The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The scripts/bootstrap.py script downloads the Lightcraft CLI binary from GitHub repositories. The download logic specifically whitelists the official upstream project (storytold/lightcraft) and the skill vendor's repository (full-aigc-skills).- [REMOTE_CODE_EXECUTION]: The skill installs and executes a remote binary. This is mitigated by a hardcoded integrity check in scripts/runtime.lock.json, which mandates a specific SHA256 hash for the binary before execution.- [COMMAND_EXECUTION]: Multiple scripts, including scripts/cli.py and scripts/artifacts.py, execute the Lightcraft CLI and macOS system tools like sips. All commands are executed using list-based arguments to prevent shell injection, and file paths are resolved and validated against allowed output roots.- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied images and JSON manifests. Mitigations include:
Ingestion points: External image files and JSON manifests read by scripts/artifacts.py and scripts/command_gateway.py.
Boundary markers: Use of structured data with strict schema validation.
Capability inventory: File system access and CLI tool execution.
Sanitization: Resolution of file paths and is_relative_to checks to prevent path traversal attacks.- [DYNAMIC_EXECUTION]: Several scripts use importlib to load local modules from the scripts/ directory at runtime. This behavior is restricted to the skill's own internal components and does not load from untrusted paths.

lightcraft-use