What it is — Compresses agent context so small windows can hold long coding sessions, claiming ~100K is enough and roughly 5× fewer tokens.
Who it is for — You're on a ~100K-window model in DeepSeek Harness, dragging one refactor across many days, with context length and token use already against the ceiling.
Teams that cannot let a plugin run shell commands should stop here; that is a policy gate, not a window-size problem.
Watch out — It installed on a fresh profile and the harness registered it. We did not measure compression ratio, month-long sessions, or the stated 95–97% prefix-cache hit rate. README budgets ≤2% for compression itself and names /acp and /acp-cache when cache health falls.
The verdict — I would rather hit the wall once, then enable it, than spend a working session proving out a compressor this run never benchmarked.