diff --git a/apps/server/src/provider/CodexChatGptAuth.test.ts b/apps/server/src/provider/CodexChatGptAuth.test.ts
index 4e84553dca26a..0f75e857bdb0f 100644
--- a/apps/server/src/provider/CodexChatGptAuth.test.ts
+++ b/apps/server/src/provider/CodexChatGptAuth.test.ts
@@ -736,7 +736,8 @@ it.effect(
         yield* h.phase("succeeded");
         assert.strictEqual(Option.getOrThrow(yield* h.auth.read).subject, "user-test");
         assert.strictEqual(h.exchanges[0]?.get("client_id"), "oaiapp_test");
-        yield* h.seedExpired;
+        // 75 s left on the one-hour token: inside the renewal margin Codex's 60 s token cache needs.
+        yield* TestClock.adjust("3525 seconds");
         const records = yield* Effect.all([h.auth.access, h.auth.access], {
           concurrency: "unbounded",
         });
diff --git a/apps/server/src/provider/CodexChatGptAuth.ts b/apps/server/src/provider/CodexChatGptAuth.ts
index ad06cdc001ae6..40d1c09de7eda 100644
--- a/apps/server/src/provider/CodexChatGptAuth.ts
+++ b/apps/server/src/provider/CodexChatGptAuth.ts
@@ -31,6 +31,7 @@ import type { ProviderAuthFlowContext } from "@t3tools/provider-core/server/prov
 import { HttpClient, HttpClientRequest } from "effect/http";
 import * as ProviderCredentialStore from "./ProviderCredentialStore.ts";
 import { withChatGptSessionLock } from "./CodexChatGptSessionLock.ts";
+import { TOKEN_RENEWAL_MARGIN_MS } from "./CodexManagedTokenProtocol.ts";
 import * as ServerSecretStore from "../auth/ServerSecretStore.ts";
 import * as ServerSettings from "../serverSettings.ts";
 import * as ServerEnvironment from "../environment/ServerEnvironment.ts";
@@ -823,7 +824,7 @@ export const makeCodexChatGptAuth = Effect.fn("makeCodexChatGptAuth")(function*
         "Token sharing is disabled. Sign in with ChatGPT and enable token sharing.",
       );
     const now = yield* Clock.currentTimeMillis;
-    if (record.expiresAt - now > 60_000) return record;
+    if (record.expiresAt - now > TOKEN_RENEWAL_MARGIN_MS) return record;
     if (record.earliestRefreshAt !== null && record.earliestRefreshAt > now) {
       if (record.expiresAt > now) return record;
       return yield* failure(
diff --git a/apps/server/src/provider/CodexManagedRuntime.test.ts b/apps/server/src/provider/CodexManagedRuntime.test.ts
index 2c697f8fa8531..0df166b6d6b05 100644
--- a/apps/server/src/provider/CodexManagedRuntime.test.ts
+++ b/apps/server/src/provider/CodexManagedRuntime.test.ts
@@ -21,6 +21,7 @@ import * as ServerSecretStore from "../auth/ServerSecretStore.ts";
 import * as ServerEnvironment from "../environment/ServerEnvironment.ts";
 import * as CodexInstallation from "./CodexInstallation.ts";
 import { makeCodexManagedRuntime } from "./CodexManagedRuntime.ts";
+import { TOKEN_RENEWAL_MARGIN_MS } from "./CodexManagedTokenProtocol.ts";
 import * as ProviderCredentialStore from "./ProviderCredentialStore.ts";
 import { codexAppServerArgs } from "./codexLaunchArgs.ts";
 
@@ -147,7 +148,9 @@ it.live.each(
           assert.include(provider, 'base_url="https://api.openai.com/v1"');
           assert.include(provider, 'model_catalog_url="https://api.openai.com/v1/models"');
           assert.notInclude(provider, "env_key");
-          assert.include(provider, "refresh_interval_ms=0");
+          const refreshIntervalMs = Number(/refresh_interval_ms=(\d+)/u.exec(provider)?.[1]);
+          // regression: #33701 0 makes Codex reuse the launch token forever for catalog requests.
+          assert.isTrue(refreshIntervalMs > 0 && refreshIntervalMs < TOKEN_RENEWAL_MARGIN_MS);
           // regression: #16132 Codex fetches each token through this helper, so a running
           // session picks up renewals instead of keeping the token it launched with.
           if (source === "managed" && account === "primary") {
diff --git a/apps/server/src/provider/CodexManagedRuntime.ts b/apps/server/src/provider/CodexManagedRuntime.ts
index 3704c919c2791..d04605e2b47ba 100644
--- a/apps/server/src/provider/CodexManagedRuntime.ts
+++ b/apps/server/src/provider/CodexManagedRuntime.ts
@@ -10,7 +10,7 @@ import * as ServerEnvironment from "../environment/ServerEnvironment.ts";
 import { resolveSelfInvocation, selfInvocationArgs } from "@t3tools/shared/nodeRuntime";
 import * as CodexInstallation from "./CodexInstallation.ts";
 import { makeCodexChatGptAuth } from "./CodexChatGptAuth.ts";
-import { COMMAND_TIMEOUT_MS } from "./CodexManagedTokenProtocol.ts";
+import { COMMAND_TIMEOUT_MS, TOKEN_REFRESH_INTERVAL_MS } from "./CodexManagedTokenProtocol.ts";
 import { materializeCodexShadowHome } from "./Drivers/CodexHomeLayout.ts";
 
 export interface CodexEffectiveRuntime {
@@ -23,7 +23,7 @@ const managedCodexLaunchArgs = (command: string, args: ReadonlyArray<string>) =>
   [
     'model_provider="openai_token_sharing"',
     // Codex persists this provider ID in rollouts and needs it unchanged to resume.
-    `model_providers.openai_token_sharing={name="OpenAI Token Sharing",base_url="https://api.openai.com/v1",model_catalog_url="https://api.openai.com/v1/models",wire_api="responses",requires_openai_auth=false,supports_websockets=false,auth={command=${JSON.stringify(command)},args=${JSON.stringify(args)},timeout_ms=${COMMAND_TIMEOUT_MS},refresh_interval_ms=0}}`,
+    `model_providers.openai_token_sharing={name="OpenAI Token Sharing",base_url="https://api.openai.com/v1",model_catalog_url="https://api.openai.com/v1/models",wire_api="responses",requires_openai_auth=false,supports_websockets=false,auth={command=${JSON.stringify(command)},args=${JSON.stringify(args)},timeout_ms=${COMMAND_TIMEOUT_MS},refresh_interval_ms=${TOKEN_REFRESH_INTERVAL_MS}}}`,
     "features.api_key_model_discovery=true",
     "features.multi_agent=false",
     "features.multi_agent_v2=false",
diff --git a/apps/server/src/provider/CodexManagedTokenProtocol.ts b/apps/server/src/provider/CodexManagedTokenProtocol.ts
index d5f9e610d2188..cacee29243f1d 100644
--- a/apps/server/src/provider/CodexManagedTokenProtocol.ts
+++ b/apps/server/src/provider/CodexManagedTokenProtocol.ts
@@ -18,6 +18,9 @@ export const RESPONSE_BYTES = 64 * 1024;
 export const SOCKET_IDLE_MS = 5_000;
 // Existing lock retry (40s), OAuth (20s), plus startup/store headroom (20s).
 export const COMMAND_TIMEOUT_MS = 80_000;
+// Codex reuses a helper token this long; the broker renews earlier, leaving room for helper latency.
+export const TOKEN_REFRESH_INTERVAL_MS = 60_000;
+export const TOKEN_RENEWAL_MARGIN_MS = TOKEN_REFRESH_INTERVAL_MS + 30_000;
 class ProtocolError extends Schema.TaggedError<ProtocolError>()("ManagedTokenProtocolError", {}) {
   override get message(): string {
     return "The managed token frame was invalid or incomplete.";