arbiter

              __    _ __
  ____ ______/ /_  (_) /____  _____
 / __ `/ ___/ __ \/ / __/ _ \/ ___/
/ /_/ / /  / /_/ / / /_/  __/ /
\__,_/_/  /_.___/_/\__/\___/_/

Route decisions through independent AI reviewers.

Quick second opinion on a diff or a formal two-round quorum vote with adversarial tiebreaker — via Codex CLI and Gemini CLI.

What it does

Code review, architecture decisions, and risky merges all suffer from the same bias: a single reviewer. Arbiter routes tasks to Codex CLI and Gemini CLI, brings their output back into your Claude Code session, and lets Claude synthesize the results. It handles everything from a quick second opinion on a diff to a formal two-round quorum vote with an adversarial tiebreaker. Arbiter uses your existing CLI authentication — no additional API keys required.

Install

claude plugin marketplace add heurema/emporium
claude plugin install arbiter@emporium

git clone https://github.com/heurema/arbiter
cd arbiter
claude plugin install .

Quick start

# Get an independent code review of the current diff
/arbiter review

# Ask both providers the same question in parallel
/arbiter panel "What are the failure modes of this approach?"

# Formal go/no-go gate before shipping
/arbiter quorum "Merge the payment refactor to main?"

Commands

Features

• Independent review routes to Codex or Gemini and filters noise before surfacing findings
• True parallel panel runs both providers simultaneously via run_in_background so neither influences the other
• Formal quorum gate enforces a deterministic policy with an adversarial tiebreaker when models split
• Adversarial claim verification decomposes any statement into per-claim verdicts across three rounds
• Isolated implementation creates a fresh git worktree per provider so changes never touch your working tree
• Parallel diverge produces three independent solutions with different strategy hints (minimal / refactor / redesign) and lets you merge only the one you choose

Provider Configuration (optional)

Create ~/.claude/emporium-providers.local.md to customize which models Claude, Codex, and Gemini use per task:

---
version: 1
defaults:
  claude:
    model: "sonnet"
  codex:
    model: "gpt-5.3-codex"
  gemini:
    model: "auto-gemini-3"
routing:
  doctor:
    claude: "haiku"
  review:
    gemini: "gemini-3-flash-preview"
  implement:
    claude: "sonnet"
    gemini: "gemini-3.1-pro-preview"
  evaluate:
    claude: "opus"
---

Without this file, arbiter uses each CLI's default model, with Claude diverge/evaluator falling back to sonnet. Verified Gemini headless strings in this environment are auto-gemini-3, gemini-3.1-pro-preview, gemini-3-flash-preview, gemini-2.5-pro, gemini-2.5-flash, pro, and flash. If Gemini 3 preview is not enabled for your account, use pro / flash or gemini-2.5-* instead. For nested Claude CLI calls, arbiter should unset ANTHROPIC_API_KEY and CLAUDECODE so subscription auth is used reliably. See forge doctor and /arbiter doctor to validate your setup.

Requirements

• Claude Code with skill support
• Codex CLI installed and authenticated (codex --version)
• Gemini CLI installed and authenticated (gemini --version) for modes that use Gemini
• Git repository for review, implement, and diverge modes
• Clean working tree for diverge (uncommitted changes must be stashed or committed first)
• 120-second timeout per provider call for most modes; diverge defaults to 300 seconds per agent (configurable via --timeout)
• Optional: Forge for forge doctor config validation

Privacy