Vibe Coding Discoverwinnow
View on GitHubA calibrated context sieve for Claude Code: every tool result is judged by a System One model before it enters context.
A Claude Code plugin that judges each large tool result with a calibrated System One model (Jev, or a Haiku adapter) and replaces unneeded blocks with a stub plus a recall key. Hooks, shadow mode, and MCP server included; local cache keeps the hidden text restorable.
Use Cases
Shrink large Read/Bash/Grep tool results before they enter Claude Code's contextStub out irrelevant tool output and recall the full text on demand via winnow_recallAuto-select relevant project memory files at prompt timeRun shadow mode to measure pruning quality without changing outputTrack judge token usage and cost with winnow statsEval and replay pruning decisions with labeled judge dataExpose winnow over MCP for other clientsCut context token spend in long Claude Code coding sessions
Built With
- Language
- Python
- Frameworks
- Claude Code Plugin · MCP · uv
Tags
claude-code · claude-code-plugin · context-management · hooks · token-reduction · context-pruning · llm-agents · relevance-judge · calibrated-model · mcp · python · shadow-mode