Community skill: Claude Code skill for NemoClaw host management

[Koneisto]

What

A Claude Code skill that teaches LLMs how the OpenShell security boundary works, so they stop giving advice that breaks against it.

Repo: https://github.com/Koneisto/nemoclaw-skill

Why

LLMs confidently suggest things like curl http://... (proxy blocks it), export API_KEY=... inside sandbox (credentials never enter), or editing openclaw.json (Landlock read-only). Each produces confusing errors that look like bugs but are the security model working as designed.

This skill encodes 19 critical rules and 6 native skill design principles learned from production operation on DGX Spark with local vLLM inference.

What it covers

• OpenShell security boundary rules (presets, Landlock, overlay filesystem, SSH bugs)
• Native skill design: why helper scripts with one exec call beat MCP bridges and proxy chains
• Automated workspace recovery (watchdog + sentinel pattern for Sandbox Not Available Anymore After Machine Restart #486)
• Signal bridge production hardening (process detection, auto-restart)
• Network policy: direct endpoint entries instead of reverse proxies
• Session snapshot timing (skill cache invalidation after restore)
• Full reference docs: CLI, policies, inference, config, troubleshooting

Also usable as plain Markdown context for any LLM tool, not just Claude Code.

Key lesson

Do NOT deviate from the standard NVIDIA/OpenShell/NemoClaw architecture. Custom proxies, protocol bridges, blanket network policies — every workaround adds a fragile link that breaks on rebuild or reboot. The architecture works when you work with it.

Related

• Workarounds for Sandbox Not Available Anymore After Machine Restart #486 (workspace persistence), Gateway restart regenerates TLS certificates, breaking existing sandbox connections #888 (SSH secret), Sandbox config file (openclaw.json) is root-owned and unwritable — no supported path to change primary model #759 (openclaw.json)
• PR to awesome-nemoclaw: Add nemoclaw-skill (Claude Code skill for NemoClaw) VoltAgent/awesome-nemoclaw#4

[wscurran]

✨ Thanks for submitting this issue, which proposes a new feature for a Claude Code skill to teach LLMs about the OpenShell security boundary.

---

Possibly related open issues:

• #486 Sandbox Not Available Anymore After Machine Restart
• #759 Sandbox config file (openclaw.json) is root-owned and unwritable — no supported path to change primary model
• #888 Gateway restart regenerates TLS certificates, breaking existing sandbox connections