I’ve noticed in the more recent versions of Cursor, maybe just 2.2 but not sure, that many TOTALLY BENIGN commands are now no longer able to be added to the allowlist. Such as this PNPM command:
There is no reason why Cursor should block this command. The agent runs this kind of command all the time, and lately, I have been on the hook to APPROVE about HALF of the benign, non risky, NEVER WILL BE RISKY commands the agent tries to run. This is ridiculous, guys. I have absolutely no interest in a nanny-IDE. I’ve been writing code since I was EIGHT!! Maybe even before that! I don’t need my hand to be held, no, slapped away, by Cursor. Please stop screwing around with the terminal allow list feature, stop trying to save me from myself, and GIVE ME CONTROL. I require the ability to be in control of my own allow list! Stop blocking commands that are utterly benign!
Steps to Reproduce
Instruct agent to execute a benign command.
Weep when its arbitrarily blocked and you have to manually approve.
Expected Behavior
Commands run ENTIRELY according to MY allowlist! I choose what runs automatically and what does not. Don’t interfere!
Operating System
MacOS
Current Cursor Version (Menu → About Cursor → Copy)
We definitely want users to be in control of what runs. I love YOLO (run everything) mode myself, and would be super frustrated if I had to approve stuff I already approved.
Looking at your screenshot, I’m wondering if one of the other commands in that chain (cd or head) might be what’s triggering the approval, rather than pnpm. The allowlist checks each command separately, so even if pnpm is allowed, another part could be the culprit.
Can you share how you have your allowlist configured? That’d help us figure out what’s actually going on here.
I guess this is the only way. As you can see, head, tail, cd, etc. are already in there. I have been running into this issue on a lot of commands lately, where the option to add to the allow list is not even present. Now, in some cases, when a command is very long and complex with lots of individual sub commands piped together, I can understand not auto-allowing it. That is fine…it is when there are simple commands like above, that are totally benign (no risk if they were executed), get blocked…it just seems arbitrary. In fact, it seems NON-DETERMINISTIC, which makes me wonder if you are running every command through a model to allow the model to determine whether to block it from the allowlist or not…
I wonder if Cursor has difficulty parsing out which commands might need to be added to the allowlist when you run multiple commands in a single go. I’ll test this out.
In the meantime, adding pnpm list to your allowlist should let it run that command. I don’t see it in your screenshot, though I’ll admit I used AI to help extract the commands from the image. I’m only human!
Ok, I was eventually able to get focus in there. It took a while for focus to actually appear, and I suspect it is a result of the performance issues that are caused by the Diff/Review mode. I think that you guys must be holding on to a ludicrous amount of data for that diff/review mode, or have some kind of actively running perpetual loop of code that operates on the diff/review data, or something.
As I work throughout the day, Cursor becomes progressively slower, to the point it can no longer be used, and I have to quit it and restart (however, if on restart, it immediately reloads all the diff data again, then it can still start out slowish). This performance issue slows down the entirety of the Cursor UI though….all typing slows down, eventually selection of ui elements slows down, it can get to the point where as you type. cursor is WORDS or even sentences behind. I’d never seen the settings ui slow down, or at least, I’d never FELT it before, because usually I’m just in there flipping switches (mostly on MCPs)….I guess its impacted as much as anything else though, but this performance issue the Diff/Review mode imposes.
FWIW, I have an existing bug report about this performance issue.
