fix: preserve protocol-aware custom guardrail streams

litellm/integrations/custom_guardrail.py

@@ -103,6 +103,7 @@ class CustomGuardrail(CustomLogger):
     use_native_during_call_hook: ClassVar[bool] = False
 
     records_own_guardrail_information: ClassVar[bool] = False
+    supported_streaming_call_types: ClassVar[frozenset[CallTypes]] = frozenset()
 
     def __init__(
         self,
@@ -625,6 +626,9 @@ class CustomGuardrail(CustomLogger):
     def uses_apply_guardrail_interface(self) -> bool:
         return type(self).apply_guardrail is not CustomGuardrail.apply_guardrail
 
+    def supports_streaming_call_type(self, call_type: CallTypes) -> bool:
+        return call_type in self.supported_streaming_call_types
+
     def _deployment_pre_call_target(self) -> "CustomLogger":
         if not self.uses_apply_guardrail_interface():
             return self

litellm/proxy/utils.py

@@ -1766,6 +1766,23 @@ class ProxyLogging:
             return False
         return call_types[0] in NON_OPENAI_STREAM_GUARDRAIL_TRANSLATION_CALL_TYPES
 
+    @staticmethod
+    def _guardrail_supports_streaming_route(
+        guardrail: CustomGuardrail,
+        user_api_key_dict: UserAPIKeyAuth,
+    ) -> bool:
+        from litellm.litellm_core_utils.api_route_to_call_types import (
+            get_call_types_for_route,
+        )
+
+        route: Final = user_api_key_dict.request_route
+        if not route:
+            return False
+        call_types: Final = get_call_types_for_route(route)
+        if not call_types:
+            return False
+        return any(guardrail.supports_streaming_call_type(call_type) for call_type in call_types)
+
     @staticmethod
     def has_post_call_response_headers_callbacks() -> bool:
         return ProxyLogging._callback_capabilities().has_post_call_response_headers
@@ -2744,6 +2761,10 @@ class ProxyLogging:
                     and isinstance(resolved_callback, CustomGuardrail)
                     and resolved_callback.uses_apply_guardrail_interface()
                     and not resolved_callback.mask_response_content
+                    and not ProxyLogging._guardrail_supports_streaming_route(
+                        resolved_callback,
+                        user_api_key_dict,
+                    )
                 )
                 else kind
             )