Changelog — July 19, 2026

A one-line release following yesterday's v2.1.210–214 hardening batch: Claude no longer invokes the /verify and /code-review skills on its own initiative. If you want either one to run, you now type /verify or /code-review explicitly.

Both skills were previously eligible for Claude to reach for autonomously — /verify exercises a change end-to-end to confirm it actually works, and /code-review scans the working diff for bugs and cleanups. Each can spawn a meaningful amount of work (and, for /code-review ultra, billed cloud agents), so an unrequested run was both a surprise and a cost. Making them explicit puts that decision back with you.

Who should care: anyone who had come to expect Claude Code to self-check its work before handing it back. It won't anymore — if verification matters for a change, ask for it. Teams that wired either skill into a habit or a written workflow should update that workflow to name the command. Nothing else in the CLI changed in this release.

Claude Code v2.1.215 release notes → Affects: /openclaw/, /openclaw/setup/, /openclaw/skills-guide/, /openclaw/configuration/

Twenty commits landed on main since v0.0.87 was tagged — no new tag yet. Most of it is DGX Station installer hardening, advisory-registry refactoring, and e2e test work, but four changes are worth knowing about:

• Messaging presets now disclose their full egress before they apply (#7187): this is the one to read. Every path that newly applies a messaging-preset policy — channels add, channels start, channels add --dry-run, onboarding widening, and rebuild reconciliation — now prints the complete effective network scope (host, port, protocol/TLS/access mode, per-rule methods and paths, and the declared binary allowlist) parsed from the exact YAML being applied, before any mutation. Previously all of those short-circuited before the host-only preview that only policy-add printed. The concrete consequence: the WhatsApp preset silently opened a narrowly scoped GET to raw.githubusercontent.com (a Baileys protocol-version file) that its own description never mentioned. The preset descriptions now name that fetch, and — importantly — the disclosure is derived from the YAML rather than from description prose, so a preset can't under-describe itself again.
• nemohermes uninstall stops the detached forward watcher it owns (#7184): a successful Hermes express onboard leaves a detached host-side Node watcher alive for the port-8642 forward, and uninstall had no cleanup path matching its process shape. So uninstall printed No local OpenShell forward processes found, exited successfully, deleted the watcher's PID state, and left the process running — which then failed the next Station express install's workload-quiescence preflight with Agent or inference workload is active. Uninstall now stops the managed watcher and its sandbox-scoped forward before removing state, signals only PIDs it can prove it owns, and exits non-zero if one it owns won't stop instead of showing a success banner.
• nemoclaw inference get sanitizes terminal output (#7165): provider and model values are now stripped of control characters before rendering the human-readable lines. JSON output and the returned route payload are deliberately unchanged, so machine consumers keep the raw gateway route contract.
• Trusted policy scanners hardened (#7159): a follow-up that closes gaps in the trusted CI ratchet and the OpenShell policy-read scanner — a locally shadowed require or lookalike resolver call can no longer satisfy the policy-read check, and a duplicate root-dependency execution path was removed. Regression tests now cover the trust-boundary and AST-decoy cases.
• Ollama onboarding explains the shared-gateway constraint (#6874): interactive onboarding could reuse a same-gateway model and skip the model menu without saying why. It now reports the required model when another sandbox already owns the shared route, and gives the exact separate-gateway command (an unused NEMOCLAW_GATEWAY_PORT plus NEMOCLAW_MODEL) for picking a different one.

Who should care: anyone running messaging channels on NemoClaw should re-read what a preset actually opens the next time they apply one — the WhatsApp case is a good reminder that a preset's description is not its policy. Operators who uninstall and reinstall Hermes on the same host were the ones hitting the mysterious "workload is active" preflight failure; that's fixed. Everyone else can wait for the next tagged release.

NemoClaw commits → Affects: /nemoclaw/, /nemoclaw/policy/, /nemoclaw/setup/, /nemoclaw/switching-providers/