{
  "markdown": "# Colosseum\n\n> Web-search-grounded adversarial debate between multiple AI agents.\n\nAgents fact-check each other's claims using live web searches, concede when proven wrong, and reach consensus — or force a verdict after 5 rounds. Final answers are based on argument quality, not majority vote.\n\n## How It Works\n\n```\nPhase 0  →  Parse question (type, stakes, debate axis)\nPhase 1  →  Build fact base with 2–3 web searches from opposing angles\nPhase 2  →  Each agent declares an initial position with evidence\nPhase 3  →  Adversarial loop: agents attack claims with live searches;\n             concede when shown wrong, hold when attacks fail\nPhase 4  →  Claude arbitrates: per-conflict verdicts → synthesized final answer\n```\n\nEarly exit when CONVERGENCE ≥ 80%. Deadlock or 5 rounds triggers forced synthesis with conditional verdicts.\n\n### Single-agent mode (zero CLIs installed)\n\nClaude runs four internal roles simultaneously:\n\n| Role | Approach |\n|------|----------|\n| Pragmatist | Fastest viable answer, real-world constraints first |\n| Skeptic | Attacks mainstream assumptions, seeks counterexamples |\n| Idealist | Best-case scenario, ideal conditions assumed |\n| Realist | Resource, time, and risk calculus |\n\nEach role web-searches the others' claims through the same round loop.\n\n## Requirements\n\n- **Brave Search MCP** — required for all web searches during debate\n- **One or more AI CLIs** — optional; enables true multi-agent mode\n\nSupported CLIs:\n\n| CLI | Install |\n|-----|---------|\n| [Gemini CLI](https://github.com/google-gemini/gemini-cli) | `npm install -g @google/gemini-cli` |\n| [aichat](https://github.com/sigoden/aichat) | `cargo install aichat` |\n| [llm](https://github.com/simonw/llm) | `pip install llm` |\n\n## Installation\n\nVia Claude Code plugin system:\n\n```bash\n/plugin install colosseum@claude-plugins-official\n```\n\nManual:\n\n```bash\ngit clone https://github.com/JinHo-von-Choi/colosseum\ncd colosseum && ./install.sh\n```\n\n## Usage\n\nTrigger the skill by including any of these phrases in your prompt:\n\n```\ncolosseum\n여러 AI에게 물어봐\nAI 토론\n비판적으로 분석\n다양한 관점\n```\n\nExample:\n\n```\ncolosseum: Is TypeScript worth adopting for a mid-size team?\n```\n\n```\n비판적으로 분석: RAG vs fine-tuning, which is better for production LLM apps?\n```\n\n## Output Format\n\n```\n=== COLOSSEUM ===\n\nQuestion     : [question]\nRounds       : N/5  |  Exit reason: [consensus / deadlock / exhausted]\n\n## Fact Base\n[Key facts gathered in Phase 1]\n\n## Round Summary\n| Round | Key Conflict | Web Search Result | Concession |\n|-------|-------------|-------------------|------------|\n| 1     | ...         | ...               | ...        |\n\n## Conflict Verdicts\n| Conflict | Agent A | Agent B | Evidence | Verdict |\n|----------|---------|---------|----------|---------|\n| #1       | ...     | ...     | [URL]    | A wins / draw |\n\n## Consensus\n- ...\n\n## Unresolved (conditional answer)\n- [If condition X]: A's position holds\n- [If condition Y]: B's position holds\n\n## Final Answer\n[Evidence-based synthesized conclusion]\n\n## Meta\n- Total web searches  : N (initial N + debate N)\n- Final convergence   : N%\n- Conditions under which this answer could be wrong: ...\n```\n\n## Constraints\n\n- Every factual attack must be backed by a live web search — no search, no claim\n- Deadlock is reported as deadlock, not disguised as consensus\n- Round cap is hard at 5; no extensions\n\n## License\n\nMIT\n",
  "bytes": 3359,
  "sha": "e7c8f9b2e9c22ba9211f33cb3dc943fb2d0651abbf2240d4290983808f0ef6c4",
  "repo_slug": "jinho-von-choi/colosseum",
  "fonte": "repo",
  "truncated": false,
  "api": "https://agentalog.com/api/listings/plg_jinho_von_choi_colosseum_colosseum_634235ae/readme"
}