The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of subprocess.run in Python and execFile in Node.js to interact with external tools.
scripts/apply_reclass.py and scripts/apply_reclass_bulk.py execute commands through a dedicated environment-loading runner from a dependent skill (soia-pkm-alipan-drive-ops).
Commands are restricted to a whitelist (mkdir, mv, rename) and validated against root directory boundaries.
scripts/gen_catalog_xlsx.py and scripts/gen_family_nav_xlsx.mjs invoke Node.js and LibreOffice (soffice) for spreadsheet generation and verification.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data in the form of file and directory names from Aliyun Drive.
There is a clear attack surface where malicious file names could influence LLM planning or script execution.
The developer has implemented defensive coding practices in scripts/catalog_xlsx/build_workbooks_fallback.py to prevent CSV/formula injection by prefixing suspicious characters (=, +, -, @) in spreadsheet cells.
Extensive preflight checks (scripts/preflight_reclass.py) verify the state of the drive against the intended plan before any modifications occur, mitigating risk from unexpected drive state changes.
[EXTERNAL_DOWNLOADS]: SKILL.md provides instructions to install the skill and its dependencies using npx skills add. These resources originate from the same author (soia-team) and represent expected vendor functionality.
[PROMPT_INJECTION]: The SKILL.md instructions include detailed rules for the agent to follow, emphasizing user confirmation for destructive actions (deletions, overwrites) and strict adherence to specific data sources. No patterns were found that attempt to bypass safety filters or override system constraints in a malicious manner.

soia-pkm-alipan-curator