The Agent Skills Directory

[DYNAMIC_EXECUTION]: The module scripts/office/soffice.py implements a sophisticated runtime execution pattern. It contains a C source code string (_SHIM_SOURCE) that is written to a temporary file and compiled into a shared object (.so) using gcc during execution. This compiled library is then injected into the LibreOffice process using the LD_PRELOAD environment variable. While this is used to work around socket restrictions in sandboxed environments, runtime compilation and process injection are high-risk patterns.
[COMMAND_EXECUTION]: Several components of the skill execute external binaries via the subprocess module:
scripts/office/soffice.py calls gcc to compile the socket shim.
scripts/thumbnail.py and scripts/office/soffice.py execute soffice (LibreOffice) for file conversion.
scripts/thumbnail.py executes pdftoppm to generate slide images.
scripts/office/validators/redlining.py executes git diff for comparing document versions.
[INDIRECT_PROMPT_INJECTION]: The skill represents an attack surface for indirect injection as it processes untrusted Office documents.
Ingestion points: Untrusted data enters the agent context through file extraction in scripts/add_slide.py and text extraction via markitdown as described in SKILL.md.
Boundary markers: Scripts do not implement explicit delimiters or instructions to ignore embedded commands within the processed slides.
Capability inventory: The skill has broad capabilities including arbitrary file system reads/writes (via zip manipulation and XML editing) and multiple subprocess execution paths.
Sanitization: The skill mitigates XML-based attacks like XXE by using defusedxml for parsing and lxml with default security settings.

pptx