Skip to navigation

Release Notes

October 26, 2026

v0.0.139

NemoClaw v0.0.139 improves sandbox recovery and native inference checks, fixes WhatsApp removal, reports host GPUs and native backup contents, qualifies a 64 GB DGX Spark vLLM profile, and publishes compatible llama.cpp image signatures.

  • Destroying a Hermes sandbox on rootless Podman now retries native-home cleanup as the image’s pinned sandbox user when privileged cleanup cannot remove agent-owned files. If that cleanup fails, NemoClaw preserves the sandbox registry entry for a retry. Related change: PR #13639.
  • When a native inference provider is missing and only the sandbox registry records its ownership, onboarding and nemoclaw inference set name the affected sandbox. If the sandbox no longer exists, run nemoclaw <name> destroy, then onboard again. Related change: PR #13626.
  • nemoclaw backup-all now reports the directory and file counts in each successful native archive, followed by its retained path. Directory counts include the archive root, and file counts include symbolic links. Related change: PR #13073.
  • nemoclaw host probe now reports detected GPUs’ indices, names, and memory in its text and JSON output, within the public evidence limit. JSON output holds at most 256 entries across GPU and other evidence, and reports gpuInventory.collected, listed, and omitted when it omits devices. Related change: PR #13070.
  • New llama.cpp image publications include Sigstore bundle signatures and a legacy image signature. The publication workflow verifies the legacy signature with Cosign 2 before it emits a receipt. Related change: PR #13667.
  • nemoclaw <name> channels remove whatsapp now succeeds without changes when the messaging plan and current policy show that WhatsApp is not configured. It does not clear session files or queue a rebuild. Related change: PR #13221.
  • Rebuilding a Hermes sandbox with native NVIDIA inference refreshes its managed credential references while preserving its model, provider, and Discord settings. nemohermes <name> status verifies that credential reference and directs you to rebuild when it is stale. Related change: PR #13651.
  • The Qwen vLLM profile for a 64 GB DGX Spark is now supported. Automatic selection accepts hosts with at least 60 GB of detected unified memory. Related change: PR #12780.
  • Operation telemetry uses the sandbox’s recorded inference route to identify managed native model selections. This corrects the reported source for hosted native models whose route does not use inference.local. Related change: PR #13589.
  • For a verified native NVIDIA provider, nemoclaw <name> status reports degraded and exits 0 after a transient inference failure or confirmed timeout. The model response remains unverified. Run nemoclaw <name> connect --probe-only to require a successful response. Rebuild stops before sandbox deletion when a temporary upstream failure prevents credential verification. Related change: PR #13642.
  • An interrupted upgrade rebuild can resume when the failed replacement has no recorded identity and the recorded gateway and Docker daemon confirm it is absent. NemoClaw keeps the original registry entry, rebuild journal, and backup for the resumed rebuild. Related change: PR #13680.
  • On a reused gateway that lacks native provider composition, local inference onboarding now prints the complete administrator recovery command. nemoclaw onboard --resume reruns native local inference setup before it creates a sandbox when NemoClaw has no recorded native local provider for that sandbox. Related change: PR #13650.
  • For NemoClaw-managed gateways, default nemoclaw uninstall backs up eligible sandbox state, then removes owned OpenClaw state volumes even without --destroy-user-data. Externally supervised gateways retain their sandboxes and state volumes. If you use --destroy-user-data, back up any state you need separately before uninstall. Related change: PR #13596.