From ab801fe48c05c57718a59ebdc09c220b20e174f6 Mon Sep 17 00:00:00 2001
From: Kevin Wang <kenwang3333@gmail.com>
Date: Sat, 19 Sep 2026 18:24:49 -0400
Subject: [PATCH] Ask the Gateway again after a 503 before paying OpenRouter

A quarter of judge calls came back 503 from the Gateway in about a second and
were paid for on OpenRouter. They arrive in short spells and clear on a wait of
under two seconds.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
---
 src/lib/jev.ts    | 23 +++++++++++++++++++++++
 tests/jev.test.ts | 23 +++++++++++++++++++++++
 2 files changed, 46 insertions(+)

diff --git a/src/lib/jev.ts b/src/lib/jev.ts
index da78305..1cf403f 100644
--- a/src/lib/jev.ts
+++ b/src/lib/jev.ts
@@ -31,6 +31,15 @@ export const GATEWAY_FREE_UNTIL = new Date("2026-09-26T07:00:00Z");
  */
 const GATEWAY_TIMEOUT_MS = 15_000;
 
+/**
+ * The waits before the Gateway is asked again after a 503. Measured 2026-09-19
+ * on 2,047 judge calls, 507 came back 503 in about a second and were paid for
+ * on OpenRouter; they arrive in short spells, whatever the request's size, and
+ * of 9 met in a probe of 80 calls 8 cleared on the first wait and the last on
+ * the second. Nothing else is retried here: a 429 or a timeout moves on.
+ */
+const GATEWAY_503_WAITS_MS = [400, 1200];
+
 const ENDPOINT = "https://openrouter.ai/api/alpha/decisions";
 
 export type NoulQuestion = {
@@ -193,6 +202,20 @@ export async function askJev(call: JevCall): Promise<Answers> {
 }
 
 async function askGateway(key: string, call: JevCall): Promise<Answers> {
+  for (let attempt = 0; ; attempt += 1) {
+    try {
+      return await askGatewayOnce(key, call);
+    } catch (error) {
+      const wait = GATEWAY_503_WAITS_MS[attempt];
+      if (wait === undefined || !GatewayError.isInstance(error) || error.statusCode !== 503) {
+        throw error;
+      }
+      await new Promise((resolve) => setTimeout(resolve, wait));
+    }
+  }
+}
+
+async function askGatewayOnce(key: string, call: JevCall): Promise<Answers> {
   const { JEV_GATEWAY_MODEL } = config();
   const gateway = createGateway({ apiKey: key });
   const startedAt = Date.now();
diff --git a/tests/jev.test.ts b/tests/jev.test.ts
index 6752645..ecee942 100644
--- a/tests/jev.test.ts
+++ b/tests/jev.test.ts
@@ -179,6 +179,29 @@ describe("the Gateway first, OpenRouter behind it", () => {
     expect(recorded[1]).toMatchObject({ provider: "typesafe", finishReason: "answered" });
   });
 
+  it("asks the Gateway again after a 503 before paying OpenRouter", async () => {
+    recorded.length = 0;
+    settings.AI_GATEWAY_API_KEY = "vck_test";
+    let asked = 0;
+    const fetch = twoRoutes(() => {
+      asked += 1;
+      return asked === 1
+        ? new Response(JSON.stringify({ error: { message: "unavailable", type: "service_unavailable" } }), { status: 503 })
+        : new Response(
+            JSON.stringify({
+              answers: { p0__audience: { type: "boolean", probability: 0.9 } },
+              usage: { inputTokens: 500, outputTokens: 0 },
+            }),
+          );
+    }, openrouterAnswer);
+    vi.stubGlobal("fetch", fetch);
+
+    const answers = await askJev(call);
+
+    expect(answers.p0__audience).toEqual({ type: "noul", noul: 0.9 });
+    expect(recorded.map((row) => `${row.provider}:${row.finishReason}`)).toEqual(["vercel:http_503", "vercel:answered"]);
+  });
+
   it("asks the Gateway first every time, even after it rate-limits", async () => {
     recorded.length = 0;
     settings.AI_GATEWAY_API_KEY = "vck_test";