NemoClaw v0.0.126 is out!!

[sandl99]

NemoClaw v0.0.126 strengthens sandbox recovery and cleanup around OpenShell's lifecycle controls. Gateway and host-forward operations now share ownership checks, and destructive operations preserve local records when cleanup cannot be confirmed. Recovery also repairs the case where a container runs while OpenShell still reports its sandbox as stopped.

This release improves Hermes Portable and managed Podman onboarding, checks remote inference credentials before rebuilding, and redacts sensitive inventory values consistently. Policy updates have enough time to complete OpenShell's activation wait. Fresh packages also preserve executable permissions for the ACP and Blueprint Runner commands.

For contributors, builds use a consistent Node.js and npm toolchain, dependency installation handles reviewed SDK archives, and review automation reports unresolved blockers. A new review coordinator runs in read-only shadow mode. The documentation describes the release's recovery behavior and experimental profile limits.

Sandbox and gateway recovery

• Host forwards now share typed OpenShell controls across onboarding, recovery, snapshots, messaging, and cleanup. Stop, destroy, and rebuild retain registry records when forward release cannot be confirmed. #11594
• Gateway lifecycle consumers share typed operations for registration, reuse, and cleanup. Ambiguous results trigger fresh observation, reducing the risk of repeated mutations or unsafe removal fallback. #11741
• Recovery no longer replaces legacy keepalive containers. OpenShell owns the supported stop/start lifecycle, keeping recovery aligned with native agent readiness and persistent state. #11776
• Gateway startup offers a service stop command only when a stable active service owns the selected state and is the sole listener. This avoids directing recovery toward an inactive or unrelated service. #11725
• recover starts stopped Docker-driver sandboxes through OpenShell, and start repairs a running container whose sandbox remains in Stopped. Both paths can restore OpenShell's Ready phase. #11797
• Sandbox log retrieval and streaming use a typed OpenShell adapter. Command validation, bounded reads, and cancellation now share one boundary while preserving existing log options and output behavior. #11835
• uninstall --yes can remove a non-default-port state root left by failed preflight before sandbox or gateway creation. Cleanup rechecks ownership under a lock and preserves uncertain or foreign state. #11602

Inference, policy, and secret handling

• rebuild checks recorded remote inference credentials before backup and again before deletion. Expired, missing, or unverifiable credentials stop the operation; --force cannot bypass this protection. Refresh the credential before retrying. #11779
• nemoclaw list applies the status redaction boundary to text, JSON, and recovery metadata, including credentials embedded in URLs. This adapts @udsy19's original contribution in #10200. #11821
• Managed vLLM adoption queries published ports using the inspected container ID, so a container-name transfer cannot redirect the decision. This carries the remaining hardening from @HOYALIM's #11530. #11815
• Policy writes allow 65 seconds for OpenShell's 60-second activation wait. Larger updates, including a second managed MCP server, can complete before the host deadline; policy reads retain their 15-second deadline. #11863

Hermes, Podman, and packaging

• Fresh Hermes Portable onboarding verifies and recovers its gateway through the Podman container identified by its authority receipt. This prevents Docker discovery from substituting for the selected Portable runtime. #11784
• Managed Podman onboarding reads OpenShell's structured main-process specification and releases the stopped original container before its replacement binds the port. Cleanup after failed onboarding also respects the selected provider when durable provider state does not yet exist. #11816
• Hermes Portable rejects non-x86_64 Linux hosts before inspecting or changing host state. Use a Linux x86_64 host for this experimental profile. #11839
• Managed image and CI builds use Node.js 24.18.1 and integrity-verified npm 12.0.2, giving builds and audit receipts a consistent toolchain. The public Node.js compatibility floor remains 22.19. #11736
• Builds set every declared package executable to mode 0755. Fresh installations can invoke nemoclaw-acp and nemoclaw-blueprint-runner directly without a permission failure. #11840

Contributor checks and review automation

• PR CI filters reviewed OpenShell SDK package runs by the PR commit, so unrelated recent runs cannot hide the required archive. #11693
• Trusted E2E jobs install the lockfile-matched OpenShell SDK through the shared dependency installer. This handles npm 12's optional-package omission without passing package credentials into candidate execution. #11836, #11731
• E2E artifact restoration selects the action that implements the updated Node.js and npm contract, preventing a stale toolchain check from stopping jobs before tests start. #11811
• SDK test-fixture packaging accepts npm 11 and npm 12 metadata formats while preserving duplicate package names. The Advisor blocker job also uses the reviewed toolchain. #11804
• Pre-push TypeScript validation reuses file digests only while file identity remains unchanged. This reduces repeated hashing while retaining checks for files that change during validation. #11785
• PR Review Advisor has an aggregate blocker check. Unresolved P0/P1 findings, outstanding E2E recommendations, or incomplete evidence prevent the workflow from reporting readiness. #11787
• A review coordinator evaluates the current PR commit and repository gates in read-only shadow mode. It makes readiness decisions inspectable without posting reviews or merging changes. #11862
• Hermes image tests explicitly select root for scenarios that exercise root startup and configuration refresh. The tests now reach those paths when the image defaults to the sandbox user. #11878
• The Jaeger runtime test resolves its propagator from the reviewed diagnostics package. It can validate malformed headers and valid traces with that package's dependency layout. #11880

Documentation and contribution guidance

• The Experimental NemoCUA reference explains that host-local inference, including local Ollama, is unsupported. Choose a non-host-local provider for this experimental route. #11095
• Issue-planning guidance identifies the component that owns requested behavior before proposing local work. Plans prefer upstream capabilities and define removal conditions for temporary integrations. #11841
• Advisor guidance explains why manually requested static review can inspect fork PRs. It documents the read-only boundary so contributors retain that review coverage. #11824
• The release changelog and command reference describe the shipped recovery, credential, policy, and packaging changes, giving operators guidance that matches this release. #11788

Thank you

Thank you to @udsy19 for the inventory-redaction contribution in #10200, adapted in #11821, and @HOYALIM for the vLLM adoption work in #11530, whose remaining hardening ships in #11815.

Full changelog: v0.0.125 → v0.0.126