grtninja/skill-arbiter

Skill Arbiter

Use this skill when the work requires:

• live skill inventory and source reconciliation
• OpenClaw / NullClaw source-risk review
• VS Code/Codex built-in baseline drift checks
• Codex app / VS Code / GitHub Copilot instruction-surface interop checks
• local Codex config and loopback LM Studio advisor checks
• stale or untracked Python detection
• quarantine/admission decisions
• public-shape privacy validation
• self-governance of this repo and its release artifacts
• ownership/legitimacy vetting for installed, candidate, and official baseline skills

Run

Install desktop launcher dependencies when using the managed security-console startup flow:

cd apps/nullclaw-desktop
npm install

Open the desktop app for developer/debug work:

python scripts/nullclaw_desktop.py

Open the desktop app through the accepted no-shell Windows launcher:

wscript.exe //B //Nologo .\scripts\launch_security_console.vbs

Run the local loopback agent without the desktop shell:

python scripts/nullclaw_agent.py

Refresh the machine-generated catalogs:

python scripts/generate_skill_catalog.py

This refresh writes both the repo-owned discovery index at skill-catalog.md and the deeper governance inventory at references/skill-catalog.md.

Run the public-shape gate:

python scripts/check_private_data_policy.py

Run the public-release gate:

python scripts/check_public_release.py

Behavior

1. Open the desktop shell first.
2. Attach or start the loopback agent.
3. Run privacy and self-governance checks.
4. Refresh the full skill/source inventory.
5. Surface the layered operator flow: startup, critical queue, active finding, mitigation, then evidence layers.
6. Keep destructive actions operator-confirmed.
7. Keep audit events in local state, not repo-tracked files.
8. Use a loopback LM Studio coding model for short coding-security guidance, preferring local Qwen when available.
9. Surface public support and security links as copy-only actions, not browser launches.
10. Keep subagent routing local-first: use healthy local OpenClaw-compatible lanes before cloud sidecars, and keep cloud sidecars on lower-reasoning bounded work unless the operator explicitly chooses otherwise.
11. Treat no-shell desktop launch acceptance as mandatory: no empty cmd.exe, powershell.exe, or pwsh.exe windows may flash or remain open during startup.
12. When governed work is substantial, record it as a quest with steps, checkpoints, deliverables, and evidence so the full request-to-result path stays human-readable and feeds cumulative agent progression.

Local advisor

Default local advisor configuration:

$env:NULLCLAW_AGENT_BASE_URL="http://127.0.0.1:9000/v1"
$env:NULLCLAW_AGENT_MODEL="radeon-qwen3.5-4b"
$env:NULLCLAW_AGENT_ENABLE_LLM="1"

The advisor must remain local-only by default. The shared app-agent lane is radeon-qwen3.5-4b; Hui Hui stays reserved for avatar-specialized endpoints. Treat http://127.0.0.1:9000/v1 as the public authoritative model plane, http://127.0.0.1:2337/v1 as the hosted large-model lane, and http://127.0.0.1:1234/v1 as a non-authoritative operator surface only.

Subagent policy:

• The user chooses the operating mode; arbiter recommendations do not override operator intent.
• Healthy local subagents should be used first for quick bounded tasks.
• Cloud subagents should default to cheaper, lower-reasoning sidecar work.
• Fast mode is not part of the governed default path.

Guardrails

• Do not launch an external browser as part of the normal app flow.
• Do not auto-install from unvetted third-party sources.
• Do not publish raw host evidence into repo-tracked files.
• Do not disable built-in VS Code/Codex skills to make overlays work.

Related references

• skill-catalog.md
• BOUNDARIES.md
• SECURITY.md
• references/skill-catalog.md
• references/skill-vetting-report.md
• references/vscode-skill-handling.md
• references/usage-chaining-multitasking.md
• references/OPENCLAW_NULLCLAW_THREAT_MATRIX_2026-03-11.md