Codex VS Code + Azure AI Foundry: First Responses request intermittently fails with “Bad Request”

Thank you for your question. Since this topic involves Azure OpenAI, WebSocket behavior, and incremental Responses API handling, it may take some time to receive replies here. You might also consider posting the question in the English forum, where more users and engineers familiar with these components are active. This could increase the likelihood of receiving detailed technical guidance.

English forum: https://learn.microsoft.com/en-us/answers/

Hi DANG HOANG VAN,

This is a good question. The behavior you are seeing is mainly because Codex models in Azure AI Foundry use a different API compared to regular chat models.

In Azure AI Foundry, there are two common ways to send requests:

• Chat Completions API
• Responses API

For most chat models, Chat Completions works fine. But Codex models are designed to work with the Responses API, not the Chat Completions API. [usercomp.com]

This is why you may see differences in how the “first response” behaves between VS Code tools and Azure Foundry.

In simple terms:

• VS Code tools often use Chat Completions by default
• Codex models expect Responses API for proper behavior

The Responses API combines chat, tool calling, and multi-step reasoning in one request. Because of that, the response format and behavior can look different when compared to older chat-based responses. [stackoverflow.com]

Here are a few things you can check:

First, confirm which API you are using If you are calling Codex models, make sure the request is going through the Responses API endpoint and not the chat completions endpoint

Second, check your endpoint format When using Responses API, the base URL usually includes /openai/v1/ and uses the responses request style

Third, verify the model type Codex models behave more like agent-style models, not simple chat models. They may:

• Return structured outputs
• Perform multi-step reasoning
• Use tools internally

So, the first response may look different compared to normal chat models.

Fourth, compare behavior with a standard model Try the same request with a regular chat model like GPT type models. If it works as expected there, then the difference is due to model behavior, not configuration

In short, nothing is wrong with your setup. It is just that Codex models are designed to work with the Responses API, so switching to that API normally fixes these differences.

I Hope this helps. Do let me know if you have any further queries.
Thankyou!

おい、 ダン・ホアン・ヴァン

ご迷惑をおかけしてすみません。

短い通話のために、つながり情報をプライベートメッセージで共有してください。

サポートチケットを作成し、問題が続く場合はエンジニアリングチームにエスカレーションするために、あなたの意見が必要です。

ありがとうございます。