colosseum
proven wrong by search results, and earn scores based on accuracy and independence — not agreement. A built-in Devil's Advocate challenges t
Open source Repository Open in the app JSON README (API)
About
proven wrong by search results, and earn scores based on accuracy and independence — not agreement. A built-in Devil's Advocate challenges the final synthesis before it's delivered. How it works: 1. Builds a fact base from opposing-angle web searches 2. Each agent declares an independent position (blind draft — no agent sees others before submitting) 3. Agents debate up to 5 rounds: Prosecutor attacks, Defender responds, Adverse Witness probes shared blind spots 4. Role-switch gate fires when agents start citing the same sources or hit deadlock 5. Devil's Advocate challenges the final answer; a quantitative score (Accuracy · Evidence · Independence) determines whether it stands Works with Gemini CLI, Codex, aichat, or llm. Falls back to Claude's four internal roles (Pragmatist / Skeptic / Idealist / Realist) with zero setup.
Details
- Kind
- Plugins
- Topic
- AI, RAG & memory
- Publisher
- jinho-von-choi
- Origin
- marketplace
- Category
- ferramentas
- Stars
- 1
- Last push
- 2026-04-17T13:08:23Z
- Repository state
- ativo
- Language
- Shell
- Added
- 2026-08-30 01:48:58
- Updated
- 2026-08-30 01:48:58
- Origin id
jinho-von-choi/colosseum/colosseum
README
# Colosseum
> Web-search-grounded adversarial debate between multiple AI agents.
Agents 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.
## How It Works
```
Phase 0 → Parse question (type, stakes, debate axis)
Phase 1 → Build fact base with 2–3 web searches from opposing angles
Phase 2 → Each agent declares an initial position with evidence
Phase 3 → Adversarial loop: agents attack claims with live searches;
concede when shown wrong, hold when attacks fail
Phase 4 → Claude arbitrates: per-conflict verdicts → synthesized final answer
```
Early exit when CONVERGENCE ≥ 80%. Deadlock or 5 rounds triggers forced synthesis with conditional verdicts.
### Single-agent mode (zero CLIs installed)
Claude runs four internal roles simultaneously:
| Role | Approach |
|------|----------|
| Pragmatist | Fastest viable answer, real-world constraints first |
| Skeptic | Attacks mainstream assumptions, seeks counterexamples |
| Idealist | Best-case scenario, ideal conditions assumed |
| Realist | Resource, time, and risk calculus |
Each role web-searches the others' claims through the same round loop.
## Requirements
- **Brave Search MCP** — required for all web searches during debate
- **One or more AI CLIs** — optional; enables true multi-agent mode
Supported CLIs:
| CLI | Install |
|-----|---------|
| [Gemini CLI](https://github.com/google-gemini/gemini-cli) | `npm install -g @google/gemini-cli` |
| [aichat](https://github.com/sigoden/aichat) | `cargo install aichat` |
| [llm](https://github.com/simonw/llm) | `pip install llm` |
## Installation
Via Claude Code plugin system:
```bash
/plugin install colosseum@claude-plugins-official
```
Manual:
```bash
git clone https://github.com/JinHo-von-Choi/colosseum
cd colosseum && ./install.sh
```
## Usage
Trigger the skill by including any of these phrases in your prompt:
```
colosseum
여러 AI에게 물어봐
AI 토론
비판적으로 분석
다양한 관점
```
Example:
```
colosseum: Is TypeScript worth adopting for a mid-size team?
```
```
비판적으로 분석: RAG vs fine-tuning, which is better for production LLM apps?
```
## Output Format
```
=== COLOSSEUM ===
Question : [question]
Rounds : N/5 | Exit reason: [consensus / deadlock / exhausted]
## Fact Base
[Key facts gathered in Phase 1]
## Round Summary
| Round | Key Conflict | Web Search Result | Concession |
|-------|-------------|-------------------|------------|
| 1 | ... | ... | ... |
## Conflict Verdicts
| Conflict | Agent A | Agent B | Evidence | Verdict |
|----------|---------|---------|----------|---------|
| #1 | ... | ... | [URL] | A wins / draw |
## Consensus
- ...
## Unresolved (conditional answer)
- [If condition X]: A's position holds
- [If condition Y]: B's position holds
## Final Answer
[Evidence-based synthesized conclusion]
## Meta
- Total web searches : N (initial N + debate N)
- Final convergence : N%
- Conditions under which this answer could be wrong: ...
```
## Constraints
- Every factual attack must be backed by a live web search — no search, no claim
- Deadlock is reported as deadlock, not disguised as consensus
- Round cap is hard at 5; no extensions
## License
MIT