LetUen / Agent Skill Anatomy

LetUen is a multi-agent research harness for deconstructing Agent Skills and agent workflows into readable manuals, reusable anchors, Obsidian knowledge vaults, graph data, and release-ready skill packs.

It is designed for people who want to understand how a skill works, extract the reusable parts, compare model outputs, and recombine selected capabilities without damaging existing user skill structures.

Why LetUen

Most skill repositories show the final SKILL.md, scripts, and assets, but not the design logic behind them. LetUen turns a skill repository into a structured anatomy:

  • What it is — identity, trigger intent, boundaries, expected user request shape.
  • How it works — workflow stages, tool usage, scripts, references, assets, and runtime assumptions.
  • Why it is trustworthy — evidence-grounded claims, deterministic quality gates, reviewer summaries, and evidence repair.
  • How to reuse it — anchors, composition plans, Obsidian notes, reusable templates, and non-destructive integration guidance.

Outputs

Surface What it gives you Status
Report A readable skill anatomy manual for humans. Working
Cinema / Repo UI Local static presentation surfaces for decomposition results. Working demo surface
Obsidian Vault Bilingual notes, MOCs, Mermaid maps, evidence notes, and reusable assets. Working
Data / Graph JSONL, CSV, Mermaid, and graph-ready artifacts. Working
Anchor Pack skills/asa-* method skills and anchor composition contracts. Developer preview
Quality Gates Deterministic checks, reviewer summaries, source-aware evidence repair. Working

Repository Scope

This repository currently contains the full LetUen project:

  • Python harness and providers
  • Static UI surfaces
  • report / vault / data / graph exporters
  • internal method skills
  • tests and sample outputs
  • packaged developer-preview skill anchor pack

A smaller skill-only repository can be published later from the package-letuen-skill artifact. The current repo intentionally keeps the full system together so the harness, UI, docs, and skill pack can evolve in sync.

Quick Start

Run the mock pipeline first. It does not require model API keys.

cp .env.example .env
$env:PYTHONPATH='src'
python -m asa run --config anatomy.config.example.yaml
python -m asa export-letuen --run runs\<run-id> --output dist\letuen-demo

Start the local visual surface:

npm install
npm run demo:export
npm run demo:serve

Open http://localhost:4173/.

Real Model Runs

The default config uses mock. For real multi-agent calls, copy one provider config and set the matching key in .env or your shell.

Provider route Config file API key env Base URL
OpenAI · gpt-5.2 anatomy.openai.example.yaml OPENAI_API_KEY https://api.openai.com/v1
Claude · Opus 4.5 anatomy.claude.example.yaml ANTHROPIC_API_KEY https://api.anthropic.com/v1
DeepSeek · V4 Pro anatomy.deepseek.example.yaml DEEPSEEK_API_KEY https://api.deepseek.com/v1
Qwen · Qwen3.7 Max anatomy.qwen.example.yaml DASHSCOPE_API_KEY https://dashscope.aliyuncs.com/compatible-mode/v1
Moonshot · moonshot-v1-32k anatomy.moonshot.example.yaml MOONSHOT_API_KEY https://api.moonshot.cn/v1

Recommended first real run:

cp anatomy.openai.example.yaml anatomy.config.yaml
cp sources.github.example.yaml sources.yaml
notepad .env
$env:PYTHONPATH='src'
python -m asa plan-run --config anatomy.config.yaml --limit-skills 1
python -m asa run --config anatomy.config.yaml --limit-skills 1
python -m asa quality-run --run runs\<run-id> --output runs\<run-id>\quality_report.json
python -m asa export-letuen --run runs\<run-id> --output dist\letuen-real

Use --limit-skills 1 until the provider route, cost, schema validation, and artifact quality look stable.

Harness Flow

GitHub URL / Local Path
  -> Collector
  -> Structure Analyst
  -> Workflow Analyst
  -> Pattern Miner
  -> Asset Builder
  -> Reviewer
  -> Quality Gates
  -> Report / Vault / Data / Anchors

Local Model Bridge

The static UI can call a local-only bridge for connection tests and homepage GitHub URL analysis.

$env:PYTHONPATH='src'
python scripts/local_bridge.py

Then open http://localhost:4173/settings/models/. The bridge does not persist keys. For CLI runs, store keys in .env using the mapped environment variable.

Skill Anchor Pack

The developer-preview method pack is available in the GitHub Release and under releases/:

  • releases/letuen-skill-anchor-pack-v0.2.0-dev.zip
  • releases/letuen-skill-anchor-pack-v0.2.0-dev.tar.gz
  • releases/SHA256SUMS.txt

Rebuild it locally:

$env:PYTHONPATH='src'
python -m asa package-letuen-skill --output releases --version v0.2.0-dev

The pack contains a coordinator SKILL.md, 9 skills/asa-* method skills, anchor schemas, non-destructive invocation policy, and sample anchor composition assets.

CLI Cheat Sheet

$env:PYTHONPATH='src'
python -m asa collect --source examples/sample-skill --output runs/manual-collect/inventory.json
python -m asa run --config anatomy.config.example.yaml
python -m asa validate --run runs/<run-id>
python -m asa review-run --run runs/<run-id> --output runs/<run-id>/review_summary.json
python -m asa repair-evidence --run runs/<run-id>
python -m asa quality-run --run runs/<run-id> --output runs/<run-id>/quality_report.json
python -m asa export-report --run runs/<run-id> --output site/report
python -m asa export-vault --run runs/<run-id> --output vault
python -m asa export-data --run runs/<run-id> --output site/data
python -m asa export-anchors --run runs/<run-id> --output dist/anchors.json
python -m asa export-letuen --run runs/<run-id> --output dist/<run-id>
python -m asa smoke-github --sources sources.smoke.github.yaml --output runs/github-smoke
python -m asa package-letuen-skill --output releases --version v0.2.0-dev

GitHub Pages Demo

The repository includes a static site under site/. After enabling GitHub Pages with GitHub Actions as the source, the Deploy GitHub Pages workflow publishes the local showcase UI.

Expected URL after Pages is enabled:

https://3yesore.github.io/LetUenforskills/

Repository Layout

src/                  Python harness, providers, exporters, quality gates
site/                 Static UI: home, cinema, repo, data/settings surfaces
skills/               LetUen internal method skills and anchor-aware workflow skills
docs/                 Design specs, harness contracts, UI/output strategy
examples/             Sample skill, sample vault, anchor composition examples
runs/demo-multi-skill Committed demo run used by export/demo scripts
releases/             Developer-preview skill anchor pack archives
tests/                Unit tests for runtime, exporters, quality, packaging

Documentation Map

  • docs/fork-ready-guide.md — clone/fork usage, CI, sample vault workflow.
  • docs/harness-spec.md — runtime state machine and quality gates.
  • docs/agent-protocol.md — agent responsibilities and boundaries.
  • docs/evidence-spec.md — evidence object contract and confidence rules.
  • docs/provider-spec.md — model provider interface, retry, and OpenAI-compatible adapter.
  • docs/letuen-usage-guide.md — end-to-end commands for report, vault, data, anchors, and composition plans.
  • docs/letuen-skill-anchor-pack-design.md — anchor-based skill decomposition and recomposition method pack.
  • docs/graph-data-surface-spec.md — data table contracts, graph nodes/edges, manifest, and UI expectations.
  • docs/ui-strategy.md — local research dashboard and public website direction.

Contributing

Use the issue templates for reproducible bugs and scoped feature requests. Pull requests should include the relevant validation commands from .github/pull_request_template.md.

Security-sensitive reports should follow SECURITY.md instead of public issues.

Release Status

Current developer preview: v0.2.0-dev.

  • CI is active on main.
  • The release asset is a method skill pack, not the full Python harness.
  • The full repository remains the source of truth for harness + UI + skill pack development.

License

MIT. See LICENSE.