skillsmith

An agent (with subskills) that audits, optimizes, tests, and self-heals other agentic skills and prompt-based tools — starting with the Claude Code SKILL.md format and the Claude Agent SDK / MCP ecosystem, architected to extend elsewhere.

• Architecture — packages, data flow, the Adapter boundary.
• Extending — write an Adapter for a new ecosystem.
• Self-healing — autonomy tiers, config, why human-in-the-loop.
• Future extensions — the roadmap.

The one-paragraph pitch

skillsmith reads a skill, classifies each work-block as deterministic-candidate or judgment-required, and only ever offers to convert the former into a tested script — cutting tokens and nondeterminism without stripping out judgment where it matters. It treats temperature: 0 as necessary-but-insufficient and measures output variance empirically. Every repair it proposes flows through detect → diagnose → propose-as-diff/PR → human-approve, with a full audit trail and a layered sandbox around all untrusted code. No silent commits, ever.

Install & run (no preinstalled runtime)

./bootstrap.sh          # installs uv, a pinned Python, and all deps from uv.lock
uv run skillsmith --help

See the README for the full quickstart.