connectOpenAi fails with server_error code 42 while OpenAI Realtime direct test succeeds

从 Next.js 应用中的 connectOpenAi 调用开始，检查所列出的 @stream-io/node-sdk 和 @stream-io/openai-realtime-api 版本以及 Vercel 部署。复现成功的 client_secrets 请求和失败的 agent 连接，然后确定该设置是否受支持，以及需要哪些 Stream 设置、权限或 entitlement。

Hi Stream team,

I’m using Stream Video with OpenAI Realtime in a Next.js app deployed on Vercel.

Packages:

• @stream-io/node-sdk@0.7.59
• @stream-io/openai-realtime-api@0.3.3
• Next.js 15.3.8
• Deployed on Vercel

OpenAI Realtime direct test succeeds:

POST https://api.openai.com/v1/realtime/client_secrets
model: gpt-4o-realtime-preview
status: 200 OK

But Stream connectOpenAi fails:

const realtimeClient = await streamVideo.video.connectOpenAi({
call,
openAiApiKey: process.env.OPENAI_API_KEY,
agentUserId: "lucy",
model: "gpt-4o-realtime-preview",
});

Error:

{
type: "error",
error: {
type: "server_error",
code: "42",
message: "Error: code ",
param: null,
event_id: ""
}
}

Additional context:

• The Stream call is created successfully.
• The user joins successfully.
• The agent user "lucy" appears to join and then leaves.
• Webhooks show call.session_participant_joined and call.session_participant_left for lucy.
• OpenAI Realtime client_secret endpoint returns 200 OK, so the OpenAI key and Realtime access appear valid.

Can you confirm whether connectOpenAi currently supports this setup, and whether any Stream app settings, permissions, or entitlements are required for OpenAI Realtime agents?