Installing Skills and Plugins into Overdeck's Distribution

This skill is for Overdeck developers only. It covers taking an external
artifact — a skill directory, a GitHub repo, or a Claude Code plugin — and
wiring it into the overdeck repo so pan sync distributes it to every
Overdeck user (or to Overdeck dev machines only).

Step 1 — Identify what you're installing

Fetch or inspect the source first. The artifact type determines the mechanism:

Artifact	How to recognize it	Distribution mechanism
Plain skill	A directory with SKILL.md (plus optional scripts/references)	Copy into sync-sources/skills/<name>/
Dev-only skill	Same shape, but only useful on an overdeck source checkout	Copy into sync-sources/dev-skills/<name>/
Claude Code plugin	A repo with .claude-plugin/marketplace.json, installed via /plugin install — often has its own runtime (TypeScript, MCP servers, commands)	Add an entry to sync-sources/plugins.json

A plugin can NOT be flattened into a skill: it carries slash commands and a
runtime that only Claude Code's plugin system can load. Conversely, a plain
skill should NOT go through the plugin manifest.

Step 2a — Bundled skill (sync-sources/skills/)

1. Copy the skill directory (the folder containing SKILL.md) into
   sync-sources/skills/<kebab-name>/. Vendor the whole directory —
   scripts, references, templates.
2. Ensure the frontmatter has name, description, and triggers. Match
   the style of two neighbors before editing.
3. Skills here reach all users on every machine and cost catalog space in
   every session — keep the description tight.

Dev-only skills follow the same shape under sync-sources/dev-skills/; they
sync only when running from an overdeck source checkout (isDevMode()).

Step 2b — Claude Code plugin (sync-sources/plugins.json)

1. Find the marketplace name and plugin name in the repo's
   .claude-plugin/marketplace.json (name at the top level is the
   marketplace; each plugins[].name is a plugin).

2. Append an entry to sync-sources/plugins.json:

   {
     "plugin": "<pluginName>@<marketplaceName>",
     "marketplace": "<owner>/<repo>"
   }

   Example — the OpenAI Codex plugin (openai/codex-plugin-cc declares
   marketplace openai-codex with plugin codex):

   { "plugin": "codex@openai-codex", "marketplace": "openai/codex-plugin-cc" }

3. Provisioning is implemented in src/lib/claude-plugins-provision.ts and
   runs as a pan sync step: for each manifest entry it checks
   claude plugin list --json, adds the marketplace if missing
   (claude plugin marketplace add <source>), then installs at user scope
   (claude plugin install <plugin> --scope user). It is idempotent and
   degrades to a warning when claude is missing or the network install
   fails — it never fails the sync.

4. Plugins install for all users on machines with the claude CLI. There
   is no dev-only plugin tier; if one is ever needed, that's a code change in
   the provisioning module, not a manifest convention.

Step 3 — Sync, verify, commit

npm run build          # provisioning runs from dist via the pan CLI
pan sync               # look for the "Provisioning Claude Code plugins" /
                       # skills sync lines in the output
claude plugin list     # plugin installs only: confirm the id appears

For skills, confirm the copy landed in ~/.claude/skills/<name>/ and
~/.agents/skills/<name>/.

Then commit and push — bundled content ships with the package, so an
uncommitted skill or manifest entry exists only on your machine. Other users
receive it on their next pan install / pan sync after release.

Related

• /pan-skill-creator — authoring a NEW skill from scratch (shapes, frontmatter, lint rules)
• /pan-sync — how the sync pipeline distributes bundles to each harness
• docs/SKILLS-CONVENTION.md — wrapper-skill naming rules and the lint that enforces them