Arbiter

**Judgement:** A well-designed, specialized skill for adding structured human-in-the-loop decision review into agent workflows. Very suitable where human approval or oversight is mandatory; unnecessary overhead for routine tasks. **Key strengths:** - Strong fit for governance, compliance, or high-risk technical decisions that should not be fully automated. - Clear JSON schema for decisions and options, with solid examples and CLI parity. - Good support for async patterns: push, status, get, and await, plus tagging and notifications for agents. **Main risks / limitations:** - Requires an external Arbiter Zebu bot and filesystem-based queue; if that infrastructure isn’t running or is misconfigured, the skill becomes non-functional. - Not appropriate for urgent real-time decisions; latency depends on human reviewers. - Potential data-exposure risk if sensitive context is pushed to Arbiter without proper access controls. **Recommended scenarios:** - Use for architectural decisions, plan reviews, and any blocking choices where you explicitly want human judgment before proceeding. - Use tags and notify fields to integrate into multi-agent or multi-project environments that have a regular “arbiter review” process. - Avoid using it for simple, low-impact questions or anything requiring immediate answers; in those cases, either decide autonomously or route via a direct human message instead.