fix(proxy): hold fallback lifecycle frames

litellm/router.py

@@ -4984,6 +4984,7 @@ class Router:
         self,
         response: AsyncIterator[bytes],
         initial_kwargs: dict[str, Any],  # mutable-ok: mutated in-place before re-entering the fallback chain
+        allow_fallback: bool = True,
     ) -> AsyncIterator[bytes]:
         """
         Wrap an anthropic_messages (/v1/messages) streaming response so a
@@ -5100,6 +5101,10 @@ class Router:
                 for buffered_chunk in buffered_lifecycle_chunks:
                     yield buffered_chunk
             except Exception as stream_error:  # noqa: BLE001  # any raised provider error must reach the fallback gate
+                if not allow_fallback:
+                    if isinstance(stream_error, MidStreamFallbackError) and stream_error.original_exception is not None:
+                        raise stream_error.original_exception from stream_error
+                    raise
                 async for item in self._aanthropic_messages_recover_stream_error(
                     stream_error,
                     has_generated_content,
@@ -5201,7 +5206,12 @@ class Router:
             wrapper.merge_fallback_hidden_params(fallback_hidden_params, fallback_headers)
             wrapper.adopt_fallback_source(fallback_response)
             if hasattr(fallback_response, "__aiter__"):
-                async for fallback_item in fallback_response:
+                fallback_stream: Final = await self._aanthropic_messages_streaming_iterator(
+                    response=cast("AsyncIterator[bytes]", fallback_response),
+                    initial_kwargs=initial_kwargs,
+                    allow_fallback=False,
+                )
+                async for fallback_item in fallback_stream:
                     yield fallback_item
             else:
                 # A fallback can resolve to a complete AnthropicMessagesResponse

tests/test_litellm/test_router.py

@@ -10686,6 +10686,38 @@ async def test_anthropic_messages_fallback_also_failing_raises_original_exceptio
     assert exc_info.value is original_exception
 
 
+@pytest.mark.asyncio
+async def test_anthropic_messages_double_failure_holds_back_fallback_lifecycle_frames():
+    router = _anthropic_messages_make_router()
+    primary_stream = _AnthropicMessagesFakeByteStream([_anthropic_messages_overloaded_error_chunk()])
+    fallback_error = litellm.APIError(
+        status_code=503,
+        message="fallback also overloaded",
+        llm_provider="bedrock",
+        model="fallback",
+    )
+    fallback_stream = _AnthropicMessagesRaisingByteStream(
+        [_anthropic_messages_message_start_chunk()], fallback_error
+    )
+
+    with patch.object(
+        router,
+        "async_function_with_fallbacks_common_utils",
+        new=AsyncMock(return_value=fallback_stream),
+    ):
+        wrapped = await router._aanthropic_messages_streaming_iterator(
+            response=primary_stream,
+            initial_kwargs={"model": "primary"},
+        )
+        collected = []
+        with pytest.raises(litellm.APIError) as exc_info:
+            async for chunk in wrapped:
+                collected.append(chunk)
+
+    assert collected == []
+    assert exc_info.value is fallback_error
+
+
 # -------- _dispatch_generic_call_type --------