ARIS

// Overview

ARIS — Auto-Research-In-Sleep — is a set of Markdown skills that turn an existing coding agent into an autonomous ML-research loop: plan, draft, adversarially review, iterate, persist. The authors are explicit that it is a methodology rather than a platform; the skills carry no runtime of their own, which is why the same bundle drops into Claude Code, Codex CLI, Cursor, Trae, Antigravity, GitHub Copilot CLI, OpenClaw and DeepSeek Harness. A standalone ARIS-Code CLI packages the same workflow for people who would rather install one thing.

The distinguishing mechanic is cross-model adversarial review: one model does the work and a different model reviews it, so the agent is not grading its own homework. Since v0.4.17 that reviewer can be wired up through Codex MCP against a ChatGPT subscription rather than an OpenAI API key. A research-wiki gives the loop long-term memory — `/research-lit` ingests papers into it, `/idea-creator` reads it before brainstorming and writes ideas back, `/result-to-claim` updates claim status, and failed ideas become anti-repetition memory so the agent stops re-proposing them.

The project has grown a family around the core loop: Anti-Autoresearch audits research output against 61 signals (46 integrity hack-patterns in eight families, 13 AI-style impressions, two advisory) and is explicitly a self-consistency and fabrication-forensics check, not an AI-text detector; HERO is a ~550-token contract for CLAUDE.md/AGENTS.md that bounds over-defensive proposals; ARIS-Anything generalises the five-step loop to non-academic research such as due diligence or investigative work. A technical report accompanies the repository, and there is an AGENT_GUIDE.md written for agents to read instead of the human README. Everything is MIT-licensed.

// What it does

• A plan → draft → adversarial review → iterate → persist research loop expressed entirely as Markdown skills
• Cross-model review: a second model verifies the work, wireable through Codex MCP against a ChatGPT subscription instead of an API key
• research-wiki long-term memory — literature ingestion, idea generation that reads and writes back, claim tracking, and anti-repetition memory for failed ideas
• Runs unmodified in Claude Code, Codex CLI, Cursor, Trae, Antigravity, GitHub Copilot CLI, OpenClaw and DeepSeek Harness, or via the standalone ARIS-Code CLI
• Companion audit tool (Anti-Autoresearch) checking 61 integrity and style signals into a reviewer-ready report
• ARIS-Monitor, a small always-on-top macOS widget that flags which agent sessions are waiting for approval
• AGENT_GUIDE.md — a machine-facing entry point structured for LLM consumption

// Getting started

ARIS is a skill bundle: you copy the skills into your agent's skill directory. There is no server and no package to run.

git clone https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep.git ~/aris_repo

mkdir -p ~/.claude/skills
cp -r ~/aris_repo/skills/* ~/.claude/skills/

mkdir -p ~/your-project/.claude/skills
bash ~/aris_repo/tools/smart_update.sh --project ~/your-project --apply

/research-wiki init

npm install -g @openai/codex
claude mcp add codex -s user -- codex mcp-server

cd ~/aris_repo && ls skills/ | xargs -I{} rm -rf ~/.claude/skills/{}

Commands and code are distilled from the project's own documentation — always check the official repo for the latest.

// When to use it

• Run an unattended ML research loop overnight and read the audited report in the morning
• Have a second model adversarially review an agent's claims before you trust a result
• Keep a persistent research wiki so an agent stops re-proposing ideas that already failed
• Carry the same research workflow across Claude Code, Codex, Cursor and other harnesses without rewriting it

// How ARIS compares

ARIS alongside other open-source agent skills & plugins tools AI/TLDR tracks, ranked by GitHub stars.