The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill downloads and executes locked CLI and desktop binaries from the author's official GitHub repositories (full-aigc-skills and storytold). Security is maintained through hardcoded SHA256 hashes in runtime.lock.json and desktop.lock.json, which are verified before any execution.
[COMMAND_EXECUTION]: Native VectorCraft tools are executed as subprocesses. The skill uses the macOS sandbox-exec utility (in scripts/filesystem_scope.py) to enforce a strict sandbox policy, restricting the tools to specific authorized filesystem roots and loopback network ports.
[DYNAMIC_EXECUTION]: The harness dynamically loads its own internal modules (e.g., input_security.py, filesystem_scope.py, artboard_mapping.py) using importlib.util. All module paths are strictly computed relative to the skill's own directory, preventing the execution of arbitrary external scripts.
[DATA_EXPOSURE]: A dedicated safety module (scripts/input_security.py) implements a CREDENTIAL_FIELDS filter that proactively identifies and scrubs API keys, tokens, and passwords from input data to prevent their exposure to the native execution environment.
[INDIRECT_PROMPT_INJECTION]: The skill processes external vector assets (SVGs) and implements a robust sanitization layer in scripts/asset_inputs.py. It explicitly blocks potentially malicious SVG features such as <script> tags, foreignObject elements, and external URL references (href, @import, or CSS url()).

vectorcraft-harness