NemoClaw v0.0.84 is out

[prekshivyas]

NemoClaw v0.0.84 strengthens the path from first setup to a healthy, recoverable sandbox. The release adds a one-confirmation DGX Station Nemotron Ultra setup, carefully bounded support for trusted private inference endpoints, and model-aware readiness checks that verify the configured local model without spending inference tokens.

The same theme carries through day-to-day operation. Resume checkpoints avoid replaying completed onboarding work, recovery messages retain the agent and sandbox choices needed to continue, per-gateway state prevents local environments from colliding, and snapshots no longer move OpenClaw's machine-local identity or pairing state between systems.

Underneath those user-facing improvements, v0.0.84 tightens runtime configuration, Shields recovery, Hermes dashboard and MCP convergence, and release validation. The release train now exercises candidate OpenShell runtimes, isolates destructive MCP lifecycles, reports mode-sharded evidence more clearly, and uses sandbox-reachable fixtures for the gateway upgrade and EXDEV paths.

Onboarding, inference, and platform

• #6883 adds a DGX Station express install for the pinned NVIDIA Nemotron 3 Ultra 550B managed-vLLM recipe. This makes the recommended Station path a single-confirmation workflow while preserving the existing DeepSeek option.
• #6985 lets automated vLLM setup continue after verified storage warnings while keeping interactive confirmation and fail-closed inconclusive checks. This separates advisory capacity guidance from conditions that genuinely make unattended setup unsafe.
• #6886 allows explicitly trusted private inference hostnames and IP literals while retaining DNS pinning and reserved-range protections. Teams can connect approved private endpoints without weakening the default SSRF boundary.
• #6873 makes Ollama onboarding prefer NEMOCLAW_MODEL and honor the compatibility provider-model variable. Environment-driven setup now selects the intended model consistently in both automated and interactive flows.
• #6767 carries Hermes' 64K Ollama context floor through daemon configuration, validation, and generated config. Hermes no longer silently falls below the context window expected by its runtime.
• #6863 makes status and doctor compare the configured Ollama or vLLM model with the provider inventory. Readiness is now model-specific without generating a completion or consuming tokens.
• #6862 treats unreachable and HTTP 5xx inference routes as not ready during onboarding finalization. The session remains resumable so operators can repair the route and finish instead of being left with a falsely healthy setup.

Recovery, sandbox state, and security boundaries

• #6934 checkpoints completed sandbox, search, messaging, and resource choices during onboarding. Resuming a failed session no longer repeats work that was already completed successfully.
• #6916 names NEMOCLAW_FRESH=1 directly in failed-session remediation. The recovery path is actionable without requiring users to infer the fresh-install switch.
• #6835 includes the sandbox name in printed rebuild-resume commands. Recovery now targets the environment that actually failed.
• #6964 prints complete agent-specific fresh-install and resume commands. Operators can copy the remediation directly without reconstructing omitted arguments.
• #6711 gives every non-default gateway port its own host state root. Multiple local gateways can coexist without sharing registry, onboarding, snapshot, migration, or inference-adapter state.
• #6882 excludes OpenClaw device identity and pairing-token directories from new snapshots and ignores them in older restores. Restoring application state can no longer overwrite the live machine-local authentication relationship.
• #6902 validates the complete OpenClaw configuration candidate before replacing the live file. Invalid, oversized, timed-out, or concurrently changed candidates leave the working configuration and gateway untouched.
• #6766 recovers Shields transition locks only when the recorded owner is definitively dead or its PID was reused. Ambiguous ownership continues to fail closed, preserving the security boundary while making stale-lock recovery practical.

Messaging and Hermes runtime stability

• #6887 adds live Telegram health states to channels status, including idle, unreachable, rejected-token, and not-started outcomes. Channel summaries now distinguish unchecked configuration from a working runtime connection.
• #6898 re-seeds the isolated Hermes dashboard configuration after an in-place inference change. Success is reported only after the dashboard converges on the new route.
• #6948 recognizes when a concurrent gateway apply already committed the intended Hermes MCP state. A successful race no longer triggers an unnecessary destructive rollback.

Docs and release surface

• #6894 moves the starter prompt to a single Markdown source. Human-facing and agent-facing onboarding guidance can no longer drift independently.
• #6900 migrates release history to the native Fern changelog. Each version now has a canonical dated entry with stable documentation routes.
• #6966 consolidates shared gateway and certificate-authority guidance. Troubleshooting paths now point to one maintained explanation instead of duplicating security-sensitive instructions.
• #6979 teaches documentation validation about native changelog routes. Release entries receive link checking without being misclassified as broken pages.
• #6988 adds the canonical v0.0.84 dated changelog entry. The shipped behavior and recovery guidance are captured at the tagged commit rather than reconstructed after release.

CI and release confidence

• #6672 adds a live inference-adapter fixture. Adapter behavior now has dedicated end-to-end evidence instead of relying only on adjacent workflows.
• #6738 validates candidate OpenShell runtimes ahead of adoption. Compatibility regressions can surface before a dependency promotion reaches users.
• #6872 tracks Ubuntu 26.04 onboarding preflight coverage. The next Ubuntu platform boundary is visible in CI rather than remaining an implicit compatibility assumption.
• #6892 isolates OpenClaw, Hermes, and DeepAgents MCP bridge lifecycles on fresh runners. Destructive agent setup can no longer contaminate another matrix variant's evidence.
• #6944 supports mode-sharded E2E evidence in the release scorecard. Multi-mode targets are represented individually instead of hiding gaps behind one aggregate result.
• #6978 links aggregate E2E failures directly to their job logs. Maintainers can move from the gate result to the failing evidence without manually searching the workflow run.
• #6989 makes inference-switch baseline fixtures reachable from the OpenShell gateway network namespace. The live checks now exercise the intended compatibility behavior instead of failing on runner loopback routing.
• #6994 aligns release fixtures with inventory-based readiness and current deployment messages. Release validation now asserts the supported contract rather than superseded completion behavior.
• #6999 gives the EXDEV and historical gateway-upgrade fixtures real, authenticated, sandbox-reachable endpoints and preserves their request logs. The release-critical lifecycle and upgrade probes now fail on product behavior, not unreachable test infrastructure.

Thank you

Thank you to external contributors @HwangJohn for #6766, #6767, #6872, and #6873; @HOYALIM for #6738, #6863, and #6886; @Hokonoken for #6835; @harjothkhara for #6882; @deepujain for #6672; and @atulya-singh for #6941.