The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple scripts within the skill utilize subprocess.run to execute local binaries for rendering and diagnostic purposes:
scripts/render_pptx.py invokes chrome.exe (or msedge.exe), soffice (LibreOffice), pdftoppm, and magick to render PowerPoint slides.
scripts/quality_audit.py executes chrome.exe (or msedge.exe) to perform headless rendering of SVG previews for quality checks.
scripts/pptx_narration.py runs ffprobe to analyze the duration of narration audio files.
scripts/generate_clean_plate.py includes a mechanism to execute an arbitrary command defined in an external JSON configuration file specified by the FIGEDIT_CLEAN_PLATE_CONFIG environment variable.
[EXTERNAL_DOWNLOADS]: The scripts/generate_clean_plate.py script performs network operations to external AI image generation providers. It makes outbound HTTP POST requests using urllib.request.urlopen to api.labnana.com and generativelanguage.googleapis.com (Google Gemini), and uses the OpenAI SDK to interact with OpenAI's API. These external calls are required for the skill's 'AI Clean Plate' and 'Chroma Regeneration' features.
[DATA_EXFILTRATION]: While performing its intended function, the skill transmits project-specific image data (encoded as base64 strings) to third-party AI services (Labnana, OpenAI, Google) via scripts/generate_clean_plate.py. Users should be aware that their visual assets are shared with these providers during processing.
[PROMPT_INJECTION]: The skill presents an indirect prompt injection surface as it ingests untrusted image data and uses it to generate prompts for downstream AI models (as described in references/image_generation.md).
Ingestion points: Source images provided in the manifest.json (e.g., assets/source.jpg).
Boundary markers: The instructions in references/image_generation.md establish structural blocks for prompts, but they do not guarantee protection against adversarial content embedded within the source images.
Capability inventory: The skill has capabilities for file system access, subprocess execution, and external network communication.
Sanitization: Standard image libraries (NumPy, Pillow) are used to manipulate the data before transmission, which does not sanitize semantic instructions embedded in the image.

figedit-v2