NemoClaw v0.0.46 is out

[ericksoa]

NemoClaw v0.0.46

In v0.0.46, NemoClaw takes another careful step from capable instrument toward more widely inhabited world. The release widens the doorway for people arriving from Windows WSL, macOS, DGX Spark, local Ollama, Model Router, and hosted uninstall flows, replacing a little more ritual knowledge with visible paths and recoverable choices. What matters here is not merely that more platforms are named; it is that their first contact with the system is less mysterious.

The messaging universe also becomes more richly populated. WhatsApp joins the channel set with in-sandbox QR pairing, WeChat becomes faster and more durable across sandbox rebuilds, Slack mention safety is repaired, and Hermes gains the managed-tool gateway broker and startup-marker recovery work needed for a steadier integration path. These are small instruments, but together they let the sandbox speak through more channels without dissolving the policy boundary that makes it trustworthy.

Beneath that visible surface, v0.0.46 sharpens the machinery of confidence. The CLI moves closer to a coherent public grammar, non-interactive sandbox execution becomes first-class, provider tool catalogs become dramatically leaner, and the E2E framework learns to describe expected failures instead of treating every negative scenario as folklore. A release is a map of accumulated attention; this one draws attention toward the places users enter, connect, automate, and ask whether the system is healthy.

Installation, Onboarding, and Platform Reach

The release broadens the paved road for Windows WSL, local model setup, hosted uninstall, and provider bootstrap paths.

• #3824 Offer express install on Windows WSL: WSL users now get the guided express path for Windows-host Ollama setup.
• #3744 Add minimal Windows bootstrap installer: Introduces the first Windows bootstrap path for getting NemoClaw started from WSL.
• #3807 Document the Windows bootstrap script option: Adds the matching get-started documentation for the Windows bootstrap flow.
• #3732 Front Ollama with the auth proxy on WSL native Docker too Issues: #3342, #3695: Extends the Ollama auth-proxy path to WSL native Docker setups.
• #3786 Probe host pythons and fall back before Model Router venv setup Issues: #3781: Makes Model Router onboarding more tolerant of host Python differences.
• #3790 Surface install-vllm when explicitly opted in via env Issues: #3684, #3765: Keeps vLLM installation discoverable when users choose the explicit opt-in path.
• #3801 Route runCaptureEx env through buildRunnerEnv to inject NO_PROXY Issues: #2616: Repairs macOS local Ollama onboarding by carrying proxy bypass env into capture probes.
• #3124 Show flag-passing form for the hosted uninstall.sh Issues: #2357: Documents the bash -s -- form for passing uninstall flags through curl | bash.

Messaging, Channels, and Agent Integrations

Messaging support moved from scattered special cases toward durable, policy-aware channel behavior across WhatsApp, WeChat, Slack, and Hermes.

• #3392 Add WhatsApp messaging channel with QR pairing Issues: #361, #513: Adds WhatsApp as a QR-paired channel with policy, onboarding, docs, and durable state handling.
• #3682 Cache WeChat plugin in the sandbox base Issues: #3677: Improves WeChat startup performance by baking the plugin into the sandbox base path.
• #3737 Align OpenClaw pairing state root Issues: #3730: Repairs pairing state placement for OpenClaw channels, including DGX Spark and Slack paths.
• #3752 Allowlisted users for Slack channel mentions Issues: #3729: Restores safe Slack mention handling for configured allowlisted users.
• #3742 Add Hermes managed tool gateway broker: Adds the managed-tool gateway broker needed for Hermes integration work.
• #3797 Recover Hermes Tirith startup marker Issues: #3793: Fixes Hermes startup marker recovery called out by UAT/NV QA.
• #3839 Reapply WeChat channel seed after config mutations: Re-seeds WeChat channel configuration after OpenClaw rewrites config during build and doctor flows.

Inference, Runtime, and Sandbox Performance

The runtime became faster and less brittle where provider discovery, sandbox execution, and model-facing tool surfaces meet.

• #3808 Reduce sandbox tool-catalog latency for Nemotron Issues: #2600: Backports a compact searchable tool catalog, cutting provider-facing catalog overhead for large tool sets.
• #3767 Auto-detect Bedrock Runtime custom endpoints: Improves AWS and Anthropic-style inference endpoint detection and configuration.

CLI, Policy, and Operator Experience

The command surface is closer to a coherent public grammar: easier to discover, easier to automate, and less dependent on migration-era shims.

• #3635 Align public dispatch with oclif conventions: Moves command discovery, help, parsing, JSON output, and failures toward oclif-native behavior.
• #3833 Reduce command adapter migration debt: Replaces argv reconstruction with typed action inputs and moves more metadata onto command classes.
• #3667 Add nemoclaw <name> exec for non-interactive in-sandbox commands Issues: #3564, #3570, #3616: Adds a direct exec path for automation that needs to run commands inside a sandbox.
• #3674 Allowlist safe run plan output Issues: #3648: Keeps safe run-plan output available while preserving the security boundary around generated artifacts.
• #3777 Register the nemoclaw command in the plugin Issues: #2021: Fixes plugin registration so the NemoClaw command is available where the extension expects it.

Test, CI, and Release Reliability

The release also invests in the machinery that tells us when a sandbox is honestly healthy, honestly broken, or honestly expected to fail.

• #3661 Model expected-failure scenarios explicitly Issues: #3608: Adds structured expected-failure contracts for negative E2E scenarios and side-effect assertions.
• #3696 Cover channel add/remove lifecycle in E2E Issues: #3462: Adds lifecycle coverage for channel add/remove flows.
• #3791 Cover onboard negative paths: Adds negative onboarding coverage for VDR/nightly E2E confidence.
• #3731 Add layered scenario model: Improves scenario composition in the E2E framework.
• #3736 Move broad Vitest ownership to the Linux PR job: Moves broad Vitest coverage to the right CI lane.
• #3735 Clarify selective E2E scorecards: Makes selective E2E reporting easier to reason about.
• #3723 Remove CodeRabbit E2E recommendations: Cleans up noisy automated E2E recommendation behavior.
• #3845 Restore extension backup compatibility: Fixes nightly failures from extension backup auditing and credential-scan false positives after extension install changes.