fix(proxy): bill a failure after a logged stream success at zero tokens

When a post-call hook raises after a stream's success log went out, the
failure row lifted no usage and fell back to the success's token counts, so
the daily usage tables counted the stream's tokens twice. Lift a zero usage
and zero cost while the stream success bills the current attempt.

litellm/proxy/utils.py

@@ -1819,7 +1819,7 @@ def _failure_fields_to_lift(
     )
     _billed: Final = provider_billed and not (_success_already_logged or _stream_success_bills_this_attempt)
     _usage_to_lift: Final = (
-        None
+        (Usage(prompt_tokens=0, completion_tokens=0, total_tokens=0), 0.0)
         if _stream_success_bills_this_attempt
         else _failure_usage_to_lift(
             model_call_details=_model_call_details,

tests/unit/proxy/utils/proxy_logging/test_streaming_hooks.py

@@ -28,6 +28,7 @@ from litellm.llms.anthropic.messages.native.streaming_iterator import (
 )
 from litellm.proxy._types import UserAPIKeyAuth
 from litellm.litellm_core_utils.streaming_handler import CustomStreamWrapper
+from litellm.proxy.spend_tracking.spend_tracking_utils import get_logging_payload
 from litellm.proxy.utils import ProxyLogging, _failure_fields_to_lift
 from litellm.types.guardrails import GuardrailEventHooks
 from litellm.types.utils import Usage
@@ -782,9 +783,11 @@ async def test_chat_stream_callback_error_after_its_success_dispatch_leaves_the_
             pass
     await asyncio.sleep(0)
 
-    lifted: Final = _failure_fields_to_lift(request_data, provider_billed=True)
+    failure_row: Final = get_logging_payload(
+        dict(_failure_fields_to_lift(request_data, provider_billed=True)), None, datetime.now(), datetime.now()
+    )
     assert events == ["success_dispatched"]
-    assert (lifted.get("combined_usage_object"), lifted.get("response_cost")) == (None, None)
+    assert (failure_row["prompt_tokens"], failure_row["completion_tokens"], failure_row["spend"]) == (0, 0, 0.0)
 
 
 @pytest.mark.asyncio

tests/unit/test_router/test_router.py

@@ -2868,6 +2868,7 @@ async def test_continued_stream_sharing_the_proxy_logging_object_bills_the_cut_t
 async def test_a_failure_after_a_continued_stream_success_bills_nothing_the_success_row_already_billed(
     monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter
 ) -> None:
+    from litellm.proxy.spend_tracking.spend_tracking_utils import get_logging_payload
     from litellm.proxy.utils import _failure_fields_to_lift
 
     monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
@@ -2880,7 +2881,12 @@ async def test_a_failure_after_a_continued_stream_success_bills_nothing_the_succ
     (success,) = recorder.successes
     assert success["cost_breakdown"]["additional_costs"]["mid_stream_partial_cost"] > 0
     late_failure: Final = _failure_fields_to_lift({"litellm_logging_obj": shared_logging}, provider_billed=True)
-    assert (late_failure.get("combined_usage_object"), late_failure.get("response_cost")) == (None, None)
+    late_failure_row: Final = get_logging_payload(dict(late_failure), None, datetime.now(), datetime.now())
+    assert (late_failure_row["prompt_tokens"], late_failure_row["completion_tokens"], late_failure_row["spend"]) == (
+        0,
+        0,
+        0.0,
+    )
 
 
 _AGENTIC_FIRST_CALL_USAGE: Final = {"prompt_tokens": 50, "completion_tokens": 7, "total_tokens": 57}
@@ -3286,6 +3292,7 @@ def _backup_responses_cost() -> float:
 async def test_responses_fallback_sharing_the_proxy_logging_object_bills_the_cut_partial_once(
     monkeypatch: pytest.MonkeyPatch, respx_mock: respx.MockRouter, failed_usage: Mapping[str, int] | None
 ) -> None:
+    from litellm.proxy.spend_tracking.spend_tracking_utils import get_logging_payload
     from litellm.proxy.utils import _failure_fields_to_lift
 
     monkeypatch.setattr(litellm, "disable_aiohttp_transport", True)
@@ -3318,7 +3325,12 @@ async def test_responses_fallback_sharing_the_proxy_logging_object_bills_the_cut
         success["response_cost"],
     )
     late_failure: Final = _failure_fields_to_lift({"litellm_logging_obj": shared_logging}, provider_billed=True)
-    assert (late_failure.get("combined_usage_object"), late_failure.get("response_cost")) == (None, None)
+    late_failure_row: Final = get_logging_payload(dict(late_failure), None, datetime.now(), datetime.now())
+    assert (late_failure_row["prompt_tokens"], late_failure_row["completion_tokens"], late_failure_row["spend"]) == (
+        0,
+        0,
+        0.0,
+    )
 
 
 @pytest.mark.asyncio