The Agent Skills Directory

Dynamic Execution via Runtime Compilation and Injection: The script scripts/office/soffice.py contains an embedded C source shim that is compiled at runtime using gcc and then injected into the LibreOffice process via the LD_PRELOAD environment variable. This is a highly sophisticated technique used here to facilitate communication in sandboxed environments where standard socket operations are restricted. While this represents a complex execution pattern, it is used to enable the core recalculation feature of the skill.
System Command Execution: The skill makes several calls to external system utilities using the subprocess module. These include invoking soffice for formula recalculation in scripts/recalc.py, gcc for the environment shim in scripts/office/soffice.py, and git for document diffing in scripts/office/validators/redlining.py. These executions are central to the skill's advertised functionality.
Indirect Prompt Injection Surface: The skill is designed to ingest and process untrusted spreadsheet and document data, creating a potential surface for indirect prompt injection.
Ingestion points: Data enters the agent context through the parsing of OOXML files (e.g., document.xml, workbook.xml) in scripts/recalc.py and various scripts in scripts/office/validators/.
Boundary markers: Explicit boundary markers for ignoring embedded instructions in ingested spreadsheets were not observed in the provided scripts.
Capability inventory: The skill possesses capabilities to execute shell commands (via soffice, gcc), write files (via openpyxl and Path.write_bytes), and perform network-related operations if LibreOffice is configured to do so.
Sanitization: The skill effectively uses the defusedxml library in scripts/office/validators/base.py and redlining.py, which is a recognized safety practice to mitigate XML-based vulnerabilities.

xlsx