fix(proxy): defer background reservation settlement until usage arrives

litellm/proxy/hooks/budget_reservation_reconcile_logger.py

@@ -8,11 +8,15 @@ from datetime import datetime
 from typing import Final
 
 from litellm._logging import verbose_proxy_logger
+from litellm.constants import BACKGROUND_INTERACTION_COST_POLLING_ENABLED
 from litellm.integrations.custom_logger import CustomLogger
+from litellm.interactions.background_cost_polling import is_pollable_background_interaction
 from litellm.litellm_core_utils.core_helpers import (
     budget_reservation_from_metadata,
-    get_litellm_metadata_from_kwargs,
+    get_litellm_metadata_from_kwargs,  # pyright: ignore[reportUnknownVariableType]  # shared helper has a legacy untyped signature
 )
+from litellm.types.interactions import InteractionsAPIResponse
+
 
 class BudgetReservationReconcileLogger(CustomLogger):
     """Settle auth's spend-counter reservation on success when ProxyDBLogger is absent.
@@ -27,35 +31,44 @@ class BudgetReservationReconcileLogger(CustomLogger):
     async def async_log_success_event(
         self, kwargs: Mapping[str, object], response_obj: object, start_time: datetime, end_time: datetime
     ) -> None:
-        del response_obj, start_time, end_time
-        metadata: Final = get_litellm_metadata_from_kwargs(kwargs=dict(kwargs))
+        metadata: Final = get_litellm_metadata_from_kwargs(kwargs=dict(kwargs))  # pyright: ignore[reportUnknownVariableType]  # shared helper returns untyped metadata
         if not isinstance(metadata, Mapping):
             return
-        budget_reservation: Final = budget_reservation_from_metadata(metadata)
+        budget_reservation: Final = budget_reservation_from_metadata(metadata)  # pyright: ignore[reportUnknownArgumentType]  # helper only reads known keys from the checked mapping
         if not isinstance(budget_reservation, dict) or budget_reservation.get("finalized") is True:
             return
 
-        sl_object = kwargs.get("standard_logging_object")
-        response_cost: object
-        if isinstance(sl_object, Mapping):
-            response_cost = sl_object.get("response_cost")
-        else:
-            response_cost = kwargs.get("response_cost")
-        actual_cost: Final = float(response_cost or 0.0)
+        sl_object: Final = kwargs.get("standard_logging_object")
+        response_cost: Final[object] = (  # pyright: ignore[reportUnknownVariableType]  # callback payloads are legacy untyped mappings
+            sl_object.get("response_cost")  # pyright: ignore[reportUnknownMemberType]  # float below validates the untyped cost value
+            if isinstance(sl_object, Mapping)
+            else kwargs.get("response_cost")
+        )
+        if (
+            response_cost is None
+            and BACKGROUND_INTERACTION_COST_POLLING_ENABLED
+            and isinstance(response_obj, InteractionsAPIResponse)
+            and response_obj.usage is None
+            and is_pollable_background_interaction(response_obj)
+        ):
+            return
+        actual_cost: Final = float(response_cost or 0.0)  # pyright: ignore[reportArgumentType]  # float validates legacy callback values at runtime
 
-        from litellm.proxy.spend_tracking.budget_reservation import reconcile_budget_reservation
+        from litellm.proxy.spend_tracking.budget_reservation import (
+            reconcile_budget_reservation,  # pyright: ignore[reportUnknownVariableType]  # shared reconciler accepts legacy reservation dictionaries
+        )
 
         try:
             await reconcile_budget_reservation(
                 budget_reservation=budget_reservation,
                 actual_cost=actual_cost,
             )
-        except Exception:
+        except Exception:  # noqa: BLE001  # a counter backend failure must still release the reservation
             verbose_proxy_logger.exception(
                 "Failed to reconcile budget reservation without a database logger; releasing instead"
             )
             from litellm.proxy.spend_tracking.budget_reservation import (
-                release_or_invalidate_budget_reservation,
+                release_or_invalidate_budget_reservation,  # pyright: ignore[reportUnknownVariableType]  # shared release accepts legacy reservation dictionaries
             )
 
             await release_or_invalidate_budget_reservation(budget_reservation=budget_reservation)

litellm/proxy/proxy_server.py

@@ -2910,9 +2910,7 @@ def cost_tracking():
             BudgetReservationReconcileLogger,
         )
 
-        litellm.logging_callback_manager.add_litellm_async_success_callback(
-            BudgetReservationReconcileLogger()
-        )
+        litellm.logging_callback_manager.add_litellm_async_success_callback(BudgetReservationReconcileLogger())
 
 
 async def _drain_spend_event_producer_on_shutdown() -> None:

tests/unit/proxy/hooks/test_budget_reservation_reconcile_logger.py

@@ -1,13 +1,18 @@
 """Unit tests for BudgetReservationReconcileLogger (#45895)."""
 
+from collections.abc import Mapping
 from datetime import UTC, datetime
+from typing import Final
 from unittest.mock import AsyncMock
 
 import pytest
 
+from litellm.caching import DualCache
+from litellm.proxy import proxy_server
 from litellm.proxy.hooks.budget_reservation_reconcile_logger import (
     BudgetReservationReconcileLogger,
 )
+from litellm.types.interactions import InteractionsAPIResponse
 
 
 def _reservation(callback_bound: bool = True) -> dict:
@@ -144,3 +149,140 @@ async def test_success_falls_back_to_release_when_reconcile_fails(
     )
 
     release.assert_awaited_once_with(budget_reservation=reservation)
+
+
+@pytest.fixture
+def spend_counter_cache(monkeypatch: pytest.MonkeyPatch) -> DualCache:
+    cache: Final = DualCache()
+    monkeypatch.setattr(proxy_server, "spend_counter_cache", cache)
+    monkeypatch.setattr(proxy_server, "prisma_client", None)
+    return cache
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize("status", ("in_progress", "queued"))
+@pytest.mark.parametrize("standard_logging", (False, True))
+async def test_background_interaction_keeps_reservation_until_final_cost(
+    spend_counter_cache: DualCache, status: str, standard_logging: bool
+) -> None:
+    reservation: Final = _reservation()
+    counter_key: Final = "spend:key:hashed"
+    spend_counter_cache.in_memory_cache.set_cache(key=counter_key, value=reservation["reserved_cost"])
+    logger: Final = BudgetReservationReconcileLogger()
+    now: Final = datetime(2026, 1, 1, tzinfo=UTC)
+    kwargs: Final = {
+        "litellm_params": {"metadata": {"user_api_key_budget_reservation": reservation}},
+        **({"standard_logging_object": {"response_cost": None}, "response_cost": 9.9} if standard_logging else {}),
+    }
+    pending: Final = InteractionsAPIResponse(id="interactions/background", status=status)
+
+    await logger.async_log_success_event(kwargs, pending, now, now)
+
+    assert reservation["finalized"] is False
+    assert spend_counter_cache.in_memory_cache.get_cache(key=counter_key) == pytest.approx(reservation["reserved_cost"])
+
+    completed: Final = pending.model_copy(update={"status": "completed", "usage": {"total_tokens": 12}})
+    final_kwargs: Final = {**kwargs, "standard_logging_object": {"response_cost": 0.004}}
+    await logger.async_log_success_event(final_kwargs, completed, now, now)
+
+    assert reservation["finalized"] is True
+    assert spend_counter_cache.in_memory_cache.get_cache(key=counter_key) == pytest.approx(0.004)
+
+    await logger.async_log_success_event(final_kwargs, completed, now, now)
+
+    assert spend_counter_cache.in_memory_cache.get_cache(key=counter_key) == pytest.approx(0.004)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("status", "interaction_id", "polling_enabled", "has_usage"),
+    (
+        ("in_progress", "interactions/background", False, False),
+        ("queued", "interactions/background", False, False),
+        ("in_progress", None, True, False),
+        ("queued", "", True, False),
+        ("completed", "interactions/background", True, False),
+        ("failed", "interactions/background", True, False),
+        ("cancelled", "interactions/background", True, False),
+        ("requires_action", "interactions/background", True, False),
+        ("in_progress", "interactions/background", True, True),
+    ),
+)
+async def test_missing_cost_releases_reservation_when_no_usage_poll_will_settle_it(
+    spend_counter_cache: DualCache,
+    monkeypatch: pytest.MonkeyPatch,
+    status: str,
+    interaction_id: str | None,
+    polling_enabled: bool,
+    has_usage: bool,
+) -> None:
+    monkeypatch.setattr(
+        "litellm.proxy.hooks.budget_reservation_reconcile_logger.BACKGROUND_INTERACTION_COST_POLLING_ENABLED",
+        polling_enabled,
+    )
+    reservation: Final = _reservation()
+    counter_key: Final = "spend:key:hashed"
+    spend_counter_cache.in_memory_cache.set_cache(key=counter_key, value=reservation["reserved_cost"])
+    logger: Final = BudgetReservationReconcileLogger()
+    now: Final = datetime(2026, 1, 1, tzinfo=UTC)
+
+    await logger.async_log_success_event(
+        {"litellm_params": {"metadata": {"user_api_key_budget_reservation": reservation}}},
+        InteractionsAPIResponse(id=interaction_id, status=status, usage={"total_tokens": 12} if has_usage else None),
+        now,
+        now,
+    )
+
+    assert reservation["finalized"] is True
+    assert spend_counter_cache.in_memory_cache.get_cache(key=counter_key) == pytest.approx(0.0)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    ("response_cost", "expected_cost"),
+    ((0.0, 0.0), (0.004, 0.004), ("0.004", 0.004), (b"0.004", 0.004), (memoryview(b"0.004"), 0.004)),
+)
+async def test_reported_interaction_cost_is_settled_immediately(
+    spend_counter_cache: DualCache, response_cost: object, expected_cost: float
+) -> None:
+    reservation: Final = _reservation()
+    counter_key: Final = "spend:key:hashed"
+    spend_counter_cache.in_memory_cache.set_cache(key=counter_key, value=reservation["reserved_cost"])
+    logger: Final = BudgetReservationReconcileLogger()
+    now: Final = datetime(2026, 1, 1, tzinfo=UTC)
+
+    await logger.async_log_success_event(
+        {
+            "litellm_params": {"metadata": {"user_api_key_budget_reservation": reservation}},
+            "response_cost": response_cost,
+        },
+        InteractionsAPIResponse(id="interactions/background", status="in_progress"),
+        now,
+        now,
+    )
+
+    assert reservation["finalized"] is True
+    assert spend_counter_cache.in_memory_cache.get_cache(key=counter_key) == pytest.approx(expected_cost)
+
+
+@pytest.mark.asyncio
+@pytest.mark.parametrize(
+    "kwargs",
+    (
+        {},
+        {"litellm_params": {"metadata": None}},
+        {"litellm_params": {"metadata": "invalid"}},
+        {"litellm_params": {"metadata": {1: "ignored"}}},
+    ),
+)
+async def test_missing_or_invalid_metadata_leaves_existing_spend_unchanged(
+    spend_counter_cache: DualCache, kwargs: Mapping[str, object]
+) -> None:
+    counter_key: Final = "spend:key:hashed"
+    spend_counter_cache.in_memory_cache.set_cache(key=counter_key, value=0.00255)
+    logger: Final = BudgetReservationReconcileLogger()
+    now: Final = datetime(2026, 1, 1, tzinfo=UTC)
+
+    await logger.async_log_success_event(kwargs, object(), now, now)
+
+    assert spend_counter_cache.in_memory_cache.get_cache(key=counter_key) == pytest.approx(0.00255)