agent-plugins: Instructions file for Claude Code

AGENTS.md

This file provides guidance to AI coding agents (Claude Code, Cursor, Copilot, etc.) for creating and modifying plugins in this repository. CLAUDE.md at the repo root points here for anything plugin-authoring related — this is the canonical source.

Repository Overview

A Claude Code plugin marketplace. Plugins live under plugins/; .claude-plugin/marketplace.json at the repo root registers them all.

Plugin Directory Structure

Each plugin lives under plugins/. Folder name matches the plugin name (plugin.json → name):

plugins/<plugin-name>/
├── .claude-plugin/
│   └── plugin.json               # Required — {name, version, description}
├── CHANGELOG.md                  # Required — version history (Keep a Changelog format)
├── skills/<plugin-name>/
│   ├── SKILL.md                  # Canonical skill definition (YAML frontmatter + body)
│   ├── scripts/                  # Optional executables invoked by the skill
│   └── references/               # Optional reference docs loaded on-demand
├── agents/                       # Optional subagent definitions (plugin root, NOT under skills/)
├── commands/                     # Optional slash command definitions (plugin root)
└── evals/                        # Optional manual test prompts (gitignored)

agents/ and commands/ live at the plugin root, not inside skills/<name>/. Only skill content lives under skills/<name>/.

Conventions

• SKILL.md is the contract. YAML frontmatter must include name, description, and allowed-tools. The name must match the folder name and the skills/<name>/ directory name.
• plugin.json is minimal. Just {name, version, description} — description copied verbatim from SKILL.md frontmatter. Version follows SemVer (MAJOR.MINOR.PATCH). Every plugin must have one — check with:

  for f in plugins/*/.claude-plugin/plugin.json; do python3 -c "import json;print('$f', json.load(open('$f')).get('version','MISSING'))"; done
  

• CHANGELOG.md is required. Lives at the plugin root. Add a ## [x.y.z] — YYYY-MM-DD entry when bumping the version. Patch = fix/tweak, Minor = new capability, Major = breaking change.
• Scripts run standalone. Only their output enters Claude's context. Bash scripts use set -e, status to stderr, machine-readable JSON to stdout.
• References stay separate. Large API docs and specs go in references/ so they aren't loaded until the skill needs them.
• Agent definitions default to background. Any subagent shipped under agents/ must include "ALWAYS launch this agent with run_in_background: true" in its description, so callers don't block the main conversation.
• No shared build step. Plugins are distributed as-is.
• No automated tests, as a rule. Evals in evals/ are manual (invoke + verify) and gitignored. Exception: side-quest ships a real pytest suite (tests/) for its ledger/sync logic — that logic has enough state-machine complexity (event dedup, XP tiers, MQTT sync, coalescing) to warrant it. Follow the no-tests default unless a plugin's logic is similarly nontrivial and stateful, not just a thin CLI wrapper.
• $CLAUDE_PLUGIN_ROOT resolves to the plugin root (the directory containing .claude-plugin/), not the skill directory. A script at plugins/<name>/skills/<name>/scripts/foo.sh must be referenced as ${CLAUDE_PLUGIN_ROOT}/skills/<name>/scripts/foo.sh — the skills/<name>/ segment is easy to drop by accident (font-extractor and wp-local both shipped with this bug: $CLAUDE_PLUGIN_ROOT/scripts/foo pointing at a path that doesn't exist). Verify with a quick sweep before shipping:

  grep -rn "CLAUDE_PLUGIN_ROOT" plugins/*/skills/*/SKILL.md plugins/*/agents/*.md plugins/*/commands/*.md
  

  and check each result's path actually exists under the plugin directory. Exception: plugins whose scripts must also run outside any live Claude Code session — as a settings.json hook command, or a background daemon (launchd/systemd) — can't rely on $CLAUDE_PLUGIN_ROOT (it's only set inside a running session). Those plugins (side-quest, self-improvement) hoist their scripts/ to the plugin root instead of nesting under skills/<name>/, and ship a setup step (see below) that deploys copies to a stable $HOME path.
• Plugins that install anything outside the plugin cache need a setup command. If a plugin merges hooks into ~/.claude/settings.json, installs a daemon, or deploys scripts to a stable path so they work from a hook/daemon context (see above), it needs: (1) a setup agent that does the actual idempotent installation work (backgrounded per the rule above), and (2) a /plugin-name:setup slash command in commands/ that wraps it, so users don't have to know to invoke the agent by name. self-improvement (/self-improvement:setup) and side-quest (/side-quest:setup) are the reference implementations. A plugin whose scripts only ever run inside a live Bash tool call (the common case) does not need this — $CLAUDE_PLUGIN_ROOT already resolves correctly at runtime, no deploy step required.