Auto-discovered marketplace from sharifli4/crucible
npx claudepluginhub sharifli4/crucibleTwo debater agents directly argue with each other across multiple rounds until they converge on the most correct answer — then an arbiter synthesizes the final result
sharifli4A Claude Code plugin that pits multiple AI agents against each other in adversarial debate. Agents independently propose solutions, self-critique, then attack and defend across adaptive rounds until they converge on the strongest answer. An arbiter scores every position and synthesizes the final result.
┌─────────┐ ┌─────────┐ ┌─────────┐
│ Agent 1 │ │ Agent 2 │ │ Agent N │
└────┬────┘ └────┬────┘ └────┬────┘
│ │ │
Round 1 ▼ ▼ ▼
Propose ┌─────────────────────────────────────┐
│ Each agent proposes independently │
└──────────────────┬──────────────────┘
│
Round 1.5 ▼
Self-Critique ┌─────────────────────────────────────┐
│ Each agent attacks its own proposal │
│ and produces a hardened position │
└──────────────────┬──────────────────┘
│
┌──────────────────────────────────────────┐
│ ADAPTIVE LOOP │
│ │
Round K │ ┌────────────────────────────────────┐ │
Cross-Attack │ │ Every agent attacks every other │ │
│ │ N×(N-1) critiques in parallel │ │
│ └─────────────────┬──────────────────┘ │
│ │ │
Defend │ ┌─────────────────▼──────────────────┐ │
│ │ Every agent defends and refines │ │
│ └─────────────────┬──────────────────┘ │
│ │ │
Converge? │ ┌─────────────────▼──────────────────┐ │
│ │ CONVERGED ──────────► exit loop │ │
│ │ DIVERGED ───► focus areas ► loop │ │
│ │ Max rounds ─────────► exit loop │ │
│ └────────────────────────────────────┘ │
└──────────────────┬───────────────────────┘
│
▼
┌─────────────────────────────────────┐
│ ARBITER │
│ │
│ • Reads full transcript │
│ • Scores all agents (/60) │
│ • May send back for 1 more round │
│ • Synthesizes final answer │
└─────────────────────────────────────┘
By default, agents alternate between Opus and Sonnet so the debate has genuine model diversity — different models have different reasoning patterns and blind spots. Use --model to force all agents to the same model. The arbiter always uses Opus.
In Claude Code, run:
/plugin marketplace add sharifli4/crucible
/plugin install crucible@crucible-marketplace
Requires Claude Code with a valid Anthropic API key.
/crucible [--agents N] [--model sonnet|opus] [--rounds N] <your task>
| Flag | Default | Description |
|---|---|---|
--agents | 2 | Number of debaters (2-5) |
--model | mixed | Force all debaters to one model (sonnet, opus, or haiku). Default: agents alternate between opus and sonnet for model diversity. |
--rounds | 5 | Max cross-attack rounds (2-5, exits early on convergence) |
Cost note: Each cross-attack round runs N×(N-1) critiques in parallel. With 2 agents that's 2 critiques/round; with 5 agents it's 20 critiques/round. A full 5-agent, 5-round run can exceed 80 agent calls. Start with 2-3 agents for most tasks.
/crucible implement a thread-safe LRU cache in Python
/crucible --agents 3 should we use event sourcing or CRUD for a high-write system?
/crucible --agents 3 --model opus design a distributed consensus algorithm
crucible/
├── .claude-plugin/
│ ├── plugin.json
│ └── marketplace.json
├── commands/crucible.md
├── agents/
│ ├── debater.md
│ └── arbiter.md
└── README.md
MIT