NemoClaw v0.0.134 is out!

[rsliter]

NemoClaw v0.0.134 is out

NemoClaw v0.0.134 strengthens the boundaries around sandbox creation, downloads, and runtime identity. Generated and custom Dockerfiles now produce a verified local image before sandbox creation. Downloads reject symbolic links anywhere in the destination parent, and provider setup checks the gateway's stored identity profile before creating a provider.

Onboarding now gives operators more information before managed model downloads and handles more real-world setup states. OpenClaw can finish a verified CLI device's write-scope upgrade after a probe refresh. Station WSL binds managed llama.cpp to the Docker-proved GB300 GPU by UUID, while serving recipes stay on the published image pending qualification.

This release also updates managed LangChain Deep Agents Code to 0.1.71, repairs Hermes backup handling, and refreshes documentation. CI and E2E fixture changes improve the evidence available for follow-up work.

Sandbox and runtime identity

• #13035 and #13077 build generated and user-supplied Dockerfiles in the selected local engine, verify the image, and stop before sandbox creation if the build or verification fails. This preserves onboarding after OpenShell's image-only creation change.
• #12870 rejects a symbolic link in any component of a download destination's parent. This prevents a sandbox download from writing through a linked host path.
• #12947 verifies the stored gateway profile after runtime identity import. A failed or mismatched export stops provider creation; a later change produces a warning without undoing completed setup.

Onboarding and inference

• #12220 shows the selected managed serving profile, recipe, runtime image, and estimated model download size before acquisition. Operators can assess capacity and wait time before they confirm.
• #13107 lets a verified OpenClaw CLI device with pairing and read permissions request write access after probe activity refreshes its metadata. The upgrade retains identity and credential checks.
• #13048 treats a free Linux port as unbound when lsof reports mount warnings but finds no listener. Other errors still leave port ownership unverified.
• #13053 binds Station WSL managed llama.cpp to the Docker-proved GB300 by UUID. The serving recipes remain pinned to the published image while the new GB300 image awaits qualification.

Agents and backup

• #12376 updates managed LangChain Deep Agents Code to 0.1.71 with a pinned Python dependency graph and matching startup and MCP behavior.
• #13049 lets Hermes backups store a sanitized configuration even when its serialized form grows beyond the source file. Credentials remain excluded, and the live configuration stays unchanged.

Documentation and release confidence

• #13065 and #13115 update the owning documentation and add the v0.0.134 changelog entry. #13117 repairs README links to the published agent-specific routes and adds a route check so stale links fail validation.
• #13038 allows a failed-job rerun to reuse a protected base artifact from an earlier attempt in the same workflow run, while preserving source and content checks.
• #13024 and #13094 tighten Hermes Slack and configuration-export E2E checks. #13097 adds regression coverage for unfinished Station setup recovery.
• #13052 repairs platform test fixtures affected by path aliases, inherited proxies, and executable names. #13041 updates the reviewed OpenShell 0.1.2 supervisor template trust record.