The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses python3 to execute internal utility scripts such as bootstrap.py, workflow.py, and commands.py. These executions utilize the -I (isolated) and -B (no bytecode) flags to mitigate environment-based attacks. It also employs npx to manage skill dependencies.
[EXTERNAL_DOWNLOADS]: The skill downloads a pinned version of its command-line interface from the vendor's repository (full-aigc-skills/vectorcraft-skills) and an upstream repository (storytold/vectorcraft). These downloads are protected by mandatory SHA-256 checksum verification and version locking to prevent supply chain attacks.
[DYNAMIC_EXECUTION]: The skill interprets a domain-specific language defined in JSON 'plans' (e.g., craft-command-plan/v1) to execute a sequence of native vector operations. This execution is gated by a catalog of allowed commands and strict schema validation.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data in the form of vector graphics files and brand assets, creating a potential injection surface.
Ingestion points: External data enters via the document.open command and the asset.place workflow, which accepts local file paths.
Boundary markers: The system uses structured JSON schemas and manifest bindings to delimit data, but does not explicitly instruct the agent to ignore instructions that might be embedded within vector metadata or text elements.
Capability inventory: The agent has the ability to write project files, execute localized scripts, and interact with a desktop environment via a loopback bridge.
Sanitization: Security is enforced through kernel-level filesystem scoping (preventing access outside authorized project roots), descriptor-based file traversal (preventing path traversal and symlink attacks), and mandatory binary integrity checks.

vectorcraft-cli