Support Codex chat models via Copilot Responses API

[PeddnadePradeep]

GPT-5.3-Codex now appears in NBI's Copilot chat model dropdown , but selecting it for chat fails with a 400 from Copilot's API. Codex models are only accessible through Copilot's /responses endpoint, not /chat/completions, so NBI's single chat code path cannot serve them.

NBI version: 4.8.0
JupyterLab: 4.x
Python: 3.11
Platform: Linux container (JupyterHub-deployed)
Copilot plan: Business with GPT-5.3-Codex policy enabled
Verified the same model works in VS Code's Copilot Chat extension on the same account

[pjdoland]

Opened #341 with a candidate fix. The change adds a second code path that routes Codex-family models through Copilot's /responses endpoint (the OpenAI Responses API mirror), while non-Codex models keep using /chat/completions. Endpoint selection is driven by each model's supported_endpoints field in Copilot's /models response, with a codex substring fallback for the offline / pre-login window.

The wire contract is implemented against codecompanion.nvim's mature Copilot/Responses adapter and OpenAI's documented Responses API. I do not have a Copilot Business token with Codex enabled in this environment, so end-to-end verification against live Copilot was not possible from my side.

Would you be willing to test the PR against your account and confirm whether gpt-5.3-codex (or whichever Codex model you have provisioned) now works for chat? To try it locally:

gh pr checkout 341
pip install -e .
jupyter labextension develop . --overwrite
jupyter lab

If anything goes sideways (a different 400, a streaming hang, missing tool calls), please paste the failure detail here. The implementation defensively handles response.failed / response.incomplete / generic error SSE events and accumulates response.function_call_arguments.delta chunks in case the terminal arguments ships empty, but real-world Copilot behavior may surface cases the documentation does not describe.

Thanks for the report.