fix(responses): forward only message cache points through the bridge on native-config routes

Routes with their own Responses config keep only chat completion request
fields on the way into the bridge. Forwarding every injection point there
let a tool_config point reach the provider body through the generic
unknown-param copy, where no transformation reads it. Forward only the
location: message points the bridge's second pass places.

Cherry-picked from 008e91a5c6, adapted to the line's test imports.

litellm/responses/main.py

@@ -66,7 +66,6 @@ from litellm.types.responses.streaming_websocket import ResponsesWebSocketReques
 from litellm.types.router import GenericLiteLLMParams
 from litellm.types.utils import all_litellm_params
 from litellm.utils import (
-    LITELLM_ONLY_PARAMS,
     ProviderConfigManager,
     client,
 )
@@ -443,6 +442,15 @@ def _bridges_to_chat_completions(
 _RESPONSES_ONLY_REQUEST_FIELDS_NEVER_BRIDGED: Final = frozenset({"client_metadata"})
 
 
+def _message_cache_control_points(kwargs: Mapping[str, object]) -> Mapping[str, object]:
+    points: Final = tuple(
+        point
+        for point in configured_injection_points(kwargs.get("cache_control_injection_points"))
+        if point.get("location") == "message"
+    )
+    return {"cache_control_injection_points": points} if points else {}
+
+
 def _bridge_kwargs(
     kwargs: Mapping[str, object],
     responses_api_provider_config: BaseResponsesAPIConfig | None,
@@ -458,11 +466,15 @@ def _bridge_kwargs(
             *DEFAULT_CHAT_COMPLETION_PARAM_VALUES,
             *all_litellm_params,
             *GenericLiteLLMParams.model_fields,
-            *LITELLM_ONLY_PARAMS,
             *(allowed_openai_params or ()),
         )
     ).difference(_RESPONSES_ONLY_REQUEST_FIELDS_NEVER_BRIDGED)
-    return MappingProxyType({key: value for key, value in kwargs.items() if key in forwarded_keys})
+    return MappingProxyType(
+        {
+            **{key: value for key, value in kwargs.items() if key in forwarded_keys},
+            **_message_cache_control_points(kwargs),
+        }
+    )
 
 
 _ResponsesCompatibilityFailure: TypeAlias = Literal["encrypted_task_unsupported"]

tests/unit/responses/test_responses_api_bridge_flag.py

@@ -429,17 +429,18 @@ class TestUseResponsesApiBridgeFlag:
                 {"type": "function_call_output", "call_id": "call_a", "output": "contents of a"},
             ],
             use_chat_completions_api=True,
-            cache_control_injection_points=[{"location": "message", "index": -1}],
+            cache_control_injection_points=[{"location": "tool_config"}, {"location": "message", "index": -1}],
             api_key="fake-provider-api-key",
             num_retries=0,
         )
 
-        sent_messages: Final = json.loads(upstream.calls[0].request.read())["messages"]
+        sent_body: Final = json.loads(upstream.calls[0].request.read())
         assert [
             message.get("tool_call_id") or message["role"]
-            for message in sent_messages
+            for message in sent_body["messages"]
             if "cache_control" in json.dumps(message)
         ] == ["call_a"]
+        assert "cache_control_injection_points" not in sent_body
 
     def test_bridge_keeps_deployment_credentials_while_dropping_unknown_params(self, respx_mock: respx.MockRouter):
         upstream: Final = respx_mock.post(

tests/unit/responses/test_responses_api_request_body.py

@@ -17,6 +17,7 @@ import pytest
 import respx
 
 import litellm
+from litellm.integrations.anthropic_cache_control_hook import MAX_CACHE_CONTROL_BLOCKS
 from litellm.llms.custom_httpx.http_handler import AsyncHTTPHandler
 
 
@@ -1015,11 +1016,11 @@ async def test_the_ordinal_point_listed_first_takes_the_last_breakpoint_slot(ind
         tools=[
             {
                 "type": "function",
-                "name": name,
+                "name": f"t{i}",
                 "parameters": {"type": "object", "properties": {}},
                 "cache_control": {"type": "ephemeral"},
             }
-            for name in ("t0", "t1", "t2")
+            for i in range(MAX_CACHE_CONTROL_BLOCKS - 1)
         ],
     )