winnow

Reviewed at commit 4547a81

Caution: Saves the full text of every tool result it hides, possibly including secrets, to ~/.winnow outside your project, and installs Python packages with uv at session start.

Sends data to TypeSafe Jev judge (typesafe.ai; receives the tool output blocks, task description and tool arguments)

A calibrated context sieve for Claude Code. Every large tool result is judged by a System One model before it enters context; irrelevant blocks are replaced…

Made by
Ghaleb Dweikat
What we checked

Sends large Read, Bash and Grep results to a local sidecar that asks a judge model which blocks matter and replaces the rest with a short summary and a recall key before Claude sees them. Tool output goes to TypeSafe by default (or to Claude with your own API key), and the hidden text is cached on disk. It also adds relevant memory files to your prompts.

Review details

Static review of the source at the pinned commit; not run. The module registers tool.call for Read, Bash and Grep (waits for the real result, and when it is at least 1500 characters POSTs it to http://127.0.0.1:47311 and may replace the result) and prompt.submit (sends the prompt to the sidecar and may append memory files as context), and uses $.http.fetch to loopback, $.process.run to start the sidecar (uv run python -m winnow serve --revive), $.fs.stat, $.ui.toast and $.ui.log. hooks.json also has a SessionStart shell hook that runs uv run with the bundled sidecar project, which syncs Python dependencies from uv.lock (typesafe-sdk, system-one-adapter, anthropic, mcp) and starts a detached resident server that exits after 45 idle minutes; .mcp.json adds a winnow MCP server (winnow_recall, winnow_stats) started the same way. The sidecar source (read in the repository, not in the bundle) reads ~/.winnow/env (TYPESAFE_API_KEY, ANTHROPIC_API_KEY), the session transcript tail and ~/.claude/projects/<project>/memory/*.md, calls the TypeSafe judge by default or the Anthropic API (summaries, adapter judge), and writes the full text of every hidden output, a decision log and serve.log under ~/.winnow, with no automatic eviction; the directory is made user-private. The server rejects requests with an Origin header. Fails open: when the sidecar is unreachable, results pass through unchanged and a toast is shown once. Needs CLAUDE_CODE_ENABLE_FUNCTION_HOOKS=1; the README states what is sent where.

What it uses
  • Reads your prompts: yes
  • Sees tool calls: yes
  • Blocks or changes tool calls: yes
  • Reads or writes files: yes
  • Runs programs on your machine: yes
  • Calls the model (uses tokens): yes
  • Sends data over the network: yes
  • Stores data: yes
  • Shows something on screen: yes

A calibrated context sieve for Claude Code. Every large tool result is judged by a System One model before it enters context; irrelevant blocks are replaced with a summary and a recall key.

Made this mod? Add a badge to your README

modsforClaude badge for winnow

Paste this into your README. The badge says reviewed once the current version has been reviewed, and links readers to this page.

[![Listed on modsforClaude](https://modsforclaude.com/badge/winnow.svg)](https://modsforclaude.com/mods/winnow/)

← All mods