Security model

Layered controls: permissions, HITL approvals, sandboxing, and installation trust checks.

Vitruvian applies a deny-by-default design across four layers.

1) Permission model

Modules declare required access via [RequiresPermission] and runtime enforces it through PermissionChecker.

ModuleAccess flags:

• Read
• Write
• Execute

Permission resolution follows user → group → other context.

2) HITL approval gate

IApprovalGate mediates side-effecting operations.

• write, delete, network, and execute paths can require approval
• timeout behavior is deny by default
• decisions are recorded as approval audit records

3) Sandboxed module execution

SandboxedModuleRunner and ISandboxPolicy provide limits for:

• CPU time
• memory
• wall-clock runtime
• file system access
• network access
• process spawning

4) Installation controls

Before install, Vitruvian enforces:

• plugin manifest (Vitruvian-manifest.json)
• signing checks (unless explicitly overridden for local development)
• required secret prompts

Security operator checklist

• Set durable memory storage for audit (VITRUVIAN_MEMORY_CONNECTION_STRING).
• Review plugin manifests before install.
• Keep approval timeouts and prompts enabled in production.
• Run doctor regularly and track findings in CI.