Skills and Plugins

Skills provide reusable instructions through SKILL.md. Claude Code-format plugins can bundle skills with commands, hooks, MCP servers, and subagents.

Skill discovery

Anvil discovers skills from:

1. $CODEX_HOME/skills (or ~/.codex/skills)
2. ~/.claude/skills and ~/.agents/skills
3. Project .claude/skills and .agents/skills directories
4. Enabled plugins

Project and user entries override plugin-provided entries. Built-in slash commands keep precedence if names collide. A discovered skill is exposed as a slash command and may also be activated by the model when its description matches the task.

A minimal skill looks like:

my-skill/
├── SKILL.md
└── scripts/

---
name: my-skill
description: Explain when this workflow should be used.
---

Follow these steps...

Instructions and scripts execute with the same workspace, permission, and sandbox boundaries as the parent session. Review a skill before installing or enabling it.

Plugin management

Anvil discovers Claude Code installations and accepts Claude Code-format manifests at .claude-plugin/plugin.json.

/plugin list
/plugin add <git-url | owner/repo | local-path> [marketplace-plugin]
/plugin update <name>
/plugin enable <name>
/plugin disable <name>
/plugin remove <name>

Plugins installed with /plugin add live in Anvil’s configuration area. A plugin may add:

• skills and slash commands;
• subagent definitions;
• lifecycle hooks;
• MCP server processes.

That makes a plugin a code-and-process trust decision, not merely a prompt library. Inspect its manifest, hooks, executable commands, and MCP configuration before enabling it. See Data and Trust Boundaries.

Hook execution

Plugins can register UserPromptSubmit, PreToolUse, and PostToolUse command hooks. Anvil runs them as sh -c on Unix or cmd /C on Windows, in the session working directory, with inherited host environment and operating-system authority. They are not routed through the model tool permission gate or the shell sandbox.

Exit code 0 accepts the hook and can contribute bounded stdout context. Exit code 2 blocks a prompt/tool or feeds post-tool feedback using stderr. Other exits and timeouts are logged as hook errors. Enabling a plugin therefore authorizes its hooks to execute at lifecycle events even when the model has not requested a shell tool.