Rollback Disabled and Suggestion to Create a New Chat
Rollback has been essential to my project. The agent always suggests rollback as one of the many options it provides, but since rollback has been disabled, I suggest the agent should not offer it as one of the options.
Next, it happens too many times that the agent, after coding for about 2-3 lines, already offers to create a new chat. This means the AI will forget your last history, which doesn’t benefit our coding experience because the AI seems to have an “Alzheimer’s mentality” where you have to inform the agent again about what the last issue was, potentially screwing up the whole architecture.
I understand that Replit needs to make money to ensure the agent keeps billing us; however, it’s really frustrating if the whole architecture gets broken, and you have to piece it back together. As a non-tech person, most of the time I feel lost.
Way back in October and November, the agent was tremendous, but today I feel like it has degraded so much. We expect AI to improve over time.
yea - i just use the assistant nowadays. if the assistant doesn’t work, i may try out the agent. otherwise, just use the agent when setting up your initial prototype.
Rollback is back, but the new problem is whenever there’s a checkpoint, agent will tell you that’s it too complex where we just had a few line items coded (just created a new chat) where it seems it’s asking you to create another new chat…I hope replit see’s it
This is relatively new behavior, and the reason behind it is probably (sadly) in an effort to reduce their LLM expense. When the chats get long, in order to not be completely useless, they have to send ALL of the context, or as much as the LLM api will let them. The more context, the more tokens, the more cost. So they’re trying to make money (or lose less money) by encouraging you to start new chats constantly (and honestly after a very minimal period of time).
YOU MUST SETUP GIT, because this new process basically ruins Rollback feature. Git is your lifeline now.
I know the guy with the great walkthrough on that will jump on here and post a link in 3, 2, 1…
yeah, I read in a couple of post that git is the lifeline, I just started exploring it now. Thanks! however there was another issue that came back, where you provided a command, it will execute then after a couple of line it comes back again with your original command,where it will mess up again the whole thought process of AI
Checkpoints are git commits that also tie in database versioning, the reason we have to disable rollbacks for old chat sessions is because for many users there’s often discontiguous changes between the database schema and the code when running multiple chat sessions. If you know what you’re doing you can do some manual code surgery to do the database rollback too after restoring from git but most of our users aren’t pro devs so it’s disabled until we can make the most common interaction with it safer and more reliable without wiping your db.
There’s a project underway right now to change, simplify, and make more reliable both the rollbacks and git experiences, should be seeing serious improvements here soon.
I just did it yesterday, After reading some threads, Git is the best tool for any rollbacks, so if the agent is annoying lolz, I just use Git it works like wonders you need to setup your git account at github then replit auto links it once you go to git
Every time the Agent or assistant take an action they will:
do a git commit to checkpoint your code before they act
take their actions
do another git commit to checkpoint your code after their action
So that you can restore the code to a working state in case the agent/assistant break something.
On why we disable rollbacks on old threads: it actually has nothing to do with costs, it’s everything to do with database and code versioning. When you do multiple chat threads, they don’t start from the same commit, git is graph based version control, and on the product side we haven’t made it easy enough for non-technical users to be able to map and resolve diverging git histories to make it safe for them to rollback a non-linear history. it was turned on for a while and caused a huge backlog of support tickets of people thinking they’d wiped their code by branching and being completely unable to manage merge conflicts.
Allowing a rollback to 1 of multiple threads at this point when people still treat the threads like ChatGPT and most of the user base is unfamiliar with graph based version controls is a huge footgun, so it’s by default disabled.
Like I said, there’s a complete redo of the product experience of checkpoints and rollbacks in the works that should be much more effective and mitigate the most common issues.
You can still restore it manually through Git or the shell.
you need to setup your git account at github then replit auto links it once you go to git