Add per-key model blacklist

apps/admin/src/lib/api/keys.test.ts

@@ -29,6 +29,7 @@ function summaryRow(
     role: 'user',
     allowed_lanes: null,
     allow_custom_model: false,
+    blocked_models: null,
     allow_fast_mode: false,
     disabled: false,
     rate_limit_rpm: null,
@@ -119,6 +120,7 @@ describe('keys api client', () => {
       role: 'user',
       allowed_lanes: ['economy', 'balanced'],
       allow_custom_model: false,
+      blocked_models: ['gpt-4o'],
       allow_fast_mode: true,
     });
 
@@ -129,6 +131,7 @@ describe('keys api client', () => {
     expect(body.role).toBe('user');
     expect(body.allowed_lanes).toEqual(['economy', 'balanced']);
     expect(body.allow_custom_model).toBe(false);
+    expect(body.blocked_models).toEqual(['gpt-4o']);
     expect(body.allow_fast_mode).toBe(true);
     expect(result.key_id).toBe('key_1');
     expect(result.plaintext).toBe('helm_live_SECRET_ONCE');
@@ -150,9 +153,20 @@ describe('keys api client', () => {
     const body = JSON.parse(init.body as string);
     expect(body).not.toHaveProperty('max_lane');
     expect(body).not.toHaveProperty('allowed_lanes');
+    expect(body).not.toHaveProperty('blocked_models');
     expect(body.role).toBe('user');
   });
 
+  it('listKeys surfaces blocked model ids as null-or-array', async () => {
+    const rows = [summaryRow('k1'), summaryRow('k2', { blocked_models: ['gpt-4o'] })];
+    (fetch as ReturnType<typeof vi.fn>).mockResolvedValue(
+      new Response(JSON.stringify(rows), { status: 200 }),
+    );
+    const keys = await listKeys();
+    expect(keys[0].blocked_models).toBeNull();
+    expect(keys[1].blocked_models).toEqual(['gpt-4o']);
+  });
+
   it('listKeys surfaces per-key rate limits (null = inherit, number = override)', async () => {
     const rows = [summaryRow('k1'), summaryRow('k2', { rate_limit_rpm: 60, rate_limit_tpm: 0 })];
     (fetch as ReturnType<typeof vi.fn>).mockResolvedValue(
@@ -267,6 +281,7 @@ describe('keys api client', () => {
     await updateKey('key_1', {
       allowed_lanes: ['economy', 'balanced'],
       allow_custom_model: true,
+      blocked_models: ['gpt-4o'],
       allow_fast_mode: true,
       rate_limit_rpm: null,
       rate_limit_tpm: 100,
@@ -277,6 +292,7 @@ describe('keys api client', () => {
     const body = JSON.parse(init.body as string);
     expect(body.allowed_lanes).toEqual(['economy', 'balanced']);
     expect(body.allow_custom_model).toBe(true);
+    expect(body.blocked_models).toEqual(['gpt-4o']);
     expect(body.allow_fast_mode).toBe(true);
     expect(body.rate_limit_rpm).toBeNull(); // explicit null = clear
     expect(body.rate_limit_tpm).toBe(100);
@@ -292,6 +308,16 @@ describe('keys api client', () => {
     expect(body.allowed_lanes).toBeNull();
   });
 
+  it('updateKey forwards null to clear the blocked-model blacklist', async () => {
+    (fetch as ReturnType<typeof vi.fn>).mockResolvedValue(
+      new Response(JSON.stringify({ key_id: 'key_1' }), { status: 200 }),
+    );
+    await updateKey('key_1', { blocked_models: null });
+    const [, init] = (fetch as ReturnType<typeof vi.fn>).mock.calls[0];
+    const body = JSON.parse(init.body as string);
+    expect(body.blocked_models).toBeNull();
+  });
+
   it('updateKey rejects on a non-2xx response (404)', async () => {
     (fetch as ReturnType<typeof vi.fn>).mockResolvedValue(
       new Response(JSON.stringify({ error: 'key not found' }), { status: 404 }),

apps/admin/src/lib/api/keys.ts

@@ -17,6 +17,7 @@ export interface ApiKeyView {
   name: string | null; // human-readable label; null = unnamed (cosmetic only)
   allowed_lanes: string[] | null; // lane whitelist (empty/null = any lane)
   allow_custom_model: boolean; // explicit client-model passthrough
+  blocked_models: string[] | null; // exact model ids denied across direct and lane routes
   allow_fast_mode: boolean; // explicit client-requested Fast mode passthrough
   disabled: boolean; // revoked state (soft)
   rate_limit_rpm: number | null; // per-key RPM override; null = inherit system default
@@ -49,6 +50,7 @@ export interface CreateKeyInput {
   name?: string;
   allowed_lanes?: string[];
   allow_custom_model?: boolean;
+  blocked_models?: string[];
   allow_fast_mode?: boolean;
   // Optional per-key rate limits at mint time. Omitted => inherit the system
   // default. 0 => explicitly unlimited for that dimension.
@@ -81,6 +83,7 @@ export interface UpdateKeyInput {
   name?: string | null;
   allowed_lanes?: string[] | null;
   allow_custom_model?: boolean;
+  blocked_models?: string[] | null;
   allow_fast_mode?: boolean;
   rate_limit_rpm?: number | null;
   rate_limit_tpm?: number | null;
@@ -177,6 +180,7 @@ async function asJson<T>(res: Response): Promise<T> {
 // leak a secret even if the server response ever changed shape (defence in depth).
 function normalizeView(raw: Record<string, unknown>): ApiKeyView {
   const allowed = raw.allowed_lanes;
+  const blocked = raw.blocked_models;
   return {
     key_id: String(raw.key_id ?? ''),
     prefix: String(raw.prefix ?? ''),
@@ -186,6 +190,7 @@ function normalizeView(raw: Record<string, unknown>): ApiKeyView {
     name: typeof raw.name === 'string' && raw.name.trim().length > 0 ? raw.name.trim() : null,
     allowed_lanes: Array.isArray(allowed) ? allowed.map(String) : null,
     allow_custom_model: raw.allow_custom_model === true,
+    blocked_models: Array.isArray(blocked) ? blocked.map(String) : null,
     allow_fast_mode: raw.allow_fast_mode === true,
     disabled: raw.disabled === true,
     // null/absent = inherit system default; a finite number (incl. 0) = override.
@@ -219,6 +224,9 @@ function toServerBody(input: CreateKeyInput): Record<string, unknown> {
   if (input.allow_custom_model !== undefined) {
     out.allow_custom_model = input.allow_custom_model;
   }
+  if (input.blocked_models && input.blocked_models.length > 0) {
+    out.blocked_models = input.blocked_models;
+  }
   if (input.allow_fast_mode !== undefined) {
     out.allow_fast_mode = input.allow_fast_mode;
   }

apps/admin/src/lib/components/CreateKeyDialog.svelte

@@ -59,6 +59,7 @@
     const trimmedName = name.trim();
     if (trimmedName.length > 0) input.name = trimmedName;
     if (form.allowedLanes.length > 0) input.allowed_lanes = [...form.allowedLanes];
+    if (form.blockedModels.length > 0) input.blocked_models = [...form.blockedModels];
     // Send a rate limit only when the operator set one; blank => inherit default.
     // `!= null` also catches the `undefined` Svelte 5 gives an emptied number input.
     if (form.rpm != null) input.rate_limit_rpm = form.rpm;
@@ -113,6 +114,7 @@
         name: name.trim().length > 0 ? name.trim() : null,
         allowed_lanes: form.allowedLanes.length > 0 ? [...form.allowedLanes] : null,
         allow_custom_model: form.allowCustomModel,
+        blocked_models: form.blockedModels.length > 0 ? [...form.blockedModels] : null,
         allow_fast_mode: form.allowFastMode,
         disabled: false,
         rate_limit_rpm: form.rpm ?? null,

apps/admin/src/lib/components/CreateKeyDialog.test.ts

@@ -41,13 +41,17 @@ describe('CreateKeyDialog', () => {
     // Tick a subset of lanes; the whitelist is the only lane cap (no max-lane field).
     await fireEvent.click(screen.getByLabelText('economy'));
     await fireEvent.click(screen.getByLabelText('balanced'));
+    await fireEvent.input(screen.getByLabelText('Blocked models'), {
+      target: { value: 'gpt-4o\nanthropic/claude-sonnet-4-6' },
+    });
     await fireEvent.click(screen.getByLabelText('allow client-requested Fast mode'));
     // allow_custom_model defaults to false; leave that checkbox unchecked.
     await fireEvent.click(screen.getByRole('button', { name: /create key/i }));
 
     await waitFor(() => expect(createKey).toHaveBeenCalledTimes(1));
     const input = createKey.mock.calls[0][0];
     expect(input.allowed_lanes).toEqual(['economy', 'balanced']);
+    expect(input.blocked_models).toEqual(['gpt-4o', 'anthropic/claude-sonnet-4-6']);
     expect(input.allow_custom_model).toBe(false);
     expect(input.allow_fast_mode).toBe(true);
   });

apps/admin/src/lib/components/EditKeyDialog.svelte

@@ -46,6 +46,7 @@
       name: name.trim().length > 0 ? name.trim() : null,
       allowed_lanes: form.allowedLanes.length > 0 ? [...form.allowedLanes] : null,
       allow_custom_model: form.allowCustomModel,
+      blocked_models: form.blockedModels.length > 0 ? [...form.blockedModels] : null,
       allow_fast_mode: form.allowFastMode,
       rate_limit_rpm: form.rpm ?? null,
       rate_limit_tpm: form.tpm ?? null,
@@ -68,6 +69,7 @@
         name: patch.name ?? null,
         allowed_lanes: patch.allowed_lanes ?? null,
         allow_custom_model: form.allowCustomModel,
+        blocked_models: patch.blocked_models ?? null,
         allow_fast_mode: form.allowFastMode,
         rate_limit_rpm: patch.rate_limit_rpm ?? null,
         rate_limit_tpm: patch.rate_limit_tpm ?? null,
@@ -128,7 +130,9 @@
         bind:value={name}
       />
       <span class="field-help"
-        >{$t('A label to help you recognize this key later — e.g. the project it belongs to.')}</span
+        >{$t(
+          'A label to help you recognize this key later — e.g. the project it belongs to.',
+        )}</span
       >
     </label>
 

apps/admin/src/lib/components/KeyCapsForm.svelte

@@ -7,6 +7,7 @@
   export type KeyCaps = {
     allowedLanes: string[];
     allowCustomModel: boolean;
+    blockedModels: string[];
     allowFastMode: boolean;
     rpm: number | null;
     tpm: number | null;
@@ -27,6 +28,7 @@
     return {
       allowedLanes: [],
       allowCustomModel: false,
+      blockedModels: [],
       allowFastMode: false,
       rpm: null,
       tpm: null,
@@ -50,6 +52,7 @@
     return {
       allowedLanes: [...(key.allowed_lanes ?? [])],
       allowCustomModel: key.allow_custom_model,
+      blockedModels: [...(key.blocked_models ?? [])],
       allowFastMode: key.allow_fast_mode,
       rpm: key.rate_limit_rpm,
       tpm: key.rate_limit_tpm,
@@ -65,6 +68,18 @@
       memoryThreadSource: key.memory_thread_source,
     };
   }
+
+  export function parseBlockedModels(value: string): string[] {
+    const out: string[] = [];
+    const seen = new Set<string>();
+    for (const raw of value.split(/[\n,;]+/)) {
+      const model = raw.trim();
+      if (model.length === 0 || seen.has(model)) continue;
+      seen.add(model);
+      out.push(model);
+    }
+    return out;
+  }
 </script>
 
 <script lang="ts">
@@ -113,6 +128,10 @@
       : form.allowedLanes.filter((l) => l !== lane);
   }
 
+  function updateBlockedModels(value: string): void {
+    form.blockedModels = parseBlockedModels(value);
+  }
+
   // One-line state recaps shown on the closed section headers.
   const ratesSummary = $derived.by(() => {
     const parts: string[] = [];
@@ -200,6 +219,23 @@
   >
 </div>
 
+<label class="flex flex-col gap-1 text-sm">
+  <span class="field-label">{$t('Blocked models')}</span>
+  <textarea
+    rows="3"
+    aria-label={$t('Blocked models')}
+    placeholder={$t('One model id per line')}
+    class="input min-h-20 resize-y"
+    value={form.blockedModels.join('\n')}
+    oninput={(e) => updateBlockedModels(e.currentTarget.value)}
+  ></textarea>
+  <span class="field-help"
+    >{$t(
+      'Blocks exact model ids for this key across direct model requests and all lane/fallback routes.',
+    )}</span
+  >
+</label>
+
 <div class="flex flex-col gap-1">
   <label class="checkbox-field">
     <input

apps/admin/src/routes/keys/[keyId]/key-detail.test.ts

@@ -55,6 +55,7 @@ function keyView(overrides: Partial<ApiKeyView> = {}): ApiKeyView {
     name: 'Prod backend',
     allowed_lanes: ['balanced'],
     allow_custom_model: false,
+    blocked_models: null,
     allow_fast_mode: false,
     disabled: false,
     rate_limit_rpm: null,

apps/admin/src/routes/keys/keys.test.ts

@@ -30,6 +30,7 @@ function key(keyId: string, overrides: Partial<ApiKeyView> = {}): ApiKeyView {
     name: null,
     allowed_lanes: null,
     allow_custom_model: false,
+    blocked_models: null,
     allow_fast_mode: false,
     disabled: false,
     rate_limit_rpm: null,
@@ -365,13 +366,17 @@ describe('keys page', () => {
     await fireEvent.input(within(dialog).getByLabelText(/requests per minute/i), {
       target: { value: '120' },
     });
+    await fireEvent.input(within(dialog).getByLabelText('Blocked models'), {
+      target: { value: 'gpt-4o' },
+    });
     await fireEvent.click(within(dialog).getByRole('button', { name: /save changes/i }));
     await waitFor(() =>
       expect(updateKey).toHaveBeenCalledWith('k1', {
         // Name untouched in this edit → still unnamed (null), never undefined.
         name: null,
         allowed_lanes: ['economy'],
         allow_custom_model: false,
+        blocked_models: ['gpt-4o'],
         allow_fast_mode: false,
         rate_limit_rpm: 120,
         rate_limit_tpm: null, // untouched → still inherit (null), not undefined

apps/admin/src/routes/memory/memory.test.ts

@@ -91,6 +91,7 @@ function key(keyId: string, overrides: Partial<ApiKeyView> = {}): ApiKeyView {
     name: null,
     allowed_lanes: null,
     allow_custom_model: false,
+    blocked_models: null,
     allow_fast_mode: false,
     disabled: false,
     rate_limit_rpm: null,

apps/admin/src/routes/requests/requests.test.ts

@@ -92,6 +92,7 @@ function apiKey(keyId: string, overrides: Partial<ApiKeyView> = {}): ApiKeyView
     name: null,
     allowed_lanes: null,
     allow_custom_model: false,
+    blocked_models: null,
     allow_fast_mode: false,
     disabled: false,
     rate_limit_rpm: null,

apps/gateway/src/middleware/auth.test.ts

@@ -15,6 +15,7 @@ function record(overrides: Partial<ApiKeyRecord> = {}): ApiKeyRecord {
     name: null,
     allowed_lanes: ["economy", "balanced"],
     allow_custom_model: false,
+    blocked_models: null,
     allow_fast_mode: false,
     disabled: false,
     rate_limit_rpm: null,

apps/gateway/src/middleware/auth.ts

@@ -17,6 +17,7 @@ export interface AuthIdentity {
   caps: {
     allowedLanes: string[] | null;
     allowCustomModel: boolean;
+    blockedModels: string[] | null;
     /** Per-key cap for client-requested Fast mode passthrough. Account-level Fast
      *  mode can still be forced by subscription account settings. */
     allowFastMode: boolean;
@@ -105,6 +106,7 @@ export function authMiddleware(deps: AuthDeps): MiddlewareHandler {
       caps: {
         allowedLanes: record.allowed_lanes,
         allowCustomModel: record.allow_custom_model,
+        blockedModels: record.blocked_models,
         allowFastMode: record.allow_fast_mode,
         rateLimit: { rpm: record.rate_limit_rpm, tpm: record.rate_limit_tpm },
         concurrencyLimit: record.concurrency_limit,