We built a tool to use Git Worktrees to manage all of our in-progress agents. It's expanded from just Claude Code to having Codex support, and I'm loving running them side by side. They see things from different angles, make different UX choices, find bugs the other one doesn't, etc.
Just having the model variety is helpful, it is my default way to work now for every prompt I submit.
I've built something very similar but am further along (https://github.com/stravu/crystal). I understand the desire to do your own thing, but if you are interested in joining forces and contributing I would love to have you. I think we were thinking along very similar lines.
I use Crystal which archives all my old claude code conversations, I've had to do this a few times when I threw out code that I later realized I needed.
It supports Claude Code and Codex, but has you constantly working on multiple features in Git worktrees. This way you are always able to stay busy while waiting on your agents.
It has built in tools for review, such as a diff viewer, and a quick button to run your application in different worktrees for testing. It has completely transformed the way I work.
We've added support for Codex in Crystal, meaning you can now run Codex sessions alongside Claude Code in isolated git worktrees for parallelized development.
This is a great list for people who want to smugly say "Um, actually" a lot in conversation.
Based on my brief stint doing data work in psychology research, amongst many other problems they are AWFUL at stats. And it isn't a skill issue as much as a cultural one. They teach it wrong and have a "well, everybody else does it" attitude towards p-hacking and other statistical malpractice.
IMO this is the best long context benchmark. Hopefully they will run it for the new models soon. Needle-in-a-haystack is useless at this point. Llama-4 had perfect needle in a haystack results but horrible real-world-performance.