The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the subprocess module to execute external binaries for screen capture and to launch the Godot editor application.
scripts/capture.py executes the grim utility on Wayland sessions and uses wmctrl and xdotool for window management on X11 sessions.
scripts/stage_editor_bridge.py uses subprocess.Popen to execute the Godot binary path provided via the --godot argument, allowing for the execution of arbitrary local binaries if the input is manipulated.
[DYNAMIC_EXECUTION]: The skill dynamically loads internal Python modules using string-based paths, which is an obfuscation-adjacent practice and a potential vector for loading unexpected code.
scripts/capture.py utilizes the __import__ function to load helper scripts such as asset_sheet.py, webp_encode.py, and ensure_gitignore.py from local paths.
[INDIRECT_PROMPT_INJECTION]: The skill possesses an attack surface for indirect prompt injection because it ingests untrusted visual data (screenshots) from the host machine and instructs the agent to process and score this data against a complex set of criteria.
Ingestion points: Screen captures are generated via mss or grim and saved as WebP files in scripts/capture.py.
Boundary markers: The instructions do not define clear delimiters or warnings to prevent the agent from following instructions embedded within the captured images.
Capability inventory: The skill includes capabilities for subprocess execution (capture.py, stage_editor_bridge.py) and project file modification (project.godot).
Sanitization: There is no filtering or sanitization of the visual content before it is read by the agent's vision model.

godot-agent-vision