The Agent Skills Directory

[SAFE]: The skill's primary function is to enhance the security and predictability of autonomous agent operations by enforcing resource constraints and termination conditions.
[PROMPT_INJECTION]: The skill utilizes strong imperative language to govern agent behavior regarding resource management. These instructions are functional directives for logic control and do not attempt to bypass core AI safety filters, identity constraints, or exfiltrate data.
[DATA_EXPOSURE]: No hardcoded credentials, sensitive file paths, or unauthorized network operations were detected. The tools requested (Bash, Read, AskUserQuestion) are consistent with the skill's administrative purpose.
[INDIRECT_PROMPT_INJECTION]: The skill processes information about other agent tasks (runs) to apply safety bounds, which constitutes a potential attack surface for indirect injection.
Ingestion points: The $run argument in SKILL.md identifies the task context to be monitored.
Boundary markers: The skill explicitly requires the declaration of five stop conditions and mandates the use of AskUserQuestion for escalation, creating clear human-in-the-loop boundaries.
Capability inventory: The skill uses Bash and Read for monitoring and AskUserQuestion for reporting/escalation.
Sanitization: The skill effectively acts as a sanitization layer itself by requiring explicit confirmation for any destructive actions identified during the monitored run.

budget-guard