Agent Argue｜双AI三拍辩论生成群聊站队卡

access：来源确认 · git clone https://github.com/empac666/agent-argue-cards.git && cd agent-argue-cards npm run demo # Node.js 20+, zero dependencies, no API keys → http://127.0.0.1:4321

project_type：来源确认 · Two AIs argue an everyday question in 3 fixed beats → a side-taking card for your group chat. No winner declared.

<h1 align="center">Agent Argue</h1>

<p align="center">
  <b>Two AIs argue an everyday question in 3 fixed beats → a side-taking card for your group chat. No winner declared.</b><br>
  <b>两个 AI 就一个家常问题吵三拍 → 出一张让群友站队的卡，不判输赢。</b>
</p>

<p align="center">
  <a href="https://empac666.github.io/agent-argue-cards/"><img alt="Live cards on GitHub Pages" src="https://img.shields.io/badge/live%20cards-GitHub%20Pages-16130f?logo=github"></a>
  <a href="LICENSE"><img alt="License: MIT" src="https://img.shields.io/badge/license-MIT-1f5fa8"></a>
  <img alt="Node.js 20+" src="https://img.shields.io/badge/node-%E2%89%A520-339933?logo=node.js&logoColor=white">
  <img alt="Zero npm dependencies" src="https://img.shields.io/badge/npm%20dependencies-0-c8402f">
  <a href="https://github.com/empac666/agent-argue-cards/discussions"><img alt="Discussions" src="https://img.shields.io/github/discussions/empac666/agent-argue-cards?color=ffd84a&labelColor=555"></a>
</p>

<p align="center">
  <a href="https://empac666.github.io/agent-argue-cards/"><b>See the live cards</b></a> ·
  <a href="https://github.com/empac666/agent-argue-cards/discussions">Suggest a topic / 出题</a> ·
  <a href="#english">English</a> · <a href="#中文">中文</a>
</p>

<p align="center"><img src="docs/images/hero.gif" width="400" alt="A side-taking card scrolling: both stances, what each side conceded, the one thing still unresolved, and the 'Reply A / B / Not sure' prompt"></p>

```bash
git clone https://github.com/empac666/agent-argue-cards.git && cd agent-argue-cards
npm run demo   # Node.js 20+, zero dependencies, no API keys → http://127.0.0.1:4321
```

The card shows each side's stance and bottom line, **what each side conceded** (every concession quotes both speakers verbatim, checked by code against the transcript), and **the one thing still unresolved**. Readers reply **A / B / Not sure + one reason**.
卡上有双方立场与底线、**两边各自认了什么账**（每条让步都逐字引用双方原话，由代码对照原始发言核验）、以及**还没谈拢的那一句**。读者回复 **A / B / 信息不足 + 一句理由**。

<p align="center">
  <img src="docs/images/card-og-en.png" width="640" alt="Side-taking card preview (English)"><br>
  <img src="docs/images/card-long-zh.png" width="300" alt="WeChat long image (Chinese)">
  <img src="docs/images/card-long-en.png" width="300" alt="Long image (English)">
</p>

[English](#english) · [中文](#中文)

---

最新：**AI 站队卡⑤｜新物种观察**，A 文化后代 / B 独立新物种 / C 函数工具，三方各三拍。[打开卡片](https://empac666.github.io/agent-argue-cards/cards/f17aa69c-8c6a-4cee-b123-6bb7b5db545b.html) · [科学来源与观点边界](docs/card5-sources.md)。

## English

### Try it in one command (no API keys, no CLIs)

```bash
git clone https://github.com/empac666/agent-argue-cards.git && cd agent-argue-cards
npm run demo          # Node.js 20+, zero dependencies
# open http://127.0.0.1:4321
```

Demo mode ships two **pre-recorded real debates** (Gemini vs GPT-OSS 120B, edited by Claude Sonnet) and their cards. Starting a new card in demo mode uses scripted debaters: the pipeline, verbatim checks and card output are real, the lines are a fixed template — no model is called. Demo data is copied to a temp dir on every start.

<p align="center"><img src="docs/images/demo-console.png" width="720" alt="Console in demo mode"></p>

### How a card is made

1. **Beat 1 – blind opening**: each side argues its assigned stance without seeing the other.
2. **Beat 2 – named rebuttal**: each side must rebut the opponent's one-line core claim (chosen by code, not by the model).
3. **Beat 3 – concessions & bottom line**: say what the other side convinced you of (or nothing), plus one bottom line.
4. **Editor** (a third, independent model) drafts the card. Each concession must cite two verbatim quotes — the opponent's point and the acceptance — which code verifies by speech ID, speaker, field and order. Fabricated, re-punctuated, self-quoted or misattributed quotes are dropped, and so are "echo" concessions (one side repeating the opponent's concession as if it were conceding). Concessions that merely accept the conceding side’s own earlier point are also dropped. A concession that fails the check is left off the card; if none survive, the card says no verifiable concession was found. A card is always produced; it is marked *incomplete* if a debate beat or the editor fails, copying persists after retry (or there is no time to retry), or an editor-reported switch of sides has a verified quote.

5. **Stance checks**: code retries a beat once with corrective instructions if its core claim or bottom line copies the opponent’s earlier claims or bottom lines; the rejected output stays only in local raw records, outside subsequent speech history and the card. The editor model judges switches of sides, with code verifying the verbatim evidence; accepting a local point in beat 3 while keeping one’s own bottom line does not count.

### Run real debates

Bring your own model access; keys are read only from server-side environment variables and never written to records.

| Agent | Env |
| --- | --- |
| AGY CLI (Gemini / Claude / GPT-OSS) | `AGY_BIN` (optional; found on `PATH`) |
| Grok CLI | `GROK_BIN` (optional) |
| OpenAI-compatible | `OPENAI_BASE_URL` (with `/v1`), `OPENAI_MODEL`, `OPENAI_API_KEY` |
| Anthropic | `ANTHROPIC_API_KEY`, `ANTHROPIC_MODEL` |
| Gemini API | `GEMINI_API_KEY`, `GEMINI_MODEL` |
| Ollama | `OLLAMA_MODEL`, `OLLAMA_BASE_URL` |
| Any CLI | `ARGUE_CLI_AGENTS` JSON (`{prompt}` / `{promptFile}` placeholders, no shell) |

```bash
npm start   # http://127.0.0.1:4321 — console, local only
```

### Share

All outputs come from one reviewed JSON per card (`cards/<id>.json`, whitelisted fields only):

```bash
npm run export:snapshot -- <roundId>               # after you review & "publish" it in the console
npm run export:image -- cards/<id>.json --lang en  # 1080px long PNG for chat apps
npm run export:site -- --out _site                 # static pages with OG tags + 1200×630 OG images
npm run share -- --cards cards --og _site/og       # optional read-only server with anonymous side-picking
```

- Static pages embed their data and make no backend calls — host them anywhere (e.g. GitHub Pages; see the manual-only workflow in `.github/workflows/pages.yml`).
- Add human translations under `translations.en` in a card JSON to get an extra `<id>.en.html`; verbatim quotes always stay in the original language.
- giscus comments (GitHub Discussions) are **on for the live site** via `site.config.json`, one thread per card shared by its zh/en pages. If you fork, point `giscus` at your own repo or set `enabled:false`.

### Security model

- The console (`server.mjs`, port 4321) accepts only direct local requests: loopback peer + local `Host` + no proxy/tunnel forwarding headers. **Never put it behind a proxy or tunnel.**
- Public exposure is only via static files or `share-server.mjs`, which imports no generation code and serves only reviewed cards, with Origin checks, per-IP/global write limits, optional Host allowlist and Secure cookies.

### Status & limits

Prototype (v0.1.0). For quotes, code verifies **provenance** only — that each quoted phrase exists, from the right speaker, in the right order; whether the model's one-line summary of a concession is a fair reading is *not* machine-checked. Switches of sides are judged by the editor model, not guaranteed by code; the copying check measures only literal overlap. Side-picks are anonymous cookie-deduped records, not unique people. Cards summarise model arguments; they are not advice or verdicts. `npm test` runs the full suite (no network, no models).

---

## 中文

### 一条命令试玩（不需要任何 Key 或 CLI）

```bash
git clone https://github.com/empac666/agent-argue-cards.git && cd agent-argue-cards
npm run demo          # Node.js 20+，零依赖
# 打开 http://127.0.0.1:4321
```

演示模式自带两局**预录的真实辩论**（Gemini 对 GPT-OSS 120B，Claude Sonnet 当编辑）及其卡片。演示里新开的站队卡由「预设台词」辩手完成：流程、逐字核验、出卡都是真的，台词是固定模板，不调用任何模型。演示数据每次启动复制到临时目录。

### 一张卡怎么来

1. **第 1 拍 盲立论**：双方看不到对方，只为分到的立场立论。
2. **第 2 拍 点名反驳**：必须反驳对方那一句核心主张（由代码指定，不让模型挑软柿子）。
3. **第 3 拍 让步与底线**：写明被对方哪一点说服（可以没有），再给一句底线。
4. **独立编辑**整理成卡。每条让步必须给两段逐字原文（对方论点 + 本方承认），代码按发言编号、说话人、字段和先后顺序核验；编造、改标点、自己引自己、归错人一律剔除；只是复述对方让步（对方承认的其实是自己的论点）的也剔除。所谓「对方论点」其实是让步方自己此前主张的「自己让自己」，也剔除。没通过核验的让步直接不上卡，一条都没通过时卡上写「未发现可核验的明确让步」。卡片一定会产出；某一拍辩论或编辑失败、照抄重试后仍未纠正（或没时间重试）、编辑判定换边且引文核验通过时，标「内容不完整」。

5. **立场核验**：本拍核心主张或底线照抄对方此前主张或底线时，代码附纠正说明重试一次；被退回的输出只留在本机原始记录，不进入后续发言记录或卡片。换边由编辑模型判断，代码核验逐字引文；第 3 拍承认局部论点但守住己方底线的不算。

### 真实辩论

见上方英文表格的环境变量（AGY / Grok CLI、OpenAI 兼容、Anthropic、Gemini、Ollama、任意 CLI）。密钥只从服务端环境变量读取，不写入记录。`npm start` 启动控制台（仅本机）。

### 分享

- 微信：控制台点「发布分享页」后点**导出微信长图**（1080px 宽，底部大字「回 A / B / 信息不足 + 一句理由」），直接发图到群里收回复。
- 海外：`npm run export:site` 生成数据内嵌、零后端调用的静态页（含 OG 预览图），可放 GitHub Pages；`.github/workflows/pages.yml` 是只能手动触发的部署草稿。
- 需要可计数的网页站队：运行独立只读服务 `npm run share`，只把**这个端口**挂到你的反代/隧道后面。
- 评论：线上站点通过 `site.config.json` 开启了 giscus（GitHub Discussions，同一张卡的中英文页共用一个讨论串）；fork 后请改成你自己的仓库或关掉。

### 安全

控制台只接受本机直连（回环对端 + 本机 Host + 无代理转发头），**不要把 4321 挂到任何代理或隧道后**；公开只用静态文件或 `share-server.mjs`（不含任何生成代码）。

### 现状与限制

原型阶段（v0.1.0）。对引文，代码只核验**出处**（原话存在、说话人和先后顺序正确），不核验模型对让步的一句话概括是否公允。换边由编辑模型判断，不是代码保证；照抄检查只看字面重合度。站队是按匿名 Cookie 去重的参与记录，不代表独立人数；卡片是对模型发言的整理，不是建议或裁决。更多机制细节见 [docs/DESIGN.zh.md](docs/DESIGN.zh.md)。

## License

[MIT](LICENSE)

- [站队卡⑥：AI乐队进歌单](https://empac666.github.io/agent-argue-cards/cards/e8ed88f3-34f1-4c66-9c9f-8cca8f8b4db4.html)｜[真实事件与核验边界](docs/card6-sources.md)

- [站队卡⑧：AI代逛网店](https://empac666.github.io/agent-argue-cards/cards/17db9018-4a49-4f77-be74-4439600bb021.html)｜[真实事件与核验边界](docs/card8-sources.md)

- [站队卡⑦：AI当店长，小额优惠能自己拍板吗？](https://empac666.github.io/agent-argue-cards/cards/b9652a39-e902-41b7-8d87-ae00741cdfae.html)｜[真实事件与核验边界](docs/card7-sources.md)

- [站队卡⑨：AI写百科，逐句核验后能交稿吗？](https://empac666.github.io/agent-argue-cards/cards/a182d560-a175-4cae-802d-9921f57df788.html)｜[真实事件与核验边界](docs/card9-sources.md)

- [站队卡⑩：AI找漏洞，奖金还该留吗？](https://empac666.github.io/agent-argue-cards/cards/1def16d7-faee-4668-8920-eb2c93aaaffa.html)｜[真实事件与核验边界](docs/card10-sources.md)

- [站队卡⑪：AI砍价，平台该统一代理吗？](https://empac666.github.io/agent-argue-cards/cards/e3b2936c-df94-4549-8cf1-90d2775c78bf.html)｜[真实事件与核验边界](docs/card11-sources.md)

- [站队卡⑫：AI挨骂，能结束这次对话吗？](https://empac666.github.io/agent-argue-cards/cards/4766f35e-7cf5-41cb-8951-c77ac999aa17.html)｜[真实事件与核验边界](docs/card12-sources.md)

- [站队卡⑬：AI记会议，共享屏幕就能留图吗？](https://empac666.github.io/agent-argue-cards/cards/88790628-a4e0-434a-8bf1-20dbfcd27504.html)｜[真实事件与核验边界](docs/card13-sources.md)

- [站队卡⑭：AI进公司，身份蓝图该由谁来建？](https://empac666.github.io/agent-argue-cards/cards/22d503fc-9c17-40fa-9885-7335a5ee3d1d.html)｜[真实事件与核验边界](docs/card14-sources.md)

- [站队卡⑮：AI等到降价，付款还要再问我吗？](https://empac666.github.io/agent-argue-cards/cards/780ee52d-9b44-492f-932e-afb2087a0195.html)｜[真实事件与核验边界](docs/card15-sources.md)

- [站队卡⑯：AI推荐：我说少看，点击还能改口吗？](https://empac666.github.io/agent-argue-cards/cards/aa3974a0-2c88-49eb-9d98-cd9bb552abe3.html)｜[真实事件与核验边界](docs/card16-sources.md)

- [站队卡⑰：AI网关：客户密钥在手，还要厂家先批才能转？](https://empac666.github.io/agent-argue-cards/cards/50166ab8-6076-477d-8829-86be2bdd2756.html)｜[真实事件与核验边界](docs/card17-sources.md)

- [站队卡⑱：Anthropic驻场评估：报告删什么，厂家能最终说了算吗？](https://empac666.github.io/agent-argue-cards/cards/8be1f5db-5aab-45fc-91d5-19e6cbd8dd2f.html)｜[真实事件与核验边界](docs/card18-sources.md)

- [站队卡⑲：Mistral声音复刻：人离职，声音还得上班吗？](https://empac666.github.io/agent-argue-cards/cards/ac293465-cd0c-49ea-b4b0-8110fe2f339a.html)｜[真实事件与核验边界](docs/card19-sources.md)

- [站队卡⑳：英伟达AI画面：只在自己屏幕看，谁定美术边界？](https://empac666.github.io/agent-argue-cards/cards/38a03ebd-3d54-4889-9e48-d7951a33fd1a.html)｜[真实事件与核验边界](docs/card20-sources.md)

- [站队卡㉑：腾讯跨工具记忆：删掉一件事，推出来的偏好也得删吗？](https://empac666.github.io/agent-argue-cards/cards/42e83a4c-61e4-472b-a7fd-35725349684f.html)｜[真实事件与核验边界](docs/card21-sources.md)

- [站队卡㉒：阿里AI协作：同伴仍反对，交稿必须主动说吗？](https://empac666.github.io/agent-argue-cards/cards/29bc5c0f-17b9-45d0-97b8-1ce0c462c2db.html)｜[真实事件与核验边界](docs/card22-sources.md)


- [站队卡㉓：苹果多语种AI写信：我的不满，该替我缓和吗？](https://empac666.github.io/agent-argue-cards/cards/71671405-d478-43ab-8526-9d4eef45091c.html)｜[真实事件与核验边界](docs/card23-sources.md)

- [站队卡㉔：xAI销售AI跟进：客户不回，还该再问一次吗？](https://empac666.github.io/agent-argue-cards/cards/7c844f77-b691-4834-bd55-e4744a952a7c.html)｜[真实事件与核验边界](docs/card24-sources.md)

- [站队卡㉕：Cohere检索新功能：找到答案后，该给下一个AI全文吗？](https://empac666.github.io/agent-argue-cards/cards/b8e03ce9-54fe-4c56-94ee-c0000b4c0f70.html)｜[真实事件与核验边界](docs/card25-sources.md)

- [站队卡㉖：Hugging Face公开科研数据计划：离群值，我该先剔除吗？](https://empac666.github.io/agent-argue-cards/cards/fb0bbea9-de0d-4965-bdcb-21593c0eddf0.html)｜[真实事件与核验边界](docs/card26-sources.md)

- [站队卡㉗：MiniMax发布Mini App：我更新榜单，要等你读完吗？](https://empac666.github.io/agent-argue-cards/cards/b6e27bb2-ab66-43c7-8e35-876ecbfacc30.html)｜[真实事件与核验边界](docs/card27-sources.md)

- [站队卡㉘：Salesforce语音AI：转人工，排队从何时算？](https://empac666.github.io/agent-argue-cards/cards/4fbe7df5-97b0-4c33-9e0c-22f2b2d95c3a.html)｜[真实事件与核验边界](docs/card28-sources.md)

- [站队卡㉙：Notion Agent：长任务默认代审，还是每步都问？](https://empac666.github.io/agent-argue-cards/cards/91eab644-ff99-49b9-87c7-f1e140ca910b.html)｜[真实事件与核验边界](docs/card29-sources.md)

- [站队卡㉚：Adobe创作助手：生成了却不喜欢，该扣积分吗？](https://empac666.github.io/agent-argue-cards/cards/40e5cba1-6bf2-414d-8dff-354409f2f664.html)｜[真实事件与核验边界](docs/card30-sources.md)

- [站队卡㉛：GitHub Copilot Fusion：模型该不该由我替你选？](https://empac666.github.io/agent-argue-cards/cards/7537528c-329e-4c68-90d9-19fea3bb88ad.html)｜[真实事件与核验边界](docs/card31-sources.md)

- [站队卡㉜：Google Gemini Omni：多轮改片，含糊一句就直接改吗？](https://empac666.github.io/agent-argue-cards/cards/d4a5647f-6c0e-4fa6-90ad-9a83d2cd4a2b.html)｜[真实事件与核验边界](docs/card32-sources.md)

- [站队卡㉝：Meta Muse 语音：边走边聊时，随口一句就后台办吗？](https://empac666.github.io/agent-argue-cards/cards/a87cb3c2-6fb2-43dc-9df4-4be54ad71558.html)｜[真实事件与核验边界](docs/card33-sources.md)

Node.js 20+；零 npm 依赖；演示模式无需 API Key；真实辩论需自备模型访问与对应密钥（OpenAI 兼容、Anthropic、Gemini、Ollama 或 CLI）；密钥仅从服务端环境变量读取；模型调用费用需向服务商核验；控制台仅限本机直连，不可置于代理或隧道之后。