NemoClaw v0.0.131 is out!

[rsliter]

NemoClaw v0.0.131 is out

NemoClaw v0.0.131 strengthens the boundaries around sandbox state, networking, and lifecycle recovery. Rebuilds now preserve the complete native agent home and workspace, while stricter pod and process identity checks prevent repair or cleanup from acting on the wrong resource.

Inference and onboarding are more reliable across local and hosted providers. This release improves OpenRouter verification, llama.cpp reachability checks, GPT-6 probes, managed vLLM behavior, and container discovery during cloud onboarding.

The release also expands portable configuration, managed web search, skill routing, diagnostics, and packaging confidence. Documentation validation now checks every published internal link, and the managed runtime dependency set includes the reviewed security update for proxy-addr.

State, sandbox, and network safety

• #12340 transfers the complete native agent home and workspace during rebuilds. It validates credentials, inspection results, and integrity before replacing the sandbox, which reduces the risk of partial state recovery.
• #12556 requires an exact sandbox pod name or a valid generated suffix during legacy DNS repair. This prevents a sandbox such as box1 from selecting an unrelated pod such as box10-xyz12. The change preserves the original contribution from @udsy19 in #11672.
• #12491 verifies a recorded cloudflared process before tunnel startup or cleanup acts on it. An inconclusive check now preserves the process and PID record for recovery.
• #12560 uses typed OpenShell observations in Docker GPU diagnostics. Diagnostic collection remains best effort and cannot hide the original failure or cleanup guidance.
• #12535 accepts an externally supervised gateway's declared mTLS state directory after ownership and permission validation. Operators no longer need a NemoClaw ownership marker for this supported lifecycle.

Inference and onboarding

• #12516 gives the managed N1x Qwen vLLM profile one corrective retry after a compaction summary fails its quality audit. Repeated rejection or generation failure still preserves the existing conversation history.
• #12633 changes the Experimental 64 GB DGX Spark Qwen recipe to lazy safetensors loading. This reduces startup memory pressure without changing the model, context, concurrency, or memory limits.
• #11709 checks sandbox reachability before accepting an operator-attached llama.cpp service. A confirmed TCP failure now stops onboarding with binding and firewall guidance.
• #12625 and #12685 verify OpenRouter readiness and model changes through the target sandbox. A failed switch restores the previous route when possible.
• #12691 uses max_completion_tokens for GPT-6 compatibility probes, including gpt-6-astra.
• #11700 accepts recognized underscore spellings such as ollama_local, nvidia_prod, and open_router, then normalizes them to canonical provider IDs.
• #12606 waits briefly for Docker to expose a newly created managed-cloud sandbox. Ownership ambiguity, identity changes, stopped containers, and inspection mismatches still fail immediately.
• #11874 checks the portable managed registry before building a sandbox image. An unreachable registry now fails early with a bounded diagnostic.

Configuration and agent workflows

• #12451 exports native NVIDIA provider bindings from the current workspace or an unscoped binding. When retained corporate CA state cannot be exported, the CLI now reports the destination trust requirement on standard error.
• #12225 and #12624 preinstall the verified Tavily plugin in managed OpenClaw images, keep it disabled by default, and preserve admitted OpenClaw and Hermes Tavily sources during configuration export.
• #12607 routes skill install, list, and remove commands using the target sandbox's registered agent. Onboarding another sandbox no longer changes routing for an existing OpenClaw sandbox.
• #12635 labels the Deep Agents Code observability-otlp-local policy preset with its agent provenance. This corrects displayed ownership without changing permissions.
• #11774 preserves remote hermes-acp exit status, including status 255, and reports SSH transport or temporary-config cleanup failures separately.
• #12000 shows usage and valid actions when a user enters an unknown global subcommand. Unknown input values remain excluded from public error text.

Packaging, documentation, and release confidence

• #12239 removes the older bundled npm tree before Brev CPU bootstrap extracts the pinned Node archive. This prevents stale npm files from remaining beside the new installation.
• #12656 updates managed runtime dependency graphs and the official Slack bundle to proxy-addr 2.0.8. It also validates internal links across every published documentation page.
• #12701 and #12702 align the reviewed MCP SDK 1.31.0 archive, checksums, and offline recovery fixtures. Managed-image builds now use the archive selected by the lockfile on both supported architectures.
• #12689 isolates portable uninstall tests from package-manager state on the developer host. This keeps lifecycle evidence tied to the scenario under test and preserves the original contribution from @HOYALIM in #11952.
• #11958 updates CodeQL Actions to 4.38.0, including native Linux Arm64 bundle support and newer default analysis tooling.

Thank you

Thank you to external contributors @udsy19 and @HOYALIM. Their original work in #11672 and #11952 shipped through the maintained replays in #12556 and #12689.