Merge 41ab7e57e8 into 3746ba58d7

litellm/proxy/management_endpoints/key_management_endpoints.py

@@ -107,6 +107,7 @@ from litellm.proxy.management_helpers.team_member_permission_checks import (
     TeamMemberPermissionChecks,
 )
 from litellm.proxy.management_helpers.utils import management_endpoint_wrapper
+from litellm.proxy.spend_tracking.budget_reservation import get_budget_window_start
 from litellm.proxy.spend_tracking.spend_tracking_utils import _is_master_key
 from litellm.proxy.ui_crud_endpoints.proxy_setting_endpoints import (
     get_ui_settings_cached,
@@ -3515,6 +3516,97 @@ async def _build_model_max_budget_usage(
     )
 
 
+def _budget_window_to_dict(window: object) -> Mapping[str, object] | None:
+    """Coerce a budget_limits entry to a dict; None when the entry is unusable."""
+    if isinstance(window, dict):
+        return window
+    model_dump: Final = getattr(window, "model_dump", None)
+    if not callable(model_dump):
+        return None
+    try:
+        dumped: Final = model_dump()
+    except Exception:  # noqa: BLE001  # model_dump implementations can raise arbitrary errors
+        return None
+    return dumped if isinstance(dumped, dict) else None
+
+
+def _coerce_budget_limits(budget_limits: object) -> Sequence[object] | None:
+    """Coerce budget_limits to a sequence of windows, parsing JSON strings; None when unusable."""
+    if isinstance(budget_limits, str):
+        try:
+            parsed: Final = json.loads(budget_limits)
+        except (TypeError, ValueError):
+            return None
+        return parsed if isinstance(parsed, list) else None
+    return budget_limits if isinstance(budget_limits, list) else None
+
+
+def _parse_window_max_budget(value: object) -> float | None:
+    """Coerce a window's max_budget to float; None when absent or unparseable."""
+    if isinstance(value, (int, float, str)):
+        try:
+            return float(value)
+        except (TypeError, ValueError):
+            return None
+    return None
+
+
+async def _budget_window_with_usage(window: Mapping[str, object], api_key_hash: str) -> Mapping[str, object]:
+    """
+    Return a copy of a budget window with current-window spend attached.
+
+    Per-window spend is not persisted in the DB; it lives in the cross-pod spend
+    counters (spend:key:{hashed_token}:window:{budget_duration}) that
+    _virtual_key_multi_budget_check enforces against, so we read the same
+    counters via get_current_spend. Passing max_budget + window_start makes the
+    read re-check against the authoritative spend-log aggregate when the counter
+    is stale-low (e.g. after a Redis flush), same as the enforcement path.
+    """
+    from litellm.proxy.proxy_server import get_current_spend
+
+    duration: Final = window.get("budget_duration")
+    if not duration:
+        return dict(window)  # mutable-ok: per-window response copy, built once per window
+    spend: Final = await get_current_spend(
+        counter_key=f"spend:key:{api_key_hash}:window:{duration}",
+        fallback_spend=0.0,
+        max_budget=_parse_window_max_budget(window.get("max_budget")),
+        window_entity_type="Key",
+        window_entity_id=api_key_hash,
+        window_start=get_budget_window_start(window),
+    )
+    return {**window, "current_spend": round(spend, 4)}  # mutable-ok: per-window response copy, built once per window
+
+
+async def _budget_limits_entry_with_usage(window: object, api_key_hash: str) -> object:
+    """Return the window as an enriched dict when dict-coercible; the original entry otherwise."""
+    coerced: Final = _budget_window_to_dict(window)
+    if not coerced:
+        return window
+    return await _budget_window_with_usage(window=coerced, api_key_hash=api_key_hash)
+
+
+async def _budget_limits_with_usage(budget_limits: object, api_key_hash: str) -> Sequence[object] | None:
+    """
+    Return budget_limits as window dicts with current-window spend attached.
+
+    None when budget_limits is not a usable (possibly JSON-encoded) list; the
+    caller keeps the original value then. Entries that are not dict-coercible
+    are preserved as-is.
+    """
+    windows: Final = _coerce_budget_limits(budget_limits)
+    if windows is None:
+        return None
+    return [  # mutable-ok: entries are awaited, so they cannot be built inside a frozen wrapper
+        await _budget_limits_entry_with_usage(window=window, api_key_hash=api_key_hash) for window in windows
+    ]
+
+
+# Caps per-request fan-out: each key with budget windows costs one spend-counter
+# read (worst case a SpendLogs aggregation) per window.
+MAX_KEY_INFO_KEYS_PER_REQUEST: Final = 100
+
+
 @router.post(
     "/v2/key/info",
     tags=["key management"],
@@ -3557,6 +3649,17 @@ async def info_key_fn_v2(
                 status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
                 detail={"message": "Malformed request. No keys passed in."},
             )
+        requested_key_count: Final = len(data.keys or ()) + len(data.key_aliases or ())
+        if requested_key_count > MAX_KEY_INFO_KEYS_PER_REQUEST:
+            raise HTTPException(
+                status_code=status.HTTP_422_UNPROCESSABLE_ENTITY,
+                detail={  # mutable-ok: one-shot HTTPException payload matching the sibling detail dict above; never mutated after construction
+                    "message": (
+                        f"Too many keys requested: {requested_key_count}. "
+                        f"At most {MAX_KEY_INFO_KEYS_PER_REQUEST} keys and key_aliases combined per request."
+                    )
+                },
+            )
 
         # Resolve key_aliases to tokens so we never pass token=None (unbounded query)
         tokens_to_query: Final = list(data.keys) if data.keys else []
@@ -3599,6 +3702,13 @@ async def info_key_fn_v2(
                     model_max_budget=model_max_budget,
                     user_api_key_cache=model_max_budget_limiter.dual_cache,
                 )
+            if k_token_hash:
+                budget_limits_usage = await _budget_limits_with_usage(
+                    budget_limits=k_dict.get("budget_limits"),
+                    api_key_hash=k_token_hash,
+                )
+                if budget_limits_usage is not None:
+                    k_dict["budget_limits"] = budget_limits_usage
 
             filtered_key_info.append(k_dict)
         return {"key": data.keys, "info": filtered_key_info}
@@ -3635,6 +3745,9 @@ async def info_key_fn(
         - model_max_budget: dict - Per-model budgets, e.g. {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}
         - model_max_budget_usage: dict | None - Current-window spend per model, present only when
           the key has per-model budgets
+        - budget_limits: list | None - Concurrent budget windows. Each entry includes
+          current_spend: spend accumulated in the window so far (read from the same cross-pod
+          spend counter the budget enforcement uses)
         - models: list - Model_name's the key is allowed to call
         - tpm_limit / rpm_limit: int | None - Tokens and requests per minute limits
         - metadata: dict - Metadata for the key, e.g. {"team": "core-infra"}
@@ -3714,6 +3827,12 @@ async def info_key_fn(
                 model_max_budget=model_max_budget,
                 user_api_key_cache=model_max_budget_limiter.dual_cache,
             )
+        budget_limits_usage: Final = await _budget_limits_with_usage(
+            budget_limits=key_info.get("budget_limits"),
+            api_key_hash=key_token_hash,
+        )
+        if budget_limits_usage is not None:
+            key_info["budget_limits"] = budget_limits_usage
 
         # Attach object_permission if object_permission_id is set
         key_info = await attach_object_permission_to_dict(key_info, prisma_client)

tests/test_litellm/proxy/management_endpoints/test_key_management_endpoints.py

@@ -538,6 +538,57 @@ async def test_key_generation_with_object_permission(monkeypatch):
     assert key_insert_calls[0]["data"].get("object_permission_id") == "objperm123"
 
 
+@pytest.mark.asyncio
+async def test_generate_key_with_soft_budget_creates_budget_row(monkeypatch):
+    """soft_budget on /key/generate must create a budget table row and link its budget_id to the key."""
+    mock_prisma_client = AsyncMock()
+    mock_prisma_client.jsonify_object = lambda data: data
+    mock_prisma_client.db = MagicMock()
+    mock_budget_create = AsyncMock(return_value=MagicMock(budget_id="budget-soft-123"))
+    mock_prisma_client.db.litellm_budgettable = MagicMock()
+    mock_prisma_client.db.litellm_budgettable.create = mock_budget_create
+
+    async def _insert_data_side_effect(*args, **kwargs):
+        if kwargs.get("table_name") == "user":
+            return MagicMock(models=[], spend=0)
+        return MagicMock(
+            token="hashed_token_soft",
+            litellm_budget_table=None,
+            object_permission=None,
+        )
+
+    mock_prisma_client.insert_data = AsyncMock(side_effect=_insert_data_side_effect)
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", mock_prisma_client)
+
+    from litellm.proxy._types import GenerateKeyRequest, LitellmUserRoles
+    from litellm.proxy.auth.user_api_key_auth import UserAPIKeyAuth
+    from litellm.proxy.management_endpoints.key_management_endpoints import (
+        generate_key_fn,
+    )
+
+    await generate_key_fn(
+        data=GenerateKeyRequest(soft_budget=5.0),
+        user_api_key_dict=UserAPIKeyAuth(
+            user_role=LitellmUserRoles.PROXY_ADMIN,
+            api_key="sk-1234",
+            user_id="admin-1",
+        ),
+    )
+
+    mock_budget_create.assert_awaited_once()
+    created_budget = mock_budget_create.call_args.kwargs["data"]
+    assert created_budget["soft_budget"] == 5.0
+    assert created_budget["created_by"] == "admin-1"
+
+    key_insert_calls = [
+        call.kwargs
+        for call in mock_prisma_client.insert_data.call_args_list
+        if call.kwargs.get("table_name") == "key"
+    ]
+    assert len(key_insert_calls) == 1
+    assert key_insert_calls[0]["data"].get("budget_id") == "budget-soft-123"
+
+
 @pytest.mark.asyncio
 async def test_generate_key_debug_log_never_contains_raw_token(monkeypatch, caplog):
     """Regression for LIT-4356: /key/generate must never emit the raw virtual key
@@ -13895,6 +13946,377 @@ async def test_info_key_fn_v2_budget_table_fallback(monkeypatch):
     mock_prisma_client.db.query_raw.assert_not_awaited()
 
 
+@pytest.mark.asyncio
+async def test_info_key_fn_budget_limits_includes_current_spend(monkeypatch):
+    """
+    /key/info should attach current_spend to each budget_limits window, read from
+    the same spend counter (spend:key:{token}:window:{duration}) that budget
+    enforcement uses.
+    """
+    from unittest.mock import AsyncMock, MagicMock
+
+    from litellm.proxy._types import LiteLLM_VerificationToken
+    from litellm.proxy.management_endpoints.key_management_endpoints import info_key_fn
+
+    test_key_token = "hashed_token_window_test"
+    budget_limits = [
+        {
+            "reset_at": "2026-08-15T18:00:00+00:00",
+            "max_budget": 2.0,
+            "budget_duration": "1h",
+        }
+    ]
+
+    mock_prisma_client = AsyncMock()
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", mock_prisma_client)
+    mock_user_api_key_cache = AsyncMock()
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.user_api_key_cache", mock_user_api_key_cache
+    )
+    mock_get_current_spend = AsyncMock(return_value=0.73)
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.get_current_spend", mock_get_current_spend
+    )
+
+    mock_key_info = MagicMock(spec=LiteLLM_VerificationToken)
+    mock_key_info.token = test_key_token
+    mock_key_info.object_permission_id = None
+    mock_key_info.user_id = "user-w"
+    mock_key_info.team_id = None
+    mock_key_info.litellm_budget_table = None
+    mock_key_info.model_dump.return_value = {
+        "token": test_key_token,
+        "budget_limits": [dict(w) for w in budget_limits],
+        "user_id": "user-w",
+        "team_id": None,
+        "object_permission_id": None,
+        "litellm_budget_table": None,
+    }
+    mock_key_info.dict.return_value = mock_key_info.model_dump.return_value
+
+    mock_prisma_client.db.litellm_verificationtoken.find_unique = AsyncMock(
+        return_value=mock_key_info
+    )
+
+    user_api_key_dict = UserAPIKeyAuth(
+        user_role=LitellmUserRoles.PROXY_ADMIN,
+        api_key="sk-test-window-key",
+    )
+
+    result = await info_key_fn(
+        key="sk-test-window-key",
+        user_api_key_dict=user_api_key_dict,
+    )
+
+    windows = result["info"]["budget_limits"]
+    assert len(windows) == 1
+    assert windows[0]["current_spend"] == 0.73
+    assert windows[0]["max_budget"] == 2.0
+    assert windows[0]["budget_duration"] == "1h"
+
+    mock_get_current_spend.assert_awaited_once()
+    call_kwargs = mock_get_current_spend.await_args.kwargs
+    assert call_kwargs["counter_key"] == f"spend:key:{test_key_token}:window:1h"
+    assert call_kwargs["max_budget"] == 2.0
+    assert call_kwargs["window_entity_type"] == "Key"
+    assert call_kwargs["window_entity_id"] == test_key_token
+    assert call_kwargs["window_start"] is not None
+
+
+@pytest.mark.asyncio
+async def test_info_key_fn_no_budget_limits_skips_spend_lookup(monkeypatch):
+    """Keys without budget_limits should not trigger window spend lookups."""
+    from unittest.mock import AsyncMock, MagicMock
+
+    from litellm.proxy._types import LiteLLM_VerificationToken
+    from litellm.proxy.management_endpoints.key_management_endpoints import info_key_fn
+
+    test_key_token = "hashed_token_no_windows"
+
+    mock_prisma_client = AsyncMock()
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", mock_prisma_client)
+    mock_user_api_key_cache = AsyncMock()
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.user_api_key_cache", mock_user_api_key_cache
+    )
+    mock_get_current_spend = AsyncMock(return_value=0.0)
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.get_current_spend", mock_get_current_spend
+    )
+
+    mock_key_info = MagicMock(spec=LiteLLM_VerificationToken)
+    mock_key_info.token = test_key_token
+    mock_key_info.object_permission_id = None
+    mock_key_info.user_id = "user-nw"
+    mock_key_info.team_id = None
+    mock_key_info.litellm_budget_table = None
+    mock_key_info.model_dump.return_value = {
+        "token": test_key_token,
+        "budget_limits": None,
+        "user_id": "user-nw",
+        "team_id": None,
+        "object_permission_id": None,
+        "litellm_budget_table": None,
+    }
+    mock_key_info.dict.return_value = mock_key_info.model_dump.return_value
+
+    mock_prisma_client.db.litellm_verificationtoken.find_unique = AsyncMock(
+        return_value=mock_key_info
+    )
+
+    user_api_key_dict = UserAPIKeyAuth(
+        user_role=LitellmUserRoles.PROXY_ADMIN,
+        api_key="sk-test-no-window-key",
+    )
+
+    result = await info_key_fn(
+        key="sk-test-no-window-key",
+        user_api_key_dict=user_api_key_dict,
+    )
+
+    assert result["info"]["budget_limits"] is None
+    mock_get_current_spend.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_info_key_fn_v2_budget_limits_includes_current_spend(monkeypatch):
+    """/v2/key/info should attach current_spend to each budget_limits window."""
+    from unittest.mock import AsyncMock, MagicMock
+
+    from litellm.proxy._types import KeyRequest, LiteLLM_VerificationToken
+    from litellm.proxy.management_endpoints.key_management_endpoints import (
+        info_key_fn_v2,
+    )
+
+    test_key_token = "hashed_token_v2_window_test"
+
+    mock_prisma_client = AsyncMock()
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", mock_prisma_client)
+    mock_user_api_key_cache = AsyncMock()
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.user_api_key_cache", mock_user_api_key_cache
+    )
+    mock_get_current_spend = AsyncMock(return_value=1.25)
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.get_current_spend", mock_get_current_spend
+    )
+
+    mock_key = MagicMock(spec=LiteLLM_VerificationToken)
+    mock_key.token = test_key_token
+    mock_key.user_id = "user-v2-w"
+    mock_key.team_id = None
+    mock_key.model_dump.return_value = {
+        "token": test_key_token,
+        "budget_limits": [
+            {
+                "reset_at": "2026-08-15T18:00:00+00:00",
+                "max_budget": 2.0,
+                "budget_duration": "1h",
+            },
+            {
+                "reset_at": "2026-08-16T00:00:00+00:00",
+                "max_budget": 20.0,
+                "budget_duration": "1d",
+            },
+        ],
+        "user_id": "user-v2-w",
+        "team_id": None,
+        "litellm_budget_table": None,
+    }
+    mock_key.dict.return_value = mock_key.model_dump.return_value
+
+    mock_prisma_client.get_data = AsyncMock(return_value=[mock_key])
+
+    user_api_key_dict = UserAPIKeyAuth(
+        user_role=LitellmUserRoles.PROXY_ADMIN,
+        api_key="sk-admin-v2-w",
+    )
+
+    result = await info_key_fn_v2(
+        data=KeyRequest(keys=[test_key_token]),
+        user_api_key_dict=user_api_key_dict,
+    )
+
+    assert len(result["info"]) == 1
+    windows = result["info"][0]["budget_limits"]
+    assert len(windows) == 2
+    assert windows[0]["current_spend"] == 1.25
+    assert windows[1]["current_spend"] == 1.25
+    assert mock_get_current_spend.await_count == 2
+    counter_keys = {
+        call.kwargs["counter_key"] for call in mock_get_current_spend.await_args_list
+    }
+    assert counter_keys == {
+        f"spend:key:{test_key_token}:window:1h",
+        f"spend:key:{test_key_token}:window:1d",
+    }
+
+
+@pytest.mark.asyncio
+async def test_info_key_fn_v2_rejects_oversized_batch(monkeypatch):
+    """/v2/key/info must reject over-cap batches before doing any DB work."""
+    from unittest.mock import AsyncMock
+
+    from litellm.proxy._types import KeyRequest, ProxyException
+    from litellm.proxy.management_endpoints.key_management_endpoints import (
+        MAX_KEY_INFO_KEYS_PER_REQUEST,
+        info_key_fn_v2,
+    )
+
+    mock_prisma_client = AsyncMock()
+    monkeypatch.setattr("litellm.proxy.proxy_server.prisma_client", mock_prisma_client)
+    monkeypatch.setattr("litellm.proxy.proxy_server.user_api_key_cache", AsyncMock())
+
+    user_api_key_dict = UserAPIKeyAuth(
+        user_role=LitellmUserRoles.PROXY_ADMIN,
+        api_key="sk-admin-batch-cap",
+    )
+
+    with pytest.raises(ProxyException) as exc_info:
+        await info_key_fn_v2(
+            data=KeyRequest(
+                keys=[f"hash-{i}" for i in range(MAX_KEY_INFO_KEYS_PER_REQUEST)],
+                key_aliases=["alias-over-cap"],
+            ),
+            user_api_key_dict=user_api_key_dict,
+        )
+
+    assert exc_info.value.code == "422"
+    mock_prisma_client.get_data.assert_not_awaited()
+
+
+@pytest.mark.asyncio
+async def test_budget_limits_with_usage_json_string_input(monkeypatch):
+    """budget_limits stored as a JSON string should be parsed and annotated."""
+    import json as json_module
+    from unittest.mock import AsyncMock
+
+    from litellm.proxy.management_endpoints.key_management_endpoints import (
+        _budget_limits_with_usage,
+    )
+
+    mock_get_current_spend = AsyncMock(return_value=0.5)
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.get_current_spend", mock_get_current_spend
+    )
+
+    raw = json_module.dumps(
+        [{"budget_duration": "1h", "max_budget": 2.0, "reset_at": None}]
+    )
+    result = await _budget_limits_with_usage(budget_limits=raw, api_key_hash="hash-1")
+
+    assert result == [
+        {
+            "budget_duration": "1h",
+            "max_budget": 2.0,
+            "reset_at": None,
+            "current_spend": 0.5,
+        }
+    ]
+    mock_get_current_spend.assert_awaited_once()
+
+
+@pytest.mark.asyncio
+async def test_budget_limits_with_usage_skips_unusable_inputs(monkeypatch):
+    """Invalid JSON strings, non-list values, and malformed windows are skipped."""
+    from unittest.mock import AsyncMock
+
+    from litellm.proxy.management_endpoints.key_management_endpoints import (
+        _budget_limits_with_usage,
+    )
+
+    mock_get_current_spend = AsyncMock(return_value=0.0)
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.get_current_spend", mock_get_current_spend
+    )
+
+    # invalid JSON string and non-list values return None: callers keep the original
+    assert await _budget_limits_with_usage(budget_limits="{not json", api_key_hash="hash-1") is None
+    assert await _budget_limits_with_usage(budget_limits={"budget_duration": "1h"}, api_key_hash="hash-1") is None
+
+    # windows that are falsy, missing budget_duration, or not dict-like
+    windows = [
+        {},
+        {"max_budget": 2.0},
+        {"budget_duration": "1h", "max_budget": "not-a-number"},
+        42,
+    ]
+    result = await _budget_limits_with_usage(budget_limits=windows, api_key_hash="hash-1")
+
+    # only the well-formed window (with unparseable max_budget coerced to None)
+    # triggers a spend lookup
+    mock_get_current_spend.assert_awaited_once()
+    call_kwargs = mock_get_current_spend.await_args.kwargs
+    assert call_kwargs["counter_key"] == "spend:key:hash-1:window:1h"
+    assert call_kwargs["max_budget"] is None
+    assert result is not None
+    assert result[0] == {}
+    assert result[1] == {"max_budget": 2.0}
+    assert result[2] == {"budget_duration": "1h", "max_budget": "not-a-number", "current_spend": 0.0}
+    assert result[3] == 42
+    # input is not mutated
+    assert windows[2] == {"budget_duration": "1h", "max_budget": "not-a-number"}
+
+
+@pytest.mark.asyncio
+async def test_budget_limits_with_usage_window_without_max_budget(monkeypatch):
+    """A window with only budget_duration still gets current_spend, read without a budget ceiling."""
+    from unittest.mock import AsyncMock
+
+    from litellm.proxy.management_endpoints.key_management_endpoints import (
+        _budget_limits_with_usage,
+    )
+
+    mock_get_current_spend = AsyncMock(return_value=0.75)
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.get_current_spend", mock_get_current_spend
+    )
+
+    result = await _budget_limits_with_usage(
+        budget_limits=[{"budget_duration": "2d"}], api_key_hash="hash-no-max"
+    )
+
+    assert result == [{"budget_duration": "2d", "current_spend": 0.75}]
+    call_kwargs = mock_get_current_spend.await_args.kwargs
+    assert call_kwargs["counter_key"] == "spend:key:hash-no-max:window:2d"
+    assert call_kwargs["max_budget"] is None
+
+
+@pytest.mark.asyncio
+async def test_budget_limits_with_usage_pydantic_windows(monkeypatch):
+    """Window objects with model_dump() are converted to dicts; failing windows pass through."""
+    from unittest.mock import AsyncMock, MagicMock
+
+    from litellm.proxy.management_endpoints.key_management_endpoints import (
+        _budget_limits_with_usage,
+    )
+
+    mock_get_current_spend = AsyncMock(return_value=1.0)
+    monkeypatch.setattr(
+        "litellm.proxy.proxy_server.get_current_spend", mock_get_current_spend
+    )
+
+    good_window = MagicMock()
+    good_window.model_dump.return_value = {
+        "budget_duration": "7d",
+        "max_budget": 10.0,
+        "reset_at": None,
+    }
+    bad_window = MagicMock()
+    bad_window.model_dump.side_effect = ValueError("boom")
+
+    result = await _budget_limits_with_usage(
+        budget_limits=[good_window, bad_window], api_key_hash="hash-2"
+    )
+
+    # good window converted to dict and annotated; failing window left as-is
+    assert result is not None
+    assert isinstance(result[0], dict)
+    assert result[0]["current_spend"] == 1.0
+    assert result[1] is bad_window
+    mock_get_current_spend.assert_awaited_once()
+
+
 @pytest.mark.asyncio
 async def test_info_key_fn_reads_the_configured_budget_model_key(monkeypatch):
     """/key/info reads the one counter enforcement reads: the configured budget model.

ui/litellm-dashboard/src/lib/http/schema.d.ts

@@ -7489,6 +7489,9 @@ export interface paths {
          *         - model_max_budget: dict - Per-model budgets, e.g. {"gpt-4": {"budget_limit": 0.0005, "time_period": "30d"}}
          *         - model_max_budget_usage: dict | None - Current-window spend per model, present only when
          *           the key has per-model budgets
+         *         - budget_limits: list | None - Concurrent budget windows. Each entry includes
+         *           current_spend: spend accumulated in the window so far (read from the same cross-pod
+         *           spend counter the budget enforcement uses)
          *         - models: list - Model_name's the key is allowed to call
          *         - tpm_limit / rpm_limit: int | None - Tokens and requests per minute limits
          *         - metadata: dict - Metadata for the key, e.g. {"team": "core-infra"}