fix(bedrock): keep chat_completions/-prefixed deployments on the native Responses surface

litellm/llms/bedrock/responses/transformation.py

@@ -71,11 +71,16 @@ BEDROCK_RUNTIME_SUPPORTED_RESPONSE_TOOL_TYPES: Final = frozenset(
     {"function", "mcp", "custom", "apply_patch", "namespace", "tool_search", "computer"}
 )
 BEDROCK_RUNTIME_UNSUPPORTED_RESPONSE_PARAMS: Final = frozenset({"background"})
+BEDROCK_CHAT_COMPLETIONS_ROUTE_PREFIX: Final = "chat_completions/"
 REMOTE_IMAGE_URL_SCHEMES: Final = ("http://", "https://")
 IMAGE_BLOCK_KEYS: Final = ("content", "output")
 IMAGE_BLOCK_TYPES: Final = frozenset({"input_image", "computer_screenshot"})
 
 
+def _without_chat_completions_route(model: str) -> str:
+    return model.removeprefix(BEDROCK_CHAT_COMPLETIONS_ROUTE_PREFIX)
+
+
 def resolve_bedrock_bearer_token(api_key: str | None) -> str | None:
     return api_key or get_secret_str("AWS_BEARER_TOKEN_BEDROCK")
 
@@ -168,9 +173,13 @@ class BedrockOpenAIResponsesConfig(BaseAWSLLM, OpenAIResponsesAPIConfig):
         The capability decision lives here rather than in the shared dispatch so that
         onboarding a model, or changing how the signal is read, stays inside the
         Bedrock adapter. ``None`` leaves the caller's existing behaviour untouched --
-        chat-only Bedrock models keep the Chat Completions bridge.
+        chat-only Bedrock models keep the Chat Completions bridge. The ``chat_completions/``
+        opt-in only moves Chat Completions calls off Converse, so a Responses call on such a
+        deployment still takes this surface instead of being bridged.
         """
-        if not bedrock_supports_openai_responses(model, litellm.model_cost):
+        if not model or not bedrock_supports_openai_responses(
+            _without_chat_completions_route(model), litellm.model_cost
+        ):
             return None
         return cls()
 
@@ -330,7 +339,7 @@ class BedrockOpenAIResponsesConfig(BaseAWSLLM, OpenAIResponsesAPIConfig):
                 rewritten_types,
             )
         return super().transform_responses_api_request(
-            model=model,
+            model=_without_chat_completions_route(model),
             input=normalized_input,
             response_api_optional_request_params=response_api_optional_request_params,
             litellm_params=litellm_params,

tests/unit/llms/bedrock/responses/test_bedrock_openai_responses.py

@@ -162,6 +162,27 @@ class TestForModelGate:
         ):
             assert BedrockOpenAIResponsesConfig.for_model(None) is None
 
+    def test_chat_completions_route_keeps_the_native_responses_surface(self):
+        with patch.object(  # test-quality-ok: the gate reads the global cost map by design; no injection point exists
+            litellm, "model_cost", {MODEL: {"supported_endpoints": ["/v1/responses"]}}
+        ):
+            cfg = BedrockOpenAIResponsesConfig.for_model(f"chat_completions/{MODEL}")
+        assert isinstance(cfg, BedrockOpenAIResponsesConfig)
+        body = cfg.transform_responses_api_request(
+            model=f"chat_completions/{MODEL}",
+            input="hi",
+            response_api_optional_request_params={},
+            litellm_params=GenericLiteLLMParams(),
+            headers={},
+        )
+        assert body["model"] == MODEL
+
+    def test_converse_route_keeps_the_chat_completions_bridge(self):
+        with patch.object(  # test-quality-ok: the gate reads the global cost map by design; no injection point exists
+            litellm, "model_cost", {MODEL: {"supported_endpoints": ["/v1/responses"]}}
+        ):
+            assert BedrockOpenAIResponsesConfig.for_model(f"converse/{MODEL}") is None
+
 
 class TestProviderResolution:
     """model_cost is patched explicitly: it is populated at import time from a GitHub