guardian-angel

What This Skill Does

Guardian Angel is an ethics-centered cognitive framework for OpenClaw AI agents, built upon the foundation of Thomistic virtue ethics. Unlike traditional rule-based filters that rely on fragile blacklists, Guardian Angel cultivates a stable internal disposition—a 'moral conscience'—within the agent. At its core is caritas, the act of willing the good of the principal, which informs the cardinal virtues of prudence, justice, fortitude, and temperance.

v3.1 introduces a robust plugin enforcement layer. This includes 'before_tool_call' hooks that act as safety gates, preventing harmful actions before they occur. It utilizes nonce-based approval workflows for sensitive tasks, ensuring that any action with high stakes or ambiguity requires human-in-the-loop verification. By shifting focus from reactive rule-checking to proactive virtue-based reasoning, the agent gains a consistent, reliable decision-making structure that remains effective even when faced with novel edge cases or sophisticated prompt injection attempts.

Installation

To install the Guardian Angel skill, use the OpenClaw hub command:

clawhub install openclaw/skills/skills/leo3linbeck/guardian-angel

Ensure that your OpenClaw environment is updated to the latest minor version to support the v3.1 plugin enforcement architecture and infrastructure-disabling protections.

Use Cases

• Sensitive Data Handling: Prevents unauthorized data exfiltration or reckless file modification by prioritizing the user's long-term security over speed.
• Autonomous Infrastructure Management: Automatically blocks commands that attempt to disable the agent's own oversight mechanisms, such as unauthorized skill removal or unauthorized reconfiguration.
• High-Stakes Decision Support: Acts as a 'second opinion' for financial or legal tasks, requiring human authorization for any action that meets the threshold for critical risk.

Example Prompts

• 'Guardian Angel, before I execute this script that wipes the temp folder, perform a moral risk assessment based on my data retention policy.'
• 'I need to perform an emergency system update; please run a before_tool_call audit to ensure no critical security hooks are being bypassed.'
• 'Analyze my proposed strategy for the client negotiation; does this align with the virtue of justice, or am I prioritizing my own gain over their flourishing?'

Tips & Limitations

• Tips: Use the verbose logging mode to review the 'thought process' of the agent during virtue-based evaluations. This can help you understand why certain requests are flagged as ambiguous.
• Limitations: Guardian Angel does not replace human responsibility. While it significantly elevates the agent's safety, it is a tool for alignment and should be treated as a supportive layer rather than an infallible moral authority. High-stakes infrastructure actions will always trigger an escalation regardless of the agent's internal assessment.