Using Codex also brings new pain points: A sluggish architecture migration is a common problem in every AI system, and it's particularly evident in GPT 5.4, with various obscure compatibility layers. Furthermore, Codex requires adding tests every time it's modified, resulting in more test files than implementations! While it seems stable, it's incredibly cumbersome. The result is that refactoring is extremely slow! And every time, I have to say "continue," even after sitting in front of my computer for two whole days, Codex is still doing the same thing! The image below is Claude's sharp review of Codex: