This fascinates me because I tend to use Claude Code in _very_ long sessions, driving it until its context window is exhausted at which point I grab most of the session history from the terminal window and paste it right back into the context. This usually fills the context window right back up to 80-100K tokens. Seems a lot more successful especially with keeping track of recent developments than built in compaction does.
For some reason I get the best results this way. I know it's unorthodox but with my approach the agent seems to learn about the ongoing concerns as it stays 'in the loop' and I prefer it to use its minion agents to do grunt work like grepping sources or log files. That way the main context is free of monotonous blobs. I like having it act as a coding/troubleshooting companion rather than a minion to delegate short bursts of work to. I believe it's because I rarely feed it a big block of data to parse in a single prompt or let it grep incessantly in the main context that I don't get hit by the dreaded 'context rot'.
This little study seems to line up with my experiences.
This fascinates me because I tend to use Claude Code in _very_ long sessions, driving it until its context window is exhausted at which point I grab most of the session history from the terminal window and paste it right back into the context. This usually fills the context window right back up to 80-100K tokens. Seems a lot more successful especially with keeping track of recent developments than built in compaction does.
For some reason I get the best results this way. I know it's unorthodox but with my approach the agent seems to learn about the ongoing concerns as it stays 'in the loop' and I prefer it to use its minion agents to do grunt work like grepping sources or log files. That way the main context is free of monotonous blobs. I like having it act as a coding/troubleshooting companion rather than a minion to delegate short bursts of work to. I believe it's because I rarely feed it a big block of data to parse in a single prompt or let it grep incessantly in the main context that I don't get hit by the dreaded 'context rot'.
This little study seems to line up with my experiences.