The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill generates and executes code at runtime through compilation and macro generation.
In scripts/office/soffice.py, the function _ensure_shim writes a C source file (lo_socket_shim.c) to a temporary directory and compiles it into a shared object using gcc.
In scripts/recalc.py, a StarBasic macro (RECALCULATE_MACRO) is dynamically written to a LibreOffice profile directory to automate document processing.
[PRIVILEGE_ESCALATION]: The skill employs process injection techniques to alter the execution environment of external binaries.
In scripts/office/soffice.py, the get_soffice_env function modifies the LD_PRELOAD environment variable to force the loading of a runtime-compiled shim into the soffice process.
[COMMAND_EXECUTION]: Several components execute arbitrary shell commands via the subprocess module.
scripts/recalc.py invokes soffice and gtimeout to manage formula recalculation.
scripts/office/soffice.py invokes gcc to compile binary shims.
scripts/office/validators/redlining.py invokes git diff for document comparison.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted spreadsheet and document data, creating a vulnerability surface for data-driven attacks.
Ingestion points: Spreadsheet and Office files (.xlsx, .csv, .docx) are opened and read using openpyxl, pandas, and markitdown as specified in SKILL.md.
Boundary markers: Absent; there are no instructions to the agent to treat data content as untrusted or to ignore embedded instructions.
Capability inventory: The skill possesses capabilities for file system modification (openpyxl), command execution (recalc.py), and runtime code generation (soffice.py).
Sanitization: Partial; while XML structures are validated against schemas in scripts/office/validators/, the textual content within those files is not sanitized before being presented to the agent context.

xlsx