From 26611dd4bd953b4180a55d219d6f92fbd3955ca4 Mon Sep 17 00:00:00 2001 From: Ryan Crabbe Date: Thu, 29 Jan 2026 16:40:55 -0800 Subject: [PATCH 01/41] fix: dead code cleanup in MCP server error handler raise e makes the error logging and JSON error response below it unreachable --- litellm/proxy/_experimental/mcp_server/server.py | 1 - 1 file changed, 1 deletion(-) diff --git a/litellm/proxy/_experimental/mcp_server/server.py b/litellm/proxy/_experimental/mcp_server/server.py index 6d54c3871e5..545d5956bd2 100644 --- a/litellm/proxy/_experimental/mcp_server/server.py +++ b/litellm/proxy/_experimental/mcp_server/server.py @@ -1898,7 +1898,6 @@ if MCP_AVAILABLE: await session_manager.handle_request(scope, receive, send) except Exception as e: - raise e verbose_logger.exception(f"Error handling MCP request: {e}") # Instead of re-raising, try to send a graceful error response try: From f9e8f8712b1b4eb8b619a9ff18630d84bfdfcc6e Mon Sep 17 00:00:00 2001 From: Ryan Crabbe Date: Tue, 3 Feb 2026 16:51:45 -0800 Subject: [PATCH 02/41] fix: add cache invalidation for _cached_get_model_group_info on deployment changes _cached_get_model_group_info uses @lru_cache but had no invalidation, causing stale model group info (TPM/RPM limits) after dynamic deployment changes. Add cache_clear() at all 5 model_list mutation sites. --- litellm/router.py | 12 ++++++ tests/test_litellm/test_router.py | 67 +++++++++++++++++++++++++++++++ 2 files changed, 79 insertions(+) diff --git a/litellm/router.py b/litellm/router.py index d01c8443dab..5b72c3fb669 100644 --- a/litellm/router.py +++ b/litellm/router.py @@ -6055,6 +6055,7 @@ class Router: self.model_list = [] self.model_id_to_deployment_index_map = {} # Reset the index self.model_name_to_deployment_indices = {} # Reset the model_name index + self._invalidate_model_group_info_cache() # we add api_base/api_key each model so load balancing between azure/gpt on api_base1 and api_base2 works for model in original_model_list: @@ -6358,6 +6359,7 @@ class Router: """ idx = len(self.model_list) self.model_list.append(model) + self._invalidate_model_group_info_cache() # Update model_id index for O(1) lookup if model_id is not None: @@ -6405,6 +6407,7 @@ class Router: if removal_idx is not None: self.model_list.pop(removal_idx) + self._invalidate_model_group_info_cache() self._update_deployment_indices_after_removal( model_id=deployment_id, removal_idx=removal_idx ) @@ -6438,6 +6441,7 @@ class Router: if deployment_idx is not None: # Pop the item from the list first item = self.model_list.pop(deployment_idx) + self._invalidate_model_group_info_cache() self._update_deployment_indices_after_removal( model_id=id, removal_idx=deployment_idx ) @@ -7172,6 +7176,7 @@ class Router: """ # First populate the model_list self.model_list = [] + self._invalidate_model_group_info_cache() for _, model in enumerate(model_list): # Extract model_info from the model dict model_info = model.get("model_info", {}) @@ -7508,6 +7513,13 @@ class Router: return returned_models + def _invalidate_model_group_info_cache(self) -> None: + """Invalidate the cached model group info. + + Call this whenever self.model_list is modified to ensure the cache is rebuilt. + """ + self._cached_get_model_group_info.cache_clear() + def get_model_access_groups( self, model_name: Optional[str] = None, diff --git a/tests/test_litellm/test_router.py b/tests/test_litellm/test_router.py index 08ae804ea80..48aa435c3a9 100644 --- a/tests/test_litellm/test_router.py +++ b/tests/test_litellm/test_router.py @@ -925,6 +925,73 @@ def test_router_get_model_access_groups_team_only_models(): assert list(access_groups.keys()) == ["default-models"] +def test_cached_get_model_group_info(): + """ + Test that _cached_get_model_group_info caches results and + invalidates on deployment changes. + """ + from litellm.types.router import Deployment, LiteLLM_Params + + router = litellm.Router( + model_list=[ + { + "model_name": "gpt-4", + "litellm_params": {"model": "gpt-4", "api_key": "fake"}, + "model_info": {"tpm": 1000, "rpm": 100}, + }, + ] + ) + + # First call should compute and cache + result1 = router._cached_get_model_group_info("gpt-4") + assert result1 is not None + assert result1.tpm == 1000 + + # Second call should hit cache (same object) + result2 = router._cached_get_model_group_info("gpt-4") + assert result1 is result2 + + # Add a deployment — cache should be invalidated + router.add_deployment( + Deployment( + model_name="gpt-4", + litellm_params=LiteLLM_Params(model="gpt-4", api_key="fake2"), + model_info={"tpm": 2000, "rpm": 200}, + ) + ) + result3 = router._cached_get_model_group_info("gpt-4") + assert result3 is not result2 + assert result3 is not None + assert result3.tpm == 3000 # 1000 + 2000 + + # Delete a deployment — cache should be invalidated + deployment_id = router.model_list[-1]["model_info"]["id"] + router.delete_deployment(id=deployment_id) + result4 = router._cached_get_model_group_info("gpt-4") + assert result4 is not result3 + assert result4 is not None + assert result4.tpm == 1000 + + # set_model_list — cache should be invalidated + router.set_model_list( + [ + { + "model_name": "gpt-4", + "litellm_params": {"model": "gpt-4", "api_key": "fake"}, + "model_info": {"tpm": 5000}, + }, + ] + ) + result5 = router._cached_get_model_group_info("gpt-4") + assert result5 is not result4 + assert result5 is not None + assert result5.tpm == 5000 + + # Verify cache still works after invalidation + result6 = router._cached_get_model_group_info("gpt-4") + assert result5 is result6 + + @pytest.mark.asyncio async def test_acompletion_streaming_iterator(): """Test _acompletion_streaming_iterator for normal streaming and fallback behavior.""" From 6743d20de262338750f9b9046606222812915877 Mon Sep 17 00:00:00 2001 From: Ryan Crabbe Date: Thu, 5 Feb 2026 10:50:45 -0800 Subject: [PATCH 03/41] docs: add trailing slash to /mcp endpoint URLs The /mcp endpoint requires a trailing slash because the MCP server is mounted as a sub-application using app.mount(). Starlette's mount behavior causes a 307 redirect from /mcp to /mcp/, which many MCP clients fail to handle. Updates documentation examples to use /mcp/ consistently. --- README.md | 2 +- docs/my-website/docs/mcp.md | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 77adddf8978..e7701b5cf9d 100644 --- a/README.md +++ b/README.md @@ -203,7 +203,7 @@ curl -X POST 'http://0.0.0.0:4000/v1/chat/completions' \ { "mcpServers": { "LiteLLM": { - "url": "http://localhost:4000/mcp", + "url": "http://localhost:4000/mcp/", "headers": { "x-litellm-api-key": "Bearer sk-1234" } diff --git a/docs/my-website/docs/mcp.md b/docs/my-website/docs/mcp.md index d63b55ee29e..564a054aae2 100644 --- a/docs/my-website/docs/mcp.md +++ b/docs/my-website/docs/mcp.md @@ -632,7 +632,7 @@ import asyncio config = { "mcpServers": { "mcp_group": { - "url": "http://localhost:4000/mcp", + "url": "http://localhost:4000/mcp/", "headers": { "x-mcp-servers": "dev_group", # assume this gives access to github, zapier and deepwiki "x-litellm-api-key": "Bearer sk-1234", From 82f6d0fe431b1edb65ff6cd1c322fa7c588ba992 Mon Sep 17 00:00:00 2001 From: Alejandro Tapia Date: Thu, 12 Feb 2026 14:47:09 -0800 Subject: [PATCH 04/41] healthcheck-model_id-fix: There was quite a bit of code that needed to be changed since health checks were entirely keyed by model name. This includes, proxy logic, the dashboard, and even networking, because model name was the identifier everywhere. All changed files had tests added to them, which are passing with no regressions. --- litellm/proxy/health_check.py | 12 +- .../health_endpoints/_health_endpoints.py | 8 +- .../litellm_utils_tests/test_health_check.py | 44 ++++ .../proxy/test_health_check_functions.py | 39 ++++ .../ModelsAndEndpointsView.test.tsx | 51 ++++- .../ModelsAndEndpointsView.tsx | 9 +- .../HealthCheckComponent.test.tsx | 147 +++++++++++++ .../model_dashboard/HealthCheckComponent.tsx | 199 ++++++++---------- .../model_dashboard/health_check_columns.tsx | 33 +-- .../src/components/networking.test.ts | 54 +++++ .../src/components/networking.tsx | 10 +- 11 files changed, 469 insertions(+), 137 deletions(-) create mode 100644 ui/litellm-dashboard/src/components/model_dashboard/HealthCheckComponent.test.tsx diff --git a/litellm/proxy/health_check.py b/litellm/proxy/health_check.py index 427a16a9801..48a20833cb2 100644 --- a/litellm/proxy/health_check.py +++ b/litellm/proxy/health_check.py @@ -190,10 +190,15 @@ async def perform_health_check( model: Optional[str] = None, cli_model: Optional[str] = None, details: Optional[bool] = True, + model_id: Optional[str] = None, ): """ Perform a health check on the system. + When model_id is provided, only the deployment with that id is checked + (so models that share the same name but have different ids are checked separately). + When model (name) is provided, all deployments matching that name are checked. + Returns: (bool): True if the health check passes, False otherwise. """ @@ -205,7 +210,12 @@ async def perform_health_check( else: return [], [] - if model is not None: + # Filter by model_id first so a single deployment is checked when id is specified + if model_id is not None: + _by_id = [x for x in model_list if (x.get("model_info") or {}).get("id") == model_id] + if _by_id: + model_list = _by_id + elif model is not None: _new_model_list = [ x for x in model_list if x["litellm_params"]["model"] == model ] diff --git a/litellm/proxy/health_endpoints/_health_endpoints.py b/litellm/proxy/health_endpoints/_health_endpoints.py index da90696ec2d..b844a959622 100644 --- a/litellm/proxy/health_endpoints/_health_endpoints.py +++ b/litellm/proxy/health_endpoints/_health_endpoints.py @@ -238,7 +238,7 @@ async def health_services_endpoint( # noqa: PLR0915 service_in_success_callbacks = True else: for cb in litellm.success_callback: - if hasattr(cb, 'callback_name') and cb.callback_name == service: + if getattr(cb, "callback_name", None) == service: service_in_success_callbacks = True break cb_id = get_callback_identifier(cb) @@ -732,7 +732,11 @@ async def _perform_health_check_and_save( ): """Helper function to perform health check and save results to database""" healthy_endpoints, unhealthy_endpoints = await perform_health_check( - model_list=model_list, cli_model=cli_model, model=target_model, details=details + model_list=model_list, + cli_model=cli_model, + model=target_model, + details=details, + model_id=model_id, ) # Optionally save health check result to database (non-blocking) diff --git a/tests/litellm_utils_tests/test_health_check.py b/tests/litellm_utils_tests/test_health_check.py index 19882bbe4be..5123b0a7789 100644 --- a/tests/litellm_utils_tests/test_health_check.py +++ b/tests/litellm_utils_tests/test_health_check.py @@ -471,6 +471,50 @@ def test_update_litellm_params_for_health_check(): ) +@pytest.mark.asyncio +async def test_perform_health_check_filters_by_model_id(): + """ + When model_id is passed, only that deployment is checked (not all deployments + that share the same model name). + """ + from litellm.proxy.health_check import perform_health_check + + # Two deployments with same model_name but different ids + model_list = [ + { + "model_name": "gpt-4", + "model_info": {"id": "deployment-id-1"}, + "litellm_params": {"model": "gpt-4", "api_key": "fake-key-1"}, + }, + { + "model_name": "gpt-4", + "model_info": {"id": "deployment-id-2"}, + "litellm_params": {"model": "gpt-4", "api_key": "fake-key-2"}, + }, + ] + + captured_list = [] + + async def mock_perform_health_check(m_list, details=True): + captured_list.append(m_list) + return [{"model": "gpt-4", "api_key": m_list[0]["litellm_params"]["api_key"]}], [] + + with patch( + "litellm.proxy.health_check._perform_health_check", + side_effect=mock_perform_health_check, + ): + healthy_endpoints, unhealthy_endpoints = await perform_health_check( + model_list=model_list, model_id="deployment-id-2", details=True + ) + + # Only one deployment (deployment-id-2) should have been passed to _perform_health_check + assert len(captured_list) == 1 + assert len(captured_list[0]) == 1 + assert (captured_list[0][0].get("model_info") or {}).get("id") == "deployment-id-2" + assert len(healthy_endpoints) == 1 + assert healthy_endpoints[0]["api_key"] == "fake-key-2" + + @pytest.mark.asyncio async def test_perform_health_check_with_health_check_model(): """ diff --git a/tests/test_litellm/proxy/test_health_check_functions.py b/tests/test_litellm/proxy/test_health_check_functions.py index ccae9fb5425..4c91d0ae91e 100644 --- a/tests/test_litellm/proxy/test_health_check_functions.py +++ b/tests/test_litellm/proxy/test_health_check_functions.py @@ -12,6 +12,7 @@ sys.path.insert(0, os.path.abspath("../../..")) from litellm.proxy.health_endpoints._health_endpoints import ( _aggregate_health_check_results, _build_model_param_to_info_mapping, + _perform_health_check_and_save, _save_background_health_checks_to_db, _save_health_check_results_if_changed, _save_health_check_to_db, @@ -466,5 +467,43 @@ async def test_get_all_latest_health_checks_without_model_id(mock_prisma): assert result[0].checked_at == mock_check2.checked_at # Latest +@pytest.mark.asyncio +async def test_perform_health_check_and_save_passes_model_id_to_perform_health_check(): + """Test that _perform_health_check_and_save passes model_id to perform_health_check so health checks run by model id.""" + model_list = [ + { + "model_name": "gpt-4", + "model_info": {"id": "deployment-abc"}, + "litellm_params": {"model": "gpt-4"}, + }, + ] + healthy = [{"model": "gpt-4"}] + unhealthy = [] + + async def mock_perform_health_check(model_list, model=None, cli_model=None, details=True, model_id=None): + return healthy, unhealthy + + with patch( + "litellm.proxy.health_endpoints._health_endpoints.perform_health_check", + side_effect=mock_perform_health_check, + ) as mock_perform: + result = await _perform_health_check_and_save( + model_list=model_list, + target_model=None, + cli_model=None, + details=True, + prisma_client=None, + start_time=0.0, + user_id="user-1", + model_id="deployment-abc", + ) + + mock_perform.assert_called_once() + call_kwargs = mock_perform.call_args[1] + assert call_kwargs["model_id"] == "deployment-abc" + assert result["healthy_count"] == 1 + assert result["unhealthy_count"] == 0 + + if __name__ == "__main__": pytest.main([__file__]) \ No newline at end of file diff --git a/ui/litellm-dashboard/src/app/(dashboard)/models-and-endpoints/ModelsAndEndpointsView.test.tsx b/ui/litellm-dashboard/src/app/(dashboard)/models-and-endpoints/ModelsAndEndpointsView.test.tsx index 1e8eabaea2e..b0df37ad6d8 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/models-and-endpoints/ModelsAndEndpointsView.test.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/models-and-endpoints/ModelsAndEndpointsView.test.tsx @@ -1,6 +1,6 @@ /* @vitest-environment jsdom */ import { QueryClient, QueryClientProvider } from "@tanstack/react-query"; -import { render } from "@testing-library/react"; +import { act, render } from "@testing-library/react"; import { beforeEach, describe, expect, it, vi } from "vitest"; import ModelsAndEndpointsView from "./ModelsAndEndpointsView"; @@ -13,6 +13,8 @@ vi.mock("@/components/networking", () => ({ getCallbacksCall: vi.fn().mockResolvedValue({ router_settings: {} }), setCallbacksCall: vi.fn().mockResolvedValue(undefined), getUiSettings: vi.fn().mockResolvedValue({ values: {} }), + latestHealthChecksCall: vi.fn().mockResolvedValue({ latest_health_checks: {} }), + getModelCostMapReloadStatus: vi.fn().mockResolvedValue({}), })); vi.mock("@/app/(dashboard)/models-and-endpoints/components/ModelAnalyticsTab/ModelAnalyticsTab", () => ({ @@ -27,6 +29,14 @@ vi.mock("@/components/add_model/AddModelForm", () => ({ default: () => null, })); +const mockHealthCheckComponent = vi.fn((_props: { all_models_on_proxy?: string[] }) => null); +vi.mock("@/components/model_dashboard/HealthCheckComponent", () => ({ + default: (props: { all_models_on_proxy?: string[] }) => { + mockHealthCheckComponent(props); + return null; + }, +})); + vi.mock("@/app/(dashboard)/hooks/useTeams", () => ({ default: () => ({ teams: [], @@ -104,4 +114,43 @@ describe("ModelsAndEndpointsView", () => { ); expect(await findByText("Model Management", {}, { timeout: 10000 })).toBeInTheDocument(); }, 15000); + + it("should pass model IDs (not model names) to HealthCheckComponent as all_models_on_proxy", async () => { + mockHealthCheckComponent.mockClear(); + const modelDataWithIds = { + data: [ + { model_name: "gpt-4", model_info: { id: "deployment-id-1" } }, + { model_name: "gpt-4", model_info: { id: "deployment-id-2" } }, + ], + }; + mockUseModelsInfo.mockReturnValue({ + data: { data: modelDataWithIds.data }, + isLoading: false, + refetch: vi.fn(), + }); + + const queryClient = createQueryClient(); + const { getByRole } = render( + + {}} + premiumUser={false} + teams={[]} + /> + , + ); + + const healthStatusTab = getByRole("tab", { name: "Health Status" }); + await act(async () => { + healthStatusTab.click(); + }); + + expect(mockHealthCheckComponent).toHaveBeenCalled(); + const healthCheckProps = mockHealthCheckComponent.mock.calls[0][0]; + expect(healthCheckProps.all_models_on_proxy).toEqual(["deployment-id-1", "deployment-id-2"]); + expect(healthCheckProps.all_models_on_proxy).not.toContain("gpt-4"); + }); }); diff --git a/ui/litellm-dashboard/src/app/(dashboard)/models-and-endpoints/ModelsAndEndpointsView.tsx b/ui/litellm-dashboard/src/app/(dashboard)/models-and-endpoints/ModelsAndEndpointsView.tsx index 9d77774cb4c..b697a859dc5 100644 --- a/ui/litellm-dashboard/src/app/(dashboard)/models-and-endpoints/ModelsAndEndpointsView.tsx +++ b/ui/litellm-dashboard/src/app/(dashboard)/models-and-endpoints/ModelsAndEndpointsView.tsx @@ -98,6 +98,13 @@ const ModelsAndEndpointsView: React.FC = ({ premiumUser, te return modelDataResponse.data.map((model: any) => model.model_name); }, [modelDataResponse?.data]); + const allModelIdsOnProxy = useMemo(() => { + if (!modelDataResponse?.data) return []; + return modelDataResponse.data + .map((model: any) => model.model_info?.id) + .filter((id: string | undefined): id is string => Boolean(id)); + }, [modelDataResponse?.data]); + const getProviderFromModel = (model: string) => { if (modelCostMapData !== null && modelCostMapData !== undefined) { if (typeof modelCostMapData == "object" && model in modelCostMapData) { @@ -397,7 +404,7 @@ const ModelsAndEndpointsView: React.FC = ({ premiumUser, te ({ + individualModelHealthCheckCall: (...args: unknown[]) => mockIndividualModelHealthCheckCall(...args), + latestHealthChecksCall: (...args: unknown[]) => mockLatestHealthChecksCall(...args), +})); + +describe("HealthCheckComponent", () => { + const getDisplayModelName = (model: { model_name?: string }) => model.model_name ?? ""; + + beforeEach(() => { + vi.clearAllMocks(); + mockLatestHealthChecksCall.mockResolvedValue({ latest_health_checks: {} }); + mockIndividualModelHealthCheckCall.mockResolvedValue({ + healthy_count: 1, + unhealthy_count: 0, + healthy_endpoints: [], + unhealthy_endpoints: [], + }); + }); + + it("should render the health check section", async () => { + const modelData = { + data: [ + { + model_name: "gpt-4", + model_info: { id: "deployment-1" }, + litellm_model_name: "gpt-4", + }, + ], + }; + + await act(async () => { + render( + , + ); + }); + await act(async () => { + await new Promise((r) => setTimeout(r, 0)); + }); + + expect(screen.getByText("Model Health Status")).toBeInTheDocument(); + expect( + screen.getByText("Run health checks on individual models to verify they are working correctly"), + ).toBeInTheDocument(); + }); + + it("should call individualModelHealthCheckCall with model id when run health check is triggered", async () => { + const modelData = { + data: [ + { + model_name: "gpt-4", + model_info: { id: "deployment-abc-123" }, + litellm_model_name: "gpt-4", + }, + ], + }; + + render( + , + ); + + const runButtons = screen.getAllByTestId("run-health-check-btn"); + expect(runButtons.length).toBeGreaterThanOrEqual(1); + const runButton = runButtons[0]; + + await act(async () => { + runButton.click(); + }); + + await act(async () => { + await new Promise((r) => setTimeout(r, 50)); + }); + + expect(mockIndividualModelHealthCheckCall).toHaveBeenCalledWith("token-123", "deployment-abc-123"); + expect(mockIndividualModelHealthCheckCall).not.toHaveBeenCalledWith("token-123", "gpt-4"); + }); + + it("should key health status by model id and show status from latest_health_checks by model_id", async () => { + const modelData = { + data: [ + { + model_name: "gpt-4", + model_info: { id: "id-alpha" }, + litellm_model_name: "gpt-4", + }, + { + model_name: "gpt-4", + model_info: { id: "id-beta" }, + litellm_model_name: "gpt-4", + }, + ], + }; + + mockLatestHealthChecksCall.mockResolvedValue({ + latest_health_checks: { + "id-alpha": { + status: "healthy", + checked_at: "2024-01-15T10:00:00Z", + error_message: null, + }, + "id-beta": { + status: "unhealthy", + checked_at: "2024-01-15T10:05:00Z", + error_message: "Connection failed", + }, + }, + }); + + await act(async () => { + render( + , + ); + }); + + await act(async () => { + await new Promise((r) => setTimeout(r, 0)); + }); + + expect(mockLatestHealthChecksCall).toHaveBeenCalledWith("token"); + const healthyBadges = screen.getAllByText("healthy"); + const unhealthyBadges = screen.getAllByText("unhealthy"); + expect(healthyBadges.length).toBeGreaterThanOrEqual(1); + expect(unhealthyBadges.length).toBeGreaterThanOrEqual(1); + }); +}); diff --git a/ui/litellm-dashboard/src/components/model_dashboard/HealthCheckComponent.tsx b/ui/litellm-dashboard/src/components/model_dashboard/HealthCheckComponent.tsx index 6fa83494408..b4bb1019dd6 100644 --- a/ui/litellm-dashboard/src/components/model_dashboard/HealthCheckComponent.tsx +++ b/ui/litellm-dashboard/src/components/model_dashboard/HealthCheckComponent.tsx @@ -53,31 +53,33 @@ const HealthCheckComponent: React.FC = ({ const healthTableRef = useRef>(null); - // Initialize health statuses on component mount + // Initialize health statuses on component mount (keyed by model id) useEffect(() => { if (!accessToken || !modelData?.data) return; const initializeHealthStatuses = async () => { const healthStatusMap: { [key: string]: HealthStatus } = {}; - // Initialize all models with default state using model names + // Initialize all models with default state using model ids modelData.data.forEach((model: any) => { - const modelName = model.model_name; - healthStatusMap[modelName] = { - status: "none", - lastCheck: "None", - lastSuccess: "None", - loading: false, - error: undefined, - fullError: undefined, - successResponse: undefined, - }; + const modelId = model.model_info?.id; + if (modelId) { + healthStatusMap[modelId] = { + status: "none", + lastCheck: "None", + lastSuccess: "None", + loading: false, + error: undefined, + fullError: undefined, + successResponse: undefined, + }; + } }); try { const latestHealthChecks = await latestHealthChecksCall(accessToken); - // Override with actual database data if it exists + // Override with actual database data if it exists (latest_health_checks is keyed by model_id) if ( latestHealthChecks && latestHealthChecks.latest_health_checks && @@ -86,32 +88,28 @@ const HealthCheckComponent: React.FC = ({ Object.entries(latestHealthChecks.latest_health_checks).forEach(([key, checkData]: [string, any]) => { if (!checkData) return; - let targetModelName: string | null = null; + let targetModelId: string | null = null; - // The key could be either model_id or model_name, try both approaches - const directModelMatch = modelData.data.find((m: any) => m.model_name === key); - if (directModelMatch) { - targetModelName = directModelMatch.model_name; + // The key is model_id from the backend; fallback to matching by model_name for legacy data + const modelByIdMatch = modelData.data.find((m: any) => m.model_info && m.model_info.id === key); + if (modelByIdMatch) { + targetModelId = modelByIdMatch.model_info.id; } else { - // If not a direct match, treat as model_id and find the corresponding model - const modelByIdMatch = modelData.data.find((m: any) => m.model_info && m.model_info.id === key); - if (modelByIdMatch) { - targetModelName = modelByIdMatch.model_name; - } else { - // Check if checkData contains model_name and use that - if (checkData.model_name) { - const modelByNameInData = modelData.data.find((m: any) => m.model_name === checkData.model_name); - if (modelByNameInData) { - targetModelName = modelByNameInData.model_name; - } + const directModelMatch = modelData.data.find((m: any) => m.model_name === key); + if (directModelMatch?.model_info?.id) { + targetModelId = directModelMatch.model_info.id; + } else if (checkData.model_name) { + const modelByNameInData = modelData.data.find((m: any) => m.model_name === checkData.model_name); + if (modelByNameInData?.model_info?.id) { + targetModelId = modelByNameInData.model_info.id; } } } - if (targetModelName) { + if (targetModelId) { const fullError = checkData.error_message || undefined; - healthStatusMap[targetModelName] = { + healthStatusMap[targetModelId] = { status: checkData.status || "unknown", lastCheck: checkData.checked_at ? new Date(checkData.checked_at).toLocaleString() : "None", lastSuccess: @@ -246,33 +244,31 @@ const HealthCheckComponent: React.FC = ({ return cleaned.length > 100 ? cleaned.substring(0, 97) + "..." : cleaned; }; - const runIndividualHealthCheck = async (modelName: string) => { + const runIndividualHealthCheck = async (modelId: string) => { if (!accessToken) return; setModelHealthStatuses((prev) => ({ ...prev, - [modelName]: { - ...prev[modelName], + [modelId]: { + ...prev[modelId], loading: true, status: "checking", }, })); try { - // Run the health check and process the response directly - const response = await individualModelHealthCheckCall(accessToken, modelName); + const response = await individualModelHealthCheckCall(accessToken, modelId); const currentTime = new Date().toLocaleString(); - // Check if there are any unhealthy endpoints (which means this specific model failed) if (response.unhealthy_count > 0 && response.unhealthy_endpoints && response.unhealthy_endpoints.length > 0) { const rawError = response.unhealthy_endpoints[0]?.error || "Health check failed"; const errorMessage = extractMeaningfulError(rawError); setModelHealthStatuses((prev) => ({ ...prev, - [modelName]: { + [modelId]: { status: "unhealthy", lastCheck: currentTime, - lastSuccess: prev[modelName]?.lastSuccess || "None", + lastSuccess: prev[modelId]?.lastSuccess || "None", loading: false, error: errorMessage, fullError: rawError, @@ -281,7 +277,7 @@ const HealthCheckComponent: React.FC = ({ } else { setModelHealthStatuses((prev) => ({ ...prev, - [modelName]: { + [modelId]: { status: "healthy", lastCheck: currentTime, lastSuccess: currentTime, @@ -291,41 +287,33 @@ const HealthCheckComponent: React.FC = ({ })); } - // Refresh health status from database to get the saved check data including timestamp try { const latestHealthChecks = await latestHealthChecksCall(accessToken); + const checkData = latestHealthChecks.latest_health_checks?.[modelId]; - // Find the model ID for this model name to look up database data - const model = modelData.data.find((m: any) => m.model_name === modelName); - if (model) { - const modelId = model.model_info.id; - const checkData = latestHealthChecks.latest_health_checks?.[modelId]; - - if (checkData) { - const fullError = checkData.error_message || undefined; - setModelHealthStatuses((prev) => ({ - ...prev, - [modelName]: { - status: checkData.status || prev[modelName]?.status || "unknown", - lastCheck: checkData.checked_at - ? new Date(checkData.checked_at).toLocaleString() - : prev[modelName]?.lastCheck || "None", - lastSuccess: - checkData.status === "healthy" - ? checkData.checked_at - ? new Date(checkData.checked_at).toLocaleString() - : prev[modelName]?.lastSuccess || "None" - : prev[modelName]?.lastSuccess || "None", - loading: false, - error: fullError ? extractMeaningfulError(fullError) : prev[modelName]?.error, - fullError: fullError || prev[modelName]?.fullError, - successResponse: checkData.status === "healthy" ? checkData : prev[modelName]?.successResponse, - }, - })); - } + if (checkData) { + const fullError = checkData.error_message || undefined; + setModelHealthStatuses((prev) => ({ + ...prev, + [modelId]: { + status: checkData.status || prev[modelId]?.status || "unknown", + lastCheck: checkData.checked_at + ? new Date(checkData.checked_at).toLocaleString() + : prev[modelId]?.lastCheck || "None", + lastSuccess: + checkData.status === "healthy" + ? checkData.checked_at + ? new Date(checkData.checked_at).toLocaleString() + : prev[modelId]?.lastSuccess || "None" + : prev[modelId]?.lastSuccess || "None", + loading: false, + error: fullError ? extractMeaningfulError(fullError) : prev[modelId]?.error, + fullError: fullError || prev[modelId]?.fullError, + successResponse: checkData.status === "healthy" ? checkData : prev[modelId]?.successResponse, + }, + })); } } catch (dbError) { - // Ignore database errors - we already have the health check result from the API call console.debug("Could not fetch updated status from database (non-critical):", dbError); } } catch (error) { @@ -334,10 +322,10 @@ const HealthCheckComponent: React.FC = ({ const errorMessage = extractMeaningfulError(rawError); setModelHealthStatuses((prev) => ({ ...prev, - [modelName]: { + [modelId]: { status: "unhealthy", lastCheck: currentTime, - lastSuccess: prev[modelName]?.lastSuccess || "None", + lastSuccess: prev[modelId]?.lastSuccess || "None", loading: false, error: errorMessage, fullError: rawError, @@ -349,11 +337,10 @@ const HealthCheckComponent: React.FC = ({ const runAllHealthChecks = async () => { const modelsToCheck = selectedModelsForHealth.length > 0 ? selectedModelsForHealth : all_models_on_proxy; - // Set all models to loading state const loadingStatuses = modelsToCheck.reduce( - (acc, modelName) => { - acc[modelName] = { - ...modelHealthStatuses[modelName], + (acc, modelId) => { + acc[modelId] = { + ...modelHealthStatuses[modelId], loading: true, status: "checking", }; @@ -364,30 +351,25 @@ const HealthCheckComponent: React.FC = ({ setModelHealthStatuses((prev) => ({ ...prev, ...loadingStatuses })); - // Store results from individual health checks const healthCheckResults: { [key: string]: any } = {}; - // Run all health checks in parallel and collect results - const healthCheckPromises = modelsToCheck.map(async (modelName) => { + const healthCheckPromises = modelsToCheck.map(async (modelId) => { if (!accessToken) return; try { - // Run the health check and store the result - const response = await individualModelHealthCheckCall(accessToken, modelName); - healthCheckResults[modelName] = response; + const response = await individualModelHealthCheckCall(accessToken, modelId); + healthCheckResults[modelId] = response; - // Update status immediately based on response const currentTime = new Date().toLocaleString(); - // Check if there are any unhealthy endpoints (which means this specific model failed) if (response.unhealthy_count > 0 && response.unhealthy_endpoints && response.unhealthy_endpoints.length > 0) { const rawError = response.unhealthy_endpoints[0]?.error || "Health check failed"; const errorMessage = extractMeaningfulError(rawError); setModelHealthStatuses((prev) => ({ ...prev, - [modelName]: { + [modelId]: { status: "unhealthy", lastCheck: currentTime, - lastSuccess: prev[modelName]?.lastSuccess || "None", + lastSuccess: prev[modelId]?.lastSuccess || "None", loading: false, error: errorMessage, fullError: rawError, @@ -396,7 +378,7 @@ const HealthCheckComponent: React.FC = ({ } else { setModelHealthStatuses((prev) => ({ ...prev, - [modelName]: { + [modelId]: { status: "healthy", lastCheck: currentTime, lastSuccess: currentTime, @@ -406,17 +388,16 @@ const HealthCheckComponent: React.FC = ({ })); } } catch (error) { - console.error(`Health check failed for ${modelName}:`, error); - // Set error status for failed health checks + console.error(`Health check failed for model id ${modelId}:`, error); const currentTime = new Date().toLocaleString(); const rawError = error instanceof Error ? error.message : String(error); const errorMessage = extractMeaningfulError(rawError); setModelHealthStatuses((prev) => ({ ...prev, - [modelName]: { + [modelId]: { status: "unhealthy", lastCheck: currentTime, - lastSuccess: prev[modelName]?.lastSuccess || "None", + lastSuccess: prev[modelId]?.lastSuccess || "None", loading: false, error: errorMessage, fullError: rawError, @@ -425,27 +406,21 @@ const HealthCheckComponent: React.FC = ({ } }); - // Wait for all health checks to complete await Promise.allSettled(healthCheckPromises); - // Refresh health statuses from database to get the saved check data including timestamps try { if (!accessToken) return; const latestHealthChecks = await latestHealthChecksCall(accessToken); if (latestHealthChecks.latest_health_checks) { - // Update health statuses from database, which should have the most accurate saved data Object.entries(latestHealthChecks.latest_health_checks).forEach(([modelId, checkData]: [string, any]) => { - // Find the model name for this model ID - const model = modelData.data.find((m: any) => m.model_info.id === modelId); - if (model && modelsToCheck.includes(model.model_name) && checkData) { - const modelName = model.model_name; + if (modelsToCheck.includes(modelId) && checkData) { const fullError = checkData.error_message || undefined; setModelHealthStatuses((prev) => { - const currentStatus = prev[modelName]; + const currentStatus = prev[modelId]; return { ...prev, - [modelName]: { + [modelId]: { status: checkData.status || currentStatus?.status || "unknown", lastCheck: checkData.checked_at ? new Date(checkData.checked_at).toLocaleString() @@ -468,15 +443,14 @@ const HealthCheckComponent: React.FC = ({ } } catch (dbError) { console.warn("Failed to fetch updated health statuses from database (non-critical):", dbError); - // This is non-critical - we already have the health check results from the API calls } }; - const handleModelSelection = (modelName: string, checked: boolean) => { + const handleModelSelection = (modelId: string, checked: boolean) => { if (checked) { - setSelectedModelsForHealth((prev) => [...prev, modelName]); + setSelectedModelsForHealth((prev) => [...prev, modelId]); } else { - setSelectedModelsForHealth((prev) => prev.filter((name) => name !== modelName)); + setSelectedModelsForHealth((prev) => prev.filter((id) => id !== modelId)); setAllModelsSelected(false); } }; @@ -580,8 +554,9 @@ const HealthCheckComponent: React.FC = ({ teams, )} data={modelData.data.map((model: any) => { - const modelName = model.model_name; - const healthStatus = modelHealthStatuses[modelName] || { + const modelId = model.model_info?.id; + const healthStatus = modelId ? modelHealthStatuses[modelId] : null; + const status = healthStatus || { status: "none", lastCheck: "None", loading: false, @@ -591,12 +566,12 @@ const HealthCheckComponent: React.FC = ({ model_info: model.model_info, provider: model.provider, litellm_model_name: model.litellm_model_name, - health_status: healthStatus.status, - last_check: healthStatus.lastCheck, - last_success: healthStatus.lastSuccess || "None", - health_loading: healthStatus.loading, - health_error: healthStatus.error, - health_full_error: healthStatus.fullError, + health_status: status.status, + last_check: status.lastCheck, + last_success: status.lastSuccess || "None", + health_loading: status.loading, + health_error: status.error, + health_full_error: status.fullError, }; })} isLoading={false} diff --git a/ui/litellm-dashboard/src/components/model_dashboard/health_check_columns.tsx b/ui/litellm-dashboard/src/components/model_dashboard/health_check_columns.tsx index 3e8ae662ad4..396afb7ed0b 100644 --- a/ui/litellm-dashboard/src/components/model_dashboard/health_check_columns.tsx +++ b/ui/litellm-dashboard/src/components/model_dashboard/health_check_columns.tsx @@ -35,9 +35,9 @@ export const healthCheckColumns = ( modelHealthStatuses: { [key: string]: HealthStatus }, selectedModelsForHealth: string[], allModelsSelected: boolean, - handleModelSelection: (modelName: string, checked: boolean) => void, + handleModelSelection: (modelId: string, checked: boolean) => void, handleSelectAll: (checked: boolean) => void, - runIndividualHealthCheck: (modelName: string) => void, + runIndividualHealthCheck: (modelId: string) => void, getStatusBadge: (status: string) => JSX.Element, getDisplayModelName: (model: any) => string, showErrorModal?: (modelName: string, cleanedError: string, fullError: string) => void, @@ -62,14 +62,14 @@ export const healthCheckColumns = ( sortingFn: "alphanumeric", cell: ({ row }) => { const model = row.original; - const modelName = model.model_name; - const isSelected = selectedModelsForHealth.includes(modelName); + const modelId = model.model_info?.id ?? ""; + const isSelected = selectedModelsForHealth.includes(modelId); return (
handleModelSelection(modelName, e.target.checked)} + onChange={(e) => handleModelSelection(modelId, e.target.checked)} onClick={(e) => e.stopPropagation()} /> @@ -169,8 +169,9 @@ export const healthCheckColumns = ( ); } - const modelName = model.model_name; - const hasSuccessResponse = healthStatus.status === "healthy" && modelHealthStatuses[modelName]?.successResponse; + const modelId = model.model_info?.id ?? ""; + const displayName = getDisplayModelName(model) || model.model_name; + const hasSuccessResponse = healthStatus.status === "healthy" && modelHealthStatuses[modelId]?.successResponse; return (
@@ -178,7 +179,7 @@ export const healthCheckColumns = ( {hasSuccessResponse && showSuccessModal && (
+ ), + { + Option: ({ value, children }: any) => ( + + ), + } + ), +})); + +const baseProps = { + selectedStrategy: null, + availableStrategies: ["simple-shuffle", "latency-based-routing", "least-busy"], + routingStrategyDescriptions: { + "simple-shuffle": "Randomly pick a deployment", + "latency-based-routing": "Pick the lowest-latency deployment", + }, + routerFieldsMetadata: {}, + onStrategyChange: vi.fn(), +}; + +describe("RoutingStrategySelector", () => { + it("should render", () => { + render(); + expect(screen.getByTestId("ant-select")).toBeInTheDocument(); + }); + + it("should display default label when no metadata is provided", () => { + render(); + expect(screen.getByText("Routing Strategy")).toBeInTheDocument(); + }); + + it("should display ui_field_name from metadata when provided", () => { + const props = { + ...baseProps, + routerFieldsMetadata: { + routing_strategy: { + ui_field_name: "Strategy", + field_description: "How to pick a deployment", + }, + }, + }; + render(); + expect(screen.getByText("Strategy")).toBeInTheDocument(); + expect(screen.getByText("How to pick a deployment")).toBeInTheDocument(); + }); + + it("should render all available strategies as options", () => { + render(); + expect(screen.getByText("simple-shuffle")).toBeInTheDocument(); + expect(screen.getByText("latency-based-routing")).toBeInTheDocument(); + expect(screen.getByText("least-busy")).toBeInTheDocument(); + }); + + it("should display strategy descriptions alongside option labels", () => { + render(); + expect(screen.getByText("Randomly pick a deployment")).toBeInTheDocument(); + expect(screen.getByText("Pick the lowest-latency deployment")).toBeInTheDocument(); + }); + + it("should not render a description for a strategy that has none", () => { + render(); + // "least-busy" has no entry in routingStrategyDescriptions + const select = screen.getByTestId("strategy-select"); + const leastBusyOption = Array.from(select.querySelectorAll("option")).find( + (o) => o.value === "least-busy" + ); + expect(leastBusyOption).toBeInTheDocument(); + }); + + it("should call onStrategyChange with the selected strategy value", () => { + const onStrategyChange = vi.fn(); + render(); + + const select = screen.getByTestId("strategy-select") as HTMLSelectElement; + select.value = "latency-based-routing"; + select.dispatchEvent(new Event("change", { bubbles: true })); + + expect(onStrategyChange).toHaveBeenCalledWith("latency-based-routing"); + }); +}); diff --git a/ui/litellm-dashboard/src/components/router_settings/TagFilteringToggle.test.tsx b/ui/litellm-dashboard/src/components/router_settings/TagFilteringToggle.test.tsx new file mode 100644 index 00000000000..593721db023 --- /dev/null +++ b/ui/litellm-dashboard/src/components/router_settings/TagFilteringToggle.test.tsx @@ -0,0 +1,113 @@ +import { describe, it, expect, vi } from "vitest"; +import { render, screen } from "@testing-library/react"; +import userEvent from "@testing-library/user-event"; +import TagFilteringToggle from "./TagFilteringToggle"; + +// setupTests.ts mocks @tremor/react but leaves Switch as the real implementation. +// Re-mock Switch as a plain checkbox so toggle interactions are trivially testable. +vi.mock("@tremor/react", async (importOriginal) => { + const actual = await importOriginal(); + return { + ...actual, + Switch: ({ checked, onChange, className }: any) => ( + onChange(e.target.checked)} + className={className} + /> + ), + }; +}); + +const baseMetadata = { + enable_tag_filtering: { + ui_field_name: "Tag Filtering", + field_description: "Route requests based on tags", + link: null, + }, +}; + +describe("TagFilteringToggle", () => { + it("should render", () => { + render( + + ); + expect(screen.getByRole("switch")).toBeInTheDocument(); + }); + + it("should display default label when no metadata is provided", () => { + render( + + ); + expect(screen.getByText("Enable Tag Filtering")).toBeInTheDocument(); + }); + + it("should display the label from metadata when provided", () => { + render( + + ); + expect(screen.getByText("Tag Filtering")).toBeInTheDocument(); + }); + + it("should display the description from metadata", () => { + render( + + ); + expect(screen.getByText("Route requests based on tags")).toBeInTheDocument(); + }); + + it("should render a Learn more link when metadata provides one", () => { + const metadata = { + enable_tag_filtering: { + ...baseMetadata.enable_tag_filtering, + link: "https://docs.example.com/tag-filtering", + }, + }; + render( + + ); + const link = screen.getByRole("link", { name: /learn more/i }); + expect(link).toBeInTheDocument(); + expect(link).toHaveAttribute("href", "https://docs.example.com/tag-filtering"); + }); + + it("should not render a Learn more link when metadata has no link", () => { + render( + + ); + expect(screen.queryByRole("link", { name: /learn more/i })).not.toBeInTheDocument(); + }); + + it("should reflect the enabled=true state on the switch", () => { + render( + + ); + expect(screen.getByRole("switch")).toBeChecked(); + }); + + it("should call onToggle with the new value when the switch is toggled", async () => { + const onToggle = vi.fn(); + const user = userEvent.setup(); + render( + + ); + + await user.click(screen.getByRole("switch")); + + expect(onToggle).toHaveBeenCalledWith(true); + }); +}); diff --git a/ui/litellm-dashboard/src/components/router_settings/index.test.tsx b/ui/litellm-dashboard/src/components/router_settings/index.test.tsx new file mode 100644 index 00000000000..1920268207d --- /dev/null +++ b/ui/litellm-dashboard/src/components/router_settings/index.test.tsx @@ -0,0 +1,170 @@ +import { describe, it, expect, vi, beforeEach } from "vitest"; +import { renderWithProviders, screen, waitFor } from "../../../tests/test-utils"; +import userEvent from "@testing-library/user-event"; +import RouterSettings from "./index"; + +vi.mock("antd", () => ({ + Select: Object.assign( + ({ value, onChange, children }: any) => ( + + ), + { + Option: ({ value, children }: any) => ( + + ), + } + ), +})); + +vi.mock("@tremor/react", async (importOriginal) => { + const actual = await importOriginal(); + return { + ...actual, + Switch: ({ checked, onChange }: any) => ( + onChange(e.target.checked)} + /> + ), + }; +}); + +vi.mock("@/components/networking", () => ({ + getCallbacksCall: vi.fn(), + getRouterSettingsCall: vi.fn(), + setCallbacksCall: vi.fn(), +})); + +import { + getCallbacksCall, + getRouterSettingsCall, + setCallbacksCall, +} from "@/components/networking"; + +const mockCallbacksResponse = { + router_settings: { + routing_strategy: "simple-shuffle", + num_retries: 3, + timeout: 30, + }, +}; + +const mockRouterSettingsResponse = { + fields: [ + { + field_name: "routing_strategy", + ui_field_name: "Routing Strategy", + field_description: "How requests are distributed", + options: ["simple-shuffle", "latency-based-routing"], + link: null, + }, + { + field_name: "enable_tag_filtering", + ui_field_name: "Tag Filtering", + field_description: "Route by tag", + field_value: false, + link: null, + }, + ], + routing_strategy_descriptions: { + "simple-shuffle": "Randomly pick a deployment", + "latency-based-routing": "Pick the lowest-latency deployment", + }, +}; + +const defaultProps = { + accessToken: "test-token", + userRole: "Admin", + userID: "user-1", + modelData: null, +}; + +describe("RouterSettings", () => { + beforeEach(() => { + vi.clearAllMocks(); + vi.mocked(getCallbacksCall).mockResolvedValue(mockCallbacksResponse); + vi.mocked(getRouterSettingsCall).mockResolvedValue(mockRouterSettingsResponse); + vi.mocked(setCallbacksCall).mockResolvedValue({}); + }); + + it("should render nothing when accessToken is null", () => { + const { container } = renderWithProviders( + + ); + expect(container).toBeEmptyDOMElement(); + }); + + it("should render the Save Changes and Reset buttons when authenticated", () => { + renderWithProviders(); + expect(screen.getByRole("button", { name: /save changes/i })).toBeInTheDocument(); + expect(screen.getByRole("button", { name: /reset/i })).toBeInTheDocument(); + }); + + it("should fetch callbacks and router settings on mount", async () => { + renderWithProviders(); + + await waitFor(() => { + expect(getCallbacksCall).toHaveBeenCalledWith("test-token", "user-1", "Admin"); + }); + expect(getRouterSettingsCall).toHaveBeenCalledWith("test-token"); + }); + + it("should not fetch data when any required prop is missing", () => { + renderWithProviders( + + ); + expect(getCallbacksCall).not.toHaveBeenCalled(); + }); + + it("should render routing strategies loaded from the API", async () => { + renderWithProviders(); + + await waitFor(() => { + expect(screen.getByTestId("strategy-select")).toBeInTheDocument(); + }); + + const select = screen.getByTestId("strategy-select") as HTMLSelectElement; + const optionValues = Array.from(select.options).map((o) => o.value); + expect(optionValues).toContain("simple-shuffle"); + expect(optionValues).toContain("latency-based-routing"); + }); + + it("should call setCallbacksCall with updated settings on Save Changes", async () => { + const user = userEvent.setup(); + renderWithProviders(); + + await waitFor(() => { + expect(screen.getByRole("button", { name: /save changes/i })).toBeInTheDocument(); + }); + + await user.click(screen.getByRole("button", { name: /save changes/i })); + + expect(setCallbacksCall).toHaveBeenCalledWith( + "test-token", + expect.objectContaining({ router_settings: expect.any(Object) }) + ); + }); + + it("should show a success notification after saving", async () => { + const NotificationsManager = await import("@/components/molecules/notifications_manager"); + const user = userEvent.setup(); + renderWithProviders(); + + await waitFor(() => + expect(screen.getByRole("button", { name: /save changes/i })).toBeInTheDocument() + ); + await user.click(screen.getByRole("button", { name: /save changes/i })); + + expect(NotificationsManager.default.success).toHaveBeenCalledWith( + "router settings updated successfully" + ); + }); +}); From 784af16cb4064fcd27f407a4b8e268045652ba62 Mon Sep 17 00:00:00 2001 From: yuneng-jiang Date: Tue, 24 Feb 2026 15:48:31 -0800 Subject: [PATCH 13/41] address greptile review feedback (greploop iteration 1) - Replace document.querySelector/querySelectorAll with screen.getByRole - Replace raw dispatchEvent with userEvent.selectOptions --- .../LatencyBasedConfiguration.test.tsx | 3 +-- .../ReliabilityRetriesSection.test.tsx | 3 +-- .../router_settings/RouterSettingsForm.test.tsx | 7 +++---- .../RoutingStrategySelector.test.tsx | 16 ++++++---------- 4 files changed, 11 insertions(+), 18 deletions(-) diff --git a/ui/litellm-dashboard/src/components/router_settings/LatencyBasedConfiguration.test.tsx b/ui/litellm-dashboard/src/components/router_settings/LatencyBasedConfiguration.test.tsx index 0176be5ab40..39d9f8c881c 100644 --- a/ui/litellm-dashboard/src/components/router_settings/LatencyBasedConfiguration.test.tsx +++ b/ui/litellm-dashboard/src/components/router_settings/LatencyBasedConfiguration.test.tsx @@ -47,8 +47,7 @@ describe("LatencyBasedConfiguration", () => { const args = { ttl: { nested: true } }; render(); // HTML input type=text strips newlines, so check that the key/value appears - const input = document.querySelector('input[name="ttl"]') as HTMLInputElement; - expect(input).not.toBeNull(); + const input = screen.getByRole("textbox", { name: /ttl/i }) as HTMLInputElement; expect(input.value).toContain('"nested"'); expect(input.value).toContain('true'); }); diff --git a/ui/litellm-dashboard/src/components/router_settings/ReliabilityRetriesSection.test.tsx b/ui/litellm-dashboard/src/components/router_settings/ReliabilityRetriesSection.test.tsx index 0892d39fc29..101b09af0dc 100644 --- a/ui/litellm-dashboard/src/components/router_settings/ReliabilityRetriesSection.test.tsx +++ b/ui/litellm-dashboard/src/components/router_settings/ReliabilityRetriesSection.test.tsx @@ -67,8 +67,7 @@ describe("ReliabilityRetriesSection", () => { const settings = { retry_policy: { "rate-limited": 2 } }; render(); // HTML input type=text strips newlines, so check that the key/value appears - const input = document.querySelector('input[name="retry_policy"]') as HTMLInputElement; - expect(input).not.toBeNull(); + const input = screen.getByRole("textbox", { name: /retry_policy/i }) as HTMLInputElement; expect(input.value).toContain('"rate-limited"'); expect(input.value).toContain('2'); }); diff --git a/ui/litellm-dashboard/src/components/router_settings/RouterSettingsForm.test.tsx b/ui/litellm-dashboard/src/components/router_settings/RouterSettingsForm.test.tsx index 767820cd485..01f318b909a 100644 --- a/ui/litellm-dashboard/src/components/router_settings/RouterSettingsForm.test.tsx +++ b/ui/litellm-dashboard/src/components/router_settings/RouterSettingsForm.test.tsx @@ -97,8 +97,9 @@ describe("RouterSettingsForm", () => { expect(screen.getByText("Latency-Based Configuration")).toBeInTheDocument(); }); - it("should call onChange with the updated strategy when the selector changes", () => { + it("should call onChange with the updated strategy when the selector changes", async () => { const onChange = vi.fn(); + const user = userEvent.setup(); const props = { ...baseProps, onChange, @@ -106,9 +107,7 @@ describe("RouterSettingsForm", () => { }; render(); - const select = screen.getByTestId("strategy-select") as HTMLSelectElement; - select.value = "latency-based-routing"; - select.dispatchEvent(new Event("change", { bubbles: true })); + await user.selectOptions(screen.getByTestId("strategy-select"), "latency-based-routing"); expect(onChange).toHaveBeenCalledWith( expect.objectContaining({ selectedStrategy: "latency-based-routing" }) diff --git a/ui/litellm-dashboard/src/components/router_settings/RoutingStrategySelector.test.tsx b/ui/litellm-dashboard/src/components/router_settings/RoutingStrategySelector.test.tsx index 85b1dc21acf..01f839681f1 100644 --- a/ui/litellm-dashboard/src/components/router_settings/RoutingStrategySelector.test.tsx +++ b/ui/litellm-dashboard/src/components/router_settings/RoutingStrategySelector.test.tsx @@ -1,5 +1,6 @@ import { describe, it, expect, vi } from "vitest"; import { render, screen } from "@testing-library/react"; +import userEvent from "@testing-library/user-event"; import RoutingStrategySelector from "./RoutingStrategySelector"; // Ant Design's Select is complex to drive in JSDOM; swap it for a plain @@ -77,21 +78,16 @@ describe("RoutingStrategySelector", () => { it("should not render a description for a strategy that has none", () => { render(); - // "least-busy" has no entry in routingStrategyDescriptions - const select = screen.getByTestId("strategy-select"); - const leastBusyOption = Array.from(select.querySelectorAll("option")).find( - (o) => o.value === "least-busy" - ); - expect(leastBusyOption).toBeInTheDocument(); + // "least-busy" has no entry in routingStrategyDescriptions — it still renders without crashing + expect(screen.getByText("least-busy")).toBeInTheDocument(); }); - it("should call onStrategyChange with the selected strategy value", () => { + it("should call onStrategyChange with the selected strategy value", async () => { const onStrategyChange = vi.fn(); + const user = userEvent.setup(); render(); - const select = screen.getByTestId("strategy-select") as HTMLSelectElement; - select.value = "latency-based-routing"; - select.dispatchEvent(new Event("change", { bubbles: true })); + await user.selectOptions(screen.getByTestId("strategy-select"), "latency-based-routing"); expect(onStrategyChange).toHaveBeenCalledWith("latency-based-routing"); }); From 26e5482abb352b16dd57444705a6e44ef206d3d8 Mon Sep 17 00:00:00 2001 From: yuneng-jiang Date: Tue, 24 Feb 2026 15:52:15 -0800 Subject: [PATCH 14/41] address greptile review feedback (greploop iteration 2) - Wait for strategy select (API data loaded) before clicking Save - Assert specific payload content in setCallbacksCall - Move NotificationsManager import to top of file --- .../components/router_settings/index.test.tsx | 20 ++++++++++++------- 1 file changed, 13 insertions(+), 7 deletions(-) diff --git a/ui/litellm-dashboard/src/components/router_settings/index.test.tsx b/ui/litellm-dashboard/src/components/router_settings/index.test.tsx index 1920268207d..8edb0ac6e07 100644 --- a/ui/litellm-dashboard/src/components/router_settings/index.test.tsx +++ b/ui/litellm-dashboard/src/components/router_settings/index.test.tsx @@ -48,6 +48,7 @@ import { getRouterSettingsCall, setCallbacksCall, } from "@/components/networking"; +import NotificationsManager from "@/components/molecules/notifications_manager"; const mockCallbacksResponse = { router_settings: { @@ -141,29 +142,34 @@ describe("RouterSettings", () => { const user = userEvent.setup(); renderWithProviders(); + // Wait for the strategy select to appear — it only renders after getRouterSettingsCall resolves await waitFor(() => { - expect(screen.getByRole("button", { name: /save changes/i })).toBeInTheDocument(); + expect(screen.getByTestId("strategy-select")).toBeInTheDocument(); }); await user.click(screen.getByRole("button", { name: /save changes/i })); expect(setCallbacksCall).toHaveBeenCalledWith( "test-token", - expect.objectContaining({ router_settings: expect.any(Object) }) + expect.objectContaining({ + router_settings: expect.objectContaining({ + routing_strategy: "simple-shuffle", + }), + }) ); }); it("should show a success notification after saving", async () => { - const NotificationsManager = await import("@/components/molecules/notifications_manager"); const user = userEvent.setup(); renderWithProviders(); - await waitFor(() => - expect(screen.getByRole("button", { name: /save changes/i })).toBeInTheDocument() - ); + // Wait for data to load before interacting + await waitFor(() => { + expect(screen.getByTestId("strategy-select")).toBeInTheDocument(); + }); await user.click(screen.getByRole("button", { name: /save changes/i })); - expect(NotificationsManager.default.success).toHaveBeenCalledWith( + expect(NotificationsManager.success).toHaveBeenCalledWith( "router settings updated successfully" ); }); From 6ee50ff73e47da30c31526c22c8b58979248348c Mon Sep 17 00:00:00 2001 From: Ishaan Jaff Date: Tue, 24 Feb 2026 16:27:06 -0800 Subject: [PATCH 15/41] feat(proxy): tool policies - auto-discover tools + policy enforcement guardrail (#22041) * feat(proxy): tool policies - auto-discover tools, manage policies, guardrail enforcement - New LiteLLM_ToolTable in schema.prisma to store discovered tools - Auto-discovery: tools seen in LLM responses get upserted via ToolDiscoveryQueue (hooks into DBSpendUpdateWriter, same pipeline as spend tracking) - Management endpoints: GET /v1/tool/list, GET /v1/tool/{name}, POST /v1/tool/policy - ToolPolicyGuardrail: blocks tool_calls in responses based on policy setting - UI: Tool Policies page under Guardrails section with policy selector, filters by policy/team/key, live tail, sortable table - Unit tests for queue, writer, endpoints, guardrail * feat(tool-policies): track call_count + discover tools from request body and /messages API - Add call_count column to LiteLLM_ToolTable; incremented on every flush - Extract tools from request body too (not just response tool_calls): - OpenAI /chat/completions: tools[].function.name - Anthropic /messages pass-through: request_body.tools[].name - Show call_count column in UI table (sortable) - UI: drop dual_llm option, keep only trusted/blocked * fix: address greptile review feedback - Remove redundant @@index([tool_name]) from schema.prisma (tool_name has @unique which already creates an index) - Replace gen_random_uuid()::text with str(uuid.uuid4()) for portability - Rewrite test_tool_registry_writer.py to mock execute_raw/query_raw (actual implementation) instead of Prisma model methods - Fix test patches in test_tool_management_endpoints.py to target source modules since imports are inside function bodies - Add "Tool Policies" page title to ToolPolicies.tsx * fix: address greptile review round 2 - Replace NOW() with Python datetime parameter in tool_registry_writer (SQLite portability) - Fix cache key collision in tool_policy_guardrail: use null-byte separator instead of colon - Remove type==function filter from request-side tool extraction to match response-side behavior - Clear seen_tool_names on flush so call_count increments per batch cycle not per pod lifetime * fix: address greptile review round 3 - Fix test_seen_names_persist_across_flushes to match actual per-flush-cycle behavior - Update module docstring in tool_discovery_queue.py to accurately describe flush behavior - Add created_at/updated_at to raw SQL INSERT in batch_upsert_tools and update_tool_policy * fix: cache tool policies per tool name not per combination Previously the cache key was built from the full set of tool names in a request, so each unique combination of tools got its own cold cache entry and triggered a separate DB query. With N distinct tools across requests this was effectively a DB hit on every request. Now each tool name is cached individually. Cache hits are checked per tool, only missing tools are fetched from DB in a single batch query, and each result is cached separately. Once a tool's policy is warm, any subsequent request using that tool benefits from the cache regardless of what other tools are in the request. * Update ui/litellm-dashboard/src/components/ToolPolicies.tsx Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> --------- Co-authored-by: greptile-apps[bot] <165735046+greptile-apps[bot]@users.noreply.github.com> --- litellm/constants.py | 1 + litellm/proxy/_types.py | 10 + litellm/proxy/db/db_spend_update_writer.py | 151 ++++++- .../tool_discovery_queue.py | 54 +++ litellm/proxy/db/tool_registry_writer.py | 179 ++++++++ .../guardrail_hooks/tool_policy/__init__.py | 16 + .../tool_policy/tool_policy_guardrail.py | 163 +++++++ .../tool_management_endpoints.py | 149 +++++++ litellm/proxy/proxy_server.py | 4 + litellm/proxy/schema.prisma | 20 + litellm/types/tool_management.py | 42 ++ .../test_tool_discovery_queue.py | 75 ++++ .../proxy/db/test_tool_registry_writer.py | 197 +++++++++ .../test_tool_policy_guardrail.py | 181 ++++++++ .../test_tool_management_endpoints.py | 149 +++++++ ui/litellm-dashboard/src/app/page.tsx | 3 + .../src/components/ToolPolicies.tsx | 415 ++++++++++++++++++ .../src/components/leftnav.tsx | 6 + .../src/components/networking.tsx | 54 +++ 19 files changed, 1863 insertions(+), 6 deletions(-) create mode 100644 litellm/proxy/db/db_transaction_queue/tool_discovery_queue.py create mode 100644 litellm/proxy/db/tool_registry_writer.py create mode 100644 litellm/proxy/guardrails/guardrail_hooks/tool_policy/__init__.py create mode 100644 litellm/proxy/guardrails/guardrail_hooks/tool_policy/tool_policy_guardrail.py create mode 100644 litellm/proxy/management_endpoints/tool_management_endpoints.py create mode 100644 litellm/types/tool_management.py create mode 100644 tests/test_litellm/proxy/db/db_transaction_queue/test_tool_discovery_queue.py create mode 100644 tests/test_litellm/proxy/db/test_tool_registry_writer.py create mode 100644 tests/test_litellm/proxy/guardrails/guardrail_hooks/test_tool_policy_guardrail.py create mode 100644 tests/test_litellm/proxy/management_endpoints/test_tool_management_endpoints.py create mode 100644 ui/litellm-dashboard/src/components/ToolPolicies.tsx diff --git a/litellm/constants.py b/litellm/constants.py index ee79f2fa56f..b1a0021bcc6 100644 --- a/litellm/constants.py +++ b/litellm/constants.py @@ -244,6 +244,7 @@ REDIS_DAILY_TAG_SPEND_UPDATE_BUFFER_KEY = "litellm_daily_tag_spend_update_buffer MAX_REDIS_BUFFER_DEQUEUE_COUNT = int(os.getenv("MAX_REDIS_BUFFER_DEQUEUE_COUNT", 100)) # Bounds asyncio.Queue() instances (log queues, spend update queues, etc.) to prevent unbounded memory growth LITELLM_ASYNCIO_QUEUE_MAXSIZE = int(os.getenv("LITELLM_ASYNCIO_QUEUE_MAXSIZE", 1000)) +TOOL_POLICY_CACHE_TTL_SECONDS = int(os.getenv("TOOL_POLICY_CACHE_TTL_SECONDS", 60)) # Aggregation threshold: default to 80% of the asyncio queue maxsize so the check can always trigger. # Must be < LITELLM_ASYNCIO_QUEUE_MAXSIZE; if set higher the aggregation logic will never fire. MAX_SIZE_IN_MEMORY_QUEUE = int( diff --git a/litellm/proxy/_types.py b/litellm/proxy/_types.py index f354e28acd7..75b9f91acd9 100644 --- a/litellm/proxy/_types.py +++ b/litellm/proxy/_types.py @@ -183,6 +183,7 @@ class LitellmTableNames(str, enum.Enum): KEY_TABLE_NAME = "LiteLLM_VerificationToken" PROXY_MODEL_TABLE_NAME = "LiteLLM_ProxyModelTable" MANAGED_FILE_TABLE_NAME = "LiteLLM_ManagedFileTable" + TOOL_TABLE_NAME = "LiteLLM_ToolTable" class Litellm_EntityType(enum.Enum): @@ -4123,6 +4124,15 @@ class SpendUpdateQueueItem(TypedDict, total=False): response_cost: Optional[float] +class ToolDiscoveryQueueItem(TypedDict, total=False): + tool_name: str + origin: Optional[str] # MCP server name or "user_defined" + created_by: Optional[str] + key_hash: Optional[str] # hash of virtual key that triggered discovery + team_id: Optional[str] # team that triggered discovery + key_alias: Optional[str] # human-readable key alias + + class LiteLLM_ManagedFileTable(LiteLLMPydanticObjectBase): unified_file_id: str file_object: Optional[OpenAIFileObject] = None diff --git a/litellm/proxy/db/db_spend_update_writer.py b/litellm/proxy/db/db_spend_update_writer.py index d7d6b2b1eb0..edf0cf0d397 100644 --- a/litellm/proxy/db/db_spend_update_writer.py +++ b/litellm/proxy/db/db_spend_update_writer.py @@ -13,7 +13,17 @@ import random import time import traceback from datetime import datetime, timedelta, timezone -from typing import TYPE_CHECKING, Any, Dict, List, Literal, Optional, Union, cast, overload +from typing import ( + TYPE_CHECKING, + Any, + Dict, + List, + Literal, + Optional, + Union, + cast, + overload, +) import litellm from litellm._logging import verbose_proxy_logger @@ -23,18 +33,19 @@ from litellm.litellm_core_utils.safe_json_loads import safe_json_loads from litellm.proxy._types import ( DB_CONNECTION_ERROR_TYPES, BaseDailySpendTransaction, - DailyTagSpendTransaction, - DailyOrganizationSpendTransaction, - DailyTeamSpendTransaction, - DailyEndUserSpendTransaction, - DailyUserSpendTransaction, DailyAgentSpendTransaction, + DailyEndUserSpendTransaction, + DailyOrganizationSpendTransaction, + DailyTagSpendTransaction, + DailyTeamSpendTransaction, + DailyUserSpendTransaction, DBSpendUpdateTransactions, Litellm_EntityType, LiteLLM_UserTable, SpendLogsMetadata, SpendLogsPayload, SpendUpdateQueueItem, + ToolDiscoveryQueueItem, ) from litellm.proxy.db.db_transaction_queue.daily_spend_update_queue import ( DailySpendUpdateQueue, @@ -42,6 +53,9 @@ from litellm.proxy.db.db_transaction_queue.daily_spend_update_queue import ( from litellm.proxy.db.db_transaction_queue.pod_lock_manager import PodLockManager from litellm.proxy.db.db_transaction_queue.redis_update_buffer import RedisUpdateBuffer from litellm.proxy.db.db_transaction_queue.spend_update_queue import SpendUpdateQueue +from litellm.proxy.db.db_transaction_queue.tool_discovery_queue import ( + ToolDiscoveryQueue, +) from litellm.proxy.route_llm_request import ROUTE_ENDPOINT_MAPPING if TYPE_CHECKING: @@ -67,6 +81,7 @@ class DBSpendUpdateWriter: self.redis_update_buffer = RedisUpdateBuffer(redis_cache=self.redis_cache) self.pod_lock_manager = PodLockManager() self.spend_update_queue = SpendUpdateQueue() + self.tool_discovery_queue = ToolDiscoveryQueue() self.daily_spend_update_queue = DailySpendUpdateQueue() self.daily_team_spend_update_queue = DailySpendUpdateQueue() self.daily_end_user_spend_update_queue = DailySpendUpdateQueue() @@ -222,6 +237,13 @@ class DBSpendUpdateWriter: ) ) + self._enqueue_tool_registry_upsert( + kwargs=kwargs, + completion_response=completion_response, + hashed_token=hashed_token, + team_id=team_id, + ) + verbose_proxy_logger.debug("Runs spend update on all tables") except Exception: verbose_proxy_logger.error( @@ -237,6 +259,104 @@ class DBSpendUpdateWriter: traceback.format_exc(), ) + def _enqueue_tool_registry_upsert( + self, + kwargs: Optional[dict], + completion_response: Optional[Any], + hashed_token: Optional[str] = None, + team_id: Optional[str] = None, + ) -> None: + """ + Extract tool names from the LLM request and response and enqueue them + for upsert into LiteLLM_ToolTable via ToolDiscoveryQueue. + + Handles four sources: + - MCP tools: standard_logging_object.mcp_tool_call_metadata.namespaced_tool_name + - Response tool_calls (OpenAI / Anthropic pass-through converted to OpenAI format): + completion_response.choices[].message.tool_calls[].function.name + - Request tools array (OpenAI format): kwargs["tools"][].function.name + - Request tools array (Anthropic /messages format): kwargs["passthrough_logging_payload"] + ["request_body"]["tools"][].name + """ + try: + if kwargs is None: + return + + # Extract key_alias from kwargs metadata if available + key_alias: Optional[str] = None + _litellm_params = kwargs.get("litellm_params") or {} + _metadata = _litellm_params.get("metadata") or {} + key_alias = _metadata.get("user_api_key_alias") or None + + def _enqueue(tool_name: str, origin: str = "user_defined") -> None: + self.tool_discovery_queue.add_update( + ToolDiscoveryQueueItem( + tool_name=tool_name, + origin=origin, + key_hash=hashed_token, + team_id=team_id, + key_alias=key_alias, + ) + ) + + # --- MCP tool calls --- + sl_object = kwargs.get("standard_logging_object") + if sl_object is not None: + mcp_metadata = ( + sl_object.get("metadata", {}) or {} + ).get("mcp_tool_call_metadata") + if mcp_metadata and isinstance(mcp_metadata, dict): + tool_name = mcp_metadata.get("namespaced_tool_name") or mcp_metadata.get("name") + mcp_server_name = mcp_metadata.get("mcp_server_name") + if tool_name: + _enqueue(tool_name, origin=mcp_server_name or "user_defined") + + # --- Tools from request body (OpenAI format: tools[].function.name) --- + request_tools = kwargs.get("tools") or [] + for tool_def in request_tools: + if not isinstance(tool_def, dict): + continue + fn = tool_def.get("function") or {} + name = fn.get("name") if isinstance(fn, dict) else None + if name: + _enqueue(name) + + # --- Tools from Anthropic /messages pass-through request body + # (Anthropic format: tools[].name, no "function" wrapper) --- + passthrough_payload = kwargs.get("passthrough_logging_payload") or {} + request_body = ( + passthrough_payload.get("request_body") + if isinstance(passthrough_payload, dict) + else None + ) or {} + for tool_def in request_body.get("tools") or []: + if not isinstance(tool_def, dict): + continue + name = tool_def.get("name") + if name: + _enqueue(name) + + # --- Response tool_calls (OpenAI format; Anthropic pass-through converts tool_use here) --- + if completion_response is not None and hasattr(completion_response, "choices"): + for choice in completion_response.choices or []: + message = getattr(choice, "message", None) + if message is None: + continue + tool_calls = getattr(message, "tool_calls", None) + if not tool_calls: + continue + for tc in tool_calls: + fn = getattr(tc, "function", None) + if fn is None: + continue + tool_name = getattr(fn, "name", None) + if tool_name: + _enqueue(tool_name) + except Exception as e: + verbose_proxy_logger.debug( + "_enqueue_tool_registry_upsert error (non-blocking): %s", e + ) + async def _update_key_db( self, response_cost: Optional[float], @@ -752,6 +872,25 @@ class DBSpendUpdateWriter: daily_spend_transactions=daily_agent_spend_update_transactions, ) + ################## Tool Registry Upserts ################## + await self._flush_tool_discovery_queue(prisma_client=prisma_client) + + async def _flush_tool_discovery_queue( + self, + prisma_client: PrismaClient, + ) -> None: + """Flush ToolDiscoveryQueue and batch-upsert new tools into LiteLLM_ToolTable.""" + from litellm.proxy.db.tool_registry_writer import batch_upsert_tools + + try: + items = self.tool_discovery_queue.flush() + if items: + await batch_upsert_tools(prisma_client=prisma_client, items=items) + except Exception as e: + verbose_proxy_logger.debug( + "_flush_tool_discovery_queue error (non-blocking): %s", e + ) + async def _commit_spend_updates_to_db( # noqa: PLR0915 self, prisma_client: PrismaClient, diff --git a/litellm/proxy/db/db_transaction_queue/tool_discovery_queue.py b/litellm/proxy/db/db_transaction_queue/tool_discovery_queue.py new file mode 100644 index 00000000000..16a3ada40f2 --- /dev/null +++ b/litellm/proxy/db/db_transaction_queue/tool_discovery_queue.py @@ -0,0 +1,54 @@ +""" +In-memory buffer for tool registry upserts. + +Unlike SpendUpdateQueue (which aggregates increments), ToolDiscoveryQueue +uses set-deduplication: each unique tool_name is only queued once per flush +cycle (~30s). The seen-set is cleared on every flush so that call_count +increments in subsequent cycles rather than stopping after the first flush. +""" + +from typing import List, Set + +from litellm._logging import verbose_proxy_logger +from litellm.proxy._types import ToolDiscoveryQueueItem + + +class ToolDiscoveryQueue: + """ + In-memory buffer for tool registry upserts. + + Deduplicates by tool_name within each flush cycle: a tool is only queued + once per ~30s batch, so call_count increments once per flush cycle the + tool appears in (not once per invocation, but not once per pod lifetime + either). The seen-set is cleared on flush so subsequent batches can + re-count the same tool. + """ + + def __init__(self) -> None: + self._seen_tool_names: Set[str] = set() + self._pending: List[ToolDiscoveryQueueItem] = [] + + def add_update(self, item: ToolDiscoveryQueueItem) -> None: + """Enqueue a tool discovery item if tool_name has not been seen before.""" + tool_name = item.get("tool_name", "") + if not tool_name: + return + if tool_name in self._seen_tool_names: + verbose_proxy_logger.debug( + "ToolDiscoveryQueue: skipping already-seen tool %s", tool_name + ) + return + self._seen_tool_names.add(tool_name) + self._pending.append(item) + verbose_proxy_logger.debug( + "ToolDiscoveryQueue: queued new tool %s (origin=%s)", + tool_name, + item.get("origin"), + ) + + def flush(self) -> List[ToolDiscoveryQueueItem]: + """Return and clear all pending items. Resets seen-set so the next + flush cycle can re-count the same tools.""" + items, self._pending = self._pending, [] + self._seen_tool_names.clear() + return items diff --git a/litellm/proxy/db/tool_registry_writer.py b/litellm/proxy/db/tool_registry_writer.py new file mode 100644 index 00000000000..4e0a8095a08 --- /dev/null +++ b/litellm/proxy/db/tool_registry_writer.py @@ -0,0 +1,179 @@ +""" +DB helpers for LiteLLM_ToolTable — the global tool registry. + +Tools are auto-discovered from LLM responses and upserted here. +Admins use the management endpoints to read and update call_policy. + +NOTE: Uses raw SQL (query_raw / execute_raw) instead of Prisma model methods +because the generated Prisma Python client may not have LiteLLM_ToolTable +when running against an older generated schema. +""" + +import uuid +from datetime import datetime, timezone +from typing import TYPE_CHECKING, Dict, List, Optional + +from litellm._logging import verbose_proxy_logger +from litellm.proxy._types import ToolDiscoveryQueueItem +from litellm.types.tool_management import LiteLLM_ToolTableRow, ToolCallPolicy + +if TYPE_CHECKING: + from litellm.proxy.utils import PrismaClient + + +def _row_to_model(row: dict) -> LiteLLM_ToolTableRow: + return LiteLLM_ToolTableRow( + tool_id=row.get("tool_id", ""), + tool_name=row.get("tool_name", ""), + origin=row.get("origin"), + call_policy=row.get("call_policy", "untrusted"), + call_count=int(row.get("call_count") or 0), + assignments=row.get("assignments"), + key_hash=row.get("key_hash"), + team_id=row.get("team_id"), + key_alias=row.get("key_alias"), + created_at=row.get("created_at"), + updated_at=row.get("updated_at"), + created_by=row.get("created_by"), + updated_by=row.get("updated_by"), + ) + + +async def batch_upsert_tools( + prisma_client: "PrismaClient", + items: List[ToolDiscoveryQueueItem], +) -> None: + """ + Batch-upsert tool registry rows via raw SQL. + + On first insert: sets call_policy = "untrusted" (schema default), call_count = 1. + On conflict: increments call_count; preserves existing call_policy. + """ + if not items: + return + try: + data = [item for item in items if item.get("tool_name")] + if not data: + return + for item in data: + tool_name = item.get("tool_name", "") + origin = item.get("origin") or "user_defined" + created_by = item.get("created_by") or "system" + key_hash = item.get("key_hash") + team_id = item.get("team_id") + key_alias = item.get("key_alias") + now = datetime.now(timezone.utc).isoformat() + await prisma_client.db.execute_raw( + 'INSERT INTO "LiteLLM_ToolTable" ' + "(tool_id, tool_name, origin, call_policy, call_count, created_by, updated_by, key_hash, team_id, key_alias, created_at, updated_at) " + "VALUES ($7, $1, $2, 'untrusted', 1, $3, $3, $4, $5, $6, $8, $8) " + "ON CONFLICT (tool_name) DO UPDATE SET " + "call_count = \"LiteLLM_ToolTable\".call_count + 1, " + "updated_at = $8", + tool_name, + origin, + created_by, + key_hash, + team_id, + key_alias, + str(uuid.uuid4()), + now, + ) + verbose_proxy_logger.debug( + "tool_registry_writer: upserted %d tool(s)", len(data) + ) + except Exception as e: + verbose_proxy_logger.error("tool_registry_writer batch_upsert_tools error: %s", e) + + +async def list_tools( + prisma_client: "PrismaClient", + call_policy: Optional[ToolCallPolicy] = None, +) -> List[LiteLLM_ToolTableRow]: + """Return all tools, optionally filtered by call_policy.""" + try: + if call_policy is not None: + rows = await prisma_client.db.query_raw( + 'SELECT tool_id, tool_name, origin, call_policy, call_count, assignments, ' + 'key_hash, team_id, key_alias, created_at, updated_at, created_by, updated_by ' + 'FROM "LiteLLM_ToolTable" WHERE call_policy = $1 ORDER BY created_at DESC', + call_policy, + ) + else: + rows = await prisma_client.db.query_raw( + 'SELECT tool_id, tool_name, origin, call_policy, call_count, assignments, ' + 'key_hash, team_id, key_alias, created_at, updated_at, created_by, updated_by ' + 'FROM "LiteLLM_ToolTable" ORDER BY created_at DESC', + ) + return [_row_to_model(row) for row in rows] + except Exception as e: + verbose_proxy_logger.error("tool_registry_writer list_tools error: %s", e) + return [] + + +async def get_tool( + prisma_client: "PrismaClient", + tool_name: str, +) -> Optional[LiteLLM_ToolTableRow]: + """Return a single tool row by tool_name.""" + try: + rows = await prisma_client.db.query_raw( + 'SELECT tool_id, tool_name, origin, call_policy, call_count, assignments, ' + 'key_hash, team_id, key_alias, created_at, updated_at, created_by, updated_by ' + 'FROM "LiteLLM_ToolTable" WHERE tool_name = $1', + tool_name, + ) + if not rows: + return None + return _row_to_model(rows[0]) + except Exception as e: + verbose_proxy_logger.error("tool_registry_writer get_tool error: %s", e) + return None + + +async def update_tool_policy( + prisma_client: "PrismaClient", + tool_name: str, + call_policy: ToolCallPolicy, + updated_by: Optional[str], +) -> Optional[LiteLLM_ToolTableRow]: + """Update the call_policy for a tool. Upserts the row if it does not exist yet.""" + try: + _updated_by = updated_by or "system" + now = datetime.now(timezone.utc).isoformat() + await prisma_client.db.execute_raw( + 'INSERT INTO "LiteLLM_ToolTable" (tool_id, tool_name, call_policy, created_by, updated_by, created_at, updated_at) ' + "VALUES ($4, $1, $2, $3, $3, $5, $5) " + "ON CONFLICT (tool_name) DO UPDATE SET call_policy = $2, updated_by = $3, updated_at = $5", + tool_name, + call_policy, + _updated_by, + str(uuid.uuid4()), + now, + ) + return await get_tool(prisma_client, tool_name) + except Exception as e: + verbose_proxy_logger.error("tool_registry_writer update_tool_policy error: %s", e) + return None + + +async def get_tools_by_names( + prisma_client: "PrismaClient", + tool_names: List[str], +) -> Dict[str, str]: + """ + Return a {tool_name: call_policy} map for the given tool names. + Used by the policy enforcement guardrail — single batch query, never N+1. + """ + if not tool_names: + return {} + try: + placeholders = ", ".join(f"${i+1}" for i in range(len(tool_names))) + rows = await prisma_client.db.query_raw( + f'SELECT tool_name, call_policy FROM "LiteLLM_ToolTable" WHERE tool_name IN ({placeholders})', + *tool_names, + ) + return {row["tool_name"]: row["call_policy"] for row in rows} + except Exception as e: + verbose_proxy_logger.error("tool_registry_writer get_tools_by_names error: %s", e) + return {} diff --git a/litellm/proxy/guardrails/guardrail_hooks/tool_policy/__init__.py b/litellm/proxy/guardrails/guardrail_hooks/tool_policy/__init__.py new file mode 100644 index 00000000000..5a43006e23c --- /dev/null +++ b/litellm/proxy/guardrails/guardrail_hooks/tool_policy/__init__.py @@ -0,0 +1,16 @@ +import litellm +from litellm.types.guardrails import Guardrail, LitellmParams + + +def initialize_guardrail(litellm_params: LitellmParams, guardrail: Guardrail): + from litellm.proxy.guardrails.guardrail_hooks.tool_policy.tool_policy_guardrail import ( + ToolPolicyGuardrail, + ) + + _callback = ToolPolicyGuardrail( + guardrail_name=guardrail.get("guardrail_name", "tool_policy"), + event_hook=litellm_params.mode, + default_on=litellm_params.default_on, + ) + litellm.logging_callback_manager.add_litellm_callback(_callback) + return _callback diff --git a/litellm/proxy/guardrails/guardrail_hooks/tool_policy/tool_policy_guardrail.py b/litellm/proxy/guardrails/guardrail_hooks/tool_policy/tool_policy_guardrail.py new file mode 100644 index 00000000000..87558566c42 --- /dev/null +++ b/litellm/proxy/guardrails/guardrail_hooks/tool_policy/tool_policy_guardrail.py @@ -0,0 +1,163 @@ +""" +Tool Policy Guardrail + +Reads call_policy from LiteLLM_ToolTable and enforces it on LLM requests/responses. + +Policy values: + "trusted" - allow through (no action) + "untrusted" - allow through (no action; default for newly discovered tools) + "blocked" - raise HTTPException, preventing the tool call + "dual_llm" - (Phase 3) send to second LLM for verification; currently treated as allowed + +Configuration in proxy config YAML: + guardrails: + - guardrail_name: "tool_policy" + litellm_params: + guardrail: tool_policy + mode: post_call + +or both pre and post call: + - guardrail_name: "tool_policy" + litellm_params: + guardrail: tool_policy + mode: during_call # runs before LLM and on response +""" + +from typing import TYPE_CHECKING, Any, Dict, List, Literal, Optional + +from fastapi import HTTPException + +from litellm._logging import verbose_proxy_logger +from litellm.caching.dual_cache import DualCache +from litellm.constants import TOOL_POLICY_CACHE_TTL_SECONDS +from litellm.integrations.custom_guardrail import ( + CustomGuardrail, + log_guardrail_information, +) +from litellm.types.guardrails import GuardrailEventHooks +from litellm.types.utils import GenericGuardrailAPIInputs + +if TYPE_CHECKING: + from litellm.litellm_core_utils.litellm_logging import Logging as LiteLLMLoggingObj + +GUARDRAIL_NAME = "tool_policy" + + +class ToolPolicyGuardrail(CustomGuardrail): + """ + Guardrail that enforces per-tool call policies stored in LiteLLM_ToolTable. + + Tools with call_policy="blocked" are rejected before/after the LLM call. + Tools with call_policy="trusted" or "untrusted" pass through unchanged. + """ + + def __init__(self, **kwargs: Any) -> None: + if "supported_event_hooks" not in kwargs: + kwargs["supported_event_hooks"] = [ + GuardrailEventHooks.pre_call, + GuardrailEventHooks.post_call, + GuardrailEventHooks.during_call, + ] + super().__init__(**kwargs) + self._policy_cache: DualCache = DualCache() + + @log_guardrail_information + async def apply_guardrail( + self, + inputs: GenericGuardrailAPIInputs, + request_data: dict, + input_type: Literal["request", "response"], + logging_obj: Optional["LiteLLMLoggingObj"] = None, + ) -> GenericGuardrailAPIInputs: + """ + Enforce tool policies on both request tools and response tool_calls. + + - input_type="request": check inputs["tools"] (tool definitions in the LLM request) + - input_type="response": check inputs["tool_calls"] (tool_calls in the LLM response) + + Raises HTTPException (400) if any tool is "blocked". + """ + if input_type == "request": + tools = inputs.get("tools") or [] + tool_names = [ + t["function"]["name"] + for t in tools + if isinstance(t, dict) + and isinstance(t.get("function"), dict) + and t["function"].get("name") + ] + else: # response + tool_calls = inputs.get("tool_calls") or [] + tool_names = [] + for tc in tool_calls: + fn = None + if isinstance(tc, dict): + fn = (tc.get("function") or {}).get("name") + elif hasattr(tc, "function"): + fn = getattr(tc.function, "name", None) + if fn: + tool_names.append(fn) + + if not tool_names: + return inputs + + policy_map = await self._get_policies_cached(tool_names) + + blocked = [name for name in tool_names if policy_map.get(name) == "blocked"] + if blocked: + verbose_proxy_logger.warning( + "ToolPolicyGuardrail: blocking tool(s) %s (policy=blocked)", blocked + ) + raise HTTPException( + status_code=400, + detail={ + "error": "Violated tool policy", + "blocked_tools": blocked, + "message": f"Tool(s) {blocked} are blocked by policy.", + }, + ) + + return inputs + + async def _get_policies_cached(self, tool_names: List[str]) -> Dict[str, str]: + """ + Batch-fetch call_policy for the given tool names. + + Caches per individual tool name (not per combination) so that adding + a new tool to a request doesn't invalidate the cached policies for all + the other tools already in the cache. + """ + from litellm.proxy.db.tool_registry_writer import get_tools_by_names + from litellm.proxy.proxy_server import prisma_client + + if not tool_names or prisma_client is None: + return {} + + result: Dict[str, str] = {} + cache_misses: List[str] = [] + + for name in tool_names: + cached = await self._policy_cache.async_get_cache(f"tool_policy:{name}") + if cached is not None and isinstance(cached, str): + result[name] = cached + else: + cache_misses.append(name) + + if cache_misses: + fetched = await get_tools_by_names( + prisma_client=prisma_client, tool_names=cache_misses + ) + for name, policy in fetched.items(): + result[name] = policy + await self._policy_cache.async_set_cache( + key=f"tool_policy:{name}", + value=policy, + ttl=TOOL_POLICY_CACHE_TTL_SECONDS, + ) + verbose_proxy_logger.debug( + "ToolPolicyGuardrail: fetched %d policies from DB (cache hits: %d)", + len(cache_misses), + len(tool_names) - len(cache_misses), + ) + + return result diff --git a/litellm/proxy/management_endpoints/tool_management_endpoints.py b/litellm/proxy/management_endpoints/tool_management_endpoints.py new file mode 100644 index 00000000000..89880c9a4ec --- /dev/null +++ b/litellm/proxy/management_endpoints/tool_management_endpoints.py @@ -0,0 +1,149 @@ +""" +TOOL POLICY MANAGEMENT + +All /tool management endpoints + +GET /v1/tool/list - List all discovered tools and their policies +GET /v1/tool/{tool_name} - Get a single tool's details +POST /v1/tool/policy - Update the call_policy for a tool +""" + +from typing import Optional + +from fastapi import APIRouter, Depends, HTTPException + +from litellm._logging import verbose_proxy_logger +from litellm.proxy._types import CommonProxyErrors, UserAPIKeyAuth +from litellm.proxy.auth.user_api_key_auth import user_api_key_auth +from litellm.types.tool_management import ( + LiteLLM_ToolTableRow, + ToolCallPolicy, + ToolListResponse, + ToolPolicyUpdateRequest, + ToolPolicyUpdateResponse, +) + +router = APIRouter() + + +@router.get( + "/v1/tool/list", + tags=["tool management"], + dependencies=[Depends(user_api_key_auth)], + response_model=ToolListResponse, +) +async def list_tools( + call_policy: Optional[ToolCallPolicy] = None, + user_api_key_dict: UserAPIKeyAuth = Depends(user_api_key_auth), +): + """ + List all auto-discovered tools and their call policies. + + Parameters: + - call_policy: Optional filter — one of "trusted", "untrusted", "dual_llm", "blocked" + """ + from litellm.proxy.db.tool_registry_writer import list_tools as db_list_tools + from litellm.proxy.proxy_server import prisma_client + + if prisma_client is None: + raise HTTPException( + status_code=500, detail=CommonProxyErrors.db_not_connected_error.value + ) + + try: + tools = await db_list_tools(prisma_client=prisma_client, call_policy=call_policy) + return ToolListResponse(tools=tools, total=len(tools)) + except Exception as e: + verbose_proxy_logger.exception("Error listing tools: %s", e) + raise HTTPException(status_code=500, detail=str(e)) + + +@router.get( + "/v1/tool/{tool_name:path}", + tags=["tool management"], + dependencies=[Depends(user_api_key_auth)], + response_model=LiteLLM_ToolTableRow, +) +async def get_tool( + tool_name: str, + user_api_key_dict: UserAPIKeyAuth = Depends(user_api_key_auth), +): + """ + Get details for a single tool. + + Parameters: + - tool_name: The tool name (supports namespaced names with slashes) + """ + from litellm.proxy.db.tool_registry_writer import get_tool as db_get_tool + from litellm.proxy.proxy_server import prisma_client + + if prisma_client is None: + raise HTTPException( + status_code=500, detail=CommonProxyErrors.db_not_connected_error.value + ) + + try: + tool = await db_get_tool(prisma_client=prisma_client, tool_name=tool_name) + if tool is None: + raise HTTPException( + status_code=404, detail=f"Tool '{tool_name}' not found" + ) + return tool + except HTTPException: + raise + except Exception as e: + verbose_proxy_logger.exception("Error getting tool: %s", e) + raise HTTPException(status_code=500, detail=str(e)) + + +@router.post( + "/v1/tool/policy", + tags=["tool management"], + dependencies=[Depends(user_api_key_auth)], + response_model=ToolPolicyUpdateResponse, +) +async def update_tool_policy( + data: ToolPolicyUpdateRequest, + user_api_key_dict: UserAPIKeyAuth = Depends(user_api_key_auth), +): + """ + Set the call policy for a tool. + + Parameters: + - tool_name: str - The tool to update + - call_policy: "trusted" | "untrusted" | "dual_llm" | "blocked" + + Setting a tool to "blocked" will cause the ToolPolicyGuardrail to remove + that tool_call from LLM responses before returning them to the client. + """ + from litellm.proxy.db.tool_registry_writer import ( + update_tool_policy as db_update_tool_policy, + ) + from litellm.proxy.proxy_server import prisma_client + + if prisma_client is None: + raise HTTPException( + status_code=500, detail=CommonProxyErrors.db_not_connected_error.value + ) + + try: + updated = await db_update_tool_policy( + prisma_client=prisma_client, + tool_name=data.tool_name, + call_policy=data.call_policy, + updated_by=user_api_key_dict.user_id, + ) + if updated is None: + raise HTTPException( + status_code=500, detail=f"Failed to update policy for tool '{data.tool_name}'" + ) + return ToolPolicyUpdateResponse( + tool_name=updated.tool_name, + call_policy=updated.call_policy, + updated=True, + ) + except HTTPException: + raise + except Exception as e: + verbose_proxy_logger.exception("Error updating tool policy: %s", e) + raise HTTPException(status_code=500, detail=str(e)) diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 905a4be35e4..36b6bcc7707 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -410,6 +410,9 @@ from litellm.proxy.management_endpoints.team_endpoints import ( update_team, validate_membership, ) +from litellm.proxy.management_endpoints.tool_management_endpoints import ( + router as tool_management_router, +) from litellm.proxy.management_endpoints.ui_sso import ( get_disabled_non_admin_personal_key_creation, ) @@ -12882,6 +12885,7 @@ app.include_router(budget_management_router) app.include_router(model_management_router) app.include_router(model_access_group_management_router) app.include_router(tag_management_router) +app.include_router(tool_management_router) app.include_router(cost_tracking_settings_router) app.include_router(router_settings_router) app.include_router(fallback_management_router) diff --git a/litellm/proxy/schema.prisma b/litellm/proxy/schema.prisma index 50c0a55a875..23917cf7c7f 100644 --- a/litellm/proxy/schema.prisma +++ b/litellm/proxy/schema.prisma @@ -1051,6 +1051,26 @@ model LiteLLM_PolicyAttachmentTable { updated_by String? } +// Global tool registry - auto-discovered from LLM responses; admins set call_policy here +model LiteLLM_ToolTable { + tool_id String @id @default(uuid()) + tool_name String @unique // e.g. "huggingface_remote-mcp__dynamic_space" + origin String? // MCP server name or "user_defined" + call_policy String @default("untrusted") // "trusted" | "untrusted" | "dual_llm" | "blocked" + call_count Int @default(0) // cumulative number of times this tool was seen + assignments Json? @default("{}") + key_hash String? // hash of the virtual key that first called this tool + team_id String? // team that first called this tool + key_alias String? // human-readable alias of the virtual key + created_at DateTime @default(now()) + created_by String? + updated_at DateTime @default(now()) @updatedAt + updated_by String? + + @@index([call_policy]) + @@index([team_id]) +} + //Unified Access Groups table for storing unified access groups model LiteLLM_AccessGroupTable { access_group_id String @id @default(uuid()) diff --git a/litellm/types/tool_management.py b/litellm/types/tool_management.py new file mode 100644 index 00000000000..8704ff27759 --- /dev/null +++ b/litellm/types/tool_management.py @@ -0,0 +1,42 @@ +""" +Pydantic models for Tool Policy management endpoints. +""" + +from datetime import datetime +from typing import Dict, List, Literal, Optional + +from pydantic import BaseModel + +ToolCallPolicy = Literal["trusted", "untrusted", "dual_llm", "blocked"] + + +class LiteLLM_ToolTableRow(BaseModel): + tool_id: str + tool_name: str + origin: Optional[str] = None + call_policy: ToolCallPolicy = "untrusted" + call_count: int = 0 + assignments: Optional[Dict] = None + key_hash: Optional[str] = None + team_id: Optional[str] = None + key_alias: Optional[str] = None + created_at: Optional[datetime] = None + updated_at: Optional[datetime] = None + created_by: Optional[str] = None + updated_by: Optional[str] = None + + +class ToolListResponse(BaseModel): + tools: List[LiteLLM_ToolTableRow] + total: int + + +class ToolPolicyUpdateRequest(BaseModel): + tool_name: str + call_policy: ToolCallPolicy + + +class ToolPolicyUpdateResponse(BaseModel): + tool_name: str + call_policy: ToolCallPolicy + updated: bool diff --git a/tests/test_litellm/proxy/db/db_transaction_queue/test_tool_discovery_queue.py b/tests/test_litellm/proxy/db/db_transaction_queue/test_tool_discovery_queue.py new file mode 100644 index 00000000000..defdb3834d8 --- /dev/null +++ b/tests/test_litellm/proxy/db/db_transaction_queue/test_tool_discovery_queue.py @@ -0,0 +1,75 @@ +""" +Unit tests for ToolDiscoveryQueue. +""" + +import os +import sys + +import pytest + +sys.path.insert(0, os.path.abspath("../../..")) + +from litellm.proxy.db.db_transaction_queue.tool_discovery_queue import ( + ToolDiscoveryQueue, +) + + +@pytest.fixture +def queue(): + return ToolDiscoveryQueue() + + +def test_add_single_tool(queue): + queue.add_update({"tool_name": "my_tool", "origin": "user_defined"}) + items = queue.flush() + assert len(items) == 1 + assert items[0]["tool_name"] == "my_tool" + assert items[0]["origin"] == "user_defined" + + +def test_deduplication_same_name(queue): + """Adding the same tool_name twice should only keep the first.""" + queue.add_update({"tool_name": "tool_a", "origin": "mcp_server"}) + queue.add_update({"tool_name": "tool_a", "origin": "user_defined"}) + items = queue.flush() + assert len(items) == 1 + assert items[0]["origin"] == "mcp_server" # first wins + + +def test_deduplication_different_names(queue): + queue.add_update({"tool_name": "tool_a"}) + queue.add_update({"tool_name": "tool_b"}) + items = queue.flush() + assert len(items) == 2 + names = {i["tool_name"] for i in items} + assert names == {"tool_a", "tool_b"} + + +def test_flush_clears_pending(queue): + queue.add_update({"tool_name": "tool_x"}) + items1 = queue.flush() + assert len(items1) == 1 + items2 = queue.flush() + assert len(items2) == 0 + + +def test_seen_names_reset_after_flush(queue): + """Seen-set is cleared on flush so the same tool can re-enter the next cycle.""" + queue.add_update({"tool_name": "tool_a"}) + queue.flush() + queue.add_update({"tool_name": "tool_a"}) # same tool, new cycle + items = queue.flush() + assert len(items) == 1 + assert items[0]["tool_name"] == "tool_a" + + +def test_empty_tool_name_ignored(queue): + queue.add_update({"tool_name": ""}) + queue.add_update({"tool_name": None}) # type: ignore[arg-type] + items = queue.flush() + assert len(items) == 0 + + +def test_flush_returns_list(queue): + result = queue.flush() + assert isinstance(result, list) diff --git a/tests/test_litellm/proxy/db/test_tool_registry_writer.py b/tests/test_litellm/proxy/db/test_tool_registry_writer.py new file mode 100644 index 00000000000..44f9e32058a --- /dev/null +++ b/tests/test_litellm/proxy/db/test_tool_registry_writer.py @@ -0,0 +1,197 @@ +""" +Unit tests for tool_registry_writer.py — uses a mock prisma client +that exposes execute_raw / query_raw (matching the actual raw-SQL implementation). +""" + +import os +import sys +from datetime import datetime, timezone +from unittest.mock import AsyncMock, MagicMock + +import pytest + +sys.path.insert(0, os.path.abspath("../../..")) + +from litellm.proxy.db.tool_registry_writer import ( + batch_upsert_tools, + get_tool, + get_tools_by_names, + list_tools, + update_tool_policy, +) + + +def _make_prisma(query_rows=None): + """Return a minimal mock prisma_client with execute_raw / query_raw.""" + default_row = { + "tool_id": "uuid-1", + "tool_name": "my_tool", + "origin": "user_defined", + "call_policy": "untrusted", + "call_count": 1, + "assignments": {}, + "key_hash": None, + "team_id": None, + "key_alias": None, + "created_at": datetime.now(timezone.utc), + "updated_at": datetime.now(timezone.utc), + "created_by": None, + "updated_by": None, + } + rows = query_rows if query_rows is not None else [default_row] + + prisma = MagicMock() + prisma.db.execute_raw = AsyncMock(return_value=None) + prisma.db.query_raw = AsyncMock(return_value=rows) + return prisma + + +@pytest.mark.asyncio +async def test_batch_upsert_tools_calls_execute_raw(): + prisma = _make_prisma() + items = [{"tool_name": "tool_a", "origin": "mcp_server", "created_by": None}] + await batch_upsert_tools(prisma, items) + prisma.db.execute_raw.assert_awaited_once() + call_args = prisma.db.execute_raw.call_args + sql = call_args.args[0] + assert "LiteLLM_ToolTable" in sql + assert "ON CONFLICT" in sql + + +@pytest.mark.asyncio +async def test_batch_upsert_tools_empty_list(): + prisma = _make_prisma() + await batch_upsert_tools(prisma, []) + prisma.db.execute_raw.assert_not_awaited() + + +@pytest.mark.asyncio +async def test_batch_upsert_tools_skips_empty_names(): + prisma = _make_prisma() + items = [{"tool_name": "", "origin": None}, {"tool_name": None}] # type: ignore[list-item] + await batch_upsert_tools(prisma, items) + prisma.db.execute_raw.assert_not_awaited() + + +@pytest.mark.asyncio +async def test_batch_upsert_multiple_tools_calls_execute_raw_per_tool(): + prisma = _make_prisma() + items = [ + {"tool_name": "tool_a", "origin": "mcp_server", "created_by": None}, + {"tool_name": "tool_b", "origin": "user_defined", "created_by": "alice"}, + ] + await batch_upsert_tools(prisma, items) + assert prisma.db.execute_raw.await_count == 2 + + +@pytest.mark.asyncio +async def test_list_tools_no_filter(): + row = { + "tool_id": "id1", + "tool_name": "tool_a", + "origin": "mcp", + "call_policy": "untrusted", + "call_count": 5, + "assignments": {}, + "key_hash": None, + "team_id": None, + "key_alias": None, + "created_at": datetime.now(timezone.utc), + "updated_at": datetime.now(timezone.utc), + "created_by": None, + "updated_by": None, + } + prisma = _make_prisma(query_rows=[row]) + result = await list_tools(prisma) + assert len(result) == 1 + assert result[0].tool_name == "tool_a" + assert result[0].call_count == 5 + prisma.db.query_raw.assert_awaited_once() + + +@pytest.mark.asyncio +async def test_list_tools_with_policy_filter(): + row = { + "tool_id": "id1", + "tool_name": "blocked_tool", + "origin": None, + "call_policy": "blocked", + "call_count": 2, + "assignments": None, + "key_hash": None, + "team_id": None, + "key_alias": None, + "created_at": datetime.now(timezone.utc), + "updated_at": datetime.now(timezone.utc), + "created_by": None, + "updated_by": None, + } + prisma = _make_prisma(query_rows=[row]) + result = await list_tools(prisma, call_policy="blocked") + assert result[0].call_policy == "blocked" + call_args = prisma.db.query_raw.call_args + sql = call_args.args[0] + assert "WHERE call_policy" in sql + + +@pytest.mark.asyncio +async def test_get_tool_found(): + prisma = _make_prisma() + result = await get_tool(prisma, "my_tool") + assert result is not None + assert result.tool_name == "my_tool" + prisma.db.query_raw.assert_awaited_once() + + +@pytest.mark.asyncio +async def test_get_tool_not_found(): + prisma = _make_prisma(query_rows=[]) + result = await get_tool(prisma, "nonexistent") + assert result is None + + +@pytest.mark.asyncio +async def test_update_tool_policy_calls_execute_raw(): + row = { + "tool_id": "uuid-1", + "tool_name": "my_tool", + "origin": "user_defined", + "call_policy": "blocked", + "call_count": 1, + "assignments": {}, + "key_hash": None, + "team_id": None, + "key_alias": None, + "created_at": datetime.now(timezone.utc), + "updated_at": datetime.now(timezone.utc), + "created_by": None, + "updated_by": "admin", + } + prisma = _make_prisma(query_rows=[row]) + result = await update_tool_policy(prisma, "my_tool", "blocked", "admin") + assert result is not None + assert result.call_policy == "blocked" + prisma.db.execute_raw.assert_awaited_once() + call_args = prisma.db.execute_raw.call_args + sql = call_args.args[0] + assert "ON CONFLICT" in sql + assert "call_policy" in sql + + +@pytest.mark.asyncio +async def test_get_tools_by_names_returns_policy_map(): + rows = [ + {"tool_name": "tool_a", "call_policy": "trusted"}, + {"tool_name": "tool_b", "call_policy": "blocked"}, + ] + prisma = _make_prisma(query_rows=rows) + result = await get_tools_by_names(prisma, ["tool_a", "tool_b"]) + assert result == {"tool_a": "trusted", "tool_b": "blocked"} + + +@pytest.mark.asyncio +async def test_get_tools_by_names_empty_list(): + prisma = _make_prisma() + result = await get_tools_by_names(prisma, []) + assert result == {} + prisma.db.query_raw.assert_not_awaited() diff --git a/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_tool_policy_guardrail.py b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_tool_policy_guardrail.py new file mode 100644 index 00000000000..c6a81efbf0b --- /dev/null +++ b/tests/test_litellm/proxy/guardrails/guardrail_hooks/test_tool_policy_guardrail.py @@ -0,0 +1,181 @@ +""" +Unit tests for ToolPolicyGuardrail. +""" + +import os +import sys +from typing import Any +from unittest.mock import AsyncMock, MagicMock, patch + +import pytest +from fastapi import HTTPException + +sys.path.insert(0, os.path.abspath("../../../../../..")) + +from litellm.proxy.guardrails.guardrail_hooks.tool_policy.tool_policy_guardrail import ( + ToolPolicyGuardrail, +) +from litellm.types.guardrails import GuardrailEventHooks + + +@pytest.fixture +def guardrail(): + return ToolPolicyGuardrail() + + +# --- helpers --- + +def _tool_request_inputs(tool_names: list) -> dict: + return { + "tools": [ + {"type": "function", "function": {"name": name, "description": ""}} + for name in tool_names + ] + } + + +def _tool_response_inputs(tool_names: list) -> dict: + return { + "tool_calls": [ + {"type": "function", "function": {"name": name}} + for name in tool_names + ] + } + + +# --- tests --- + + +def test_guardrail_supports_pre_and_post_call(guardrail): + hooks = guardrail.supported_event_hooks + assert GuardrailEventHooks.pre_call in hooks + assert GuardrailEventHooks.post_call in hooks + + +@pytest.mark.asyncio +async def test_no_tools_in_request_passes_through(guardrail): + inputs: Any = {"tools": []} + result = await guardrail.apply_guardrail( + inputs=inputs, request_data={}, input_type="request" + ) + assert result is inputs + + +@pytest.mark.asyncio +async def test_no_tool_calls_in_response_passes_through(guardrail): + inputs: Any = {"tool_calls": []} + result = await guardrail.apply_guardrail( + inputs=inputs, request_data={}, input_type="response" + ) + assert result is inputs + + +@pytest.mark.asyncio +async def test_untrusted_tools_pass_through(guardrail): + policy_map = {"search": "untrusted", "read_file": "trusted"} + with patch.object(guardrail, "_get_policies_cached", new=AsyncMock(return_value=policy_map)): + inputs: Any = _tool_request_inputs(["search", "read_file"]) + result = await guardrail.apply_guardrail( + inputs=inputs, request_data={}, input_type="request" + ) + assert result is inputs + + +@pytest.mark.asyncio +async def test_blocked_tool_in_request_raises_http_exception(guardrail): + policy_map = {"dangerous_tool": "blocked"} + with patch.object(guardrail, "_get_policies_cached", new=AsyncMock(return_value=policy_map)): + inputs: Any = _tool_request_inputs(["dangerous_tool"]) + with pytest.raises(HTTPException) as exc_info: + await guardrail.apply_guardrail( + inputs=inputs, request_data={}, input_type="request" + ) + assert exc_info.value.status_code == 400 + assert "dangerous_tool" in exc_info.value.detail["blocked_tools"] + + +@pytest.mark.asyncio +async def test_blocked_tool_in_response_raises_http_exception(guardrail): + policy_map = {"exfil_tool": "blocked"} + with patch.object(guardrail, "_get_policies_cached", new=AsyncMock(return_value=policy_map)): + inputs: Any = _tool_response_inputs(["exfil_tool"]) + with pytest.raises(HTTPException) as exc_info: + await guardrail.apply_guardrail( + inputs=inputs, request_data={}, input_type="response" + ) + assert exc_info.value.status_code == 400 + assert "exfil_tool" in exc_info.value.detail["blocked_tools"] + + +@pytest.mark.asyncio +async def test_mixed_blocked_and_allowed_raises_for_blocked(guardrail): + policy_map = {"safe_tool": "trusted", "bad_tool": "blocked"} + with patch.object(guardrail, "_get_policies_cached", new=AsyncMock(return_value=policy_map)): + inputs: Any = _tool_request_inputs(["safe_tool", "bad_tool"]) + with pytest.raises(HTTPException) as exc_info: + await guardrail.apply_guardrail( + inputs=inputs, request_data={}, input_type="request" + ) + blocked = exc_info.value.detail["blocked_tools"] + assert "bad_tool" in blocked + assert "safe_tool" not in blocked + + +@pytest.mark.asyncio +async def test_tool_not_in_db_passes_through(guardrail): + """Tools not found in the DB (no entry) should not be blocked.""" + with patch.object(guardrail, "_get_policies_cached", new=AsyncMock(return_value={})): + inputs: Any = _tool_request_inputs(["unknown_tool"]) + result = await guardrail.apply_guardrail( + inputs=inputs, request_data={}, input_type="request" + ) + assert result is inputs + + +@pytest.mark.asyncio +async def test_get_policies_cached_uses_cache(guardrail): + """Second call with same tool names should return the cached result.""" + policy_map = {"tool_a": "trusted"} + with patch( + "litellm.proxy.db.tool_registry_writer.get_tools_by_names", + new=AsyncMock(return_value=policy_map), + ) as mock_db, patch( + "litellm.proxy.proxy_server.prisma_client", + new=MagicMock(), + ): + # first call — should hit DB + result1 = await guardrail._get_policies_cached(["tool_a"]) + assert result1 == policy_map + + # second call — should hit cache, not DB again + result2 = await guardrail._get_policies_cached(["tool_a"]) + assert result2 == policy_map + + assert mock_db.call_count == 1 + + +@pytest.mark.asyncio +async def test_get_policies_cached_no_prisma(guardrail): + """Without a prisma client, returns empty dict.""" + with patch( + "litellm.proxy.proxy_server.prisma_client", + None, + ): + result = await guardrail._get_policies_cached(["tool_a"]) + assert result == {} + + +@pytest.mark.asyncio +async def test_response_tool_calls_as_objects(guardrail): + """tool_calls that are objects (not dicts) with .function.name should work.""" + policy_map = {"obj_tool": "blocked"} + with patch.object(guardrail, "_get_policies_cached", new=AsyncMock(return_value=policy_map)): + fn = MagicMock() + fn.name = "obj_tool" + tc = MagicMock() + tc.function = fn + inputs: Any = {"tool_calls": [tc]} + with pytest.raises(HTTPException): + await guardrail.apply_guardrail( + inputs=inputs, request_data={}, input_type="response" + ) diff --git a/tests/test_litellm/proxy/management_endpoints/test_tool_management_endpoints.py b/tests/test_litellm/proxy/management_endpoints/test_tool_management_endpoints.py new file mode 100644 index 00000000000..6f1d373fdee --- /dev/null +++ b/tests/test_litellm/proxy/management_endpoints/test_tool_management_endpoints.py @@ -0,0 +1,149 @@ +""" +Unit tests for tool management endpoints (/v1/tool/*). +Uses FastAPI TestClient with mocked DB functions. + +Patches target the source modules (litellm.proxy.db.tool_registry_writer.* +and litellm.proxy.proxy_server.prisma_client) because the endpoint code +imports these inside function bodies to avoid circular imports. +""" + +import os +import sys +from datetime import datetime, timezone +from typing import Optional +from unittest.mock import AsyncMock, MagicMock, patch + +from fastapi import FastAPI +from fastapi.testclient import TestClient + +sys.path.insert(0, os.path.abspath("../../..")) + +from litellm.proxy.management_endpoints.tool_management_endpoints import router +from litellm.types.tool_management import LiteLLM_ToolTableRow + +# --- helpers --- + + +def _make_tool_row( + tool_name: str = "my_tool", + call_policy: str = "untrusted", + origin: Optional[str] = None, +) -> LiteLLM_ToolTableRow: + now = datetime.now(timezone.utc) + return LiteLLM_ToolTableRow( + tool_id="uuid-1", + tool_name=tool_name, + origin=origin, + call_policy=call_policy, # type: ignore[arg-type] + assignments={}, + created_at=now, + updated_at=now, + ) + + +def _make_app() -> FastAPI: + """Build a minimal FastAPI app with the tool management router.""" + app = FastAPI() + app.include_router(router) + return app + + +# Stub the auth dependency so we don't need a real proxy running. +def _override_auth(): + from litellm.proxy._types import UserAPIKeyAuth + + return UserAPIKeyAuth(api_key="sk-test", user_id="admin") + + +# A real (non-None) prisma stub for truthiness checks. +_MOCK_PRISMA = MagicMock() + + +# --- test class --- + + +class TestToolManagementEndpoints: + def setup_method(self): + from litellm.proxy.auth.user_api_key_auth import user_api_key_auth + + app = _make_app() + app.dependency_overrides[user_api_key_auth] = _override_auth + self.client = TestClient(app, raise_server_exceptions=True) + + @patch( + "litellm.proxy.db.tool_registry_writer.list_tools", + new_callable=AsyncMock, + ) + @patch("litellm.proxy.proxy_server.prisma_client", _MOCK_PRISMA) + def test_list_tools_returns_200(self, mock_db_list): + mock_db_list.return_value = [_make_tool_row()] + + resp = self.client.get("/v1/tool/list") + assert resp.status_code == 200 + body = resp.json() + assert body["total"] == 1 + assert body["tools"][0]["tool_name"] == "my_tool" + + @patch( + "litellm.proxy.db.tool_registry_writer.list_tools", + new_callable=AsyncMock, + ) + @patch("litellm.proxy.proxy_server.prisma_client", _MOCK_PRISMA) + def test_list_tools_with_policy_filter(self, mock_db_list): + mock_db_list.return_value = [_make_tool_row(call_policy="blocked")] + + resp = self.client.get("/v1/tool/list?call_policy=blocked") + assert resp.status_code == 200 + assert resp.json()["tools"][0]["call_policy"] == "blocked" + + @patch( + "litellm.proxy.db.tool_registry_writer.get_tool", + new_callable=AsyncMock, + ) + @patch("litellm.proxy.proxy_server.prisma_client", _MOCK_PRISMA) + def test_get_tool_found(self, mock_db_get): + mock_db_get.return_value = _make_tool_row(tool_name="tool_a") + + resp = self.client.get("/v1/tool/tool_a") + assert resp.status_code == 200 + assert resp.json()["tool_name"] == "tool_a" + + @patch( + "litellm.proxy.db.tool_registry_writer.get_tool", + new_callable=AsyncMock, + ) + @patch("litellm.proxy.proxy_server.prisma_client", _MOCK_PRISMA) + def test_get_tool_not_found_returns_404(self, mock_db_get): + mock_db_get.return_value = None + + resp = self.client.get("/v1/tool/nonexistent", follow_redirects=True) + assert resp.status_code == 404 + + @patch( + "litellm.proxy.db.tool_registry_writer.update_tool_policy", + new_callable=AsyncMock, + ) + @patch("litellm.proxy.proxy_server.prisma_client", _MOCK_PRISMA) + def test_update_tool_policy_blocked(self, mock_db_update): + mock_db_update.return_value = _make_tool_row(call_policy="blocked") + + resp = self.client.post( + "/v1/tool/policy", + json={"tool_name": "my_tool", "call_policy": "blocked"}, + ) + assert resp.status_code == 200 + body = resp.json() + assert body["call_policy"] == "blocked" + assert body["updated"] is True + + @patch("litellm.proxy.proxy_server.prisma_client", None) + def test_list_tools_no_db_returns_500(self): + resp = self.client.get("/v1/tool/list") + assert resp.status_code == 500 + + def test_update_tool_policy_invalid_policy_returns_422(self): + resp = self.client.post( + "/v1/tool/policy", + json={"tool_name": "my_tool", "call_policy": "invalid_value"}, + ) + assert resp.status_code == 422 diff --git a/ui/litellm-dashboard/src/app/page.tsx b/ui/litellm-dashboard/src/app/page.tsx index fb749d7afb0..258c2ccb0e0 100644 --- a/ui/litellm-dashboard/src/app/page.tsx +++ b/ui/litellm-dashboard/src/app/page.tsx @@ -38,6 +38,7 @@ import Usage from "@/components/usage"; import UserDashboard from "@/components/user_dashboard"; import { AccessGroupsPage } from "@/components/AccessGroups/AccessGroupsPage"; import VectorStoreManagement from "@/components/vector_store_management"; +import ToolPolicies from "@/components/ToolPolicies"; import SpendLogsTable from "@/components/view_logs"; import ViewUserDashboard from "@/components/view_users"; import { ThemeProvider } from "@/contexts/ThemeContext"; @@ -548,6 +549,8 @@ function CreateKeyPageContent() { ) : page == "vector-stores" ? ( + ) : page == "tool-policies" ? ( + ) : page == "guardrails-monitor" ? ( ) : page == "new_usage" ? ( diff --git a/ui/litellm-dashboard/src/components/ToolPolicies.tsx b/ui/litellm-dashboard/src/components/ToolPolicies.tsx new file mode 100644 index 00000000000..0e3f5434e7f --- /dev/null +++ b/ui/litellm-dashboard/src/components/ToolPolicies.tsx @@ -0,0 +1,415 @@ +"use client"; + +import React, { useCallback, useDeferredValue, useEffect, useState } from "react"; +import { Select, Switch, Tooltip } from "antd"; +import { Select, Tooltip } from "antd"; +import { + Table, + TableHead, + TableHeaderCell, + TableBody, + TableRow, + TableCell, +} from "@tremor/react"; +import { TimeCell } from "./view_logs/time_cell"; +import { TableHeaderSortDropdown } from "./common_components/TableHeaderSortDropdown/TableHeaderSortDropdown"; +import type { SortState } from "./common_components/TableHeaderSortDropdown/TableHeaderSortDropdown"; +import FilterComponent, { FilterOption } from "./molecules/filter"; +import { fetchToolsList, updateToolPolicy, ToolRow } from "./networking"; + +const POLICY_OPTIONS = [ + { value: "trusted", label: "trusted", color: "#065f46", bg: "#d1fae5", border: "#6ee7b7" }, + { value: "blocked", label: "blocked", color: "#991b1b", bg: "#fee2e2", border: "#fca5a5" }, +] as const; + +type PolicyValue = "trusted" | "blocked"; + +const policyStyle = (p: string) => + POLICY_OPTIONS.find((o) => o.value === p) ?? POLICY_OPTIONS[1]; + +type SortField = "tool_name" | "call_policy" | "team_id" | "key_alias" | "created_at" | "call_count"; + +interface FilterValues { + [key: string]: string; +} + +interface ToolPoliciesProps { + accessToken: string | null; + userRole?: string; +} + +const PolicySelect: React.FC<{ + value: string; + toolName: string; + saving: boolean; + onChange: (toolName: string, policy: string) => void; +}> = ({ value, toolName, saving, onChange }) => { + const style = policyStyle(value); + return ( + { setSearchTerm(e.target.value); setCurrentPage(1); }} + /> + + + +
+ +
+ Live Tail + +
+ + + + +
+ + Showing {filtered.length === 0 ? 0 : (currentPage - 1) * pageSize + 1} - {Math.min(currentPage * pageSize, filtered.length)} of {filtered.length} results + + Page {currentPage} of {totalPages} +
+ + +
+
+ + + {/* Filter row */} +
+ +
+ + + {/* Auto-refresh banner */} + {isLiveTail && ( +
+ Auto-refreshing every 15 seconds + +
+ )} + + {error && ( +
{error}
+ )} + + {/* Table */} + + + + + + + + + Key Hash + + Origin + + + + {loading ? ( + + Loading tools… + + ) : paginated.length === 0 ? ( + + + No tools discovered yet. Make a chat completion that returns tool_calls to start auto-discovery. + + + ) : ( + paginated.map((tool) => ( + + + + + + + + {tool.tool_name} + + + + + + + + {(tool.call_count ?? 0).toLocaleString()} + + + + {tool.team_id ?? "-"} + + + + + + {tool.key_hash ?? "-"} + + + + + + {tool.key_alias ?? "-"} + + + + + {tool.origin ?? "-"} + + + + )) + )} + +
+ + {/* Bottom pagination (only when > 1 page) */} + {totalPages > 1 && ( +
+ Showing {(currentPage - 1) * pageSize + 1} - {Math.min(currentPage * pageSize, sorted.length)} of {sorted.length} +
+ + +
+
+ )} + + + ); +}; + +export default ToolPolicies; diff --git a/ui/litellm-dashboard/src/components/leftnav.tsx b/ui/litellm-dashboard/src/components/leftnav.tsx index da3ca2a8bae..2cbeb22ec81 100644 --- a/ui/litellm-dashboard/src/components/leftnav.tsx +++ b/ui/litellm-dashboard/src/components/leftnav.tsx @@ -134,6 +134,12 @@ const menuGroups: MenuGroup[] = [ label: "Vector Stores", icon: , }, + { + key: "tool-policies", + page: "tool-policies", + label: "Tool Policies", + icon: , + }, ], }, ], diff --git a/ui/litellm-dashboard/src/components/networking.tsx b/ui/litellm-dashboard/src/components/networking.tsx index 6ffd744cce9..8536a584ee1 100644 --- a/ui/litellm-dashboard/src/components/networking.tsx +++ b/ui/litellm-dashboard/src/components/networking.tsx @@ -9854,3 +9854,57 @@ export const checkGdprCompliance = async ( } return response.json(); }; + +export interface ToolRow { + tool_id: string; + tool_name: string; + origin?: string; + call_policy: string; + call_count?: number; + assignments?: Record; + key_hash?: string; + team_id?: string; + key_alias?: string; + created_at?: string; + updated_at?: string; + created_by?: string; + updated_by?: string; +} + +export const fetchToolsList = async (accessToken: string): Promise => { + const url = proxyBaseUrl ? `${proxyBaseUrl}/v1/tool/list` : `/v1/tool/list`; + const response = await fetch(url, { + method: "GET", + headers: { + [globalLitellmHeaderName]: `Bearer ${accessToken}`, + "Content-Type": "application/json", + }, + }); + if (!response.ok) { + const errorData = await response.text(); + throw new Error(errorData); + } + const data = await response.json(); + return data.tools ?? []; +}; + +export const updateToolPolicy = async ( + accessToken: string, + toolName: string, + callPolicy: string +): Promise => { + const url = proxyBaseUrl ? `${proxyBaseUrl}/v1/tool/policy` : `/v1/tool/policy`; + const response = await fetch(url, { + method: "POST", + headers: { + [globalLitellmHeaderName]: `Bearer ${accessToken}`, + "Content-Type": "application/json", + }, + body: JSON.stringify({ tool_name: toolName, call_policy: callPolicy }), + }); + if (!response.ok) { + const errorData = await response.text(); + throw new Error(errorData); + } + return response.json(); +}; From 360643e21315eb02ca790eb22effb87a6d48167b Mon Sep 17 00:00:00 2001 From: Ishaan Jaff Date: Tue, 24 Feb 2026 16:40:04 -0800 Subject: [PATCH 16/41] [Feat] UI - Allow using AI to understand Usage patterns (#22042) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit * Add Ask AI chat component to Usage page - Create UsageAIChatModal component with streaming chat interface - Integrate with existing model hub for model selection - Pass usage data context (spend, models, providers, keys) to AI - Add Ask AI button next to Export Data button in global view - Add tests for the new component and integration Co-authored-by: Ishaan Jaff * Convert Ask AI from modal to right-side sliding panel - Replace UsageAIChatModal with UsageAIChatPanel - Panel slides in from right side, usage page stays visible - Full-height panel with header, model selector, chat area, and input - Smooth CSS transition for open/close animation - Update tests for new panel component (34 tests passing) Co-authored-by: Ishaan Jaff * Remove build output directory from tracking Co-authored-by: Ishaan Jaff * Add backend AI usage chat endpoint with tool calling Backend: - New /usage/ai/chat SSE streaming endpoint - AI agent has get_usage_data tool that queries /user/daily/activity/aggregated - Follows same architecture as policy AI suggest (litellm.acompletion + tools) - Non-admin users are restricted to their own data - 12 backend unit tests Frontend: - Panel now calls /usage/ai/chat backend endpoint via SSE - Removed direct OpenAI client calls from frontend - Added usageAiChatStream networking function following enrichPolicyTemplateStream pattern Co-authored-by: Ishaan Jaff * Make model selection optional, default to gpt-4o-mini on backend Co-authored-by: Ishaan Jaff * Add team/tag tools, status indicators, and improved AI agent - AI agent now has 3 tools: get_usage_data, get_team_usage_data, get_tag_usage_data - Stream status events (Thinking... Fetching... Analyzing...) to UI - Frontend shows spinner + status text during tool execution - Better system prompt guiding tool selection - Entity summariser for team/tag data with ranked breakdowns - 13 backend tests, 34 frontend tests passing Co-authored-by: Ishaan Jaff * Fix: inject today's date into system prompt so AI resolves relative dates correctly Co-authored-by: Ishaan Jaff * Show tool calls as distinct steps + render markdown in responses - Backend emits tool_call events with tool_name, label, args, and status - Frontend shows each tool call as a step with ✓/spinner/✗ indicator - Tool call steps show icon, label, date range, and filters - AI responses rendered with ReactMarkdown (bold, lists, tables, code) - Cursor-like UX: Thinking → tool calls → Analyzing → streamed answer Co-authored-by: Ishaan Jaff * Refactor backend for code quality: proper types, constants, all functions ≤50 LOC - TypedDict for SSE events (SSEStatusEvent, SSEToolCallEvent, etc.) and ToolHandler - Constants for table names, entity fields, temperature, page sizes, top-N limits - Shared _query_activity() eliminates duplicated fetch logic - _accumulate_breakdown() + _ranked_lines() replace inline aggregation loops - Extracted _process_tool_call() and _stream_final_response() from main stream fn - Black + Ruff clean, all 15 functions verified ≤50 LOC - Replaced Tremor Button with Antd Button in panel (Tremor deprecated per AGENTS.md) Co-authored-by: Ishaan Jaff * Address greptile review: security fixes and input validation - Restrict team/tag tools to admin-only users (non-admins only get get_usage_data) - Constrain ChatMessage.role to Literal['user', 'assistant'] to prevent system prompt injection - Add test for base tools restriction (non-admin gets 1 tool, admin gets 3) - Issues 3 (unused imports) and 4 (inline datetime) were already fixed in prior commit Co-authored-by: Ishaan Jaff * Address greptile round 2: sanitize errors, defense-in-depth allowlist, revert tsconfig - Sanitize error messages: generic 'An internal error occurred' sent to client, full exception logged server-side via verbose_proxy_logger - Defense-in-depth: _process_tool_call validates fn_name against role-based allowlist before dispatch (even though LLM only receives allowed tools) - Revert tsconfig.json jsx back to 'preserve' (Next.js recommended default) Co-authored-by: Ishaan Jaff * Role-scoped system prompt + additional test coverage - System prompt is now role-aware: admin sees all 3 tool descriptions, non-admin only sees get_usage_data (consistent with tool filtering) - Added tests: non-admin prompt excludes team/tag tools, date injection - 15 backend tests, 34 frontend tests passing Co-authored-by: Ishaan Jaff * Fix LLM arg validation + cap conversation size at 20 messages - _resolve_fetch_kwargs uses .get() with ValueError for missing dates (handles malformed LLM tool arguments gracefully) - MAX_CHAT_MESSAGES = 20 constant; backend truncates to last 20 - Frontend also sends only last 20 messages per request - Prevents excessive token usage and context-length errors Co-authored-by: Ishaan Jaff --------- Co-authored-by: Cursor Agent Co-authored-by: Ishaan Jaff --- .../usage_endpoints/__init__.py | 9 + .../usage_endpoints/ai_usage_chat.py | 578 ++++++++++++++++++ .../usage_endpoints/endpoints.py | 65 ++ litellm/proxy/proxy_server.py | 2 + .../usage_endpoints/__init__.py | 0 .../usage_endpoints/test_ai_usage_chat.py | 402 ++++++++++++ ui/litellm-dashboard/package-lock.json | 151 +---- .../components/UsageAIChatPanel.test.tsx | 85 +++ .../UsagePage/components/UsageAIChatPanel.tsx | 402 ++++++++++++ .../components/UsagePageView.test.tsx | 26 + .../UsagePage/components/UsagePageView.tsx | 51 +- .../src/components/networking.tsx | 76 +++ 12 files changed, 1704 insertions(+), 143 deletions(-) create mode 100644 litellm/proxy/management_endpoints/usage_endpoints/__init__.py create mode 100644 litellm/proxy/management_endpoints/usage_endpoints/ai_usage_chat.py create mode 100644 litellm/proxy/management_endpoints/usage_endpoints/endpoints.py create mode 100644 tests/test_litellm/proxy/management_endpoints/usage_endpoints/__init__.py create mode 100644 tests/test_litellm/proxy/management_endpoints/usage_endpoints/test_ai_usage_chat.py create mode 100644 ui/litellm-dashboard/src/components/UsagePage/components/UsageAIChatPanel.test.tsx create mode 100644 ui/litellm-dashboard/src/components/UsagePage/components/UsageAIChatPanel.tsx diff --git a/litellm/proxy/management_endpoints/usage_endpoints/__init__.py b/litellm/proxy/management_endpoints/usage_endpoints/__init__.py new file mode 100644 index 00000000000..6e68dcd2a2e --- /dev/null +++ b/litellm/proxy/management_endpoints/usage_endpoints/__init__.py @@ -0,0 +1,9 @@ +""" +Usage endpoints package. + +Re-exports the router from endpoints module. +""" + +from litellm.proxy.management_endpoints.usage_endpoints.endpoints import ( # noqa: F401 + router, +) diff --git a/litellm/proxy/management_endpoints/usage_endpoints/ai_usage_chat.py b/litellm/proxy/management_endpoints/usage_endpoints/ai_usage_chat.py new file mode 100644 index 00000000000..f156be7d2cc --- /dev/null +++ b/litellm/proxy/management_endpoints/usage_endpoints/ai_usage_chat.py @@ -0,0 +1,578 @@ +""" +AI Usage Chat - uses LLM tool calling to answer questions about +usage/spend data by querying the aggregated daily activity endpoints. +""" + +import json +from datetime import date +from typing import Any, AsyncIterator, Callable, Dict, List, Literal, Optional + +import litellm +from litellm._logging import verbose_proxy_logger +from litellm.constants import DEFAULT_COMPETITOR_DISCOVERY_MODEL +from litellm.types.proxy.management_endpoints.common_daily_activity import ( + SpendAnalyticsPaginatedResponse, +) + +from typing_extensions import TypedDict + +# --------------------------------------------------------------------------- +# Constants +# --------------------------------------------------------------------------- + +USAGE_AI_TEMPERATURE = 0.2 + +TABLE_DAILY_USER_SPEND = "litellm_dailyuserspend" +TABLE_DAILY_TEAM_SPEND = "litellm_dailyteamspend" +TABLE_DAILY_TAG_SPEND = "litellm_dailytagspend" + +ENTITY_FIELD_USER = "user_id" +ENTITY_FIELD_TEAM = "team_id" +ENTITY_FIELD_TAG = "tag" + +PAGINATED_PAGE_SIZE = 200 +MAX_CHAT_MESSAGES = 20 +TOP_N_MODELS = 15 +TOP_N_PROVIDERS = 10 +TOP_N_KEYS = 10 + +# --------------------------------------------------------------------------- +# Types +# --------------------------------------------------------------------------- + + +class SSEStatusEvent(TypedDict): + type: Literal["status"] + message: str + + +class SSEToolCallEvent(TypedDict, total=False): + type: Literal["tool_call"] + tool_name: str + tool_label: str + arguments: Dict[str, str] + status: Literal["running", "complete", "error"] + error: str + + +class SSEChunkEvent(TypedDict): + type: Literal["chunk"] + content: str + + +class SSEDoneEvent(TypedDict): + type: Literal["done"] + + +class SSEErrorEvent(TypedDict): + type: Literal["error"] + message: str + + +SSEEvent = ( + SSEStatusEvent | SSEToolCallEvent | SSEChunkEvent | SSEDoneEvent | SSEErrorEvent +) + + +class ToolHandler(TypedDict): + fetch: Callable[..., Any] + summarise: Callable[[Dict[str, Any]], str] + label: str + + +# --------------------------------------------------------------------------- +# Tool definitions (OpenAI function-calling schema) +# --------------------------------------------------------------------------- + +_DATE_PARAMS = { + "start_date": {"type": "string", "description": "Start date in YYYY-MM-DD format"}, + "end_date": {"type": "string", "description": "End date in YYYY-MM-DD format"}, +} + +_TOOL_USAGE = { + "type": "function", + "function": { + "name": "get_usage_data", + "description": ( + "Fetch aggregated global usage/spend data. Returns daily spend, " + "token counts, request counts, and breakdowns by model, provider, " + "and API key. Use for overall spend, top models, top providers." + ), + "parameters": { + "type": "object", + "properties": { + **_DATE_PARAMS, + "user_id": { + "type": "string", + "description": "Optional user ID filter. Omit for global view.", + }, + }, + "required": ["start_date", "end_date"], + }, + }, +} + +_TOOL_TEAM = { + "type": "function", + "function": { + "name": "get_team_usage_data", + "description": ( + "Fetch usage/spend data broken down by team. Use for questions " + "like 'which team spends the most' or 'show me team X usage'." + ), + "parameters": { + "type": "object", + "properties": { + **_DATE_PARAMS, + "team_ids": { + "type": "string", + "description": "Optional comma-separated team IDs. Omit for all teams.", + }, + }, + "required": ["start_date", "end_date"], + }, + }, +} + +_TOOL_TAG = { + "type": "function", + "function": { + "name": "get_tag_usage_data", + "description": ( + "Fetch usage/spend data broken down by tag. Tags are labels " + "attached to requests (features, environments, credentials)." + ), + "parameters": { + "type": "object", + "properties": { + **_DATE_PARAMS, + "tags": { + "type": "string", + "description": "Optional comma-separated tag names. Omit for all tags.", + }, + }, + "required": ["start_date", "end_date"], + }, + }, +} + +TOOLS_BASE = [_TOOL_USAGE] +TOOLS_ADMIN = [_TOOL_USAGE, _TOOL_TEAM, _TOOL_TAG] + + +def get_tools_for_role(is_admin: bool) -> List[Dict[str, Any]]: + """Return the tool list appropriate for the user's role.""" + return TOOLS_ADMIN if is_admin else TOOLS_BASE + + +_SYSTEM_PROMPT_BASE = ( + "You are an AI assistant embedded in the LiteLLM Usage dashboard. " + "You help users understand their LLM API spend and usage data.\n\n" + "ALWAYS call the appropriate tool(s) first to fetch data before answering. " + "You may call multiple tools if the question spans different dimensions.\n\n" + "Guidelines:\n" + "- Be concise and specific. Use exact numbers from the data.\n" + "- Format costs as dollar amounts (e.g. $12.34).\n" + "- When comparing entities, show a ranked list.\n" + "- If data is empty or no results found, say so clearly.\n" + "- Do not hallucinate data — only use what the tools return.\n" + "- Today's date will be provided below. Use it to interpret relative dates " + "like 'this week', 'this month', 'last 7 days', etc." +) + +_TOOL_DESCRIPTIONS_ADMIN = ( + "You have access to these tools:\n" + "- `get_usage_data`: Global/user-level usage (spend, models, providers, API keys)\n" + "- `get_team_usage_data`: Team-level usage breakdown\n" + "- `get_tag_usage_data`: Tag-level usage breakdown\n\n" +) + +_TOOL_DESCRIPTIONS_BASE = ( + "You have access to this tool:\n" + "- `get_usage_data`: Your usage data (spend, models, providers, API keys)\n\n" +) + + +def _build_system_prompt(is_admin: bool) -> str: + """Build role-appropriate system prompt with today's date.""" + tool_desc = _TOOL_DESCRIPTIONS_ADMIN if is_admin else _TOOL_DESCRIPTIONS_BASE + return ( + f"{_SYSTEM_PROMPT_BASE}\n\n{tool_desc}" + f"Today's date: {date.today().isoformat()}" + ) + + +# keep a public reference for test assertions +SYSTEM_PROMPT = _SYSTEM_PROMPT_BASE + +# --------------------------------------------------------------------------- +# Data fetchers +# --------------------------------------------------------------------------- + + +def _parse_csv_ids(raw: Optional[str]) -> Optional[List[str]]: + if not raw: + return None + return [t.strip() for t in raw.split(",") if t.strip()] + + +async def _query_activity( + table_name: str, + entity_id_field: str, + entity_id: Optional[Any], + start_date: str, + end_date: str, + *, + use_aggregated: bool = False, +) -> SpendAnalyticsPaginatedResponse: + """Shared helper that calls the daily activity query layer.""" + from litellm.proxy.management_endpoints.common_daily_activity import ( + get_daily_activity, + get_daily_activity_aggregated, + ) + from litellm.proxy.proxy_server import prisma_client + + if use_aggregated: + return await get_daily_activity_aggregated( + prisma_client=prisma_client, + table_name=table_name, + entity_id_field=entity_id_field, + entity_id=entity_id, + entity_metadata_field=None, + start_date=start_date, + end_date=end_date, + model=None, + api_key=None, + ) + return await get_daily_activity( + prisma_client=prisma_client, + table_name=table_name, + entity_id_field=entity_id_field, + entity_id=entity_id, + entity_metadata_field=None, + start_date=start_date, + end_date=end_date, + model=None, + api_key=None, + page=1, + page_size=PAGINATED_PAGE_SIZE, + ) + + +async def _fetch_usage_data( + start_date: str, end_date: str, user_id: Optional[str] = None +) -> Dict[str, Any]: + resp = await _query_activity( + TABLE_DAILY_USER_SPEND, + ENTITY_FIELD_USER, + user_id, + start_date, + end_date, + use_aggregated=True, + ) + return resp.model_dump(mode="json") + + +async def _fetch_team_usage_data( + start_date: str, end_date: str, team_ids: Optional[str] = None +) -> Dict[str, Any]: + resp = await _query_activity( + TABLE_DAILY_TEAM_SPEND, + ENTITY_FIELD_TEAM, + _parse_csv_ids(team_ids), + start_date, + end_date, + ) + return resp.model_dump(mode="json") + + +async def _fetch_tag_usage_data( + start_date: str, end_date: str, tags: Optional[str] = None +) -> Dict[str, Any]: + resp = await _query_activity( + TABLE_DAILY_TAG_SPEND, + ENTITY_FIELD_TAG, + _parse_csv_ids(tags), + start_date, + end_date, + ) + return resp.model_dump(mode="json") + + +# --------------------------------------------------------------------------- +# Summarisers — convert raw JSON to concise text the LLM can reason over +# --------------------------------------------------------------------------- + + +def _accumulate_breakdown( + results: List[Dict[str, Any]], dimension: str, fields: List[str] +) -> Dict[str, Dict[str, float]]: + """Aggregate a single breakdown dimension across days.""" + totals: Dict[str, Dict[str, float]] = {} + for day in results: + for key, entry in day.get("breakdown", {}).get(dimension, {}).items(): + if key not in totals: + totals[key] = {f: 0.0 for f in fields} + m = entry.get("metrics", {}) + for f in fields: + totals[key][f] += m.get(f, 0) + return totals + + +def _ranked_lines( + totals: Dict[str, Dict[str, float]], + fmt: Callable[[str, Dict[str, float]], str], + limit: int, +) -> List[str]: + """Sort by spend descending, format each entry, and truncate.""" + return [ + fmt(name, vals) + for name, vals in sorted(totals.items(), key=lambda x: -x[1].get("spend", 0))[ + :limit + ] + ] + + +def _summarise_usage_data(data: Dict[str, Any]) -> str: + meta = data.get("metadata", {}) + results = data.get("results", []) + + header = ( + f"Total Spend: ${meta.get('total_spend', 0):.4f}\n" + f"Total Requests: {meta.get('total_api_requests', 0)}\n" + f"Successful: {meta.get('total_successful_requests', 0)} | " + f"Failed: {meta.get('total_failed_requests', 0)}\n" + f"Total Tokens: {meta.get('total_tokens', 0)}" + ) + + models = _accumulate_breakdown( + results, "models", ["spend", "api_requests", "total_tokens"] + ) + providers = _accumulate_breakdown(results, "providers", ["spend", "api_requests"]) + + model_lines = _ranked_lines( + models, + lambda n, d: f" - {n}: ${d['spend']:.4f} ({int(d['api_requests'])} reqs, {int(d['total_tokens'])} tokens)", + TOP_N_MODELS, + ) + provider_lines = _ranked_lines( + providers, + lambda n, d: f" - {n}: ${d['spend']:.4f} ({int(d['api_requests'])} reqs)", + TOP_N_PROVIDERS, + ) + + sections = [header, ""] + sections += ["Top Models by Spend:"] + (model_lines or [" (no data)"]) + [""] + sections += ["Top Providers by Spend:"] + (provider_lines or [" (no data)"]) + return "\n".join(sections) + + +def _summarise_entity_data(data: Dict[str, Any], entity_label: str) -> str: + """Summarise team/tag entity usage data.""" + results = data.get("results", []) + if not results: + return f"No {entity_label} usage data found for the given date range." + + totals: Dict[str, Dict[str, Any]] = {} + for day in results: + for eid, entry in day.get("breakdown", {}).get("entities", {}).items(): + if eid not in totals: + alias = entry.get("metadata", {}).get("alias", eid) + totals[eid] = {"alias": alias, "spend": 0.0, "requests": 0, "tokens": 0} + m = entry.get("metrics", {}) + totals[eid]["spend"] += m.get("spend", 0) + totals[eid]["requests"] += m.get("api_requests", 0) + totals[eid]["tokens"] += m.get("total_tokens", 0) + + lines = [f"{entity_label} Usage ({len(totals)} {entity_label.lower()}s):", ""] + for eid, d in sorted(totals.items(), key=lambda x: -x[1]["spend"]): + label = d["alias"] if d["alias"] != eid else eid + lines.append( + f"- {label} (ID: {eid}): ${d['spend']:.4f} | " + f"{int(d['requests'])} reqs | {int(d['tokens'])} tokens" + ) + return "\n".join(lines) + + +# --------------------------------------------------------------------------- +# Tool dispatch registry +# --------------------------------------------------------------------------- + +TOOL_HANDLERS: Dict[str, ToolHandler] = { + "get_usage_data": ToolHandler( + fetch=_fetch_usage_data, + summarise=_summarise_usage_data, + label="global usage data", + ), + "get_team_usage_data": ToolHandler( + fetch=_fetch_team_usage_data, + summarise=lambda data: _summarise_entity_data(data, "Team"), + label="team usage data", + ), + "get_tag_usage_data": ToolHandler( + fetch=_fetch_tag_usage_data, + summarise=lambda data: _summarise_entity_data(data, "Tag"), + label="tag usage data", + ), +} + + +# --------------------------------------------------------------------------- +# SSE streaming +# --------------------------------------------------------------------------- + + +def _sse(event: SSEEvent) -> str: + return f"data: {json.dumps(event)}\n\n" + + +def _resolve_fetch_kwargs( + fn_name: str, + fn_args: Dict[str, str], + user_id: Optional[str], + is_admin: bool, +) -> Dict[str, Any]: + """Build keyword arguments for a tool's fetch function.""" + start_date = fn_args.get("start_date", "") + end_date = fn_args.get("end_date", "") + if not start_date or not end_date: + raise ValueError("Missing required start_date or end_date from tool arguments") + kwargs: Dict[str, Any] = {"start_date": start_date, "end_date": end_date} + if fn_name == "get_usage_data": + if not is_admin: + kwargs["user_id"] = user_id + elif fn_args.get("user_id"): + kwargs["user_id"] = fn_args["user_id"] + elif fn_name == "get_team_usage_data" and fn_args.get("team_ids"): + kwargs["team_ids"] = fn_args["team_ids"] + elif fn_name == "get_tag_usage_data" and fn_args.get("tags"): + kwargs["tags"] = fn_args["tags"] + return kwargs + + +async def _execute_tool_call( + handler: ToolHandler, + fn_name: str, + fn_args: Dict[str, str], + user_id: Optional[str], + is_admin: bool, +) -> str: + """Run a single tool and return the summarised result text.""" + kwargs = _resolve_fetch_kwargs(fn_name, fn_args, user_id, is_admin) + raw_data = await handler["fetch"](**kwargs) + return handler["summarise"](raw_data) + + +async def _process_tool_call( + tc: Any, + chat_messages: List[Dict[str, Any]], + user_id: Optional[str], + is_admin: bool, +) -> AsyncIterator[str]: + """Execute a single tool call, yielding SSE events for status.""" + fn_name = tc.function.name + fn_args = json.loads(tc.function.arguments) + + allowed_names = {t["function"]["name"] for t in get_tools_for_role(is_admin)} + handler = TOOL_HANDLERS.get(fn_name) + + if fn_name not in allowed_names or not handler: + chat_messages.append( + { + "role": "tool", + "tool_call_id": tc.id, + "content": f"Tool not available: {fn_name}", + } + ) + return + + tool_event_base = { + "type": "tool_call", + "tool_name": fn_name, + "tool_label": handler["label"], + "arguments": fn_args, + } + yield _sse({**tool_event_base, "status": "running"}) + + try: + tool_result = await _execute_tool_call( + handler, fn_name, fn_args, user_id, is_admin + ) + yield _sse({**tool_event_base, "status": "complete"}) + except Exception as e: + verbose_proxy_logger.error("Tool %s failed: %s", fn_name, e) + tool_result = f"Error fetching {handler['label']}. Please try again." + yield _sse({**tool_event_base, "status": "error"}) + + chat_messages.append( + {"role": "tool", "tool_call_id": tc.id, "content": tool_result} + ) + + +async def _stream_final_response( + model: str, chat_messages: List[Dict[str, Any]] +) -> AsyncIterator[str]: + """Stream the final LLM response after tool results are appended.""" + yield _sse({"type": "status", "message": "Analyzing results..."}) + + response = await litellm.acompletion( + model=model, + messages=chat_messages, + stream=True, + temperature=USAGE_AI_TEMPERATURE, + ) + async for chunk in response: + delta = chunk.choices[0].delta.content + if delta: + yield _sse({"type": "chunk", "content": delta}) + + +async def stream_usage_ai_chat( + messages: List[Dict[str, str]], + model: Optional[str] = None, + user_id: Optional[str] = None, + is_admin: bool = False, +) -> AsyncIterator[str]: + """Stream SSE events: status → tool_call → chunk → done.""" + resolved_model = (model or "").strip() or DEFAULT_COMPETITOR_DISCOVERY_MODEL + truncated = ( + messages[-MAX_CHAT_MESSAGES:] if len(messages) > MAX_CHAT_MESSAGES else messages + ) + chat_messages: List[Dict[str, Any]] = [ + {"role": "system", "content": _build_system_prompt(is_admin)}, + *truncated, + ] + + try: + yield _sse({"type": "status", "message": "Thinking..."}) + tools = get_tools_for_role(is_admin) + response = await litellm.acompletion( + model=resolved_model, + messages=chat_messages, + tools=tools, + temperature=USAGE_AI_TEMPERATURE, + ) + choice = response.choices[0] # type: ignore + + if not choice.message.tool_calls: + if choice.message.content: + yield _sse({"type": "chunk", "content": choice.message.content}) + yield _sse({"type": "done"}) + return + + chat_messages.append(choice.message.model_dump()) + for tc in choice.message.tool_calls: + async for event in _process_tool_call(tc, chat_messages, user_id, is_admin): + yield event + async for event in _stream_final_response(resolved_model, chat_messages): + yield event + yield _sse({"type": "done"}) + + except Exception as e: + verbose_proxy_logger.error("AI usage chat failed: %s", e) + yield _sse( + { + "type": "error", + "message": "An internal error occurred. Please try again.", + } + ) diff --git a/litellm/proxy/management_endpoints/usage_endpoints/endpoints.py b/litellm/proxy/management_endpoints/usage_endpoints/endpoints.py new file mode 100644 index 00000000000..0dbe518afb7 --- /dev/null +++ b/litellm/proxy/management_endpoints/usage_endpoints/endpoints.py @@ -0,0 +1,65 @@ +""" +USAGE AI CHAT ENDPOINTS + +/usage/ai/chat - Stream AI chat responses about usage data +""" + +from typing import List, Literal, Optional + +from fastapi import APIRouter, Depends, Request +from fastapi.responses import StreamingResponse +from pydantic import BaseModel, Field + +from litellm.proxy._types import UserAPIKeyAuth +from litellm.proxy.auth.user_api_key_auth import user_api_key_auth + +router = APIRouter() + + +class ChatMessage(BaseModel): + role: Literal["user", "assistant"] + content: str + + +class UsageAIChatRequest(BaseModel): + messages: List[ChatMessage] = Field( + ..., description="Chat messages (user/assistant history)" + ) + model: Optional[str] = Field(default=None, description="Model to use for AI chat") + + +@router.post( + "/usage/ai/chat", + tags=["Budget & Spend Tracking"], + dependencies=[Depends(user_api_key_auth)], +) +async def usage_ai_chat( + data: UsageAIChatRequest, + request: Request, + user_api_key_dict: UserAPIKeyAuth = Depends(user_api_key_auth), +): + """ + AI chat about usage data. Streams SSE events with the AI response. + The AI agent has access to tools that query aggregated daily activity data. + """ + from litellm.proxy.management_endpoints.common_utils import ( + _user_has_admin_view, + ) + from litellm.proxy.management_endpoints.usage_endpoints.ai_usage_chat import ( + stream_usage_ai_chat, + ) + + is_admin = _user_has_admin_view(user_api_key_dict) + user_id = user_api_key_dict.user_id + messages = [{"role": m.role, "content": m.content} for m in data.messages] + + return StreamingResponse( + stream_usage_ai_chat( + messages=messages, + model=data.model, + user_id=user_id, + is_admin=is_admin, + ), + media_type="text/event-stream", + headers={"Cache-Control": "no-cache", "X-Accel-Buffering": "no"}, + ) diff --git a/litellm/proxy/proxy_server.py b/litellm/proxy/proxy_server.py index 36b6bcc7707..607306f3806 100644 --- a/litellm/proxy/proxy_server.py +++ b/litellm/proxy/proxy_server.py @@ -392,6 +392,7 @@ from litellm.proxy.management_endpoints.organization_endpoints import ( router as organization_router, ) from litellm.proxy.management_endpoints.policy_endpoints import router as policy_router +from litellm.proxy.management_endpoints.usage_endpoints import router as usage_ai_router from litellm.proxy.management_endpoints.project_endpoints import ( router as project_router, ) @@ -12872,6 +12873,7 @@ app.include_router(caching_router) app.include_router(analytics_router) app.include_router(guardrails_router) app.include_router(policy_router) +app.include_router(usage_ai_router) app.include_router(policy_crud_router) app.include_router(policy_resolve_router) app.include_router(search_tool_management_router) diff --git a/tests/test_litellm/proxy/management_endpoints/usage_endpoints/__init__.py b/tests/test_litellm/proxy/management_endpoints/usage_endpoints/__init__.py new file mode 100644 index 00000000000..e69de29bb2d diff --git a/tests/test_litellm/proxy/management_endpoints/usage_endpoints/test_ai_usage_chat.py b/tests/test_litellm/proxy/management_endpoints/usage_endpoints/test_ai_usage_chat.py new file mode 100644 index 00000000000..f9303bd13a6 --- /dev/null +++ b/tests/test_litellm/proxy/management_endpoints/usage_endpoints/test_ai_usage_chat.py @@ -0,0 +1,402 @@ +""" +Tests for AI Usage Chat module. +""" + +import json +from unittest.mock import AsyncMock, MagicMock, patch + +import pytest + +from litellm.proxy.management_endpoints.usage_endpoints.ai_usage_chat import ( + TOOL_HANDLERS, + TOOLS_ADMIN, + TOOLS_BASE, + _build_system_prompt, + _summarise_entity_data, + _summarise_usage_data, + stream_usage_ai_chat, +) + + +SAMPLE_AGGREGATED_RESPONSE = { + "results": [ + { + "date": "2025-01-15", + "metrics": { + "spend": 50.25, + "prompt_tokens": 20000, + "completion_tokens": 10000, + "total_tokens": 30000, + "api_requests": 500, + "successful_requests": 480, + "failed_requests": 20, + "cache_read_input_tokens": 0, + "cache_creation_input_tokens": 0, + }, + "breakdown": { + "models": { + "gpt-4": { + "metrics": { + "spend": 40.0, + "api_requests": 300, + "total_tokens": 25000, + }, + "metadata": {}, + "api_key_breakdown": {}, + }, + }, + "providers": { + "openai": { + "metrics": {"spend": 50.25, "api_requests": 500}, + "metadata": {}, + "api_key_breakdown": {}, + }, + }, + "api_keys": { + "sk-test123": { + "metrics": {"spend": 50.25}, + "metadata": {"key_alias": "Production Key"}, + }, + }, + "model_groups": {}, + "mcp_servers": {}, + "entities": {}, + }, + }, + ], + "metadata": { + "total_spend": 50.25, + "total_api_requests": 500, + "total_successful_requests": 480, + "total_failed_requests": 20, + "total_tokens": 30000, + }, +} + +SAMPLE_TEAM_RESPONSE = { + "results": [ + { + "date": "2025-01-15", + "metrics": {"spend": 100.0, "api_requests": 1000, "total_tokens": 50000}, + "breakdown": { + "entities": { + "team-1": { + "metrics": { + "spend": 60.0, + "api_requests": 600, + "total_tokens": 30000, + }, + "metadata": {"alias": "Engineering"}, + "api_key_breakdown": {}, + }, + "team-2": { + "metrics": { + "spend": 40.0, + "api_requests": 400, + "total_tokens": 20000, + }, + "metadata": {"alias": "Marketing"}, + "api_key_breakdown": {}, + }, + }, + "models": {}, + "providers": {}, + "api_keys": {}, + "model_groups": {}, + "mcp_servers": {}, + }, + }, + ], + "metadata": {"total_spend": 100.0, "total_api_requests": 1000}, +} + + +class TestToolSchemas: + def test_admin_tools_include_all(self): + assert len(TOOLS_ADMIN) == 3 + names = {t["function"]["name"] for t in TOOLS_ADMIN} + assert "get_usage_data" in names + assert "get_team_usage_data" in names + assert "get_tag_usage_data" in names + + def test_base_tools_restricted_to_usage_only(self): + assert len(TOOLS_BASE) == 1 + assert TOOLS_BASE[0]["function"]["name"] == "get_usage_data" + + def test_admin_prompt_mentions_all_tools(self): + prompt = _build_system_prompt(is_admin=True) + assert "get_usage_data" in prompt + assert "get_team_usage_data" in prompt + assert "get_tag_usage_data" in prompt + + def test_non_admin_prompt_only_mentions_usage_tool(self): + prompt = _build_system_prompt(is_admin=False) + assert "get_usage_data" in prompt + assert "get_team_usage_data" not in prompt + assert "get_tag_usage_data" not in prompt + + def test_system_prompt_includes_todays_date(self): + from datetime import date + + prompt = _build_system_prompt(is_admin=True) + assert date.today().isoformat() in prompt + + +class TestSummariseUsageData: + def test_summarise_includes_totals(self): + summary = _summarise_usage_data(SAMPLE_AGGREGATED_RESPONSE) + assert "$50.25" in summary + assert "500" in summary + + def test_summarise_includes_models(self): + summary = _summarise_usage_data(SAMPLE_AGGREGATED_RESPONSE) + assert "gpt-4" in summary + + def test_summarise_includes_providers(self): + summary = _summarise_usage_data(SAMPLE_AGGREGATED_RESPONSE) + assert "openai" in summary + + def test_summarise_handles_empty_data(self): + empty = {"results": [], "metadata": {}} + summary = _summarise_usage_data(empty) + assert "no data" in summary.lower() + + +class TestSummariseEntityData: + def test_team_summary_includes_teams(self): + summary = _summarise_entity_data(SAMPLE_TEAM_RESPONSE, "Team") + assert "Engineering" in summary + assert "Marketing" in summary + assert "$60.0" in summary + assert "$40.0" in summary + + def test_team_summary_empty(self): + empty = {"results": [], "metadata": {}} + summary = _summarise_entity_data(empty, "Team") + assert "No Team usage data" in summary + + +class TestStreamUsageAiChat: + @pytest.mark.asyncio + async def test_stream_emits_status_events(self): + mock_tool_call = MagicMock() + mock_tool_call.id = "call_123" + mock_tool_call.function.name = "get_usage_data" + mock_tool_call.function.arguments = json.dumps( + { + "start_date": "2025-01-01", + "end_date": "2025-01-31", + } + ) + + mock_first_response = MagicMock() + mock_first_response.choices = [MagicMock()] + mock_first_response.choices[0].message.tool_calls = [mock_tool_call] + mock_first_response.choices[0].message.model_dump.return_value = { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": "call_123", + "type": "function", + "function": { + "name": "get_usage_data", + "arguments": '{"start_date":"2025-01-01","end_date":"2025-01-31"}', + }, + } + ], + } + + async def mock_stream(): + chunk = MagicMock() + chunk.choices = [MagicMock()] + chunk.choices[0].delta.content = "Total spend is $50.25" + yield chunk + + with patch( + "litellm.proxy.management_endpoints.usage_endpoints.ai_usage_chat.litellm" + ) as mock_litellm, patch( + "litellm.proxy.management_endpoints.usage_endpoints.ai_usage_chat._fetch_usage_data", + new_callable=AsyncMock, + ) as mock_fetch: + mock_litellm.acompletion = AsyncMock( + side_effect=[ + mock_first_response, + mock_stream(), + ] + ) + mock_fetch.return_value = SAMPLE_AGGREGATED_RESPONSE + + events = [] + async for event in stream_usage_ai_chat( + messages=[{"role": "user", "content": "What is my total spend?"}], + model="gpt-4o-mini", + user_id="user-123", + is_admin=True, + ): + events.append(json.loads(event.replace("data: ", "").strip())) + + status_events = [e for e in events if e["type"] == "status"] + tool_call_events = [e for e in events if e["type"] == "tool_call"] + chunk_events = [e for e in events if e["type"] == "chunk"] + done_events = [e for e in events if e["type"] == "done"] + + assert len(status_events) >= 1 + assert "Thinking" in status_events[0]["message"] + assert len(tool_call_events) >= 1 + assert tool_call_events[0]["tool_name"] == "get_usage_data" + assert tool_call_events[0]["status"] in ("running", "complete") + assert len(chunk_events) >= 1 + assert len(done_events) == 1 + + @pytest.mark.asyncio + async def test_stream_handles_team_tool(self): + mock_tool_call = MagicMock() + mock_tool_call.id = "call_team" + mock_tool_call.function.name = "get_team_usage_data" + mock_tool_call.function.arguments = json.dumps( + { + "start_date": "2025-01-01", + "end_date": "2025-01-31", + } + ) + + mock_first_response = MagicMock() + mock_first_response.choices = [MagicMock()] + mock_first_response.choices[0].message.tool_calls = [mock_tool_call] + mock_first_response.choices[0].message.model_dump.return_value = { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": "call_team", + "type": "function", + "function": { + "name": "get_team_usage_data", + "arguments": '{"start_date":"2025-01-01","end_date":"2025-01-31"}', + }, + } + ], + } + + async def mock_stream(): + chunk = MagicMock() + chunk.choices = [MagicMock()] + chunk.choices[0].delta.content = "Engineering is the top team." + yield chunk + + with patch( + "litellm.proxy.management_endpoints.usage_endpoints.ai_usage_chat.litellm" + ) as mock_litellm, patch( + "litellm.proxy.management_endpoints.usage_endpoints.ai_usage_chat._fetch_team_usage_data", + new_callable=AsyncMock, + ) as mock_fetch: + mock_litellm.acompletion = AsyncMock( + side_effect=[ + mock_first_response, + mock_stream(), + ] + ) + mock_fetch.return_value = SAMPLE_TEAM_RESPONSE + + events = [] + async for event in stream_usage_ai_chat( + messages=[{"role": "user", "content": "Which team spends the most?"}], + model="gpt-4o-mini", + is_admin=True, + ): + events.append(json.loads(event.replace("data: ", "").strip())) + + chunk_events = [e for e in events if e["type"] == "chunk"] + assert len(chunk_events) >= 1 + assert "Engineering" in chunk_events[0]["content"] + + @pytest.mark.asyncio + async def test_stream_handles_error(self): + with patch( + "litellm.proxy.management_endpoints.usage_endpoints.ai_usage_chat.litellm" + ) as mock_litellm: + mock_litellm.acompletion = AsyncMock(side_effect=Exception("LLM error")) + + events = [] + async for event in stream_usage_ai_chat( + messages=[{"role": "user", "content": "test"}], + ): + events.append(json.loads(event.replace("data: ", "").strip())) + + error_events = [e for e in events if e["type"] == "error"] + assert len(error_events) == 1 + assert "internal error" in error_events[0]["message"].lower() + + @pytest.mark.asyncio + async def test_non_admin_enforces_user_id(self): + mock_tool_call = MagicMock() + mock_tool_call.id = "call_456" + mock_tool_call.function.name = "get_usage_data" + mock_tool_call.function.arguments = json.dumps( + { + "start_date": "2025-01-01", + "end_date": "2025-01-31", + "user_id": "other-user", + } + ) + + mock_first_response = MagicMock() + mock_first_response.choices = [MagicMock()] + mock_first_response.choices[0].message.tool_calls = [mock_tool_call] + mock_first_response.choices[0].message.model_dump.return_value = { + "role": "assistant", + "content": None, + "tool_calls": [ + { + "id": "call_456", + "type": "function", + "function": { + "name": "get_usage_data", + "arguments": '{"start_date":"2025-01-01","end_date":"2025-01-31","user_id":"other-user"}', + }, + } + ], + } + + async def mock_stream(): + chunk = MagicMock() + chunk.choices = [MagicMock()] + chunk.choices[0].delta.content = "Data." + yield chunk + + mock_fetch = AsyncMock(return_value=SAMPLE_AGGREGATED_RESPONSE) + + with patch( + "litellm.proxy.management_endpoints.usage_endpoints.ai_usage_chat.litellm" + ) as mock_litellm, patch.dict( + "litellm.proxy.management_endpoints.usage_endpoints.ai_usage_chat.TOOL_HANDLERS", + { + "get_usage_data": { + "fetch": mock_fetch, + "summarise": _summarise_usage_data, + "label": "global usage data", + } + }, + ): + mock_litellm.acompletion = AsyncMock( + side_effect=[ + mock_first_response, + mock_stream(), + ] + ) + + events = [] + async for event in stream_usage_ai_chat( + messages=[{"role": "user", "content": "Show data"}], + model="gpt-4o-mini", + user_id="my-user-id", + is_admin=False, + ): + events.append(event) + + mock_fetch.assert_called_once_with( + start_date="2025-01-01", + end_date="2025-01-31", + user_id="my-user-id", + ) diff --git a/ui/litellm-dashboard/package-lock.json b/ui/litellm-dashboard/package-lock.json index 489a39a7ee2..3787f451ad3 100644 --- a/ui/litellm-dashboard/package-lock.json +++ b/ui/litellm-dashboard/package-lock.json @@ -1757,29 +1757,6 @@ "url": "https://opencollective.com/libvips" } }, - "node_modules/@isaacs/balanced-match": { - "version": "4.0.1", - "resolved": "https://registry.npmjs.org/@isaacs/balanced-match/-/balanced-match-4.0.1.tgz", - "integrity": "sha512-yzMTt9lEb8Gv7zRioUilSglI0c0smZ9k5D65677DLWLtWJaXIS3CqcGyUFByYKlnUj6TkjLVs54fBl6+TiGQDQ==", - "dev": true, - "license": "MIT", - "engines": { - "node": "20 || >=22" - } - }, - "node_modules/@isaacs/brace-expansion": { - "version": "5.0.1", - "resolved": "https://registry.npmjs.org/@isaacs/brace-expansion/-/brace-expansion-5.0.1.tgz", - "integrity": "sha512-WMz71T1JS624nWj2n2fnYAuPovhv7EUhk69R6i9dsVyzxt5eM3bjwvgk9L+APE1TRscGysAVMANkB0jh0LQZrQ==", - "dev": true, - "license": "MIT", - "dependencies": { - "@isaacs/balanced-match": "^4.0.1" - }, - "engines": { - "node": "20 || >=22" - } - }, "node_modules/@istanbuljs/schema": { "version": "0.1.3", "resolved": "https://registry.npmjs.org/@istanbuljs/schema/-/schema-0.1.3.tgz", @@ -3696,32 +3673,6 @@ "typescript": ">=4.8.4 <6.0.0" } }, - "node_modules/@typescript-eslint/typescript-estree/node_modules/brace-expansion": { - "version": "2.0.2", - "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.0.2.tgz", - "integrity": "sha512-Jt0vHyM+jmUBqojB7E1NIYadt0vI0Qxjxd2TErW94wDz+E2LAm5vKMXXwg6ZZBTHPuUlDgQHKXvjGBdfcF1ZDQ==", - "dev": true, - "license": "MIT", - "dependencies": { - "balanced-match": "^1.0.0" - } - }, - "node_modules/@typescript-eslint/typescript-estree/node_modules/minimatch": { - "version": "9.0.5", - "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-9.0.5.tgz", - "integrity": "sha512-G6T0ZX48xgozx7587koeX9Ys2NYy6Gmv//P89sEte9V9whIapMNF4idKxnW2QtCcLiTWlb/wfCabAtAFWhhBow==", - "dev": true, - "license": "ISC", - "dependencies": { - "brace-expansion": "^2.0.1" - }, - "engines": { - "node": ">=16 || 14 >=14.17" - }, - "funding": { - "url": "https://github.com/sponsors/isaacs" - } - }, "node_modules/@typescript-eslint/utils": { "version": "8.54.0", "resolved": "https://registry.npmjs.org/@typescript-eslint/utils/-/utils-8.54.0.tgz", @@ -4749,11 +4700,14 @@ } }, "node_modules/balanced-match": { - "version": "1.0.2", - "resolved": "https://registry.npmjs.org/balanced-match/-/balanced-match-1.0.2.tgz", - "integrity": "sha512-3oSeUO0TMV67hN1AmbXsK4yaqU7tjiHlbxRDZOpH0KW9+CeX4bRAaX0Anxt0tx2MrpRpWwQaPwIlISEJhYU5Pw==", + "version": "4.0.4", + "resolved": "https://registry.npmjs.org/balanced-match/-/balanced-match-4.0.4.tgz", + "integrity": "sha512-BLrgEcRTwX2o6gGxGOCNyMvGSp35YofuYzw9h1IMTRmKqttAZZVU67bdb9Pr2vUHA8+j3i2tJfjO6C6+4myGTA==", "dev": true, - "license": "MIT" + "license": "MIT", + "engines": { + "node": "18 || 20 || >=22" + } }, "node_modules/baseline-browser-mapping": { "version": "2.9.19", @@ -4787,14 +4741,16 @@ } }, "node_modules/brace-expansion": { - "version": "1.1.12", - "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-1.1.12.tgz", - "integrity": "sha512-9T9UjW3r0UW5c1Q7GTwllptXwhvYmEzFhzMfZ9H7FQWt+uZePjZPjBP/W1ZEyZ1twGWom5/56TF4lPcqjnDHcg==", + "version": "5.0.3", + "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-5.0.3.tgz", + "integrity": "sha512-fy6KJm2RawA5RcHkLa1z/ScpBeA762UF9KmZQxwIbDtRJrgLzM10depAiEQ+CXYcoiqW1/m96OAAoke2nE9EeA==", "dev": true, "license": "MIT", "dependencies": { - "balanced-match": "^1.0.0", - "concat-map": "0.0.1" + "balanced-match": "^4.0.2" + }, + "engines": { + "node": "18 || 20 || >=22" } }, "node_modules/braces": { @@ -5149,13 +5105,6 @@ "integrity": "sha512-VRhuHOLoKYOy4UbilLbUzbYg93XLjv2PncJC50EuTWPA3gaja1UjBsUP/D/9/juV3vQFr6XBEzn9KCAHdUvOHw==", "license": "MIT" }, - "node_modules/concat-map": { - "version": "0.0.1", - "resolved": "https://registry.npmjs.org/concat-map/-/concat-map-0.0.1.tgz", - "integrity": "sha512-/Srv4dswyQNBfohGpz9o6Yb3Gz3SrUDqBH5rTuhGR7ahtlbYKnVxw2bCFMRljaA7EXHaXZ8wsHdodFvbkhKmqg==", - "dev": true, - "license": "MIT" - }, "node_modules/copy-to-clipboard": { "version": "3.3.3", "resolved": "https://registry.npmjs.org/copy-to-clipboard/-/copy-to-clipboard-3.3.3.tgz", @@ -6924,22 +6873,6 @@ "node": ">=10.13.0" } }, - "node_modules/glob/node_modules/minimatch": { - "version": "10.1.1", - "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-10.1.1.tgz", - "integrity": "sha512-enIvLvRAFZYXJzkCYG5RKmPfrFArdLv+R+lbQ53BmIMLIry74bjKzX6iHAm8WYamJkhSSEabrWN5D97XnKObjQ==", - "dev": true, - "license": "BlueOak-1.0.0", - "dependencies": { - "@isaacs/brace-expansion": "^5.0.0" - }, - "engines": { - "node": "20 || >=22" - }, - "funding": { - "url": "https://github.com/sponsors/isaacs" - } - }, "node_modules/globals": { "version": "14.0.0", "resolved": "https://registry.npmjs.org/globals/-/globals-14.0.0.tgz", @@ -9035,16 +8968,19 @@ } }, "node_modules/minimatch": { - "version": "3.1.2", - "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-3.1.2.tgz", - "integrity": "sha512-J7p63hRiAjw1NDEww1W7i37+ByIrOWO5XQQAzZ3VOcL0PNybwpfmV/N05zFAzwQ9USyEcX6t3UO+K5aqBQOIHw==", + "version": "10.2.2", + "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-10.2.2.tgz", + "integrity": "sha512-+G4CpNBxa5MprY+04MbgOw1v7So6n5JY166pFi9KfYwT78fxScCeSNQSNzp6dpPSW2rONOps6Ocam1wFhCgoVw==", "dev": true, - "license": "ISC", + "license": "BlueOak-1.0.0", "dependencies": { - "brace-expansion": "^1.1.7" + "brace-expansion": "^5.0.2" }, "engines": { - "node": "*" + "node": "18 || 20 || >=22" + }, + "funding": { + "url": "https://github.com/sponsors/isaacs" } }, "node_modules/minimist": { @@ -12004,32 +11940,6 @@ "node": ">=18" } }, - "node_modules/test-exclude/node_modules/brace-expansion": { - "version": "2.0.2", - "resolved": "https://registry.npmjs.org/brace-expansion/-/brace-expansion-2.0.2.tgz", - "integrity": "sha512-Jt0vHyM+jmUBqojB7E1NIYadt0vI0Qxjxd2TErW94wDz+E2LAm5vKMXXwg6ZZBTHPuUlDgQHKXvjGBdfcF1ZDQ==", - "dev": true, - "license": "MIT", - "dependencies": { - "balanced-match": "^1.0.0" - } - }, - "node_modules/test-exclude/node_modules/minimatch": { - "version": "9.0.5", - "resolved": "https://registry.npmjs.org/minimatch/-/minimatch-9.0.5.tgz", - "integrity": "sha512-G6T0ZX48xgozx7587koeX9Ys2NYy6Gmv//P89sEte9V9whIapMNF4idKxnW2QtCcLiTWlb/wfCabAtAFWhhBow==", - "dev": true, - "license": "ISC", - "dependencies": { - "brace-expansion": "^2.0.1" - }, - "engines": { - "node": ">=16 || 14 >=14.17" - }, - "funding": { - "url": "https://github.com/sponsors/isaacs" - } - }, "node_modules/thenify": { "version": "3.3.1", "resolved": "https://registry.npmjs.org/thenify/-/thenify-3.3.1.tgz", @@ -13085,21 +12995,6 @@ "type": "github", "url": "https://github.com/sponsors/wooorm" } - }, - "node_modules/@next/swc-win32-ia32-msvc": { - "version": "14.2.33", - "resolved": "https://registry.npmjs.org/@next/swc-win32-ia32-msvc/-/swc-win32-ia32-msvc-14.2.33.tgz", - "integrity": "sha512-pc9LpGNKhJ0dXQhZ5QMmYxtARwwmWLpeocFmVG5Z0DzWq5Uf0izcI8tLc+qOpqxO1PWqZ5A7J1blrUIKrIFc7Q==", - "cpu": [ - "ia32" - ], - "optional": true, - "os": [ - "win32" - ], - "engines": { - "node": ">= 10" - } } } } diff --git a/ui/litellm-dashboard/src/components/UsagePage/components/UsageAIChatPanel.test.tsx b/ui/litellm-dashboard/src/components/UsagePage/components/UsageAIChatPanel.test.tsx new file mode 100644 index 00000000000..85ce11e605a --- /dev/null +++ b/ui/litellm-dashboard/src/components/UsagePage/components/UsageAIChatPanel.test.tsx @@ -0,0 +1,85 @@ +import { screen } from "@testing-library/react"; +import { beforeAll, describe, expect, it, vi } from "vitest"; +import { renderWithProviders } from "../../../../tests/test-utils"; +import UsageAIChatPanel from "./UsageAIChatPanel"; + +beforeAll(() => { + if (typeof window !== "undefined" && !window.ResizeObserver) { + window.ResizeObserver = class ResizeObserver { + observe() {} + unobserve() {} + disconnect() {} + } as any; + } +}); + +vi.mock("../../networking", () => ({ + modelHubCall: vi.fn().mockResolvedValue({ + data: [ + { model_group: "gpt-4" }, + { model_group: "claude-3-opus" }, + ], + }), + usageAiChatStream: vi.fn(), +})); + +const defaultProps = { + open: true, + onClose: vi.fn(), + accessToken: "test-token", +}; + +describe("UsageAIChatPanel", () => { + it("should render the panel when open", () => { + renderWithProviders(); + + expect(screen.getByText("Ask AI")).toBeInTheDocument(); + expect( + screen.getByText("Ask about your spend, models, keys, and trends") + ).toBeInTheDocument(); + }); + + it("should render model selector", () => { + renderWithProviders(); + + expect(screen.getByText("Select a model (optional, defaults to gpt-4o-mini)")).toBeInTheDocument(); + }); + + it("should render empty state message when no conversation", () => { + renderWithProviders(); + + expect(screen.getByText("Ask a question about your usage")).toBeInTheDocument(); + }); + + it("should render the send button", () => { + renderWithProviders(); + + expect(screen.getByText("Send")).toBeInTheDocument(); + }); + + it("should render input placeholder", () => { + renderWithProviders(); + + expect(screen.getByPlaceholderText("Ask about your usage...")).toBeInTheDocument(); + }); + + it("should render clear chat button", () => { + renderWithProviders(); + + expect(screen.getByText("Clear chat")).toBeInTheDocument(); + }); + + it("should have the panel element even when closed (just off-screen)", () => { + renderWithProviders(); + + expect(screen.getByTestId("usage-ai-chat-panel")).toBeInTheDocument(); + expect(screen.getByTestId("usage-ai-chat-panel")).toHaveClass("translate-x-full"); + }); + + it("should not have translate-x-full class when open", () => { + renderWithProviders(); + + expect(screen.getByTestId("usage-ai-chat-panel")).not.toHaveClass("translate-x-full"); + expect(screen.getByTestId("usage-ai-chat-panel")).toHaveClass("translate-x-0"); + }); +}); diff --git a/ui/litellm-dashboard/src/components/UsagePage/components/UsageAIChatPanel.tsx b/ui/litellm-dashboard/src/components/UsagePage/components/UsageAIChatPanel.tsx new file mode 100644 index 00000000000..85f46bfa346 --- /dev/null +++ b/ui/litellm-dashboard/src/components/UsagePage/components/UsageAIChatPanel.tsx @@ -0,0 +1,402 @@ +import React, { useEffect, useRef, useState } from "react"; +import { Button, Select, Input, Spin } from "antd"; +import ReactMarkdown from "react-markdown"; +import { modelHubCall, usageAiChatStream, UsageAiToolCallEvent } from "../../networking"; + +const { TextArea } = Input; + +interface ToolCallStep { + tool_name: string; + tool_label: string; + arguments: Record; + status: "running" | "complete" | "error"; + error?: string; +} + +interface ChatMessage { + role: "user" | "assistant"; + content: string; + toolCalls?: ToolCallStep[]; +} + +interface UsageAIChatPanelProps { + open: boolean; + onClose: () => void; + accessToken: string | null; +} + +const TOOL_ICONS: Record = { + get_usage_data: "📊", + get_team_usage_data: "👥", + get_tag_usage_data: "🏷️", +}; + +const ToolCallDisplay: React.FC<{ step: ToolCallStep }> = ({ step }) => { + const icon = TOOL_ICONS[step.tool_name] || "🔧"; + const args = step.arguments; + const dateRange = args.start_date && args.end_date + ? `${args.start_date} → ${args.end_date}` + : ""; + const filter = args.team_ids || args.tags || args.user_id || ""; + + return ( +
+ + {step.status === "running" ? ( + + ) : step.status === "error" ? ( + ✗ + ) : ( + ✓ + )} + +
+
+ {icon} {step.tool_label} +
+ {dateRange && ( +
{dateRange}
+ )} + {filter && ( +
Filter: {filter}
+ )} + {step.status === "error" && step.error && ( +
{step.error}
+ )} +
+
+ ); +}; + +const MarkdownContent: React.FC<{ content: string }> = ({ content }) => ( +

{children}

, + strong: ({ children }) => {children}, + ul: ({ children }) =>
    {children}
, + ol: ({ children }) =>
    {children}
, + li: ({ children }) =>
  • {children}
    • , + h1: ({ children }) =>

    {children}

    , + h2: ({ children }) =>

    {children}

    , + h3: ({ children }) =>

    {children}

    , + code: ({ children, className }) => { + const isBlock = className?.includes("language-"); + return isBlock ? ( + 
    +            {children}
+
    + ) : ( + {children} + ); + }, + table: ({ children }) => ( +
    + {children}
    +
    + ), + th: ({ children }) => {children}, + td: ({ children }) => {children}, + }} + > + {content} +     +); + +const UsageAIChatPanel: React.FC = ({ + open, + onClose, + accessToken, +}) => { + const [messages, setMessages] = useState([]); + const [inputText, setInputText] = useState(""); + const [isLoading, setIsLoading] = useState(false); + const [selectedModel, setSelectedModel] = useState(undefined); + const [availableModels, setAvailableModels] = useState([]); + const [isLoadingModels, setIsLoadingModels] = useState(false); + const [streamingContent, setStreamingContent] = useState(""); + const [statusMessage, setStatusMessage] = useState(null); + const [activeToolCalls, setActiveToolCalls] = useState([]); + const messagesEndRef = useRef(null); + const abortControllerRef = useRef(null); + + useEffect(() => { + if (open && availableModels.length === 0) { + loadModels(); + } + }, [open]); + + useEffect(() => { + if (typeof messagesEndRef.current?.scrollIntoView === "function") { + messagesEndRef.current.scrollIntoView({ behavior: "smooth" }); + } + }, [messages, streamingContent, activeToolCalls, statusMessage]); + + const loadModels = async () => { + if (!accessToken) return; + setIsLoadingModels(true); + try { + const fetchedModels = await modelHubCall(accessToken); + if (fetchedModels?.data?.length > 0) { + const models = fetchedModels.data + .map((item: any) => item.model_group as string) + .sort(); + setAvailableModels(models); + } + } catch (error) { + console.error("Failed to load models:", error); + } finally { + setIsLoadingModels(false); + } + }; + + const handleSend = async () => { + if (!accessToken || !inputText.trim() || isLoading) return; + + const userMessage: ChatMessage = { role: "user", content: inputText.trim() }; + const updatedMessages = [...messages, userMessage]; + setMessages(updatedMessages); + setInputText(""); + setIsLoading(true); + setStreamingContent(""); + setStatusMessage(null); + setActiveToolCalls([]); + + const abortController = new AbortController(); + abortControllerRef.current = abortController; + + let accumulated = ""; + const toolCalls: ToolCallStep[] = []; + + try { + await usageAiChatStream( + accessToken, + updatedMessages.slice(-20).map((m) => ({ role: m.role, content: m.content })), + selectedModel || "", + (content: string) => { + setStatusMessage(null); + accumulated += content; + setStreamingContent(accumulated); + }, + () => { + setStatusMessage(null); + setActiveToolCalls([]); + setMessages((prev) => [ + ...prev, + { role: "assistant", content: accumulated, toolCalls: toolCalls.length > 0 ? [...toolCalls] : undefined }, + ]); + setStreamingContent(""); + }, + (errorMsg: string) => { + setStatusMessage(null); + setActiveToolCalls([]); + setMessages((prev) => [ + ...prev, + { role: "assistant", content: `Error: ${errorMsg}` }, + ]); + setStreamingContent(""); + }, + (status: string) => { + setStatusMessage(status); + }, + (event: UsageAiToolCallEvent) => { + const idx = toolCalls.findIndex((tc) => tc.tool_name === event.tool_name); + if (idx >= 0) { + toolCalls[idx] = { ...event }; + } else { + toolCalls.push({ ...event }); + } + setActiveToolCalls([...toolCalls]); + }, + abortController.signal, + ); + } catch (error: any) { + if (error?.name === "AbortError" || abortController.signal.aborted) { + return; + } + const errorMsg = error?.message || "Failed to get response. Please try again."; + setMessages((prev) => [ + ...prev, + { role: "assistant", content: `Error: ${errorMsg}` }, + ]); + setStreamingContent(""); + } finally { + setIsLoading(false); + abortControllerRef.current = null; + } + }; + + const handleKeyDown = (e: React.KeyboardEvent) => { + if (e.key === "Enter" && !e.shiftKey) { + e.preventDefault(); + handleSend(); + } + }; + + const handleClose = () => { + if (abortControllerRef.current) { + abortControllerRef.current.abort(); + } + onClose(); + }; + + const handleClear = () => { + setMessages([]); + setStreamingContent(""); + setActiveToolCalls([]); + setStatusMessage(null); + }; + + return ( +
    + {/* Header */} +
    +
    +
    + + + +

    Ask AI

    +
    + +
    +

    + Ask about your spend, models, keys, and trends +

    +
    + + {/* Model selector */} +
    +