NemoClaw v0.0.132 is out!

[rsliter]

NemoClaw v0.0.132 is out

NemoClaw v0.0.132 strengthens sandbox recovery, qualified image handling, and the boundaries around destructive operations. Rootless Podman can now use exact-digest external agent images, while rebuild and cleanup paths make ownership checks and recovery actions more explicit.

Inference and onboarding also get more reliable. Newly created NVIDIA Endpoints sandboxes use a sandbox-attached OpenShell provider, eligible Station WSL hosts can select a managed llama.cpp recipe, and resumed or recreated sandboxes receive more precise validation before NemoClaw reports success.

This release updates operational guidance and qualification infrastructure. The documentation covers the new paths and their recovery steps. Release checks distinguish some candidate failures from missing or stale test prerequisites.

Sandbox Lifecycle and Recovery

• #12538 adds rootless Podman onboarding, resume, rebuild, and cleanup for exact-digest external OpenClaw and Hermes images. NemoClaw verifies image identity and leaves shared images it does not own in place.
• #12893 lets operators destroy a retained live sandbox after interrupted onboarding when its recorded OpenShell identity matches. Keep other clients from replacing that sandbox name during cleanup because OpenShell deletion is name-addressed.
• #12801 checks applicable host API keys before rebuild preparation and again before sandbox deletion. An invalid key stops the rebuild while preserving the existing sandbox.
• #12767 restores recovery of legacy sandboxes after a confirmed failed probe, and #12762 lets source-installed rebuilds use their installed catalog instead of an unpublished release version.
• #12769 checks provider and model changes before fresh onboarding reuses a Ready sandbox. Changed selections follow the recreation path; unchanged selections keep the sandbox.
• #12626 reuses a healthy retained gateway registration when it matches the declared endpoint and improves the rebuild recovery command after a failed replacement.
• #12774 identifies the native-state archive after uninstall and accepts a completed empty Docker registry after the final sandbox is destroyed.
• #12721 preserves ordinary Deep Agents Code interpreter checkpoints during rebuild after observability probes.
• #12608 protects images still used by sibling gateways during cleanup. #12835 avoids detaching a provider before a fresh sandbox exists.

Inference and GPU Onboarding

• #12562 moves new and recreated NVIDIA Endpoints sandboxes to a sandbox-attached OpenShell provider. #12826 and #12895 repair native requests and agent credential handling. Older sandboxes keep their shared route until recreated; Deep Agents Code requires recreation, while OpenClaw and Hermes can switch an existing native attachment.
• #12557 enables qualified Station GB300 WSL hosts to select the managed Qwen 3.6 35B-A3B llama.cpp recipe. #12609 preserves a completed Docker GPU proof for managed vLLM, and #12856 makes an unattempted container CUDA proof visible in readiness output.
• #12720 waits for custom-image OpenClaw startup before gateway configuration. Rebuild with the Other OpenAI-compatible endpoint provider verifies the restored inference route and reports the preserved backup when verification fails.
• #12681 restarts a stopped OpenRouter adapter during resumed Deep Agents Code onboarding and requires a bounded inference request before reporting success.
• #12793 preserves the OpenShell sandbox during CPU Deep Agents Code startup and restores native NVIDIA credential handling for onboarding, health checks, and rebuild preflight.
• #12764 preserves managed vLLM serving-profile provenance in session, registry, and status data, while avoiding a managed-install claim for an endpoint NemoClaw did not install. This carries forward @harjothkhara's contribution from #12346.
• #12755 checks the local inference route during GPU finalization, and #12372 retains a completed GPU proof across resumed sandbox creation.
• #12551 rejects unsupported Gemini configuration export instead of producing unusable output.

Networking and Messaging

• #12490 makes tunnels use the selected OpenShell gateway port. #12566 keeps a gateway-scoped tunnel available when one sandbox stops, is destroyed, or is recreated.
• #12763 keeps unprovable forwarded ports occupied instead of choosing a conflicting port. This carries forward @sandeepstele's contribution from #12352.
• #12841 preserves Hermes messaging configuration through rebuild.
• #12706 adds a reviewed offline cache for selected official messaging plugins during OpenClaw image builds.
• #12846 validates managed Slack installation and policy boundaries before remediation; #12662 reports remediation command failures.

Diagnostics and Operations

• #12759 identifies a failed custom endpoint without exposing credential-bearing URL components. It carries forward @sandeepstele's contribution from #12043.
• #12705 reports the gateway restart failure category during inference switching.
• #12796 gives recovery guidance for the operation that actually failed instead of GPU advice after a non-GPU startup failure. It carries forward @scs0209's contribution from #12098.
• #12758 retries an empty executable-readiness response within the existing bounded wait while keeping explicit failures terminal.

Documentation and Release Confidence

• #12735 updates the release history and owning guides for inference routing, rebuild recovery, retained cleanup, and uninstall archives.
• #12847 qualifies candidate E2E dependencies through trusted controllers, making prerequisite failures easier to distinguish from product failures.
• #12926 accepts digest-pinned image identity in concurrent gateway cleanup tests.
• #12790 removes a workflow-run pagination cap from E2E evidence discovery, and #12853 carries PR authorization through qualification and retries.

Thank You

Thank you to @harjothkhara, @sandeepstele, and @scs0209 for the original contributions carried into this release.