Plugin Creator

Use this skill when a user wants a local Codewhale plugin bundle. Trusted and enabled bundles may add declarative Skills, commands, agents, hooks, and MCP servers (stdio and remote) through the existing engines. LSP, filesystem roots, and lifecycle mutation are inventory-only. Native extensions (host code) are inventory-only unless the user has turned on the experimental [features] extension_host flag.

Workflow

1. Pick a Codewhale-owned location:
   • User bundle: ~/.codewhale/plugins/<plugin-name>/
   • Workspace bundle: <workspace>/.codewhale/plugins/<plugin-name>/
2. Normalize the bundle name to lowercase hyphen-case.
3. Create plugin.json (Agent Plugins v1.0.0; a legacy plugin.toml stays readable, but new bundles use plugin.json):

{
  "$schema": "https://agent-plugins.org/schemas/plugin.json",
  "name": "my-plugin",
  "version": "0.1.0",
  "description": "What this bundle provides"
}

4. Put each Skill under skills/<skill-name>/SKILL.md; Codewhale finds skills/ automatically and exposes each as my-plugin:<skill-name>, never as an unqualified command.
5. Add MCP servers in a sibling mcp.json only when the bundle needs an existing MCP engine. Keep stdio commands and paths inside the bundle. Map local environment values only as exact ${SOURCE_ENV} references. For remote MCP, use HTTPS (or loopback HTTP), forbid URL user information/query/fragment, use only environment-backed headers or bearer tokens, and declare the exact normalized endpoint host set in capabilities.network_hosts under extensions["net.codewhale"]. Never place credentials in the manifest.
6. Commands (commands/*.md), agents (agents/*.toml), and hooks (hooks/*.toml), declared under extensions["net.codewhale"], activate under the current policy — workspace bundles win same-name collisions over user and built-in bundles. LSP, filesystem roots, and lifecycle mutation are inventory-only: declare them only when inventorying future work. A native entry runs only under the experimental extension host; there it must be one .mjs, .js or .mts ES module file, /plugin validate rejects anything else, and its tools always use Required approval, never a plugin's read-only hint. Full Access, Bypass, or an exact session grant for the reviewed build can satisfy that gate without a prompt. A bundle that declares only unsupported surfaces cannot be enabled.
7. Validate and review without executing bundle content:
   • /plugin validate <plugin-name>
   • /plugin show <plugin-name>
   • stop and present these results; the person runs /plugin enable <plugin-name> to open the content/capability review, reviews it, runs the exact /plugin trust ... confirmation shown, then enables the bundle
8. Verify /skills inspect reports plugin provenance and /plugin list reports the expected trust and activation state. Trust stages the reviewed content but does not activate it. After enablement, follow the host's reload notice: use /reload or a new session to apply changes to a live session's pinned skills and tools.

Every user and workspace bundle starts untrusted and disabled. Reuse the existing /plugin marketplace, install, update, review and reload surfaces; do not add a parallel installer, registry or automatic trust flow. Catalog membership alone never installs, trusts or enables a plugin.

Experimental host code

Only scaffold host code when the person explicitly uses the experimental extension-host feature. Start from the tested hello-extension example and docs/EXTENSIONS.md in the Codewhale repository. A typed .mts entry may use Node's erasable TypeScript syntax; bundle dependencies locally. Register tools with a plugin-specific prefix and an object input schema, propagate exec.signal, and use ctx.effect for bounded asynchronous cleanup. The current execution context exposes signal, callId and args; it does not expose the calling workspace path. Do not change the shared process cwd.

Stop after install, validate and show; never automate the trust token. A person reviews, trusts and enables the bundle. /plugin show <name> reports owner state, live tools and recent attributed diagnostics. Recovery may create fresh registrations, but never replays an interrupted tool call. Explain the shared-process and current platform sandbox limits without claiming isolation.