NemoClaw v0.0.128 is out! 🎉

[rsliter]

NemoClaw v0.0.128 is out

NemoClaw v0.0.128 moves more sandbox creation and recovery behavior behind OpenShell's typed lifecycle boundary. Configuration export now produces the supported v1alpha1 shape for single-agent sandboxes, while recovery paths preserve enough identity and diagnostics for operators to act safely when creation, forwarding, or upgrade work is interrupted.

This release also refreshes the managed OpenClaw and Hermes runtimes. It improves inference readiness, MCP credential handling, onboarding validation, and Deep Agents Code compatibility without weakening the sandbox controls that protect agent workloads.

The release adds stronger evidence around configuration export, managed images, cleanup, secret isolation, and live runtime checks. The command reference and release documentation now describe the new status, export, recovery, and troubleshooting behavior.

Configuration export and status

• #11986, #12057, #12013, #12168, and #12164 make nemoclaw config export emit the supported nemoclaw.nvidia.com/v1alpha1 shape for hosted OpenClaw, Hermes, and Deep Agents sandboxes. The export now preserves supported corporate CA state, uses the singular agent field, and leaves managed image selection to the importer.
• #11919 and #11918 preserve supported read-only secondary-agent configurations that share managed vLLM or attached Ollama. Unsupported rosters still fail before publication instead of losing agents.
• #11953 adds automatic E2E export evidence. Release testing can now compare exported configuration with independently observed sandbox state and retain bounded diagnostics when export is refused.
• #12233 separates each sandbox's configured inference provider and model from the shared gateway's live route. Operators can now see route drift, and JSON consumers receive a redacted configuredInference object without a schema-version-1 break. This replay preserves work originally contributed by @Sehastrajit-S in #11962 and @sylvesterkaczmarek in #11437.
• #11770 and #12094 move inference route reads to the typed adapter and remove legacy helpers. Status and provider operations now use one route-observation boundary.
• #12052 validates OpenShell sandbox names before submission, and #12095 documents the limit. Users now receive an early error instead of a later OpenShell rejection.

Sandbox lifecycle and recovery

• #11909, #12121, and #12236 route standard Docker, native Podman, OpenClaw Portable, and Hermes Portable creation through the typed OpenShell lifecycle adapter. This removes duplicate command construction while preserving existing receipts and recovery behavior.
• #12083 and #12082 reconcile retained identities before publication and preserve route observation bounds. Ambiguous creation outcomes can now be checked without an unsafe duplicate mutation.
• #12055 reserves routes before resumed sandbox recreation, and #10498 restores Portable provider state. Interrupted onboarding is less likely to publish conflicting or incomplete state.
• #11927 preserves N1x recreation source proof, while #11924 preserves managed vLLM profile provenance. Recovery can verify what it is recreating instead of inferring identity from mutable host state.
• #12090 retains bounded forward-startup diagnostics, #11956 restores a missing named gateway registration during upgrade, and #12173 closes gateway recovery review gaps. These changes give operators clearer recovery evidence without relaxing lifecycle checks.
• #12061 distinguishes an unverified replacement version from a confirmed mismatch after rebuild failure. It also prints the preserved backup and recovery commands so operators can choose the correct next action.
• #12015 supports protected Docker 27 receipt transfer and adds a force-fresh macOS reset that retains model caches. #10733 measures host reuse from the end of state collection, which avoids expiring a valid reuse window during slow probes.
• #11875 rejects leading-zero port overrides, and #11887 checks Portable Ollama publication ports before mutation. Invalid port input now fails before onboarding changes state.
• #12234 preserves corporate CA validation errors through managed image lookup and reports a path-free correction. Users now see the invalid bundle reason instead of an unrelated image-catalog failure.
• #12074 reports the live process that owns an onboarding state lock. Operators can identify real contention without guessing which process to stop.
• #10489 validates the selected Teams webhook port before channel mutation or rebuild deletion. A bad override can no longer remove working state before the replacement is known to be usable.

OpenClaw, Hermes, and Deep Agents Code

• #11105 updates the managed OpenClaw runtime to 2026.9.1. #12178 then makes device approval exit after local fallback, and #12018 preserves N1x compaction liveness under the refreshed runtime.
• #11817 updates Hermes to 0.21.3. #12047 restores native startup health checks, #12101 recovers restart after credential changes, and #12102 preserves forwards created during fresh onboarding.
• #11972 lets Deep Agents Code complete its first interactive model turn while OpenShell continues to deny memfd_create. #12092 verifies the image dependencies required for that runtime before publication.

MCP and inference

• #11917 rejects endpoint aliases that request conflicting credential providers before policy or provider state changes. #12175 runs credential probes through the adapter executable boundary, and #12091 requires Hermes MCP reload health before accepting an expected relay disconnect.
• #12056 refreshes the managed runtime digest used by MCP validation. Tests now exercise the runtime image that the managed workflow actually publishes.
• #12007 uses the documented reasoning_effort: "none" control for Nemotron 3 Super validation. This keeps readiness checks aligned with the supported NVIDIA Endpoints request contract.
• #11955 honors the configured llama.cpp readiness budget, and #12096 waits for each Model Router semantic health response within the existing startup budget. Slow but healthy local services no longer fail because an inner probe used a shorter timeout.
• #11617 admits qualifying Windows WSL N1x hosts to installed Ollama model ranking when the chassis exposes an OEM model string. Supported systems no longer depend on one firmware reporting convention.

CI, E2E, and release confidence

• #11970 preserves destroy failures and separate cleanup artifacts. Live E2E can now show both the original failure and final cleanup result instead of overwriting evidence.
• #12087 builds the shared managed-image boundary before direct image contracts, and #12222 installs the reviewed OpenShell SDK before MCP bridge execution. Candidate tests now receive the trusted artifacts they depend on.
• #12150 preserves required Advisor evidence paths, #12188 lets Advisor select Model Router E2E, and #12190 accepts clean stale-base approvals. These changes keep automated review decisions tied to reviewable evidence.
• #12177 rebuilds incompatible Deep Agents Code PR bases, #12196 bounds the Bedrock secret-isolation audit, and #12214 waits for descendant process reaping. The affected lanes now fail on product behavior instead of setup drift or cleanup races.
• #12232 aligns hosted-turn and pairing fixtures with OpenClaw 2026.9.1. The live checks now use the managed gateway topology and canonical package root.
• #11881 adds the v0.0.128 changelog and refreshes the generated command reference. The documented release surface matches the shipped CLI behavior.

Thank you

Thank you to external contributors @ALDRIN121 for the lock-owner diagnostics in #12074 and @Adr1an04 for the rebuild recovery guidance in #12061.

Thank you also to @Sehastrajit-S and @sylvesterkaczmarek. Their original work in #11962 and #11437 was preserved in the configured-versus-live inference status change shipped through #12233.

[sylvesterkaczmarek]

Thanks for the mention and for preserving the contribution through #12233. Glad to see the configured-versus-live inference route separation ship in v0.0.128.

[Sehastrajit-S]

Great to see NemoClaw v0.0.128 out! Thanks for the mention and for carrying forward my contribution through #11962!