- cpptools is kind of amazing but also pretty bad. It regularly malfunctions for me. It’s essentially undebuggable. I have less experience with the other extensions, but I don’t expect that they that much better.
- The VSCode security story is very, very weak. Extensions are not sandboxed. The client, accessing remote repos, is wildly insecure, by design.
And that second point is a big deal. Maybe when poking at your own company’s code, you aren’t that concerned about your repo attacking you. You probably should be concerned about malicious extensions, but we’re all far too used to trusting dev tooling. (Don’t forget that MS’s extensions are just this side of malicious.)
But, with AI, you should absolutely not trust your LLM. It is entirely unsafe to give an LLM that might try to exploit you the ability to write into your repo directly or to run JS code in the context of any portion of VSCode. And it’s really quite easy to convince LLMs to engage in all manner of shenanigans.
There’s an opportunity to make a better ecosystem. In a better ecosystem, the equivalent of cpptools would not have telemetry, because the equivalent of cpptools would not have Internet access. It would have the ability to read the workspace, to create cache files for its own data, and to operate its UI, and that’s all.