policy-engine

What This Skill Does

The Policy Engine for OpenClaw is a production-hardened governance layer designed to bring deterministic control to autonomous agent operations. It functions as a middleware hook that intercepts all tool invocations before they execute. By enforcing granular control through tool allowlists, path restrictions, and risk-based escalation policies, it transforms unpredictable LLM behavior into a structured, audit-ready workflow. It is built to prevent common security risks such as prompt-injected path traversals, unintended shell execution, and unauthorized file system modifications, all while maintaining high performance and system stability.

Installation

To install, run the following command in your terminal:

clawhub install openclaw/skills/skills/joetomasone/policy-engine

After installation, activate it by updating your openclaw.json configuration file:

{
  "plugins": { "policy-engine": { "enabled": true } }
}

Use Cases

• Sandboxed Agent Environments: Define restrictive tool profiles for sub-agents (e.g., a research agent restricted to read-only browser tools).
• Regulatory Compliance: Audit every agent decision automatically to meet logging requirements for sensitive enterprise environments.
• Risk Mitigation: Use T1 and T2 risk tiers to gate sensitive system-level commands, preventing agents from executing risky shell commands without prior escalation or oversight.
• Path Guardrails: Protect critical directories (e.g., /etc, .ssh) from unauthorized write access by enforcing strictly defined workspace path allowlists.

Example Prompts

1. "Agent, please scan the current directory for todo.md files and summarize the remaining tasks, ensuring you only use read-only tools."
2. "Update the project configuration file in the safe workspace path with the new API endpoint."
3. "Attempt a system diagnostics check; if the policy engine blocks this due to risk tiering, inform me with the specific reason for denial."

Tips & Limitations

• Dry-Run Mode: Utilize "dryRun": true during initial setup to observe how the engine handles your current agent workflows without interrupting actual operations.
• Fail-Open Design: Note that the engine is designed to fail-open. While this ensures system availability, audit logs remain the primary source for identifying missed blocks.
• Canonicalization: Always use absolute paths for your pathAllowlists to avoid ambiguity in resolution.
• Break-Glass: Remember that the OPENCLAW_POLICY_BYPASS environment variable is strictly for emergencies; avoid its use in automated production pipelines to maintain audit integrity.