I cannot begin to tell you how many places I tell the stupid agent not to ever edit replit.md. Including more warnings than you can shake a stick at inside replit.md itself.
So what does it go and f(&( do??? Edit it. It is even so brainless that it copies over the line that says “do not edit this file”. FFS!
Will someone at Replit please add one line to your system instructions:
if the user says never to edit replit.md, then never edit replit.md
PS, the agent followed my rules until version 3 (autonomy) was launched. It is only since then that it ignores me.
Unfortunately, the system is so incredibly close to magic in many ways (more than not, tbh), yet in some ways, it totally makes you want to use profanities and just scream in agony….
I do agree that there should be a way to fully stop Agent from touching certain things….I have had environmental issues several times that it introduced but could then not resolve anymore….
I will take the good with the bad any day tho.
It seems that it isn’t agent modifying replit.md, but Replit’s automatic checkpoint system. Well, that is what the agent has just told me - essentially it said:
“I didn’t change replit.md and if you don’t want me to then I promise I never will. But I have just investigated and discovered it is that nasty automatic checkpoint system doing it. Sadly I have no control over that tool, so you’re screwed mate.”
Totally, there should be a way to revoke write access so the Agent still has context but can’t overwrite anything. If something in that part of the app has recommended changes, it could give you code snippets to recommend replacements or something.
Any quick way for Replit to implement would be to let us make any file in the file tree read-only.
And I love your idea of agent or other tools outputting: “here is some useful info that you may wish to add to your replit.md, readme, or other documentation”