What it is
Goal-based chunk filtering before an agent reads.
What it does: Skims pages, files, grep output or lists with one Jev relevance question per chunk and returns only the kept chunks, cited by path and line range.
Why it belongs in the showcase: A small, installable tool-result sieve that any coding agent can call from AGENTS.md or CLAUDE.md to cut context size.
Caveats: No labelled recall benchmark; the 0.5 threshold is a default, not a tuned value.
Evidence
Public MIT repo with Homebrew, apt, Scoop and go install packages. Text is cut into about 2,000-character chunks with 10% overlap; Jev returns P(relevant) per chunk or list item in parallel batches, and chunks at or above the threshold (default 0.5) are kept with path:line citations. When Jev cannot judge a chunk, skimmer keeps it instead of dropping it. Prompts are editable templates. It is the chunk filter from the author's webctl tool, split out for any material.
Caveats
- The README shows example savings, not a measured recall rate on a labelled set
- Needs a TypeSafe API key or a credential-injecting proxy
- Content of every chunk is sent to Jev
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.