Autonomous Adversarial Code Review System
Your AI says "LGTM". MDTalk says "Prove it."
Features • Interface • Quick Start • 简体中文
MDTalk is an uncompromising, multi-agent code auditing framework. It replaces passive AI compliance with rigorous adversarial debate, ensuring mission-critical code quality through cross-examination.
The Chinese name 左右互搏 (Zuo You Hu Bo) originates from the classic Wuxia novel "The Legend of the Condor Heroes". It describes an advanced martial arts technique where a warrior's left and right hands fight each other independently—perfectly symbolizing our dual-agent adversarial architecture.
You push a complex feature. You ask a single AI model to review it. In seconds, it replies: "Looks great to me!"
Production breaks. Why? Single-agent AI is inherently compliant and prone to hallucination. It misses edge cases, race conditions, and semantic flaws because it lacks a verification mechanism.
MDTalk deploys a dual-agent architecture to completely eliminate AI complacency.
- The Auditor (Agent A): Systematically analyzes your codebase, proposing vulnerabilities, anti-patterns, and logical flaws.
- The Verifier (Agent B): Skeptically validates every claim made by the Auditor against the actual source code.
They engage in an autonomous, rigorous debate. No human intervention is required. They cross-examine, push back, and iterate until a mathematically hard consensus is reached. Only then are verified fixes applied directly to your codebase.
In our internal benchmarks, a single AI missed 5 critical bugs. MDTalk found all of them, debated the optimal fix, and applied the changes across 9 files autonomously.
- Adversarial Debate Engine: Two independent AI models (e.g., Claude & Codex) auditing each other to guarantee zero hallucinations.
- Strict Phase Separation: Agents are forbidden from modifying code during discussion. Code changes only happen in the dedicated apply phase after consensus.
- Autonomous Resolution: Once consensus is achieved, verified fixes are surgically applied to your source files in real-time. Three apply levels: high-priority only, high+medium, or all.
- Universal LLM Integration: Seamlessly connect with Claude Code, Codex, Gemini CLI, or any standard CLI-based AI agent.
- Advanced Consensus Detection: Keyword matching with negation awareness, word-boundary detection, turning-word rejection (but/however/但是), and partial-qualifier filtering (部分/partially). Supports full and partial agreement.
- Interactive Start Screen: Configure agents, timeouts, rounds, exchanges, apply level, language, and branch mode — all from the TUI before launch.
- Git Branch Mode: Optionally run reviews on an isolated
mdtalk/review-*branch, with filtered commits and optional merge back. - Bilingual (EN/ZH): All prompts, TUI labels, and logs fully support English and Chinese.
- Premium Terminal UI: Monitor the entire auditing process through a beautifully crafted, highly responsive Ratatui dashboard.
Experience the auditing process in real-time with our state-of-the-art terminal dashboard.
Prerequisites: Rust 1.75+ and at least one AI CLI (e.g., Claude Code).
# 1. Install MDTalk
cargo install --git https://github.com/cloveric/mdtalk --tag v0.1.9 mdtalk
# 2. Enter your target repo and open the start screen (recommended)
cd /path/to/your/project
mdtalk --project .
# 3. Override agents and review depth from CLI
mdtalk --project . --agent-a claude --agent-b codex --max-rounds 2 --max-exchanges 6
# 4. Load an explicit config file (CLI flags still override overlapping fields)
mdtalk --config ./mdtalk.toml --project .
# 5. Non-interactive safe run (CI-friendly)
mdtalk --project . --no-dashboard --no-apply --apply-level 1
# 6. Preview dashboard / print version
mdtalk --demo
mdtalk -VIn dashboard mode, overlapping settings use this priority: built-in defaults < mdtalk.toml < CLI flags < start-screen selections.
MDTalk uses a three-rule consensus model — the judgment logic adapts based on where you are in the debate:
| Situation | Consensus Rule |
|---|---|
Exchange 1, max_exchanges = 1 (only shot) |
Agent B alone decides — full or partial agreement |
Exchange 1, max_exchanges > 1 (more rounds available) |
Agent B alone — only full agreement; partial → keep debating |
| Exchange 2+ (not the last) | Both A and B must express agreement (full or partial) |
Last exchange (exchange == max_exchanges) |
Agent B alone — full or partial agreement |
Consensus detection features:
- Keyword matching — checks for
agree / confirmed / 同意 / 成立 / 认可(and 50+ variants) with negation and word-boundary awareness - Turning-word rejection —
but / however / although / 但是 / 不过after an affirmative keyword invalidates full agreement - Partial-qualifier filtering —
部分成立/partially agreeis recognized as partial, not full agreement
flowchart LR
A["🔍 Auditor (A)"] -->|proposes| MD["📄 conversation.md"]
MD -->|reads| B["✅ Verifier (B)"]
B --> CK{"Consensus?"}
CK -->|"No"| A
CK -->|"Full / Partial"| FIX["🔧 Apply Fixes"]
FIX --> NR{"Next Round?"}
NR -->|Yes| A
NR -->|No| Done["🛡️ Secured"]
拒绝虚假的"代码不错 (LGTM)"。用交叉验证重塑代码审查。
"左右互搏" 这一命名源自金庸武侠经典《射雕英雄传》，指的是双手互搏、自我对弈的绝顶武艺——这完美契合了本系统双重 AI 智能体相互盘问、交叉验证的架构理念。
当你向单一的 AI 助手提交数千行的复杂功能进行审查时，它往往会在几秒钟内回复："逻辑严密，代码优秀 (LGTM)"。 然而，潜藏在代码深处的并发死锁、参数语义错误和提权漏洞依然存在。单一 AI 模型天生具有"迎合性"和"幻觉"倾向，缺乏严谨的自我验证闭环。
左右互搏 (MDTalk) 彻底颠覆了传统的 AI 审查模式。它引入了双重独立的 AI 智能体，让它们在你的代码库中进行高强度的交叉盘问：
- 审计者 (Agent A)： 负责深度扫描、主动出击，无情地挖掘代码库中的每一个潜在缺陷。
- 验证者 (Agent B)： 充当怀疑论者，结合真实的源代码，逐一验证审计者提出的每一个问题。
它们会在没有人类干预的情况下展开激烈的技术辩论。互相反驳、不断推演，直至达成不可辩驳的共识。最终，验证者会自动将无争议的极致修复方案精准应用到你的代码中。
在我们的内部测试中，单一 AI 遗漏了 5 个核心漏洞；而 MDTalk 成功捕获了全部问题，不仅推演出了最佳修复方案，还全自动完成了 9 个文件的修改。
- 对抗性验证引擎：双模型互为攻守，从根本上消除 AI 幻觉，确保审查结果的绝对可靠。支持全部同意与部分同意两种共识结果，精准应用已达成一致的修改。
- 严格阶段隔离：讨论阶段严禁修改代码，代码修改仅在达成共识后的专门 apply 阶段执行。
- 零妥协自动修复：一旦达成技术共识，系统将以"手术刀级别"的精度自动修改源码。三档修复级别：仅高优先级、高+中优先级、全部修复。
- 无缝接入任意大模型：完美适配 Claude Code, Codex, Gemini CLI 或任何基于命令行的 AI 工具。
- 精准共识检测：关键词匹配 + 否定词识别 + 词边界检查 + 转折词拦截（but/however/但是）+ 部分限定词过滤（部分/partially）。
- 交互式启动屏：在 TUI 中配置 Agent、超时、轮次、讨论次数、修复级别、语言、分支模式，一屏搞定。
- Git 分支模式：可选在独立的
mdtalk/review-*分支上运行审查，过滤提交，可选合并回主分支。 - 双语支持（中/英）：所有提示词、界面标签、日志均支持中英文切换。
- 极客级终端大屏 (TUI)：基于 Ratatui 构建的实时终端可视化面板，以极其优雅的方式呈现 AI 间的思维碰撞。
MDTalk 使用三规则共识模型，根据当前讨论的位置动态调整判断逻辑：
| 情况 | 共识规则 |
|---|---|
第 1 次讨论，max_exchanges = 1（唯一一次） |
只看 B，全部同意或部分同意都算 |
第 1 次讨论，max_exchanges > 1（还有机会） |
只看 B，仅全部同意才算；部分同意继续讨论 |
| 第 2 次及以上（非最后一次） | A 和 B 都需要表达认可（全/部分均可） |
| 最后一次讨论（用尽次数） | 只看 B，全部同意或部分同意都算 |
共识检测能力：
- 关键词匹配 — 检测
agree / confirmed / 同意 / 成立 / 认可等 50+ 种表达，带否定词和词边界识别 - 转折词拦截 —
but / however / although / 但是 / 不过出现在肯定关键词之后，则不算完全同意 - 部分限定词过滤 —
部分成立/partially agree被识别为部分同意，不算完全同意
环境要求： Rust 1.75+ 以及至少一个可用的 AI 命令行工具（如 Claude Code）。
# 1. 安装 MDTalk
cargo install --git https://github.com/cloveric/mdtalk --tag v0.1.9 mdtalk
# 2. 进入待审查项目并打开启动页（推荐）
cd /path/to/your/project
mdtalk --project .
# 3. 通过 CLI 覆盖 agent 与审查深度
mdtalk --project . --agent-a claude --agent-b codex --max-rounds 2 --max-exchanges 6
# 4. 显式读取配置文件（同名项仍可被 CLI 覆盖）
mdtalk --config ./mdtalk.toml --project .
# 5. 非交互安全模式（适合 CI）
mdtalk --project . --no-dashboard --no-apply --apply-level 1
# 6. 预览仪表盘 / 查看版本
mdtalk --demo
mdtalk -V在 Dashboard 模式下，同一配置项优先级为：内置默认 < mdtalk.toml < CLI 参数 < 启动页选择。