The Rack now handles summaries and context selection
While the work platform runs from the Rack, long chats are summarised and trimmed by the models you put on those tasks, all on this Mac.
Side tasks get their own models
When a chat grows long, Bandicot condenses the older part. Running from the Rack, that summary is written by the model on Text, Summarization, or by your Text, Causal LM model when Summarization is empty. Picking what to keep goes to the model on System One, Context Selection, when you set one.
Each side model runs beside the main one instead of replacing it, so the model doing the work is never reloaded to write a summary.
Never a quiet detour
If the Rack's summary model cannot start, Bandicot skips that summary and tries again later. It never hands a Rack chat to a hosted model you did not choose.