Before you dive in
What you’ll find in the original
- Compaction should preserve a coherent account of what the agent tried and learned; judging each tool call without its result or surrounding thread can make the agent repeat failed work.
- Some providers require an intact history to reuse encrypted reasoning data, so deleting earlier messages can remove information a separate classifier cannot inspect.
- Changing an early message invalidates cached content after it; when cache writes cost more than reads, deleting text can increase rather than reduce total spend.
Worth knowing
This is a critique of one fast-jev-compaction design, not evidence that every classifier-assisted context strategy fails. The cost observations come from the author's personal Claude Code and Codex usage.