jevbooks

← All projects

winnow

GhalebDweikat/winnow

Calibrated context sieve that judges tool results before they enter Claude Code context.

What it is

A Claude Code plugin for developers running long agent sessions. It intercepts large Read, Bash, and Grep results, splits them into ~25-line blocks, and replaces blocks the judge is confident you don't need with a three-line stub that can be restored on demand.

How it uses Jev

For each block, winnow asks Jev one typed yes/no question — "is this block needed for the current task?" — using the block text, a short task description from the transcript, and the tool's arguments as state. Jev returns calibrated probabilities in one parallel call; confident-yes and uncertain blocks stay verbatim, confident-no blocks are cached, summarized, and stubbed.

Primitives:choice

Technique worth stealing

Batch many per-block yes/no questions into a single System One call and threshold the calibrated probabilities to decide what enters context.

Try it

uv run --project winnow/sidecar winnow demo --fake

View on GitHub

judged by Jevjev-1.13.0

Evidence

Each line is one question put to Jev about the README. ≥ 0.60 reads as yes, ≤ 0.40 as no; in between Jev is not making a call.

  • Jev-centricyes0.70
  • Shows a System One patternyes0.73
  • Handles uncertaintyno0.11
  • Measuredno0.07
  • Runnableyes0.84
  • Worth recommendingunclear0.43
  • Model replicano0.05
  • Problem scopescore on a 0–2 scale1.08
  • About Jevyes0.96

Signals by Jev jev-1.13.0, card written by DeepSeek V4.1 Flash from the README on 20 Sept 2026.