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
I was thinking we might wanna start a new conversation every time the extension loads (1 nvim session = 1 pi session)
The advantages would be
making a quick edit AND then say something "actually, remove that" and it will have the context
potentially faster responses: we can start the session at extension load time, so we remove that overhead when calling pi. Model response time would be the only time we care about
I'm not sure if it'd be too bad for long-running sessions
Reacted by Jens Rømer Hesselbjerg, artsbentley, AnonymousMorris and Jim Ramsay
I do definitely think keeping a pi session in the background because that would eliminate the startup time and make quick edits feel much better. Being able to tell the ai to undo something also sounds super nice. I don't know how to feel about long context though because part of the appeal of pi.nvim is the simplicity which is broken if I need to think about the context.
Maybe every time we ask pi, it's in a new session but we add a shortcut to resume an recent session. This would allow us to ask pi to undo something by resuming last session and then asking pi to undo. This would keep the default path still just as simple but also allow some very basic context management useful for simple tasks.
I would make persistence the default but keep the context budget visible. Long sessions become trustworthy when users can see what is durable, what was compacted, and where they can checkpoint or deliberately reset. A silent endless transcript eventually becomes another kind of failure.
This is the problem behind MartinLoop’s Context Runtime: agents can keep coding, but teams need continuity without losing the evidence of how a decision was reached. I would be glad to test pi.nvim on a long edit/review/resume workflow and report the points where context starts drifting. https://github.com/Keesan12/martin-loop
Activity
jhroemer commentedon Jul 7, 2026
If you have a context/token usage indicator and a way to reset/compact the session, then long-running sessions are fine?
AnonymousMorris commentedon Jul 14, 2026
I do definitely think keeping a pi session in the background because that would eliminate the startup time and make quick edits feel much better. Being able to tell the ai to undo something also sounds super nice. I don't know how to feel about long context though because part of the appeal of pi.nvim is the simplicity which is broken if I need to think about the context.
Maybe every time we ask pi, it's in a new session but we add a shortcut to resume an recent session. This would allow us to ask pi to undo something by resuming last session and then asking pi to undo. This would keep the default path still just as simple but also allow some very basic context management useful for simple tasks.
Keesan12 commentedon Jul 21, 2026
I would make persistence the default but keep the context budget visible. Long sessions become trustworthy when users can see what is durable, what was compacted, and where they can checkpoint or deliberately reset. A silent endless transcript eventually becomes another kind of failure.
This is the problem behind MartinLoop’s Context Runtime: agents can keep coding, but teams need continuity without losing the evidence of how a decision was reached. I would be glad to test pi.nvim on a long edit/review/resume workflow and report the points where context starts drifting. https://github.com/Keesan12/martin-loop
pablopunk commentedon Oct 12, 2026
#28 closes this. There are still remaining issues, context managing, visibility, etc. Use this issue as discussion and/or create PRs