echoskill CLI: skill marketplace management tool (eskill)

[krisrowe]

Vision

A standalone CLI for managing AI agent skill marketplaces. Install, list, publish, and curate skills across Claude Code and Gemini CLI. Published to PyPI as echoskill with alias eskill.

This is NOT an extraction from echomodel/aicfg. It is a fresh design informed by best practices research across existing ecosystems.

Design inputs (must research before building)

1. Gemini CLI skills commands

• Locate the .md documentation for Gemini's built-in skill management
• Understand: install, list, search, marketplace interaction
• What works well, what's limited

2. Claude Code plugin marketplace commands

• claude plugin marketplace add/update/remove
• claude plugin install <name>@<marketplace> --scope user
• Marketplace registry model: marketplace.json, git-based, source types
• How this translates to skills (vs plugins)

3. Vercel Labs skills-npm

• https://github.com/vercel-labs/skills
• What model do they use for skill packaging and distribution?
• How do they handle dependencies, versioning, discovery?

4. What echomodel/aicfg got right and wrong

• em skills install/list/publish/marketplace exists but echomodel is currently broken
• Don't replicate — learn from it and start fresh
• Key features to preserve: marketplace registration, cross-platform install (claude + gemini), install manifests with provenance tracking

Key design concept: skill curation as marketplace

A marketplace is a JSON file pointing to skills across arbitrary repos:

{
  "skills": [
    {
      "name": "sociable-unit-tests",
      "source": {
        "url": "https://github.com/echo-skill/echoskill.git",
        "path": "coding/sociable-unit-tests"
      }
    }
  ]
}

This decouples curation from source:

• Anyone can create a marketplace (just a JSON file in a repo)
• Skills can live in any repo
• Multiple marketplaces can curate overlapping sets
• Many-to-many: one skill in many marketplaces, one marketplace curating from many repos

CLI design (tentative, pending research)

eskill marketplace add <alias> <url>      # register a marketplace
eskill marketplace list                    # show registered marketplaces
eskill list                                # list available skills across marketplaces
eskill install <name>                      # install to claude + gemini
eskill install <name> --target <dir>       # install to arbitrary directory (for plugin builds)
eskill publish <name> --marketplace <alias> # publish a skill to a marketplace
eskill show <name>                         # show skill details + install status

PyPI details

• Package: echoskill
• Entry points: echoskill and eskill (same binary)
• Both names verified available on PyPI
• GitHub org: echo-skill (existing)
• Repo: echo-skill/echoskill (this repo, or a new echo-skill/eskill repo)

Process

Before building:

1. Research all four design inputs above
2. /identify-best-practices on the proposed CLI design
3. Capture decision record in CONTRIBUTING.md of the target repo
4. Then build

Dogfooding

The claude-coding-plugin build script (./build) is the primary consumer. Currently uses raw git clone — will migrate to eskill install --target skills/ once this ships. Keeping the daily-use plugin dependent on this tool ensures it stays reliable.

[krisrowe]

Additional capability: skill integrity validation

The echoskill CLI should include a lint or validate command that checks the integrity of a skill collection (plugin, marketplace, or any directory with skills):

What it validates

1. Skill reference detection — scan all .md files for lower-case-hyphen patterns that match skill names. Skill names follow agentskills.io convention: lowercase letters, numbers, hyphens, max 64 chars.

2. Registry validation — every detected skill reference exists in the marketplace config or is present in skills/

3. Source validation — every registered marketplace source URL is alive and the skill exists at the declared path. Efficient: clone unique repos with --depth 1, or use gh api to check file existence if faster.

4. Build output logging — the build/sync process produces a detailed JSONL log (no json lib dependency — manual serialization) of what was found, what was copied, what matched, SHAs of .md files before and after.

5. Orphan detection — skills in skills/ that have no marketplace source and are not declared as native

6. Dead link detection — marketplace entries pointing to repos/paths that don't exist or can't be cloned

Usage

eskill validate .                    # validate current directory
eskill validate --marketplace all    # validate all registered marketplaces
eskill validate --log build.jsonl    # produce detailed JSONL build log

Reusability

This is a general-purpose pattern — any plugin, marketplace, or skill collection benefits from it. The claude-coding-plugin consumes it in its build/test process, but it's not plugin-specific.

Related

• echoskill#25 — skill reference syntax (@skill:name inline + metadata.references frontmatter) feeds into what the validator scans for
• echoskill#22 — allowed-tools testing is adjacent (validating frontmatter correctness)