fix(search): give the alpha/search relay its own total deadline (search.timeoutMs)

connectTimeoutMs is documented as the DNS/TCP/TLS/header-arrival budget, but the
relay reused it as the TOTAL upstream deadline. alpha/search is non-streaming —
headers arrive only when the search completes — so a typical 10s connect budget
killed every long-running search. Mirror OcxImagesConfig.timeoutMs with a
dedicated config.search.timeoutMs (default 200s), move the 504 test to the new
knob, and add a regression test proving a short connectTimeoutMs no longer cuts
a slow search. (PR #89 follow-up)

Author: lidge-jun

src/server/search.ts

@@ -26,6 +26,13 @@ import { readJsonRequestBody } from "./request-decompress";
 import type { RequestLogContext } from "./request-log";
 import { codexLogAccountId, decodeRequestErrorResponse, sidecarOutcomeRecorder } from "./responses";
 
+/**
+ * Default TOTAL deadline for one search relay. alpha/search is non-streaming JSON — response
+ * headers arrive only when the search finishes — so the budget must cover the whole request.
+ * Overridable via config.search.timeoutMs; never config.connectTimeoutMs, whose documented
+ * contract is the DNS/TCP/TLS/header-arrival budget (a 10s connect budget would kill every
+ * long-running search).
+ */
 const SEARCH_UPSTREAM_TIMEOUT_MS = 200_000;
 const SEARCH_RESPONSE_MAX_BYTES = 16 * 1024 * 1024;
 
@@ -103,7 +110,7 @@ export async function handleSearch(
   if (upstream.provider.headers) Object.assign(headers, upstream.provider.headers);
   for (const [name, value] of authHeaders) headers[name] = value;
   const url = `${upstream.provider.baseUrl}/alpha/search`;
-  const timeoutMs = config.connectTimeoutMs ?? SEARCH_UPSTREAM_TIMEOUT_MS;
+  const timeoutMs = config.search?.timeoutMs ?? SEARCH_UPSTREAM_TIMEOUT_MS;
   const linkedSignal = signalWithTimeout(timeoutMs, req.signal);
   const sidecarExit = sidecarEnter("search");
   try {

src/types.ts

@@ -328,6 +328,8 @@ export interface OcxConfig {
   visionSidecar?: OcxVisionSidecarConfig;
   /** /v1/images relay for codex's built-in image_gen tool. */
   images?: OcxImagesConfig;
+  /** /v1/alpha/search relay for codex's built-in web search client. */
+  search?: OcxSearchConfig;
   /** Codex multi-account pool. */
   codexAccounts?: CodexAccount[];
   /** Active pool account id for next session. undefined = main (passthrough as-is). */
@@ -378,6 +380,15 @@ export interface OcxImagesConfig {
   timeoutMs?: number;
 }
 
+export interface OcxSearchConfig {
+  /**
+   * Total upstream deadline (ms) for one /v1/alpha/search relay. Default 200000. The endpoint
+   * is non-streaming JSON (headers arrive only when the search completes), so this is a whole-
+   * request budget — deliberately NOT connectTimeoutMs, which is a header-arrival budget.
+   */
+  timeoutMs?: number;
+}
+
 export interface OcxVisionSidecarConfig {
   /** Master switch. Default: enabled when a forward (ChatGPT) provider exists and the caller is logged in. */
   enabled?: boolean;

tests/server-search.test.ts

@@ -254,7 +254,7 @@ test("relays search upstream error status and body verbatim", async () => {
   }
 });
 
-test("a hung search upstream times out with 504 after config.connectTimeoutMs", async () => {
+test("a hung search upstream times out with 504 after config.search.timeoutMs", async () => {
   const upstream = Bun.serve({
     port: 0,
     fetch(req) {
@@ -265,7 +265,7 @@ test("a hung search upstream times out with 504 after config.connectTimeoutMs",
   });
   saveConfig({
     ...forwardConfig(upstream.url.toString().replace(/\/$/, "")),
-    connectTimeoutMs: 100,
+    search: { timeoutMs: 100 },
   } as OcxConfig);
 
   const server = startServer(0);
@@ -284,6 +284,37 @@ test("a hung search upstream times out with 504 after config.connectTimeoutMs",
   }
 }, 5_000);
 
+test("a short connectTimeoutMs does NOT cut a slow search (total deadline is search.timeoutMs)", async () => {
+  // Regression: alpha/search is non-streaming, so its headers arrive only when the search
+  // completes. Reusing connectTimeoutMs as the relay deadline killed every search longer than
+  // the header-arrival budget (often ~10s in real configs).
+  const upstream = Bun.serve({
+    port: 0,
+    async fetch() {
+      await new Promise(resolve => setTimeout(resolve, 300));
+      return Response.json({ output: "slow but fine" });
+    },
+  });
+  saveConfig({
+    ...forwardConfig(upstream.url.toString().replace(/\/$/, "")),
+    connectTimeoutMs: 50,
+  } as OcxConfig);
+
+  const server = startServer(0);
+  try {
+    const response = await fetch(new URL("/v1/alpha/search", server.url), {
+      method: "POST",
+      headers: { "content-type": "application/json", authorization: "Bearer chatgpt-user-token" },
+      body: JSON.stringify({ id: "search-session", model: "gpt-test" }),
+    });
+    expect(response.status).toBe(200);
+    expect(await response.json()).toEqual({ output: "slow but fine" });
+  } finally {
+    await server.stop(true);
+    await upstream.stop(true);
+  }
+}, 5_000);
+
 test("GET /v1/alpha/search still falls through to the JSON 404 guard", async () => {
   saveConfig(forwardConfig("https://chatgpt.example/backend-api/codex"));