NVIDIA / NVIDIA/NemoClaw

[Kubernetes] Run one buildless OpenClaw lifecycle through an external gateway

Open
#9,873 0 comments 0 reactions 0 assignees View on GitHub
integration: openclaw platform: k8s
Dominant language
TypeScript
Stars
22.5k
Forks
3.1k
Avg merge
1d 1h
Merged PRs (30d)
715

Description

Parent: #9816

## Outcome

A non-root NemoClaw Blueprint Runner completes one CPU/cloud OpenClaw workflow through an externally managed OpenShell gateway using a released, digest-pinned managed image and no Docker or Podman operations.

## Scope

- Resolve and verify one stock managed OpenClaw image as an exact `repository@sha256:` reference.
- Create the sandbox, configure provider/policy/inference, complete one agent turn, inspect status/log or exec, and clean up.
- Apply and verify the managed image startup profile through an OpenShell-owned mechanism rather than Docker `inspect`, `cp`, rename, or image mutation.
- Record the exact gateway, workspace, image provenance, and remote resources created by the run.
- Reconcile ambiguous create, configure, and delete results before retry or cleanup.

## Acceptance

- [ ] The complete flow succeeds without a runtime socket, image build, privileged container, or host mount.
- [ ] Wrong image digest, platform, agent identity, release, provenance, or startup profile fails closed.
- [ ] A timeout after mutation triggers an exact workspace and run-identity lookup; name-only matches are not adopted and mutation is not blindly repeated.
- [ ] Cleanup removes only resources created by the run and never changes the external gateway, controller, chart release, shared image, or consumer-owned Secret.
- [ ] NVIDIA/OpenShell#1849 is reproduced or dispositioned. Prefer SDK/gRPC create plus standalone exec; otherwise declare and preflight the required SSH client.
- [ ] Credentials are absent from arguments, logs, plans, receipts, and retained artifacts.

## Dependencies

Reuse #9140 and the applicable typed OpenShell operation slices under #9802. Stop for an upstream or maintainer decision if the managed startup profile cannot be applied and verified without Docker.

## Out of scope

Hermes and other agents, GPU or local inference, arbitrary custom images, rebuild/clone, multi-tenancy, and production support.

Contributor guide

Open the contributing guide

Research direction

Start with dependency #9140 and the applicable typed OpenShell operation slices under #9802; reproduce or disposition NVIDIA/OpenShell#1849 and determine whether SDK/gRPC create with standalone exec is available. Done means one non-root, digest-pinned CPU/cloud OpenClaw lifecycle completes through the external gateway without runtime sockets or Docker/Podman operations, with exact identity-based cleanup and no credential leakage.

Written by the indexing model from the issue text.

Assessment

Tech stack
docker, grpc, kubernetes, typescript
Domain
ai-infra-agents, backend-api-design, cloud, infrastructure
Issue type
Feature
Difficulty
5/5
Estimated time
Over a week
Activity status
Active
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.