NemoClaw v0.0.69 is out!

[cv]

NemoClaw 0.0.69 makes day-to-day sandbox operation more resilient. Destroy, recreate, rebuild, and recovery paths now preserve the right state, remove the wrong state, and fail with clearer, safer behavior when configuration or infrastructure is incomplete.

This release also deepens the Deep Agents Code and Hermes integrations. Terminal workflows are more predictable, managed Python and skills survive the workflows that users expect, messaging configuration is derived from manifests, and inference setup now handles hosted, custom, DGX Station, and Ollama paths more explicitly.

Release confidence received the same attention as runtime behavior. The live release gates now exercise the public and hosted NVIDIA inference paths correctly, Deep Agents terminal behavior is covered end to end, CI resource use is bounded, and the documentation includes more copyable policy and approval-flow examples.

Sandbox, Onboarding, and Recovery

• Sandbox destroy now wipes persistent workspace and state directories before deletion, preventing old PVC data from unexpectedly reappearing in a later sandbox #5455.
• Docker Desktop on WSL no longer selects an unavailable CDI GPU path, and an early GPU-recreate failure now rolls back the pre-patch sandbox instead of stranding it under a backup name #5537 #5541.
• Hosted endpoint validation failures, including rejected credentials, retain a nonzero exit status so automation cannot mistake a failed onboarding flow for success #5738.
• Recovery can rediscover live sandboxes when the local registry is empty, reducing manual repair after local state loss #5786.
• Fresh and recreate onboarding flows isolate ambient environment state and ignore stale provider recovery when the user has chosen a new configuration #5787 #5829.
• Apple Silicon installs can resolve Homebrew OpenShell correctly, removing an architecture-specific onboarding failure #5791.
• Backup-all recovery now handles orphaned manifests explicitly, making recovery results more complete and diagnosable #5819.
• Rebuild now aborts when its gateway provider is missing and pins the validated endpoint over stale session state, preventing recovery from silently targeting the wrong inference service #5831 #5879.

Deep Agents Code and CLI Experience

• Session export reports actionable guidance for dash-prefixed arguments instead of leaving users with an opaque parser failure #5514.
• The temporary onboarding warm-up session is hidden from normal list and export workflows, keeping session history focused on user-created work #5533.
• Deep Agents Code can opt into Tavily through a narrowly scoped network policy while continuing to deny unrelated telemetry destinations #5651.
• Dashboard guidance now accurately explains that the terminal runtime has no dashboard URL #5731.
• Empty non-interactive prompts fail immediately, avoiding launches that cannot produce useful work #5758.
• Rebuild warns before dropping user-managed files, making the persistence boundary visible before data is lost #5765.
• Secret-shaped environment injection is rejected before Deep Agents Code starts, closing an accidental credential-exposure path #5767.
• Skills created with skill-creator persist across rebuilds, so user-authored agent capabilities survive routine maintenance #5785.
• Agent dispatch now targets the terminal runtime directly and places the managed Python environment on PATH, making CLI and Python behavior consistent inside the sandbox #5815 #5816.
• Deep Agents Code reports a useful degradation warning after an out-of-memory event rather than failing without context #5818.
• Snapshot operations refuse to capture an active Deep Agents Code session, protecting snapshot consistency #5824.
• Status probes the live terminal runtime version, and abrupt disconnects restore the terminal cleanly for reconnection #5833 #5834.
• The nemo-deepagents launcher and its supported aliases now share a cleaner, consistent command surface #5881 #5883.

Hermes, Messaging, and Inference

• Hermes Agent moves to the 2026.6.19 release line with bearer-token API support, dashboard and model-picker integration, workspace builds, terminal support, required multipart handling, and a tighter secret boundary #5594.
• Invalid messaging channel values now fail fast instead of being coerced to a default channel #5707.
• Supported messaging channels are derived from their manifests, reducing drift between runtime behavior and channel declarations #5777.
• Reused sandboxes refresh stale messaging plans before applying them, so channel changes are not lost during onboarding #5788.
• DeepSeek V4 Flash compatibility and a DGX Station default recipe expand the validated local-inference options #5825.
• Hermes preserves messaging after upgrades through manifest runtime aliases and hardened environment refresh behavior, including Slack-compatible aliases #5837 #5866.
• Provider switches now retain explicit endpoint, credential, and inference-API metadata, making custom-to-managed transitions deterministic #5877.
• Ollama setup records explicit interaction state, improving recovery and status reporting for local model flows #5898.
• Inference smoke validation now retries only recognized transient failures, uses status-aware boundaries, and redacts sensitive error details #5932 #5934.

Security, Policy, and Observability

• Auto-restore seals the expected configuration hash, detecting configuration drift before restored state is trusted #5681.
• Sandbox connect and login flows apply process and file-descriptor limits through a POSIX-compatible, quiet shell helper, strengthening runtime containment without noisy startup output #5682 #5840 #5865.
• JSON agent failures preserve parseable output and attach provenance warnings for failed or untrusted child results #5683.
• Log collection retains per-source breadcrumbs even when a source is quiet, making multi-component failures easier to trace #5711.
• In-sandbox shields status no longer claims host posture it cannot verify; the host CLI remains the authoritative security view #5838.

Docs, CI, and Release Confidence

• The messaging architecture now has a contributor-facing manifest guide, clarifying how channels, credentials, policies, and runtime aliases fit together #5586.
• Documentation cleanup removes a duplicate quickstart heading, makes how-to examples easier to copy, and establishes a consistent technical style #5700 #5828 #5875.
• New policy round-trip and network-request approval examples make it easier to understand and verify sandbox egress decisions #5880 #5886.
• Automated review is a harder merge-gate signal and now includes a Nemotron Ultra advisor lane, improving correctness feedback before changes land #5601 #5873.
• Live Vitest jobs now distinguish the public nvapi- NVIDIA API from the project's hosted sk- inference service, with shared routing, environment, model, and assertion fixes across the release gate #5698 #5889 #5890 #5891 #5893 #5894.
• Deep Agents Code gained headless inference, TUI startup, secret-boundary, Python-egress, Tavily, marker, and first-run coverage that matches current OpenShell command constraints #5789 #5839 #5884 #5899 #5902 #5903 #5905 #5908 #5912 #5914.
• The release-gate harness now preserves failure outcomes, covers residual assertions and TUI correlation, serializes hosted agent proofs, and bounds full-gate resource usage #5887 #5897 #5911 #5917 #5920 #5928 #5931.
• Pre-commit and pre-push responsibilities are separated, while source-level tests and coverage checks protect the release automation itself #5901 #5904 #5906.

Thank you to external contributor @abhi-0906 for the WSL GPU, rollback, Tavily, and headless inference improvements #5537 #5541 #5651 #5789, and to @fbratten for the LangChain quickstart cleanup #5700.