The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes content from live, uncurated websites such as Reddit and GitHub, which introduces a risk of the agent following malicious instructions embedded in the external content. \n
Ingestion points: External data enters the agent's context through tools like snapshot(), ref(), and screenshotWithLabels() while navigating live websites (SKILL.md). \n
Boundary markers: There are no specific delimiters or instructions provided to isolate or ignore potentially adversarial instructions within the retrieved web data. \n
Capability inventory: The agent is instructed to modify local source files in src/ and execute shell commands for building and testing, providing a significant capability tier that could be targeted (SKILL.md). \n
Sanitization: The workflow does not mention sanitizing, escaping, or validating content retrieved from the external sites. \n- [COMMAND_EXECUTION]: The skill utilizes local shell commands for build management, deployment, and testing, including pnpm runtime:prepare, pnpm test, and browser-control relay restart. It also includes a security-positive step of auditing for duplicate dependencies using pnpm audit:duplicates (SKILL.md).

gauntlet