According to Agent Enforcer 2, CI should enforce checks for file length and folder size. I already had the file limit enabled—it had been triggering warnings for a few days—but I only just enabled the folder size check in this project.
So I asked the agents to run the CI runner and fix all warnings they found (one file-related warning remained, plus several about the number of executable files per folder).
- First, I launched
Composer 2. After all, it’s supposed to be quite smart. And indeed it was — it assessed the scale of the issues and decided it was easier to raise the CI limits than to refactor the project. - I figured it was just being lazy, so I escalated the task to
GPT-5.1-High. Which did exactly the same thing. - I added a strict prompt prohibition against modifying CI config and re-ran the task using
Composer 2. In the end, it invoked an audit fromGPT-5.1-Codex-Max-XHighand ignored most of the feedback. I didn’t like this approach and the fact that tests intests/were stupidly scattered across folders named “Alpha”, “Beta”, (…). - I rolled back everything again and escalated once more to
GPT-5.1-High. It carefully reviewed the context. Drafted a plan. Made 5 changes. Rolled back the last one. Then decided it didn’t want to deal with all this crap itself — and summoned a Subagent “Senior SWE” powered byGPT-5.4-Highto do the whole job for it
For context: GPT-5.1-High is now choosed as my “Middle SWE” Subagent.