Concepts

Artifact kinds

Dex Trove has seven kinds. Each keeps its upstream format, maps to an Ananke kind, and installs into the native layout of each tool that supports it.

KindUpstream shapeAnanke kindClaude CodeGitHub CopilotTrove count
skillfolder with SKILL.mdskillskills/<name>/.github/skills/<name>/468
agent*.agent.md or frontmatter .mdagentagents/<name>.md.github/agents/<name>.agent.md222
instruction*.instructions.md with applyToprompt—.github/instructions/192
prompt*.prompt.mdpromptcommands/<name>.md.github/prompts/0
hookfolder with README.md + hooks.jsontool—.github/hooks/<name>/8
workflowagentic workflow .mdworkflow—.github/workflows/<name>.md8
pluginplugin.json listing membersbundlemembersmembers72

”—” means the tool has no native equivalent; use --target dir. Counts are for the 0.1.0 trove.

Skill

A skill is a folder of instructions plus bundled resources (scripts, references, templates, assets) that an agent loads on demand when the task matches its description. It follows the Agent Skills specification:

my-skill/
├── SKILL.md          frontmatter: name, description (+ license, allowed-tools, compatibility, metadata…)
├── references/       docs the agent reads only when needed
├── scripts/          deterministic helpers the agent runs
└── assets/           templates, examples, images

The description is what makes a skill trigger: the harness shows the agent only names and descriptions, and the agent opens SKILL.md when one matches. Good descriptions say what the skill does and when to use it.

  • Claude Code: dex install <ref> --target claude → ~/.claude/skills/<name>/ (user) or ./.claude/skills/<name>/ (--scope project). Claude Code discovers skills there automatically.
  • Copilot: --target copilot → .github/skills/<name>/.
  • Ananke: kind skill; activate with ananke apm install <ns>/<name> after learning.

Skills that contain scripts are tagged executes-code and show runs code on their page. Read the scripts before you let an agent run them.

Agent

An agent (custom agent, subagent or “chat mode”) is a persona with its own system instructions and, optionally, a tool allow-list, a preferred model and MCP servers. Upstream frontmatter commonly includes:

KeyMeaning
name, descriptionidentity and when to use it
toolstools the agent may use (e.g. codebase, editFiles, runCommands)
modelpreferred model
mcp-serversMCP servers the agent expects
argument-hint, user-invocable, handoffsinvocation details

The item page shows all of these under Upstream metadata.

  • Claude Code: agents/<name>.md (a subagent). Tool names from Copilot agents may not match Claude Code’s tool names. Review the frontmatter after installing.
  • Copilot: .github/agents/<name>.agent.md.
  • Ananke: kind agent; the manifest’s instructions.ref points at the entry file.

Instruction

An instruction file holds coding standards that the harness applies automatically to files matching its applyTo glob (e.g. '**/*.py'). Item pages show the glob.

  • Copilot: .github/instructions/<name>.instructions.md.
  • Claude Code: no native equivalent. Install with --target dir and reference the file from CLAUDE.md, or turn it into a skill.
  • Ananke: kind prompt, tag instruction.

Prompt

A reusable prompt file (*.prompt.md), invoked explicitly by the user. The current sources ship none, but the awesome-copilot adapter supports them.

  • Claude Code: commands/<name>.md (a slash command).
  • Copilot: .github/prompts/<name>.prompt.md.

Hook

A hook runs a command automatically when an agent session event fires (sessionStart, sessionEnd, preToolUse, …). It is a folder with a README.md (frontmatter name, description, tags), a hooks.json and usually a script:

{
  "version": 1,
  "hooks": {
    "sessionEnd": [
      { "type": "command", "bash": ".github/hooks/secrets-scanner/scan-secrets.sh", "timeoutSec": 30 }
    ]
  }
}

Item pages list each event with its command and timeout.

Hooks run code without asking, so Dex never installs one unless you pass --allow-exec. Read the script first.

  • Copilot: .github/hooks/<name>/ (with --allow-exec).
  • Ananke: kind tool, tags hook and executes-code.

Workflow

An agentic workflow (GitHub Agentic Workflows) is a Markdown file whose frontmatter declares triggers (on), permissions, safe-outputs, tools and engine, and whose body is natural-language instructions. Item pages show the triggers and permissions.

  • Copilot / GitHub: .github/workflows/<name>.md, then compile with the gh aw CLI.
  • Ananke: kind workflow.

Plugin

A plugin groups related skills and agents around a theme (e.g. arize-ax bundles 9 Arize skills). Upstream it is a plugin.json (Agent Plugins) whose extensions.com.github.awesome-copilot lists skills and agents.

In Dex a plugin keeps its plugin.json and README.md, and its members are resolved to trove items:

  • dex install <plugin> installs every member into the target.
  • In Ananke the plugin becomes a bundle whose members are ananke://<kind>/<ns>/<name> references. ananke registry resolve <ns>/<plugin> --explain shows the member graph.
  • Members excluded for licence or other reasons are dropped. A plugin with no admissible members, such as awesome-copilot’s canvas-only plugins, is excluded as out-of-scope.

Edit this page on GitHub ↗