You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
One feature that is available in other agents that would be useful here is the option to go back to a previous message and edit it and restart the conversation from that point.
This would drop the history that comes after that message, but would also have to undo the changes made in the workspace.
Here's an example of how this is implemented in something like Copilot:
Screencast.from.2026-03-09.11-03-42.mp4
In this example, I prompt the agent to add some console logs, then when I modify that message I get a modal to confirm that I want to restore the history and code to that state.
We could do something similar, or perhaps even better would be to implement a branching history, so users don't lose their previous conversations or modifications when editing their messages.
I imagine the hardest part is undoing all of the different commands an agent can execute and I'm not yet sure how feasible this even is.
We could look at this PR for inspiration.
We might also be able to use the Jupyter checkpoint system (.ipynb_checkpoints).
One feature that is available in other agents that would be useful here is the option to go back to a previous message and edit it and restart the conversation from that point.
This would drop the history that comes after that message, but would also have to undo the changes made in the workspace.
Here's an example of how this is implemented in something like Copilot:
Screencast.from.2026-03-09.11-03-42.mp4
In this example, I prompt the agent to add some console logs, then when I modify that message I get a modal to confirm that I want to restore the history and code to that state.
We could do something similar, or perhaps even better would be to implement a branching history, so users don't lose their previous conversations or modifications when editing their messages.
I imagine the hardest part is undoing all of the different commands an agent can execute and I'm not yet sure how feasible this even is.
We could look at this PR for inspiration.
We might also be able to use the Jupyter checkpoint system (
.ipynb_checkpoints).