Release Notes

August 5, 2026

v0.0.103

NemoClaw v0.0.103 adds nemoclaw launch <name> and its agent-specific CLI variants so you can start an agent directly after the standard sandbox preflight and recovery checks. It improves onboarding, managed inference, installation, uninstall, snapshot, rebuild, and Shields failure handling with more accurate recovery behavior and diagnostics. It also strengthens endpoint validation, sandbox privilege separation, and the managed Hermes and LangChain Deep Agents Code lifecycle.

  • The new nemoclaw launch <name> command starts the sandbox agent in one host-side step while preserving the same preflight, process recovery, inference reconciliation, and pairing checks used by connect. The command resolves the trusted interactive command from the selected agent manifest and supports OpenClaw, Hermes, and LangChain Deep Agents Code through their corresponding NemoClaw CLI variants. For more information, refer to the NemoClaw CLI Commands Reference and Quickstart.
  • Managed vLLM onboarding now rejects a model before image or checkpoint download when the host GPU does not meet its minimum compute capability. The readiness poller also stops a repeatedly restarting vLLM container instead of waiting for the full load timeout, and the DGX Spark Qwen profile no longer enables multi-token prediction by default. Non-interactive onboarding preserves an explicit model selection when you switch providers. For more information, refer to Set Up vLLM and Use Ollama.
  • Windows Subsystem for Linux Express setup now waits until Node.js can read Docker Desktop configuration before selecting the Ollama provider. Re-running the installer at the installed revision reuses a clean managed checkout instead of recloning it after dependency installation, and packaged-service teardown can complete through the expected standalone fallback. Scoped uninstall also treats an already-removed sandbox as completed cleanup while retaining retry state for unreachable or rejected deletions. For more information, refer to Additional Setup for Windows Machines, Gateway Lifecycle Authority, and Uninstall NemoClaw.
  • Onboarding and rebuild recovery now preserve the original container-start failure diagnosis, require journal-backed authority before repairing a not-ready sandbox, and print a structured safe-abort diagnostic for rebuild preflight failures. Top-level CLI and MCP bridge destruction failures return one redacted error with a nonzero exit status instead of exposing an uncaught Node.js stack trace. For more information, refer to Recover and Rebuild Sandboxes and Troubleshooting.
  • A failed snapshot create now removes its incomplete capture so later list and restore operations do not treat it as complete. Cross-sandbox restore keeps documented policy reconciliation best-effort, warns when preset mutation fails, and continues to gateway pairing when the restored sandbox remains usable. The documentation also clarifies that OpenClaw snapshot persistence follows the manifest-defined /sandbox/.openclaw/workspace paths rather than arbitrary files under /sandbox. For more information, refer to Create and Restore Snapshots and Understand Sandbox State.
  • Shields transitions now preserve independently managed MCP policy entries, retain a fresh OpenClaw sandbox’s mutable-default rollback state, and reject corrupt persisted transition state before mutation. A rejected policy transition no longer leaves shields status claiming a state that was not applied, and lock inspection remains bound to the verified descriptor. For more information, refer to Understand Runtime Changes, Manage MCP Servers, and the NemoClaw CLI Commands Reference.
  • Hermes now stores dashboard state in the canonical ~/.hermes/profiles/dashboard-home profile and safely migrates an existing legacy dashboard home when the destination is empty. Rebuild restore keeps script-backed cron jobs undispatchable until restored state and scripts validate against the same gateway, and successful gateway recovery resets the consecutive health-failure budget. Hermes build stages also apply configured corporate certificate authority trust before supported npm, uv, and Python dependency downloads. For more information, refer to Create and Restore Snapshots, Understand Runtime Changes, and Configure Corporate CA Trust.
  • Managed LangChain Deep Agents Code non-interactive failures now use a bounded error classification instead of exposing exception messages or reporting every failure as unknown. Rebuilds reuse the published Deep Agents Code base image, while the documented recovery guidance reflects the classified failure and managed-image path. For more information, refer to Run LangChain Deep Agents Code and Troubleshooting.
  • CLI and plugin endpoint validation now share expanded special-purpose address protections and reject endpoint URLs containing embedded credentials before DNS resolution. Managed images use the Debian-pinned setpriv implementation for reviewed user and group transitions and fail closed instead of starting an agent service as root when that transition is unavailable. Shields-protected confidentiality roots remain unlistable while allowing the sandbox group to resolve an absent top-level path without breaking gateway startup. For more information, refer to Meet Custom Endpoint Security Requirements, Review Sandbox Hardening, and Security Best Practices.
  • Documentation now removes the unsupported logs --audit example, places network-policy walkthrough prerequisites before the command sequence, and clarifies snapshot restore selection and platform runtime validation. For more information, refer to Troubleshoot MCP Servers, Approve Network Requests, and Create and Restore Snapshots.