The Agent Skills Directory

[COMMAND_EXECUTION]: The script references/vr_calibrate.py uses subprocess.run to call its companion tool vr_score.py. This is used to sweep directories of renders through vision models for calibration. The execution path is resolved safely using Path(__file__).resolve().parent, targeting a known script within the skill's own directory.
[DYNAMIC_EXECUTION]: The skill provides extensive Python code blocks for use with execute_blender_code. These scripts automate scene manipulation, rendering, and diagnostic data extraction. While powerful, this behavior is central to the skill's purpose of automating Blender tasks.
[INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it ingests and processes external model data and generated images. These are analyzed by a vision model (VLM) using queries from references/checks_hardsurface.json. The risk is mitigated by the model being hosted locally (Ollama) and the specialized nature of the geometric auditing tasks.
[DATA_EXFILTRATION]: The script references/vr_score.py communicates with a local vision model via urllib.request targeting 127.0.0.1:11434. This is the standard port for local Ollama instances and does not involve external network requests.

blender-visual-review