Skill Security
Domain: Skills System
Motivation
Skill security governs how untrusted or third-party skills are handled. In dsh, a skill’s effects run under the session’s file/process sandbox, its invocation policy controls which surfaces may load it, and the registry re-validates a loaded definition before accepting it. Bundled providers like dsh-skill-badge are immutable and register a fixed candidate, limiting attack surface.
Concrete Example
A dsh-skill-office script that writes a file is confined by the workspace-write sandbox, and a skill with disable-model-invocation: true cannot be triggered by the model at all.
Analogy
A locked pantry so a visiting cook can only use what you’ve left out.