ai_stp

# `plugin` A `plugin` is a native extension of a harness. It can add commands, UI, integrations, MCP configuration, skills or other surfaces, where the particular harness supports them. `ai_stp` treats a plugin as a component of its own, because installing one usually carries more supply-chain and runtime risk than a Markdown instruction. ## Plugin or another component? | If the extension | Kind | | --- | --- | | only gives the agent rules | `instruction` | | describes a workflow | `skill` | | connects one MCP server | `mcp` | | adds a harness package or extension | `plugin` | | changes parameters only | `setting` | ## What is checked | Check | Why | | --- | --- | | source and version | so a substituted package is not installed | | supported harness | the plugin format is not universal | | post-install behaviour | to know what changes after installing | | dependencies | to judge the supply chain | | provider plan | to see the exact changes to the target | === "Catalog" Shows what it is for, the author, the version, compatibility and the trust line. === "Plan" Shows which files, indexes or package surfaces the provider will change. === "Rollback" Must return the target to its state before the install, as far as that harness's provider allows. !!! danger "A verified author is not a safe plugin" A verified author helps you understand provenance. It does not prove the plugin's content is safe. Look at the component's own independent checks.