There is an issue with when the cursor agent executes powershell commands in chat, were it cannot correctly read the results of what it executed… this often causes the AI to try again, or worse come to the wrong conclusion and start coding under that assumption. I have posted about it before… but perhaps a screen shot with the cursor agent attempting to explain it will work better.. Please please fix this…
I think its related to this file.. but not sure
I have a suggestion:
What I have found to be helpful for me - is to direct the bot to always wrap its agentic commands in .ps1 scripts and execute those and store them in a /scripts/ directory.
This allows for reusable actions - but it gets beyond buggy behaviour such as the bot attempting && chaining of commands.
Plus it allows for logging properly, and structure.
So tell youre bot to:
(I akso find that the following phrase is powerful:
GIVE ME XYZ thing – the ‘give me’ part seems to be a best in class prompting intro for great justice. (if you know the meme)
but dont let the agent just run stuff - tell the agent to GIVE YOU scripts which it then runs. Log what it does, build proper headers in the scripts with description and invocations
Make the bot auto-dialog.
Tell it to state its “give me your understanding of entanglements and directives”
Change your terminal to cmd in cursor.