User repository on GitHub
A calibrated context sieve for Claude Code: every tool result is judged by a System One model before it enters context.
- Forks
- 14
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Tools & CLIs
Winnow — GhalebDweikat
A Claude Code context sieve: large Read, Bash, and Grep results are split into ~25-line blocks and judged by Jev before they enter context, with confident-no blocks replaced by a recallable stub.
- Stars
- 112
- Language
- Python
- License
- MIT
- Reading time
- 3 min
- Published by
- Community
- Author
- GhalebDweikat
- Last verified
- Sep 2026
About this project
A calibrated context sieve for Claude Code. Every large Read, Bash, or Grep result is judged before it enters Claude's context. Blocks the judge is confident you don't need are replaced with a three-line stub: what was hidden, a one-paragraph summary from a cheap model, and a key that restores the full text on demand. Nothing is lost; it just stops costing tokens until you ask for it.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
Highlights
- Large Read, Bash, and Grep results are split into ~25-line blocks, and each block gets one yes/no Jev question.
- Hidden blocks become a three-line stub: what was hidden, a cheap summary or deterministic digest, and a key that restores the text.
- Two safety rules are fixed: output judged to show an error is never hidden, and uncertain blocks between WINNOW_DROP and WINNOW_KEEP are kept.
- A resident loopback server makes a judged call cost about 16 ms plus the judge call, and winnow never judges its own files or commands.
- The default WINNOW_DROP of 0.1 came back clean on hand-labeled replay and hides about 5% of text there.
Quickstart
bash
claude plugin marketplace add GhalebDweikat/winnow
claude plugin install winnow@winnow
uv run --project winnow/sidecar winnow demo --fake # pipeline preview, no keyWatch out
MIT-licensed. Needs Python 3.10+, uv, and Claude Code 2.1.260 or newer with function hooks enabled; Jev is in early access and the uncalibrated system-one-adapter can stand in for it.
Questions about this entry
Short answers drawn from the linked project or page.
Hidden text is not destroyed. Each stub carries a key that restores the full text on demand, and the full text stays in ~/.winnow/cache locally. Two safety rules are built in and not optional: output the judge believes shows an error is never hidden, and blocks sitting between WINNOW_DROP and WINNOW_KEEP are always kept. The default 0.1 drop threshold hid about 5% of text on hand-labeled replay.
Per judged result: the tool output in blocks, a short task description taken from the transcript (last user request, last assistant sentence), and the tool's arguments. Hidden blocks alone go to Anthropic when summaries are on, and each prompt sends your text plus the first 600 characters of every candidate memory file. Nothing is sent when the judge is off or the output is under the size threshold.
Yes. Set WINNOW_JUDGE=adapter and it sends the identical request to Claude Haiku 4.5 through your Anthropic credentials, though the README is explicit that those probabilities are not calibrated. Switching to Jev later is a one-line change. With no key at all, every hook passes results through untouched and tells you so once per session.
It changes only what Claude sees: files on disk, the commands that ran, and Claude Code's own transcript are untouched. It does need Claude Code 2.1.260 or newer with function hooks enabled, which is early access and currently one env line in settings.json; winnow doctor checks it and the plugin does nothing without it. Installed plugins are copied rather than linked, so run claude plugin update after pulling changes.
First set CLAUDE_CODE_ENABLE_FUNCTION_HOOKS to 1 in ~/.claude/settings.json, then git clone, claude plugin marketplace add ./winnow and claude plugin install winnow@winnow. Open a new session and read a file longer than about 1,500 characters; a [winnow] line means it is working.
It ranks the memory files Claude Code keeps for the project under ~/.claude/projects/<project>/memory/*.md, excluding the MEMORY.md index, plus any directories in WINNOW_CONTEXT_DIRS against your prompt, and injects the relevant ones so Claude reads what it needs without a round of Read calls.
For a 10,000-token output, Jev at $0.042 per million input tokens is about $0.0004 for the judge call plus about $0.005 for up to four Haiku summaries, under a cent. The adapter path on Haiku 4.5 is about $0.01 for the judge plus the same summaries, a few cents. winnow stats reports actual usage and cost.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: The compaction plugin, in Chinese
YouTube: Ray Amjad puts Jev inside an agentic coding loop and looks at what it costs to run.
YouTube: RoboNuggets puts Jev to work inside Claude Code.
X: Decision outputs in Grep
Other listings of this project
Forks and copies of this project in the directory. This entry is the canonical one.
A Chrome extension that scores articles, videos and feed links with a fixed set of typed Jev questions and turns the probabilities into a read, skim, save or skip verdict, with the thresholds in options.
Jev Demos & Experiments#community#awesomejev#catalog-import
Community
ThinkyMiner/WinnowTypeScript
ThinkyMiner/WinnowTypeScriptMore Jev resources like this
An experimental local context-pruning proxy for Claude Code and Codex that asks Jev, through the Vercel AI Gateway, which history spans to omit, then forwards a rewritten request with omission markers.
Jev Tools & CLIs#community#typescript#compaction
Community
compozy/yoshiTypeScript
compozy/yoshiTypeScriptA Claude Code plugin and npm library that replaces transcript summarization with Jev keep-or-drop judgments, deleting stale tool calls and results while kept text stays verbatim.
Jev Tools & CLIs#community#typescript#claude-code
A Python asyncio runtime that puts Jev in a supervision loop above Codex or OpenCode workers: it judges completion, test sufficiency, and stuck loops, then steers, retries, or stops the job.
Jev Tools & CLIs#community#python#coding-agent
Community
thruwire/foremanPython
thruwire/foremanPythonA Jev-guided context compaction library for coding agents: instead of summarizing old context, Jev decides which tool calls and results still matter while user and assistant text is kept verbatim.
Jev Tools & CLIs#community#typescript#compaction
An Agent Skill that sends closed coding-agent judgments to TypeSafe Jev and acts on the probabilities, asking for yes/no, choice, and score answers in about 250 ms while the main coding model keeps writing code.
Jev Tools & CLIs#community#python#skills
Community
HyunjunJeon/jev-judgmentPython
HyunjunJeon/jev-judgmentPythonA Stop hook that keeps coding agents from quitting early: rules written in plain language are scored by Jev in about 0.7 seconds, and a violated rule sends the agent back to work instead of letting it stop.
Jev Tools & CLIs#community#python#claude-code
Community
noplan-inc/limpetPython
noplan-inc/limpetPythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.