fix(proxy): retain the addressed comparison response id and authorize only the effective prompt_cache_options

The hook keeps the encrypted comparison id under _litellm_addressed_comparison_response_id, a registered LiteLLM-owned kwarg that never reaches a provider, so a repeated hook run re-checks ownership for the current caller. When extra_body carries prompt_cache_options it replaces the top-level value as a whole, so only that effective value is authorized and rewritten

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/proxy/hooks/responses_id_security.py

@@ -24,7 +24,13 @@ from litellm.types.llms.openai import (
     BaseLiteLLMOpenAIResponseObject,
     ResponsesAPIResponse,
 )
-from litellm.types.utils import ADDRESSED_RESPONSE_ID_FIELD, CallTypesLiteral, LLMResponseTypes, SpecialEnums
+from litellm.types.utils import (
+    ADDRESSED_COMPARISON_RESPONSE_ID_FIELD,
+    ADDRESSED_RESPONSE_ID_FIELD,
+    CallTypesLiteral,
+    LLMResponseTypes,
+    SpecialEnums,
+)
 
 if TYPE_CHECKING:
     from litellm.caching.caching import DualCache
@@ -155,37 +161,29 @@ class ResponsesIDSecurity(CustomLogger):
         data: Mapping[str, object],
         user_api_key_dict: "UserAPIKeyAuth",
     ) -> Mapping[str, object]:
-        options: Final = self._authorized_prompt_cache_options(data.get("prompt_cache_options"), user_api_key_dict)
         extra_body: Final = as_str_mapping(data.get("extra_body"))
-        extra_body_options: Final = (
-            None
-            if extra_body is None
-            else self._authorized_prompt_cache_options(extra_body.get("prompt_cache_options"), user_api_key_dict)
+        overriding_extra_body: Final = (
+            extra_body if extra_body is not None and "prompt_cache_options" in extra_body else None
         )
-        options_field: Final[Mapping[str, object]] = {} if options is None else {"prompt_cache_options": options}
-        extra_body_field: Final[Mapping[str, object]] = (
-            {}
-            if extra_body is None or extra_body_options is None
-            else {"extra_body": {**extra_body, "prompt_cache_options": extra_body_options}}
+        options: Final = as_str_mapping(
+            data.get("prompt_cache_options")
+            if overriding_extra_body is None
+            else overriding_extra_body.get("prompt_cache_options")
         )
-        return {**options_field, **extra_body_field}
-
-    def _authorized_prompt_cache_options(
-        self,
-        prompt_cache_options: object,
-        user_api_key_dict: "UserAPIKeyAuth",
-    ) -> Mapping[str, object] | None:
-        options: Final = as_str_mapping(prompt_cache_options)
-        comparison_response_id: Final = None if options is None else options.get("comparison_response_id")
-        if (
-            options is None
-            or not isinstance(comparison_response_id, str)
-            or not self._is_encrypted_response_id(comparison_response_id)
-        ):
-            return None
-        original_response_id, user_id, team_id = self._decrypt_response_id(comparison_response_id)
+        retained_id: Final = data.get(ADDRESSED_COMPARISON_RESPONSE_ID_FIELD)
+        requested_id: Final = None if options is None else options.get("comparison_response_id")
+        addressed_id: Final = retained_id if isinstance(retained_id, str) and retained_id else requested_id
+        if options is None or not isinstance(addressed_id, str) or not self._is_encrypted_response_id(addressed_id):
+            return {}
+        original_response_id, user_id, team_id = self._decrypt_response_id(addressed_id)
         self.check_user_access_to_response_id(user_id, team_id, user_api_key_dict)
-        return {**options, "comparison_response_id": original_response_id}
+        authorized_options: Final = {**options, "comparison_response_id": original_response_id}
+        authorized_field: Final[Mapping[str, object]] = (
+            {"prompt_cache_options": authorized_options}
+            if overriding_extra_body is None
+            else {"extra_body": {**overriding_extra_body, "prompt_cache_options": authorized_options}}
+        )
+        return {**authorized_field, ADDRESSED_COMPARISON_RESPONSE_ID_FIELD: addressed_id}
 
     def authorize_response_id(
         self,

litellm/types/litellm_params.py

@@ -55,6 +55,7 @@ RoutingStrategyName: TypeAlias = Literal[
 
 TRUSTED_CALLBACK_VARS_FIELD: Final = "litellm_trusted_callback_vars"
 ADDRESSED_RESPONSE_ID_FIELD: Final = "_litellm_addressed_response_id"
+ADDRESSED_COMPARISON_RESPONSE_ID_FIELD: Final = "_litellm_addressed_comparison_response_id"
 STREAM_USAGE_INJECTED_FIELD: Final = "_litellm_stream_usage_injected"
 
 WIRE_NAME: Final = "wire_name"
@@ -406,6 +407,9 @@ class ProxyRequestState:
     secret_fields: "SecretFields | None" = None
     trusted_callback_vars: Mapping[str, str] | None = field(default=None, metadata=wire(TRUSTED_CALLBACK_VARS_FIELD))
     addressed_response_id: str | None = field(default=None, metadata=wire(ADDRESSED_RESPONSE_ID_FIELD))
+    addressed_comparison_response_id: str | None = field(
+        default=None, metadata=wire(ADDRESSED_COMPARISON_RESPONSE_ID_FIELD)
+    )
     strip_stream_usage: bool | None = field(default=None, metadata=wire("_litellm_strip_stream_usage"))
     stream_usage_injected: bool | None = field(default=None, metadata=wire(STREAM_USAGE_INJECTED_FIELD))
     client_side_timeout: bool | None = None

litellm/types/utils.py

@@ -4334,6 +4334,7 @@ bedrock_batch_litellm_params: Final = BEDROCK_BATCH_KWARG_NAMES
 
 TRUSTED_CALLBACK_VARS_FIELD: Final = _litellm_params.TRUSTED_CALLBACK_VARS_FIELD
 ADDRESSED_RESPONSE_ID_FIELD: Final = _litellm_params.ADDRESSED_RESPONSE_ID_FIELD
+ADDRESSED_COMPARISON_RESPONSE_ID_FIELD: Final = _litellm_params.ADDRESSED_COMPARISON_RESPONSE_ID_FIELD
 
 anthropic_wif_litellm_params: Final = tuple(sorted(ANTHROPIC_WIF_KWARGS_KEYS))
 openai_wif_litellm_params: Final = tuple(sorted(OPENAI_WIF_KWARGS_KEYS))