The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external vector project files (.vectorcraft) and JSON operation plans, creating a surface for indirect prompt injection via embedded metadata or design instructions.
Ingestion points: The skill ingests user-provided vector projects and JSON plans through workflow.py and commands.py as seen in SKILL.md and references/workflow.md.
Boundary markers: Employs SHA-256 digest binding (expectedProjectSha256) and structured JSON schemas (craft-command-plan/v1) to differentiate instructions from project data.
Capability inventory: The skill can perform file system writes, native CLI command execution, and local process management for GUI operations.
Sanitization: Validates commands against an internal whitelist, checks for unique JSON keys, and verifies project integrity before modification.
[EXTERNAL_DOWNLOADS]: The skill automates the download and setup of the VectorCraft CLI and related desktop components.
Evidence: SKILL.md and references/desktop-install.md describe a bootstrap process that fetches pinned binaries for the macOS arm64 architecture.
Security context: Downloads originate from vendor-managed sources and are verified against hardcoded SHA-256 checksums and architecture signatures before execution.
[COMMAND_EXECUTION]: Triggers local execution of Python scripts and native engine commands to process vector tasks.
Evidence: SKILL.md uses python3 -I -B to run bootstrap.py, cli.py, and commands.py.
Security context: Native command execution is restricted by a kernel-level policy (filesystem_scope.py) that enforces explicit file roots, limiting the engine's access to authorized project directories.

vectorcraft-cli-export