Login_chatgpt_device Fails Due To Incorrect Request Encoding And Mismatched OAuth Contract

The implementation uses an outdated device authorization API contract: the initial POST to /deviceauth/usercode sends form-encoded data instead of JSON, and subsequent steps (polling, token exchange) use incorrect request formats and status code handling, mismatching OpenAI's Codex device-login implementation.

Run the following Python code with langchain-openai 1.6.0: ```python from pathlib import Path from tempfile import TemporaryDirectory from langchain_openai.chatgpt_oauth import login_chatgpt_device with TemporaryDirectory() as directory: login_chatgpt_device(store_path=Path(directory) / "auth.json") ``` Observe the HTTP 400 error from https://auth.openai.com/api/accounts/deviceauth/usercode.

The endpoint URLs (poll and token) are guessed from the description and may need adjustment to match the actual OpenAI API. Test thoroughly before merging. The function uses a new httpx.Client per call, which is safe but not optimal for high concurrency; consider reusing a client. Rollback: if this fix causes issues, revert to the previous version (though it is already broken), or consult the official OpenAI Codex implementation for the exact endpoints and payloads.