arbiter

What This Skill Does

The Arbiter skill provides a robust bridge between OpenClaw AI agents and human oversight via the Arbiter Zebu interface. It allows agents to pause complex workflows and solicit structured human input on architectural decisions, implementation plans, or strategic pivots. By offloading critical judgment calls to a human-in-the-loop, the Arbiter skill ensures that agents remain aligned with high-level goals while reducing the risk of autonomous error. It transforms opaque decision-making processes into transparent, reviewable batches, capturing context and potential options for human evaluation before execution resumes.

Installation

For most users, the recommended installation method is through the ClawHub package manager. Run the following command in your terminal: clawhub install arbiter. Alternatively, if you have a Node.js environment, you can install it globally via bun add -g arbiter-skill. For those who prefer managing their own build processes, the manual installation involves cloning the repository from https://github.com/5hanth/arbiter-skill.git, running npm install, and symlinking the directory into your ~/.clawdbot/skills/ folder. Ensure the Arbiter Zebu bot is active by running bunx arbiter-zebu to process the incoming queue from ~/.arbiter/queue/.

Use Cases

Use the Arbiter skill for any task requiring significant human oversight or judgment. Primary use cases include reviewing architectural trade-offs (e.g., database selection, API authentication protocols), validating long-term project plans before resource-heavy implementation begins, and grouping related decisions into a single review session for efficiency. It is also invaluable when an agent encounters ambiguous requirements that could result in costly technical debt if handled incorrectly without human alignment.

Example Prompts

1. "I've drafted three potential database architectures for the new telemetry microservice. Please use the Arbiter skill to create a review plan for the lead developer to choose between PostgreSQL and MongoDB."
2. "Before I begin implementing the authentication module, I need sign-off on the proposed JWT vs OAuth strategy. Push a decision request to Arbiter Zebu for human approval."
3. "The current roadmap for the refactoring task has multiple branching paths. Create an Arbiter review request for the stakeholders to decide which component to tackle first to minimize production risk."

Tips & Limitations

To maximize the effectiveness of the Arbiter skill, be specific in your decision contexts. Always provide clear, distinct options to the human reviewer; avoid vague requests that don't include sufficient context for an informed choice. Note that this tool is designed for async communication and is not suitable for real-time, urgent interruptions. Do not overuse it for binary, trivial decisions that the agent is already capable of handling via its baseline logic. Finally, ensure your CLAWDBOT_AGENT and CLAWDBOT_SESSION environment variables are configured correctly to ensure the feedback flows back to the appropriate agent instance.