I used a claude pro subscription back in the opus 4.x days. If the usage limits filled up you just have to wait until next week. I wasn't in a hurry so this was fine. For testing, I wrote (this was hand coded by me since I found that part mentally stimulating) a couple of scripts that the clanker could use to drive the image editor headlessly: A headless sway instance, a daemon implementing the wayland virtual keyboard and mouse protocols that accepts commands via an unix socket, and a couple of bash scripts to tie everything together. Then I gave the agent instructions to use this setup in claude.md, the basic command set is `./headless.sh {start COMMAND|screenshot|mouse move X Y|mouse click X Y|type STUFF}`. My first attempt was using accessibility APIs by giving the agent access to pyatspi2, but there were a lot of things that didn't work properly. Using input mocking was far more reliable.
So the development was done with opus, the interactive testing was done with a sonnet subagent for cost reasons (cheaper model & doesn't inherit the long opus context). Most of this testing was completely automatic, sonnet caught most bugs itself. I still had to test everything manually however because headless sway doesn't behave exactly like a real DE with kde's platform theme, but this was just a cursory check. Most of the work was me specifying how I want every tool and panel to work.