fix(bedrock): map Bedrock's 200 error body to a 400 on the Anthropic invoke /v1/messages route

litellm/llms/bedrock/chat/invoke_handler.py

@@ -48,6 +48,7 @@ from litellm.types.utils import GenericStreamingChunk as GChunk
 from ..common_utils import (
     BedrockError,
     BedrockEventStreamResponseDict,
+    bedrock_answered_json,
     bedrock_error_in_success_body,
     bedrock_event_stream_header,
     bedrock_event_stream_response,
@@ -193,10 +194,6 @@ def response_format_schema_from_request(data: str, json_mode: bool | None) -> Re
     return JsonModeRequest.model_validate(body).tool_schema()
 
 
-def _answered_json(response: httpx.Response) -> bool:
-    return "application/json" in (response.headers.get("content-type") or "")
-
-
 def _stream_decoder(
     bedrock_invoke_provider: litellm.BEDROCK_INVOKE_PROVIDERS_LITERAL | None,
     *,
@@ -264,7 +261,7 @@ async def make_call(
                 response=response,
             )
 
-        if _answered_json(response):
+        if bedrock_answered_json(response):
             await response.aread()
             success_body_error: Final = bedrock_error_in_success_body(response)
             if success_body_error is not None:
@@ -366,7 +363,7 @@ def make_sync_call(
                 response=response,
             )
 
-        if _answered_json(response):
+        if bedrock_answered_json(response):
             response.read()
             success_body_error: Final = bedrock_error_in_success_body(response)
             if success_body_error is not None:

litellm/llms/bedrock/common_utils.py

@@ -10,7 +10,7 @@ import json
 import os
 import re
 import warnings
-from collections.abc import Iterator, Mapping, Sequence
+from collections.abc import AsyncIterator, Iterator, Mapping, Sequence
 from dataclasses import dataclass
 from types import MappingProxyType
 from typing import TYPE_CHECKING, Any, Final, Literal, TypeAlias
@@ -147,6 +147,22 @@ def bedrock_error_in_success_body(response: httpx.Response) -> BedrockError | No
     )
 
 
+def bedrock_answered_json(response: httpx.Response) -> bool:
+    return "application/json" in (response.headers.get("content-type") or "")
+
+
+async def _json_body_or_bedrock_error(response: httpx.Response) -> AsyncIterator[bytes]:
+    body: Final = await response.aread()
+    error: Final = bedrock_error_in_success_body(response)
+    if error is not None:
+        raise error
+    yield body
+
+
+def bedrock_event_stream_bytes(response: httpx.Response) -> AsyncIterator[bytes]:
+    return _json_body_or_bedrock_error(response) if bedrock_answered_json(response) else response.aiter_bytes()
+
+
 _BEDROCK_AWS_AUTH_PARAMETER_KEYS: Final[tuple[str, ...]] = (*AWS_AUTH_PARAM_KEYS, "aws_region_name")
 
 

litellm/llms/bedrock/messages/invoke_transformations/anthropic_claude_transformation.py

@@ -39,6 +39,8 @@ from litellm.llms.bedrock.chat.invoke_transformations.base_invoke_transformation
 from litellm.llms.bedrock.common_utils import (
     BedrockError,
     apply_bedrock_invoke_structured_output,
+    bedrock_error_in_success_body,
+    bedrock_event_stream_bytes,
     bedrock_supports_tool_search,
     ensure_bedrock_anthropic_messages_tool_names,
     get_anthropic_beta_from_headers,
@@ -62,6 +64,7 @@ from litellm.types.llms.anthropic import (
     ANTHROPIC_FINE_GRAINED_TOOL_STREAMING_BETA_HEADER,
     ANTHROPIC_TOOL_SEARCH_BETA_HEADER,
 )
+from litellm.types.llms.anthropic_messages.anthropic_response import AnthropicMessagesResponse
 from litellm.types.llms.bedrock import BedrockInvokeAnthropicMessagesRequest
 from litellm.types.llms.openai import AllMessageValues
 from litellm.types.router import GenericLiteLLMParams
@@ -868,6 +871,17 @@ class AmazonAnthropicClaudeMessagesConfig(
 
         return anthropic_messages_request
 
+    def transform_anthropic_messages_response(
+        self,
+        model: str,
+        raw_response: httpx.Response,
+        logging_obj: LiteLLMLoggingObj,
+    ) -> AnthropicMessagesResponse:
+        success_body_error: Final = bedrock_error_in_success_body(raw_response)
+        if success_body_error is not None:
+            raise success_body_error
+        return super().transform_anthropic_messages_response(model, raw_response, logging_obj)
+
     def get_async_streaming_response_iterator(
         self,
         model: str,
@@ -879,7 +893,7 @@ class AmazonAnthropicClaudeMessagesConfig(
             model=model,
         )
         completion_stream: Final = aws_decoder.aiter_bytes(
-            httpx_response.aiter_bytes(), response_headers=httpx_response.headers
+            bedrock_event_stream_bytes(httpx_response), response_headers=httpx_response.headers
         )
         # Convert decoded Bedrock events to Server-Sent Events expected by Anthropic clients.
         return self.bedrock_sse_wrapper(

tests/unit/llms/bedrock/messages/invoke_transformations/test_anthropic_claude_transformation.py

@@ -33,6 +33,7 @@ from litellm.llms.anthropic.messages.native.mid_conversation_system import (
     as_system_content_blocks,
 )
 from litellm.llms.bedrock.common_utils import (
+    BedrockError,
     ensure_bedrock_anthropic_messages_tool_names,
     normalize_custom_field_on_tools,
     normalize_tool_input_schema_types_for_bedrock_invoke,
@@ -4553,3 +4554,42 @@ async def test_bedrock_chat_fallback_to_anthropic_sends_the_tool_bedrock_cleaned
 
     (body,) = upstream.anthropic_bodies
     assert body["tools"] == [{"name": "get_weather", "input_schema": schema, **sent_extras}]
+
+
+_UNKNOWN_OPERATION_BODY: Final = {
+    "Output": {"__type": "com.amazon.coral.service#UnknownOperationException"},
+    "Version": "1.0",
+}
+
+
+@pytest.mark.usefixtures("local_model_cost_map")
+@pytest.mark.parametrize(
+    ("stream", "raised"),
+    [(False, litellm.BadRequestError), (True, BedrockError)],
+    ids=["non-stream", "stream"],
+)
+async def test_bedrock_messages_raise_a_400_when_bedrock_answers_200_with_an_unknown_operation_body(
+    stream: bool,
+    raised: type[litellm.BadRequestError] | type[BedrockError],
+) -> None:
+    bedrock: Final = ScriptedBedrockInvoke(replies=((200, _UNKNOWN_OPERATION_BODY),))
+
+    async def _call_and_drain() -> None:
+        response: Final = await litellm.anthropic_messages(
+            model="bedrock/invoke/anthropic/us.anthropic.claude-haiku-5-5",
+            messages=[{"role": "user", "content": "Say hi in three words"}],
+            max_tokens=20,
+            stream=stream,
+            aws_access_key_id="test-access-key",
+            aws_secret_access_key="test-secret-key",
+            aws_region_name="us-east-1",
+            client=AsyncHTTPHandler(transport=httpx.MockTransport(bedrock)),
+        )
+        if isinstance(response, AsyncIterator):
+            _ = [chunk async for chunk in response]
+
+    with pytest.raises(raised) as error:
+        await _call_and_drain()
+
+    assert error.value.status_code == 400
+    assert "UnknownOperationException" in str(error.value)