Reflections on AI behavior and a significant improvement to the mindspace tree editing code that removed over 300 lines of brittle code with a much more robust solution.
How to confuse an LLM
I'd been asking Claude Sonnet 4 to help fix an issue in the mindspace code, but it seems that resulted in it getting slightly confused. The code contains a "delegate" component, but when I asked for an analysis this seemed to have caused Claude to want to use the "delegate" tool. Could just be coincidence, but those delegates also spun up more child LLM instances too.
Fixing tree editing problems
A few days ago I realized that scrolling the files tree very quickly would result in any open editor for a filename ending up in the wrong place.
After a long discussion with Claude, it turns out the entire approach the code had been using was very brittle. It had taken several days to get it working and had felt somewhat clunky when I first did it.
What's frustrating is that original approach was also built by Claude, but somehow it had not understood Qt well enough at the time.
The new code is far more robust and a lot smaller! This change removed a net 300+ lines of code.