Nemoclaw v0.0.114 is out!

[jyaunches]

NemoClaw v0.0.114 expands deterministic automation while making routine sandbox operations easier to diagnose. LangChain Deep Agents Code can now invoke one approved read-only MCP tool without model participation, and launch-readiness probes report timing for each stage.

This release also strengthens the boundaries around destructive operations, configuration recovery, installation, and updates. Users receive clearer warnings and recovery instructions, while NemoClaw verifies more state before it reports success.

Local inference, Portable onboarding, Hermes, and messaging workflows receive focused reliability improvements. The release also improves documentation automation and keeps agent-specific guidance aligned with supported behavior.

Automation and Readiness Diagnostics

• #9894 adds dcode tools call-read-only TOOL --json for deterministic, bounded invocation of one approved read-only MCP tool. Automation can now require a tool call without relying on model behavior.
• #9865 adds credential-free elapsed time for each connect --probe-only readiness stage. Operators can identify where launch readiness is delayed without changing the command result.
• #9964 adds a fixed-stage timing summary for Portable lifecycle recovery. Operators can distinguish container, OpenShell, local inference, startup, and gateway delays.
• #10031 derives loaded Deep Agents Code MCP tools from the executable catalog and rejects duplicate implementations. Tools discovered after startup remain searchable until the runtime loads them.

Sandbox Safety and Recovery

• #9858 warns before sandbox destruction terminates active SSH sessions, including unattended --yes and --force runs. Users can account for connected work before destruction proceeds.
• #9782 reports a retained in-sandbox session export artifact when cleanup fails and gives its removal command. This helps users remove an artifact that can contain pasted secrets.
• #9787 reports the actual condition that makes a host-mount directory unusable, while #9853 names the device-review path for a pending OpenClaw scope upgrade. Both changes make recovery actions specific.
• #9231 preserves sealed OpenClaw configuration during snapshot restore. #9791 verifies final configuration integrity after rebuild, and #9866 retains timer authority until lifecycle gates finish.
• #10053 binds legacy GPU sandbox replacement handoff to the transaction-owned container instead of stale name metadata. Healthy replacements no longer wait through the full handoff deadline when legacy metadata differs.

Installation and Onboarding

• #9776 stops installation when neither sha256sum nor shasum can verify the nvm installer. A required Node.js installation or upgrade no longer continues without verification.
• #9822 retries throttled Hermes source archive downloads. Hermes installation can recover from this bounded upstream response.
• #9323 lets ordinary Docker onboarding use the release-managed prebuilt path and fall back to the shipped Dockerfile when that path is unavailable. Explicit --from and Portable onboarding retain their existing paths.
• #10018 produces a missing OpenClaw CLI pairing request before final observation. #10022 recognizes the same pending request during Portable settlement, which prevents duplicate requests.
• #9867 preserves the Portable Podman proxy setting across process boundaries. Later lifecycle operations retain the proxy behavior selected during onboarding.
• #9862 requires HTTPS for the maintained nemoclaw update installer request and each redirect. Update downloads no longer accept a protocol downgrade.

Local Inference

• #9843 strengthens managed vLLM selection and resumed-session validation, while #9852 restores large-model selection on DGX Station. Interrupted onboarding retains the validated model selection.
• #9906 gives the first Portable Podman inference completion a bounded 120-second request deadline within a 150-second wait. Cold local models get time to respond without weakening other readiness bounds.
• #9868 distinguishes recovery instructions for Ollama, compatible endpoints, and unavailable backend types. #9836 also names the endpoint behind a protected loopback route.
• #10025 adds --vllm-gpu-device for selecting a managed vLLM host GPU by index or full NVIDIA GPU UUID. The choice persists across resume and remains independent of sandbox GPU access.
• #9821, #9849, and #9857 strengthen managed llama.cpp dependencies and lifecycle provenance. Operators get a more consistent local inference runtime across supported platforms.

Messaging and Documentation

• #9913 binds static messaging credentials to validated endpointless provider profiles. Onboarding and channel lifecycle operations now reject incompatible provider state before reuse.
• #10021 stabilizes detailed status for paused channels, while #10026 defers credential-bound policy activation until rebuild attaches the provider. Paused status remains informative, and channel restart preserves policy ordering.
• #10047 and #10052 retain the exact validated static provider for a stopped Hermes Discord channel without starting the channel or recreating credentials.
• #9785 limits model-setting guidance to the agents that support each setting. Deep Agents Code documentation now states when no equivalent operation exists.

Thank You

Thank you to external contributors @HOYALIM for sealed configuration recovery in #9231, @udsy19 for installation, session export, host-mount, Portable, and update improvements in #9776, #9782, #9787, #9862, and #9867, and @harjothkhara for inference diagnostics and agent-specific documentation in #9785 and #9836.