fix(proxy): satisfy type discipline for late error frames

litellm/proxy/common_request_processing.py

@@ -1297,7 +1297,7 @@ _ANTHROPIC_LATE_ERROR_BODY_LIMIT: Final = 64 * 1024
 
 async def _read_late_error_body(response: Response) -> bytes:
     """Read at most one page of an error body. A read that fails yields what it got."""
-    read: bytes = b""
+    read: bytes = b""  # rebind-ok: accumulates streamed chunks, then takes the buffered body
     try:
         if isinstance(response, StreamingResponse):
             async for chunk in response.body_iterator:
@@ -1305,20 +1305,20 @@ async def _read_late_error_body(response: Response) -> bytes:
                 if len(read) >= _ANTHROPIC_LATE_ERROR_BODY_LIMIT:
                     break
         else:
-            read = bytes(response.body or b"")
+            read = bytes(response.body or b"")  # rebind-ok: the buffered body arrives whole, nothing to accumulate
     except Exception:  # noqa: BLE001  # the stream is already broken; the caller still needs a frame
         verbose_proxy_logger.exception("could not read the body of an upstream late failure")
     return read[:_ANTHROPIC_LATE_ERROR_BODY_LIMIT]
 
 
 class _LateErrorDetail(BaseModel):
-    model_config = ConfigDict(extra="ignore")
+    model_config = ConfigDict(extra="ignore", frozen=True)
 
     message: str | None = None
 
 
 class _LateErrorEnvelope(BaseModel):
-    model_config = ConfigDict(extra="ignore")
+    model_config = ConfigDict(extra="ignore", frozen=True)
 
     error: _LateErrorDetail | None = None
 

tests/unit/proxy/test_passthrough_late_error_frames.py

@@ -11,7 +11,7 @@ import json
 from collections.abc import AsyncIterator
 
 import pytest
-from starlette.responses import StreamingResponse
+from starlette.responses import Response, StreamingResponse
 
 from litellm.proxy._types import ProxyException
 from litellm.proxy.common_request_processing import open_sse_before_first_byte
@@ -36,8 +36,8 @@ async def _drain(response: StreamingResponse) -> str:
     return b"".join(chunks).decode("utf-8")
 
 
-async def _relay(produced: StreamingResponse, *, answers_as_anthropic: bool = False) -> str:
-    async def deliver() -> StreamingResponse:
+async def _relay(produced: Response, *, answers_as_anthropic: bool = False) -> str:
+    async def deliver() -> Response:
         await asyncio.sleep(0.05)
         return produced
 
@@ -126,6 +126,20 @@ async def test_a_large_error_body_is_read_with_a_bound():
     assert len(body) < 128 * 1024
 
 
+@pytest.mark.asyncio
+async def test_an_error_body_that_arrived_whole_still_reaches_the_client_as_an_event():
+    produced = Response(
+        content=json.dumps({"error": {"message": "buffered upstream said no", "type": "server_error"}}),
+        status_code=502,
+        media_type="application/json",
+    )
+
+    body = await _relay(produced, answers_as_anthropic=True)
+
+    assert body.count("event: error") == 1
+    assert "buffered upstream said no" in body
+
+
 FAKE_KEY = "sk-test-ABCDEFGHIJKLMNOPQRSTUVWXYZ0123456789"
 INTERNAL_PATH = "/home/runner/litellm/proxy/internal_state.py"