The Agent Skills Directory

[DYNAMIC_EXECUTION]: The module scripts/office/soffice.py generates and compiles C source code at runtime using gcc. The resulting shared object is then loaded into the execution flow of the soffice binary using the LD_PRELOAD environment variable. This mechanism is employed to shim AF_UNIX socket calls in restricted environments, but involves high-risk runtime code generation and process injection techniques.\n- [COMMAND_EXECUTION]: Multiple scripts execute shell commands via subprocess.run(): \n
scripts/office/soffice.py invokes gcc to compile binary components.\n
scripts/recalc.py and scripts/office/soffice.py execute soffice (LibreOffice) to perform spreadsheet operations.\n
scripts/office/validators/redlining.py executes git diff for text content comparison.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted spreadsheet data from .xlsx, .csv, and .tsv files, creating a vulnerability surface.\n
Ingestion points: Data enters the agent context via pandas.read_excel (documented in SKILL.md) and openpyxl.load_workbook (in scripts/recalc.py).\n
Boundary markers: The instructions do not specify the use of delimiters or warnings for the agent to ignore instructions embedded in cell data.\n
Capability inventory: The skill possesses significant capabilities, including shell execution and filesystem writes, which could be abused if the agent obeys malicious data content.\n
Sanitization: While defusedxml is correctly used for safe XML parsing, there is no logic to sanitize or filter potential prompt injection instructions within cell values before processing.

xlsx