The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill manages authentication tokens and API keys in a local file (credentials.json). The manifest.yaml correctly identifies these as sensitive fields and ensures they are not committed to version control, following standard secret management practices for local CLI tools.
[COMMAND_EXECUTION]: In scripts/_auth.py, the subprocess.run function is used to execute a local script (open_login_page.py) to assist with user authentication. This execution is limited to the skill's own directory and uses the current Python interpreter.
[EXTERNAL_DOWNLOADS]: The scripts/download_result.py script allows the user to download generated media files from remote URLs to the local outputs directory. This is an expected feature for retrieving assets generated by the API.
[DYNAMIC_EXECUTION]: The scripts/_task_api.py file uses __import__('os') to dynamically load the os module for environment variable access. This is a common coding pattern and does not incorporate untrusted input into the import path.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data that is passed to its generative functions.
Ingestion points: User-provided CLI arguments in scripts/create_photo_task.py and scripts/create_motion_task.py (e.g., --detail, --background).
Boundary markers: None identified; input is interpolated directly into API payloads.
Capability inventory: Network requests via urllib.request and local file writing via scripts/download_result.py.
Sanitization: Arguments are passed as string literals in JSON payloads without escaping or validation.

chanjing-text-to-digital-person