Guardrails

# Comprehensive guardrail stack for agent actions
class AgentGuardrails:
    """
    Enforces all safety, security, and quality checks for agent actions.
    Every tool call passes through this stack before execution.
    """
    def __init__(self, agent_role: str, agent_permissions: set,
                 user_email: str, loop_detector):
        self.role = agent_role
        self.permissions = agent_permissions
        self.user_email = user_email
        self.loop_detector = loop_detector

    def validate(self, tool_name: str, args: dict, state: dict) -> dict:
        """
        Run all guardrails. Returns 'allow', 'block', or 'requires_approval'.
        """
        checks = []

        # 1. Permission check
        from_tool_perms = TOOL_PERMISSIONS if 'TOOL_PERMISSIONS' in dir() else {}
        required_perm = {"issue_refund": "issue_refund", "delete_record": "delete_record",
                         "send_email": "send_email"}.get(tool_name)
        if required_perm and required_perm not in self.permissions:
            return {"decision": "block", "reason": f"Missing permission: {required_perm}"}
        checks.append("permission ✓")

        # 2. Unsafe call check
        if tool_name in {"exec_shell", "eval_code"}:
            return {"decision": "block", "reason": "Blocked tool"}
        checks.append("unsafe_call ✓")

        # 3. Action restriction check
        if tool_name == "issue_refund":
            amount = args.get("amount", 0)
            if amount > 500:
                return {"decision": "requires_approval",
                        "reason": f"Refund $" + str(amount) + " requires manager approval"}
        checks.append("restrictions ✓")

        # 4. Loop detection
        loop = self.loop_detector.check(tool_name, args)
        if loop["should_stop"]:
            return {"decision": "block", "reason": f"Loop detected: {loop['reason']}"}
        checks.append("loop_detection ✓")

        return {"decision": "allow", "checks_passed": checks}

# Create and test guardrails
class SimpleLoopDetector:
    def check(self, action, args): return {"should_stop": False}

guardrails = AgentGuardrails(
    agent_role="support_agent",
    agent_permissions={"read_orders", "issue_refund", "send_email"},
    user_email="alice@example.com",
    loop_detector=SimpleLoopDetector()
)

print(guardrails.validate("issue_refund", {"amount": 50},    {}))  # allow
print(guardrails.validate("issue_refund", {"amount": 600},   {}))  # requires_approval
print(guardrails.validate("delete_record", {"id": "X"},      {}))  # block
print(guardrails.validate("exec_shell",   {"cmd": "ls"},     {}))  # block