Mostly E2E/CI stabilization, but with a handful of changes readers will actually feel. No new version tag landed (NemoClaw stays at v0.0.67):
- DGX Station defaults to DeepSeek V4 Flash (#5825): the DGX Station managed-vLLM profile now defaults to
deepseek-ai/DeepSeek-V4-Flash, and the DeepSeek V4 Flash serve recipe drops a synthetic sampler setting from its speculative-decoding config to fix an observed accuracy issue. If you run NemoClaw inference on a DGX Station, expect a different default model and a quality fix to the speculative path. - Non-interactive onboarding now fails loudly (#5738): when endpoint validation rejects the selected provider — e.g. an invalid NVIDIA API key returning HTTP 403 — onboarding now sets
process.exitCode = 1instead of printing the error but exiting0. Anyone scripting headless onboarding should re-check their exit-code handling: failures that previously slipped through as "success" will now correctly stop a pipeline. - Onboarding oclif flag migration finished (#5916): parsed flags now flow through a typed command boundary instead of being reconstructed into argv and re-parsed. GPU flag exclusivity and device dependencies move into oclif metadata,
setupandsetup-sparkroute through the shared onboard action with oclif's native alias-aware deprecation, and the legacy manual parser/usage renderer is deleted (804 net lines removed). Behavior and documented semantics are preserved — deprecated aliases just surface oclif's standard deprecation notices now. - Internal hardening (no user-facing behavior change): a large wave of E2E release-gate stabilization across the Deep Agents Code TUI, Hermes-dashboard, Slack-policy, and Spark/Docker lanes (#5928, #5917, others); a CI migration so ordinary CLI tests run against TypeScript
src/rather than potentially-staledist/output (#5904/#5906); complexity-reducing refactors of thedoctor(#5910) and sandbox-status(#5909) commands and the onboarding sandbox-state handler (#5913); and a new package-level messaging architecture README (#5586).