The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Documentation within the skill (e.g., in B1_变量更新规则.md and mvu_zod_cn.js) encourages the retrieval of JavaScript bundles from external domains, specifically cdn.jsdelivr.net and testingcf.jsdelivr.net. These resources originate from repositories belonging to unverified GitHub users ("MagicalAstrogy", "StageDog") rather than trusted organizations.
[REMOTE_CODE_EXECUTION]: The guides provided with the skill instruct the agent or user to utilize the await import() function to dynamically load and execute the aforementioned remote JavaScript files at runtime. This practice bypasses local code auditing and creates a significant execution risk if the remote repositories are compromised.
[COMMAND_EXECUTION]: The skill defines several operational procedures that require the agent to execute local Node.js scripts using shell commands. These include query-library.mjs for routing task logic, manage-guidance-preference.mjs for configuring user experience levels, and snapshot-library.mjs for library maintenance.
[PERSISTENCE]: The manage-guidance-preference.mjs script is designed to modify the agent's permanent environment by writing to rule files such as CLAUDE.md or AGENTS.md. While intended to store user-selected "guidance levels" (e.g., 新人, 入门), this mechanism provides a path to persistently alter the agent's core instructions and constraints.
[INDIRECT_PROMPT_INJECTION]: The skill processes a large library of over 200 files, including markdown, HTML, and JSON. Since the agent uses the content of these files to guide its actions and provide engineering advice, it presents an expansive surface area for indirect injection if malicious instructions are embedded within the documentation or templates.

consult-tavernweave-library