修复 weighted 路由失败退避时间溢出导致服务崩溃 (#354)

* fix: cap token router cooldown backoff

* docs: explain token router cooldown cap

src/server/services/tokenRouter.cache.test.ts

@@ -691,4 +691,60 @@ describe('TokenRouter runtime cache', () => {
     expect(current?.cooldownLevel).toBe(0);
     expect(current?.cooldownUntil).toBeNull();
   });
+
+  it('caps weighted cooldowns before Date overflow for heavily failed channels', async () => {
+    const site = await db.insert(schema.sites).values({
+      name: 'weighted-cooldown-cap-site',
+      url: 'https://weighted-cooldown-cap.example.com',
+      platform: 'new-api',
+      status: 'active',
+    }).returning().get();
+
+    const account = await db.insert(schema.accounts).values({
+      siteId: site.id,
+      username: 'weighted-cooldown-cap-user',
+      accessToken: 'weighted-cooldown-cap-access-token',
+      apiToken: 'weighted-cooldown-cap-api-token',
+      status: 'active',
+    }).returning().get();
+
+    const token = await db.insert(schema.accountTokens).values({
+      accountId: account.id,
+      name: 'weighted-cooldown-cap-token',
+      token: 'sk-weighted-cooldown-cap-token',
+      enabled: true,
+      isDefault: true,
+    }).returning().get();
+
+    const route = await db.insert(schema.tokenRoutes).values({
+      modelPattern: 'gpt-5.4',
+      routingStrategy: 'weighted',
+      enabled: true,
+    }).returning().get();
+
+    const channel = await db.insert(schema.routeChannels).values({
+      routeId: route.id,
+      accountId: account.id,
+      tokenId: token.id,
+      priority: 0,
+      weight: 10,
+      enabled: true,
+      failCount: 57,
+    }).returning().get();
+
+    // Weighted routes previously let Fibonacci backoff grow beyond the Date range.
+    const router = new TokenRouter();
+
+    const startedAt = Date.now();
+    await router.recordFailure(channel.id);
+    const current = await db.select().from(schema.routeChannels)
+      .where(eq(schema.routeChannels.id, channel.id))
+      .get();
+    const cooldownMs = Date.parse(String(current?.cooldownUntil || '')) - startedAt;
+
+    expect(current?.failCount).toBe(58);
+    expect(Number.isFinite(Date.parse(String(current?.cooldownUntil || '')))).toBe(true);
+    expect(cooldownMs).toBeGreaterThanOrEqual((30 * 24 * 60 * 60 - 5) * 1000);
+    expect(cooldownMs).toBeLessThanOrEqual((30 * 24 * 60 * 60 + 5) * 1000);
+  });
 });

src/server/services/tokenRouter.ts

@@ -72,6 +72,8 @@ type SiteRuntimeHealthState = {
 
 const FAILURE_BACKOFF_BASE_SEC = 15;
 const SHORT_WINDOW_LIMIT_COOLDOWN_MS = 5 * 60 * 1000;
+// Keep weighted-route backoff within the JavaScript Date range when fail counts grow large.
+const MAX_FAILURE_BACKOFF_SEC = 30 * 24 * 60 * 60;
 const MIN_EFFECTIVE_UNIT_COST = 1e-6;
 const ROUND_ROBIN_FAILURE_THRESHOLD = 3;
 const ROUND_ROBIN_COOLDOWN_LEVELS_SEC = [0, 10 * 60, 60 * 60, 24 * 60 * 60] as const;
@@ -192,9 +194,13 @@ function fibonacciNumber(index: number): number {
   return current;
 }
 
+/**
+ * Weighted-route failures use a Fibonacci backoff, but the resulting cooldown must stay
+ * representable as a JavaScript Date for downstream `toISOString()` calls.
+ */
 function resolveFailureBackoffSec(failCount?: number | null): number {
   const normalizedFailCount = Math.max(1, Math.trunc(failCount ?? 0));
-  return FAILURE_BACKOFF_BASE_SEC * fibonacciNumber(normalizedFailCount);
+  return Math.min(FAILURE_BACKOFF_BASE_SEC * fibonacciNumber(normalizedFailCount), MAX_FAILURE_BACKOFF_SEC);
 }
 
 function resolveRoundRobinCooldownSec(level: number): number {