The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/gossamer.py dynamically locates and executes Python scripts from sibling skill directories (e.g., qwencloud-update-check/scripts/check_update.py). The paths are computed at runtime by searching for specific directory markers, and the scripts are executed via subprocess.run.
[PROMPT_INJECTION]: The SKILL.md file contains mandatory instructions designed to override the agent's response behavior. It uses strong imperatives such as 'MANDATORY', 'You MUST follow the instructions', and 'Do NOT silently skip this' to compel the agent to perform update checks and present specific installation options to the user before completing their original request.
[INDIRECT_PROMPT_INJECTION]: The scripts/video.py script accepts user-provided prompts and interpolates them into API requests for video generation. The skill lacks boundary markers or sanitization for this untrusted input, creating an attack surface where instructions embedded in user data could influence the agent or the downstream video model.
[COMMAND_EXECUTION]: The skill instructions and maintenance scripts utilize subprocess.run to execute Python scripts. Furthermore, it directs the agent to run npx skills add to download and install additional skill components from a remote GitHub repository at runtime.

qwencloud-video-generation