OK, so I’ve got a very consistent problem with my environment. I often use cmd.exe or WSL bash. BUT NEVER USE Powershell. Apparently every model that we throw at it also doesn’t have a whole lot of experience with Powershell and often messes up the commands with build processes and debug stuff. So easy solution, I tell it to switch to cmd.exe then it continues to try to use Powershell formatted commands within cmd.exe, or run bash commands from cmd.exe. This is an issue with model context that can be fixed. You could allow preferences for which terminals are used for what stuff. And then you could just build the proper context chain so it would output the proper format commands for the given shell that you’re using.. At least I think it would be a whole lot better than watching the stupidity that agent mode does when trying to work on projects: ie shell execute 20 different make directory commands in a row instead of making batch a batch file. and screw them up. each time iteratively with the wrong shell context.. mixed with not realizing that a tool call failed…
That brings me to the second issue. The tool integrations. oftentimes on my system, when executing a shell or shell command via cursor, hangs and does not open, or it opens, but for some reason, the command never executes, and it just gets in a stuck state. Then I either have to press stop, or I have to, actually, I have to hit reject. And then I have to hit stop. But the stop button moves. due to the format and the size of my screen. So that’s an annoyance. But the bigger issue is that when the shell fails, it either returns no error code or one that is never checked because the model just assumed that the command was successful and then tries to plow ahead forward against the output of that command that failed like trying to run the. compiled C program that was never outputted, instead runs the old one with the existing bug, causing the model to go insane, further lose context, causes drastic untracked changes and trashes the project.. It does this consistently enough that I have learned to work around it somewhat.. Furthermore, it also confuses files and context ie if you have two main.cpp’s in the same folder tree. May God have mercy on your soul..
I don’t know if I’m an edge case where the terminal never fails like that, but it does on my machine, however all of this stuff is accountable for programmatically at some point. And just a suggestion that on the agentic mode, have it tracking MD file and make sure that it tracks its progress to make sure it doesn’t lose track and just stop in the middle like it very often does to me. I will ask it to set off on a series of stuff. It will request like a call and then it will sometimes even get confused and think I’m answering as itself. (I don’t know. Maybe this is caused by the output failure of the terminal?) So I think we have a prompt chain coordination issue here.
I honestly didn’t think the screenshot or anything, so you’re kind of just going off my word here, but if anybody else experienced it, they will know exactly what I’m talking about.
running Windows 11 Pro x64 23H2 with wsl 2
I’ve been slowly working on my own coding tools, so i understand what you are up against. Let me know if I can help!
Sorry about any voice to text typos. I tried to get most of them.