NemoClaw v0.0.94 is out!

[prekshivyas]

NemoClaw v0.0.94 strengthens the boundaries around sandbox recovery, network policy updates, and security evidence. Snapshot restores now preserve state that was not safely captured, cross-sandbox restores re-establish authenticated gateway access, and policy updates distinguish a matching live preset from one that genuinely needs to change.

Onboarding and inference workflows are more observable and more predictable. Automation can consume a versioned, redacted JSONL event stream, DGX Spark resume preserves managed vLLM intent, invalid platform/model combinations fail before downloads, and Hermes image assembly is substantially faster without changing its ownership or scanning guarantees.

This release also improves the signal behind those workflows. Long-running E2E scenarios are isolated, cancellation and runner-pressure evidence is clearer, privileged Hermes setup stays tied to exact-head workflow code, and documentation checks remain useful without mutating user state or blocking on an absent preview.

Sandbox, onboarding, inference, and policy

• #6820 adds --events=jsonl, a versioned and redacted stream of onboarding state-machine events, so automation can track progress without scraping terminal prose.
• #7185 selects an IPv4 fallback DNS resolver and probes that same address, so compatibility recovery does not validate one resolver and configure another.
• #7234 preserves DGX Spark managed-vLLM Express intent across onboard --resume, so recovery continues with the deployment path the user selected.
• #7284 keeps upgrade-sandboxes --check read-only and targets the gateway recorded by the sandbox, so an inspection cannot start or recover infrastructure as a side effect.
• #7322 tags plugin registration output with its gateway source, so users can tell which gateway produced each activation message.
• #7352 detects edited policy presets and previews their reapplication while documenting bounded tls: skip use, so policy changes are intentional and their security tradeoffs are visible.
• #7359 rejects managed-vLLM model overrides that do not support the detected platform before downloading artifacts, so an invalid configuration fails early and cheaply.
• #7371 verifies that session-export downloads created the expected host file before recording a bundle, so backup metadata cannot claim an artifact that is missing.
• #7437 re-establishes gateway pairing after a cross-sandbox snapshot restore and verifies it with an authenticated agent run, so the restored OpenClaw sandbox is usable rather than merely copied.

Security and release safeguards

• #7345 adds reviewed npm-audit exceptions with scanner and advisory provenance, so accepted findings remain traceable and stale exceptions cannot silently mask new results.
• #7354 adds read-only advisory correlation and audit provenance records, so maintainers can identify exact package-range matches before they become blocking audit evidence.
• #7356 redacts token-shaped values from both sandbox doctor --json output boundaries, so diagnostics retain their exit semantics without exposing credential-like data.
• #7444 ties privileged Hermes swap setup to trusted exact-head workflow code, so E2E preparation cannot execute candidate-controlled host mutations.

Hermes and sandbox performance

• #7436 clears only declared stale restore directories and repairs Hermes virtual-environment access, so restores preserve unrelated state and the sandbox user can run the restored environment.
• #7461 collapses Hermes repository payloads into five ownership-preserving BuildKit layers, cutting the first hosted production build comparison from 205 seconds to 101 seconds and export from 178.8 seconds to 70.4 seconds.

Documentation and platform guidance

• #7334 documents the qualified DGX Station OTA metadata fallback, so GB300 systems with the recognized shortened release marker can proceed without weakening hardware checks.
• #7340 repairs the absolute route to the CLI command reference, so readers reach the intended canonical documentation.
• #7385 keeps the documented agent-variant check read-only, so following validation guidance cannot change a sandbox.
• #7455 treats an absent Fern preview as non-blocking when no preview exists, so documentation validation reports the condition without failing an otherwise valid change.
• #7458 updates the reviewed Fern CLI dependency to 5.80.1, keeping the documentation toolchain current with its security review recorded.
• #7468 adds the canonical dated v0.0.94 changelog entry, so the tagged release and user-facing documentation share one reconciled record.

CI and E2E confidence

• #7277 makes gateway TCP-refusal coverage deterministic, so onboarding readiness tests fail on the intended condition instead of timing variance.
• #7448 isolates long-running E2E scenarios into independent lanes, so runner loss has a smaller blast radius and each adapter retains distinct evidence.
• #7449 parallelizes OpenClaw plugin EXDEV coverage, so the same cross-filesystem runtime contract completes with less wall-clock delay.
• #7450 stabilizes cancelled live-test child lifecycles, so cancellation is reported as cancellation rather than an unrelated cleanup failure.
• #7457 records periodic runner-pressure telemetry, so slow or failed E2E jobs can be separated from genuine product regressions.
• #7460 removes duplicate candidate-specific Hermes swap setup after the trusted exact-head lane took ownership, avoiding redundant privileged host preparation.

Thank you

Thank you to external contributors @HOYALIM for the machine-readable onboarding event stream in #6820, @RUSHcitizen for the CLI documentation route fix in #7340, @andrewwhitecdw for early platform validation of managed-vLLM model overrides in #7359, and @HwangJohn for the read-only agent-variant documentation check in #7385.

See the full comparison.