v0.0.134
NemoClaw v0.0.134 updates managed LangChain Deep Agents Code and strengthens onboarding, backup, download, and provider setup. Station WSL onboarding also binds managed llama.cpp to the GPU that passed host checks.
- Managed LangChain Deep Agents Code now uses version 0.1.71 with a pinned Python dependency graph and matching startup and MCP behavior. Related change: PR #12376.
- Generated and custom Dockerfile recipes on a local Docker gateway build and verify an image before sandbox creation. A remote gateway, engine mismatch, failed build, or missing image stops creation. Related changes: PR #13035 and PR #13077.
- OpenClaw onboarding can request write access for a verified matching CLI device with pairing/read permissions after a probe refreshes its connection metadata. Related change: PR #13107.
- Managed vLLM and llama.cpp installers declare the selected profile, recipe, runtime image, and model download size before acquisition. Managed vLLM declares the image download size when the image is not cached. Related change: PR #12220.
- Hermes backups accept a sanitized configuration that grows beyond the source file size. The archive still excludes credentials, and the live configuration remains unchanged. Related change: PR #13049.
nemoclaw <sandbox> downloadrejects a symbolic link in any component of the host destination parent. Related change: PR #12870.- Runtime identity provider setup verifies the stored gateway profile after import and stops before provider creation if the export fails or differs. A later profile change or failed export produces a warning without changing the completed operation. Related change: PR #12947.
- Linux port checks ignore known
lsofmount warnings when no listener exists. Otherlsoferrors still leave port ownership unverified. Related change: PR #13048. - Station WSL managed llama.cpp selects the Docker-proved GB300 by UUID. The serving recipes remain pinned to the published image while the new GB300 image awaits qualification. Related change: PR #13053.