Audits your entire Claude Code harness against its purpose: enumerates every configurable surface (settings scopes, permission rules and modes, sandboxing, hooks, CLAUDE.md/rules hierarchy, auto memory, skills, subagents, MCP servers), then judges it through five lenses — goal alignment, Bitter-Lesson overengineering, stale self-model, memory compounding, and autonomy calibration — producing severity-ranked findings with specific edits and a PASS/CONCERNS/FAIL verdict per lens. Invoke with /harness-audit [lens|surfaces].
