🌙 ARIS-Code — Auto Research in Sleep

    ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
    ░  █████╗ ██████╗ ██╗███████╗            ░
    ░ ██╔══██╗██╔══██╗██║██╔════╝            ░
    ░ ███████║██████╔╝██║███████╗            ░
    ░ ██╔══██║██╔══██╗██║╚════██║            ░
    ░ ██║  ██║██║  ██║██║███████║            ░
    ░ ╚═╝  ╚═╝╚═╝  ╚═╝╚═╝╚══════╝           ░
    ░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░░
         🟦 [Claude]    🟩 [GPT 🕶️]
         executor  ←→  reviewer
         Let AI do research while you sleep

Adversarial · Multi-Agent Research Automation CLI Executor acts · Reviewer critiques · Iterate to excellence

✨ What is ARIS-Code?

ARIS-Code (Auto Research in Sleep) is a terminal-based AI research assistant built for academic researchers. Its core philosophy:

• 🤖 Executor: The primary LLM — writes code, surveys literature, drafts papers, plans experiments
• 🔍 Reviewer: An independent LLM that adversarially critiques the Executor's output via the LlmReview tool
• 🔄 Iterate: Executor writes → Reviewer critiques → Executor revises → loop until quality converges

With 42 bundled research skills, ARIS covers the full pipeline from idea discovery to paper submission.

🚀 Installation (macOS Apple Silicon)

curl -fsSL https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep/releases/download/v0.1.0/aris-code-darwin-arm64.tar.gz | tar xz
sudo mv aris-code /usr/local/bin/aris
aris

Currently supports macOS Apple Silicon (M1/M2/M3/M4) only. Support for other platforms is on the roadmap.

⚙️ First-Run Setup

The first time you run aris, an interactive setup wizard launches automatically:

🌙 ARIS-Code Setup Wizard

[1/3] Choose Executor provider (primary LLM)
  > Anthropic Claude
    OpenAI GPT
    Google Gemini
    Zhipu GLM
    MiniMax
Enter API Key: sk-...

[2/3] Choose Reviewer provider (adversarial LLM)
  > OpenAI GPT
    Google Gemini
    Zhipu GLM
    MiniMax
Enter API Key: sk-...

[3/3] Choose language preference
    中文 (CN)
  > English (EN)

✅ Config saved to ~/.config/aris/config.json

After setup you drop straight into the REPL. Run /setup at any time to reconfigure without restarting.

🤖 Supported Providers

Design note: Anthropic Claude is Executor-only; all other providers can serve as both Executor and Reviewer. The classic pairing is Claude Executor + GPT/GLM Reviewer for true adversarial multi-agent research.

🎯 Key Features

1. 🔄 Adversarial Multi-Agent Architecture

User input
    ↓
[Executor LLM]  ──── calls ────→  LlmReview Tool
  write / code                         ↓
  research / analyze             [Reviewer LLM]
    ↑                             independent critique
    └──────── review feedback ───┘
              iterate until quality target met

LlmReview in action:

❯ Please review this paper for me
# ARIS reads the paper, calls LlmReview to get GPT-5.4/GLM-5/MiniMax's
# independent assessment — multi-round adversarial dialogue ensues

❯ Use LlmReview to say hello to the reviewer
# Direct LlmReview tool invocation

2. 📚 42 Bundled Research Skills

Use /skills to list all available skills:

/research-lit        — Literature search & survey
/idea-discovery      — Full idea discovery pipeline
/research-review     — GPT xhigh deep review
/paper-write         — LaTeX paper drafting
/paper-compile       — Paper compilation & error fixing
/auto-review-loop    — Autonomous multi-round review loop
/experiment-plan     — Experiment roadmap generation
/run-experiment      — Remote GPU deployment
/peer-review         — Conference reviewer simulation
/rebuttal            — Submission rebuttal generation
...  (42 total)

Three-tier skill priority (higher overrides lower):

~/.config/aris/skills/   [user custom — highest priority]
~/.claude/skills/        [Claude Code compatible]
bundled skills           [42 out-of-the-box skills]

3. 🖥️ REPL Commands

4. 🌐 Language Preference

Your chosen language (CN/EN) is injected into the system prompt so ARIS always responds in your preferred language — no per-message configuration needed.

5. 🛡️ Anti-Hallucination Design

The system prompt explicitly informs the model of its exact identity (ARIS-Code), preventing role confusion in multi-agent scenarios where the Executor and Reviewer are different models from different providers.

📖 Usage Examples

Literature Survey

❯ /research-lit find the latest work on diffusion models for protein design

Autonomous Review Loop

❯ /auto-review-loop
# ARIS reads the paper in the current directory and runs:
# draft → review → revise → review → ... until quality converges

Switch Executor Model

❯ /model
  Current Executor: claude-sonnet-4-5
  Switch to:
  > claude-opus-4
    gpt-5.4
    gemini-2.5-pro

Switch Reviewer

❯ /reviewer
  Current Reviewer: gpt-5.4
  Switch to:
  > glm-5
    gemini-2.5-pro
    minimax-m2.7

Direct Adversarial Review

❯ Review my method section — be brutal
# Executor reads the section, calls LlmReview,
# receives an independent adversarial critique, and iterates

📁 Configuration

~/.config/aris/
├── config.json        # Main config (provider, API keys, language)
└── skills/            # Custom user skills (override bundled skills)

Example config.json:

{
  "executor": {
    "provider": "anthropic",
    "model": "claude-sonnet-4-5",
    "api_key": "sk-ant-..."
  },
  "reviewer": {
    "provider": "openai",
    "model": "gpt-5.4",
    "api_key": "sk-..."
  },
  "language": "EN"
}

🗺️ Roadmap

• [x] Phase 0: Rust fork foundation (based on claw-code)
• [x] Phase 1: Multi-provider support (Anthropic / OpenAI / Gemini / GLM / MiniMax)
• [x] Phase 1: LlmReview adversarial critique tool
• [x] Phase 1: 42 bundled research skills
• [x] Phase 1: Language preference & anti-hallucination system prompt
• [ ] Phase 2: Skills system polish (three-tier priority UI)
• [ ] Phase 2: Web UI dashboard
• [ ] Phase 3: Linux / Windows support
• [ ] Phase 3: Local model integration (Ollama)

🙏 Credits & Acknowledgements

ARIS-Code is built on the excellent foundation of claw-code.

claw-code is an open-source Rust reimplementation of Claude Code. It provided the REPL framework, tool-calling infrastructure, and cross-platform compilation that made ARIS-Code possible. Huge thanks to the ultraworkers team for their outstanding work!

• 🔗 claw-code: https://github.com/ultraworkers/claw-code
• 🔗 ARIS-Code: https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep

📄 License

MIT License © 2025 ARIS-Code Contributors

_{🌙 Let AI do research while you sleep · Built with ❤️ and Rust}