What feature would you like to see?
Hey I was thinking about the compaction mechanism and many times when I'm not in Autodrive but prompting myself, the best conclusions and ideas happen when the context window is almost full. Compaction ruins the nuances there of course (especially as the latest encrypted reasoning tokens that the idea/concept/pan crystalizes with get lost by that) so I was thinking what if you made compaction more configurable?
For example, you could implement it so that when the whole conversation has been compacted, the user could select a number of last turns that (with full inclusion of their reasoning tokens) are added back to the context window.
What feature would you like to see?
Hey I was thinking about the compaction mechanism and many times when I'm not in Autodrive but prompting myself, the best conclusions and ideas happen when the context window is almost full. Compaction ruins the nuances there of course (especially as the latest encrypted reasoning tokens that the idea/concept/pan crystalizes with get lost by that) so I was thinking what if you made compaction more configurable?
For example, you could implement it so that when the whole conversation has been compacted, the user could select a number of last turns that (with full inclusion of their reasoning tokens) are added back to the context window.