NemoClaw v0.0.82 is an operator-control and boundary-hardening release for long-running sandboxes. New non-destructive stop and start commands free container resources without discarding workspace state, backup and snapshot flows handle stopped and cloned sandboxes more safely, and managed vLLM now checks image and model storage before large downloads begin.

Onboarding now preserves intent and tells the truth more often at failure boundaries: gateway port conflicts fail fast, Ollama status and compatible-endpoint guidance reflect what is actually usable, loopback probes bypass host proxies, and resumed setup retains its selected route. The same release rejects catch-all custom-policy destinations, keeps messaging build plans out of final image environments, and locks the managed WeChat dependency graph.

Underneath those user-facing changes, v0.0.82 strengthens the evidence and supply-chain paths used to ship NemoClaw. Node and advisor dependencies are aligned and locked, Ubuntu, WSL, and macOS contract lanes are more reliable, exact-diff E2E gates now require reviewed authorization and close stale outcomes cleanly, and release promotion dispatches immutable semver data to the Brev image pipeline.

Sandbox lifecycle, backup, and onboarding

• #6721 runs managed Deep Agents Code headless sessions through the bounded supervisor and reaps completed DCode and LangGraph descendants, preventing process leaks and keeping sandboxes reusable.
• #6723 temporarily starts eligible stopped Docker sandboxes for backup and returns them to the stopped state, so strict upgrades preserve recoverable state instead of skipping it.
• #6742 resolves a complete secret-free sandbox-create intent before destructive or provider mutations and rebinds credentials immediately before effects, reducing partial-state drift during onboarding and resume.
• #6748 adds non-destructive nemoclaw <name> stop and nemoclaw <name> start commands that preserve workspace and registration state while repairing runtime health, gateways, and forwards on restart.
• #6749 allocates snapshot clones their own dashboard ports, preventing clone URLs and later rebuilds from colliding with the source sandbox.
• #6753 probes local inference daemons independently of route preflight and distinguishes stopped from running Ollama, making the interactive inference menu truthful and actionable.
• #6759 detects a foreign process on the OpenShell gateway port before reuse probes, replacing onboarding hangs with immediate process details and recovery guidance.
• #6764 stops treating route-only reservations from interrupted onboarding as backupable sandboxes, preventing placeholders from blocking installer upgrades while keeping real sandboxes fail closed.
• #6768 repairs or reserves the selected inference route when resumed onboarding skips provider setup, preventing the later sandbox-profile step from losing its route.
• #6769 explains when a Chat-Completions-only agent needs an OpenAI-compatible endpoint or OpenClaw, turning an otherwise opaque endpoint mismatch into a concrete recovery choice.
• #6776 bypasses host proxies for bounded loopback readiness probes, preventing healthy local Ollama services from being reported unavailable.

Inference, messaging, and policy security

• #6730 lets the immediate host smoke check consume a narrowly matched one-shot Chat Completions capability receipt, shortening onboarding without caching credentials or weakening endpoint and model checks.
• #6731 applies one semantic-validation pipeline to repository and runtime custom policies, closing the gap that allowed catch-all egress destinations through runtime policy application.
• #6737 keeps the messaging build plan available during image construction but removes it from final OpenClaw and Hermes environments, reducing persisted configuration exposure while preserving the validated runtime plan.
• #6739 locks the managed WeChat plugin and its transitive graph behind a NemoClaw-owned offline cache with integrity verification, making that runtime supply chain reproducible and reviewable.
• #6771 checks Docker image storage and the Hugging Face cache before managed vLLM downloads and fails closed on ambiguous capacity unless explicitly accepted, avoiding unsafe large pulls that can exhaust the host.

Packaging, dependencies, and platform confidence

• #6729 raises the supported Node.js minimum to 22.19 and aligns OpenClaw and advisor SDK requirements while removing obsolete SDK advisories, giving hosts and contributors one clear compatible runtime contract.
• #6733 reorders image-build layers so plugin dependencies stay cached across source and blueprint rebuilds, shortening rebuilds without weakening ownership or policy checks.
• #6735 adds a digest-pinned Ubuntu 26.04 userspace contract lane for the CLI, preflight, installer, and platform logic, providing early compatibility evidence without claiming host Docker, AppArmor, Landlock, or live-E2E support.
• #6741 consolidates reviewed npm archive and lockfile audits, emits raw evidence, and enforces the configured severity threshold, strengthening production-dependency provenance.
• #6744 moves OpenShell checksum manifests, consumed assets, and archive-safety checks into a base-trusted verifier, making future release-pin updates reviewed data instead of mutable installer logic.
• #6822 runs WSL platform tests as a non-root user and isolates root- or host-dependent contracts, restoring meaningful WSL validation instead of permission-driven noise.
• #6825 provisions required macOS and WSL tools and handles real Darwin and WSL timing and path differences, restoring reliable cross-platform test signal.

Skills, documentation, and release operations

• #6727 treats a successful empty release-label lookup as no match, allowing post-tag carry-forward to create a missing next-version label instead of aborting.
• #6740 adds a dependency-upgrade skill requiring an adjacent-release ledger, upstream source and tests, downstream-contract analysis, and provenance evidence, making upgrades semantic migrations rather than version-only bumps.
• #6774 makes lkg promotion dispatch the downstream Brev production-image workflow with the immutable semver tag and record non-secret correlation data, enabling deterministic image builds without relying on a mutable tag.
• #6775 routes independent integrations, images, and recipes to Community Solutions and adds a product-scope gate, clearly separating community work from supported canonical behavior.
• #6777 adds a documentation-refactor skill that preserves canonical ownership, routes, anchors, and redirects while splitting large sections, making reorganizations consistent and maintainable.
• #6782 gives the PR Review Advisor correctness stage purpose-driven adversarial checks for fail-closed behavior, locality, ordering, and capacity guarantees, making bypasses and unsafe assumptions more systematically visible.
• #6789 refactors sandbox documentation into focused lifecycle, messaging, MCP, plugin, backup, and recovery pages with journey-based navigation and redirects, making operator tasks easier to find without breaking routes.
• #6824 publishes the canonical v0.0.82 documentation and updates affected guides for stopped backups, snapshot ports, Ollama, and custom policies, keeping detailed user guidance aligned with shipped behavior.

CI, advisor, and release confidence

• #6745 combines E2E Advisor into PR Review Advisor under the trusted read-only workflow, giving maintainers one coherent review and E2E-guidance conversation with safer fork and internal-PR handling.
• #6786 installs advisor dependencies from the repository lockfile through the trusted dependency graph, making the secret-bearing job reproducible and fail closed.
• #6788 binds the E2E gate to the exact PR, head, and base identity, preventing stale events or base retargets from reusing or overwriting another diff's verdict.
• #6807 reports both advisor model lanes without giving Nemotron authority over GPT and closes only seeded stale checks, keeping review and health state accurate.
• #6810 replaces a SHA-copying no-secret exception with protected-environment approval on the exact gate run, providing auditable maintainer authorization without weakening identity or risk checks.
• #6821 separates controller health from the exact-diff merge verdict and adds diagnostic and exception links, so ordinary test failures no longer masquerade as control-plane outages.
• #6826 automatically dispatches selected exact-SHA E2E jobs for controller-only internal PRs and requires authorization for PR-controlled execution or evidence changes, ensuring reviewed control-plane changes are validated instead of waived.

Thank you

Thank you to external contributors @DisturbedSage5840C for final-image state-directory metadata coverage in #6717, @Hokonoken for making strict backups handle eligible stopped sandboxes safely in #6723, @HOYALIM for improvements across inference, validation, onboarding, image performance, messaging, dependency security, platform coverage, E2E budgets, guardrails, and Hermes metadata in #6730, #6731, #6732, #6733, #6735, #6737, #6739, #6741, #6742, #6776, #6780, and #6781, and @HwangJohn for synchronizing the full-E2E budget documentation in #6762.

Full comparison: v0.0.81...v0.0.82