autoharness: Automatically learn and organize skills from Claude Code sessions

This layer automatically extracts and manages skills by analyzing actual work sessions during Claude Code usage. Without separate data collection or benchmarks, it identifies and stores patterns from user conversations and tool call logs.

Skills for similar scenarios are merged into one instead of accumulating duplicates, and skills with low usage frequency are periodically removed to keep the library clean. A case study showing performance improvement from 42% to 78% on CORE-Bench demonstrates that harness management impacts performance more than the model itself.

It runs automatically in the background, but when immediate learning is needed, the /learn command can instantly convert the current session into a skill. It never intervenes in skills manually created by users, targeting only skills generated by this tool.