dsh-plugin-reload

Under DeepSeek Harness's everything-is-a-plugin architecture, changing a single line in an MCP server or an in-process plugin often means restarting the entire host process, interrupting every unrelated Loader entry and slowing the dev loop dramatically. dsh-plugin-reload gives the agent a reload_plugin tool: supply an entry ID, module name, or MCP serverName, and exactly that one entry is restarted while all others keep running.

The strategy adapts to the entry kind. MCP-client entries respawn the server child process, pick up the latest on-disk code, and re-register their tools. In-process plugin entries undergo a hard reload: the ESM/CJS module caches for the entry and its local source files are busted, the module is re-imported from disk, and its fibers are swapped onto the fresh copy—all without touching the host process. A dry_run flag previews the matched entry and the strategy that would execute, while auto, soft, and hard modes let you dial in the level of cache invalidation. If a re-import or re-apply fails, the entry automatically rolls back to its previous code.

This plugin is aimed at developers who iterate on MCP servers or Harness plugins at high frequency and want a change-to-effect cycle measured in seconds rather than minutes. As long as your profile ships with the web or headless bundle, the agent can call the tool directly with no manual patch editing. Keep in mind that the entry's tools are briefly unavailable during the swap, and the hard reload covers only the plugin's own source—changes to node_modules dependencies still require a host restart.