The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill fetches external content from a user-provided post_url via the WebFetch tool. This content is then processed by sub-agents (hook-strength-agent and eval-runner-agent) for analysis. While this ingestion of untrusted data represents an attack surface for indirect prompt injection, the sub-agents are specifically instructed to perform literal observations and adhere to a structured rubric, which serves as a mitigation against following malicious instructions embedded in the external content.
[COMMAND_EXECUTION]: The skill includes several Bun scripts (manifest-sync.ts, scaffold-eval-loop.ts, append-loop-result.ts, bootstrap-experience.ts) to manage workspace scaffolding, result logging, and metadata indexing. These scripts perform filesystem operations using the Bash and Write tools. Analysis of the script code reveals robust security checks, such as using realpathSync to prevent directory traversal and explicitly checking for (and refusing to follow) symlinks during file operations. This follows the principle of least privilege within the skill's local directory.
[DATA_EXFILTRATION]: No data exfiltration patterns were detected. The skill's network operations are restricted to retrieving post content from the specified URL for analysis. Access to sensitive data is not requested, and local file operations are confined to the skills-resources/loops/ and .agents/ directories.

short-form-eval