The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/office/soffice.py implements a process injection technique where C source code is generated at runtime, compiled into a shared library using gcc, and loaded into the LibreOffice process via the LD_PRELOAD environment variable. This is used to bypass socket restrictions in sandboxed environments but constitutes high-risk dynamic execution behavior.
[COMMAND_EXECUTION]: The script scripts/accept_changes.py programmatically generates and executes LibreOffice macros (StarBasic) by writing them to a temporary profile on disk and invoking them through the soffice binary.
[PROMPT_INJECTION]: The skill serves as an ingestion point for untrusted data by extracting text from user-provided documents using pandoc or raw XML access. 1. Ingestion points: Document text extraction in SKILL.md and raw XML reading in scripts/office/unpack.py. 2. Boundary markers: Absent; extracted content is interpolated into the agent's context without delimiters or instructions to ignore embedded commands. 3. Capability inventory: Extensive use of subprocess.run() across multiple files (scripts/accept_changes.py, scripts/office/soffice.py, scripts/office/validators/redlining.py) and runtime compilation capabilities. 4. Sanitization: Absent; content from external documents is processed without filtering or validation of the text itself.
[COMMAND_EXECUTION]: Multiple components utilize subprocess.run() to execute system tools including pandoc, pdftoppm, git, and soffice, which could be exploited if input handling is not strictly controlled.

docx-official