The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill implements an 'Overlay' mechanism that dynamically loads and prioritizes instructions from a path specified by the $POWERSUIT_OVERLAY environment variable. This allows external files to override the core logic of the skill.
[INDIRECT_PROMPT_INJECTION]: The overlay system creates a surface for indirect prompt injection. The skill explicitly instructs the agent to read and follow a README.md and other files from an external directory ($POWERSUIT_OVERLAY/aieo/), stating 'Where the overlay and this file disagree, follow the overlay.' If these external files are compromised or manipulated, they can alter the agent's behavior and bypass the original skill constraints.
[COMMAND_EXECUTION]: The skill executes multiple shell commands to manage files and git state, including mkdir, cp, shasum/sha256sum, and git. While these are used for the stated purpose of the skill, they operate on user-provided inputs (like <slug>) which could be exploited for path traversal if not properly handled by the agent platform.
[COMMAND_EXECUTION]: The skill executes a local Python script (python3 .github/actions/gate/run.py) to verify the prediction hashes. Executing scripts from the repository assumes the repository environment is trusted.

prediction-hash