I think the confounding issue is that by now, millions of sessions of Claude Code and Codex are now in the training set for these models. So they have been trained to work the way these harnesses are configured, and at least in the case of Claude Code the harness itself is greatly stripped down because the model has absorbed it.