The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill downloads binary runtimes and desktop applications from GitHub and executes them locally. It uses SHA256 integrity checks to verify the binaries before execution.
[EXTERNAL_DOWNLOADS]: Fetches executable binaries, application packages, and machine learning models (e.g., Whisper) from GitHub releases and official repositories.
[COMMAND_EXECUTION]: Uses the Python subprocess module to run the native CLI, interact with system tools like hdiutil and codesign, and enforce macOS-specific process isolation via sandbox-exec.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process external, potentially untrusted JSON-based command plans and media assets (video, audio, LUTs).
Ingestion points: Files like scripts/workflow.py, scripts/commands.py, and scripts/desktop.py read plan files and resolve asset paths into the execution environment.
Boundary markers: The skill enforces strict JSON validation, including duplicate key detection, and verifies all command IDs against a hardcoded catalog.
Capability inventory: The tool possesses the capability to write to the filesystem within defined roots, execute the native binary, and access the network for model downloads.
Sanitization: Security is managed by executing subprocesses within a macOS sandbox using /usr/bin/sandbox-exec, which limits the process to a deny-by-default policy, only allowing access to specific authorized directories and resources. Paths are further validated against user-defined read/write roots.
[DYNAMIC_EXECUTION]: The scripts/cli.py and scripts/workflow.py scripts dynamically load internal modules and policies from the skill's own directory using importlib.

filmcraft-cli-motion