Automatic context compaction

[ethanndickson]

Right now users have to run /compact themselves, and they have to do so before the model quality degrades, or they reach the model's context limit.

Automatic context compaction would simply run /compact for the user at a configurable percentage of the model's context limit. Users should also be able to turn it off, as power users may prefer to manage plans and context themselves.

The user should be given a warning when the next message would result in a context compaction. The implementation will likely create the compaction request with that next message as the continue message.

[ammario]

Just to clarify, we would /compact <message> so that the conversation flows naturally on the auto-compact message?

Also, I was thinking we could have auto-compaction on two triggers:

1. Context usage as a percentage
2. Chat idle time (configurable)

Re Chat Idle time, this solves the problem reported in Slack of it not being clear to users what exactly a Mux workspace was for if left unattended for a day. It also has the potential to save users from waiting for imminent compaction in some cases. I haven't fully thought through it but could be easier to implement with your frontend -> backend refactor

[ethanndickson]

Just to clarify, we would /compact <message> so that the conversation flows naturally on the auto-compact message?

Yep