NemoClaw v0.0.123 is out!

[cv]

NemoClaw v0.0.123 expands configuration export for managed OpenClaw and Hermes sandboxes. Exports preserve more of the settings operators use, while verifying the running configuration and keeping credential values out of the document.

This release also improves recovery and day-to-day operation. Intentionally stopped sandboxes report their state without triggering recovery, Portable Hermes operations share consistent lifecycle locks, and installer recovery runs before generic host admission. Fresh Linux onboarding gains an experimental contract for registering one external host component.

Contributor tooling now avoids repeated local validation when the inputs have not changed. Documentation links stay within the selected agent's pages, and review tooling retains structured execution evidence for the commit under review. The complete range is v0.0.122…v0.0.123.

Configuration export

• Managed Hermes sandboxes can use the shared configuration export format through nemoclaw config export and nemohermes config export. This brings the existing read-only export workflow to the canonical Hermes hosted-inference profile. #11323
• Hermes export preserves verified Nous API-key authentication intent without exporting the key. OAuth authentication and Hermes tool gateways remain outside this export contract. #11450
• Hermes export retains nondefault dashboard, browser TUI, and allocated API-port settings so the document describes the selected interfaces. #11477
• OpenClaw export preserves nondefault managed proxy settings and verified Brave Search credential references, covering configurations that previously could not be exported faithfully. #11384, #11387
• OpenClaw export retains inference tuning, reasoning, execution timeouts, and heartbeat settings while omitting canonical defaults. #11383
• OpenClaw export preserves local conversation telemetry, direct tool disclosure, and dashboard port and bind settings. Enabling these supported settings no longer makes the export incomplete. #11429, #11430, #11431
• Local Ollama export preserves the distinct daemon and authentication-proxy ports after verifying the active route. Missing or changing evidence prevents publication. #11473
• OpenClaw export supports one read-only secondary agent sharing the primary hosted inference route, preserving its name and restricted tool selection. #11478
• The compiled CLI preserves the OpenShell SDK's import conditions. This repairs connection and policy operations that could fail with ERR_PACKAGE_PATH_NOT_EXPORTED even when the SDK was installed. #11472

Sandbox lifecycle and recovery

• Intentionally stopped sandboxes report Stopped after provider confirmation, without unnecessary recovery or inference probes. #11211 carries forward @HOYALIM's original contribution in #11091.
• Portable Hermes start, stop, connect, status, and recovery use one host-scoped lock policy. This removes conflicting lock order across commands, including configurations with nondefault gateway ports. #11336, #11482, #11490
• Hermes recovery waits for the authenticated health response owned by the sandbox. An unrelated listener returning HTTP 200 cannot end the wait early. #10998
• Hermes SSH version checks preserve the image's state-directory permissions. Rebuilds also retain the original Dockerfile identity when reusing a prepared image, avoiding an incomplete restore. #11297
• Verified OpenClaw layouts where the agent and sandbox run as the same user retain native private state permissions. Separate-user and unverified layouts keep the existing shared permissions. #11309
• Portable OpenClaw recovery and probe-only connect report timing receipts when existing readiness evidence avoids lifecycle work, making that reuse visible without repeating recovery. #10647

Installation and external components

• The installer attempts registered-sandbox recovery before generic host admission and preserves recorded GPU opt-outs. Incomplete recovery remains distinguishable from successful recovery; fresh onboarding and required DGX Station reconciliation still require admission. #10397
• Fresh Linux onboarding can register one experimental external host component with a NemoClaw-managed Docker-driver gateway. Activation requires verified sandbox identity and effective policy. NemoClaw does not manage the component service or credentials, and failed or ambiguous activation leaves explicit incomplete state. #11366, #11471
• Explicit providerless onboarding can use that external-component activation contract. It verifies activation before reporting completion and does not create an inference provider for the component. #11508
• Non-interactive onboarding resolves the sandbox name before provider selection, allowing managed runtime setup to receive the identity it requires. #11442
• Station Express recognizes both GB300 DGX Station and NVIDIA DGX Station GB300 firmware naming, while retaining the existing hardware and runtime checks. #11480
• Docker readiness diagnostics identify authority conflicts and preserve the underlying Docker error during GPU-mode probing, giving operators a more useful repair direction. #11393, #11280

Inference and messaging

• NVIDIA Endpoints onboarding validates the selected chat route. Matching model-access 404 responses produce an account-access explanation instead of an unexplained inference failure. #10957
• Hermes Tavily setup can select the compatible provider profile explicitly and preserves the credential reference issued by OpenShell. Its probe reports request status without returning credential values or upstream response bodies. #11419
• Teams removal reconciles its owned login bindings across the matching Outlook policy endpoints, avoiding partial removal while preserving bindings owned by another integration. #10846
• OpenClaw onboarding retains the Open tier's default messaging egress presets when no channels are enabled. Explicitly disabled channels keep their existing cleanup behavior. #11457

Contributor tooling and documentation

• Local hooks select affected repository checks and reuse successful compiler validation only while its inputs still match. Publication validation checks formatting without rewriting the committed files. #11491
• Independent CLI contract and MCP crash-consistency cases run concurrently within bounded limits, shortening local feedback while retaining their assertions. #11396, #11399, #11463
• Maintained JavaScript and TypeScript use consistent formatting and a shared lint configuration. Correctness errors, warnings, and unused suppressions are enforced during validation. #11402, #11403, #11420
• Local review works with macOS temporary-directory and Homebrew layouts, and owned review gateways use isolated provider state to avoid collisions across reviews. #11500, #11501
• Review queues can consume validated E2E recommendations and terminal execution results without treating missing or stale evidence as a passing empty selection. #11495
• Image verification can reuse a trusted canonical manual publication while retaining provenance checks and push-publication preference. #11289
• Hermes recovery and enterprise-readiness links now reach sections present in the selected agent's documentation. The release command reference also covers the expanded export and external-component behavior. #11182, #11189, #11447

Thank you

Thank you to @HOYALIM for the original stopped-sandbox behavior in #11091, carried forward in #11211, and @squarepots for trusted manual image-publication reuse in #11289.