fix(bedrock): read an inference profile's base_model from model_info and keep shadow eval's effort with thinking

litellm/integrations/shadow_eval_logger.py

@@ -134,7 +134,9 @@ def _chat_request_from_anthropic_messages(
         )
     )
     translated, _ = adapter.translate_anthropic_to_openai(request)  # pyright: ignore[reportArgumentType]  # wire-body mapping is the surface's native request shape; the adapter is duck-typed and read-only here
-    return translated
+    if request.get("thinking"):
+        return translated
+    return MappingProxyType({k: v for k, v in translated.items() if k != "output_config"})
 
 
 def _chat_request_from_responses(

litellm/llms/anthropic/messages/adapters/handler.py

@@ -607,6 +607,7 @@ class LiteLLMMessagesToCompletionTransformationHandler:
         from litellm.litellm_core_utils.litellm_logging import (
             Logging as LiteLLMLoggingObject,
         )
+        from litellm.responses.bridge_check import str_keyed_mapping
 
         request_data: Final[dict[str, object]] = {
             "model": model,
@@ -646,7 +647,8 @@ class LiteLLMMessagesToCompletionTransformationHandler:
             request_data["output_config"] = extra_kwargs["output_config"]
 
         custom_llm_provider: Final = extra_kwargs.get("custom_llm_provider")
-        base_model: Final = extra_kwargs.get("base_model")
+        model_info: Final = str_keyed_mapping(extra_kwargs.get("model_info"))
+        base_model: Final = extra_kwargs.get("base_model") or (model_info.get("base_model") if model_info else None)
         (
             openai_request,
             tool_name_mapping,

tests/unit/integrations/test_shadow_eval_logger.py

@@ -405,6 +405,34 @@ class TestSurfaceNormalization:
         assert "stream" not in shadow_call
         assert shadow_call["metadata"][INTERNAL_CALL_ORIGIN_METADATA_KEY] == SHADOW_EVAL_ROUTER_CALL_ORIGIN
 
+    @pytest.mark.parametrize(
+        ("thinking", "forwards_effort"),
+        [
+            pytest.param(None, False, id="no_thinking"),
+            pytest.param({"type": "enabled", "budget_tokens": 1024}, True, id="thinking"),
+        ],
+    )
+    async def test_anthropic_arm_sends_a_bedrock_claude_effort_to_the_shadow_model_only_with_thinking(
+        self, thinking: Mapping[str, object] | None, forwards_effort: bool
+    ):
+        hook_kwargs = _success_kwargs(call_type="anthropic_messages")
+        hook_kwargs["model"] = "converse/us.anthropic.claude-sonnet-4-6"
+        hook_kwargs["messages"] = [{"role": "user", "content": [{"type": "text", "text": "hi"}]}]
+        hook_kwargs["litellm_params"]["proxy_server_request"] = {
+            "body": {
+                "model": "converse-sonnet",
+                "messages": [],
+                "max_tokens": 2048,
+                "output_config": {"effort": "high"},
+                **({"thinking": thinking} if thinking else {}),
+            }
+        }
+
+        _, router = await self._drive(hook_kwargs, RESPONSE)
+
+        shadow_call = router.acompletion.call_args_list[0].kwargs
+        assert shadow_call.get("output_config") == ({"effort": "high"} if forwards_effort else None)
+
     async def test_responses_arm_translates_wire_body_params_and_drops_surface_only_keys(self):
         from litellm.types.llms.openai import ResponsesAPIResponse
 

tests/unit/llms/anthropic/messages/adapters/test_anthropic_messages_adapters_transformation.py

@@ -5560,19 +5560,28 @@ _APPLICATION_INFERENCE_PROFILE: Final = (
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize(
-    ("thinking", "drop_params", "base_model", "forwards_effort"),
+    ("thinking", "drop_params", "deployment_kwargs", "forwards_effort"),
     [
-        pytest.param(None, False, None, True, id="no_thinking"),
-        pytest.param(None, True, None, False, id="no_thinking_drop_params"),
-        pytest.param({"type": "enabled", "budget_tokens": 1024}, True, None, True, id="thinking_drop_params"),
-        pytest.param(None, True, _CONVERSE_EFFORT_MODEL, True, id="no_thinking_drop_params_base_model"),
+        pytest.param(None, False, {}, True, id="no_thinking"),
+        pytest.param(None, True, {}, False, id="no_thinking_drop_params"),
+        pytest.param({"type": "enabled", "budget_tokens": 1024}, True, {}, True, id="thinking_drop_params"),
+        pytest.param(
+            None, True, {"base_model": _CONVERSE_EFFORT_MODEL}, True, id="no_thinking_drop_params_base_model"
+        ),
+        pytest.param(
+            None,
+            True,
+            {"model_info": {"base_model": _CONVERSE_EFFORT_MODEL}},
+            True,
+            id="no_thinking_drop_params_model_info_base_model",
+        ),
     ],
 )
 async def test_bedrock_application_inference_profile_loses_the_effort_only_when_unresolved_without_thinking(
     _converse_route: respx.Route,
     thinking: Mapping[str, object] | None,
     drop_params: bool,
-    base_model: str | None,
+    deployment_kwargs: Mapping[str, object],
     forwards_effort: bool,
 ) -> None:
     await litellm.anthropic.messages.acreate(
@@ -5582,7 +5591,7 @@ async def test_bedrock_application_inference_profile_loses_the_effort_only_when_
         output_config={"effort": "high", "format": _CITY_FORMAT},
         drop_params=drop_params,
         **({"thinking": thinking} if thinking else {}),
-        **({"base_model": base_model} if base_model else {}),
+        **deployment_kwargs,
     )
 
     sent: Final = _sent_converse_fields(_converse_route)