What it is
Fork of fast-jev-compaction for Claude Code: when /compact runs, Jev answers a yes/no per old tool result (“does the assistant still need this output word for word?”) with a preview of the result in the question.
What it does: Kept results stay verbatim; dropped ones move to .jev-drawer/<time>/ with an INDEX.md and a one-line pointer in the session; tiny results are always kept; falls back to the built-in summary if Jev fails or saves <25%.
Why it belongs in the showcase: A second compaction shape next to the parent fast-jev-compaction, differing in what Jev is shown (preview vs one-line note) and in the recoverable drawer.
Caveats: Author says ~0.5 s vs ~35 s built-in — one session, one quiz, author-reported; Claude Code only.
Evidence
Public MIT README; explicit fork of tamaratran/fast-jev-compaction (listed here as fast-jev-compaction). Jev sees a preview of each result; dropped output goes to a readable drawer with INDEX.md; falls back to the built-in summary.
Caveats
- Fork of fast-jev-compaction (already in the showcase) — compare both before choosing; this one quotes a preview of each result to Jev and adds a readable drawer
- Author says ~0.43–0.74 s vs 35.1 s built-in on one 591k-token session; README itself calls it an example, not a benchmark
- Claude Code specific; drawer files are written on disk next to the project
Unofficial case page. Stars / list position observed for heat; do not treat heat as a quality SLA. Review the upstream README before you depend on it.