The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script 'scripts/office/soffice.py' implements a mechanism to work around restricted environments that block AF_UNIX sockets. It dynamically writes a C source file to a temporary directory, compiles it into a shared object using 'gcc', and then uses 'LD_PRELOAD' to inject this library into the LibreOffice ('soffice') process. This enables system call hooking of 'socket', 'listen', and 'accept'.
[COMMAND_EXECUTION]: The skill uses 'subprocess.run' to execute external binaries including 'gcc' for shim compilation, 'soffice' for document conversion, 'pdftoppm' for image rendering, and 'git diff' for document comparison.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted PowerPoint files and extracts content for analysis, creating a surface for indirect prompt injection. Mandatory Evidence Chain: (1) Ingestion points: 'scripts/thumbnail.py' and 'markitdown' extract content from user-provided PPTX files. (2) Boundary markers: No explicit delimiters or instructions to ignore embedded commands are present in the extraction logic. (3) Capability inventory: 'subprocess.run' for arbitrary commands, file-system access, and runtime compilation are available across the skill scripts. (4) Sanitization: While 'defusedxml' is used for some parsing, content is processed directly for analysis without validation against prompt injection patterns.

pptx