The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs its primary functions by executing external CLI tools through shell commands.
Evidence: Multiple scripts including scripts/apply_reclass.py, scripts/preflight_reclass.py, and scripts/apply_reclass_bulk.py use subprocess.run to call a local runner script which executes the aliyunpan CLI tool.
Evidence: scripts/gen_catalog_xlsx.py and scripts/gen_family_nav_xlsx.mjs execute node and soffice (LibreOffice) to render workbooks and perform formula recalculations.
[REMOTE_CODE_EXECUTION]: The skill utilizes dynamic Python module loading to link various auditing and gating components.
Evidence: scripts/apply_reclass.py and scripts/apply_reclass_bulk.py use importlib.util and exec_module to load scripts from paths computed at runtime (relative to __file__).
[PROMPT_INJECTION]: The skill has an attack surface for indirect prompt injection as it processes file names and paths retrieved from external cloud storage scans.
Ingestion points: Reads from initial.scan.jsonl and plan.jsonl files in scripts/apply_reclass.py and scripts/audit_structure.py.
Capability inventory: Shell command execution via subprocess.run to manipulate cloud storage entities (move/rename/mkdir).
Boundary markers: While structured JSON is used, there are no explicit delimiters or warnings to the model to ignore potential instructions embedded in cloud folder names.
Sanitization: The skill employs path normalization and validation to ensure operations remain within a 'run bundle', and includes a defensive implementation in scripts/catalog_xlsx/build_workbooks_fallback.py that strips formula trigger characters (=, +, -, @) from Excel cells to prevent CSV/formula injection.

soia-pkm-alipan-curator