colibri-ai-agents

Colibri AI Agents

Centralized agents and coding standards for the Colibri team, usable as both a GitHub Copilot plugin and a Claude Code plugin from this single repository.

What's shared vs tool-specific

Component	Path	Copilot	Claude Code
Coding standards (skills)	skills/*/SKILL.md	✅ same files	✅ same files
Agents (source of truth)	agents/*.md	generated → copilot/agents/	✅ read directly
Manifest	plugin.json / .claude-plugin/plugin.json	plugin.json	.claude-plugin/plugin.json

The skills are identical for both tools — one set of files, no duplication. The agents have one source of truth (agents/*.md, Claude format); the Copilot versions in copilot/agents/ are generated, never edited by hand.

Install

Claude Code

This repo doubles as its own Claude Code marketplace (.claude-plugin/marketplace.json).

/plugin marketplace add <git-url-of-this-repo>
/plugin install colibri-ai-agents@colibri
/reload-plugins

Azure DevOps note: the owner/repo shorthand only works for GitHub. Since this repo is hosted on Azure DevOps, pass the full git URL to /plugin marketplace add.

Or test locally without installing:

claude --plugin-dir /path/to/ColibriAIAgents

To auto-prompt teammates to install when they trust the project folder, add the marketplace to a shared .claude/settings.json:

{
  "extraKnownMarketplaces": {
    "colibri": { "source": { "source": "git", "url": "<git-url-of-this-repo>" } }
  },
  "enabledPlugins": { "colibri-ai-agents@colibri": true }
}

Agents appear in /agents; skills load automatically by description.

GitHub Copilot (VS Code)

1. Ctrl+Shift+P → Chat: Install Plugin From Source
2. Paste the repository URL.
3. Agents appear in the Chat view agent picker; skills load automatically.

Repository structure

├── plugin.json                  # Copilot manifest (agents → copilot/agents/, skills → skills/)
├── .claude-plugin/
│   ├── plugin.json              # Claude Code manifest
│   └── marketplace.json         # Claude Code marketplace catalog (install source)
├── skills/                      # SHARED coding standards (both tools read these)
│   ├── csharp/SKILL.md
│   ├── dotnet-framework/SKILL.md
│   ├── frontend/SKILL.md
│   └── sql/SKILL.md
├── agents/                      # SOURCE OF TRUTH — Claude subagents (.md)
│   └── *.md
├── copilot/agents/             # GENERATED Copilot agents (.agent.md) — do not edit
│   └── *.agent.md
├── scripts/
│   ├── build-copilot-agents.mjs # regenerates copilot/agents/ from agents/
│   └── agents.meta.json         # display names used for the Copilot picker
└── Documentations/              # verbose human-facing standards (skill sources)

Maintenance — the only workflow you need

1. Edit a skill → change skills/<topic>/SKILL.md. Done; both tools pick it up.
2. Edit / add an agent → change agents/<name>.md (Claude format), then run:

   node scripts/build-copilot-agents.mjs
   

   This regenerates the Copilot equivalents.
3. Bump version in both plugin.json and .claude-plugin/plugin.json.
4. Commit & push.

Adding a brand-new agent

• Create agents/<slug>.md with frontmatter name, description, tools (Claude tool names: Read, Grep, Glob, Edit, Write, Bash).
• Add a display name in scripts/agents.meta.json: {"<slug>": {"display": "Nice Name"}}.
• Run the build script.

CI guard (recommended)

A GitHub Action (.github/workflows/sync-check.yml) runs node scripts/build-copilot-agents.mjs --check and fails the PR if anyone edited a Copilot agent by hand or forgot to regenerate. This prevents drift between the two tools.

Notes

• Agent bodies are shared verbatim and tool-agnostic: they reference coding standards by skill name (e.g. "the frontend standards"), which both Copilot and Claude auto-load by description — no hardcoded file paths. References to Colibri.Web/... are intentional (they point at the consuming project, not at this plugin).
• If you later add a 3rd tool (Cursor, Codex…), consider rulesync to generate every format from one source. For just Copilot + Claude, this script is enough.