Suspendedremoval display in Frontend

src/app/[locale]/dashboard/page.tsx

@@ -2,7 +2,10 @@ import { getSessionUser, canMutate } from "@/lib/session";
 import { getTranslations, getFormatter } from "next-intl/server";
 import { redirect } from "next/navigation";
 import { listTenants } from "@/lib/k8s";
-import { listActiveTenantRequestsByOrgId } from "@/lib/db";
+import {
+  listActiveTenantRequestsByOrgId,
+  syncProvisioningStatuses,
+} from "@/lib/db";
 import {
   listVisibleTenants,
   canSeeInflightRequests,
@@ -11,6 +14,7 @@ import {
 import { personalAccountAtCapacity } from "@/lib/personal-org";
 import { Card, CardHeader } from "@/components/ui/card";
 import { StatusBadge } from "@/components/ui/status-badge";
+import { WarningBadge } from "@/components/ui/warning-badge";
 import { OnboardingFlow } from "@/components/onboarding/onboarding-flow";
 import { ProvisioningStatus } from "@/components/onboarding/provisioning-status";
 import { formatDateTime } from "@/lib/format";
@@ -159,6 +163,23 @@ export default async function DashboardPage() {
 
   // Pending/in-flight requests are only shown to roles that can act on
   // them. `user`-role customers see no request cards.
+  //
+  // syncProvisioningStatuses runs on every dashboard load: it walks
+  // active and provisioning rows and reconciles them against the
+  // current cluster state. Without this, the operator-initiated
+  // 60-day TTL deletion (Bug 37b) leaves the portal showing "Your
+  // assistant is ready!" cards for tenants that no longer exist —
+  // the operator deletes the CR, but the DB row stays at active=true
+  // until something updates it. Running the sync at every dashboard
+  // load keeps the portal eventually consistent with the cluster
+  // without needing a separate cron/job.
+  //
+  // Cost: one K8s GET per row in (active, provisioning) status. At
+  // pilot scale this is small; if it grows we'd cache or move to a
+  // periodic background job.
+  if (canSeeInflightRequests(user)) {
+    await syncProvisioningStatuses();
+  }
   const orgRequests = canSeeInflightRequests(user)
     ? await listActiveTenantRequestsByOrgId(user.orgId)
     : [];
@@ -173,7 +194,16 @@ export default async function DashboardPage() {
     (t) => t.metadata.labels?.["pieced.ch/zitadel-org-id"] === user.orgId
   );
   const inflightRequests = orgRequests.filter(
-    (r) => !r.tenantName || !orgScopedTenants.some((t) => t.metadata.name === r.tenantName)
+    (r) =>
+      // Only show provision (initial creation) requests on the
+      // dashboard. Resume requests (Bug 37a) belong with their
+      // specific tenant — the SubscriptionToggle on the tenant
+      // detail page renders the pending state there. Showing them
+      // on the dashboard too would duplicate the surface and
+      // confuse customers about which tenant they refer to.
+      r.requestType !== "resume" &&
+      (!r.tenantName ||
+        !orgScopedTenants.some((t) => t.metadata.name === r.tenantName))
   );
 
   // Slice 5: only owners (and platform users, who'd typically be using
@@ -348,7 +378,10 @@ export default async function DashboardPage() {
                         {tenant.metadata.name}
                       </div>
                     </div>
-                    <StatusBadge phase={tenant.status?.phase ?? "Pending"} />
+                    <div className="flex items-center gap-2 shrink-0">
+                      <StatusBadge phase={tenant.status?.phase ?? "Pending"} />
+                      <WarningBadge warnings={tenant.status?.warnings ?? []} />
+                    </div>
                   </div>
 
                   {tenant.spec.agentName && (

src/app/[locale]/tenants/[name]/page.tsx

@@ -3,7 +3,9 @@ import { getTranslations, getFormatter } from "next-intl/server";
 import { redirect, notFound } from "next/navigation";
 import { getTenant } from "@/lib/k8s";
 import { canUserSeeTenant } from "@/lib/visibility";
+import { getPendingResumeRequestForTenant } from "@/lib/db";
 import { StatusBadge } from "@/components/ui/status-badge";
+import { WarningBadge } from "@/components/ui/warning-badge";
 import { UsageDisplay } from "@/components/dashboard/usage-display";
 import { PackageList } from "@/components/packages/package-list";
 import { WorkspaceEditor } from "@/components/packages/workspace-editor";
@@ -46,6 +48,13 @@ export default async function TenantDetailPage({
   // The current state comes from spec.suspend on the CR.
   const isSuspended = Boolean(tenant.spec.suspend);
 
+  // Bug 37a: when the tenant is suspended, an owner can request
+  // reactivation (admin-gated). Look up whether one is in flight so
+  // the SubscriptionToggle can render the right state.
+  const pendingResumeRequest = isSuspended
+    ? await getPendingResumeRequestForTenant(name)
+    : null;
+
   // Bug 7: assigned-users panel is meaningless for personal tenants
   // (sole-owner by definition; the only "assignee" is the owner
   // themselves). We hide the panel when EITHER the CR carries the
@@ -66,18 +75,12 @@ export default async function TenantDetailPage({
   );
   const channelUsers = tenant.spec.channelUsers || {};
 
-  // Admins inspecting another tenant's usage: pass teamId AND keyAlias so
-  // the backend filters spend logs by this specific tenant's virtual key.
-  // Without keyAlias the response would include sibling tenants in the
-  // same org, since teams are now shared (Slice 2).
-  // Customers viewing their own: pass nothing — backend resolves both
-  // from the session-bound tenant.
-  const usageTeamId = user.isPlatform
-    ? tenant.status?.litellmTeamId || undefined
-    : undefined;
-  const usageKeyAlias = user.isPlatform
-    ? tenant.status?.litellmKeyAlias || undefined
-    : undefined;
+  // Bug 19 fix: every viewer (customer or admin) passes the tenant
+  // name to UsageDisplay. The /api/usage route resolves team+alias
+  // from the tenant CR's status and applies the visibility check, so
+  // no per-role branching is needed here. Previous version only
+  // passed identifiers for platform admins; customers got "the first
+  // visible tenant" by API fallback, mingling siblings.
 
   return (
     <div>
@@ -88,6 +91,7 @@ export default async function TenantDetailPage({
             {tenant.spec.displayName || name}
           </h1>
           <StatusBadge phase={tenant.status?.phase ?? "Pending"} />
+          <WarningBadge warnings={tenant.status?.warnings ?? []} />
         </div>
         {tenant.spec.agentName && (
           <p className="text-sm text-text-secondary mt-3">
@@ -148,7 +152,7 @@ export default async function TenantDetailPage({
         <h2 className="text-xs font-semibold uppercase tracking-wider text-text-muted mb-3">
           {t("usage")}
         </h2>
-        <UsageDisplay teamId={usageTeamId} keyAlias={usageKeyAlias} />
+        <UsageDisplay tenant={name} />
       </section>
 
       {/* Packages */}
@@ -212,7 +216,19 @@ export default async function TenantDetailPage({
               ? t("subscriptionDescriptionSuspended")
               : t("subscriptionDescriptionActive")}
           </p>
-          <SubscriptionToggle tenantName={name} suspended={isSuspended} />
+          <SubscriptionToggle
+            tenantName={name}
+            suspended={isSuspended}
+            isPlatform={user.isPlatform}
+            pendingResumeRequest={
+              pendingResumeRequest
+                ? {
+                    id: pendingResumeRequest.id,
+                    createdAt: pendingResumeRequest.createdAt,
+                  }
+                : null
+            }
+          />
         </section>
       )}
     </div>

src/app/api/admin/requests/[id]/approve/route.ts

@@ -5,7 +5,7 @@ import {
   updateTenantRequestStatus,
   clearEncryptedSecrets,
 } from "@/lib/db";
-import { createTenant } from "@/lib/k8s";
+import { createTenant, patchTenantSpec, setTenantAnnotation } from "@/lib/k8s";
 import { sendApprovalEmail } from "@/lib/email";
 import { decryptSecrets } from "@/lib/crypto";
 import { writePackageSecrets } from "@/lib/openbao";
@@ -19,14 +19,26 @@ import { safeError } from "@/lib/errors";
 
 /**
  * POST /api/admin/requests/[id]/approve
- * Approve a tenant request:
- * 1. Decrypt stored package secrets (if any)
- * 2. Write each package's secrets to OpenBao at secret/data/tenants/{tenant-name}/{package}
- * 3. Null the encrypted_secrets column
- * 4. Build workspace files (SOUL.md, AGENTS.md, TOOLS.md)
- * 5. Create PiecedTenant CR
- * 6. Update request status, notify customer.
- * Also supports re-approving a previously rejected request (clears admin notes).
+ *
+ * Approve a request. Two paths depending on request_type:
+ *
+ * Provision (the original purpose):
+ *  1. Decrypt stored package secrets (if any)
+ *  2. Write each package's secrets to OpenBao
+ *  3. Null the encrypted_secrets column
+ *  4. Build workspace files (SOUL.md, AGENTS.md, TOOLS.md)
+ *  5. Create PiecedTenant CR
+ *  6. Update request status, notify customer.
+ *  Supports re-approving a previously rejected request (clears admin notes).
+ *
+ * Resume (Bug 37a):
+ *  1. PATCH spec.suspend=false on the existing PiecedTenant CR.
+ *  2. Clear the `pieced.ch/resume-request-pending` annotation so the
+ *     operator knows the request is settled (and doesn't pause its
+ *     60-day TTL forever — though now that the tenant isn't suspended,
+ *     the timer is moot).
+ *  3. Mark request approved, notify customer.
+ *  No CR creation, no secret materialisation, no workspace files.
  */
 export async function POST(
   request: Request,
@@ -60,6 +72,58 @@ export async function POST(
     );
   }
 
+  // Resume request: short path. Just patch the existing tenant, clear
+  // the annotation, mark approved.
+  if (tenantRequest.requestType === "resume") {
+    if (!tenantRequest.tenantName) {
+      // Shouldn't happen — resume requests are created with tenant_name
+      // set. Defensive 500 if it does.
+      return NextResponse.json(
+        { error: "Resume request has no tenant_name" },
+        { status: 500 }
+      );
+    }
+    try {
+      await patchTenantSpec(tenantRequest.tenantName, { suspend: false });
+      // Clear the annotation that pauses the operator's 60-day TTL.
+      // Best-effort — annotation cleanup is also done by the operator
+      // when it sees suspend=false on the next reconcile (it clears
+      // status.suspendedAt), but explicitly clearing here keeps the
+      // CR clean.
+      try {
+        await setTenantAnnotation(
+          tenantRequest.tenantName,
+          "pieced.ch/resume-request-pending",
+          null
+        );
+      } catch (e) {
+        console.warn(
+          "post-approve annotation clear failed; not blocking",
+          e
+        );
+      }
+
+      await updateTenantRequestStatus(id, "approved", { adminNotes });
+
+      await sendApprovalEmail(
+        tenantRequest.contactEmail,
+        tenantRequest.contactName,
+        tenantRequest.companyName
+      ).catch((e) => console.error("approval email failed:", e));
+
+      return NextResponse.json({
+        message: "Resume approved. Tenant is reactivating.",
+        tenantName: tenantRequest.tenantName,
+      });
+    } catch (e: any) {
+      console.error("Resume approval failed:", e);
+      return NextResponse.json(
+        { error: safeError(e, "Failed to approve resume") },
+        { status: 500 }
+      );
+    }
+  }
+
   const isReApproval = tenantRequest.status === "rejected";
 
   // Build the CR name: see `lib/tenant-naming.ts` for the format spec.

src/app/api/admin/requests/[id]/reject/route.ts

@@ -1,11 +1,19 @@
 import { NextResponse } from "next/server";
 import { requirePlatformRole } from "@/lib/session";
 import { getTenantRequestById, updateTenantRequestStatus } from "@/lib/db";
+import { setTenantAnnotation } from "@/lib/k8s";
 import { sendRejectionEmail } from "@/lib/email";
 
 /**
  * POST /api/admin/requests/[id]/reject
  * Reject a tenant request and notify the customer.
+ *
+ * For resume requests (Bug 37a): also clears the
+ * `pieced.ch/resume-request-pending` annotation on the tenant CR.
+ * The operator's 60-day TTL then resumes counting from the original
+ * suspendedAt — rejection doesn't reset it. The customer can submit
+ * a fresh resume request later if circumstances change, but that
+ * starts a new pending row and re-stamps the annotation.
  */
 export async function POST(
   request: Request,
@@ -37,6 +45,26 @@ export async function POST(
     adminNotes,
   });
 
+  // Resume rejection: clear the annotation so the operator's TTL
+  // resumes. Best-effort — failure is logged, not propagated.
+  if (
+    tenantRequest.requestType === "resume" &&
+    tenantRequest.tenantName
+  ) {
+    try {
+      await setTenantAnnotation(
+        tenantRequest.tenantName,
+        "pieced.ch/resume-request-pending",
+        null
+      );
+    } catch (e) {
+      console.warn(
+        "post-reject annotation clear failed; operator's TTL will pause until annotation removed by admin",
+        e
+      );
+    }
+  }
+
   // Notify customer
   await sendRejectionEmail(
     tenantRequest.contactEmail,

src/app/api/onboarding/[id]/route.ts

@@ -6,6 +6,7 @@ import {
   updateTenantRequestEditableFields,
 } from "@/lib/db";
 import { encryptSecrets } from "@/lib/crypto";
+import { setTenantAnnotation } from "@/lib/k8s";
 import { onboardingSchema } from "@/lib/validation";
 import { safeError } from "@/lib/errors";
 
@@ -91,6 +92,25 @@ export async function DELETE(
 
   try {
     await updateTenantRequestStatus(id, "cancelled");
+
+    // Customer cancels their own pending resume request: clear the
+    // operator-side annotation so the 60-day TTL resumes counting.
+    // Best-effort — the operator handles missing annotation gracefully.
+    if (tr.requestType === "resume" && tr.tenantName) {
+      try {
+        await setTenantAnnotation(
+          tr.tenantName,
+          "pieced.ch/resume-request-pending",
+          null
+        );
+      } catch (e) {
+        console.warn(
+          "post-cancel annotation clear failed; not blocking",
+          e
+        );
+      }
+    }
+
     return NextResponse.json({ message: "Request cancelled.", id });
   } catch (e: any) {
     console.error("Failed to cancel request:", e);

src/app/api/tenants/[name]/resume-request/route.ts

@@ -0,0 +1,153 @@
+import { NextRequest, NextResponse } from "next/server";
+import { getSessionUser, canMutate } from "@/lib/session";
+import { getTenant, setTenantAnnotation } from "@/lib/k8s";
+import { canUserSeeTenant } from "@/lib/visibility";
+import {
+  createResumeRequest,
+  getPendingResumeRequestForTenant,
+  getTenantRequestByTenantName,
+} from "@/lib/db";
+import { safeError } from "@/lib/errors";
+
+/**
+ * POST /api/tenants/[name]/resume-request
+ *
+ * Owner-initiated request to reactivate a suspended tenant (Bug 37a).
+ * Creates a pending tenant_request of type 'resume' for admin review,
+ * and stamps the PiecedTenant CR with an annotation that pauses the
+ * operator's 60-day deletion timer.
+ *
+ * Why a request flow at all
+ * -------------------------
+ * Customers can self-serve cancel; resume requires admin oversight.
+ * Reactivation may involve re-validating billing, confirming the
+ * customer still wants to be active, or other manual steps. The
+ * request flow gives admins a queue to review, with the same approve/
+ * reject UX as initial provision requests.
+ *
+ * Authorization
+ * -------------
+ * Owners and platform admins. Platform admins shouldn't normally use
+ * this endpoint — they have direct PATCH suspend access — but it's
+ * permissive in case admin tooling pivots.
+ *
+ * Validation
+ * ----------
+ *  - Tenant must exist and be visible to the caller.
+ *  - Tenant must be currently suspended. Resuming an active tenant
+ *    is meaningless.
+ *  - At most one pending resume request per tenant. Enforced by the
+ *    DB's partial unique index, but we also check explicitly here to
+ *    return a friendly 409 instead of a 500.
+ *
+ * Side effects on success
+ * -----------------------
+ *  - INSERT into tenant_requests (request_type='resume', status='pending')
+ *  - PATCH annotation `pieced.ch/resume-request-pending=<request-id>` on
+ *    the CR. This is the operator's signal to pause its 60-day deletion
+ *    timer until the request transitions to terminal.
+ *
+ * The annotation set is best-effort: if the K8s PATCH fails after the
+ * DB insert, the row exists without the annotation. The customer
+ * sees the request as pending; admin can still approve. The only
+ * functional consequence is the 60-day timer doesn't pause until the
+ * next request transition, which is fine in practice (admin response
+ * times are dramatically shorter than 60 days).
+ */
+export async function POST(
+  req: NextRequest,
+  { params }: { params: Promise<{ name: string }> }
+) {
+  const user = await getSessionUser();
+  if (!user) {
+    return NextResponse.json({ error: "Unauthorized" }, { status: 401 });
+  }
+  if (!canMutate(user)) {
+    return NextResponse.json({ error: "Forbidden" }, { status: 403 });
+  }
+
+  const { name } = await params;
+  const tenant = await getTenant(name);
+  if (!tenant) {
+    return NextResponse.json({ error: "Not found" }, { status: 404 });
+  }
+  if (!(await canUserSeeTenant(user, tenant))) {
+    return NextResponse.json({ error: "Not found" }, { status: 404 });
+  }
+
+  if (!tenant.spec.suspend) {
+    return NextResponse.json(
+      { error: "Tenant is not suspended; nothing to resume." },
+      { status: 409 }
+    );
+  }
+
+  // Already a pending request? Don't duplicate.
+  const existing = await getPendingResumeRequestForTenant(name);
+  if (existing) {
+    return NextResponse.json(
+      {
+        error: "A resume request for this tenant is already pending.",
+        request: { id: existing.id, createdAt: existing.createdAt },
+      },
+      { status: 409 }
+    );
+  }
+
+  // Pull traceability fields (companyName, agentName) from the original
+  // provision request. The schema marks these NOT NULL, so we have to
+  // populate them; copying from the provision row keeps the resume
+  // row navigable in the admin UI without making up values.
+  const provision = await getTenantRequestByTenantName(name);
+
+  try {
+    const resumeRequest = await createResumeRequest({
+      tenantName: name,
+      zitadelOrgId:
+        tenant.metadata.labels?.["pieced.ch/zitadel-org-id"] ?? user.orgId,
+      zitadelUserId: user.id,
+      contactName: user.name,
+      contactEmail: user.email,
+      companyName: provision?.companyName ?? tenant.spec.displayName ?? name,
+      agentName: provision?.agentName ?? "Assistant",
+    });
+
+    // Stamp the annotation so the operator pauses its TTL. If this
+    // fails the request still exists; surface the error so admin
+    // tooling can re-stamp if needed, but don't roll back.
+    try {
+      await setTenantAnnotation(
+        name,
+        "pieced.ch/resume-request-pending",
+        resumeRequest.id
+      );
+    } catch (e) {
+      console.warn(
+        "resume request created but annotation could not be set; operator's 60-day timer will not pause until next reconcile triggered by request transition",
+        e
+      );
+    }
+
+    return NextResponse.json(
+      {
+        message: "Resume request submitted. An admin will review shortly.",
+        request: { id: resumeRequest.id, status: resumeRequest.status },
+      },
+      { status: 201 }
+    );
+  } catch (e: any) {
+    // Unique violation (a pending row already exists for this tenant)
+    // is friendly-handled above; this catches everything else.
+    if (e.code === "23505") {
+      return NextResponse.json(
+        { error: "A resume request for this tenant is already pending." },
+        { status: 409 }
+      );
+    }
+    console.error("Resume request creation failed:", e);
+    return NextResponse.json(
+      { error: safeError(e, "Failed to submit resume request") },
+      { status: 500 }
+    );
+  }
+}

src/app/api/tenants/[name]/suspend/route.ts

@@ -1,7 +1,7 @@
 import { NextRequest, NextResponse } from "next/server";
 import { z } from "zod";
 import { getSessionUser, canMutate } from "@/lib/session";
-import { getTenant, patchTenantSpec } from "@/lib/k8s";
+import { getTenant, patchTenantSpec, setTenantAnnotation } from "@/lib/k8s";
 import { canUserSeeTenant } from "@/lib/visibility";
 import { safeError } from "@/lib/errors";
 
@@ -12,37 +12,38 @@ const patchSchema = z.object({
 /**
  * PATCH /api/tenants/[name]/suspend
  *
- * Customer-side "Cancel subscription" / "Resume" toggle (Bug 31).
+ * Direct suspend control on the PiecedTenant CR. Sets `spec.suspend`
+ * to true (cancel) or false (resume).
  *
- * Sets `spec.suspend` on the PiecedTenant CR. The operator interprets
- * this flag as "stop reconciling this tenant" — workloads, packages,
- * and channel-user changes are no longer applied. Existing data is
- * preserved (namespace, ConfigMaps, OpenBao secrets, CNPG database,
- * billing records). Resuming sets the flag back to false and the
- * operator picks up reconciliation on the next loop.
+ * Authorization (Bug 37a)
+ * -----------------------
+ * - suspend=true  → owners and platform admins may call.
+ * - suspend=false → platform admins ONLY. Owners must go through the
+ *   resume-request flow (POST /api/tenants/[name]/resume-request),
+ *   which creates a pending request for admin approval. This
+ *   asymmetry is by design: cancellation is self-service (low risk;
+ *   reversible by request); reactivation requires admin oversight
+ *   (e.g. to re-validate billing, confirm intent).
  *
- * Authorization
- * -------------
- * - Customer-side: only an `owner` of the tenant's org may call this.
- *   `canMutate` is the right gate (mirrors the rest of the customer
- *   API surface). User-role members cannot cancel a subscription.
- * - Platform staff: allowed via `canMutate`'s isPlatform branch, but
- *   in practice they should use admin tooling for this — the action
- *   is exposed here for the customer's benefit.
+ * Customer flow:
+ *   - Cancel:  PATCH suspend=true  here
+ *   - Resume:  POST /resume-request — creates a 'resume' tenant_request,
+ *              admin approves via /api/admin/requests/[id]/approve which
+ *              then PATCHes suspend=false here as a platform user.
  *
- * Visibility check is via `canUserSeeTenant` — same notFound() trick
- * as the detail page, so we don't leak existence of tenants the
- * caller can't see.
+ * Workload behaviour
+ * ------------------
+ * On suspend=true the operator deletes the OpenClawInstance, stopping
+ * the pod within seconds. Tenant data — namespace, ConfigMaps,
+ * OpenBao secrets, CNPG database, LiteLLM team — is retained.
  *
- * Note on workload teardown
- * -------------------------
- * As of this writing, the operator's `suspend` handling is "skip
- * reconciliation and set status.phase to Suspended". The underlying
- * StatefulSet keeps running until next reconciliation, which won't
- * happen while suspended. Group D will add scale-to-zero so cancelled
- * subscriptions actually stop incurring compute. Until then, an
- * operator following up with a `kubectl scale` is the workaround.
- * Customer data is preserved either way.
+ * Suspended tenants enter a 60-day retention window (operator
+ * constant `retentionAfterSuspend`); after that, the tenant is fully
+ * deleted unless a pending resume request exists. The operator
+ * checks the `pieced.ch/resume-request-pending` annotation to know
+ * about pending requests; we set it here when admin approves the
+ * resume (transitively, via the admin-approve endpoint), and clear
+ * it when the request reaches a terminal state.
  */
 export async function PATCH(
   req: NextRequest,
@@ -76,6 +77,18 @@ export async function PATCH(
   }
   const { suspend } = parsed.data;
 
+  // Bug 37a: resume (suspend=false) is platform-admin only via this
+  // endpoint. Owners must go through the resume-request flow.
+  if (!suspend && !user.isPlatform) {
+    return NextResponse.json(
+      {
+        error:
+          "Resume requires platform-admin approval. Submit a resume request via /api/tenants/[name]/resume-request.",
+      },
+      { status: 403 }
+    );
+  }
+
   // No-op early exit. Avoids a needless K8s patch + status churn when
   // the user double-clicks the button or the UI is briefly out of sync.
   if (Boolean(tenant.spec.suspend) === suspend) {
@@ -87,10 +100,32 @@ export async function PATCH(
 
   try {
     await patchTenantSpec(name, { suspend });
+
+    // On admin-side resume, also clear the pending-resume-request
+    // annotation if it exists. Belt-and-suspenders: the admin-approve
+    // endpoint already clears it on its happy path, but a platform
+    // user resuming directly via this endpoint shouldn't leave the
+    // annotation behind. Best-effort: failure to clear the annotation
+    // is logged but doesn't fail the resume.
+    if (!suspend) {
+      try {
+        await setTenantAnnotation(
+          name,
+          "pieced.ch/resume-request-pending",
+          null
+        );
+      } catch (e) {
+        console.warn(
+          "failed to clear resume-request-pending annotation; operator will see it stale until next request transition",
+          e
+        );
+      }
+    }
+
     return NextResponse.json(
       {
         message: suspend
-          ? "Subscription cancelled. Your data is preserved."
+          ? "Subscription cancelled. Your data is preserved for 60 days."
           : "Subscription resumed.",
         suspend,
       },

src/app/api/usage/route.ts

@@ -8,64 +8,109 @@ import { safeError } from "@/lib/errors";
 /**
  * GET /api/usage
  *
- * Customers: tenant resolved server-side from the user's orgId. The
- *            response is filtered by the tenant's `litellmKeyAlias` so
- *            sibling tenants in the same org don't bleed into the total.
- * Platform admins: may pass ?teamId=... to inspect any team. They may
- *                  also pass ?keyAlias=... to scope to a single tenant.
+ * Per-tenant spend/token usage for a given month.
  *
- * Slice 2 note
- * ------------
- * LiteLLM teams are now shared across all tenants of an org. The team's
- * `/team/info` budget is the *company* budget; the per-tenant numbers
- * come from filtering spend logs by `key_alias`. If a tenant has no
- * `litellmKeyAlias` in status (transitional state right after upgrade,
- * before the operator has reconciled), we fall back to team-level
- * filtering — the numbers will be slightly inflated for that one
- * reconcile cycle.
+ * Resolution rules (in priority order)
+ * ------------------------------------
+ *  1. `?tenant=<name>` query param — the canonical path. The route
+ *     looks up the PiecedTenant CR by name, runs it through the
+ *     viewer's visibility filter, and reads `status.litellmTeamId` +
+ *     `status.litellmKeyAlias`. This is what the tenant-detail page
+ *     calls with for both customers and admins.
+ *  2. `?teamId=<id>` (+ optional `?keyAlias=<alias>`) — admin escape
+ *     hatch for debugging across orgs (e.g. opening the platform
+ *     panel without a specific tenant in mind). Platform-only;
+ *     ignored for customer sessions.
+ *  3. No params — 400. We deliberately do NOT fall back to "the
+ *     first visible tenant". Bug 19: that fallback meant siblings
+ *     in the same org showed identical numbers because the API
+ *     always picked the same "first" tenant regardless of which
+ *     detail page the customer was viewing. Forcing callers to be
+ *     explicit makes the bug structurally impossible to reintroduce.
+ *
+ * Filtering
+ * ---------
+ * LiteLLM's `/spend/logs/v2` accepts a server-side `key_alias` filter.
+ * We pass it through directly — no more "fetch all team pages and
+ * post-filter in JS" (which was O(team_total) memory per request and
+ * masked the routing bug above by being slow enough that nobody
+ * noticed which alias was actually being used).
+ *
+ * The team-level budget is still surfaced as the *org* budget, since
+ * teams are org-scoped post-Slice-2. That's intentional: the customer
+ * sees "your company has X budget remaining" alongside "this tenant
+ * cost Y this month".
  */
 export async function GET(req: NextRequest) {
   const user = await getSessionUser();
   if (!user)
     return NextResponse.json({ error: "Unauthorized" }, { status: 401 });
 
+  const tenantName = req.nextUrl.searchParams.get("tenant");
   let teamId: string | null = null;
   let keyAlias: string | null = null;
 
-  if (user.isPlatform) {
-    teamId = req.nextUrl.searchParams.get("teamId") ?? null;
-    keyAlias = req.nextUrl.searchParams.get("keyAlias") ?? null;
-  }
-
-  // For customers (or admins without explicit params): resolve from
-  // the user's *visible* tenants. With Slice 6, a `user`-role member
-  // can only see usage for tenants they're assigned to — a non-assigned
-  // user defaults to "no active tenant" (404).
-  //
-  // Owner and platform get the full org-scoped list and pick the first
-  // tenant, matching the dashboard's "current instance" semantics.
-  if (!teamId) {
+  if (tenantName) {
+    // Path 1: resolve from tenant name with visibility check.
+    //
+    // listVisibleTenants enforces the same visibility rules as every
+    // other read endpoint:
+    //   - platform admins see everything
+    //   - owners see all tenants in their org
+    //   - users see only the tenants they're assigned to (Slice 6)
+    //
+    // Filtering through that list rather than reading the CR directly
+    // means a malicious caller can't probe arbitrary tenant names to
+    // learn what exists in other orgs.
     const allTenants = await listTenants();
     const visible = await listVisibleTenants(user, allTenants);
-    const orgTenant = visible.find((t) => !!t.status?.litellmTeamId);
+    const tenant = visible.find((t) => t.metadata.name === tenantName);
 
-    if (!orgTenant?.status?.litellmTeamId) {
+    if (!tenant) {
       return NextResponse.json(
-        { error: "No active tenant found for your organization" },
+        { error: "Tenant not found or not accessible" },
         { status: 404 }
       );
     }
-    teamId = orgTenant.status.litellmTeamId;
-
-    // If the operator has populated the per-tenant key alias, filter by it.
-    // Falling back to team-level (no alias) will return the org total, which
-    // is acceptable transitionally but means siblings' usage shows up here.
-    if (orgTenant.status.litellmKeyAlias) {
-      keyAlias = orgTenant.status.litellmKeyAlias;
+    if (!tenant.status?.litellmTeamId) {
+      // Tenant exists but the operator hasn't reconciled it yet.
+      // Common right after onboarding; the customer should see a
+      // friendly empty state, not a 500.
+      return NextResponse.json(
+        { error: "Tenant is still provisioning, no usage data yet" },
+        { status: 409 }
+      );
     }
+    teamId = tenant.status.litellmTeamId;
+    // litellmKeyAlias is set by the operator's LiteLLM reconcile step
+    // alongside litellmTeamId, so if teamId is present this should be
+    // too. Defensive fallback to team-level if missing — in that case
+    // the customer briefly sees company totals until the next operator
+    // reconcile, which is better than 500.
+    keyAlias = tenant.status.litellmKeyAlias ?? null;
+  } else if (user.isPlatform) {
+    // Path 2: admin escape hatch.
+    teamId = req.nextUrl.searchParams.get("teamId");
+    keyAlias = req.nextUrl.searchParams.get("keyAlias");
+    if (!teamId) {
+      return NextResponse.json(
+        {
+          error:
+            "Either ?tenant=<name> or ?teamId=<id> (admin) must be provided",
+        },
+        { status: 400 }
+      );
+    }
+  } else {
+    // Path 3: no resolution possible. See doc above for why we don't
+    // pick a default.
+    return NextResponse.json(
+      { error: "Tenant must be specified via ?tenant=<name>" },
+      { status: 400 }
+    );
   }
 
-  // Month param: YYYY-MM, defaults to current month
+  // Month param: YYYY-MM, defaults to current month.
   const now = new Date();
   const monthParam =
     req.nextUrl.searchParams.get("month") ||
@@ -81,11 +126,11 @@ export async function GET(req: NextRequest) {
   try {
     const teamInfo = await getTeamInfo(teamId);
 
-    // Fetch all pages from the team. We always query at the team level —
-    // LiteLLM's /spend/logs/v2 doesn't filter by key_alias reliably across
-    // versions, so we paginate and post-filter in code. For pilot scale
-    // this is cheap; if a single team ever exceeds ~10k entries/month we
-    // can revisit.
+    // 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
+    // a month. With server-side filtering this stays cheap regardless
+    // of how busy sibling tenants in the same team are.
     const allRequests: any[] = [];
     let page = 1;
     while (true) {
@@ -94,33 +139,25 @@ export async function GET(req: NextRequest) {
         startStr,
         endStr,
         page,
-        100
+        100,
+        keyAlias
       );
       allRequests.push(...(result.data || []));
       if (page >= (result.total_pages || 1)) break;
       page++;
+      // Defensive cap. A pathological response with bogus total_pages
+      // shouldn't be able to spin us forever. 50 pages × 100 = 5000
+      // entries/month/tenant is well above any realistic usage at
+      // pilot scale.
+      if (page > 50) break;
     }
 
-    // Apply key_alias post-filter when scoping to a single tenant. Match
-    // both `key_alias` (newer LiteLLM) and `metadata.user_api_key_alias`
-    // (older builds nest it inside metadata).
-    const scoped = keyAlias
-      ? allRequests.filter((r) => {
-          const alias =
-            r.key_alias ??
-            r.metadata?.user_api_key_alias ??
-            r.api_key_alias ??
-            null;
-          return alias === keyAlias;
-        })
-      : allRequests;
-
-    // Aggregate by day
+    // Aggregate by day.
     const byDay: Record<
       string,
       { inputTokens: number; outputTokens: number; spend: number }
     > = {};
-    for (const r of scoped) {
+    for (const r of allRequests) {
       const day = (r.startTime || r.endTime || "").slice(0, 10);
       if (!day) continue;
       if (!byDay[day])
@@ -134,30 +171,30 @@ export async function GET(req: NextRequest) {
       .sort(([a], [b]) => a.localeCompare(b))
       .map(([date, d]) => ({ date, ...d }));
 
-    const totalInput = scoped.reduce(
+    const totalInput = allRequests.reduce(
       (s, r) => s + (r.prompt_tokens || 0),
       0
     );
-    const totalOutput = scoped.reduce(
+    const totalOutput = allRequests.reduce(
       (s, r) => s + (r.completion_tokens || 0),
       0
     );
-    const totalSpend = scoped.reduce((s, r) => s + (r.spend || 0), 0);
+    const totalSpend = allRequests.reduce((s, r) => s + (r.spend || 0), 0);
 
     return NextResponse.json({
       teamId,
-      keyAlias, // null when not filtering — useful for the client to know it sees company-wide data
+      keyAlias, // null when admin queries team-wide (no specific tenant)
       month: monthParam,
       currentPeriod: {
         inputTokens: totalInput,
         outputTokens: totalOutput,
         totalSpend,
-        requestCount: scoped.length,
+        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".
+      // 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,

src/components/admin/admin-panel.tsx

@@ -362,9 +362,28 @@ export function AdminPanel({ initialTenants }: AdminPanelProps) {
                         className="border-b border-border last:border-0 hover:bg-surface-2/50 transition-colors"
                       >
                         <td className="px-4 py-3">
-                          <div className="font-medium text-text-primary text-sm">
+                          <div className="font-medium text-text-primary text-sm flex items-center gap-2">
                             {req.companyName}
+                            {/* Bug 37a: distinguish resume requests in the
+                                queue. Provision and resume share status
+                                semantics but very different action
+                                consequences — a resume approval just
+                                un-suspends an existing tenant, no
+                                provisioning workflow runs. */}
+                            {req.requestType === "resume" && (
+                              <span
+                                className="px-1.5 py-0.5 text-[10px] font-semibold rounded uppercase tracking-wider bg-success/15 text-success"
+                                title={t("resumeRequestTooltip")}
+                              >
+                                {t("resumeRequestBadge")}
+                              </span>
+                            )}
                           </div>
+                          {req.requestType === "resume" && req.tenantName && (
+                            <div className="text-text-muted text-xs font-mono mt-0.5">
+                              {req.tenantName}
+                            </div>
+                          )}
                         </td>
                         <td className="px-4 py-3">
                           <div className="text-text-primary text-sm">

src/components/dashboard/usage-display.tsx

@@ -94,17 +94,27 @@ function UsageChart({ data }: { data: DailyUsage[] }) {
 /**
  * Usage display widget.
  *
- * - Customers: don't pass teamId or keyAlias — the backend resolves both
- *   from the session-bound tenant.
- * - Admins inspecting a specific tenant: pass `teamId` (the org-level
- *   LiteLLM team id) AND `keyAlias` (the tenant's virtual-key alias).
- *   Without `keyAlias`, the response includes spend from sibling tenants
- *   in the same org, since teams are shared since Slice 2.
+ * Pass `tenant=<name>` for the canonical path — works for both
+ * customers and admins, the API resolves team+alias from the tenant
+ * CR's status. The visibility check on the API ensures users can't
+ * query tenants they shouldn't see.
+ *
+ * `teamId`/`keyAlias` remain available as a platform-admin escape
+ * hatch for cross-org debugging, but the tenant-detail and dashboard
+ * paths should always use `tenant`.
+ *
+ * Bug 19 fix: previous version omitted both props for customer
+ * sessions, expecting the API to "figure it out". The API's fallback
+ * was "first visible tenant", which meant siblings in the same org
+ * showed identical numbers regardless of which detail page was open.
+ * Now the page passes the tenant name explicitly; no fallback exists.
  */
 export function UsageDisplay({
+  tenant,
   teamId,
   keyAlias,
 }: {
+  tenant?: string | null;
   teamId?: string | null;
   keyAlias?: string | null;
 }) {
@@ -121,11 +131,13 @@ export function UsageDisplay({
     setError(null);
 
     const params = new URLSearchParams({ month });
-    if (teamId) {
+    if (tenant) {
+      params.set("tenant", tenant);
+    } else if (teamId) {
+      // Admin escape hatch — only honoured by the API when the
+      // viewer is platform-role.
       params.set("teamId", teamId);
-    }
-    if (keyAlias) {
-      params.set("keyAlias", keyAlias);
+      if (keyAlias) params.set("keyAlias", keyAlias);
     }
 
     fetch(`/api/usage?${params}`)
@@ -133,7 +145,7 @@ export function UsageDisplay({
       .then(setData)
       .catch((e) => setError(e.message))
       .finally(() => setLoading(false));
-  }, [teamId, keyAlias, month]);
+  }, [tenant, teamId, keyAlias, month]);
 
   useEffect(() => { fetchUsage(); }, [fetchUsage]);
 

src/components/onboarding/provisioning-status.tsx

@@ -5,6 +5,7 @@ import Link from "next/link";
 import { useRouter } from "next/navigation";
 import { useTranslations, useFormatter } from "next-intl";
 import { Card } from "@/components/ui/card";
+import { Modal } from "@/components/ui/modal";
 import { StatusBadge } from "@/components/ui/status-badge";
 import { formatDateTime, formatRelative } from "@/lib/format";
 
@@ -250,43 +251,38 @@ export function ProvisioningStatus({ requestId, canAct }: Props) {
         </div>
 
         {confirmCancel && (
-          <div
-            role="dialog"
-            aria-modal="true"
-            className="fixed inset-0 z-50 flex items-center justify-center p-4 bg-black/60 backdrop-blur-sm"
-            onClick={(e) => {
-              if (e.target === e.currentTarget) setConfirmCancel(false);
-            }}
+          <Modal
+            open={confirmCancel}
+            onClose={() => setConfirmCancel(false)}
+            ariaLabel={t("cancelConfirmRequestTitle")}
           >
-            <div className="bg-surface-1 border border-border rounded-xl p-6 max-w-md w-full">
-              <h3 className="font-display text-lg font-semibold text-text-primary mb-2">
-                {t("cancelConfirmRequestTitle")}
-              </h3>
-              <p className="text-sm text-text-secondary mb-5">
-                {t("cancelConfirmRequestDescription")}
-              </p>
-              <div className="flex justify-end gap-2">
-                <button
-                  type="button"
-                  onClick={() => setConfirmCancel(false)}
-                  disabled={actionPending}
-                  className="text-sm px-4 py-2 rounded-lg border border-border text-text-secondary hover:text-text-primary transition-colors"
-                >
-                  {tCommon("cancel")}
-                </button>
-                <button
-                  type="button"
-                  onClick={handleCancel}
-                  disabled={actionPending}
-                  className="text-sm px-4 py-2 rounded-lg bg-red-500 text-white hover:bg-red-600 transition-colors disabled:opacity-50"
-                >
-                  {actionPending
-                    ? tCommon("loading")
-                    : t("cancelRequestConfirm")}
-                </button>
-              </div>
+            <h3 className="font-display text-lg font-semibold text-text-primary mb-2">
+              {t("cancelConfirmRequestTitle")}
+            </h3>
+            <p className="text-sm text-text-secondary mb-5">
+              {t("cancelConfirmRequestDescription")}
+            </p>
+            <div className="flex justify-end gap-2">
+              <button
+                type="button"
+                onClick={() => setConfirmCancel(false)}
+                disabled={actionPending}
+                className="text-sm px-4 py-2 rounded-lg border border-border text-text-secondary hover:text-text-primary transition-colors"
+              >
+                {tCommon("cancel")}
+              </button>
+              <button
+                type="button"
+                onClick={handleCancel}
+                disabled={actionPending}
+                className="text-sm px-4 py-2 rounded-lg bg-red-500 text-white hover:bg-red-600 transition-colors disabled:opacity-50"
+              >
+                {actionPending
+                  ? tCommon("loading")
+                  : t("cancelRequestConfirm")}
+              </button>
             </div>
-          </div>
+          </Modal>
         )}
       </Card>
     );

src/components/tenants/subscription-toggle.tsx

@@ -2,48 +2,74 @@
 
 import { useState } from "react";
 import { useRouter } from "next/navigation";
-import { useTranslations } from "next-intl";
+import { useTranslations, useFormatter } from "next-intl";
+import { Modal } from "@/components/ui/modal";
+import { formatRelative } from "@/lib/format";
 
 interface Props {
   tenantName: string;
   /**
-   * Current suspend state — server-derived. The control toggles this
-   * via PATCH /api/tenants/[name]/suspend, then refreshes the route
-   * so server-component-side data (status badge, suspended notice)
-   * re-renders.
+   * Current suspend state — server-derived. Drives which control the
+   * customer sees: "Cancel subscription" while active, the
+   * resume-request flow while suspended.
    */
   suspended: boolean;
+  /**
+   * True when the viewer has platform admin role. Platform users are
+   * the only ones who can directly resume a tenant via PATCH; owners
+   * must go through the resume-request flow. We use this in the
+   * suspended branch to decide whether to render a direct "Resume"
+   * button or the "Request reactivation" workflow.
+   */
+  isPlatform: boolean;
+  /**
+   * If a resume request is currently pending for this tenant, its
+   * id and submitted-at. The component renders an info card with
+   * a cancel-request button instead of the request-reactivation
+   * button. Only meaningful when `suspended === true`.
+   */
+  pendingResumeRequest: { id: string; createdAt: string } | null;
 }
 
 /**
- * SubscriptionToggle — owner-side cancel/resume control (Bug 31).
+ * SubscriptionToggle — owner-side cancel/resume control.
  *
- * Renders a single button that toggles between "Cancel subscription"
- * (when active) and "Resume subscription" (when suspended). Cancellation
- * is gated behind a confirmation modal because it's destructive
- * looking from the user's POV — even though no data is lost, the
- * AI assistant becomes unavailable until they resume. Resume has no
- * modal; it's a strict subset of cancellation in terms of risk.
+ * Three render states:
+ *   1. Active: "Cancel subscription" button + confirmation modal
+ *      (mentions 60-day retention before permanent deletion).
+ *   2. Suspended, no pending resume request: "Request reactivation"
+ *      button + simple confirmation modal explaining admin review.
+ *   3. Suspended, pending resume request: status card "Reactivation
+ *      requested X days ago" + "Cancel request" button.
  *
- * The control intentionally lives at the bottom of the tenant detail
- * page rather than next to the status badge — putting it near the
- * top would invite mis-clicks. Customers who want to cancel scroll
- * past the running configuration, billing-relevant info, and assigned
- * users first; that's the right friction level.
+ * Platform admins viewing a suspended tenant get a fourth state in
+ * place of #2/#3: a direct "Resume now" button (no admin queue, no
+ * request flow). This is the admin escape hatch.
  *
- * Suspended tenants render a top-of-page banner separately (see the
- * detail page); this component focuses on the action itself.
+ * The control intentionally lives at the bottom of the tenant
+ * detail page rather than near the top — putting it next to the
+ * status badge would invite mis-clicks.
  */
-export function SubscriptionToggle({ tenantName, suspended }: Props) {
+export function SubscriptionToggle({
+  tenantName,
+  suspended,
+  isPlatform,
+  pendingResumeRequest,
+}: Props) {
   const t = useTranslations("tenantDetail");
   const tCommon = useTranslations("common");
+  const f = useFormatter();
   const router = useRouter();
 
-  const [confirmOpen, setConfirmOpen] = useState(false);
+  const [confirmCancelOpen, setConfirmCancelOpen] = useState(false);
+  const [confirmResumeOpen, setConfirmResumeOpen] = useState(false);
   const [submitting, setSubmitting] = useState(false);
   const [error, setError] = useState("");
 
-  const toggleSuspend = async (next: boolean) => {
+  // Customer-side cancel: PATCH suspend=true. Same path as before.
+  // The 60-day retention copy in the modal is the new bit (Bug 37b);
+  // mechanics are unchanged.
+  const cancel = async () => {
     setSubmitting(true);
     setError("");
     try {
@@ -52,18 +78,14 @@ export function SubscriptionToggle({ tenantName, suspended }: Props) {
         {
           method: "PATCH",
           headers: { "Content-Type": "application/json" },
-          body: JSON.stringify({ suspend: next }),
+          body: JSON.stringify({ suspend: true }),
         }
       );
       if (!res.ok) {
         const data = await res.json().catch(() => ({}));
         throw new Error(data.error || t("subscriptionUpdateFailed"));
       }
-      setConfirmOpen(false);
-      // The status badge + suspended banner are server-rendered, so
-      // a route refresh is the simplest way to reflect the new state.
-      // Optimistic local toggle would diverge from the actual CR if
-      // the operator hasn't observed the patch yet.
+      setConfirmCancelOpen(false);
       router.refresh();
     } catch (e: any) {
       setError(e.message);
@@ -72,56 +94,164 @@ export function SubscriptionToggle({ tenantName, suspended }: Props) {
     }
   };
 
+  // Owner-side resume request: POST a 'resume' tenant_request that
+  // sits pending until admin acts. Different from cancel: no PATCH
+  // on the CR — that happens only when admin approves.
+  const requestResume = async () => {
+    setSubmitting(true);
+    setError("");
+    try {
+      const res = await fetch(
+        `/api/tenants/${encodeURIComponent(tenantName)}/resume-request`,
+        {
+          method: "POST",
+          headers: { "Content-Type": "application/json" },
+        }
+      );
+      if (!res.ok) {
+        const data = await res.json().catch(() => ({}));
+        throw new Error(data.error || t("subscriptionUpdateFailed"));
+      }
+      setConfirmResumeOpen(false);
+      router.refresh();
+    } catch (e: any) {
+      setError(e.message);
+    } finally {
+      setSubmitting(false);
+    }
+  };
+
+  // Customer cancels their own pending resume request.
+  const cancelResumeRequest = async () => {
+    if (!pendingResumeRequest) return;
+    setSubmitting(true);
+    setError("");
+    try {
+      const res = await fetch(
+        `/api/onboarding/${pendingResumeRequest.id}`,
+        { method: "DELETE" }
+      );
+      if (!res.ok) {
+        const data = await res.json().catch(() => ({}));
+        throw new Error(data.error || t("subscriptionUpdateFailed"));
+      }
+      router.refresh();
+    } catch (e: any) {
+      setError(e.message);
+    } finally {
+      setSubmitting(false);
+    }
+  };
+
+  // Platform admin: direct resume, bypassing the request flow.
+  const adminResume = async () => {
+    setSubmitting(true);
+    setError("");
+    try {
+      const res = await fetch(
+        `/api/tenants/${encodeURIComponent(tenantName)}/suspend`,
+        {
+          method: "PATCH",
+          headers: { "Content-Type": "application/json" },
+          body: JSON.stringify({ suspend: false }),
+        }
+      );
+      if (!res.ok) {
+        const data = await res.json().catch(() => ({}));
+        throw new Error(data.error || t("subscriptionUpdateFailed"));
+      }
+      router.refresh();
+    } catch (e: any) {
+      setError(e.message);
+    } finally {
+      setSubmitting(false);
+    }
+  };
+
+  // ─── Suspended branch ───────────────────────────────────────────────
+
   if (suspended) {
+    // Platform admin sees direct resume. Independent of pending
+    // resume — admin can always resume immediately.
+    if (isPlatform) {
+      return (
+        <div>
+          <button
+            type="button"
+            onClick={adminResume}
+            disabled={submitting}
+            className="text-sm font-medium px-4 py-2 rounded-lg border border-success/30 text-success hover:bg-success/10 transition-colors disabled:opacity-50"
+          >
+            {submitting
+              ? tCommon("loading")
+              : t("resumeSubscription")}
+          </button>
+          {pendingResumeRequest && (
+            <p className="text-xs text-text-muted mt-2">
+              {t("resumeRequestPendingNoteAdmin")}
+            </p>
+          )}
+          {error && <p className="text-xs text-red-400 mt-2">{error}</p>}
+        </div>
+      );
+    }
+
+    // Owner with pending resume request: render the request status
+    // card with cancel.
+    if (pendingResumeRequest) {
+      return (
+        <div>
+          <div className="rounded-xl border border-amber-500/30 bg-amber-500/5 px-4 py-3">
+            <div className="text-sm font-medium text-amber-400 mb-1">
+              {t("resumeRequestPendingTitle")}
+            </div>
+            <div className="text-xs text-text-secondary">
+              {t("resumeRequestPendingDescription", {
+                when: formatRelative(pendingResumeRequest.createdAt, f),
+              })}
+            </div>
+            <button
+              type="button"
+              onClick={cancelResumeRequest}
+              disabled={submitting}
+              className="mt-3 text-xs px-3 py-1.5 rounded-lg border border-border text-text-secondary hover:text-text-primary transition-colors disabled:opacity-50"
+            >
+              {submitting
+                ? tCommon("loading")
+                : t("cancelResumeRequest")}
+            </button>
+          </div>
+          {error && <p className="text-xs text-red-400 mt-2">{error}</p>}
+        </div>
+      );
+    }
+
+    // Owner with no pending request: offer to create one.
     return (
       <div>
         <button
           type="button"
-          onClick={() => toggleSuspend(false)}
-          disabled={submitting}
-          className="text-sm font-medium px-4 py-2 rounded-lg border border-success/30 text-success hover:bg-success/10 transition-colors disabled:opacity-50"
+          onClick={() => setConfirmResumeOpen(true)}
+          className="text-sm font-medium px-4 py-2 rounded-lg border border-success/30 text-success hover:bg-success/10 transition-colors"
         >
-          {submitting ? tCommon("loading") : t("resumeSubscription")}
+          {t("requestReactivation")}
         </button>
-        {error && <p className="text-xs text-red-400 mt-2">{error}</p>}
-      </div>
-    );
-  }
+        {error && !confirmResumeOpen && (
+          <p className="text-xs text-red-400 mt-2">{error}</p>
+        )}
 
-  return (
-    <div>
-      <button
-        type="button"
-        onClick={() => setConfirmOpen(true)}
-        className="text-sm font-medium px-4 py-2 rounded-lg border border-border text-text-secondary hover:text-text-primary hover:border-text-secondary transition-colors"
-      >
-        {t("cancelSubscription")}
-      </button>
-      {error && !confirmOpen && (
-        <p className="text-xs text-red-400 mt-2">{error}</p>
-      )}
-
-      {confirmOpen && (
-        <div
-          role="dialog"
-          aria-modal="true"
-          className="fixed inset-0 z-50 flex items-center justify-center p-4 bg-black/60 backdrop-blur-sm"
-          onClick={(e) => {
-            if (e.target === e.currentTarget) setConfirmOpen(false);
-          }}
-        >
-          <div className="bg-surface-1 border border-border rounded-xl p-6 max-w-md w-full">
+        {confirmResumeOpen && (
+          <Modal
+            open={confirmResumeOpen}
+            onClose={() => setConfirmResumeOpen(false)}
+            ariaLabel={t("requestReactivationConfirmTitle")}
+          >
             <h3 className="font-display text-lg font-semibold text-text-primary mb-2">
-              {t("cancelConfirmTitle")}
+              {t("requestReactivationConfirmTitle")}
             </h3>
-            <p className="text-sm text-text-secondary mb-3">
-              {t("cancelConfirmDescription")}
+            <p className="text-sm text-text-secondary mb-5">
+              {t("requestReactivationConfirmDescription")}
             </p>
-            <ul className="text-xs text-text-muted list-disc list-inside space-y-1 mb-5">
-              <li>{t("cancelConfirmBullet1")}</li>
-              <li>{t("cancelConfirmBullet2")}</li>
-              <li>{t("cancelConfirmBullet3")}</li>
-            </ul>
 
             {error && (
               <div className="text-xs text-red-400 bg-red-400/10 border border-red-400/20 rounded-lg px-3 py-2 mb-3">
@@ -132,7 +262,7 @@ export function SubscriptionToggle({ tenantName, suspended }: Props) {
             <div className="flex justify-end gap-2">
               <button
                 type="button"
-                onClick={() => setConfirmOpen(false)}
+                onClick={() => setConfirmResumeOpen(false)}
                 disabled={submitting}
                 className="text-sm px-4 py-2 rounded-lg border border-border text-text-secondary hover:text-text-primary transition-colors"
               >
@@ -140,17 +270,87 @@ export function SubscriptionToggle({ tenantName, suspended }: Props) {
               </button>
               <button
                 type="button"
-                onClick={() => toggleSuspend(true)}
+                onClick={requestResume}
                 disabled={submitting}
-                className="text-sm px-4 py-2 rounded-lg bg-amber-500 text-white hover:bg-amber-600 transition-colors disabled:opacity-50"
+                className="text-sm px-4 py-2 rounded-lg bg-success text-white hover:bg-success/90 transition-colors disabled:opacity-50"
               >
                 {submitting
                   ? tCommon("loading")
-                  : t("cancelSubscriptionConfirm")}
+                  : t("requestReactivationConfirm")}
               </button>
             </div>
+          </Modal>
+        )}
+      </div>
+    );
+  }
+
+  // ─── Active branch ──────────────────────────────────────────────────
+
+  return (
+    <div>
+      <button
+        type="button"
+        onClick={() => setConfirmCancelOpen(true)}
+        className="text-sm font-medium px-4 py-2 rounded-lg border border-border text-text-secondary hover:text-text-primary hover:border-text-secondary transition-colors"
+      >
+        {t("cancelSubscription")}
+      </button>
+      {error && !confirmCancelOpen && (
+        <p className="text-xs text-red-400 mt-2">{error}</p>
+      )}
+
+      {confirmCancelOpen && (
+        <Modal
+          open={confirmCancelOpen}
+          onClose={() => setConfirmCancelOpen(false)}
+          ariaLabel={t("cancelConfirmTitle")}
+        >
+          <h3 className="font-display text-lg font-semibold text-text-primary mb-2">
+            {t("cancelConfirmTitle")}
+          </h3>
+          <p className="text-sm text-text-secondary mb-3">
+            {t("cancelConfirmDescription")}
+          </p>
+          <ul className="text-xs text-text-muted list-disc list-inside space-y-1 mb-3">
+            <li>{t("cancelConfirmBullet1")}</li>
+            <li>{t("cancelConfirmBullet2")}</li>
+            <li>{t("cancelConfirmBullet3")}</li>
+          </ul>
+          {/* Bug 37b: 60-day retention warning. Distinct paragraph so it
+              reads as a separate, more serious commitment than the
+              regular bullets above. */}
+          <div className="text-xs text-amber-400 bg-amber-400/10 border border-amber-400/20 rounded-lg px-3 py-2 mb-5">
+            {t("cancelConfirmRetentionWarning")}
           </div>
-        </div>
+
+          {error && (
+            <div className="text-xs text-red-400 bg-red-400/10 border border-red-400/20 rounded-lg px-3 py-2 mb-3">
+              {error}
+            </div>
+          )}
+
+          <div className="flex justify-end gap-2">
+            <button
+              type="button"
+              onClick={() => setConfirmCancelOpen(false)}
+              disabled={submitting}
+              className="text-sm px-4 py-2 rounded-lg border border-border text-text-secondary hover:text-text-primary transition-colors"
+            >
+              {tCommon("cancel")}
+            </button>
+            <button
+              type="button"
+              onClick={cancel}
+              disabled={submitting}
+              className="text-sm px-4 py-2 rounded-lg bg-amber-500 text-white hover:bg-amber-600 transition-colors disabled:opacity-50"
+            >
+              {submitting
+                ? tCommon("loading")
+                : t("cancelSubscriptionConfirm")}
+            </button>
+          </div>
+        </Modal>
       )}
     </div>
   );

src/components/ui/modal.tsx

@@ -0,0 +1,89 @@
+"use client";
+
+import { useEffect, useRef } from "react";
+import { createPortal } from "react-dom";
+
+interface Props {
+  open: boolean;
+  /** Called when user clicks the backdrop or presses Escape. */
+  onClose: () => void;
+  children: React.ReactNode;
+  /**
+   * ARIA label fallback when no labelled element exists inside.
+   * Optional; if you have a heading inside the modal with id, set
+   * `aria-labelledby` on a wrapper instead.
+   */
+  ariaLabel?: string;
+}
+
+/**
+ * Portal-based modal.
+ *
+ * Why a portal
+ * ------------
+ * `position: fixed` becomes positioned relative to a transformed
+ * ancestor's containing block, not the viewport, when ANY ancestor
+ * has a `transform`, `perspective`, or `filter` applied. Our
+ * `animate-in` utility sets `transform: translateY(0)` on a lot of
+ * dashboard/tenant-detail containers (because of the fade-up
+ * animation, which uses `animation-fill-mode: both` to keep the
+ * transform on after the animation finishes). That broke modals
+ * rendered as in-place children — they centred to the panel they
+ * lived in, not to the page.
+ *
+ * Rendering at `document.body` via `createPortal` escapes every
+ * containing-block ancestor and gives us true viewport coordinates.
+ *
+ * UX details
+ * ----------
+ * - Backdrop click triggers `onClose`. (Bubbling check: only fires
+ *   when the click target IS the backdrop, not the panel inside.)
+ * - Escape key triggers `onClose`. Standard modal expectation.
+ * - `body` overflow is locked while open so background content
+ *   doesn't scroll behind the modal.
+ * - Renders nothing on first paint server-side, then mounts on
+ *   client. `useEffect` gating ensures `document.body` is available;
+ *   without it Next.js SSR would throw on `document` reference.
+ */
+export function Modal({ open, onClose, children, ariaLabel }: Props) {
+  const closeRef = useRef(onClose);
+  closeRef.current = onClose;
+
+  useEffect(() => {
+    if (!open) return;
+
+    // Lock background scroll. Restore on unmount/close.
+    const previousOverflow = document.body.style.overflow;
+    document.body.style.overflow = "hidden";
+
+    const onKey = (e: KeyboardEvent) => {
+      if (e.key === "Escape") closeRef.current();
+    };
+    window.addEventListener("keydown", onKey);
+
+    return () => {
+      document.body.style.overflow = previousOverflow;
+      window.removeEventListener("keydown", onKey);
+    };
+  }, [open]);
+
+  if (!open) return null;
+  if (typeof document === "undefined") return null;
+
+  return createPortal(
+    <div
+      role="dialog"
+      aria-modal="true"
+      aria-label={ariaLabel}
+      className="fixed inset-0 z-50 flex items-center justify-center p-4 bg-black/60 backdrop-blur-sm"
+      onClick={(e) => {
+        if (e.target === e.currentTarget) onClose();
+      }}
+    >
+      <div className="bg-surface-1 border border-border rounded-xl p-6 max-w-md w-full max-h-[90vh] overflow-y-auto">
+        {children}
+      </div>
+    </div>,
+    document.body
+  );
+}

src/components/ui/warning-badge.tsx

@@ -0,0 +1,118 @@
+"use client";
+
+import { useTranslations } from "next-intl";
+
+/**
+ * Tenant warning shape received from the operator's status.warnings.
+ * Mirror of the operator's `TenantWarning` type. See
+ * pieced-operator/api/v1alpha1/piecedtenant_types.go.
+ */
+export interface TenantWarning {
+  source: string;
+  reason?: string;
+  message?: string;
+  since?: string;
+}
+
+interface Props {
+  warnings: TenantWarning[];
+}
+
+/**
+ * Renders a small amber warning badge if there are any non-fatal
+ * warnings on the tenant. The badge sits visually next to the phase
+ * StatusBadge — they're separate concepts (phase = lifecycle, warnings
+ * = observed sub-issues) and may both be present at once (e.g. tenant
+ * is `Ready` but has a SkillPacksReady=False warning).
+ *
+ * Hover/focus reveals the warning detail. We don't truncate the message
+ * inside the tooltip; OCI/CRD condition messages tend to be short and
+ * include the actionable detail (which skill, which secret, which
+ * resolver). If a future warning source has a 5-line stacktrace as a
+ * message we'll need a different treatment; cross that bridge then.
+ *
+ * Returns null when there are no warnings — keep render-call sites
+ * simple, they don't have to gate on length themselves.
+ */
+export function WarningBadge({ warnings }: Props) {
+  const t = useTranslations("warnings");
+  if (!warnings || warnings.length === 0) return null;
+
+  const tooltipLabel = (() => {
+    try {
+      return warnings.length === 1
+        ? t("oneTooltip")
+        : t("manyTooltip", { count: warnings.length });
+    } catch {
+      return warnings.length === 1
+        ? "1 warning"
+        : `${warnings.length} warnings`;
+    }
+  })();
+
+  return (
+    <span className="relative group inline-flex">
+      <button
+        type="button"
+        // Button is non-actionable in itself — it exists purely to get
+        // keyboard focus for screen readers and keyboard users, so the
+        // tooltip isn't pointer-only. `aria-label` carries the summary;
+        // the full content is in the tooltip below for sighted users.
+        aria-label={tooltipLabel}
+        className="inline-flex items-center gap-1 rounded-full border border-amber-500/30 bg-amber-500/10 px-2 py-0.5 text-xs font-medium text-amber-400 hover:bg-amber-500/20 focus:outline-none focus:ring-1 focus:ring-amber-400 cursor-help"
+        // No onClick — this is informational, not actionable. Pure
+        // hover/focus widget. tabIndex defaults to 0 for buttons.
+      >
+        <svg
+          viewBox="0 0 24 24"
+          width={12}
+          height={12}
+          fill="none"
+          stroke="currentColor"
+          strokeWidth={2}
+          strokeLinecap="round"
+          strokeLinejoin="round"
+          aria-hidden="true"
+        >
+          <path d="M12 9v4" />
+          <path d="M12 17h.01" />
+          <path d="M10.29 3.86 1.82 18a2 2 0 0 0 1.71 3h16.94a2 2 0 0 0 1.71-3L13.71 3.86a2 2 0 0 0-3.42 0Z" />
+        </svg>
+        <span>{warnings.length}</span>
+      </button>
+
+      {/*
+        Tooltip. Hidden by default; shown on hover OR focus of the
+        sibling button. Positioned below-right so it doesn't collide with
+        the StatusBadge that typically sits left of this. Constrained
+        width so long messages wrap.
+        z-50 keeps it above table rows / cards.
+      */}
+      <div
+        role="tooltip"
+        className="invisible group-hover:visible group-focus-within:visible absolute left-0 top-full mt-1 z-50 w-72 rounded-lg border border-border bg-surface-1 p-3 shadow-lg text-left"
+      >
+        <div className="text-[10px] uppercase tracking-wider text-text-muted mb-2">
+          {tooltipLabel}
+        </div>
+        <ul className="space-y-2">
+          {warnings.map((w, i) => (
+            <li key={i} className="text-xs">
+              <div className="font-mono text-amber-400 break-all">
+                {w.source}
+              </div>
+              {w.reason && (
+                <div className="text-text-secondary">{w.reason}</div>
+              )}
+              {w.message && (
+                <div className="text-text-secondary mt-0.5 break-words">
+                  {w.message}
+                </div>
+              )}
+            </li>
+          ))}
+        </ul>
+      </div>
+    </span>
+  );
+}

src/lib/db.ts

@@ -63,9 +63,14 @@ const MIGRATION_SQL = `
   CREATE INDEX IF NOT EXISTS idx_tenant_requests_status ON tenant_requests(status);
   CREATE INDEX IF NOT EXISTS idx_tenant_requests_org_id ON tenant_requests(zitadel_org_id);
   CREATE INDEX IF NOT EXISTS idx_tenant_requests_org_status ON tenant_requests(zitadel_org_id, status);
-  CREATE UNIQUE INDEX IF NOT EXISTS uniq_tenant_requests_tenant_name
-    ON tenant_requests(tenant_name)
-    WHERE tenant_name IS NOT NULL;
+  -- Note: the unique constraint on tenant_name is NOT created here.
+  -- Pre-Bug-37 we had a non-partial UNIQUE on tenant_name, which is
+  -- incompatible with resume requests (same tenant_name, different
+  -- request_type). The new partial unique indexes are created
+  -- further down in the migration block, after the request_type
+  -- column has been added and backfilled. This bootstrap section
+  -- only creates indexes that are safe regardless of request_type
+  -- semantics.
 
   -- Idempotent column adds for existing databases
   ALTER TABLE tenant_requests ADD COLUMN IF NOT EXISTS encrypted_secrets BYTEA;
@@ -78,6 +83,42 @@ const MIGRATION_SQL = `
   -- is only meaningful for rejected and cancelled rows.
   ALTER TABLE tenant_requests ADD COLUMN IF NOT EXISTS dismissed_at TIMESTAMPTZ;
 
+  -- Bug 37a: resume requests use the same table as provision requests so
+  -- the customer dashboard and admin queue share rendering. Discriminator
+  -- is request_type. Default 'provision' on backfill keeps existing rows
+  -- working without explicit migration.
+  --
+  -- Resume rows have:
+  --   request_type     = 'resume'
+  --   tenant_name      = the existing tenant being requested for reactivation
+  --   zitadel_org_id   = the org owning that tenant
+  --   zitadel_user_id  = the requesting customer
+  --   status           = pending → approved/rejected (or cancelled by customer)
+  --   most provision-only fields (packages, billing_address, etc.) are NULL
+  ALTER TABLE tenant_requests ADD COLUMN IF NOT EXISTS request_type TEXT NOT NULL DEFAULT 'provision';
+  -- Constrain to the known set so a future code change can't accidentally
+  -- write a third type without first widening this constraint.
+  DO $$ BEGIN
+    ALTER TABLE tenant_requests ADD CONSTRAINT tenant_requests_request_type_check
+      CHECK (request_type IN ('provision', 'resume'));
+  EXCEPTION WHEN duplicate_object THEN NULL;
+  END $$;
+
+  -- Tenant_name uniqueness was originally meant for "one tenant CR per
+  -- approved provision request". Resume requests reuse a tenant_name,
+  -- so the uniqueness must now be scoped to provision rows only.
+  DROP INDEX IF EXISTS uniq_tenant_requests_tenant_name;
+  CREATE UNIQUE INDEX IF NOT EXISTS uniq_tenant_requests_tenant_name_provision
+    ON tenant_requests(tenant_name)
+    WHERE tenant_name IS NOT NULL AND request_type = 'provision';
+
+  -- Only one pending resume request per tenant at a time. Otherwise a
+  -- customer could spam-create resume requests (the admin queue would
+  -- bloat) or two admins might race on approving duplicates.
+  CREATE UNIQUE INDEX IF NOT EXISTS uniq_tenant_requests_pending_resume
+    ON tenant_requests(tenant_name)
+    WHERE tenant_name IS NOT NULL AND request_type = 'resume' AND status = 'pending';
+
   -- Slice 3: drop the legacy 1-org-1-request constraint if it exists
   ALTER TABLE tenant_requests DROP CONSTRAINT IF EXISTS tenant_requests_zitadel_org_id_key;
 
@@ -381,6 +422,111 @@ export async function clearEncryptedSecrets(requestId: string): Promise<void> {
  * Caller is responsible for verifying the row belongs to the user's
  * org before calling.
  */
+/**
+ * Create a resume request (Bug 37a). Used when an owner of a suspended
+ * tenant wants to reactivate it. Resume is admin-gated — the request
+ * sits as `pending` until a platform admin approves or rejects it.
+ *
+ * Tenant-name uniqueness is enforced for `pending` resume rows by a
+ * partial unique index, so a customer can't spam the queue with
+ * duplicate resume requests for the same tenant. The DB throws a
+ * unique-violation if they try; callers should catch that and translate
+ * to a 409.
+ *
+ * Why this lives in tenant_requests instead of a separate table:
+ *   - the lifecycle is identical (pending → approved/rejected, plus
+ *     customer-side cancel and dismiss-after-terminal)
+ *   - the customer dashboard renders pending+resume cards from the
+ *     same `listActiveTenantRequestsByOrgId` query — adding a separate
+ *     table would mean two queries and union-merging in the UI
+ *   - the admin queue likewise treats them uniformly
+ * The cost is a discriminator column (`request_type`) and most
+ * provision-only fields being null on resume rows. That's a tradeoff
+ * I think is worth it.
+ */
+export async function createResumeRequest(params: {
+  tenantName: string;
+  zitadelOrgId: string;
+  zitadelUserId: string;
+  contactName: string;
+  contactEmail: string;
+  // Provision-only fields default sensibly. company_name + agent_name
+  // are NOT NULL in the original schema; we copy them from the existing
+  // tenant request for traceability rather than storing dummy values.
+  companyName: string;
+  agentName: string;
+}): Promise<TenantRequest> {
+  await ensureSchema();
+  const result = await getPool().query(
+    `INSERT INTO tenant_requests (
+        zitadel_org_id, zitadel_user_id, company_name,
+        contact_name, contact_email, agent_name,
+        tenant_name, request_type, status
+     ) VALUES ($1, $2, $3, $4, $5, $6, $7, 'resume', 'pending')
+     RETURNING *`,
+    [
+      params.zitadelOrgId,
+      params.zitadelUserId,
+      params.companyName,
+      params.contactName,
+      params.contactEmail,
+      params.agentName,
+      params.tenantName,
+    ]
+  );
+  return mapRow(result.rows[0]);
+}
+
+/**
+ * Get the most recent provision request for a tenant_name. Used by
+ * Bug 37a's resume-request creation to populate company_name and
+ * agent_name (NOT NULL columns) from the original provision row
+ * rather than make up values.
+ *
+ * Returns null when no such row exists — should be impossible in
+ * normal flow (resume requests are only created for already-existing
+ * tenants whose CR was created via approving a provision request),
+ * but the caller should guard against it for safety.
+ */
+export async function getTenantRequestByTenantName(
+  tenantName: string
+): Promise<TenantRequest | null> {
+  await ensureSchema();
+  const result = await getPool().query(
+    `SELECT * FROM tenant_requests
+      WHERE tenant_name = $1
+        AND request_type = 'provision'
+      ORDER BY created_at DESC
+      LIMIT 1`,
+    [tenantName]
+  );
+  return result.rows.length > 0 ? mapRow(result.rows[0]) : null;
+}
+
+/**
+ * Return the in-flight (pending) resume request for a given tenant, if
+ * any. Used both to gate the customer's "Request reactivation" button
+ * (don't allow a second when one's already pending) and by the admin
+ * UI to navigate from the tenant detail page to the awaiting request.
+ *
+ * Returns null when no pending resume exists. Approved/rejected rows
+ * are never returned — they're terminal.
+ */
+export async function getPendingResumeRequestForTenant(
+  tenantName: string
+): Promise<TenantRequest | null> {
+  await ensureSchema();
+  const result = await getPool().query(
+    `SELECT * FROM tenant_requests
+      WHERE tenant_name = $1
+        AND request_type = 'resume'
+        AND status = 'pending'
+      LIMIT 1`,
+    [tenantName]
+  );
+  return result.rows.length > 0 ? mapRow(result.rows[0]) : null;
+}
+
 export async function dismissTenantRequest(id: string): Promise<void> {
   await ensureSchema();
   await getPool().query(
@@ -498,8 +644,33 @@ export async function deleteTenantRequest(id: string): Promise<void> {
 }
 
 /**
- * Sync provisioning statuses: for all requests with status "provisioning",
- * check if the PiecedTenant CR has reached "Ready" and update to "active".
+ * Reconcile the portal's tenant_requests table against actual cluster
+ * state. Three passes, walking only rows with `tenant_name` set:
+ *
+ *   1. provisioning → active: when a tenant CR's phase reaches Ready
+ *      or Running, the portal flips the row to active so the
+ *      "provisioning…" card transitions into the running tenant view.
+ *
+ *   2. active/provisioning → deleted: when the corresponding CR no
+ *      longer exists in the cluster (404), or is mid-deletion (has
+ *      metadata.deletionTimestamp set), the row gets flipped to
+ *      `deleted`. The DB is otherwise blind to operator-initiated
+ *      deletions — when the 60-day TTL fires (Bug 37b) and the
+ *      operator deletes a suspended tenant, the portal would happily
+ *      keep showing the "Your assistant is ready!" card forever.
+ *      Without this reconciliation the dashboard drifts from reality.
+ *
+ *   3. pending resume → cancelled: when a pending resume request's
+ *      tenant is no longer suspended (admin resumed it directly,
+ *      tenant was deleted, or it was never suspended in the first
+ *      place), the request is moot. Flip to 'cancelled' so the
+ *      pending-resume unique index releases for any future genuine
+ *      resume request. We pick `cancelled` over `rejected` because
+ *      the customer didn't do anything wrong — circumstances just
+ *      changed.
+ *
+ * Errors are tolerated per-row: a transient API hiccup on one tenant
+ * shouldn't fail the whole sweep. Skipped rows get retried next call.
  *
  * Slice 3 note: with multi-tenant per org, this iterates each row
  * individually (keyed by its own tenant_name), so multiple in-flight
@@ -507,24 +678,78 @@ export async function deleteTenantRequest(id: string): Promise<void> {
  */
 export async function syncProvisioningStatuses(): Promise<void> {
   await ensureSchema();
+  // Active+provisioning rows: status reflects "the tenant should
+  // exist and be running".
+  // Pending resume rows: status reflects "the tenant is suspended,
+  // awaiting reactivation".
+  // Both need cluster-side validation; we fetch them in one query
+  // and dispatch on (status, request_type).
   const result = await getPool().query<TenantRequest>(
-    "SELECT * FROM tenant_requests WHERE status = 'provisioning'"
+    `SELECT * FROM tenant_requests
+       WHERE tenant_name IS NOT NULL
+         AND (
+           status IN ('provisioning', 'active')
+           OR (status = 'pending' AND request_type = 'resume')
+         )`
   );
 
   for (const row of result.rows) {
     const mapped = mapRow(row);
     if (!mapped.tenantName) continue;
 
+    let tenant: Awaited<ReturnType<typeof getTenant>> = null;
     try {
-      const tenant = await getTenant(mapped.tenantName);
-      if (
-        tenant?.status?.phase === "Ready" ||
-        tenant?.status?.phase === "Running"
-      ) {
-        await updateTenantRequestStatus(mapped.id, "active");
-      }
+      tenant = await getTenant(mapped.tenantName);
     } catch {
-      // Tenant might not exist yet — skip
+      // Transient API error — skip this row, retry on next sweep.
+      continue;
+    }
+
+    // Pending resume request: validity hinges on tenant being suspended.
+    if (
+      mapped.status === "pending" &&
+      mapped.requestType === "resume"
+    ) {
+      // Tenant doesn't exist or is being deleted: cancel the resume
+      // request (it can never be fulfilled). Don't fall through to
+      // the "deleted" branch below — that would also flip the
+      // provision row, which is the right thing for a CR-level
+      // deletion but we want this resume row specifically resolved
+      // here.
+      if (!tenant || tenant.metadata.deletionTimestamp) {
+        await updateTenantRequestStatus(mapped.id, "cancelled");
+        continue;
+      }
+      // Tenant is no longer suspended: the request is moot.
+      // Cancel it (the customer didn't do anything wrong; the
+      // condition the request was about no longer applies).
+      if (!tenant.spec.suspend) {
+        await updateTenantRequestStatus(mapped.id, "cancelled");
+        continue;
+      }
+      // Tenant still suspended, request still relevant. Leave as-is.
+      continue;
+    }
+
+    // Active or provisioning row: CR gone, or mid-deletion. Flip the
+    // row to 'deleted'. `markTenantRequestDeletedByTenantName` flips
+    // every row with this tenant_name (provision + any resume rows),
+    // which is the right thing for a CR-level deletion.
+    if (!tenant || tenant.metadata.deletionTimestamp) {
+      await markTenantRequestDeletedByTenantName(mapped.tenantName);
+      continue;
+    }
+
+    // CR exists and is healthy. Promote provisioning → active when
+    // the operator reports the tenant has reached steady state.
+    // Keep `active` rows on `active` regardless of phase — a
+    // temporarily-Reconfiguring tenant is still active from the
+    // portal's billing/visibility perspective.
+    if (
+      mapped.status === "provisioning" &&
+      (tenant.status?.phase === "Ready" || tenant.status?.phase === "Running")
+    ) {
+      await updateTenantRequestStatus(mapped.id, "active");
     }
   }
 }
@@ -555,6 +780,9 @@ function mapRow(row: any): TenantRequest {
     isPersonal: row.is_personal ?? false,
     dismissedAt:
       row.dismissed_at?.toISOString?.() ?? row.dismissed_at ?? null,
+    requestType: (row.request_type ?? "provision") as
+      | "provision"
+      | "resume",
     createdAt: row.created_at?.toISOString?.() ?? row.created_at,
     updatedAt: row.updated_at?.toISOString?.() ?? row.updated_at,
   };

src/lib/k8s.ts

@@ -130,3 +130,46 @@ export async function patchTenantSpec(
   }
   return res.json() as Promise<PiecedTenant>;
 }
+
+/**
+ * Set or clear an annotation on a PiecedTenant CR.
+ *
+ * Pass `value=null` to remove the annotation. K8s merge-patch removes
+ * a key when its value is null in the patch — that's exactly the
+ * semantic we want.
+ *
+ * Used by the resume-request flow (Bug 37a): the portal sets
+ * `pieced.ch/resume-request-pending` when a customer creates a
+ * resume request, and clears it when the request transitions to a
+ * terminal state. The operator reads this annotation to pause its
+ * 60-day deletion timer while a resume request is in flight.
+ *
+ * Annotations are namespaced informally — we use `pieced.ch/...` for
+ * everything we own, mirroring the labels.
+ */
+export async function setTenantAnnotation(
+  name: string,
+  key: string,
+  value: string | null
+): Promise<PiecedTenant> {
+  const url = `${getBaseUrl()}/apis/${API_VERSION}/${PLURAL}/${name}`;
+  const res = await fetch(url, {
+    method: "PATCH",
+    headers: {
+      Accept: "application/json",
+      "Content-Type": "application/merge-patch+json",
+      ...getAuthHeaders(),
+    },
+    body: JSON.stringify({
+      metadata: { annotations: { [key]: value } },
+    }),
+  });
+
+  if (!res.ok) {
+    const text = await res.text();
+    const err = new Error(`K8s annotate /${name}: ${res.status} ${text}`);
+    (err as any).statusCode = res.status;
+    throw err;
+  }
+  return res.json() as Promise<PiecedTenant>;
+}

src/lib/litellm.ts

@@ -32,12 +32,43 @@ export async function getTeamSpendLogs(
   return litellmFetch(`/global/spend/logs?${params}`);
 }
 
+/**
+ * Fetch one page of spend logs for a team, optionally narrowed to a
+ * single virtual key by alias.
+ *
+ * Slice 2 / Bug 19 context
+ * ------------------------
+ * Teams in LiteLLM are now org-scoped (one team per org), and each
+ * tenant in the org has its own virtual key with `key_alias = tenant
+ * CR name`. Without `keyAlias`, this returns the full team's spend —
+ * which mingles every tenant in the org. The portal's per-tenant
+ * usage view passes `keyAlias` to filter server-side via LiteLLM's
+ * native `key_alias` query param. Confirmed available on the
+ * `/spend/logs/v2` endpoint via OpenAPI introspection — no need to
+ * page-and-post-filter as the previous slice did.
+ *
+ * Why this matters
+ * ----------------
+ * Previous implementation fetched all team pages, then post-filtered
+ * by alias in JS. Two problems: (1) at any reasonable scale this is
+ * O(team_total) memory per request even when only one tenant's data
+ * is needed; (2) more importantly, when called from the customer
+ * dashboard without an explicit alias, the route's "pick the first
+ * visible tenant" fallback meant both Acme tenants showed identical
+ * numbers — the alias used was always the first tenant in the
+ * visible list, regardless of which tenant page was being viewed.
+ *
+ * The route layer above is responsible for resolving the tenant
+ * identity correctly and passing the right alias here. This
+ * function's only job is to pass it through to LiteLLM.
+ */
 export async function getTeamSpendLogsV2(
   teamId: string,
   startDate: string,
   endDate: string,
   page: number = 1,
-  pageSize: number = 100
+  pageSize: number = 100,
+  keyAlias?: string | null
 ) {
   const params = new URLSearchParams({
     team_id: teamId,
@@ -46,6 +77,9 @@ export async function getTeamSpendLogsV2(
     page: String(page),
     page_size: String(pageSize),
   });
+  if (keyAlias) {
+    params.set("key_alias", keyAlias);
+  }
   return litellmFetch(`/spend/logs/v2?${params}`);
 }
 

src/messages/de.json

@@ -157,7 +157,16 @@
     "cancelConfirmBullet3": "Rechnungsdaten bleiben gespeichert",
     "subscriptionUpdateFailed": "Abonnement konnte nicht aktualisiert werden.",
     "suspendedTitle": "Abonnement gekündigt",
-    "suspendedDescription": "Ihr Assistent ist pausiert. Konfiguration und Daten bleiben erhalten. Verwenden Sie die Reaktivierungs-Schaltfläche unten auf dieser Seite, um ihn wieder online zu bringen."
+    "suspendedDescription": "Ihr Assistent ist pausiert. Konfiguration und Daten bleiben erhalten. Verwenden Sie die Reaktivierungs-Schaltfläche unten auf dieser Seite, um ihn wieder online zu bringen.",
+    "requestReactivation": "Reaktivierung anfragen",
+    "requestReactivationConfirmTitle": "Reaktivierung anfragen?",
+    "requestReactivationConfirmDescription": "Ein Administrator prüft Ihre Anfrage und reaktiviert Ihren Tenant. Sie erhalten eine E-Mail, sobald die Anfrage genehmigt wurde.",
+    "requestReactivationConfirm": "Anfrage senden",
+    "cancelResumeRequest": "Anfrage stornieren",
+    "resumeRequestPendingTitle": "Reaktivierungsanfrage ausstehend",
+    "resumeRequestPendingDescription": "Eingereicht {when}. Ein Administrator wird die Anfrage in Kürze prüfen.",
+    "resumeRequestPendingNoteAdmin": "Ein Inhaber hat eine Reaktivierung angefragt; Sie können direkt oben fortfahren oder die Anfrage in der Admin-Warteschlange bearbeiten.",
+    "cancelConfirmRetentionWarning": "Ihre Daten bleiben nach der Kündigung 60 Tage lang erhalten. Danach werden alle Tenant-Daten – Konfiguration, Geheimnisse, Konversationen und Dateien – endgültig gelöscht."
   },
   "usage": {
     "inputTokens": "Input-Tokens",
@@ -297,7 +306,9 @@
     "loadingHealth": "Statusdaten werden geladen…",
     "statusHealthy": "OK",
     "statusDown": "Ausgefallen",
-    "spendChf": "Kosten (CHF)"
+    "spendChf": "Kosten (CHF)",
+    "resumeRequestBadge": "Wieder",
+    "resumeRequestTooltip": "Reaktivierungsanfrage für einen bestehenden Tenant. Bei Genehmigung wird der Tenant wieder aktiviert; keine Provisionierung läuft."
   },
   "channelUsers": {
     "title": "Autorisierte Benutzer",
@@ -361,5 +372,9 @@
     "Error": "Fehler",
     "Deleting": "Wird gelöscht",
     "Reconfiguring": "Wird neu konfiguriert"
+  },
+  "warnings": {
+    "oneTooltip": "1 Warnung",
+    "manyTooltip": "{count} Warnungen"
   }
 }

src/messages/en.json

@@ -157,7 +157,16 @@
     "cancelConfirmBullet3": "Billing information is kept on file",
     "subscriptionUpdateFailed": "Could not update subscription.",
     "suspendedTitle": "Subscription cancelled",
-    "suspendedDescription": "Your assistant is paused. Configuration and data are preserved. Use the Resume control at the bottom of this page to bring it back online."
+    "suspendedDescription": "Your assistant is paused. Configuration and data are preserved. Use the Resume control at the bottom of this page to bring it back online.",
+    "requestReactivation": "Request reactivation",
+    "requestReactivationConfirmTitle": "Request reactivation?",
+    "requestReactivationConfirmDescription": "An administrator will review your request and reactivate your tenant. You'll be notified by email once it's approved.",
+    "requestReactivationConfirm": "Submit request",
+    "cancelResumeRequest": "Cancel request",
+    "resumeRequestPendingTitle": "Reactivation request pending",
+    "resumeRequestPendingDescription": "Submitted {when}. An administrator will review it shortly.",
+    "resumeRequestPendingNoteAdmin": "An owner has requested reactivation; you can resume directly above or process the request from the admin queue.",
+    "cancelConfirmRetentionWarning": "Your data is preserved for 60 days after cancellation. After that, all tenant data — configuration, secrets, conversations, and files — will be permanently deleted."
   },
   "usage": {
     "inputTokens": "Input Tokens",
@@ -297,7 +306,9 @@
     "loadingHealth": "Loading health data…",
     "statusHealthy": "Healthy",
     "statusDown": "Down",
-    "spendChf": "Spend (CHF)"
+    "spendChf": "Spend (CHF)",
+    "resumeRequestBadge": "Resume",
+    "resumeRequestTooltip": "Reactivation request for an existing tenant. Approving will un-suspend the tenant; no provisioning runs."
   },
   "channelUsers": {
     "title": "Authorized Users",
@@ -361,5 +372,9 @@
     "Error": "Error",
     "Deleting": "Deleting",
     "Reconfiguring": "Reconfiguring"
+  },
+  "warnings": {
+    "oneTooltip": "1 warning",
+    "manyTooltip": "{count} warnings"
   }
 }

src/messages/fr.json

@@ -157,7 +157,16 @@
     "cancelConfirmBullet3": "Les informations de facturation sont conservées",
     "subscriptionUpdateFailed": "Impossible de mettre à jour l'abonnement.",
     "suspendedTitle": "Abonnement annulé",
-    "suspendedDescription": "Votre assistant est en pause. La configuration et les données sont préservées. Utilisez le contrôle Reprendre en bas de cette page pour le remettre en ligne."
+    "suspendedDescription": "Votre assistant est en pause. La configuration et les données sont préservées. Utilisez le contrôle Reprendre en bas de cette page pour le remettre en ligne.",
+    "requestReactivation": "Demander la réactivation",
+    "requestReactivationConfirmTitle": "Demander la réactivation ?",
+    "requestReactivationConfirmDescription": "Un administrateur examinera votre demande et réactivera votre locataire. Vous recevrez un e-mail dès que la demande sera approuvée.",
+    "requestReactivationConfirm": "Envoyer la demande",
+    "cancelResumeRequest": "Annuler la demande",
+    "resumeRequestPendingTitle": "Demande de réactivation en attente",
+    "resumeRequestPendingDescription": "Soumise {when}. Un administrateur l'examinera sous peu.",
+    "resumeRequestPendingNoteAdmin": "Un propriétaire a demandé la réactivation ; vous pouvez reprendre directement ci-dessus ou traiter la demande depuis la file d'attente d'administration.",
+    "cancelConfirmRetentionWarning": "Vos données sont conservées pendant 60 jours après l'annulation. Passé ce délai, toutes les données du locataire — configuration, secrets, conversations et fichiers — seront définitivement supprimées."
   },
   "usage": {
     "inputTokens": "Tokens d'entrée",
@@ -297,7 +306,9 @@
     "loadingHealth": "Chargement des données de santé…",
     "statusHealthy": "OK",
     "statusDown": "Hors service",
-    "spendChf": "Coûts (CHF)"
+    "spendChf": "Coûts (CHF)",
+    "resumeRequestBadge": "Reprise",
+    "resumeRequestTooltip": "Demande de réactivation d'un locataire existant. L'approbation le réactivera ; aucun provisionnement ne s'exécute."
   },
   "channelUsers": {
     "title": "Utilisateurs autorisés",
@@ -361,5 +372,9 @@
     "Error": "Erreur",
     "Deleting": "Suppression",
     "Reconfiguring": "Reconfiguration"
+  },
+  "warnings": {
+    "oneTooltip": "1 avertissement",
+    "manyTooltip": "{count} avertissements"
   }
 }

src/messages/it.json

@@ -157,7 +157,16 @@
     "cancelConfirmBullet3": "Le informazioni di fatturazione sono mantenute",
     "subscriptionUpdateFailed": "Impossibile aggiornare l'abbonamento.",
     "suspendedTitle": "Abbonamento annullato",
-    "suspendedDescription": "Il suo assistente è in pausa. Configurazione e dati sono preservati. Usi il controllo Riprendi in fondo a questa pagina per riportarlo online."
+    "suspendedDescription": "Il suo assistente è in pausa. Configurazione e dati sono preservati. Usi il controllo Riprendi in fondo a questa pagina per riportarlo online.",
+    "requestReactivation": "Richiedi riattivazione",
+    "requestReactivationConfirmTitle": "Richiedere la riattivazione?",
+    "requestReactivationConfirmDescription": "Un amministratore esaminerà la tua richiesta e riattiverà il tuo tenant. Riceverai un'email non appena la richiesta sarà approvata.",
+    "requestReactivationConfirm": "Invia richiesta",
+    "cancelResumeRequest": "Annulla richiesta",
+    "resumeRequestPendingTitle": "Richiesta di riattivazione in sospeso",
+    "resumeRequestPendingDescription": "Inviata {when}. Un amministratore la esaminerà a breve.",
+    "resumeRequestPendingNoteAdmin": "Un proprietario ha richiesto la riattivazione; puoi riprendere direttamente sopra o elaborare la richiesta dalla coda di amministrazione.",
+    "cancelConfirmRetentionWarning": "I tuoi dati sono conservati per 60 giorni dopo l'annullamento. Trascorso tale periodo, tutti i dati del tenant — configurazione, segreti, conversazioni e file — verranno eliminati definitivamente."
   },
   "usage": {
     "inputTokens": "Token di input",
@@ -297,7 +306,9 @@
     "loadingHealth": "Caricamento dati di stato…",
     "statusHealthy": "OK",
     "statusDown": "Non disponibile",
-    "spendChf": "Costi (CHF)"
+    "spendChf": "Costi (CHF)",
+    "resumeRequestBadge": "Ripresa",
+    "resumeRequestTooltip": "Richiesta di riattivazione di un tenant esistente. L'approvazione lo riattiverà; non viene eseguito alcun provisioning."
   },
   "channelUsers": {
     "title": "Utenti autorizzati",
@@ -361,5 +372,9 @@
     "Error": "Errore",
     "Deleting": "Eliminazione",
     "Reconfiguring": "Riconfigurazione"
+  },
+  "warnings": {
+    "oneTooltip": "1 avviso",
+    "manyTooltip": "{count} avvisi"
   }
 }

src/types/index.ts

@@ -103,6 +103,21 @@ export interface PiecedTenantStatus {
   litellmKeyAlias?: string;
   tenantNamespace?: string;
   enabledPackages?: string[];
+  /**
+   * Non-fatal issues from downstream resources surfaced by the operator
+   * (e.g. an OpenClawInstance sub-condition reporting failure). The
+   * tenant is still usable — these are informational, rendered as a
+   * warning badge alongside the phase.
+   *
+   * `source` is "<Kind>/<ConditionType>" e.g. "OpenClawInstance/SkillPacksReady".
+   * `message` is shown in the tooltip when the user hovers the badge.
+   */
+  warnings?: Array<{
+    source: string;
+    reason?: string;
+    message?: string;
+    since?: string;
+  }>;
   conditions?: Array<{
     type: string;
     status: string;
@@ -119,6 +134,15 @@ export interface PiecedTenant {
     name: string;
     namespace?: string;
     creationTimestamp?: string;
+    /**
+     * Set by the API server when something issues a Delete on the CR.
+     * The CR continues to exist while finalizers run cleanup; once
+     * they all remove themselves, the API server permanently removes
+     * the CR. Used by the portal's status sync to detect tenants
+     * being torn down — the customer should see "Deleted" rather
+     * than "Ready" while the cleanup runs.
+     */
+    deletionTimestamp?: string;
     labels?: Record<string, string>;
     annotations?: Record<string, string>;
   };
@@ -212,6 +236,19 @@ export interface TenantRequest {
    * login). Always null for non-rejected statuses.
    */
   dismissedAt?: string | null;
+  /**
+   * Bug 37a: discriminator between provision (initial tenant creation,
+   * the original purpose of this table) and resume (admin-gated
+   * reactivation of a suspended tenant). Default 'provision' for all
+   * pre-existing rows; resume rows have most provision fields null
+   * but tenant_name set to the tenant being requested.
+   *
+   * Optional on the TS type so provision-only callers (like the
+   * onboarding wizard's create flow) don't need to know about resume
+   * requests. The DB column is NOT NULL DEFAULT 'provision', so rows
+   * loaded via `mapRow` always have a value populated.
+   */
+  requestType?: "provision" | "resume";
   createdAt: string;
   updatedAt: string;
 }