Skip to content

feat(sandbox): automated image builds, in-place sandbox updates, and LLM response streaming on Azure #5680

Open

Description

@NVIDIA-AI-Labs

User Story

An NVIDIA AI lab partner uses OpenShell for running an agent application inside an OpenShell sandbox, and is validating that deployment on Azure Kubernetes Service (AKS). The partner would like OpenShell to provide automated sandbox image build/publish on Azure, a supported way to update sandbox code and dependencies without full image rebuilds, and end-to-end LLM response streaming for chat and voice workflows.

Problem Statement

Three gaps were reported:

  1. Sandbox image build automation on Azure — there is no automated path to build and publish sandbox images on Azure to produce repeatable, deployable images.
  2. Sandbox update workflow — there is no supported workflow for updating sandbox code, configuration, and dependencies; today changes appear to require repeatedly rebuilding the sandbox image.
  3. LLM response streaming — end-to-end streaming of LLM responses is not available for chat and voice reception workflows.

The streaming request is a capability gap and does not correspond to a reproduced OpenShell defect.

Impact / Why This Matters

The missing build automation, update workflow, and streaming support add integration and development effort while the partner validates its product on OpenShell running on AKS. Each sandbox code or dependency change currently costs a full image rebuild cycle, and the lack of streaming constrains interactive chat and voice reception use cases.

Proposed Design

Not provided in detail by the reporter. The requested capabilities are:

  • Automated sandbox image build and publishing on Azure, producing repeatable and directly deployable images.
  • A supported workflow to update sandbox code, configuration, and dependencies without requiring a full image rebuild for every change.
  • End-to-end LLM response streaming through OpenShell, usable by chat and voice reception workflows.

Acceptance Criteria

  • Sandbox images can be built and published on Azure through an automated, documented pipeline that yields repeatable, deployable images.
  • A documented, supported workflow exists for updating sandbox code, configuration, and dependencies without rebuilding the sandbox image for every change.
  • LLM responses can be streamed end to end through OpenShell and consumed by chat and voice reception workflows.
  • All three capabilities are verified for an OpenShell sandbox deployment on Azure Kubernetes Service (AKS).

Alternatives Considered

Not provided.

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