Performance getting worse and worse on a long-term project (workspace rules, NOTES/LOCKs, French prose), partial delivery, doesn’t follow the rules, misleading in its output, extremely difficult to keep working? The French is just gibberish. Seriously, it’s really bad.
Steps to Reproduce
Performance getting worse and worse on a long-term
Expected Behavior
an agent who does the bare minimum, doesn’t make a fuss, and does what is asked of them
Hey, thanks for the report. I get it, when quality drops on a big long-running project, it really gets in the way.
A couple things to check:
Auto picks a model dynamically. If you need predictable behavior on a complex project with strict rules, try selecting a specific model manually instead of Auto. It should be more consistent. More on how model selection works: Cursor Router | Cursor Docs
Long context. In long sessions, the model can start to drift. It follows rules worse and forgets earlier instructions sooner. Starting a fresh session for a new task helps, and keeping the context focused helps too, instead of pulling the whole project history into one chat.
Rules. Make sure the rules are actually being applied Cursor Settings > Rules, and that they are set to always-on where needed: Rules | Cursor Docs
To dig into it, can you share:
The Request ID from one of the problematic answers right top corner of the chat > Copy Request ID. This works if Privacy Mode is off.
A short example of which exact rule got ignored, and where the French gibberish showed up
With that, we can look into the specific case. Let me know how it goes.