Since May, I've been using Cursor to build some projects that I'll certainly be sharing more here. It's been pretty invigorating. I've learned more in the last six months than I can remember learning at any time. And that includes grad school.
Cursor has a variety of features to help you build software, but ultimately the AI is always constrained by something called a context window (basically the maximum number of tokens it can process at once). When you hit that limit, the software automatically summarizes the conversation to that point so you can get back to a smaller batch of data and keep working.
Summaries naturally lose detail so the quality of the coding agent can go down. But the context windows are usually big enough that you don't need to summarize until you're pretty far into a conversation.
Until today. This is easily the weirdest and most unsettling bug I've encountered in Cursor. All of a sudden, the context window is filling up instantly, a summary is generated, and the AI starts working on something I never asked for it to do (a flask app in Python).
Is this data contamination thing? A prompt injection in Cursor's harness? Luckily I was working on something trivial, so nothing important has been lost or damaged. Also, I'm apparently not the only one to have this happen.
I'm going to work on other things and definitely leave my other projects closed for now.