The Agent Skills Directory

[COMMAND_EXECUTION]: Runtime compilation of system utilities. The file scripts/office/soffice.py contains an embedded C source string that is written to a temporary file and compiled at runtime into a shared library (lo_socket_shim.so) using gcc via subprocess.run. This is used to facilitate socket-level communication for LibreOffice in restricted environments.\n- [COMMAND_EXECUTION]: Process library injection. The skill utilizes the LD_PRELOAD environment variable to load the dynamically compiled library into the soffice process, hooking networking-related system calls such as socket, listen, and accept.\n- [COMMAND_EXECUTION]: Execution of external binaries. The skill frequently invokes system binaries including soffice, gcc, and git for formula processing, component compilation, and document diffing through subprocess.run calls.\n- [PROMPT_INJECTION]: Indirect prompt injection surface.\n
Ingestion points: Untrusted data enters the agent's context when reading external spreadsheet files (.xlsx, .csv) via pandas and openpyxl.\n
Boundary markers: The instructions lack explicit requirements for delimiters or warnings to ignore embedded instructions within cell content.\n
Capability inventory: The ability to execute shell commands and write to the filesystem provides a significant surface that could be exploited if malicious instructions are processed from ingested files.\n
Sanitization: While the skill mitigating XML-level threats using defusedxml is a secure practice, it does not include semantic filtering for natural language instructions in data cells.

xlsx