fix(ai): keep header ingests from clearing usage-fetch cooldowns

packages/ai/src/auth-storage.ts

@@ -3172,7 +3172,8 @@ export class AuthStorage {
 		}
 
 		if (this.#fetchUsageReportsOverride || this.#store.fetchUsageReports) return false;
-		const prior = this.#usageCache.getStale<UsageReport | null>(cacheKey)?.value;
+		const priorEntry = this.#usageCache.getStale<UsageReport | null>(cacheKey);
+		const prior = priorEntry?.value;
 		let merged = report;
 		if (prior && Array.isArray(prior.limits)) {
 			const headerLimitsById = new Map(report.limits.map(limit => [limit.id, limit]));
@@ -3203,9 +3204,12 @@ export class AuthStorage {
 		}
 
 		// Header-only reports are last-good hints, not completed usage fetches.
-		// Keep them durable but stale so the next poll probes the provider; a
-		// failed probe still applies the normal short backoff in #fetchUsageCached.
-		const expiresAt = merged.metadata?.source === "ratelimit-headers" ? now - 1 : now + USAGE_REPORT_TTL_MS;
+		// Keep them durable but stale so the next poll probes the provider, while
+		// preserving any active failure backoff installed by #fetchUsageCached.
+		const expiresAt =
+			merged.metadata?.source === "ratelimit-headers"
+				? Math.max(priorEntry?.expiresAt ?? now - 1, now - 1)
+				: now + USAGE_REPORT_TTL_MS;
 		this.#usageCache.set(cacheKey, { value: merged, expiresAt });
 		this.#usageHeaderIngestAt.set(cacheKey, now);
 		return true;

packages/ai/test/auth-storage-usage-cache.test.ts

@@ -460,6 +460,38 @@ describe("AuthStorage usage cache: header ingestion", () => {
 		expect(storage.ingestUsageHeaders("anthropic", usageHeaders("0.05", "0.6"), { sessionId: "s" })).toBe(false);
 	});
 
+	it("preserves a failed fetch cooldown across exhausted header ingestion, then retries after expiry", async () => {
+		const start = Date.now();
+		const now = vi.spyOn(Date, "now").mockReturnValue(start);
+		let calls = 0;
+		vi.spyOn(claudeUsage.claudeUsageProvider, "fetchUsage").mockImplementation(async () => {
+			calls += 1;
+			return null;
+		});
+
+		expect(await storage.getApiKey("anthropic", "cooldown-session")).toBe("oat-1");
+		expect(
+			storage.ingestUsageHeaders("anthropic", usageHeaders("0.02", "0.3"), {
+				sessionId: "cooldown-session",
+			}),
+		).toBe(true);
+		expect(await storage.fetchUsageReports()).toHaveLength(1);
+		expect(calls).toBe(1);
+
+		now.mockReturnValue(start + 1_000);
+		expect(
+			storage.ingestUsageHeaders("anthropic", usageHeaders("1", "0.3"), {
+				sessionId: "cooldown-session",
+			}),
+		).toBe(true);
+		expect(await storage.fetchUsageReports()).toHaveLength(1);
+		expect(calls).toBe(1);
+
+		now.mockReturnValue(start + 12_501);
+		expect(await storage.fetchUsageReports()).toHaveLength(1);
+		expect(calls).toBe(2);
+	});
+
 	it("merges header umbrella windows onto the last real report and preserves tier limits", async () => {
 		const realReport = makeTieredReport("a@example.com");
 		let calls = 0;