Supercharge your AI agents with data from the web and beyond. Building the library for superintelligence. 🔥
TypeScript aiai-agentsai-crawlerai-scrapingai-searchcrawlerdata-extractionhtml-to-markdownllmmarkdownscraperscrapingweb-crawlerweb-dataweb-data-extractionweb-scraperweb-scrapingweb-searchwebscraping
fix(branding): fail fast when TypeSafe asks Jev calls to back off (#5179)
Abimael Martell9d18f53ecf28
diff --git a/apps/api/src/__tests__/lib/branding/jev.test.ts b/apps/api/src/__tests__/lib/branding/jev.test.tsindex e8f4f4041..8db817cbd 100644--- a/apps/api/src/__tests__/lib/branding/jev.test.ts+++ b/apps/api/src/__tests__/lib/branding/jev.test.ts@@ -307,7 +307,17 @@ describe("branding with Jev", () => {expect(request.model).toBe("jev-latest");expect(options).toEqual({timeout: 5000,- retry: { maxRetries: 1, apiTimeoutError: false },+ retry: {+ maxRetries: 1,+ apiTimeoutError: false,+ httpStatuses: new Set([+ 408,+ ...Array.from({ length: 100 }, (_, i) => 500 + i).filter(+ s => s !== 529,+ ),+ ]),+ maxRetryAfterMs: 1000,+ },});expect(Object.keys(request.questions.logo.criteria)).toEqual(["logo_0",@@ -401,6 +411,46 @@ describe("branding with Jev", () => {expect(result.buttonClassification.primaryButtonIndex).toBe(-1);});+ it("fails fast when TypeSafe asks to back off, and retries a 503 once without waiting it out", async () => {+ // The options Jev passes, run through the real SDK against a fake server.+ mocks.systemOne.mockResolvedValueOnce(jevResponse());+ await enhanceBrandingWithLLM(baseInput(new CostTracking()));+ const [request, options] = mocks.systemOne.mock.calls[0];+ const { TypeSafeClient } =+ await vi.importActual<typeof import("@typesafe-ai/sdk")>(+ "@typesafe-ai/sdk",+ );+ const call = async (status: number) => {+ const fetch = vi.fn(+ async () =>+ new Response(JSON.stringify({ error: { message: "busy" } }), {+ status,+ headers: {+ "content-type": "application/json",+ "retry-after": "10",+ },+ }),+ );+ const client = new TypeSafeClient({+ apiKey: "test",+ logLevel: "off",+ fetch,+ });+ const started = Date.now();+ await expect(client.systemOne(request, options)).rejects.toThrow();+ return { attempts: fetch.mock.calls.length, ms: Date.now() - started };+ };++ for (const status of [529, 429]) {+ const r = await call(status);+ expect(r.attempts).toBe(1);+ expect(r.ms).toBeLessThan(500);+ }+ const r = await call(503);+ expect(r.attempts).toBe(2);+ expect(r.ms).toBeLessThan(2000); // not the 10 s Retry-After+ });+it("stops calling Jev for a minute after repeated failures, then tries again", async () => {vi.useFakeTimers({ toFake: ["Date"] });vi.setSystemTime(Date.parse("2026-10-08T12:00:00Z"));diff --git a/apps/api/src/lib/branding/jev.ts b/apps/api/src/lib/branding/jev.tsindex 9dcdfea81..cd691eeaf 100644--- a/apps/api/src/lib/branding/jev.ts+++ b/apps/api/src/lib/branding/jev.ts@@ -3,6 +3,7 @@ import type {JsonValue,NoulResponse,Question,+ RequestOptions,SystemOneResult,} from "@typesafe-ai/sdk";import { formatHex, hsl, parse } from "culori";@@ -32,6 +33,21 @@ const JEV_MODEL = "jev-latest";const JEV_INPUT_USD_PER_MTOK = 0.042;const DEFAULT_TIMEOUT_MS = 5000;+// One quick retry, for a transient server error only. A call that ran out of+// time, or a 429 or 529 asking us to back off, is answered by the LLM instead:+// waiting can take longer than the whole LLM fallback, and adds load to a+// struggling TypeSafe. A server-requested delay over a second is not waited+// out either (the SDK backs off ~0.5 s instead).+const JEV_RETRY: RequestOptions["retry"] = {+ maxRetries: 1,+ apiTimeoutError: false,+ httpStatuses: new Set([+ 408,+ ...Array.from({ length: 100 }, (_, i) => 500 + i).filter(s => s !== 529),+ ]),+ maxRetryAfterMs: 1000,+};+// Circuit breaker: after this many calls in a row fail (timeouts, 5xx), skip// Jev for the cooldown and answer with the LLM, so an outage doesn't add the// timeout to every branding request. After the cooldown calls go through@@ -936,9 +952,7 @@ export async function enhanceBrandingWithJev(},{timeout: config.BRANDING_JEV_TIMEOUT_MS ?? DEFAULT_TIMEOUT_MS,- // A call that ran out of time is answered by the LLM instead:- // retrying it doubles the wait, and the load on a slow TypeSafe.- retry: { maxRetries: 1, apiTimeoutError: false },+ retry: JEV_RETRY,},);setSpanAttributes(span, {