The Agent Skills Directory

[DYNAMIC_EXECUTION]: The script scripts/inspect-threejs-canvas.mjs includes a loadDependency function that uses dynamic import() to load Node.js dependencies (@playwright/test and pngjs). The resolution path for these imports is computed at runtime using process.cwd(), allowing the skill to execute code from the user's project directory, which could be exploited in a compromised environment.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from external URLs via browser automation. It captures console output and page errors which are then included in a technical report that influences the agent's release decisions.
Ingestion points: Data is ingested from the URL target in the inspectPage function within scripts/inspect-threejs-canvas.mjs.
Boundary markers: No explicit boundary markers or safety instructions are provided to the agent to treat external log content as data rather than instructions.
Capability inventory: The skill possesses capabilities for shell command execution, browser automation, and local file system interaction.
Sanitization: Log and error content is captured directly without filtering or sanitization.
[COMMAND_EXECUTION]: The skill requires running shell commands such as npm install, npx playwright, and a custom Node.js script. These commands interact with the local operating system to manage dependencies and execute testing logic.
[EXTERNAL_DOWNLOADS]: The documentation references using npx playwright install chromium to fetch browser binaries. While the source is a well-known service (Microsoft/Playwright), it involves the acquisition of external executable code.

threejs-qa-release