The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides a controlled wrapper script, scripts/run_with_env.py, which uses subprocess.run to execute the aliyunpan command-line interface. It includes a whitelist (ALLOWED_EXECUTABLES) to prevent the execution of arbitrary commands, ensuring the agent's capabilities are restricted to the intended cloud drive operations.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the cloud drive in the form of filenames and directory listings, which are processed by the agent (e.g., during ls operations or when using scripts/scan_drive.py). The absence of boundary markers or instructions to disregard instructions potentially embedded within these filenames presents a surface for indirect prompt injection attacks. Evidence: scripts/scan_drive.py parses file and directory names into JSONL format which is then processed by other skills.
[EXTERNAL_DOWNLOADS]: Instructions are provided to install the aliyunpan CLI tool via Homebrew and other vendor-specific components via npx and the Claude plugin system. These downloads originate from well-known registries and established developer tools.
[CREDENTIALS_UNSAFE]: The skill manages sensitive authentication materials for Aliyunpan. It does not hardcode credentials but provides explicit operational guidelines to prevent the accidental exposure of tokens or session data. It uses a private configuration file for environment variables to keep secrets out of the codebase and includes a wrapper script designed to avoid echoing command-line arguments that might contain secrets.

soia-pkm-alipan-drive-ops