NemoClaw v0.0.124 is out!!

[jyaunches]

NemoClaw v0.0.124 updates the managed OpenShell runtime to 0.0.116 and improves upgrade and sandbox recovery. Gateway readiness, forward ownership, and backup cleanup receive focused fixes to make interrupted lifecycle operations easier to recover from.

This release also improves local inference and agent interaction. llama.cpp uses the server's served context window, managed vLLM can recover an interrupted installation, and Deep Agents Code restores native interactive execution features inside OpenShell. Message dispatch and OpenClaw heartbeats receive fixes that preserve conversation behavior.

The documentation now covers these changes, including recovery limits and inference configuration. Image builds consume verified audit evidence during offline rebuilds, and fatal diagnostics redact sensitive process details. The changes below cover the user-facing improvements since v0.0.123.

Upgrades and sandbox recovery

• #11251 updates the managed OpenShell runtime and SDK to 0.0.116, aligning installation and runtime behavior with its credential handling.
• #11710 waits for the replaced packaged gateway to bind before recovery checks, avoiding a startup race during upgrades.
• #11589 removes incomplete backups after required pre-upgrade backup failures. If cleanup fails, diagnostics identify the retained path.
• #11686 preserves the recorded serving-profile selection during rebuild. This carries @Adr1an04's contribution from #11461.
• #11215 settles OpenClaw pairing before rebuild completes, so the next agent command can proceed. This carries @jiaqiang000's contribution from #10493.
• #11445 routes named sandbox commands through the recorded gateway across gateway-port registry roots, fixing commands issued from another gateway context.
• #11406 allows recovery-backup retirement after the sandbox registry entry is removed, making the documented cleanup sequence usable.
• #11657 releases lifecycle locks after interactive launch dispatch, allowing other lifecycle commands during the session.

Hermes and forward recovery

• #11196 requires proof that a listener belongs to the expected OpenShell forward before recovery accepts it. Unverified listeners remain untouched.
• #11427 reuses verified Hermes dashboard and API forwards during repeated onboarding, avoiding conflicts with the sandbox's existing ports.
• #11578 restores required Hermes Portable forwards before interactive launch and preserves auxiliary request token limits, preventing unnecessarily long title generation.
• #11661 reports forward spawn errors and early child exits through the owning operation, replacing misleading bind timeouts.
• #11668 reconciles a saved stopped Hermes Portable state with an already-running owned container, while refusing unsupported state combinations.
• #11670 removes forwards created during a partially failed recovery. It preserves pre-existing forwards and reports cleanup it cannot verify.
• #11675 shares one startup deadline across Portable recovery stages, preventing sequential waits from exceeding the operation's allowance.
• #11666 preserves redacted Portable startup diagnostics and distinguishes saved lifecycle state from unprobed health, improving offline troubleshooting.

Local inference and onboarding

• #11544 uses the authenticated llama.cpp server's served context window when no valid explicit override exists, avoiding an unrelated default.
• #11444 checks the llama.cpp host-loopback listener from the WSL CLI process on Docker Desktop, before checking the sandbox bridge route.
• #11464 suggests supported inference providers registered on the selected gateway. Failed discovery falls back to the supported-provider list.
• #11521 increases the Gemini probe response budget so reasoning can precede a visible reply. This carries @deepujain's contribution from #10311.
• #11587 recovers an interrupted managed vLLM container only when ownership, local Docker connection, and port mapping match. This carries @HOYALIM's contribution from #11530.
• #11580 accepts the canonical managed vLLM endpoint in the existing N1x preview flow, preserving the recorded preview decision through registration and rebuild.
• #11550 permits native Linux Docker OpenClaw configuration export with the verified Ollama route created during onboarding. The required proxy credential value is omitted.
• #11656 allows Ollama model unload when matching siblings are stopped or incomplete. Unverified sibling state prevents unload and makes stop fail.
• #11631 reuses Windows-hosted Ollama after the final protected-route probe succeeds, even when initial discovery missed it, avoiding an unnecessary installation path.
• #11584 shares providerless onboarding lifecycle handling across OpenClaw and Hermes, allowing deferred inference setup.
• #11633 represents absent inference explicitly during fresh providerless onboarding, avoiding empty or placeholder model configuration. Configure a managed provider before requesting inference.

Agent interaction and messaging

• #11683 restores native interactive Deep Agents Code shell, startup-command, interpreter, and hook behavior inside OpenShell. Managed headless shell execution remains disabled.
• #11276 repairs managed non-interactive Deep Agents Code replies with NVIDIA Nemotron, including successful JSON response handling.
• #11546 verifies Deep Agents Code inference before opening a session, surfacing route problems before interactive use.
• #11382 preserves piped message input and closes unused input for inline messages, preventing dispatch from waiting on an idle pipe.
• #11487 gives managed OpenClaw heartbeats isolated sessions, preventing heartbeat activity from changing the main conversation.
• #11404 rejects the literal Discord bot-token placeholder before storing credentials or configuring the channel. This carries @deepujain's contribution from #10705.
• #11624 reuses a successful Google Chat refresh registration during onboarding, avoiding a second configuration call that resets refresh readiness.

Images, diagnostics, and documentation

• #11156 consumes verified npm audit evidence during offline image rebuilds, while rejecting incomplete or mismatched evidence.
• #11634 refreshes unavailable Debian Trixie package pins, restoring image builds from the configured repositories.
• #11674 redacts fatal messaging-build and npm remediation diagnostics, preventing sensitive process details from reaching terminal and CI output.
• #11474 explains how to restore Docker access after uninstall fails, while retaining gateway state for a retry.
• #11456 reports successful documentation draft publication as success and links the draft for maintainer review.
• #11538 updates documentation for the shipped recovery, inference, and agent changes. It also states that v2 external-component declarations remain unqualified against a released OpenShell runtime.

Thank you

Thank you to @deepujain, @jiaqiang000, @HOYALIM, and @Adr1an04 for their contributions, including the original changes carried into this release.

Full comparison: v0.0.123…v0.0.124