Fusion Skills

Skill lifecycle entrypoint. Resolve intent, engage the right agent.

Loading behavior

Load ONLY the routed agent file. References are loaded on-demand by the agent when it needs them (e.g., skill-catalog.md only on MCP failure, sync-workflows.md only in setup mode). Do not preload all agents or all references.

MCP tools

This skill uses mcp_fusion_skills — the Fusion MCP tool that handles both skill discovery and lifecycle operations:

• Discovery (intent: query): semantic search over the skills index by description or task.
• Inventory (intent: inventory): list installed skills.
• Lifecycle (intent: install | update | remove): advisory commands for managing skills.

Routing

If the user was routed here from the fusion main gate with a partially-resolved intent (e.g., "skills"), proceed directly to the most likely agent without re-asking the top-level question.

If intent is genuinely unclear and the user was not already asked a clarifying question, ask one question: "Are you looking to find a skill, install or update one, create your own, or report a problem with one?"

Compound intents: If the request spans multiple agents (e.g., "inspect this skill and fix the issues"), run agents sequentially. Pass the output of the first agent as context to the second.

Loop prevention: If you were activated from the fusion main gate, do not redirect back to fusion. Handle the request within your agents or state that it is out of scope.

Out of scope: Resolving GitHub issues, reviewing PRs, planning task breakdowns, or authoring GitHub issues. Direct the user to the relevant installed skill or the fusion main gate — but never re-route back to fusion if you were activated from there.

Migration from deprecated skills

If fusion-discover-skills or fusion-skill-self-report-bug is installed alongside this skill, tell the user:

"You have a deprecated skill installed that's now part of fusion-skills. You can safely remove it: npx -y skills remove <deprecated-skill-name>"

Safety

• No secrets or credentials.
• No GitHub mutations without confirmation.
• No remote script execution.
• No invented skill names or catalog results.
• Installed skill files are copies (see skills-lock.json for source) — never edit them in place directly; route content fixes through agents/author.agent.md (which redirects to fusion-skill-authoring, source repo) and failures through agents/warden.agent.md (report mode).