The Agent Skills Directory

[COMMAND_EXECUTION]: Runtime compilation and process injection. The file scripts/office/soffice.py contains a C source code string (_SHIM_SOURCE) that is compiled into a shared object using gcc at runtime. This object is then injected into the soffice process using the LD_PRELOAD environment variable to intercept and redirect socket calls. This mechanism is used to bypass environment-specific socket restrictions.
[COMMAND_EXECUTION]: Dynamic script generation. The script scripts/accept_changes.py generates a LibreOffice Basic macro (ACCEPT_CHANGES_MACRO) and writes it to a temporary user profile directory before executing it via the soffice command line to accept tracked changes.
[COMMAND_EXECUTION]: The skill invokes several system utilities via subprocess.run(), including soffice for document conversion, gcc for shim compilation, and git for generating diffs during validation in scripts/office/validators/redlining.py.
[PROMPT_INJECTION]: Surface for indirect prompt injection. The skill is designed to ingest and process external Word documents using pandoc and custom XML parsing logic. Maliciously crafted content or metadata within these documents could potentially override agent instructions during analysis.
Ingestion points: pandoc extraction commands in SKILL.md and scripts/office/unpack.py.
Boundary markers: No explicit markers are used to isolate extracted text when interpolated into agent prompts.
Capability inventory: Shell command execution (soffice, gcc), file system access, and package management.
Sanitization: Uses the defusedxml library to mitigate standard XML-based attacks.
[EXTERNAL_DOWNLOADS]: The skill instructions include a global installation command for the docx package via npm install -g docx. This is a standard dependency for the skill's functionality.

alterlab-docx