The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted metadata, such as book titles and descriptions, from external web novel platforms (e.g., Qidian, Fanqie, JJWXC) into the agent's context. This creates a surface for potential indirect prompt injection attacks where instructions could be hidden in book descriptions. * Ingestion points: Data extracted from web pages in the scripts/ directory. * Boundary markers: Absent; data is processed as raw text within Markdown reports. * Capability inventory: The skill possesses file-writing capabilities (fs.writeFileSync), network access for scraping, and browser control via agent-browser. * Sanitization: Content is truncated and basic whitespace cleaning is applied, but there is no specific sanitization to prevent the agent from interpreting instructions in the ingested text.
[DYNAMIC_EXECUTION]: The scraper scripts utilize a utility function evalJSONBase64 to execute JavaScript code within the browser context via the Chrome DevTools Protocol (CDP). This is a standard and necessary function for the skill's intended purpose of data extraction from complex web pages.
[COMMAND_EXECUTION]: The skill invokes the agent-browser executable using Node.js execFileSync to manage browser automation sessions. The implementation includes logic to safely resolve executable shims on Windows systems.

story-long-scan