The Agent Skills Directory

[DYNAMIC_EXECUTION]: Runtime library compilation and process injection.
Evidence: The file scripts/office/soffice.py contains embedded C source code designed to intercept socket calls. This code is written to a temporary file, compiled using gcc into a shared object, and loaded into the LibreOffice process space via the LD_PRELOAD environment variable.
Intent: This technique is used to provide fallback AF_UNIX socket support in restricted or sandboxed environments.
[COMMAND_EXECUTION]: Execution of system-level binaries.
Evidence: Multiple scripts (recalc.py, scripts/recalc.py, scripts/office/soffice.py, scripts/office/validators/redlining.py) utilize subprocess.run to call external tools including soffice (LibreOffice), gcc, git, and timeout. These are used for tasks like formula recalculation, diffing spreadsheet content, and managing runtime environment shims.
[INDIRECT_PROMPT_INJECTION]: Potential risk surface through untrusted data ingestion.
Ingestion points: Data enters the agent context via pandas.read_excel and openpyxl.load_workbook (referenced in SKILL.md) when processing user-provided spreadsheet files.
Boundary markers: Absent; no specific delimiters or instructions to ignore embedded content are provided for the data being processed.
Capability inventory: The skill has access to shell execution (subprocess.run in recalc.py and soffice.py) and file system write operations.
Sanitization: Content from processed Excel files is not explicitly sanitized before being handled by the agent context.

xlsx