Limit by tenant and org

src/app/api/usage/route.ts

@@ -2,7 +2,11 @@ import { NextRequest, NextResponse } from "next/server";
 import { getSessionUser } from "@/lib/session";
 import { listTenants } from "@/lib/k8s";
 import { listVisibleTenants } from "@/lib/visibility";
-import { getTeamInfo, getTeamSpendLogsV2 } from "@/lib/litellm";
+import {
+  getTeamInfo,
+  getTeamSpendLogsV2,
+  findKeyByAlias,
+} from "@/lib/litellm";
 import { safeError } from "@/lib/errors";
 
 /**
@@ -126,6 +130,16 @@ export async function GET(req: NextRequest) {
   try {
     const teamInfo = await getTeamInfo(teamId);
 
+    // Per-tenant budget lives on the virtual key, not the team
+    // (Feature 7 fix). When the request is scoped to a specific
+    // tenant (keyAlias provided), look up the key so we can return
+    // the per-tenant cap. Tolerate failure — older LiteLLM builds
+    // or short-lived race conditions during provisioning shouldn't
+    // 500 the whole usage page; we degrade to "no key info".
+    const keyInfo = keyAlias
+      ? await findKeyByAlias(teamId, keyAlias).catch(() => null)
+      : null;
+
     // Page through results — server-side filtered by key_alias when
     // provided. Pagination still needed because LiteLLM caps
     // page_size at 100, and a busy tenant can easily exceed that in
@@ -191,17 +205,38 @@ export async function GET(req: NextRequest) {
         totalSpend,
         requestCount: allRequests.length,
       },
-      // Budget is always team-level (= company budget). Spend reported
-      // here is the team total, not the per-key total — the customer
-      // wants to see "how much of our company budget is left", not
-      // just "how much has this one tenant cost".
-      budget: {
-        maxBudget: teamInfo?.team_info?.max_budget ?? null,
-        spend: teamInfo?.team_info?.spend ?? 0,
-        remaining: teamInfo?.team_info?.max_budget
-          ? teamInfo.team_info.max_budget - (teamInfo.team_info.spend ?? 0)
-          : null,
-      },
+      // Budget reporting (Feature 7).
+      //
+      // When the caller scopes to a specific tenant (keyAlias set),
+      // we report THAT tenant's per-key budget — that's what the
+      // tenant detail page renders, and what the customer expects
+      // when they see "Budget" on a tenant's page.
+      //
+      // When unscoped (admin / org-wide view), we fall back to the
+      // team budget — that's the org-wide cap, conceptually different
+      // but the only thing meaningful at that scope.
+      //
+      // The two cases display the same way; the editor button gates
+      // on whether we know which tenant we're on (= keyAlias set).
+      budget: keyAlias && keyInfo
+        ? {
+            maxBudget: keyInfo.maxBudget,
+            spend: keyInfo.spend,
+            remaining:
+              keyInfo.maxBudget !== null
+                ? keyInfo.maxBudget - keyInfo.spend
+                : null,
+            budgetDuration: keyInfo.budgetDuration,
+          }
+        : {
+            maxBudget: teamInfo?.team_info?.max_budget ?? null,
+            spend: teamInfo?.team_info?.spend ?? 0,
+            remaining: teamInfo?.team_info?.max_budget
+              ? teamInfo.team_info.max_budget -
+                (teamInfo.team_info.spend ?? 0)
+              : null,
+            budgetDuration: teamInfo?.team_info?.budget_duration ?? null,
+          },
       rateLimits: {
         rpm: teamInfo?.team_info?.rpm_limit ?? null,
         tpm: teamInfo?.team_info?.tpm_limit ?? null,