NemoClaw v0.0.88 is out

[ericksoa]

NemoClaw v0.0.88 focuses on making the boundaries around host preparation and sandbox operation more dependable. DGX Station setup now preserves safe stopped containers, offers precise OpenIB remediation, and can restore the packaged CDI refresh lifecycle without reaching beyond the accepted factory-image boundary.

The release also makes runtime health claims more honest and recovery more predictable. Authenticated model probes distinguish a reachable route from a healthy upstream model, sandbox commands consistently select the owning gateway, port-forward recovery verifies live ownership, and rebuilds can reuse the exact web-search credential already registered with the gateway.

Finally, v0.0.88 tightens the machinery that supports those user-facing changes. Onboarding policy defaults now follow the selected web-search provider, repository and ShellCheck validation have clearer trust boundaries, E2E authentication comes from an immutable action, and the release workflow carries unfinished work forward before retiring a completed target label.

DGX Station preparation

• #7152 adds RDMA-aware guidance when openibd.service fails, helping Station users distinguish an expected networking dependency from a generic host-preparation failure.
• #7155 preserves stopped containers whose restart policy is no while continuing to fail closed on active workloads, inventory drift, or containers that could restart automatically, so preparation is safer without needlessly rejecting an idle host.
• #7158 restores the packaged CDI refresh lifecycle for the exact NVIDIA AI Developer Tools Station profile when GPU CDI inventory is missing, recovering the supported configuration without generating CDI directly or restarting Docker or containerd.

Inference, onboarding, and rebuilds

• #7007 adds the serving-process health leg to sandbox status and doctor, preventing a fresh diagnostic exec from being presented as proof that the long-running agent process has equivalent inference access.
• #7074 replaces endpoint-only inference checks with bounded authenticated model-invocation probes for supported remote providers, so healthy means the configured model can actually answer rather than merely having a reachable route.
• #7129 lets rebuild preflight reuse the sandbox's exact gateway-registered Brave or Tavily credential binding, avoiding unnecessary key re-entry while preserving fail-closed behavior for missing or mismatched bindings.
• #7133 aligns fresh onboarding policy defaults with the selected web-search provider, avoiding Brave egress when search is disabled and selecting only the maintained preset that matches Brave or Tavily.

Sandbox and gateway recovery

• #7092 treats an already-active port forward as recovery success only after live inventory proves that the target sandbox owns the reachable listener, making repeated recovery safe without accepting an unrelated process.
• #7113 routes sandbox-scoped status and exec operations through the sandbox's owning OpenShell gateway, preventing another active gateway on the host from returning stale state or receiving the command.

Documentation and release operations

• #7118 teaches the release workflow to move remaining open work to the next patch target before deleting the released label, keeping release planning accurate after a tag is cut.
• #7134 documents the agentic documentation engineering workflow, giving contributors a repeatable path for planning, validating, and reviewing documentation changes.
• #7160 adds the canonical v0.0.88 changelog entry, preserving the user-facing release record in the tagged source tree alongside the product documentation it references.

CI and release confidence

• #7047 migrates the repository check suite to native TypeScript modules, reducing mixed-module complexity in the validation path contributors and CI rely on.
• #7057 enforces the trusted createRequire ratchet, preventing new unreviewed CommonJS escape hatches from weakening that module boundary.
• #7060 extracts ShellCheck SARIF conversion into a focused component, making shell-analysis reporting easier to test and maintain independently.
• #7079 moves Docker Hub authentication for E2E into an immutable pinned action, keeping credentials on a trusted workflow-controlled path instead of PR-modifiable code.

Thank you

Thank you to external contributors @agrovr for the ShellCheck SARIF refactor in #7060 and @DisturbedSage5840C for the authenticated model-invocation health probes in #7074.

See the complete comparison: v0.0.87...v0.0.88.

[convexchaos]

Is there a framework or thought process around when/how often to update the openclaw release that packages with nemoclaw?

[ericksoa]

Yes, generally, we are trying to get the most recent stable release, after we've had a chance to run it thoroughly against our regression suite and make sure we do not need further adjustments on our side. Major dependency updates tend to happen after a declared lkg so that any issues that emerge post testing have a few days to stabilize before the next lkg.