I use pi+codex and OpenAI codex models do much less of this in my experience.