atlas/home-voice-agent/safety/boundaries/README.md

Boundary Enforcement

Enforces strict separation between work and family agents to ensure privacy and safety.

Features

• Path Whitelisting: Restricts file system access to allowed directories
• Tool Access Control: Limits which tools each agent can use
• Network Separation: Controls network access
• Config Validation: Ensures config files don't mix work/family data

Usage

from safety.boundaries.policy import get_enforcer

enforcer = get_enforcer()

# Check path access
allowed, reason = enforcer.check_path_access(
    agent_type="family",
    path=Path("/home/beast/Code/atlas/home-voice-agent/data/tasks/home")
)
if not allowed:
    raise PermissionError(reason)

# Check tool access
allowed, reason = enforcer.check_tool_access(
    agent_type="family",
    tool_name="add_task"
)
if not allowed:
    raise PermissionError(reason)

# Check network access
allowed, reason = enforcer.check_network_access(
    agent_type="family",
    target="10.0.30.63"
)
if not allowed:
    raise PermissionError(reason)

Policies

Family Agent Policy

Allowed Paths:

• data/tasks/home/ - Home task Kanban
• data/notes/home/ - Family notes
• data/conversations.db - Conversation history
• data/timers.db - Timers and reminders

Forbidden Paths:

• Work repositories
• Work-specific data directories

Allowed Tools:

• All home management tools (time, weather, timers, tasks, notes)
• No work-specific tools

Network Access:

• Localhost only (by default)
• Can be configured for specific networks

Work Agent Policy

Allowed Paths:

• All family paths (read-only)
• Work-specific data directories

Forbidden Paths:

• Family notes (should not modify)

Network Access:

• Broader access including GPU VM

Integration

In MCP Tools

Tools should check boundaries before executing:

from safety.boundaries.policy import get_enforcer

enforcer = get_enforcer()

def execute(self, agent_type: str, **kwargs):
    # Check tool access
    allowed, reason = enforcer.check_tool_access(agent_type, self.name)
    if not allowed:
        raise PermissionError(reason)
    
    # Check path access if applicable
    if "path" in kwargs:
        allowed, reason = enforcer.check_path_access(agent_type, Path(kwargs["path"]))
        if not allowed:
            raise PermissionError(reason)
    
    # Execute tool...

In Router

The router can enforce network boundaries:

from safety.boundaries.policy import get_enforcer

enforcer = get_enforcer()

# Before routing, check network access
allowed, reason = enforcer.check_network_access(agent_type, target_url)

Static Policy Checks

For CI/CD, create a script that validates:

• Config files don't mix work/family paths
• Code doesn't grant cross-access
• Path whitelists are properly enforced

Future Enhancements

• Container/namespace isolation
• Firewall rule generation
• Runtime monitoring and alerting
• Audit logging for boundary violations