Create Plugin Scaffold

Create plugin scaffold

Contract

Inputs

• Plugin name (required): the directory name and manifest identifier for the new plugin.
• Target root directory (required): the filesystem path under which the plugin tree is created.
• Target ecosystem (optional): the host agent platform the scaffold targets. Determines the manifest filename, schema, and registration API. Defaults to Claude Code (plugin.json with .claude-plugin/ schema).
• Entry point file name (optional): defaults to index.ts; supplied when a non-default entry is needed.
• Plugin manifest fields (optional): author, version, description, and capability list. Apply defaults to omitted fields.

Procedure

1. Resolve the target path and verify write access. Confirm the plugin name and target root are present. Resolve the target root to an absolute path and verify it exists and is writable. Compute the plugin directory as <target root>/<plugin name>. If it already exists and is non-empty, stop and report the collision rather than overwriting. Done when: the plugin directory path is computed and confirmed not to collide with an existing non-empty directory.
2. Create the standard directory tree. Create the plugin root, an agents/ subdirectory, and a skills/ subdirectory. Done when: the plugin root, agents/, and skills/ directories exist.
3. Write the host-specific plugin manifest(s) matching the target ecosystem. For Claude Code: write two manifests kept in parity: a root plugin.json carrying the agent-plugins.org schema ($schema, name, version default 0.1.0, description, author as an object, and an extensions block), and a .claude-plugin/plugin.json carrying the Claude manifest schema ($schema, name matching the root, displayName, version, description, author); the repo's gates require matching name and parity between the two, so populate both from the same values. For other ecosystems: write the manifest filename and schema that ecosystem requires. Done when: the manifest(s) are written with all fields populated from supplied values or defaults, conforming to the target ecosystem's schema.
4. Provide the entry surface the target ecosystem requires. For Claude Code: no entry point is needed; the plugin is declarative, loaded by the host from skills/ and agents/. For other ecosystems: write the entry point file at the plugin root using the supplied or default name, exporting the registration function the host calls to load the plugin. Done when: the entry surface the ecosystem requires is present, or a declarative host is confirmed to need none.
5. Validate the layout against the manifest. Confirm the manifest parses as valid JSON, every declared entry point file exists, the agents/ and skills/ directories are present, and no required manifest field is missing. Done when: every validation check passes or every failure is identified.

Failure and recovery

• Missing required input: stop before any write; report which input is missing. No files are created.
• Target root missing or not writable: stop before any write; report the path and the access error. No files are created.
• Plugin directory already exists and is non-empty: stop before any write; report the collision. No files are overwritten.
• Manifest write or parse failure: delete any files already written under the plugin path for this run, then report the failure. The plugin tree is left absent rather than partial.
• Validation failure: report every failing check with the offending path or field. Do not claim the done predicate holds. Leave the created tree in place only if every validation check passed; otherwise remove the partial tree and report the failure.
• Blocked or non-converged result: return the partial validation report and the exact blocker; never swallow the error or pretend success.

Output

A plugin scaffold tree under <target root>/<plugin name> containing the manifest, entry point, agents/, and skills/ directories, and a validation report listing each check result and the absolute path of every created artifact.