NemoClaw v0.0.76 is out

[apurvvkumaria]

NemoClaw v0.0.76 strengthens the boundaries around shared OpenShell infrastructure. Shared gateways now refuse incompatible inference-route mutations, legacy sandbox upgrades gain a guarded recovery path, and startup can repair one narrowly defined OpenClaw configuration drift state while ambiguous states still fail closed.

The release also makes managed agent workflows more predictable. LangChain Deep Agents Code gains opt-in OTLP tracing and a stable Nemotron 3 Ultra default, local and custom inference paths receive earlier compatibility checks, and everyday operations improve across MCP diagnostics, messaging setup, pairing, sessions, garbage collection, and non-interactive sandbox execution.

The full v0.0.75...v0.0.76 comparison contains 43 merged changes. Alongside the runtime work, this release expands the Deep Agents documentation surface, sharpens troubleshooting guidance, and makes local and live-E2E validation faster and more consistent.

Deep Agents Code and observability

• #6414 makes nvidia/nemotron-3-ultra-550b-a55b the managed NVIDIA Endpoints default for Deep Agents Code and prevents dependency or model pickers from overriding NemoClaw's selection. This gives new managed sandboxes a predictable model and startup path.
• #6418 preserves the optional first-run name prompt while continuing to suppress dependency and model selection. Users keep first-run personalization without losing deterministic managed configuration.
• #6416 moves the integration to the released deepagents-code[nvidia]==0.1.34 Nemotron Ultra profile and removes the temporary local profile. This uses the upstream packaged contract while preserving managed aliases, model IDs, and observability behavior.
• #6340 adds opt-in, backend-neutral OTLP/HTTP tracing for Deep Agents Code through a fixed host-local receiver. Operators can inspect bounded model, tool, subagent, and LangGraph activity without placing remote telemetry credentials inside the sandbox or making agent execution depend on delivery.
• #6344 adds Deep Agents Code as a first-class documentation variant with dedicated navigation, quickstart routing, ecosystem coverage, and generated command references. The third supported agent surface is now discoverable and checked for route consistency.

Inference and onboarding safety

• #6338 detects and serializes provider, model, endpoint, API-family, metadata, and binding conflicts between sandboxes on the same gateway. Onboarding, inference set, and connect-time repair can no longer silently replace another registered sandbox's route.
• #6378 lets inference set accept onboarding provider aliases, guides safe same-provider model switches, and gives Deep Agents users an actionable re-onboarding path. This removes common command dead ends while retaining DNS-pinning and endpoint-authorization safeguards.
• #6297 validates streaming SSE sequences from custom Anthropic-compatible endpoints during onboarding. Malformed endpoints now fail before sandbox creation with actionable diagnostics instead of surfacing as cryptic runtime errors.
• #6326 enables automatic tool choice with the qwen3_coder parser for the generic-Linux Nemotron 3 Nano vLLM default. Agents using the managed default can call tools without the previous HTTP 400 failure.
• #5868 warns Linux arm64 DGX Spark and DGX Station users when a selected Local NIM image may lack an arm64 manifest. Likely image incompatibility is visible before a confusing pull or startup failure, without blocking images that do support the platform.
• #6362 adds guarded recovery for pre-fingerprint legacy sandboxes, requiring complete backups, exact-name managed-image confirmation, delete-edge revalidation, and authoritative route restoration. Existing installs can recover without weakening fail-closed rebuild and custom-image protections.
• #6008 points interrupted resumable onboarding at nemoclaw onboard --resume. Users get a direct continuation path instead of assuming they must restart or reinstall.

Sandbox recovery and day-two commands

• #6307 lets startup repair the exact conservative signature of a root-owned mutable OpenClaw configuration while linked, mounted, unsafe, or ambiguous trees remain blocked. Normal sandbox writes recover without weakening shields or transaction-recovery boundaries.
• #5388 closes non-terminal sandbox-exec stdin by default and adds --stdin for intentional pipes. Automated commands no longer hang on an input stream that never reaches EOF, while scripted input remains explicit and available.
• #6306 makes nemoclaw gc scan both gateway-built and locally prebuilt sandbox image repositories. Cleanup can reclaim orphaned images and disk space across Linux and macOS development paths.

Messaging, MCP, and agent runtime

• #6382 adds a redacted wire-level credential-resolution probe to mcp add and single-server mcp status. Operators can distinguish verified replacement, failure, and inconclusive results when metadata appears healthy but OpenShell has not resolved a credential placeholder.
• #6396 gives the in-sandbox WhatsApp pairing QR code a four-module quiet zone through a hash-gated runtime preload. Terminal pairing becomes reliably scannable without writing credential-derived image data to logs.
• #6390 adds a bounded 16 MiB buffer for OpenClaw messaging package metadata and archive commands. Verbose Teams-plugin metadata no longer triggers ENOBUFS, while archive validation remains intact.
• #6007 removes the unavailable qqbot plugin from newly generated default OpenClaw configuration. First launch no longer reports a misleading missing-plugin warning.
• #6249 routes session passthrough through the sandbox's own agent binary. Hermes session listing now invokes Hermes instead of failing because an OpenClaw executable is absent, while OpenClaw behavior is preserved.
• #6303 uses Hermes' native non-interactive resume path for continued one-shot turns. New prompts append to the selected session instead of silently creating fragmented sessions.
• #6330 bootstraps the persisted same-device OpenClaw CLI pairing request when device listing is itself pairing-gated. This breaks the initial pairing deadlock while granting only the narrowly required pairing permission.

Documentation and troubleshooting

• #5539 documents direct DNS limitations inside Docker-driver GPU sandboxes and the safe diagnostic path. Users can separate resolver limitations from healthy policy-mediated inference, messaging, and search traffic without resorting to unsafe host-network workarounds.
• #6385 fixes the published route to the OpenClaw plugin-installation guide and adds route validation to strict documentation checks. The setup link works again, and future source-path drift is caught automatically.
• #6425 aligns the release highlights, arm64 Local NIM guidance, expanded image-cleanup scope, and agent-specific generated command headings with the shipped behavior. Operators and coding agents now receive the same current guidance across documentation variants.

CI and release confidence

• #6357, #6358, #6359, #6361, #6363, #6365, and #6368 consolidate live-E2E results, opt-in gating, redacted evidence, state snapshots, command environments, Docker prerequisites, and fake-provider protocols. Release validation now produces more consistent failures and artifacts with less duplicated harness behavior.
• #6420 adds bounded four-worker scheduling for local integration tests while keeping safety-sensitive routes serial and isolating onboarding fixtures from host-global ports. Clean local validation fell from roughly fourteen minutes to under four without weakening CI, coverage, or focused-test determinism.

Thank you

Thank you to external contributors @HwangJohn for #5388 and #5539, @deepujain for #5868, @abhi-0906 for #6007 and #6008, and @atulya-singh for the JSON-boundary regression coverage in #5598.