Tarsk now has additional tooling to keep long running session from stopping when the context window limit is met.
Tarsk now makes room and keeps going. Compaction shrinks what the model sees next.
The context window fills
A model has a fixed amount of room for the live request. As it works to complete your task that context window is slowly eaten up until exhausted and would typically error out. To prevent this a technique called compaction can be used to summarize the work done so far so that it can be completed.
Tarsk has always had a “microcompaction” step that minimizes context at appropriate times to reduce the chance of exhausting the context window. It also has had a feature to auto compact when you give it its next prompt.
The new technique now detects when exhaustion is going to happen and takes steps to fully summarise the conversation so that if it has not yet completed a task that it will have the opportunity to.
How to use it
Give Tarsk a multi-step coding job. Watch the context meter if you want a progress check.
Compaction will happen automatically if it is needed. You can also type /compact if you ever want to manually compact the conversation.