Skip to content

AGENTS.md

Latest commit

 

History

History
92 lines (72 loc) · 4.72 KB

AGENTS.md

File metadata and controls

92 lines (72 loc) · 4.72 KB
Raw

AGENTS.md

Working context and rules for any AI agent (or human) contributing to this repository: the agent-toolkit repo itself, not the components it ships. CLAUDE.md imports this file and adds Claude-Code-specific notes. Keep it lean: it loads into an agent's context every session.

What this repo is

agent-toolkit is DiUS's shared collection of skills, commands, agents and hooks for AI coding agents. It ships two ways: as portable components installable with npx skills, and as a single Claude Code plugin bundling skills, commands and agents.

For the catalogue of what's currently in the toolkit, see README.md. For how any one component works, read that component's own entry point (a skill's SKILL.md, a command or agent's frontmatter). This file is about contributing to the repo.

Repo map

skills/<name>/                 Self-contained skills; everything a skill needs lives inside it.
commands/                      Slash commands — single Markdown files with frontmatter.
agents/                        Claude Code subagents a skill can delegate work to.
hooks/                         Optional, opt-in Claude Code hooks — not wired into the plugin.
.claude-plugin/                Claude plugin + marketplace manifests.
docs/getting-started.md        Per-tool install notes.
scripts/validate.js            Verification gate (run before committing).
.github/workflows/validate.yml CI that runs the gate on push/PR.

Principles

  • Host-agnostic first. Components should run on any capable agent. Do not add hard dependencies on hooks, MCP servers, or a specific runtime. Optional integrations must degrade gracefully: used when present, skipped cleanly when absent. Claude-specific wiring stays confined to .claude-plugin/, commands/, agents/ and hooks/.
  • Skills are self-contained. Everything a skill's entry point references (playbooks, references, templates) lives inside its own skills/<name>/ directory, which is the unit npx skills and the plugin install. Never point a skill file at something outside its own directory.
  • Lean docs. Docs (including this file) earn their length by making a contributor productive. Prefer prose and small tables; push depth into the specific file, not the index.

How to add each component type

  • Skill → new skills/<name>/SKILL.md (or equivalent entry point) with name + description frontmatter, self-contained. The ./skills path in .claude-plugin/plugin.json already covers new subdirectories; add a row to the skills table in README.md.
  • Command → new commands/<name>.md with name + description frontmatter. It ships automatically via the commands path in .claude-plugin/plugin.json. See commands/README.md.
  • Agent → new agents/<name>.md with name + description frontmatter and a tools list scoped to the minimum it needs (read-only where the agent only inspects code). Add it to the agents array in .claude-plugin/plugin.json to ship it in the plugin.
  • Hook → new hooks/<name>/ with a hooks.json and its script(s), following hooks/README.md. Do not wire it into .claude-plugin/plugin.json; hooks stay opt-in, and consumers copy them into their own project.

See CONTRIBUTING.md for the full mechanics.

Verification gate (required)

Before committing, run:

node scripts/validate.js

It checks that the manifests are valid JSON with required keys, that every SKILL.md, agents/*.md and commands/*.md (except its README) has name + description frontmatter, that every path referenced by plugin.json exists, that relative markdown links resolve and a skill's links stay inside that skill, and that any hooks/**/hooks.json is valid JSON. CI runs the same script on every push and PR, and a red gate blocks merge.

validate.js itself stays component-agnostic: it validates the repo's format, never one component's content. A component with its own invariants (a rule worded identically in several files, a closed vocabulary) puts them in scripts/checks/<component>.js, which the gate loads automatically. See scripts/checks/README.md.

Do not

  • Add hooks/MCP as a hard requirement, or otherwise break host-agnostic behaviour.
  • Reference files outside a skill's own directory from within that skill.
  • Wire a hook into .claude-plugin/plugin.json; hooks are opt-in only.
  • Bloat the repo docs or this file.
  • Commit with a failing scripts/validate.js.

Pointers