obj=data.get("litellm_logging_obj"),

249249                 user_api_key_dict=user_api_key_dict,

250+                request_data=data,💬 4

💬4 comments— click to expand▶

250251             )

251252             # Add guardrail to applied guardrails header

Hover a line and click + to ask about specific lines

fix(guardrail): pass request_data to output response handlers

Add request_data to A2A output handler

Update base class abstract signatures

Update OpenAI handler + use local_request_data

Forward request_data from unified guardrail