Not something I wanna decide now, but I do wonder if that's not a better model going forward. e.g. we added GLM 5 support and also closed ~6 PRs of varying quality that all tried to solve the problem in partial ways. https://github.com/openai/codex/discussi…