Skip to content

feat(sandbox): support running the supervisor in full macOS virtual machines #5681

Open

Description

@NVIDIA-AI-Labs

User Story

An NVIDIA AI lab partner uses OpenShell for sandboxing agent workloads, but their platform runs long-lived GUI agents inside full macOS virtual machines on Mac mini hardware rather than in containers. They need a documented and validated integration path for OpenShell in that environment, including how the OpenShell supervisor executes inside a macOS VM and how control is shared with their own agent control plane.

Problem Statement

OpenShell's integration story today assumes a container-based runtime. The partner's isolation requirements for long-lived GUI agents cannot be met by a container-only approach, so the workloads run in full macOS VMs. It is currently unresolved:

  • how the OpenShell supervisor would run inside a full macOS VM,
  • how control would be shared between the OpenShell controller and the partner's existing agent control plane (i.e. who owns the controller role),
  • what telemetry coverage is available in that deployment shape,
  • how TLS is terminated and trusted between the components.

Without answers to these, the proposed integration has no validated deployment path.

Impact / Why This Matters

Full-VM macOS environments are the only option for workloads whose isolation requirements exceed what containers provide, and GUI agent workloads of this kind are long-lived rather than short-lived task containers. A specific runtime and control-plane compatibility assessment would let OpenShell define a supported deployment design for VM-based hosts instead of leaving integrators to improvise, and is a prerequisite before any proof of concept can be scoped.

Proposed Design

No concrete design has been proposed yet. The requested first step is a runtime and control-plane compatibility assessment covering:

  1. Supervisor execution in a full macOS VM: confirm whether the OpenShell supervisor can run natively inside the guest, what its host/guest dependencies are, and what the lifecycle looks like for long-lived (non-ephemeral) agent sessions.
  2. Controller ownership: define whether the OpenShell controller or an external agent control plane owns scheduling and session control, and how the two interoperate when both are present.
  3. Telemetry coverage: identify which OpenShell telemetry/audit signals are available when the sandbox boundary is a VM rather than a container, and what gaps exist.
  4. TLS handling: define certificate issuance, trust and termination between supervisor, controller and the external control plane in a VM deployment.

Acceptance Criteria

  • A single representative GUI agent workflow is selected and used as the reference case for the assessment.
  • Supervisor execution inside a full macOS VM is validated and the result (supported / unsupported / requires changes) is documented.
  • Controller ownership in a split control-plane deployment is defined and documented, including interaction with an external agent control plane.
  • Telemetry coverage for a VM-based deployment is documented, including any gaps relative to the container-based deployment.
  • TLS trust and termination for the VM deployment are specified.
  • A supported deployment path for full macOS VMs on Apple hardware is documented, or the limitation is explicitly recorded if it is not supportable.

Alternatives Considered

  • Container-only deployment (the current supported approach): rejected by the partner because it does not meet the isolation requirements of long-lived GUI agents.

Reported to NVIDIA by an AI lab partner.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    state:triage-neededOpened without agent diagnostics and needs triage

    Type

    No type

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions