fix(proxy): end failed responses streams with [DONE]

Emit data: [DONE] after event: response.failed, and after a late failure
when a terminal event already went out, so OpenAI SDK clients see the
same stream end as a completed response. Restore the lazy OpenAPI
snapshot to its Python 3.12 rendering, which is what CI regenerates.

litellm/proxy/_lazy_openapi_snapshot.json

@@ -19394,7 +19394,7 @@
                   }
                 }
               },
-              "description": "\nUnified rate-limit error.\n\nEvery rate-limit condition surfaced by litellm \u2014 whether it originated from\nan upstream LLM provider, a vendor batch endpoint, or one of litellm's own\nproxy-side limiters (parallel-requests, dynamic-rate, batch-rate, budget,\nmax-iterations, etc.) \u2014 is raised as an instance of this class.\n\nThe :attr:`category` attribute lets callers distinguish the source. See\n:class:`RateLimitErrorCategory` for the available values.\n"
+              "description": "\n    Unified rate-limit error.\n\n    Every rate-limit condition surfaced by litellm \u2014 whether it originated from\n    an upstream LLM provider, a vendor batch endpoint, or one of litellm's own\n    proxy-side limiters (parallel-requests, dynamic-rate, batch-rate, budget,\n    max-iterations, etc.) \u2014 is raised as an instance of this class.\n\n    The :attr:`category` attribute lets callers distinguish the source. See\n    :class:`RateLimitErrorCategory` for the available values.\n    "
             },
             "500": {
               "content": {

litellm/proxy/proxy_server.py

@@ -8865,6 +8865,7 @@ def _format_streaming_sse_chunk(chunk: str | bytes) -> str | bytes:
 
 
 _SSE_FRAME_DELIMITERS: Final = ("\r\n\r\n", "\n\n", "\r\r")
+_OPENAI_STREAM_DONE_FRAME: Final = "data: [DONE]\n\n"
 _MAX_RAW_SSE_BUFFER_CHARS: Final = 8 * 1024 * 1024
 
 
@@ -9272,8 +9273,7 @@ async def async_data_generator(
             yield error_message
         # OpenAI-compatible streams terminate with data: [DONE]; Google GenAI (?alt=sse) does not.
         if not request_data.get("_litellm_skip_openai_stream_done"):
-            done_message: Final = "[DONE]"
-            yield f"data: {done_message}\n\n"
+            yield _OPENAI_STREAM_DONE_FRAME
     except (asyncio.CancelledError, GeneratorExit):
         # Client disconnected mid-stream. CancelledError / GeneratorExit are
         # BaseException, so they bypass the success/failure logging callbacks
@@ -9303,6 +9303,8 @@ async def async_data_generator(
             error_frame: Final = error_state.format_failure(e)
             if error_frame is not None:
                 yield error_frame
+            if not request_data.get("_litellm_skip_openai_stream_done"):
+                yield _OPENAI_STREAM_DONE_FRAME
             return
         if isinstance(e, HTTPException):
             raise e

tests/test_litellm/proxy/proxy_server/test_streaming_helpers.py

@@ -995,6 +995,8 @@ async def test_responses_stream_keeps_tool_deltas_and_only_emits_a_valid_termina
         for frame in event_frames
     )
 
+    assert decoded[-1] == "data: [DONE]\n\n"
+    assert len(decoded) == len(event_frames) + 1
     assert payloads[0]["response"]["id"] == "resp_visible"
     assert payloads[1] == tool_delta.model_dump()
     assert len(payloads) == 3

tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py

@@ -95,7 +95,8 @@ async def test_streaming_upstream_errors_keep_the_client_protocol(
     assert result.status_code == 200, result.text
     assert message in result.text
     if path == "/v1/responses":
-        assert frames[-1].startswith("event: response.failed\n"), result.text
+        assert frames[-1] == "data: [DONE]", result.text
+        assert frames[-2].startswith("event: response.failed\n"), result.text
         if partial:
             assert [event["type"] for event in events] == [
                 "response.created", "response.output_item.added",