fix(ai-gateway): fit generation routes within Vercel timeout

apps/web/src/app/api/gateway/[...path]/route.ts

@@ -1,3 +1,3 @@
 export { POST } from '@/app/api/openrouter/[...path]/route';
 
-export const maxDuration = 1800;
+export const maxDuration = 800;

apps/web/src/app/api/gateway/audio/transcriptions/route.ts

@@ -1,3 +1,3 @@
 export { POST } from '@/app/api/openrouter/audio/transcriptions/route';
 
-export const maxDuration = 1800;
+export const maxDuration = 800;

apps/web/src/app/api/gateway/v1/audio/transcriptions/route.ts

@@ -1,3 +1,3 @@
 export { POST } from '@/app/api/openrouter/audio/transcriptions/route';
 
-export const maxDuration = 1800;
+export const maxDuration = 800;

apps/web/src/app/api/openrouter/[...path]/route.ts

@@ -107,7 +107,7 @@ import {
 import { redactProviderHints } from '@kilocode/auto-routing-contracts';
 import { logExceptInTest } from '@/lib/utils.server';
 
-export const maxDuration = 1800;
+export const maxDuration = 800;
 
 const MAX_TOKENS_LIMIT = 99999999999; // GPT4.1 default is ~32k
 
@@ -621,7 +621,7 @@ export async function POST(request: NextRequest): Promise<NextResponseType<unkno
   // previously blocking/quarantine decision wait for a fresh abuse-service result.
   const shouldBlockOnClassify = isRulesEngineBlockingAction(cachedRulesEngineAction);
 
-  // Large responses may run longer than the 1800s serverless function timeout.
+  // Large responses may run longer than the 800s serverless function timeout.
   const requestMaxTokens = getMaxTokens(requestBodyParsed);
   if (requestMaxTokens && requestMaxTokens > MAX_TOKENS_LIMIT) {
     console.warn(`SECURITY: Max tokens limit exceeded: ${user.id}`, {

apps/web/src/app/api/openrouter/audio/transcriptions/route.test.ts

@@ -85,7 +85,7 @@ describe('POST /api/gateway/v1/audio/transcriptions', () => {
       import('@/app/api/gateway/v1/audio/transcriptions/route'),
     ]);
 
-    expect(routes.map(route => route.maxDuration)).toEqual([1800, 1800, 1800, 1800]);
+    expect(routes.map(route => route.maxDuration)).toEqual([800, 800, 800, 800]);
   });
 
   it('proxies transcription requests to OpenRouter', async () => {

apps/web/src/app/api/openrouter/audio/transcriptions/route.ts

@@ -31,7 +31,7 @@ import {
 } from '@/lib/ai-gateway/transcriptions/transcription-request';
 import type { Provider } from '@/lib/ai-gateway/providers/types';
 
-export const maxDuration = 1800;
+export const maxDuration = 800;
 
 const PAID_MODEL_AUTH_REQUIRED = 'PAID_MODEL_AUTH_REQUIRED';
 

apps/web/src/app/api/openrouter/v1/audio/transcriptions/route.ts

@@ -1,3 +1,3 @@
 export { POST } from '@/app/api/openrouter/audio/transcriptions/route';
 
-export const maxDuration = 1800;
+export const maxDuration = 800;

apps/web/src/lib/ai-gateway/providers/upstream-request.generation.test.ts

@@ -36,7 +36,7 @@ describe('fetchGeneration', () => {
     jest.useRealTimers();
   });
 
-  it('uses a longer, slower retry window', async () => {
+  it('limits generation polling to about one minute', async () => {
     mockFetchWithBackoff.mockResolvedValue(
       new Response(JSON.stringify({ id: 'generation-id' }), {
         status: 200,
@@ -56,7 +56,7 @@ describe('fetchGeneration', () => {
       },
       expect.objectContaining({
         baseDelayMs: 5_000,
-        maxDelayMs: 5 * 60 * 1_000,
+        maxDelayMs: 75 * 1_000,
       })
     );
 

apps/web/src/lib/ai-gateway/providers/upstream-request.ts

@@ -23,9 +23,10 @@ type UpstreamFetchFailureFamily =
   | 'abort'
   | 'unknown';
 
-// Longer than Vercel AI Gateway's 13min timeout, shorter than Vercel Function's 30min timeout.
-const TIMEOUT_MS = 15 * 60 * 1000;
-const GENERATION_FETCH_MAX_DELAY_MS = 5 * 60 * 1000;
+// Leave 200s of the Vercel function budget for post-stream work.
+const TIMEOUT_MS = 10 * 60 * 1000;
+// fetchWithBackoff reserves the next delay before retrying, so 75s yields about one minute.
+const GENERATION_FETCH_MAX_DELAY_MS = 75 * 1000;
 
 function getProviderTargetHost(apiUrl: string): string {
   try {