NemoClaw v0.0.86 is out!

[ericksoa]

NemoClaw v0.0.86 brings the reviewed DGX Station GB300 Express path to stock DGX OS. The installer now recognizes qualified DGX OS 7.2, 7.4, and 7.5 systems, validates the factory GPU and container stack in place, and grants direct-GPU sandboxes only the narrow sysfs access CUDA needs. A clean physical DGX OS 7.5 validation completed with local Nemotron 3 Ultra inference, sandbox CUDA, and real Hermes TUI tool use while unrelated sysfs paths remained denied. DGX Station support remains Deferred while broader clean-host qualification continues.

This release also makes the first-run experience more recoverable. Station Express remembers its validated recipe across an interrupted managed-vLLM step, model-cache checks ignore unrelated model artifacts, Gemini manual model IDs are checked against Google's native catalog, and restrictive host permissions no longer produce unreadable scripts inside a sandbox build.

Behind those user-facing changes, v0.0.86 strengthens the boundaries around destructive upgrades, platform-specific coding-agent guidance, and the E2E release plan itself. The result is a smaller patch release aimed squarely at making installation and recovery more predictable on the systems people are using today.

DGX Station and local inference

• #7108 enables Station Express on qualified stock DGX OS GB300 systems. It fails closed on unsafe or unknown identities, preserves the factory driver and Docker configuration, and limits sandbox sysfs access to the exact GPU and supporting paths required for CUDA. This matters because an out-of-the-box Station can now reach local Nemotron 3 Ultra and Hermes without trading away the host or sandbox security boundary.
• #7059 persists the secret-free Station Express provider, model, sandbox, and interaction choices before managed-vLLM setup can fail. nemoclaw onboard --resume returns to the same validated recipe instead of asking the operator to reconstruct it.
• #7099 scopes Hugging Face cache writability checks to the selected managed-vLLM model. Root-owned artifacts from an unrelated model no longer block a model switch, and recovery guidance identifies the exact path that needs attention.

Hosted inference

• #6980 validates manual Google Gemini model IDs against Google's native catalog while retaining the OpenAI-compatible runtime probe. IDs with or without the models/ prefix now resolve consistently, and catalog pagination and credential redaction remain bounded and tested.

Sandbox builds and upgrade safety

• #7072 normalizes staged script permissions before Docker consumes either sandbox build context. Fresh installs under a restrictive umask no longer carry root-only modes into non-root image-build stages.
• #7093 warns before a legacy sandbox recreate that managed recovery restores only the recorded .openclaw state. The warning gives operators time to back up paths such as /sandbox/user-data before the destructive step.

Platform guidance

• #7061 moves DGX Spark, DGX Station, and Windows WSL Express instructions into revision-pinned assets loaded only after platform detection. General coding-agent guidance stays focused while every platform-specific model, policy, storage, approval, and validation safeguard remains explicit.

CI and release confidence

• #6970 moves selector, inference-mode, schema, and Hermes-selection validation into the typed E2E planner before GitHub Actions emits its matrix and summary. Unsafe test paths, duplicate IDs, and incompatible selector combinations now fail before execution begins.

Thank you

Thank you to external contributors @harjothkhara for the E2E planner refactor in #6970 and @HwangJohn for Gemini native-catalog validation in #6980.

Compare the complete release: v0.0.85...v0.0.86.