Codex

On Windows you must spell out the OS and PowerShell version in AGENTS.md, otherwise it mixes syntax from different versions in its commands and you get a pile of errors. I also added a rule: use Python whenever possible, which is much more stable across environments.

When the desktop app runs tasks, the whole computer crawls: the browser barely opens and typing in a document lags. Switching back to the CLI fixed it at once, so the problem must be the app itself.

I don't just hand it the keys. First I set up guardrails (tests, permissions and rollback) to keep the blast radius of mistakes small enough, and only then let it build, debug and deploy on its own in the dev environment. With those in place, that part really can be handed off.

On a bad day, adding five to seven lines of code can make me wait twenty minutes, and opening a simple PR takes twenty minutes or more, with 429 retry failures along the way. A one-minute job dragged into twenty; that day one release took me four hours.

The first time felt like discovering a new world; the other AI coding tools I'd used weren't in the same league on real projects. But after two months I started hunting everywhere for cheaper usage plans: at this intensity on the official plans, the money just won't hold out.

Hitting the Plus limit isn't a matter of waiting a few hours; I couldn't use it for two or three days straight. Only later did I figure out there are two limits stacked: a five-hour window, plus a weekly total on top. Tasks already running before the limit kicked in could still be checked on and run to completion.

I used to run five CLI sessions at once without hitting limits. After the update, ten or twenty prompts and it's empty. High-reasoning mode, long sessions, Skills and MCP all eat quota, so now I have to watch for when to switch to a cheaper model.

After a few months it's almost unrecognizable. Since the CLI's engineering ability improved, I've been alternating between it and Claude Code, splitting projects between them, and I can no longer tell which is clearly stronger. The early version was rough; reviews from a year ago can basically be ignored.