architecture/universal-skill-forge-v2.md

# Luuna Skill Forge v2 — Universal Skill Layer

## Goal

Luuna Skill Forge must work as a universal capability layer across providers, agent runtimes, skill registries, MCP marketplaces and local automation systems.

It should not be tied to one model, one vendor, one prompt format or one tool runtime.

## Core idea

Luuna does not store only prompts. Luuna stores capabilities.

A capability can be exported as:

- AgentSkills folder with SKILL.md
- OpenClaw skill
- Hermes skill and toolset profile
- Claude project or cookbook-style instruction
- OpenAI tool/function schema
- Gemini instruction pack
- LM Studio local system prompt
- Ollama Modelfile prompt block
- OpenRouter routing profile
- Hugging Face repo or Space package
- MCP server install profile
- n8n, CrewAI or LangGraph workflow
- Tauri/FastAPI/Node/Python local module

## 10-step future design

### 1. Universal Capability Object
Every ability becomes one normalized object:

- id
- name
- purpose
- trigger phrases
- input schema
- output schema
- permissions
- dependencies
- target runtimes
- provider adapters
- install commands
- tests
- trust level
- source registry

### 2. Registry Radar
Skill Forge scans public and private capability sources:

- AgentSkills
- OpenClaw / ClawHub
- Hermes Agent tools and skills
- MCP marketplaces
- NVIDIA Skills
- Claude cookbook and agent patterns
- ModelScope skills
- MiniMax Agent / MCP
- GitHub issues, repos and discussions
- Hugging Face models, datasets and Spaces
- local LuunaOS skill folders

### 3. Capability Classifier
Every found tool or skill is classified into one of these groups:

- browser
- files
- terminal
- code
- media
- voice
- memory
- search
- business
- finance
- automation
- messaging
- calendar
- local device control
- developer tools
- AI model routing
- MCP connector

### 4. Compatibility Mapper
For every capability, Skill Forge creates compatibility output:

- works in OpenClaw
- works in Hermes
- works as MCP
- works as AgentSkills folder
- works in Claude
- works in OpenAI tools
- works in LM Studio / Ollama prompt-only mode
- requires API key
- requires binary
- requires Docker
- requires browser session
- requires approval before execution

### 5. Adapter Generator
Skill Forge generates target-specific adapters:

- OpenClaw: SKILL.md with YAML frontmatter and metadata.openclaw
- Hermes: toolset profile plus skill instructions
- AgentSkills: portable SKILL.md folder
- MCP: server.json, install command and client config
- OpenAI: tool schema plus developer prompt
- Claude: project instructions and XML-like sections
- Gemini: compact system instruction and JSON schema
- Ollama: Modelfile SYSTEM block
- LM Studio: OpenAI-compatible system prompt

### 6. Trust and Security Layer
Every discovered skill gets a trust score:

- official vendor source
- verified repository
- signed release
- readable source
- recent maintenance
- dependency risk
- permission scope
- secret handling
- local filesystem access
- network access
- command execution access

Unknown third-party skills are never auto-enabled. They go to review queue first.

### 7. Skill Workshop Queue
New skills are not written directly into active runtime.
They are drafted into a workshop queue:

- proposed
- inspected
- tested
- approved
- installed
- monitored
- versioned

### 8. Local LuunaOS Sync
Approved capabilities sync into:

- C:/LuunaOS/skills
- C:/LuunaOS/providers
- C:/LuunaOS/mcp
- C:/LuunaOS/workflows
- C:/Luuna/memory/skills
- Obsidian skill notes

### 9. Runtime Router
When the user asks for work, Luuna chooses:

- use existing skill
- install missing skill
- create new local skill
- call MCP tool
- route to provider model
- call workflow engine
- ask for approval before risky action

### 10. Self-Improving Skill Graph
Every failure becomes training data:

- missing tool
- weak prompt
- bad adapter
- failed dependency
- wrong provider
- blocked permission
- slow runtime
- bad output format

Skill Forge turns this into an improved capability package.

## Main architecture

User goal
  -> Capability Detector
  -> Registry Radar
  -> Capability Classifier
  -> Trust Scanner
  -> Adapter Generator
  -> Test Runner
  -> Workshop Queue
  -> Runtime Installer
  -> Luuna Memory
  -> Runtime Router

## Rule

Luuna Skill Forge must never be only a prompt generator.
It must be a capability compiler.