Skillforge

Skillforge turns successful agent traces into reusable skills.

It compiles a source run into:

• a portable skill.contract.json
• an OpenClaw-friendly SKILL.md
• a verification.report.json
• example inputs for reuse

The goal is simple: stop losing good agent work in transcripts. Capture what worked, parameterize it, attach approval gates, and make it reusable across OpenClaw or your own agent stack.

Why it exists

Most agent tooling can execute one-off tasks, but successful runs usually die as logs. Skillforge promotes a successful trace into a reusable asset:

• extract inputs like paths, URLs, repositories, dates, and emails
• convert raw steps into a reusable execution plan
• infer tool requirements and approval gates
• export an installable markdown skill plus a portable JSON contract
• statically verify that the generated skill can be re-rendered safely

Install

npm install
npm run build

Or run the CLI directly in development:

npm run dev -- inspect examples/fix-flaky-test.trace.json

CLI

Compile a trace into a skill bundle:

npm run dev -- compile examples/fix-flaky-test.trace.json --out generated-skills

Inspect a trace without writing files:

npm run dev -- inspect examples/publish-weekly-report.trace.json

Verify a generated contract:

npm run dev -- verify generated-skills/fix-flaky-auth-test/skill.contract.json

List a local skill registry:

npm run dev -- list generated-skills

Supported input shapes

Skillforge accepts:

• normalized trace JSON with objective + steps
• objects containing messages
• objects containing events, entries, or trace
• JSONL event streams

The compiler is intentionally conservative. It aims to create a useful reusable skill from incomplete data without executing any part of the source trace.

Output structure

Each compiled skill bundle contains:

• skill.contract.json: portable contract for any agent runtime
• SKILL.md: Markdown skill for OpenClaw-style skill registries
• verification.report.json: static verification result
• inputs.example.json: extracted example values

Development

npm run check
npm test
npm run build

Roadmap

• adapter plugins for more agent trace formats
• richer policy engines for command risk classification
• replay-backed verification harnesses
• registry publishing and trust receipts