The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes local command-line tools such as qsv, DuckDB, and LibreOffice via subprocess.run calls in its helper scripts (peek.py, profile.py, and recalc.py). These executions are used exclusively for data analysis and transformation tasks, using list-based arguments to ensure safe handling of file paths and delimiters.
[DYNAMIC_EXECUTION]: In recalc.py, the skill generates a temporary StarBasic macro and executes it within a headless LibreOffice instance to perform spreadsheet recalculation. The macro is hardcoded to perform only benign operations: calculating formulas, saving the workbook, and closing the application.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external spreadsheet data, which is an inherent vector for CSV injection or indirect prompt injection. The skill mitigates this through a mandatory evidence chain: it identifies ingestion points in peek.py and profile.py, establishes trust boundaries in SKILL.md, and provides explicit sanitization rules (Invariant 6) to escape formula-prefix characters (=, +, @) in external datasets.

spreadsheets