Vox Skill Marketplace

The Vox skill marketplace (vox-skills crate) provides a plugin system

What is a Skill?

A skill is a self-contained bundle containing:

A SKILL.md manifest (TOML frontmatter + markdown body)
Optional code or instructions
Declared dependencies and permissions

SKILL.md Format

---
name = "web-search"
version = "1.0.0"
description = "Adds the ability to search the web"
author = "vox-team"
tags = ["search", "web"]
permissions = ["network"]
---

## Instructions

Use this skill to perform web searches...

MCP Tools

Tool	Description
`vox_skill_install`	Install a skill from a VoxSkillBundle JSON payload
`vox_skill_uninstall`	Uninstall an installed skill by ID
`vox_skill_list`	List all installed skills
`vox_skill_search`	Search installed skills by keyword
`vox_skill_info`	Get detailed info on a specific skill by ID
`vox_skill_parse`	Preview a SKILL.md manifest before installing

Built-in Skills

The following skills ship pre-installed in vox-skills/skills/:

File	Purpose
`compiler.SKILL.md`	Vox compiler integration
`testing.SKILL.md`	Test runner integration
`docs.SKILL.md`	Documentation generation
`deploy.SKILL.md`	Deployment automation
`refactor.SKILL.md`	Code refactoring helper

Plugin System

Skills are backed by the Plugin trait and managed by PluginManager:

trait Plugin: Send + Sync {
    fn id(&self) -> &str;
    fn on_event(&self, event: &HookEvent) -> Result<(), PluginError>;
}

Hook System

Skills can register lifecycle hooks via HookRegistry:

registry.register(HookEvent::TaskCompleted, |event| {
    // react to task completion
});

Available events: TaskCompleted, TaskFailed, AgentStarted, AgentStopped, MemoryFlushed.

Skills vs. Deep Execution Libraries (PyTorch)

Skills are designed exclusively for loose-coupled orchestration (e.g., connecting to a database, resolving a Git conflict, routing workflows) and are not meant to replace high-speed, tight-loop native ML execution like PyTorch. Because Skills function as discrete event-driven bundles with high-latency JSON/IPC handoffs, they should never be used for per-pixel or per-node calculations. For heavy loop execution or math operations, rely on native, highly coupled Rust constructs (such as vox-tensor), avoiding external library spaghetti.