日本語

Tracking issues: #55, #56, #57, and #58.

The repository's agent instructions currently repeat first-run decisions on every task. License selection, missing governance files, and initial policy generation belong to repository creation or deliberate initialization. Ongoing work needs only the resulting language and project rules.

repository-init is available only through an explicit $repository-init invocation. It inspects the target root, preserves clear existing licenses and project-specific rules, creates only missing AGENTS.md and English SECURITY.md files, and stores the resolved decision in .repository-init.json.

The marker contains only schema_version, status, license, and language_profile. in_progress allows a later run to resume after interruption. Invalid records and conflicting choices fail closed. A completed run performs no writes. Semantic verification is performed before the helper changes the marker to complete.

The skill does not initialize Git, create remotes, create Issues or Work Items, publish Wiki pages, commit, or push. These remain separate repository operations.

MIT repositories use Japanese agent responses, English comments, documents, and commit messages, with _ja.md translations where appropriate. Non-MIT repositories use Japanese comments, documents, and commit messages without _ja.md translations. Security and license files remain English for both profiles.

• Add Claude/Codex plugin manifests and register the plugin in both repository marketplaces.
• Provide English and Japanese skill instructions, MIT and non-MIT AGENTS.md templates, a Japanese reference translation, and an English SECURITY.md template.
• Implement an atomic, lock-protected state helper with fail-closed parsing and no-op completed transitions.
• Test missing, in-progress, complete, interrupted, conflicting, malformed, locked, atomic-write failure, alternate license filename, and repeated invocation cases.
• Run the repository marketplace validator and the skill validator.