NemoClaw v0.0.70 is here

[cv]

NemoClaw v0.0.70 completes the move to one end-to-end testing system. Vitest now owns scenario execution, GitHub Actions owns orchestration and reporting, and the legacy shell-runner lanes are gone. The cutover keeps scheduled failure routing, scorecards, trace-timing summaries, docs validation, and two-agent security-posture coverage while tightening artifact publication and selector boundaries.

The release also makes everyday operation more forgiving. You can select a default sandbox with nemoclaw use, mistyped sandbox names receive a useful suggestion, Docker Desktop onboarding tolerates the short gateway bridge startup race, reasoning-oriented compatible endpoints validate correctly, and Windows bootstrap failures produce clearer WSL recovery guidance. Hermes rebuilds and messaging channel restarts now repair stale state instead of carrying it forward.

Under the surface, provider handoff, connect recovery, registry projections, and messaging collections have smaller and more cohesive ownership boundaries. The accompanying docs explain network approvals and Teams group allowlists, while the maintainer skills and release checks now describe the same workflow the repository actually enforces. Together, these changes make the code easier to evolve and make confidence in each release depend on executable behavior rather than duplicated migration machinery.

Onboarding, inference, and CLI

• #5754 retries the Docker Desktop and VM-backed host-gateway bridge probe during onboarding. This absorbs the brief interval where the OpenShell gateway is listening but is not yet reachable from sandbox containers, avoiding a false terminal failure.
• #5930 adds nemoclaw use <name> for selecting the default registered sandbox. Users no longer need to edit ~/.nemoclaw/sandboxes.json by hand.
• #5941 routes both onboarding provider/sandbox paths through shared phase factories. The handoff now has one typed boundary instead of duplicated projections and assertions.
• #5942 projects registry inference once at each connect-recovery boundary. Route repair, reset, reconciliation, and display now consume one authoritative provider/model pair.
• #5945 removes unused sandbox-registry projection helpers. The smaller registry surface makes the fail-closed gateway and inference paths easier to reason about.
• #5948 supports reasoning-mode validation for custom OpenAI-compatible endpoints through NEMOCLAW_REASONING. The setting survives provider handoff and resume, and reasoning-only responses no longer fail an incompatible tool-call probe. This replays and completes the original contribution from @deepujain in #3286.
• #5964 suggests the closest registered sandbox when a name is mistyped. A small typo now yields an actionable command instead of only the full registry list.

Sandboxes, Hermes, messaging, and policy

• #5856 writes the effective rebuilt policy set back to the registry. Disabled channel presets no longer remain as stale policy state after rebuild.
• #5882 removes inherited /sandbox/.openclaw state from stale Hermes base images. Older published layers can now be repaired before the sandbox drops root privileges.
• #5946 makes active and disabled messaging channel collections consistently non-null. Callers lose redundant fallback logic and share a simpler contract.
• #5949 advances Hermes to the corrected base digest and retires the temporary stale-layout guard. Onboarding and live image builds work with the new base while current layout checks remain fail-closed.
• #6016 reapplies a channel's network-policy preset when channels start re-enables it. If policy restoration fails, the persisted plan rolls back to disabled instead of rebuilding into a partially active state.

Windows and user guidance

• #5950 verifies the WSL runtime before installing or launching Ubuntu. When WSL is unavailable, the bootstrap attempts the standard repair path and preserves actionable output for manual or offline recovery.
• #5859 adds the two-terminal egress-approval workflow to the OpenClaw quickstart. Readers now know how to keep chat attached while approving or denying blocked network requests from the host.
• #5863 separates Teams direct-message allowlists from group/channel policy and documents the group-specific inputs. NemoClaw renders restrictive group configuration only when a non-empty allowlist is supplied.
• #5999 publishes the v0.0.69 release notes in the documentation site. The release history now links users from shipped themes to the deeper operational guidance.

Testing, security, and release confidence

• #5926 makes the Vitest sandbox-operations replacement default-enabled and selectively dispatchable. It preserves the real Ubuntu, Docker, and OpenShell boundary with structured artifacts and cleanup reporting.
• #5927 restores the source/package testing boundary for onboarding validation. Ordinary tests import source and inject collaborators instead of reaching into compiled dist artifacts and process caches.
• #5947 restores OpenClaw and Hermes Anthropic-switch modes as distinct Vitest matrix executions. Protocol, routing, configuration, and agent behavior cannot silently pass through coverage from another provider mode.
• #6010 adds a dedicated two-agent security-posture lane. OpenClaw and Hermes now share live non-root, capability, policy, startup-log, and secret-boundary proof.
• #6011 makes docs validation a first-class Vitest E2E job. Local link and CLI parity checks remain deterministic while participating in normal selection, artifacts, and PR aggregation.
• #6013 encodes trusted multiline sandbox scripts into one newline-free command argument. The shared client works with current OpenShell validation without forcing each scenario to invent its own minifier.
• #5756 removes the remaining legacy E2E workflows, shell entrypoints, and compatibility support. One typed workflow now owns default, selective, explicit-only, reporting, scorecard, cleanup, and artifact behavior, backed by job-level parity evidence from the frozen v0.0.69 suite.

Maintainer workflow

• #5953 aligns triage, review priority, stale verification, post-release retargeting, and approval skills with the canonical maintainer policy. Executable regressions keep those workflows from drifting into conflicting process rules.

Thank you

Thank you to external contributor @deepujain for the original reasoning-compatible endpoint work in #3286, replayed and completed in #5948, and for the quickstart and Teams documentation improvements in #5859 and #5863.

Compare the full release range: v0.0.69...v0.0.70.