Dabbling in GEPA. Codex's /goal on GPT 5.5 high is still surprisingly reward-hacking I had set a /goal before I slept to implement a plan. It ended the loop after doing just 1 iteration It feels like the model is following the path of least resistance and slacking off. Though it could also be me putting "try to make good progress in 8 hour's time" in the prompt, can't be sure Lesson: When you are doing such a solver loop, always specify min_iter and max_iter