| name | ecosystem-authoritative-sources |
| description | Provides information about how to create, structure, install, and audit Agent Skills, Plugins, Antigravity Workflows, and Sub-agents. Trigger this when specifications, rules, or best practices for the ecosystem are required. |
| allowed-tools | Bash, Read, Write |
Dependencies
This skill requires Python 3.8+ and standard library only. No external packages needed.
To install this skill's dependencies:
pip-compile ./requirements.in
pip install -r ./requirements.txt
See ../../requirements.txt for the dependency lockfile (currently empty — standard library only).
Ecosystem Authoritative Sources
Official Open Standard Recognition
Important: This reference library draws heavy inspiration and structural standards directly from the Anthropic Claude Plugins official repositories. Please refer to:
- Foundational Specification:
https://github.com/anthropics/claude-plugins-official/tree/main/plugins/plugin-dev
- L4 Interaction & Execution Patterns: Derived from
https://github.com/anthropics/claude-knowledgework-plugins (specifically the Legal and Bio-Research plugins).
- Skill Creator 2.0 (Anthropic, March 2026):
https://github.com/anthropics/skills/tree/main/skills/skill-creator
- Anthropic Official Claude Code docs:
https://docs.anthropic.com/en/docs/claude-code/
Agentic OS / Agent Harness Pattern
The Agentic OS is a synthesized runtime environment pattern built from Anthropic primitives:
The context/ folder structure, heartbeat.md, START_HERE.md, and memory log patterns
are community conventions built on top of these primitives (not Anthropic-official).
For bootstrapping Agentic OS in a project, use the agentic-os-init skill:
.agents/skills/agentic-os-init/SKILL.md
The Library
The following open standards are available for review:
This skill provides comprehensive information and reference guides about the conventions and constraints defining the extensibility ecosystem.
Because of the Progressive Disclosure architecture, you should selectively read the reference files below only when you need detailed information on that specific topic.
Installation
Claude Plugin Marketplace (Claude Code native — verified 2.1.81+)
Repos with a .claude-plugin/marketplace.json at the root can be registered as a marketplace:
/plugin marketplace add owner/repo
/plugin install <plugin-name>
Claude Code fetches from the default branch. The marketplace.json must be merged to main before consumers can install.
/plugin install <name> opens an interactive Plugins panel (not plain stdout) showing plugin details and a scope picker (user / project / local). The command returns no terminal output — UI renders in the panel.
Source types in marketplace.json:
- Relative path (same-repo monorepo):
"source": "./plugins/my-plugin" — resolved from repo root
- Git subdirectory (sparse clone):
"source": { "type": "git-subdir", "url": "...", "subdir": "..." }
- npm package:
"source": { "type": "npm", "package": "@scope/pkg", "version": "^1.0.0" }
- Non-GitHub git:
"source": { "type": "url", "url": "...", "ref": "main" }
strict field: true = plugin's own plugin.json wins; false = marketplace entry is the full definition (no plugin.json needed).
Validation before publishing: /plugin validate . or claude plugin validate .
Team/enterprise distribution (.claude/settings.json):
extraKnownMarketplaces — auto-register marketplace for the whole team
enabledPlugins — default-enable specific plugins for all team members
strictKnownMarketplaces: true — lock to approved marketplaces only (enterprise)
CLAUDE_CODE_PLUGIN_SEED_DIR env var — pre-populate plugins in containers/CI
Note: /plugin marketplace browse returns no content in 2.1.81 — use /plugin help to discover available subcommands.
npx skills CLI (cross-agent)
Universal CLI to install across all supported agents (Claude Code, Copilot, Gemini CLI, etc).
Quick Reference:
For all installation, update, and management commands, strictly refer to the authoritative project hub:
This hub contains the definitive logic for uvx, bootstrap.py, and npx skills methods.
Table of Contents
To read any of the reference guides, use your file system tools to cat or view the relevant file.
- Agent Skills: Definition, lifecycle, progressive disclosure, and constraints of
.claude/skills/ (and equivalents like .agents/skills/ and .github/skills/). Custom agents deployed as Skills are stored here as <plugin>-<agent>/SKILL.md.
- Claude Plugins: Specification for the
.claude-plugin architecture, manifest setup, and distribution.
- Antigravity Workflows & Rules (and Legacy Commands): Specifications for global/workspace Rules, deterministic trajectory Workflows, and the critical distinction between deploying Skills vs. Legacy Commands.
- Sub-Agents: Definition, setup, and orchestration of nested contextual LLM boundaries. Sub-Agents are deployed structurally as pure Skills (mapped to
skills/<agent-name>/SKILL.md).
- GitHub Copilot Prompts (Models): Documentation on the exact YAML schema, dynamic variables, and exclusion logic (
exclude-targets) used by GitHub Copilot chat environments.
- GitHub Agentic Workflows: Documentation on the "Continuous AI" autonomous agents responding to CI/CD events.
Anthropic Plugin-Dev Key Learnings (March 2026)
The following patterns were confirmed from Anthropic's official plugin-dev plugin:
${CLAUDE_PLUGIN_ROOT}: Absolute path to the plugin's installation directory (read-only during updates).
${CLAUDE_PLUGIN_DATA}: Persistent directory for plugin state (e.g., node_modules, venv) that survives updates.
- Prompt-based hooks: LLM-driven hooks run in parallel; use for semantic decisions. Command-based hooks for deterministic validation.
model: inherit: Default for sub-agents. Explicitly setting a model pins the agent to a version.
<example> blocks: Use 2-4 <example> blocks in agent frontmatter descriptions for richer trigger context (including proactive and negative instructions).
.local.md settings: Store per-project agent config in .claude/<plugin-name>.local.md (always gitignore).
examples/ directory: Add to skills alongside references/ for working code samples.
- SKILL.md target: 1500-2000 words / under 500 lines. Third-person description, imperative body.
- Description checklist: 3-5 triggers, 100-word limit, precise action verbs, 1 negative boundary.
Usage Instruction
Never guess the specifics of SKILL.md frontmatter, plugin directory limits, or workflow sizes. Read the exact specifications linked above before constructing new ecosystem extensions.