I’ve run into a similar issue, compacting is pretty much worthless and leads to a lot of churn.
I have started having Claude pay constant attention to usage and assess whether the next batch of tasks flirts with compaction.
If we’re anywhere close I have Claude create a next steps doc with meticulous notes along with a prompt that I can give to the next llm to hit the ground running.
Then kill the session and start fresh.
At least this way I know there’s a fighting chance of not falling into a death spiral of the llm guessing about what it’s supposed to be doing as if it hasn’t just had a self inflicted stroke.
Then kill the session and start fresh. At least this way I know there’s a fighting chance of not falling into a death spiral of the llm guessing about what it’s supposed to be doing as if it hasn’t just had a self inflicted stroke.