feat(budgets): reset tokens_used on the team budget window tick

On the budget_reset_at tick that zeroes spend, hard-reset the team's
tokens_used and invalidate the tokens:team:{id} counter so the next read
re-seeds from the freshly-zeroed column. The reset is team-scoped (token
budgets only apply to teams), so key/user resets are untouched.

litellm/proxy/common_utils/reset_budget_job.py

@@ -1336,6 +1336,9 @@ class ResetBudgetJob:
                         team_id = getattr(t.row, "team_id", None)
                         if team_id:
                             await self._invalidate_spend_counter(f"spend:team:{team_id}")
+                            # Drop the token-budget counter too, so the next read
+                            # re-seeds from the freshly-zeroed tokens_used column.
+                            await self._invalidate_spend_counter(f"tokens:team:{team_id}")
 
             end_time = time.time()
             outcome: Final = _ChunkOutcome(

litellm/repositories/unit_of_work.py

@@ -56,9 +56,15 @@ class TeamSpendResetWrites:
     table: BatchTable
 
     def queue_spend_reset(self, team_id: str, budget_reset_at: datetime | None, spend_decrement: float) -> None:
+        data = dict(_spend_reset_data(budget_reset_at, spend_decrement))
+        # Token budgets are team-only, so zero the cumulative tokens_used on the
+        # same tick that resets spend. A hard reset (vs spend's decrement) is
+        # acceptable at a budget boundary: a token allowance need not carry the
+        # handful of tokens that can race the reset the way billing-exact spend does.
+        data["tokens_used"] = 0
         self.table.update(
             where={"team_id": team_id},
-            data=_spend_reset_data(budget_reset_at, spend_decrement),
+            data=data,
         )