Cursor has removed these options from 2.0, and it seems “others” now includes Rules. Could users be given the ability to opt out of this new “self-gathering context”?
I have Rules (and other workflows) that I intentionally attach manually only when needed - this helps control context window size and workflow behavior.
That flexibility has been removed in 2.0, yet there’s no guidance explaining how self-gathering works, how to disable it, or how to adapt existing setups.
Current state:
Similar issue here. I have to manually attach a cursor rule and code files on every new message I send. Cursor 2 has been unable to account for any of this and it’s been a nightmare. It has completely disrupted our company flow today.
Cursor has been our platform of choice up until now because it allowed us to use agents in a way that best suited our needs. We’re now being funneled into this overly-simplified approach that straight up removed functionality for our use cases and it’s left us scratching our heads.
hi @denizengt @j-nen all the options you mentioned (automatic, manual, …) for rules still exist and work. You can also go to chat and type @ followed by the first letters of rule name.
In case some feature does not work please post a full separate Bug Report with more info so we can check and fix it Create Bug Report
Hi @condor ,
I appreciate the quick response. Unfortunately, this is not a fix for us. We implement “beta” rules all the time for our team to test internally before applying them intelligently company-wide. This allows for us to test rules by requiring team members to intentionally include them in the context window of any given chat. This flow has worked great up until today.
The methodology you provided would mean that every employee is immediately affected by these changes to their context window, and we no longer have the means of actually testing if the rules are tuned enough to serve their narrow purpose for a company-wide deployment.
Deploying a change like this to every developer’s context window, with no prior internal testing, is tantamount to deploying a one-shot prompt to main. It’s just not practical.
EDIT: For some additional context, if I include a rule manually in a new chat, the rule is instantly removed from the context window in every subsequent message. This means that I need to manually add it on every new message.
From my understanding and usage a rule is applied when attached and remains in chat history. That AI models have a higher preference for data appearing closer to end of a thread is a model training feature. Adding a rule on each single request in a thread would significantly increase the cost and may result in reduced adherence.
A likely factor that contributes to a model not following rules is context rather than the rule disappearing.
I would be happy to look into a bug report for possible causes and reproduce it.
Hi @condor , thanks again for swift response. Perhaps I can give some additional context here with an example use case of what I mean:
We are currently testing a new python-standards.mdc rules file. In Cursor 1.x, we would persist this in the context window using the @ dropdown (not the @ reference within the chat input field). This rule would then persist for our chat session and we would not have to add it back (it was visible in Active Rules).
Firstly, in Cursor 2.x, this reference no longer gets added to Active Rules (whereas, in Cursor 1.x, it would be present):
Secondly, when a conversation goes through compactification, it loses all context of the rule’s contents; whereas it has full access to rules that we are applying intelligently:
Previously, I believe that manually-applied rules would persist these chat-summary transitions. It no longer persists and the manually-applied rules get lost, requiring us to manually add them back each time this happens.
Thanks @j-nen this outlines my current workflow issues really well. Similar corporate working fallout.
I have a folder where csv log files are outputted and I’m trying to get the AI to read and summerize findings from it, but I can no longer control the context and it keeps reading my code and getting distracted by it. Even when I say do not read my code, only read the csv in this folder and I only include the folder of csv as context, it still starts reading my code and telling me how the CSVs are made instead of summerizing them. you screwed up cursor in 2.0, I’ll have to go back to 1.7 when I’m actually in control of these ADD agents. Also I have a python skd I had it convert to a java skd for my java program, now it keeps seeing the python and trying to write python code; when before I had the option of excluding it from context, but now I don’t. I want to keep the python sdk there, in case I need to have it use it as an example of more things to add to the java sdk.
I rolled back to 1.7 with the same AI model and the same prompt and it works fine; so can confirm that 2.0 is screwed up.
Wow, that is pretty bad. I have not experienced that yet, but maybe there are some specific situations like with your csv file that the new context system totally fails. Very interesting. They are definitely going to have to fix the context.
