The Agent Skills Directory

[PROMPT_INJECTION]: The skill processes untrusted data from external sources, which creates a vulnerability surface for indirect prompt injection. The agent is directed to read PR review comments and CI failure logs (referenced in SKILL.md). While the instructions include a weak cognitive boundary by advising the agent to validate bot findings and avoid mass-resolving threads, there are no technical delimiters or sanitization steps to prevent embedded instructions in these external sources from influencing the agent's behavior. The agent possesses capabilities to resolve merge conflicts, write code to fix CI failures, and perform git push operations.
[COMMAND_EXECUTION]: The skill implements a persistence mechanism through a recurring execution loop. It attempts to establish a background heartbeat or use a platform-provided loop primitive to run iterations every 5 minutes by default. This persistence is a core component of the skill's intended functionality. The risk is partially mitigated by specific safety instructions that forbid the agent from modifying workflow configuration files to bypass security checks.

babysit