> This page is for version Latest (default). > For other versions, use one of these documentation indexes: > - Latest (default): https://docs.nvidia.com/nemoclaw/latest/llms.txt > For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.nvidia.com/nemoclaw/llms.txt. ## v0.0.141 NemoClaw v0.0.141 adds adaptive routing configuration for managed Hermes and improves sandbox recovery, inference validation, and credential handling. Eligible installer and CLI operations now send operation telemetry to NVIDIA's Production receiver by default. ### Hermes and Inference * Managed Hermes onboarding accepts `--switchyard-config ` to generate routing configuration for efficient and capable hosted models. The capable target is also configured as the classifier. Provider credential attachment for the routed targets remains tracked in [issue #13863](https://github.com/NVIDIA/NemoClaw/issues/13863); this configuration does not establish production routing. Targets must use approved fixed hosted providers and different model IDs; Portable and custom images are not supported. Refer to the [onboarding command reference](/nemoclaw/user-guide/hermes/reference/commands#--switchyard-config-file). Related change: [PR #13982](https://github.com/NVIDIA/NemoClaw/pull/13982). * Native inference recreation defers the source sandbox's route reservation until after its backup succeeds, avoiding a conflict with its own registry entry. Related change: [PR #13844](https://github.com/NVIDIA/NemoClaw/pull/13844). * Custom inference endpoint setup and switching refuse conflicting network policy rules before changing the provider or route. Compatible-endpoint onboarding probes every resolved IPv4 and IPv6 address to confirm direct host inference is unavailable. Related changes: [PR #13927](https://github.com/NVIDIA/NemoClaw/pull/13927) and [PR #13767](https://github.com/NVIDIA/NemoClaw/pull/13767). * Managed vLLM startup failures include available log excerpts from before the latest container restart as well as the current log tail. Related change: [PR #14045](https://github.com/NVIDIA/NemoClaw/pull/14045). * OpenClaw web-research guidance requires the agent to disclose unavailable search and avoid presenting unverified results as current research. Related change: [PR #13607](https://github.com/NVIDIA/NemoClaw/pull/13607). ### Sandbox Recovery * Rebuild refuses a retained recovery journal while the original source sandbox still exists or its identity cannot be established. Early rebuild preflight failures also identify that the rebuild stopped before replacement. Related changes: [PR #14113](https://github.com/NVIDIA/NemoClaw/pull/14113) and [PR #14058](https://github.com/NVIDIA/NemoClaw/pull/14058). * Hermes rebuild archives preserve verified native NVIDIA credential references in supported configuration fields while continuing to remove raw credentials. Related change: [PR #14030](https://github.com/NVIDIA/NemoClaw/pull/14030). * Lifecycle commands detect native `openshell sandbox connect` sessions as well as sessions started through NemoClaw. Destroy and rebuild warn before interrupting detected active sessions, including when confirmation is skipped. Related change: [PR #14000](https://github.com/NVIDIA/NemoClaw/pull/14000). * OpenClaw `start` waits for the agent gateway after Docker recovery and checks inference before reporting readiness. Related change: [PR #14006](https://github.com/NVIDIA/NemoClaw/pull/14006). * On macOS, NemoClaw passes the selected Colima Docker socket to OpenShell. Connect timeouts identify a foreign container when container identity prevents NemoClaw from matching the sandbox. Related changes: [PR #13906](https://github.com/NVIDIA/NemoClaw/pull/13906) and [PR #13672](https://github.com/NVIDIA/NemoClaw/pull/13672). ### Security and Credentials * Auxiliary agent forwards bind to loopback. If NemoClaw cannot prove ownership of a legacy wildcard forward, it refuses migration and leaves that listener unchanged; it can remain remotely reachable. Related change: [PR #13515](https://github.com/NVIDIA/NemoClaw/pull/13515). * Removing the Tavily preset revokes its provider endpoint grants as well as the sandbox network rules. Related change: [PR #13662](https://github.com/NVIDIA/NemoClaw/pull/13662). * Discord enrollment verifies authentication through the selected agent runtime, including the managed Hermes Discord path. Related change: [PR #13873](https://github.com/NVIDIA/NemoClaw/pull/13873). * Managed runtime dependencies include reviewed vulnerability remediations for affected npm packages. Related change: [PR #14096](https://github.com/NVIDIA/NemoClaw/pull/14096). * Hermes permits read-only Direct Rendering Manager (DRM) metadata discovery without granting access to GPU device nodes. Related change: [PR #14027](https://github.com/NVIDIA/NemoClaw/pull/14027). ### Operation Telemetry and Diagnostics * Eligible installer and supported CLI operations attempt to send completed-operation telemetry to NVIDIA's fixed Production receiver by default. Set `NEMOCLAW_DISABLE_TELEMETRY=1` in each shell before installation or CLI use to disable collection and delivery. Refer to [Operation Telemetry](/nemoclaw/user-guide/openclaw/security/best-practices#operation-telemetry) for collected fields and exclusions. Related change: [PR #14106](https://github.com/NVIDIA/NemoClaw/pull/14106). * Untagged Classic builds report version `0.0.0` instead of `0.1.0`. Tagged releases retain their release version. Related change: [PR #14108](https://github.com/NVIDIA/NemoClaw/pull/14108). * Configuration export explains unsupported image input, strict backup explains why the unreachable-sandbox opt-in does not apply, and `providers_v2` refusal messages put the recovery command on its own line. Related changes: [PR #13789](https://github.com/NVIDIA/NemoClaw/pull/13789), [PR #13791](https://github.com/NVIDIA/NemoClaw/pull/13791), and [PR #13812](https://github.com/NVIDIA/NemoClaw/pull/13812).