i-Lang/iReview

---
license: mit
tags:
- code-review
- ai-to-ai
- i-lang
- claude-code
- cursor
- multi-model
- ilang-protocol
pretty_name: iReview - AI-to-AI Code Review Protocol
configs:
- config_name: default
data_files:
- split: train
path: data/*.jsonl
---

# iReview

> **AI-to-AI code review, powered by I-Lang protocol.**
> Any model reviews your code. Structured instructions in, structured findings out.

[](LICENSE)
[](https://ilang.ai)
[](https://github.com/ilang-ai/ilang-spec)

---

## Why

Every AI-to-AI code review tool today sends English prose between models. "You are a senior code reviewer. Please review this diff..." — and the response is unstructured text that has to be guessed and parsed.

iReview uses **I-Lang v3.0** as the communication protocol between models. Claude Code sends structured instructions:

```ilang
[EVAL:@DIFF|focus=security,bugs]=>[SCAN]=>[CLSF|typ=severity]=>[OUT]
```

The review model returns structured declarations:

```ilang
::REVIEW{id:20260430|model:deepseek-chat|decision:fail}
::FINDING{id:IR-001|severity:critical|file:src/auth.ts|line:42}
issue: JWT token accessed before validation
fix: Add token.verify() before accessing claims
::END{REVIEW}
```

Two AIs speaking a protocol. Not prose. Not guesswork. Parseable, repeatable, model-agnostic.

| | iReview | codex-plugin-cc |
|---|---|---|
| Install | One config file | Codex CLI + ChatGPT login |
| Models | Any OpenAI-compatible API | OpenAI only |
| Dependencies | Git + Bash + Python3 | Node.js 18.18+ + Codex CLI |
| Review modes | Standard + Adversarial | Standard + Adversarial + Background |
| Multi-model | Yes (chain security→perf→arch) | No |
| Protocol | I-Lang structured declarations | English prose |
| Cross-tool | CC + Cursor + Codex + Copilot + Gemini | Claude Code only |
| Stop gate | Diff-hash aware (no infinite loop) | Full state machine |

## Requirements

- `python3` (for reliable JSON handling in API calls)
- `git`
- `curl` or `urllib` (Python stdlib)

## Install (seconds)

```bash
/plugin marketplace add ilang-ai/iReview
/plugin install ireview@ilang-plugins
```

Create your config:

```bash
cp .ireview.example.json .ireview.json
# Set model and api_key — two fields, done
```

Or run `/ireview:setup` for interactive configuration.

## Config

```json
{
"model": "deepseek/deepseek-chat",
"api_key": "",
"base_url": "https://openrouter.ai/api/v1",
"focus": ["bugs", "security"],
"auto_review": false
}
```

API key priority: `CLAUDE_PLUGIN_OPTION_API_KEY` > `IREVIEW_API_KEY` env var > `api_key` field in config.

| Provider | model | base_url |
|---|---|---|
| OpenRouter (any model) | `deepseek/deepseek-chat` | `https://openrouter.ai/api/v1` |
| DeepSeek direct | `deepseek-chat` | `https://api.deepseek.com/v1` |
| OpenAI direct | `gpt-4o` | `https://api.openai.com/v1` |
| Local Ollama | `llama3` | `http://localhost:11434/v1` |

**Change model = change one line.**

## Commands

```bash
/ireview:review # Review uncommitted changes
/ireview:review --base main # Review branch vs main
/ireview:review --full # Run all .ireview-*.json configs
/ireview:adversarial # Devil's advocate review
/ireview:setup # Interactive configuration
/ireview:status # Review history and unresolved findings
/ireview:result # Show latest review details
/ireview:cancel # Cancel pending review, allow stop
```

## Auto Review Gate

Set `"auto_review": true` in config. When the session stops:

1. Stop hook detects file changes and computes diff hash
2. If this diff wasn't reviewed yet: blocks stop with instructions
3. You run `/ireview:review` — review executes, results saved
4. On next stop: diff hash matches passed review — stop allowed

**No infinite loops.** Diff-hash tracking ensures the same changes aren't re-reviewed.

## Multi-Model Review

```bash
echo '{"model":"gpt-4o","api_key":"sk-xxx","base_url":"https://api.openai.com/v1","focus":["security"]}' > .ireview-security.json
echo '{"model":"deepseek-chat","api_key":"sk-xxx","base_url":"https://api.deepseek.com/v1","focus":["performance"]}' > .ireview-perf.json
```

`/ireview:review --full` — runs all configs. Each produces its own review file.

## Architecture

```
You ──→ Claude Code ──I-Lang──→ Review Model ──I-Lang──→ Claude Code ──→ You
(implementer) (reviewer) (presents findings)
```

```
hooks/stop-gate.sh ← Fast gating. Checks diff hash, blocks or allows.
scripts/call-api.py ← I-Lang protocol layer. Sends I-Lang instructions,
parses I-Lang responses. Falls back to JSON/text.
commands/*.md ← Slash commands. Orchestrate the review flow.
skills/ireview/SKILL.md ← Protocol definition. I-Lang request/response formats.
```

The key insight: API calls go through `call-api.py` which sends I-Lang instructions as system prompts and parses `::REVIEW{}`/`::FINDING{}` responses. Models that understand I-Lang return structured declarations. Models that don't return JSON or text, and the script handles the fallback. Either way, the protocol is the interface.

## Works Across Tools

| Tool | File | Hook support |
|---|---|---|
| Claude Code | `CLAUDE.md` | Full (stop gate + commands) |
| Cursor | `.cursorrules` | Manual review only |
| Codex | `AGENTS.md` | Manual review only |
| Copilot | `.github/copilot-instructions.md` | Manual review only |
| Gemini CLI | `GEMINI.md` | Manual review only |

## Privacy

Your code goes to whatever API you configure. iReview collects nothing. For private code, use local Ollama.

Add `.ireview.json` to `.gitignore` — it contains your API key.

## License

MIT

---

_{Built by <a href="https://ilang.ai">I-Lang Protocol</a> · The native language of artificial intelligence}

iReview is an AI-to-AI code review dataset based on the I-Lang v3.0 protocol, designed to improve the code review process through structured instructions and outputs. It supports multiple models and tools, offering standard, adversarial, and multi-model review modes to ensure parseable and repeatable results. The dataset includes training data, is compatible with any OpenAI-compatible API, and emphasizes privacy protection, with user code only sent to the configured API.

搜集汇总

数据集介绍

构建方式

iReview数据集基于I-Lang v3.0协议构建，旨在实现AI与AI之间的结构化代码审查。其核心在于，审查指令以高度结构化的I-Lang语言编写，而非传统的自然语言描述。例如，使用类似[EVAL:@DIFF|focus=security,bugs]=>[SCAN]=>[CLSF|typ=severity]=>[OUT]的指令模板，明确指定审查焦点为安全性和错误。审查模型则返回以::REVIEW{}、::FINDING{}等标记的声明，内容可解析、可重复，且与模型无关。该数据集包含多模型配置的审查结果，并采用diff哈希跟踪机制避免无限循环。

使用方法

使用iReview数据集时，用户首先通过`/plugin marketplace add ilang-ai/iReview`命令安装插件，并基于`.ireview.example.json`模板创建配置文件，指定目标模型和API密钥。核心交互通过一系列斜杠指令实现：`/ireview:review`用于审查未提交的变更，`/ireview:adversarial`执行对抗性审查，`/ireview:status`查看历史记录。启用`auto_review`功能后，停止钩子会自动检测文件更改并触发审查。对于多模型审查，用户可创建多个独立配置文件（如`.ireview-security.json`），并通过`/ireview:review --full`命令统一执行，每个模型生成独立的审查文件。

背景与挑战

背景概述

iReview数据集诞生于2024年，由I-Lang Protocol团队创建，旨在解决人工智能系统间代码审查的通信效率与结构化问题。传统的AI代码审查工具依赖自然语言进行交互，导致输出结果非结构化、难以解析且复用性差。iReview采用I-Lang v3.0协议作为模型间通信的桥梁，使得审查请求与结果通过形式化的声明式语言传递，从而实现了跨模型、跨工具的标准化审查流程。该数据集包含经过协议标记的训练样本，为多模型协作的代码审查提供了基础语料，对推动AI辅助软件开发的质量保证与自动化流程具有里程碑意义。

当前挑战

iReview数据集面临的核心挑战在于多模型异构环境下的协议兼容性与审查一致性。不同大语言模型对I-Lang协议的理解与执行能力参差不齐，部分模型仅能输出非结构化文本，需依赖回退机制进行解析，这增加了结果的不确定性。数据集构建中的难点包括：原始代码差异的多样性导致标注共识难以达成，跨工具（如Claude Code、Cursor、Codex）的钩子系统支持程度不一，限制了数据采集范围。此外，协议版本的演进要求数据集持续更新以维持语义对齐，而隐私保护诉求（如企业代码使用本地模型）进一步压缩了可用训练数据的规模。

常用场景

经典使用场景

在软件工程与人工智能交叉领域中，iReview数据集主要服务于AI驱动的代码审查范式转换。其经典使用场景聚焦于构建模型间的结构化通信协议，利用I-Lang v3.0取代传统的自然语言文本交互，使得一个AI模型能够向另一个AI模型发送格式化的审查指令，并接收同样结构化的审查结果。开发者可以通过配置不同的模型（如DeepSeek、GPT-4o或本地Ollama）作为审查方，针对未提交的代码变更或分支对比发起安全、性能或架构维度的专项审查，从而实现跨模型、跨工具的标准化代码质量评估流程。

解决学术问题

iReview数据集着力破解当前AI协作审查中普遍存在的非结构化通信瓶颈——传统方法依赖英语散文式指令与响应，导致解析歧义性高、结果不可复现且难以被自动化流水线消费。通过引入I-Lang协议作为模型间的通用语言，该数据集使得审查指令与发现结果均可被精确解析、重复执行且独立于底层模型，为可验证、可审计的AI代码审查方法论奠定基础。这一创新解决了多模型协作时信息传递的模糊性与工具链碎片化问题，推动了从单模型评估向多智能体协同审查范式的学术演进，具有重要的方法论意义。

实际应用

在实际工程环境中，iReview数据集支撑起一套轻量级、可嵌入现有开发工作流的AI代码审查系统。开发者只需通过简单的插件安装与配置文件设置，即可在Claude Code、Cursor、Codex等主流开发工具中启用基于I-Lang协议的代码审查功能。系统支持自动审查门控机制，能够在代码提交前基于差异哈希追踪自动触发审查拦截，确保每次变更均经过结构化审查后方可继续；同时支持敌对性审查模式，以质疑现有实现的方式暴露潜在缺陷。这使得团队无需依赖单一模型供应商，能够灵活组合不同AI模型进行多层次审查，有效提升代码交付质量。

数据集最近研究

最新研究方向

iReview数据集聚焦于AI-to-AI代码审查的前沿方向，通过引入I-Lang v3.0协议作为模型间的结构化通信语言，彻底摒弃了传统英文散文式的无序交互。该研究顺应了多模型协作与智能体系统快速发展的热潮，通过格式化指令与声明式输出（如::REVIEW和::FINDING），实现了跨模型、跨工具（Claude Code、Cursor、Codex等）的高效代码审查，显著提升了解析性、可重复性与模型无关性。其影响在于为AI代码审查领域树立了统一协议标准，解决了长期以来模型间沟通混乱的痛点，推动了自动化安全审计、性能优化等热点任务的精准执行，对构建可靠、可扩展的AI软件工程生态具有里程碑意义。