Merge pull request #33767 from BerriAI/litellm_lit_4561_bedrock_passthrough_content_type

fix(proxy): forward Bedrock event-stream content-type on unbuffered passthrough

litellm/proxy/common_request_processing.py

@@ -2039,6 +2039,7 @@ class ProxyBaseLLMRequestProcessing:
                         return StreamingResponse(
                             content=generator,
                             status_code=status.HTTP_200_OK,
+                            media_type=self._passthrough_event_stream_media_type(),
                             headers=custom_headers,
                         )
                     else:
@@ -2494,10 +2495,16 @@ class ProxyBaseLLMRequestProcessing:
 
     def _passthrough_event_stream_media_type(self) -> str | None:
         """
-        Content-type for a buffered passthrough event-stream response, resolved
-        from the provider handler so the proxy stays provider-agnostic. Mirrors
-        the upstream content-type the non-streaming path forwards, since the
-        buffered streaming generator carries no headers of its own.
+        Content-type for a passthrough event-stream response, resolved from the
+        provider handler so the proxy stays provider-agnostic. Mirrors the
+        upstream content-type the non-streaming path forwards, since the
+        streaming generator carries no headers of its own. Used for both the
+        buffered (guardrail-rewritten) and the unbuffered relay paths so
+        clients that enforce the event-stream content-type (e.g. Claude Code on
+        Bedrock invoke-with-response-stream) see the correct header instead of
+        no content-type at all, which they fall back to reading as
+        application/octet-stream. Returns None for providers with no
+        event-stream media type, leaving the response headers unchanged.
         """
         from litellm.llms.pass_through.guardrail_translation.handler import (
             LlmPassthroughRouteHandler,

tests/test_litellm/proxy/test_common_request_processing.py

@@ -4511,6 +4511,61 @@ class TestAllmPassthroughStreamingProviderGate:
         assert streamed == chunks
         mock_handler.assert_not_awaited()
 
+    @pytest.mark.asyncio
+    async def test_bedrock_invoke_stream_sets_event_stream_content_type(self, monkeypatch):
+        """
+        Regression for LIT-4561. The unbuffered Bedrock event-stream relay
+        (invoke-with-response-stream, no post-call guardrail rewriting) must set
+        content-type: application/vnd.amazon.eventstream instead of emitting no
+        content-type header at all, which trips Claude Code's content-type guard
+        added in 2.1.208
+        """
+        processing_obj = self._build_processing_obj(
+            "bedrock", "model/us.anthropic.claude-sonnet-4-20250514-v1:0/invoke-with-response-stream"
+        )
+        chunks = [b"raw-1", b"raw-2"]
+
+        with patch.object(
+            ProxyBaseLLMRequestProcessing,
+            "_has_post_call_guardrails",
+            return_value=False,
+        ), patch.object(
+            ProxyBaseLLMRequestProcessing,
+            "_has_post_call_guardrails_for_passthrough",
+            return_value=False,
+        ):
+            result = await self._run(processing_obj, monkeypatch, chunks)
+
+        assert isinstance(result, StreamingResponse)
+        assert result.media_type == "application/vnd.amazon.eventstream"
+        assert result.headers["content-type"] == "application/vnd.amazon.eventstream"
+        streamed = [chunk async for chunk in result.body_iterator]
+        assert streamed == chunks
+
+    @pytest.mark.asyncio
+    async def test_non_bedrock_stream_keeps_default_content_type(self, monkeypatch):
+        """
+        A provider with no registered event-stream media type must not have one
+        forced onto its unbuffered stream, so the response default is unchanged
+        """
+        processing_obj = self._build_processing_obj("anthropic")
+        chunks = [b"chunk-1", b"chunk-2"]
+
+        with patch.object(
+            ProxyBaseLLMRequestProcessing,
+            "_has_post_call_guardrails",
+            return_value=False,
+        ), patch.object(
+            ProxyBaseLLMRequestProcessing,
+            "_has_post_call_guardrails_for_passthrough",
+            return_value=False,
+        ):
+            result = await self._run(processing_obj, monkeypatch, chunks)
+
+        assert isinstance(result, StreamingResponse)
+        assert result.media_type is None
+        assert "content-type" not in result.headers
+
 
 class TestResponseCostHeaderForTypedDictResponses:
     """