@thsottiaux 1. Let planning and execution use different model settings, similar to Cursor’s Opus Plan.
I’d love to use a smarter/slower model for planning or hard reasoning, and a faster/cheaper model for routine execution.
@thsottiaux Also, for non-English-speaking developers like us, it’s unfortunate that Codex Cloud code reviews are only written in English. Being able to configure the review language would be very helpful.
@OpenAI Inspired by this video:
https://t.co/yO7pNhZSee
What if the environment blocked illegal/impossible chess moves before the agent could choose them?
Play chess with Codex, @OpenAI .
I built chessctl, a tiny hobby project for testing whether coding-agent guardrails can help LLMs play chess: inspect the board, verify legal moves, then explain the move.
https://t.co/AJQsRHB8Ju
@OpenAI key idea: the next context only contains legal moves.
The agent does not need to remember every chess rule from weights. It queries the board state, sees valid moves, picks one, and explains the choice.