The Agent Skills Directory

[DYNAMIC_EXECUTION]: In scripts/office/soffice.py, the skill includes a complete C source code string (_SHIM_SOURCE) that is written to a temporary file and compiled at runtime using gcc into a shared object. The skill then uses the LD_PRELOAD environment variable to inject this library into the soffice process, hijacking system calls like socket, listen, and accept.
[COMMAND_EXECUTION]: The skill performs several direct command executions via the subprocess module:
scripts/recalc.py and scripts/office/soffice.py execute soffice to perform formula calculations.
scripts/office/soffice.py executes gcc to compile the dynamic shim library.
scripts/office/validators/redlining.py executes git diff to compare document versions.
[PERSISTENCE_MECHANISMS]: The scripts/recalc.py script modifies the local environment by writing a persistent StarBasic macro (Module1.xba) to the LibreOffice user configuration directory on both Linux and macOS.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and process untrusted spreadsheet data which may contain embedded instructions.
Ingestion points: Data enters the agent context via pandas and openpyxl when reading .xlsx, .csv, and .tsv files as specified in SKILL.md and processed in various scripts.
Boundary markers: No boundary markers or 'ignore' instructions are used when interpolating file data into the agent's reasoning path.
Capability inventory: The skill has the capability to write files, modify system configurations, and execute arbitrary commands via the recalculation and shim compilation scripts.
Sanitization: No evidence of data sanitization or validation of cell contents before they are processed by the agent.

xlsx