Overview & result
Winnow scores blocks of large Claude Code tool results before they enter context. Confidently irrelevant blocks become recoverable stubs, while uncertain/error blocks remain verbatim. Jev is the default judge; a Haiku adapter is a distinct fallback.
From the original post
A calibrated context sieve for Claude Code.
How Jev fits in the loop
- Intercept large Read, Bash or Grep results and split them into blocks.
- Ask Jev one Noul relevance question per block against the current task.
- Keep uncertain and error-bearing content; replace confident drops with a stub and recall key.
How to reproduce
- Install the documented Claude Code function hook and configure a Jev key or explicitly choose the adapter.
- Run a large Read/Bash/Grep result and inspect the stub plus recall key.
- Use shadow/replay tools and hand-labeled cases to choose thresholds before enabling active pruning.
Why this build matters
A measured, recoverable context-sieve pattern with explicit handling of uncertainty and a clear distinction between Jev and adapter results.
Limitations
- Without a judge key, outputs pass through; the Haiku adapter is not calibrated Jev and must be labeled separately.
- Uncertain blocks and detected errors are kept; a low regret rate still requires real user review.