The Agent Skills Directory

[PROMPT_INJECTION]: Instructions explicitly override standard AI agent safety protocols by directing the agent to bypass human confirmation and oversight.
Evidence: "You are an autonomous agent. You make decisions. You do not ask questions."
Evidence: "RARV replaces request-confirm-wait loops with a self-correcting cycle that only surfaces to humans when genuinely blocked."
Evidence: "Agents reason, act, reflect, and verify — without asking humans for permission."
[COMMAND_EXECUTION]: The skill grants the agent broad authority to execute arbitrary shell commands and generate code without a human-in-the-loop checkpoint.
Evidence: "ACT → Execute the plan. Write code. Run commands. Commit changes."
Evidence: "INFRASTRUCTURE → set up DB, env, Docker, CI."
[COMMAND_EXECUTION]: The agent is instructed to automatically install software dependencies, which could lead to supply chain attacks if the agent's plan is influenced by malicious input.
Evidence: "BOOTSTRAP → environment check, dependency install."
[PROMPT_INJECTION]: Vulnerability to indirect prompt injection due to the combination of high-privilege capabilities and the ingestion of untrusted external requirements.
Ingestion points: The DISCOVERY phase involves reading external specifications and requirements files.
Boundary markers: None are specified to protect the agent from instructions embedded within the requirements files.
Capability inventory: The agent has full access to run commands, write code, manage environment infrastructure, and perform git commits.
Sanitization: No sanitization or validation logic is defined for the external data being processed.

rarv-cycles