dsh-claude-plugin-loader

The Claude Code ecosystem already ships a growing set of composite plugins, each with a plugin.json manifest, a skills/ directory, and optionally a commands/ directory, that dsh had no native way to discover or load. dsh-claude-plugin-loader fills that gap: it is a runtime loader that scans the local Claude plugin cache (~/.claude/plugins/cache by default), de-duplicates by name, keeps the highest version, and only surfaces plugins whose manifest declares dsh support, dropping existing Claude-style plugins straight into dsh skill and command pipelines.

Three channels carry the core functionality. The skills channel registers the skills/ directory of each plugin as a dsh skill provider, so the model can discover and invoke SKILL.md files through the native dsh skill / skill_search path. The commands channel turns commands/*.md files of each plugin into dsh slash commands with roughly 1.5-second hot-reload: new, modified, or removed commands are picked up automatically, and commands are cleaned up the moment a plugin is deleted or disabled. A default-off SessionStart hook bridge also parses command hooks from hooks.json and injects their stdout into the session context at agent start-up, useful for side-effect-sensitive workflows.

If you already use Claude Code, have a folder of composite plugins you want to keep, and also run dsh, this loader lets you reuse them without touching a single plugin file. If you are a dsh ecosystem developer, it doubles as a reference implementation for how dsh consumes the Claude-style plugin structure.