diff --git a/apps/web/src/__tests__/proxy-rate-limit-classification.test.ts b/apps/web/src/__tests__/proxy-rate-limit-classification.test.ts
new file mode 100644
index 000000000..31b49c350
--- /dev/null
+++ b/apps/web/src/__tests__/proxy-rate-limit-classification.test.ts
@@ -0,0 +1,92 @@
+import { afterEach, describe, expect, it, vi } from 'vitest';
+
+/**
+ * Rate-limit route classification (PR #1507 review — VADE "logic" finding).
+ *
+ * The durable video-to-actions workflow polls GET
+ * /api/workflows/video-to-actions/:runId every ~1.5s (studio-workflow.ts).
+ * When that poll is classified as an "AI route" it draws from the strict
+ * 12/min AI budget, so the client poller is 429'd mid-run and the UI hangs on
+ * "still running". Only the mutating start (POST /api/workflows/...) does real
+ * AI work and should be AI-limited; the status poll belongs on the general
+ * budget.
+ *
+ * proxy.ts reads limits + env into module-level constants, so each case
+ * re-imports the module after stubbing env (mirrors proxy-auth-gate.test.ts).
+ * The applied budget is read off the X-RateLimit-Limit response header.
+ */
+
+vi.mock('next-auth/jwt', () => ({
+  getToken: vi.fn(async () => null),
+}));
+
+// NODE_ENV=development => auth gate disabled + in-memory limiter (emits headers).
+// All limit/redis knobs are pinned so the result never depends on ambient env.
+const ENV: Record<string, string | undefined> = {
+  NODE_ENV: 'development',
+  NEXTAUTH_SECRET: undefined,
+  INTERNAL_REQUEST_TOKEN: undefined,
+  UPSTASH_REDIS_REST_URL: undefined,
+  UPSTASH_REDIS_REST_TOKEN: undefined,
+  UVAI_RATE_LIMIT_DISABLED: undefined,
+  UVAI_API_RATE_LIMIT_PER_MINUTE: undefined, // default GENERAL_LIMIT = 60
+  UVAI_AI_RATE_LIMIT_PER_MINUTE: undefined, // default AI_LIMIT = 12
+};
+
+const GENERAL = '60';
+const AI = '12';
+
+afterEach(() => {
+  vi.unstubAllEnvs();
+  vi.resetModules();
+});
+
+async function loadProxy() {
+  for (const [key, value] of Object.entries(ENV)) {
+    vi.stubEnv(key, value);
+  }
+  vi.resetModules();
+  const { NextRequest } = await import('next/server');
+  const { proxy } = await import('@/proxy');
+  return { proxy, NextRequest };
+}
+
+describe('rate-limit classification for /api/workflows (PR #1507)', () => {
+  it('puts the workflow status poll (GET) on the general budget, not the AI budget', async () => {
+    const { proxy, NextRequest } = await loadProxy();
+    const res = await proxy(
+      new NextRequest(
+        'http://localhost:3000/api/workflows/video-to-actions/run_abc123',
+        { method: 'GET' },
+      ),
+    );
+    // The bug: before the fix this header reads "12" and the poller 429s mid-run.
+    expect(res.headers.get('X-RateLimit-Limit')).toBe(GENERAL);
+  });
+
+  it('keeps the workflow start (POST) on the strict AI budget', async () => {
+    const { proxy, NextRequest } = await loadProxy();
+    const res = await proxy(
+      new NextRequest('http://localhost:3000/api/workflows/video-to-actions', {
+        method: 'POST',
+      }),
+    );
+    expect(res.headers.get('X-RateLimit-Limit')).toBe(AI);
+  });
+
+  it('leaves other AI routes on the AI budget regardless of method', async () => {
+    const { proxy, NextRequest } = await loadProxy();
+    const res = await proxy(
+      new NextRequest('http://localhost:3000/api/chat', { method: 'POST' }),
+    );
+    expect(res.headers.get('X-RateLimit-Limit')).toBe(AI);
+  });
+
+  it('keeps ordinary API routes on the general budget', async () => {
+    const { proxy, NextRequest } = await loadProxy();
+    const res = await proxy(
+      new NextRequest('http://localhost:3000/api/health', { method: 'GET' }),
+    );
+    expect(res.headers.get('X-RateLimit-Limit')).toBe(GENERAL);
+  });
+});
diff --git a/apps/web/src/proxy.ts b/apps/web/src/proxy.ts
index da97c6914..55c5baa60 100644
--- a/apps/web/src/proxy.ts
+++ b/apps/web/src/proxy.ts
@@ -38,9 +38,16 @@ const AI_ROUTE_PREFIXES = [
   '/api/training',
   '/api/transcribe',
   '/api/video',
-  '/api/workflows',
 ];
 
+// Durable-workflow routes are a start (POST /api/workflows/...) plus a status
+// long-poll (GET /api/workflows/.../:runId). Only the start does real AI work;
+// the poll fires every ~1.5s (studio-workflow.ts) and must NOT draw from the
+// strict AI budget, or the client poller is 429'd mid-run and the UI hangs on
+// "still running" (PR #1507 review). So these are classified by method, not by
+// a blanket prefix: mutating start = AI budget, GET status poll = general.
+const WORKFLOW_ROUTE_PREFIX = '/api/workflows';
+
 // Login gating (activate-when-configured) + server-to-server bypass.
 const INTERNAL_TOKEN = process.env.INTERNAL_REQUEST_TOKEN;
 const AUTH_SECRET = process.env.NEXTAUTH_SECRET;
@@ -121,7 +128,11 @@ function getRedisClient(): Promise<Redis | null> {
   return redisClientPromise;
 }
 
-function isAiRoute(pathname: string): boolean {
+function isAiRoute(pathname: string, method: string): boolean {
+  if (pathname.startsWith(WORKFLOW_ROUTE_PREFIX)) {
+    // Only the mutating start is AI-heavy; GET status polls use the general budget.
+    return method === 'POST';
+  }
   return AI_ROUTE_PREFIXES.some((prefix) => pathname.startsWith(prefix));
 }
 
@@ -144,8 +155,8 @@ function getClientIp(request: NextRequest): string {
   return 'unknown';
 }
 
-function getRateLimit(pathname: string): number {
-  return isAiRoute(pathname) ? AI_LIMIT : GENERAL_LIMIT;
+function getRateLimit(pathname: string, method: string): number {
+  return isAiRoute(pathname, method) ? AI_LIMIT : GENERAL_LIMIT;
 }
 
 async function checkRedisLimit(redisClient: Redis, key: string, limit: number): Promise<RateLimitResult> {
@@ -188,9 +199,10 @@ function checkMemoryLimit(key: string, limit: number): RateLimitResult {
 
 async function checkRateLimit(request: NextRequest): Promise<RateLimitResult> {
   const pathname = request.nextUrl.pathname;
-  const limit = getRateLimit(pathname);
+  const method = request.method;
+  const limit = getRateLimit(pathname, method);
   const clientIp = getClientIp(request);
-  const routeClass = isAiRoute(pathname) ? 'ai' : 'api';
+  const routeClass = isAiRoute(pathname, method) ? 'ai' : 'api';
   const key = `${routeClass}:${clientIp}`;
 
   const redisClient = await getRedisClient();