The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied HTML DOM or DSL strings to render video components. This ingestion of untrusted external content represents a potential indirect prompt injection surface.
Ingestion points: Untrusted data enters via the --input, --dom, or --dsl arguments in scripts/render_wechat_channel.py.
Boundary markers: The skill relies on specific HTML comments () to identify insertion points in existing files, but does not use delimiters for the injection payload itself.
Capability inventory: The skill uses Bash and Read tools; the rendering script performs file system operations (read/write).
Sanitization: The implementation includes robust sanitization. scripts/render_wechat_channel.py uses html.escape for HTML rendering and a dedicated _escape_dsl function for Markdown output, ensuring that special characters are escaped and preventing malicious breakout from the component structure.
[COMMAND_EXECUTION]: The unit test file tests/test_render_wechat_channel.py uses subprocess.run to execute the skill's primary rendering script. This is restricted to testing purposes, uses controlled arguments, and does not involve the execution of untrusted shells in the runtime skill workflow.
[DATA_EXPOSURE]: The scripts/profile_store.py utility manages local user preferences by writing to a JSON profile in the user's config directory. The script implements a proactive security filter (SENSITIVE_PARTS) that prevents the persistence of tokens, secrets, or credentials within the profile records, following best practices for secret management.
[SAFE]: The skill adheres to a 'no-network' policy, uses only the Python standard library for its core functionality, and employs atomic file replacement to ensure data integrity during writes.

lov-lovpen-add-wechat-channel-comp