I’m working with Cursor on Windows and I’ve been having a repeated issue with terminal commands not running properly.
This seems to happen most often for me with docker commands. The preview in the terminal command area appears correct, but when the tool is actually called, it is missing the leading character. It does this to me probably 40% of the time, resulting in a lot more tool calls than should be necessary.
For example, if the agent determines that it should run a command like docker compose up, it will display that as the command to be run, but when it actually runs the command it will run ocker compose up, with the d missing.
I’m using the claude-3.7-sonnet model with thinking enabled.
The AI always catches on that the last command run had a typo and attempts to fix it, though even the attempts to fix it can fail with the same problem. The AI itself clearly knows what its supposed to run, but the execution of that is what fails.
