The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/office/soffice.py performs runtime compilation of C source code into a shared object library using gcc. This compiled library is subsequently injected into the soffice (LibreOffice) process via the LD_PRELOAD environment variable. This technique is used to shim system calls to handle socket restrictions in sandboxed environments.
[COMMAND_EXECUTION]: The skill uses subprocess.run to execute several system binaries: soffice for formula recalculation (scripts/recalc.py), gcc for compilation tasks (scripts/office/soffice.py), and git for comparing document content (scripts/office/validators/redlining.py).
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to its handling of untrusted spreadsheet data combined with its command execution capabilities:
Ingestion points: Untrusted data from .xlsx, .csv, and .tsv files enters the agent context through pandas.read_excel and openpyxl.load_workbook as described in SKILL.md.
Boundary markers: There are no instructions or delimiters defined to distinguish data content from instructions, nor are there warnings to the agent to ignore embedded commands.
Capability inventory: The skill has the ability to execute arbitrary system commands via subprocess.run calls in multiple supporting scripts.
Sanitization: No evidence of content sanitization or validation is present before data interpolation or process invocation.

xlsx