Skip to content

PRD.md

Latest commit

 

History

History
135 lines (89 loc) · 4 KB

PRD.md

File metadata and controls

135 lines (89 loc) · 4 KB
Raw

Product Requirements Document: Skill System

Overview

OmniAgent needs to support loading and executing skills compatible with OpenClaw/ClawHub to enable extensible AI agent capabilities without skill fragmentation across the ecosystem.

Problem Statement

Currently, OmniAgent has:

  • A static tool registration system
  • No dynamic skill/plugin loading
  • No way to leverage the growing ClawHub skill ecosystem

Users want to:

  • Install skills from ClawHub (bunx clawhub install sonoscli)
  • Use the same skills across OpenClaw, NanoClaw, IronClaw, and OmniAgent
  • Avoid maintaining separate skill versions for each platform

Goals

  1. Full ClawHub Compatibility: Load and use ClawHub skills without modification
  2. Security: Execute skill code in sandboxed environments (Deno for TypeScript/shell, WASM for tools)
  3. Minimal Overhead: Markdown-only skills should load with zero runtime cost
  4. Incremental Adoption: Users can start with simple skills and adopt hooks/scripts later

Non-Goals

  • Creating a competing skill format
  • Building a skill marketplace (use ClawHub)
  • Docker-based sandboxing (use Deno + WASM instead)

User Stories

US-1: Load Markdown-Only Skills

As a user, I want to install a ClawHub skill and have OmniAgent automatically include its instructions in my agent's context.

Acceptance Criteria:

  • Skills are discovered from ~/.omniagent/skills/ and ./skills/
  • SKILL.md files are parsed (YAML frontmatter + Markdown body)
  • Skill content is injected into the LLM system prompt
  • Requirements (bins, env vars) are checked before loading

US-2: Use CLI-Based Skills

As a user, I want to use skills like sonoscli or github that teach my agent how to use CLI tools.

Acceptance Criteria:

  • Skill requirements are validated (e.g., gh binary exists)
  • Missing requirements show clear error messages with install instructions
  • Agent can execute CLI commands via the shell tool

US-3: Run Skill Hooks (Phase 2)

As a user, I want skills with hooks (like self-improving-agent) to execute their TypeScript/shell hooks.

Acceptance Criteria:

  • TypeScript hooks run via Deno with restricted permissions
  • Shell hooks run via Deno with --allow-run restrictions
  • OpenClaw hook API is compatible via import map

US-4: Secure Tool Execution (Phase 3)

As a user, I want tool execution to be sandboxed for security.

Acceptance Criteria:

  • Built-in tools run in WASM sandbox (wazero)
  • Capability-based permissions for file/network access
  • Memory and CPU limits enforced

Feature Phases

Phase 1: Skill Loader (Markdown-Only)

  • Parse SKILL.md format
  • Discover skills from configured directories
  • Check requirements (bins, env)
  • Inject into LLM system prompt
  • CLI command: omniagent skills list, omniagent skills info <name>

Phase 2: Hook Runner (Deno)

  • Run TypeScript hooks via Deno
  • Run shell hooks via Deno
  • OpenClaw hook API compatibility layer
  • Permission restrictions per skill

Phase 3: Tool Sandbox (WASM)

  • WASM runtime via wazero
  • Capability-based permissions
  • Built-in tools (file, http) in WASM
  • Memory/CPU limits

Test Skills

Phase 1 Test Cases

Skill Type Requirements Source
sonoscli CLI sonos binary OpenClaw bundled
github CLI gh binary OpenClaw bundled
weather CLI weather binary OpenClaw bundled
slack CLI None (API-based) OpenClaw bundled

Phase 2 Test Cases

Skill Type Requirements Source
self-improving-agent Hooks + Scripts None ClawHub

Success Metrics

  • 100% of ClawHub Markdown-only skills load without modification
  • <100ms skill loading time for 50 skills
  • Zero security incidents from sandboxed execution

Dependencies

  • Deno runtime (for Phase 2)
  • wazero library (for Phase 3)

Timeline

Phase Description Target
1 Skill Loader v0.3.0
2 Hook Runner v0.4.0
3 Tool Sandbox v0.5.0