The Agent Skills Directory

[COMMAND_EXECUTION]: The skill invokes the pdftoppm utility via subprocess.run for PDF page rasterization in scripts/_raster.py. The command is constructed using a list of arguments rather than a single shell string, which prevents command injection from malicious file paths or parameters.
[DYNAMIC_EXECUTION]: Helper scripts like scripts/pdf_read.py and scripts/pdf_form_layout.py use __import__ to lazily load optional dependencies or resolve paths. These calls use hardcoded module names and do not involve untrusted input, adhering to safe dynamic loading practices.
[INDIRECT_PROMPT_INJECTION]: As the skill extracts text and data from external PDF files for agent processing (e.g., in scripts/pdf_read.py and scripts/extract_marker.py), it possesses an indirect prompt injection surface. The skill has access to file writing and command execution tools, but it mitigates associated risks (like path traversal) by sanitizing filenames using os.path.basename when extracting embedded attachments in scripts/pdf_meta.py.
[EXTERNAL_DOWNLOADS]: The marker-pdf tool, referenced in scripts/extract_marker.py, is used for high-quality OCR. This library is documented to download pre-trained machine learning models from HuggingFace upon first use, which is a standard behavior for this well-known service.

pdf