Skills Bridge

Official Agent Skills (agentskills.io) and Agent Plugins (agent-plugins.org) bridge for Mvge.

Discovers skill packages across scopes and loads them with progressive disclosure, so the model sees a compact catalog — not the full text of every skill — until it actually needs one.

Spells

Spell	Description
activate_skill	Loads a skill's full content on demand (<skill_content name="…">).

Progressive disclosure

1. Tier 1 — session start injects an <available_skills> catalog (~50–100 tokens per skill).
2. Tier 2 — activate_skill(name) loads the full skill body when the model asks for it.
3. Tier 3 — <skill_resources> enumerates a skill's bundled resources without eager loading.

Anti-bloat turn updates keep the catalog deduplicated as skills activate.

Discovery & precedence

Scans PROJECT, USER, AGENT, and PLUGINS (plugin.json) scopes. Precedence is PROJECT > USER > AGENT, deterministic first-wins with shadow tracking. Parsing is tolerant (YAML colon repair, lenient validation).

Commands

• mvge skill list — show discovered skills.
• mvge skill dedupe — report shadowed/duplicated skills.

Hooks

session_start, before_mvge_start, context_transform, session_shutdown.

Dependencies

pyyaml>=6.0, typer>=0.12.