compaction

About

Compaction is a local tool that runs underneath Claude Code, Codex, and Cursor to shrink the input and output context sent to the model provider before it leaves the machine, without requiring a new editor or agent. Its own acceptance run reported an 8.39M-to-7.21M token input reduction (about 14%) and roughly 25% less generated output, with the measured before/after numbers shown inside the existing tools.