fix(responses): encrypt replayed stream ids and resume Azure AI Foundry streams

A replayed background stream now gets the same response id encryption as a
created one, so the key that made the response can reuse the id. Azure AI
Foundry deployments resume streams like Azure ones, and the chat-bridged
Responses iterator closes its wrapped stream instead of raising.

litellm/llms/base_llm/responses/transformation.py

@@ -70,7 +70,7 @@ class BaseResponsesAPIConfig(ABC):
         return False
 
     def supports_streamed_retrieval(self) -> bool:
-        return self.custom_llm_provider in (LlmProviders.OPENAI, LlmProviders.AZURE)
+        return self.custom_llm_provider in (LlmProviders.OPENAI, LlmProviders.AZURE, LlmProviders.AZURE_AI)
 
     def sign_request(
         self,

litellm/proxy/hooks/responses_id_security.py

@@ -31,7 +31,7 @@ if TYPE_CHECKING:
 
 
 _RESPONSES_API_PROVIDER_PREFIX: Final = "/openai"
-_RESPONSES_API_CREATE_ROUTES: Final = frozenset({"/v1/responses", "/responses"})
+_RESPONSES_API_ROUTES: Final = frozenset({"/v1/responses", "/responses"})
 _ADDRESSED_RESPONSE_ID_FIELD_BY_CALL_TYPE: Final[Mapping[str, str]] = MappingProxyType(
     {
         "aresponses": "previous_response_id",
@@ -104,7 +104,7 @@ def _rewrite_advertised_id(
     return event
 
 
-def _is_responses_api_create_route(request_route: str | None) -> bool:
+def _streams_response_ids(request_route: str | None) -> bool:
     if request_route is None:
         return False
     canonical: Final = (
@@ -112,7 +112,8 @@ def _is_responses_api_create_route(request_route: str | None) -> bool:
         if request_route.startswith(_RESPONSES_API_PROVIDER_PREFIX + "/")
         else request_route
     )
-    return canonical in _RESPONSES_API_CREATE_ROUTES
+    parent, _, response_id = canonical.rpartition("/")
+    return canonical in _RESPONSES_API_ROUTES or (parent in _RESPONSES_API_ROUTES and response_id != "")
 
 
 class ResponsesIDSecurity(CustomLogger):
@@ -339,7 +340,7 @@ class ResponsesIDSecurity(CustomLogger):
         async for chunk in response:
             if (
                 isinstance(chunk, BaseLiteLLMOpenAIResponseObject)
-                and _is_responses_api_create_route(user_api_key_dict.request_route)
+                and _streams_response_ids(user_api_key_dict.request_route)
                 and not general_settings.get("disable_responses_id_security", False)
             ):
                 chunk = self._encrypt_response_id(chunk, user_api_key_dict, request_encryption_cache)

litellm/responses/litellm_completion_transformation/streaming_iterator.py

@@ -192,6 +192,9 @@ class LiteLLMCompletionStreamingIterator(ResponsesAPIStreamingIterator):
         self._tool_search_requested: bool = has_server_tool_search(request_tools)
         self._tool_search_call_ids: frozenset[str] = frozenset()
 
+    async def aclose(self) -> None:
+        await self.litellm_custom_stream_wrapper.aclose()
+
     def _is_tool_search_call(self, call_id: str, fn_name: str) -> bool:
         if call_id in self._tool_search_call_ids:
             return True

tests/unit/responses/litellm_completion_transformation/test_streaming_iterator_transformation.py

@@ -702,6 +702,33 @@ async def test_streaming_events_share_the_chat_completion_response_id():
     assert decoded["custom_llm_provider"] == "anthropic"
 
 
+class _ClosableStreamWrapper(_FakeStreamWrapper):
+    def __init__(self, chunks):
+        super().__init__(chunks)
+        self.closed = False
+
+    async def aclose(self):
+        self.closed = True
+
+
+@pytest.mark.asyncio
+async def test_aclose_closes_the_bridged_chat_completion_stream():
+    wrapper: Final = _ClosableStreamWrapper([_chunk("Hello"), _chunk("!", finish_reason="stop")])
+    iterator: Final = LiteLLMCompletionStreamingIterator(
+        model="claude-haiku-4-5",
+        litellm_custom_stream_wrapper=wrapper,
+        request_input="What is the weather in San Francisco?",
+        responses_api_request={},
+        custom_llm_provider="anthropic",
+        litellm_metadata={},
+    )
+
+    await iterator.__anext__()
+    await iterator.aclose()
+
+    assert wrapper.closed is True
+
+
 def test_sync_streaming_events_share_the_chat_completion_response_id():
     iterator = _build_iterator([_chunk("Hello"), _chunk("!", finish_reason="stop")])
 

tests/unit/responses/test_responses_api_lifecycle.py

@@ -118,7 +118,7 @@ async def test_cancel_responses_404_surfaces_openai_api_error_with_exact_url(
     assert str(requests[0].request.url) == expected_url
 
 
-StreamProvider: TypeAlias = Literal["openai", "azure"]
+StreamProvider: TypeAlias = Literal["openai", "azure", "azure_ai"]
 _STORED_RESPONSE_ID: Final = "resp_background"
 _STREAM_HEADERS: Final = {"content-type": "text/event-stream"}
 
@@ -168,6 +168,10 @@ def _retrieve_route(mock_router: respx.MockRouter, provider: StreamProvider) ->
             return mock_router.get(f"https://api.openai.com/v1/responses/{_STORED_RESPONSE_ID}")
         case "azure":
             return mock_router.get(f"https://example-resource.openai.azure.com/openai/responses/{_STORED_RESPONSE_ID}")
+        case "azure_ai":
+            return mock_router.get(
+                f"https://example-resource.services.ai.azure.com/openai/v1/responses/{_STORED_RESPONSE_ID}"
+            )
 
 
 def _retrieve_kwargs(provider: StreamProvider) -> dict[str, object]:
@@ -181,6 +185,12 @@ def _retrieve_kwargs(provider: StreamProvider) -> dict[str, object]:
                 "api_base": "https://example-resource.openai.azure.com",
                 "api_version": "2025-03-01-preview",
             }
+        case "azure_ai":
+            return {
+                "custom_llm_provider": "azure_ai",
+                "api_key": "sk-test",
+                "api_base": "https://example-resource.services.ai.azure.com",
+            }
 
 
 async def _retrieve(sync_mode: bool, **kwargs: object) -> object:
@@ -197,7 +207,7 @@ async def _event_types(stream: object) -> tuple[str, ...]:
 
 
 @pytest.mark.parametrize("sync_mode", (True, False))
-@pytest.mark.parametrize("provider", ("openai", "azure"))
+@pytest.mark.parametrize("provider", ("openai", "azure", "azure_ai"))
 @pytest.mark.parametrize(
     ("resume_kwargs", "expected_query"),
     (

tests/unit/test_responses_id_security.py

@@ -11,7 +11,7 @@ from fastapi import HTTPException
 
 from litellm.proxy.hooks.responses_id_security import (
     ResponsesIDSecurity,
-    _is_responses_api_create_route,
+    _streams_response_ids,
 )
 from litellm.types.llms.openai import (
     GenericEvent,
@@ -527,7 +527,7 @@ class TestAsyncPreCallHook:
                     assert "team" in exc_info.value.detail.lower()
 
 
-class TestIsResponsesApiCreateRoute:
+class TestStreamsResponseIds:
     """Test the route gate that decides whether a streamed response id is encrypted."""
 
     @pytest.mark.parametrize(
@@ -536,10 +536,13 @@ class TestIsResponsesApiCreateRoute:
             "/v1/responses",
             "/responses",
             "/openai/v1/responses",
+            "/v1/responses/resp_abc",
+            "/responses/resp_abc",
+            "/openai/v1/responses/resp_abc",
         ],
     )
-    def test_create_routes_match(self, route):
-        assert _is_responses_api_create_route(route) is True
+    def test_create_and_retrieve_routes_match(self, route):
+        assert _streams_response_ids(route) is True
 
     @pytest.mark.parametrize(
         "route",
@@ -547,14 +550,14 @@ class TestIsResponsesApiCreateRoute:
             None,
             "/chat/completions",
             "/openai/v1/chat/completions",
-            "/v1/responses/{response_id}",
-            "/openai/v1/responses/{response_id}",
+            "/v1/responses/",
+            "/v1/responses/resp_abc/input_items",
             "/v1/responsesX",
             "/responsesX",
         ],
     )
-    def test_non_create_routes_do_not_match(self, route):
-        assert _is_responses_api_create_route(route) is False
+    def test_other_routes_do_not_match(self, route):
+        assert _streams_response_ids(route) is False
 
 
 class TestAsyncPostCallStreamingIteratorHook:
@@ -609,7 +612,13 @@ class TestAsyncPostCallStreamingIteratorHook:
     @pytest.mark.asyncio
     @pytest.mark.parametrize(
         "route",
-        ["/v1/responses", "/responses", "/openai/v1/responses"],
+        [
+            "/v1/responses",
+            "/responses",
+            "/openai/v1/responses",
+            "/v1/responses/resp_rawprovider123",
+            "/openai/v1/responses/resp_rawprovider123",
+        ],
     )
     async def test_streamed_id_encrypted_on_all_responses_routes(
         self, responses_id_security, route, monkeypatch