Idk how you debug tricky bugs, but I usually don't just stare at the code and take a guess. So I don't fault LLMs too much for being wrong in that scenario. So a couple months ago I built a vs code extension to let the llm debug the same way humans do- placing breakpoints and evaluating expressions in stack frame.
Works pretty well.It's an mcp server so works with any mcp client and because it uses vs code debugging API, it's language agnostic.
MIT Open Source - Contributions welcome
https://marketplace.visualstudio.com/items?itemName=JasonMcG...