Log Reducer

Sits between your logs and Claude's context window, running 18 deterministic transforms that collapse stack traces, deduplicate repeating blocks, factor common prefixes, and strip noise like health checks and UUIDs. Exposes a `reduce_log` tool that typically cuts 70-90% of tokens while preserving errors, warnings, and your application frames. The transforms are rule-based and instant (no API calls), handling Python exception groups, Java stack traces, Docker layers, and pip output. Works as an MCP server or CLI. Most useful when debugging production errors where raw logs are hundreds of lines but the signal is buried in framework noise. Supports a multi-turn funnel pattern where Claude surveys a large file's tail first, then targets specific sections or levels without loading everything into context.