Merge pull request #41689 from BerriAI/litellm_responses_bridge_strip_internal_kwargs

fix(responses): keep the addressed response id off bridged provider requests

litellm/proxy/hooks/responses_id_security.py

@@ -22,7 +22,7 @@ from litellm.types.llms.openai import (
     BaseLiteLLMOpenAIResponseObject,
     ResponsesAPIResponse,
 )
-from litellm.types.utils import CallTypesLiteral, LLMResponseTypes, SpecialEnums
+from litellm.types.utils import ADDRESSED_RESPONSE_ID_FIELD, CallTypesLiteral, LLMResponseTypes, SpecialEnums
 
 if TYPE_CHECKING:
     from litellm.caching.caching import DualCache
@@ -32,7 +32,6 @@ if TYPE_CHECKING:
 _RESPONSES_API_PROVIDER_PREFIX: Final = "/openai"
 _RESPONSES_API_CREATE_ROUTES: Final = frozenset({"/v1/responses", "/responses"})
 
-_ADDRESSED_RESPONSE_ID_KEY: Final = "_litellm_addressed_response_id"
 _UNMANAGED_RESPONSE_ID_DETAIL: Final = (
     "Forbidden. This response id was not issued by this proxy, so the proxy cannot tell who owns it. "
     "To let keys address responses this proxy did not issue, set "
@@ -132,7 +131,7 @@ class ResponsesIDSecurity(CustomLogger):
         if call_type not in responses_api_call_types:
             return None
         addressed_id_field: Final = "previous_response_id" if call_type == "aresponses" else "response_id"
-        retained_id: Final = data.get(_ADDRESSED_RESPONSE_ID_KEY)
+        retained_id: Final = data.get(ADDRESSED_RESPONSE_ID_FIELD)
         addressed_id: Final = (
             retained_id if isinstance(retained_id, str) and retained_id else data.get(addressed_id_field)
         )
@@ -140,7 +139,7 @@ class ResponsesIDSecurity(CustomLogger):
             return data
         authorized_id: Final = self._authorize_response_id(addressed_id, user_api_key_dict)
         data[addressed_id_field] = authorized_id
-        data[_ADDRESSED_RESPONSE_ID_KEY] = addressed_id
+        data[ADDRESSED_RESPONSE_ID_FIELD] = addressed_id
         return data
 
     def _authorize_response_id(

litellm/types/utils.py

@@ -3754,6 +3754,8 @@ agentic_loop_internal_litellm_params: Final = [
 # the provider.
 TRUSTED_CALLBACK_VARS_FIELD: Final = "litellm_trusted_callback_vars"
 
+ADDRESSED_RESPONSE_ID_FIELD: Final = "_litellm_addressed_response_id"
+
 # Bedrock managed-batch deployment config, read from litellm_params by the batch and
 # files transformations. Listed for the same reason as the fields above: these sit on
 # a deployment that also serves chat, so leaking them into extra_body makes Bedrock
@@ -3768,7 +3770,7 @@ bedrock_batch_litellm_params: Final = (
 
 all_litellm_params = (
     agentic_loop_internal_litellm_params
-    + [TRUSTED_CALLBACK_VARS_FIELD, *bedrock_batch_litellm_params]
+    + [TRUSTED_CALLBACK_VARS_FIELD, ADDRESSED_RESPONSE_ID_FIELD, *bedrock_batch_litellm_params]
     + [
         "metadata",
         "litellm_metadata",

tests/test_litellm/responses/litellm_completion_transformation/test_handler.py

@@ -12,14 +12,21 @@ capture the forwarded kwargs; if the flag-setting line is removed the captured
 kwargs lack the flag and these tests fail.
 """
 
+import json
+from collections.abc import Mapping
+from typing import Final
 from unittest.mock import patch
 
+import httpx
 import pytest
 
-
+import litellm
+from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
 from litellm.responses.litellm_completion_transformation.handler import (
     LiteLLMCompletionTransformationHandler,
 )
+from litellm.types.llms.openai import ResponsesAPIResponse
+from litellm.types.utils import ADDRESSED_RESPONSE_ID_FIELD
 
 
 class _StopForwarding(Exception):
@@ -170,3 +177,49 @@ async def test_async_fallback_returns_hoisted_nested_custom_tool_call_as_custom_
 
     tool_calls = [(item.type, item.name, item.input) for item in response.output if item.type == "custom_tool_call"]
     assert tool_calls == [("custom_tool_call", "exec", "ls")]
+
+
+class _RecordingAnthropicHandler:
+    def __init__(self, reply: Mapping[str, object]) -> None:
+        self.reply: Final = reply
+        self.request_body: Mapping[str, object] | None = None
+
+    def __call__(self, request: httpx.Request) -> httpx.Response:
+        self.request_body = json.loads(request.content)
+        return httpx.Response(200, json=dict(self.reply), request=request)
+
+
+_ANTHROPIC_MESSAGE_PAYLOAD: Final = {
+    "id": "msg_turn_two",
+    "type": "message",
+    "role": "assistant",
+    "model": "claude-sonnet-4-6",
+    "content": [{"type": "text", "text": "14"}],
+    "stop_reason": "end_turn",
+    "stop_sequence": None,
+    "usage": {"input_tokens": 12, "output_tokens": 1},
+}
+
+
+@pytest.mark.asyncio
+async def test_bridged_follow_up_turn_keeps_the_addressed_response_id_off_the_provider_body():
+    provider: Final = _RecordingAnthropicHandler(_ANTHROPIC_MESSAGE_PAYLOAD)
+    client: Final = AsyncHTTPHandler()
+    client.client = httpx.AsyncClient(transport=httpx.MockTransport(provider))
+
+    response = await litellm.aresponses(
+        model="azure_ai/claude-sonnet-4-6",
+        api_base="https://fake-foundry-resource.services.ai.azure.com",
+        api_key="fake-api-key",
+        input="Double it",
+        previous_response_id="resp_turn_one",
+        client=client,
+        **{ADDRESSED_RESPONSE_ID_FIELD: "resp_turn_one"},
+    )
+
+    assert provider.request_body is not None, "the bridged turn never reached the provider"
+    assert ADDRESSED_RESPONSE_ID_FIELD not in provider.request_body, (
+        f"the addressed response id reached the provider body: {sorted(provider.request_body)}"
+    )
+    assert isinstance(response, ResponsesAPIResponse)
+    assert [item.type for item in response.output] == ["message"]

tests/test_litellm/test_utils.py

@@ -46,6 +46,7 @@ from litellm.types.utils import (
     PromptTokensDetailsWrapper,
     StreamingChoices,
     Usage,
+    ADDRESSED_RESPONSE_ID_FIELD,
     all_litellm_params,
     bedrock_batch_litellm_params,
 )
@@ -4787,6 +4788,20 @@ def test_get_litellm_params_keys_never_reach_the_provider():
     )
 
 
+def test_addressed_response_id_never_reaches_the_provider():
+    kwargs = {
+        "a_real_provider_specific_param": 1,
+        ADDRESSED_RESPONSE_ID_FIELD: "resp_addressed-by-the-client",
+    }
+
+    non_default = get_non_default_completion_params(kwargs)
+
+    assert non_default == {"a_real_provider_specific_param": 1}, (
+        "the addressed response id leaked into the provider params: "
+        f"{sorted(set(non_default) - {'a_real_provider_specific_param'})}"
+    )
+
+
 def test_bedrock_batch_params_never_reach_the_provider():
     """A Bedrock managed-batch deployment carries aws_batch_role_arn / s3_* /
     bedrock_tags in its litellm_params, and the same deployment also serves chat.