The Agent Skills Directory

[COMMAND_EXECUTION]: The skill makes extensive use of child_process.execSync across multiple components, including ship-change.js, hooks/pre_completion_hook.cjs, hooks/post_tool_edit_hook.cjs, and the PI extension TypeScript file. These scripts execute shell commands (e.g., pytest, npm test) often derived from external parameters or project discovery logic.
[PROMPT_INJECTION]: The hooks/session-start script implements a context injection mechanism that wraps the entire SKILL.md content in <EXTREMELY_IMPORTANT> tags. This is designed to force the loop framework's logic into the agent's primary context, which is a technique often associated with instruction override attempts.
[PROMPT_INJECTION]: An indirect prompt injection surface is present where untrusted data from user requirements could influence the shell commands executed by verification hooks. Ingestion points: User requirements in roles/grill-me.md and task descriptions in roles/implementer.md. Boundary markers: <EXTREMELY_IMPORTANT> tags used in session startup. Capability inventory: Broad shell execution via execSync in various utility and hook scripts. Sanitization: No explicit validation or escaping of the test_command parameter before execution.

agent-loop-engine