NemoClaw v0.0.85 is out!

[cjagwani]

NemoClaw v0.0.85 strengthens the boundary between the agent experience and the infrastructure that keeps it dependable. This release adopts OpenShell v0.0.85 with reviewed immutable artifacts, tighter credential and archive handling, and safer gateway recovery, so upgrades preserve both the intended runtime and the security properties around it.

Day-to-day workflows are steadier too. Custom endpoint onboarding can recover from DNS failures, inference route changes verify the sandbox before mutation, managed vLLM routes validate the model they actually serve, MCP setup fails earlier on reserved credentials, and rebuild protects the original sandbox when state backup is incomplete.

The release also lays careful groundwork for DGX Station setup, improves Hermes topology attestation and DCode lifecycle handling, and sharpens the live E2E machinery used to qualify releases. Together, these changes make failures more specific, recovery more deliberate, and release evidence easier to trust.

OpenShell, sandbox, and recovery

• #6726 upgrades the supported OpenShell integration to v0.0.85, preserves multiline command arguments, validates archives before extraction, blocks supervisor TLS credentials from child processes, and adds specific credential-rewrite diagnostics. This makes the platform upgrade both reproducible and easier to recover when a security boundary cannot be established.
• #7069 pins the reviewed OpenShell v0.0.85 release manifests after publication. Immutable identities prevent later upstream changes from silently altering a released NemoClaw installation.
• #7028 recovers upgrades when OpenShell is installed in a user-local path. Existing installations no longer fail merely because the binary is outside a system-wide location.
• #7078 retires an out-of-range running gateway only after a strict backup, while preserving gateways inside the supported OpenShell range and failing closed on invalid compatibility data. This keeps historical upgrades from presenting an old gateway schema to the new CLI without sacrificing recoverability.
• #7005 treats a rebuild backup as failed when every state directory fails, even if loose files were captured. The original sandbox is therefore retained unless the operator explicitly accepts the force-recovery path.
• #7021 recognizes dashboardPort: 0 as an intentional no-dashboard state. Status and lifecycle operations now preserve a valid headless configuration instead of misclassifying it.
• #7029 narrows gateway discovery's zombie-process exception to descriptor-pinned, single-thread zombies and remains fail-closed for ambiguous processes. This avoids stale processes blocking normal operation without weakening process identity checks.

Onboarding, MCP, and inference

• #6865 routes custom-endpoint DNS failures into the transport recovery flow with retry, back, and exit choices. Users can correct connectivity problems without silently losing their place in onboarding.
• #6986 clarifies that the managed MCP HTTPS requirement applies to every supported agent, while the OpenShell v0.0.85 integration rejects reserved revisioned credential names before provider mutation. This makes the shared security contract explicit and keeps partial MCP configuration from being written.
• #6992 serves managed Nemotron 3 Nano with its required reasoning parser, and #7001 removes an unsupported top-level thinking field from the Nemotron Ultra Build route. Both changes align generated requests with the model-specific contracts the endpoints actually accept.
• #7006 reads and validates sandbox state before mutating the live inference route. A failed preflight now leaves the existing working route untouched.
• #7044 adopts a vLLM served-model alias only when /v1/models reports the registered model as its exact root. This distinguishes a safe alias from a genuine model mismatch.
• #7010 separates DGX Station notice acceptance from non-interactive mode, and #7045 requires an interactive terminal for --station-deepseek. Express setup can no longer silently imply consent or continue through a prompt it cannot safely present.

DGX Station, Hermes, and agent runtimes

• #6991 adds guarded prerequisite preparation for qualified DGX Station GB300 hosts, including exact-version reuse, Docker rollback, CDI and GPU validation, and reboot resume. It reduces manual setup while keeping unsupported version drift visible; DGX Station remains deferred pending physical end-to-end validation.
• #7049 attests the private mutable Hermes runtime topology before shields transitions. Security-sensitive operations now verify the topology they are about to change rather than trusting stale state.
• #7032 closes an inherited DCode route-probe descriptor, and #7036 covers the bare default-connect path. Together they reduce lifecycle leaks and protect the simplest DCode connection contract.
• #6968 shares trusted sandbox base-image resolution primitives across CI consumers. A single validated resolution path reduces disagreement about which runtime image is being tested.

Skills, docs, and maintainability

• #6990 improves the starter prompt for guided installation. New users and their agents get clearer sequencing and fewer assumptions during setup.
• #6942, #6950, and #6967 move diagnostic, E2E scorecard, chat-send, and device self-approval tools onto explicit .mts boundaries. The migrations make module loading predictable and reduce CommonJS/ESM ambiguity in trusted tooling.
• #6993 extracts codebase-growth guardrails into trusted tools, and #7024 makes maintainer skills use the latest check-run attempts. Both changes make automated maintenance decisions reflect the code and CI state that actually shipped.
• #7076 adds the canonical v0.0.85 changelog entry and links each release theme to its focused user guide. The tagged source and public Announcement now share a durable, navigable record.

CI and release confidence

• #6981 accelerates validation and stabilizes live E2E execution. Faster, more deterministic feedback shortens the path from a candidate change to trustworthy release evidence.
• #6995, #6996, #6998, and #7000 move polling, live-Vitest invocation, result reporting, and scorecard orchestration behind validated trusted helpers. This reduces workflow-shell complexity and gives the E2E gate one coherent control path.
• #7025, #7027, #7042, #7046, #7054, and #7065 repair migrated entrypoint selection, shared-helper loading, history availability, typed-target selection, check reopening, and running-check serialization. These fixes ensure PR E2E reports the current attempt and does not confuse an in-progress check with a completed result.
• #7035, #7038, #7040, and #7041 add stronger evidence for GPU recovery, private Anthropic onboarding, native GPU sandbox topology, and the vLLM Docker storage socket alias. The release matrix now checks more of the real platform boundaries users depend on.
• #7055 aligns historical gateway-upgrade evidence, #7077 repairs deterministic release-blocker regressions, and #7094 stabilizes gateway and TUI fixture lifecycles. These changes make the qualification suite prove the intended upgrade and sandbox identity instead of fixture accidents.

Full comparison

Review the complete release range in v0.0.84...v0.0.85.

Thank you

Thank you to external contributors @HwangJohn for custom-endpoint DNS recovery in #6865, @HOYALIM for shared sandbox base-image resolution in #6968, and @atulya-singh for the diagnostics tool module migration in #6942.