The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content from third-party websites during its audit process, which constitutes an attack surface for indirect prompt injection.
Ingestion points: The observe-runtime.mjs script (via Playwright page.goto and page.evaluate) and pages.mjs (via crawling) ingest HTML, network traffic, and storage data from untrusted external URLs.
Boundary markers: The instructions do not define specific delimiters or warnings to wrap the scraped content when it is returned to the agent's context.
Capability inventory: The skill possesses the ability to read local files (e.g., steps.json), write files (screenshots), perform network operations via browser automation, and execute shell commands via distributed Node.js scripts.
Sanitization: The observe-runtime.mjs script includes logic to redact potential tokens or secrets (based on length and character patterns) from browser storage dumps before they are reported.
[DYNAMIC_EXECUTION]: The library scripts/lib/browser.mjs uses dynamic import() and module.createRequire to resolve and load the Playwright automation library from the local environment at runtime.
[COMMAND_EXECUTION]: The skill instructs the agent to execute local Node.js scripts (scripts/observe-runtime.mjs) to drive a headless browser, perform network requests, and manage local filesystem operations like screenshot generation.

failure-resilience