Feature Request: Support the new OpenAI /v1/responses API route

[WASIDJ]

Background

When using clients such as lazyllm, the default OpenAI API path is the new /v1/responses endpoint (introduced in OpenAI’s 2024 API).
However, the Copilot Bridge currently only implements the traditional:

• /v1/chat/completions
• /v1/models
• /health

Because of this, any request sent to /v1/responses results in:

404 route_not_found

Even though the bridge is advertised as implementing “OpenAI-style endpoints”, it only supports the legacy Chat Completions API.

---

Problem

Many modern OpenAI-compatible clients — including lazyllm, LangChain, LlamaIndex, and the official OpenAI SDK — now use /v1/responses by default, with an input field instead of a messages array.

This makes the Copilot Bridge incompatible with them out of the box.

---

Feature Request

I would like to request:

✔ 1. Add a new /v1/responses endpoint

Mirror the OpenAI behavior, even if it internally maps to the chat completions API.

✔ 2. Automatically convert the new request format:

Example new format:

{
  "model": "gpt-4.1",
  "input": "Hello",
  "stream": false
}

Convert internally to legacy format:

{
  "model": "gpt-4.1",
  "messages": [
    {"role": "user", "content": "Hello"}
  ],
  "stream": false
}

✔ 3. (Optional) Document the limitation in README

If the bridge does not intend to support the new API, please clarify this in the README so developers know what to expect.

---

Why this matters

Supporting /v1/responses would:

• Enable automatic compatibility with modern OpenAI SDKs
• Let clients like lazyllm, LangChain, etc. work without custom patches
• Improve VS Code integration and reduce error reports
• Future-proof the bridge as OpenAI continues phasing out the chat completions format

---

Thanks

Thanks for maintaining this project — it’s extremely useful for local model proxying inside VS Code.
I’d be happy to help contribute a PR implementing this if a design direction is confirmed.

[larsbaunwall]

Thank you so much for you thorough issue and suggestion.

Would love to see a PR for this 🤩