Auto-claude-code-research-in-sleep (ARIS ⚔️🌙)

💡 Use ARIS in Claude Code / Cursor / Trae as a skill-based workflow, or get the full experience with the standalone CLI — enjoy any way you like!

🤖 AI agents: Read AGENT_GUIDE.md instead — structured for LLM consumption, not human browsing.

🔥 ARIS-Code CLI — 独立安装版 · English | ⬇️ Download

📰 ARIS-Code v0.3.11 (2026-04-11) — Proxy/custom base URL (CCSwitch) | Local models (LM Studio/Ollama) | Research Wiki | Meta-Optimize | Atomic sessions | Bash safety | Windows (experimental)

Previous versions

v0.3.0 (2026-04-03) — Multi-file memory index | Rich task system (TodoWrite) | /plan | Security hardening

v0.2.2 (2026-04-03) — /plan step-by-step planning | /tasks persistent tracking

v0.2.1 (2026-04-03) — Persistent Memory | Kimi K2.5 multi-turn fix | CJK cursor fix

v0.2.0 (2026-04-02) — Open source | Kimi + MiniMax + GLM support | Smart LlmReview routing | CI/CD

v0.1.0 (2026-04-02) — Initial release | Multi-executor & reviewer | 42 bundled skills

中文版 README | English

🌙 Let Claude Code do research while you sleep. Wake up to find your paper scored, weaknesses identified, experiments run, and narrative rewritten — autonomously.

🪶 Radically lightweight — zero dependencies, zero lock-in. The entire system is plain Markdown files. No framework to learn, no database to maintain, no Docker to configure, no daemon to babysit. Every skill is a single SKILL.md readable by any LLM — swap Claude Code for Codex CLI, OpenClaw, Cursor, Trae, Antigravity, Windsurf, or your own agent and the workflows still work. Fork it, rewrite it, adapt it to your stack.

💡 ARIS is a methodology, not a platform. What matters is the research workflow — take it wherever you go. 🌱

· · · · 💬 Join Community ·

Custom Claude Code skills for autonomous ML research workflows. These skills orchestrate cross-model collaboration — Claude Code drives the research while an external LLM (via Codex MCP) acts as a critical reviewer. 🔀 Also supports alternative model combinations (Kimi, LongCat, DeepSeek, etc.) — no Claude or OpenAI API required. For example, MiniMax-M2.7 + GLM-5 or GLM-5 + MiniMax-M2.7. 🤖 Codex CLI native — full skill set also available for OpenAI Codex. 🖱️ Cursor — works in Cursor too. 🖥️ Trae — ByteDance AI IDE. 🚀 Antigravity — Google's agent-first IDE. 🆓 Free tier via ModelScope — zero cost, zero lock-in.

💭 Why not self-play with a single model? Using Claude Code subagents or agent teams for both execution and review is technically possible, but tends to fall into local minima — the same model reviewing its own patterns creates blind spots.

Think of it like adversarial vs. stochastic bandits: a single model self-reviewing is the stochastic case (predictable reward noise), while cross-model review is adversarial (the reviewer actively probes weaknesses the executor didn't anticipate) — and adversarial bandits are fundamentally harder to game.

💭 Why two models, not more? Two is the minimum needed to break self-play blind spots, and 2-player games converge to Nash equilibrium far more efficiently than n-player ones. Adding more reviewers increases API cost and coordination overhead with diminishing returns — the biggest gain is going from 1→2, not 2→4.

Claude Code's strength is fast, fluid execution; Codex (GPT-5.4 xhigh) is slower but more deliberate and rigorous in critique. These complementary styles — speed × rigor — produce better outcomes than either model talking to itself.

🧿 Want the strongest possible reviewer? Add — reviewer: oracle-pro to any skill to route reviews through GPT-5.4 Pro via Oracle MCP. Pro-level reasoning for proof verification, experiment auditing, and final stress tests. Works with API key or free browser mode. Setup →

🎯 More Than Just a Prompt

These are full pipelines — you can also use each workflow independently. Already have an idea? Skip to Workflow 1.5. Have results? Jump to Workflow 3. Got reviews? Jump to Workflow 4. Want persistent memory? Enable Research Wiki. See Quick Start for all commands and Workflows for the full breakdown.

Basic mode — give ARIS a research direction, it handles everything:

/research-pipeline "factorized gap in discrete diffusion LMs"

🔥 Targeted mode — got a paper you want to improve? Give ARIS the paper + the code:

/research-pipeline "improve method X" — ref paper: https://arxiv.org/abs/2406.04329, base repo: https://github.com/org/project

ARIS reads the paper → finds its weaknesses → clones the codebase → generates ideas that specifically fix those weaknesses with that code → runs experiments → writes your paper. Like telling a research assistant: "read this paper, use this repo, find what's missing, and fix it."

Mix and match: ref paper only = "what can be improved?", base repo only = "what can I build with this code?", both = "improve this paper using this code."

🔥 Rebuttal mode — reviews just dropped? Don't panic. ARIS reads every concern, builds a strategy, and drafts a rebuttal that's grounded, structured, and under the character limit:

/rebuttal "paper/ + reviews" — venue: ICML, character limit: 5000

Parameter	Default	What it does
`venue`	`ICML`	Target venue (ICML/NeurIPS/ICLR/CVPR/ACL/AAAI/ACM)
`character limit`	—	Required. Hard character limit for rebuttal text
`quick mode`	`false`	Stop after parsing + strategy (Phase 0-3). See what reviewers want before drafting
`auto experiment`	`false`	Auto-run supplementary experiments via `/experiment-bridge` when reviewers ask for new evidence
`max stress test rounds`	`1`	How many times GPT-5.4 xhigh stress-tests the draft
`max followup rounds`	`3`	Per-reviewer follow-up round limit

Three safety gates — rebuttal will NOT finalize if any fails:

🔒 No fabrication — every claim maps to paper/review/user-confirmed result
🔒 No overpromise — every promise is user-approved
🔒 Full coverage — every reviewer concern is tracked

Two outputs: PASTE_READY.txt (exact char count, paste to venue) + REBUTTAL_DRAFT_rich.md (extended version for manual editing).

After acceptance — your paper is in, now prepare the presentation:

/paper-slides "paper/"     # → Beamer PDF + PPTX + speaker notes + Q&A prep
/paper-poster "paper/"     # → A0/A1 poster PDF + editable PPTX + SVG

💡 From idea to paper to podium — one toolchain. 🌱

🏆 Papers Built with ARIS

Paper	Score	Venue	Author	Stack
CS Paper	8/10 "clear accept"	CS Conference	@DefanXue & @Monglitay	Claude Code + GPT-5.4
AAAI Paper	7/10 "good paper, accept"	AAAI 2026 Main Technical	@xinbo820-web	Pure Codex CLI
UAV-CC	Under review	IEEE TGRS	@wxx827	Claude Opus 4.6 + Codex 5.4 xhigh + Cursor

🎉 Built with ARIS — from idea to submission. Full details + PDFs →

📢 What's New

2026-04-14 — 📋 /paper-claim-audit — zero-context paper-to-evidence verification. Fresh reviewer with NO prior context compares every number in the paper against raw result files. Catches rounding inflation, best-seed cherry-pick, config mismatch, delta errors, scope overclaim. Auto-integrated into Workflow 3 (Phase 4.7). Completes the 3-layer audit chain: /experiment-audit (code) → /result-to-claim (science) → /paper-claim-audit (reporting). 👁️ Visual PDF review also added to improvement loop — reviewer now sees compiled PDF, not just LaTeX source. Inspired by Hermes Agent
2026-04-13 — 🧿 GPT-5.4 Pro via Oracle — — reviewer: oracle-pro on any skill for the strongest available reviewer. API mode (fast) or browser mode (free). Supported on: /research-review, /auto-review-loop, /experiment-audit, /proof-checker, /rebuttal, /idea-creator, /research-lit. Default stays Codex xhigh. Not installed = zero impact. Setup →
2026-04-13 — 🔬 /proof-checker — rigorous mathematical proof verification via cross-model review. 20-category issue taxonomy, two-axis severity, side-condition checklists (DCT/MCT/Fubini/IFT/...), counterexample red team, proof-obligation ledger. Auto-integrated into Workflow 3: detects \begin{theorem} and runs before improvement loop. Complements /proof-writer
2026-04-10 — ⚡ Effort Levels — — effort: lite | balanced | max | beast. Controls work intensity across all skills: papers found, ideas generated, review rounds, writing depth. Codex reasoning stays xhigh always. beast = every knob to maximum for top-venue sprints. Default balanced = zero change for existing users. Details →
2026-04-10 — 🔎 DeepXiv integration — progressive paper retrieval via DeepXiv CLI. Opt-in: — sources: deepxiv or — sources: all, deepxiv. Staged reading: search → brief → head → section. pip install deepxiv-sdk to enable. Community contribution by @DreamEnding
2026-04-10 — 🛡️ /experiment-audit — cross-model experiment integrity verification. GPT-5.4 reads your eval scripts and results directly, checks for fake ground truth, self-normalized scores, phantom results, and scope inflation (#131, #57). Advisory — warns loudly, never blocks. /result-to-claim auto-reads audit if present. New experiment-integrity.md shared reference. The executor must never judge its own integrity.
2026-04-10 — 🧠 tools/smart_update.sh — intelligent skill updater. Compares local vs upstream, detects personal customizations (server paths, API keys), only updates safe skills. bash tools/smart_update.sh --apply
2026-04-10 — 🏆 Community paper: UAV-CC — first community paper with full PDF archived. UAV change captioning benchmark for IEEE TGRS by @wxx827. Stack: Claude Opus 4.6 + Codex 5.4 xhigh + Cursor. Papers now archived in community_papers/
2026-04-08 — 📚 /research-wiki — persistent research knowledge base inspired by Karpathy's LLM Wiki. Accumulates papers, ideas, experiments, and claims across the entire research lifecycle with typed relationships. Wiki-aware hooks in /research-lit (ingest papers), /idea-creator (read wiki + write ideas back), and /result-to-claim (update claim status + trigger re-ideation). Failed ideas become anti-repetition memory. ARIS now learns from its mistakes.
2026-04-05 — 🧬 /meta-optimize — outer-loop harness optimization for ARIS. Passively logs skill invocations, tool calls, failures, and parameter overrides via Claude Code hooks. Run /meta-optimize to analyze accumulated usage data and propose SKILL.md improvements — reviewer-gated, user-approved. Inspired by Meta-Harness (Lee et al., 2026). ARIS now optimizes itself.
2026-04-04 — 🔧 Codex Plugin deep integration — /codex:rescue now auto-invoked when experiments fail (Workflow 1.5) or LaTeX won't compile (Workflow 3). GPT independently diagnoses the bug before Claude retries — two AI debuggers are better than one. Optional: codex exec powers nightmare review, /codex:rescue powers auto-debug. Setup →
2026-04-03 — ☁️ Modal serverless GPU — no GPU? gpu: modal in CLAUDE.md, one command (modal run launcher.py), no SSH, no Docker, auto scale-to-zero. $30/month free tier — enough to try ARIS experiments without any hardware. pip install modal && modal setup and go. Community contribution by @zeyuzhangzyz
2026-04-03 — 🎮 Reviewer Difficulty Levels — medium (default, unchanged), hard (reviewer memory + debate protocol), nightmare (GPT reads repo directly via codex exec — Claude can't hide anything). — difficulty: nightmare for maximum stress test before submission

Earlier updates (2026-03-12 — 2026-03-30, 22 entries)

2026-03-30 — 🔥 Auto-debug & exhaust-before-surrender — experiment-bridge auto-diagnoses failures (OOM, import, CUDA, NaN) and retries up to 3×. Inspired by PUA
2026-03-30 — ☁️ Vast.ai GPU rental — gpu: vast auto-rents cheapest GPU. By @YIHONG-JIN. 🔧 MiniMax M2.7 upgrade by @octo-patch
2026-03-27 — 📄 IEEE venue support (9 families). 🔎 Semantic Scholar. By @ypd666
2026-03-26 — 📄 Document-based input — RESEARCH_BRIEF.md auto-detect
2026-03-24 — 📝 Workflow 4: /rebuttal — 7-phase pipeline, 3 safety gates
2026-03-23 — 🔧 /training-check, /result-to-claim, /ablation-planner integrated. 📦 compact mode. By @JingxuanKang & @couragec
2026-03-22 — 📋 Templates — input templates for every workflow. 📄 7 venue templates — CVPR, ACL, AAAI, ACM MM added. 🛡️ Anti-hallucination fix — Workflow 2 enforces DBLP → CrossRef → [VERIFY]. 🔗 base repo — clone a GitHub repo as base codebase (— base repo: https://github.com/org/project)
2026-03-22 — 🔍 Codex + Gemini review guide — Codex executes, Gemini reviews via local gemini-review MCP bridge. CN
2026-03-20 — 🚀 Antigravity adaptation guide — use ARIS skills in Google Antigravity (agent-first IDE). Community contribution by @PeppaPigw
2026-03-20 — 🖥️ Trae adaptation guide — use ARIS skills in Trae (ByteDance AI IDE). Community contribution by @Prometheus-cotigo. 🔢 formula-derivation — Community contribution by @Falling-Flower
2026-03-19 — 🖼️ paper-poster — Conference poster. Community contribution by @dengzhe-hou
2026-03-19 — 🔗 Workflow 1.5 upgraded — /experiment-bridge GPT-5.4 code review. 📊 W&B fix
2026-03-18 — 🎤 paper-slides + 🔁 Codex+Claude bridge + 🖱️ Cursor guide + 🤖 Codex CLI skills + 📝 grant-proposal + 🎨 paper-illustration (Gemini) + 📊 CitationClaw
2026-03-17 — 🔧 Git code sync + 🆓 ModelScope guide + parameter pass-through
2026-03-16 — 🔬 research-refine + experiment-plan — turn vague ideas into problem-anchored proposals with claim-driven experiment roadmaps. Now integrated into Workflow 1 (/idea-discovery). Community contribution by @zjYao36
2026-03-16 — 🇨🇳 Alibaba Coding Plan guide — one API key, 4 models (Kimi-K2.5 + Qwen3.5+ + GLM-5 + MiniMax-M2.7), dual-endpoint setup. Community contribution by @tianhao909
2026-03-15 — 🔀 Bring your own model! Any OpenAI-compatible API now works as reviewer via llm-chat MCP server. GLM, MiniMax, Kimi, LongCat, DeepSeek all tested — zero Claude or OpenAI API needed
2026-03-15 — 🐾 OpenClaw adaptation guide — use ARIS research workflows in OpenClaw without Claude Code slash skills
2026-03-15 — 📐 proof-writer — community skill for rigorous theorem proof drafting. 📚 Anti-hallucination citations — /paper-write now fetches real BibTeX from DBLP/CrossRef instead of LLM-generated entries — on by default, zero install
2026-03-14 — 📱 Feishu/Lark integration: three modes (off/push/interactive), mobile notifications for experiments, reviews, and checkpoints
2026-03-13 — 🛑 Human-in-the-loop: configurable AUTO_PROCEED checkpoints across all workflows. Full autopilot or step-by-step approval
2026-03-12 — 🔗 Zotero + Obsidian + local PDFs + arXiv/Scholar: multi-source literature search with cross-model novelty verification
2026-03-12 — 🚀 Three end-to-end workflows complete: one prompt → top-venue-style paper. /research-pipeline chains idea discovery → auto review → paper writing autonomously
2026-03-12 — 📝 /paper-writing workflow: narrative report → structured outline → figures → LaTeX → compiled PDF → 2-round auto-improvement (4/10 → 8.5/10)

🚀 Quick Start

# 1. Install skills
git clone https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep.git
mkdir -p ~/.claude/skills/    # create if it doesn't exist (new Claude Code versions)
cp -r Auto-claude-code-research-in-sleep/skills/* ~/.claude/skills/

# 1b. Update skills (when upstream has new versions)
cd Auto-claude-code-research-in-sleep && git pull
bash tools/smart_update.sh          # dry-run: shows what's new/changed/safe
bash tools/smart_update.sh --apply  # apply: adds new + updates safe ones

# 2. Set up Codex MCP (for review skills)
npm install -g @openai/codex
codex setup                    # set model to gpt-5.4 when prompted
claude mcp add codex -s user -- codex mcp-server

# 3. Use in Claude Code
claude
> /idea-discovery "your research direction"  # Workflow 1 — be specific! not "NLP" but "factorized gap in discrete diffusion LMs"
> /experiment-bridge                         # Workflow 1.5 — have a plan? implement + deploy + collect results
> /auto-review-loop "your paper topic or scope"  # Workflow 2: review → fix → re-review overnight
> /paper-writing "NARRATIVE_REPORT.md"       # Workflow 3: narrative → polished PDF
> /rebuttal "paper/ + reviews" — venue: ICML    # Workflow 4: parse reviews → draft rebuttal → follow-up
> /research-pipeline "your research direction"  # Full pipeline: Workflow 1 → 1.5 → 2 → 3 end-to-end
> /research-wiki init                           # 📚 Enable persistent research memory (one-time)
> /meta-optimize                                # Meta: analyze usage logs → propose skill improvements

📚 Research Wiki (optional): Give ARIS persistent memory across sessions. Papers, ideas, failed experiments — nothing is forgotten:

# In Claude Code:
> /research-wiki init                         # creates research-wiki/ in your project
# That's it. From now on, /research-lit auto-ingests papers, /idea-creator reads
# the wiki before brainstorming (and writes ideas back), /result-to-claim updates
# claim status. Failed ideas become anti-repetition memory for future ideation.

See Research Wiki for the full guide.

🧬 Meta-optimization (optional): Run these in your normal terminal (not inside Claude Code) to enable passive usage logging:
# One-time setup in your project directory
mkdir -p .claude .aris/meta tools/meta_opt
cp Auto-claude-code-research-in-sleep/templates/claude-hooks/meta_logging.json .claude/settings.json
cp Auto-claude-code-research-in-sleep/tools/meta_opt/*.sh tools/meta_opt/
chmod +x tools/meta_opt/*.sh
# Then start Claude Code — hooks are active immediately
claude
Events are logged to both project-level (.aris/meta/events.jsonl) and global (~/.aris/meta/events.jsonl) logs. After 5+ workflow runs, run /meta-optimize to see data-driven improvement proposals. Use /meta-optimize --global to analyze trends across all your projects. See Workflow M for details.

📝 Templates available! See templates/ for ready-to-use input templates for every workflow — research brief (Workflow 1), experiment plan (Workflow 1.5), narrative report (Workflow 3), paper plan (Workflow 3).

🔎 Optional: DeepXiv progressive retrieval
pip install deepxiv-sdk
Then use /deepxiv directly or opt into it from /research-lit with — sources: deepxiv or — sources: all, deepxiv.

🔎 Optional: Exa AI-powered web search
pip install exa-py
export EXA_API_KEY=your-key-here
Then use /exa-search directly or opt into it from /research-lit with — sources: exa or — sources: all, exa. Covers blogs, docs, news, and research papers with built-in content extraction.

🗑️ Uninstall: To remove ARIS skills without affecting your own personal skills:
cd Auto-claude-code-research-in-sleep && ls skills/ | xargs -I{} rm -rf ~/.claude/skills/{}

Tip: All pipeline behaviors are configurable via inline overrides — append — key: value to any command:

Parameter Default What it does

AUTO_PROCEED true Auto-continue at idea selection gate. Set false to manually pick which idea to pursue before committing GPU time

human checkpoint false Pause after each review round so you can read the score, give custom modification instructions, skip specific fixes, or stop early

sources all Which literature sources to search: zotero, obsidian, local, web, semantic-scholar, deepxiv, exa, or all. Note: semantic-scholar, deepxiv, and exa must be explicitly listed — not included in all

arxiv download false Download top relevant arXiv PDFs during literature survey. When false, only fetches metadata (title, abstract, authors)

DBLP_BIBTEX true Fetch real BibTeX from DBLP/CrossRef instead of LLM-generated entries. Eliminates hallucinated citations. Zero install

code review true GPT-5.4 xhigh reviews experiment code before GPU deployment. Set false to skip

wandb false Auto-add W&B logging to experiment scripts. Set true + configure wandb_project in CLAUDE.md. /monitor-experiment pulls training curves from W&B

illustration gemini AI illustration in Workflow 3: gemini (default, needs GEMINI_API_KEY), mermaid (free), or false (skip)

venue ICLR Target venue: ICLR, NeurIPS, ICML, CVPR, ACL, AAAI, ACM. Determines LaTeX style file and page limit

base repo false GitHub repo URL to clone as base codebase (e.g., — base repo: https://github.com/org/project). No code? Build on top of an open-source project

gpu local GPU target: local (default), remote (SSH server), or vast (rent on-demand from Vast.ai — auto-provision, auto-destroy)

compact false Generate compact summary files (IDEA_CANDIDATES.md, findings.md, EXPERIMENT_LOG.md) for short-context models and session recovery

ref paper false Reference paper to build on (PDF path or arXiv URL). Summarized first, then ideas extend/improve it. Combine with base repo for paper+code workflows

effort balanced Work intensity: lite (0.4x tokens), balanced (default), max (2.5x), beast (5-8x). Controls breadth/depth/iterations. Codex reasoning always xhigh. See Effort Levels

reviewer codex Reviewer backend: codex (GPT-5.4 xhigh, default), oracle-pro (GPT-5.4 Pro via Oracle — strongest reasoning). See Setup →

difficulty medium Reviewer adversarial level: medium (default), hard (+ memory + debate), nightmare (+ GPT reads repo via codex exec)
/research-pipeline "your topic" — AUTO_PROCEED: false                          # pause at idea selection gate
/research-pipeline "your topic" — human checkpoint: true                       # pause after each review round to give feedback
/research-pipeline "your topic" — sources: zotero, web                         # only search Zotero + web (skip local PDFs)
/research-pipeline "your topic" — sources: all, deepxiv                        # default sources plus DeepXiv progressive retrieval
/research-pipeline "your topic" — sources: all, exa                            # default sources plus Exa AI-powered web search
/research-pipeline "your topic" — arxiv download: true                         # download top arXiv PDFs during literature survey
/research-pipeline "your topic" — difficulty: nightmare                        # maximum adversarial review before submission
/research-pipeline "your topic" — effort: beast                               # all knobs to maximum — top-venue sprint
/research-pipeline "your topic" — effort: beast, reviewer: oracle-pro         # beast + GPT-5.4 Pro reviewer — ultimate mode
/research-pipeline "your topic" — effort: lite                                # quick exploration, save tokens
/research-pipeline "your topic" — effort: max, review_rounds: 3               # max effort but cap review at 3 rounds
/research-pipeline "your topic" — AUTO_PROCEED: false, human checkpoint: true  # combine options
/proof-checker "paper/" — reviewer: oracle-pro                                # Pro-level proof verification

Important: Codex MCP uses the model from ~/.codex/config.toml, not from skill files. Make sure it says model = "gpt-5.4" (recommended). Other options: gpt-5.3-codex, gpt-5.2-codex, o3. Run codex setup or edit the file directly.

Want Codex to execute but Claude Code to review? See docs/CODEX_CLAUDE_REVIEW_GUIDE.md. That path installs the base skills/skills-codex/*, then overlays skills/skills-codex-claude-review/*, and routes review-heavy skills through the local claude-review MCP bridge.

Want Codex to execute but Gemini to review locally? See docs/CODEX_GEMINI_REVIEW_GUIDE.md and CN. That path installs the base skills/skills-codex/*, then overlays skills/skills-codex-gemini-review/*, and routes the reviewer-aware predefined skills through the local gemini-review MCP bridge using direct Gemini API by default.

See full setup guide for details and alternative model combinations if you don't have Claude/OpenAI API.

🧠 Update skills later? Smart update analyzes what's safe:
cd Auto-claude-code-research-in-sleep
git pull
bash tools/smart_update.sh          # dry-run: shows what's new/changed/safe
bash tools/smart_update.sh --apply  # apply: adds new + updates safe ones
Compares local skills with upstream, detects personal customizations (server paths, API keys, etc.), and only updates skills that are safe to replace. Skills with your personal info are flagged for manual review.

✨ Features

📊 31 composable skills — mix and match, or chain into full pipelines (/idea-discovery, /auto-review-loop, /paper-writing, /research-pipeline)
🔍 Literature & novelty — multi-source paper search (Zotero + Obsidian + local PDFs + arXiv/Scholar) + cross-model novelty verification
💡 Idea discovery — literature survey → brainstorm 8-12 ideas → novelty check → GPU pilot experiments → ranked report
🔄 Auto review loop — 4-round autonomous review, 5/10 → 7.5/10 overnight with 20+ GPU experiments
📝 Paper writing — narrative → outline → figures → LaTeX → PDF → auto-review (4/10 → 8.5/10), one command. Anti-hallucination citations via DBLP/CrossRef
🤖 Cross-model collaboration — Claude Code executes, GPT-5.4 xhigh reviews. Adversarial, not self-play. Optional upgrade: — reviewer: oracle-pro for GPT-5.4 Pro (strongest reasoning) via Oracle
📝 Peer review — review others' papers as a conference reviewer, with structured scoring and meta-review
🖥️ Review-driven experiments — when GPT-5.4 says "run an ablation", Claude Code automatically writes the script, rsyncs to your GPU server, launches in screen, collects results, and folds them back into the paper. Just configure your server in CLAUDE.md (setup guide). No GPU? Use gpu: vast to rent one from Vast.ai on demand
🔀 Flexible models — default Claude × GPT-5.4, also supports GLM, MiniMax, Kimi, LongCat, DeepSeek, etc. — no Claude or OpenAI API required
🛑 Human-in-the-loop — configurable checkpoints at key decisions. AUTO_PROCEED=true for full autopilot, false to approve each step
📱 Feishu/Lark notifications — three modes: off (default, strongly recommended for most users), push-only (webhook, mobile alerts), interactive (approve/reject from Feishu). Zero impact when unconfigured

Preview: Push cards (group) & Interactive chat (private)

Push Only — group chat cards (experiment done, checkpoint, error, pipeline complete):

Interactive — private chat with Claude Code (approve/reject, custom instructions):
📚 Research Wiki — persistent knowledge base that accumulates papers, ideas, experiments, and claims across the research lifecycle. Failed ideas become anti-repetition memory. ARIS learns from its mistakes and gets smarter with every run. Inspired by Karpathy's LLM Wiki
🧩 Extensible — domain-specific skills welcome! Add a SKILL.md and open a PR. See community skills like dse-loop (architecture/EDA)

📈 Score Progression (Real Run)

A real overnight 4-round run on an ML research project, from borderline reject to submission-ready:

Round	Score	What Happened
Initial	5.0/10	Borderline reject
Round 1	6.5/10	Added standard metrics, discovered metric decoupling
Round 2	6.8/10	Key claim failed to reproduce, pivoted narrative
Round 3	7.0/10	Large seed study killed main improvement claim
Round 4	7.5/10 ✅	Diagnostic evidence solidified, submission ready

The loop autonomously ran 20+ GPU experiments, rewrote the paper's narrative framing, and killed claims that didn't hold up — all without human intervention.

🏆 Community Showcase — Papers Built with ARIS

Real projects where the ARIS pipeline was used end-to-end. If you've used ARIS to complete a paper, we'd love to feature it here — open an issue or PR!

Paper	Rating	Venue	Built by	Notes
CS Paper	8/10 — "Top 50% of accepted papers, clear accept"	CS Conference	@DefanXue & @Monglitay	Full ARIS pipeline: idea → experiments → auto-review → paper writing. Reviewer: "empirical findings are stark, well-supported, and expose a fundamental flaw"
AAAI 2026 Paper	7/10 — "Good paper, accept"	AAAI 2026 Main Technical	@xinbo820-web	Pure Codex CLI (ARIS-Codex skills). Accepted at AAAI 2026
UAV-CC	Under review	IEEE TGRS	@wxx827	UAV change captioning benchmark. Claude Opus 4.6 (executor) + Codex GPT-5.4 xhigh (reviewer) + Cursor Opus 4.6 (assist). PDF →

Reviewer screenshots

🎉 Papers built entirely with ARIS — from idea to acceptance. Know more? Let us know!

🧩 Awesome Community Skills & Extensions

Domain-specific skills and external projects contributed by the community. PRs welcome — just add a skills/your-skill/SKILL.md and open a PR!

💡 How to use: Community skills are not auto-wired into core workflows. To use one, ask your executor (Claude Code / OpenClaw / etc.) to read the skill's SKILL.md, then plug it into the appropriate workflow stage based on the description below.

🎉 Community Skills (13): research-refine · experiment-plan · grant-proposal · paper-poster · paper-slides · mermaid-diagram · proof-writer · comm-lit-review · dse-loop · idea-discovery-robot · formula-derivation · paper-illustration · writing-systems-papers

🌐 External Projects & Docs (10): open-source-hardening-skills · CitationClaw · auto-hparam-tuning · paper-to-course · Antigravity Adaptation Guide · OpenClaw Adaptation Guide · Cursor Adaptation Guide · Codex+Claude Review Bridge · Trae Adaptation Guide · paper-illustration

🙌 Thanks to every contributor! We fold the tables below to keep the README readable — but every skill and project here is equally valued. PRs always welcome!

🎉 Community Skills (13) — click to expand

Name	Domain	Description	Codex MCP?
🔬 `research-refine`	General	Turn a vague idea into a problem-anchored, implementation-oriented method proposal. Best inserted between `/idea-discovery` and `/auto-review-loop`	Yes
🧪 `experiment-plan`	General	Turn a refined proposal into a claim-driven experiment roadmap with ablations, budgets, and run order	No
🧭 `research-refine-pipeline`	General	One-shot chain: `/research-refine` → `/experiment-plan` for method refinement plus experiment planning	Yes
📝 `grant-proposal`	General	Grant proposal drafting (KAKENHI/NSF/NSFC/ERC/DFG/SNSF/ARC/NWO). Chains `/research-lit` → `/novelty-check` → `/research-review` → `/paper-illustration`	Yes
🎤 `paper-slides`	General	Conference talk slides (beamer → PDF + PPTX) with speaker notes, full talk script + Q&A prep. Auto slide count from talk type	Yes
🖼️ `paper-poster`	General	Conference poster (article + tcbposter → A0/A1 PDF + component PPTX + SVG). Venue-specific colors, visual review loop, Codex MCP review	Yes
📐 `proof-writer`	ML Theory	Rigorous theorem/lemma proof drafting — feasibility triage, dependency maps, honest blockage reports	No
📡 `comm-lit-review`	Communications / Wireless	Domain-specific literature review — IEEE/ACM/ScienceDirect priority, venue tiering, PHY/MAC/transport/NTN taxonomy	No
🏗️ `dse-loop`	Architecture / EDA	Autonomous design space exploration — iteratively run, analyze, and tune parameters (gem5, Yosys, etc.)	No
🤖 `idea-discovery-robot`	Robotics / Embodied AI	Workflow 1 adaptation — grounds idea discovery in embodiment, benchmark, sim2real path, and real-robot safety constraints	Yes
📐 `mermaid-diagram`	General	Mermaid diagrams (20+ types) — free alternative to `paper-illustration`, no API key needed	No
🔢 `formula-derivation`	General	Research formula development — derivation, verification, and LaTeX formatting	No
🖥️ `writing-systems-papers`	Systems	Paragraph-level blueprint for 10-12 page systems papers (OSDI/SOSP/ASPLOS/NSDI/EuroSys) — page allocation, writing patterns, self-check	Yes

🌐 External Projects & Docs (10) — click to expand

Name	Domain	Description
🛡️ open-source-hardening-skills	DevOps / OSS	10-skill pipeline to harden research code into production-ready open-source projects — audit, refactor, test, CI, docs, review
📊 CitationClaw	General	Citation impact analysis — input paper title → citation crawling, scholar identification, tiered analysis, HTML dashboard
🚀 Antigravity Adaptation Guide	General	Use ARIS skills in Google Antigravity — native SKILL.md support, dual model (Claude Opus 4.6 / Gemini 3.1 Pro), MCP setup, EN + CN guides
🐾 OpenClaw Adaptation Guide	General	Use ARIS workflow methodology in OpenClaw — skill-to-stage mapping, file-based orchestration, no Claude Code CLI needed
🖱️ Cursor Adaptation Guide	General	Use ARIS skills in Cursor — `@`-reference skills, MCP setup, workflow mapping, state file recovery across sessions
🖥️ Trae Adaptation Guide	General	Use ARIS skills in Trae (ByteDance AI IDE) — EN + CN guides
🎨 `paper-illustration`	General	AI-generated architecture diagrams via Gemini. Built on PaperBanana. Integrated into Workflow 3
🤖 `skills-codex`	General	Codex CLI sync pack for the main research skills, now including `training-check`, `result-to-claim`, `ablation-planner`, `rebuttal`, plus the `shared-references/` support directory
🎛️ auto-hparam-tuning	General	Automatic hyperparameter tuning — AI agent reads project, plans strategy, runs experiments, analyzes TensorBoard, learns from results. Hydra-based
🔁 Codex+Claude Review Bridge	General	Codex executes + Claude reviews via local `claude-review` MCP bridge with async polling
📚 paper-to-course	Education	Convert research papers (PDF/LaTeX) into interactive six-module HTML courses with formula breakdowns, literature timelines, quizzes, and glossary tooltips — single bundled file, no server needed

🔄 Workflows

These skills compose into a full research lifecycle. The four workflows can be used independently or chained together:

Exploring a new area (e.g., writing a survey)? Start with Workflow 1 → /idea-discovery
Have a plan, need to implement and run? Workflow 1.5 → /experiment-bridge
Already have results, need iterative improvement? Workflow 2 → /auto-review-loop
Ready to write the paper? Workflow 3 → /paper-writing (or step by step: /paper-plan → /paper-figure → /paper-write → /paper-compile → /auto-paper-improvement-loop)
Got reviews back? Need to rebuttal? Workflow 4 → /rebuttal — parse reviews, draft safe rebuttal, follow-up rounds
Full pipeline? Workflow 1 → 1.5 → 2 → 3 → submit → 4 → /research-pipeline + /rebuttal — from idea to acceptance
Want ARIS to remember and learn? 📚 /research-wiki init — persistent memory across sessions. Papers, ideas, failed experiments compound over time
Want ARIS to improve itself? Workflow M → /meta-optimize — analyze usage logs, propose skill improvements, reviewer-gated

⚠️ Important: These tools accelerate research, but they don't replace your own critical thinking. Always review generated ideas with your domain expertise, question the assumptions, and make the final call yourself. The best research comes from human insight + AI execution, not full autopilot.

Full Pipeline 🚀

/research-lit → /idea-creator → /novelty-check → /research-refine → /experiment-bridge → /auto-review-loop → /paper-writing → submit → /rebuttal → accept! 🎉
  (survey)      (brainstorm)    (verify novel)   (refine method)   (implement+deploy)  (review & fix)      (write paper)   (send)   (reply to reviewers)
  ├────────────── Workflow 1: Idea Discovery ──────────────┤ ├ Workflow 1.5 ─┤ ├── Workflow 2 ──┤ ├── Workflow 3 ──┤         ├── Workflow 4 ──┤

                                     📚 research-wiki (persistent memory — papers, ideas, experiments, claims)
                                        ↕ reads before ideation, writes after every stage, failed ideas = anti-repetition memory

                                              /meta-optimize (Workflow M — runs independently, improves ARIS itself)
                                                 ↑ reads .aris/meta/events.jsonl (accumulated from all runs above)

📝 Blog post: 梦中科研全流程开源

Workflow 1: Idea Discovery & Method Refinement 🔍

"What's the state of the art? Where are the gaps? How do we solve it?"

Don't have a concrete idea yet? Just give a research direction — /idea-discovery handles the rest:

📚 Survey the landscape (recent papers, open problems, recurring limitations)
🧠 Brainstorm 8-12 concrete ideas via GPT-5.4 xhigh
🔍 Filter by feasibility, compute cost, and quick novelty search
🛡️ Validate top ideas with deep novelty check + devil's advocate review
🧪 Pilot top 2-3 ideas in parallel on different GPUs (30 min - 2 hr each)
🏆 Rank by empirical signal — ideas with positive pilot results rise to the top
🔬 Refine the top idea into a problem-anchored proposal via iterative GPT-5.4 review
🧪 Plan claim-driven experiments with ablations, budgets, and run order

The output is a ranked IDEA_REPORT.md plus a refined proposal (refine-logs/FINAL_PROPOSAL.md) and experiment plan (refine-logs/EXPERIMENT_PLAN.md) for the top idea. Dead-end ideas are documented too, saving future exploration.

┌─────────────────────────────────────────────────────────────────┐
│              Idea Discovery & Method Refinement                  │
│                                                                  │
│   /research-lit    /idea-creator    /novelty-check               │
│   (find papers)    (brainstorm)     (verify novelty)             │
│         │               │                │                       │
│         ▼               ▼                ▼                       │
│   ┌──────────┐    ┌──────────┐     ┌──────────┐                │
│   │ Scan     │───▶│ Generate │────▶│ Check if │                │
│   │ local    │    │ 8-12     │     │ idea is  │                │
│   │ papers + │    │ ideas    │     │ novel    │                │
│   │ search   │    │ + rank   │     │          │                │
│   └──────────┘    └──────────┘     └──────────┘                │
│                         │                │                       │
│                         ▼                ▼                       │
│                   ┌──────────┐     ┌──────────┐                │
│                   │ Filter   │────▶│ External │                │
│                   │ by cost, │     │ LLM      │                │
│                   │ novelty  │     │ evaluates│                │
│                   └──────────┘     └──────────┘                │
│                                          │                       │
│                   /research-refine       ▼                       │
│                   (refine method)   ┌──────────┐                │
│                         │          │ Freeze   │                │
│                         ▼          │ problem  │                │
│                   ┌──────────┐     │ anchor + │                │
│                   │ Iterate  │◀───▶│ refine   │                │
│                   │ until    │     │ method   │                │
│                   │ score≥9  │     └──────────┘                │
│                   └──────────┘          │                       │
│                         │               ▼                       │
│                   /experiment-plan  ┌──────────┐                │
│                         │          │ Claim-   │                │
│                         ▼          │ driven   │                │
│                   ┌──────────┐     │ experiment│               │
│                   │ Plan     │────▶│ roadmap  │                │
│                   │ runs     │     └──────────┘                │
│                   └──────────┘                                  │
│                                                                  │
│   Typical flow:                                                  │
│   1. /research-lit "discrete diffusion models"                   │
│   2. /idea-creator "DLLMs post training"                         │
│   3. Review ranked ideas, pick top 2-3                           │
│   4. /novelty-check "top idea" (deep verification)               │
│   5. /research-review "top idea" (critical feedback)             │
│   6. /research-refine "top idea" (problem anchor + method)       │
│   7. /experiment-plan (claim-driven roadmap)                     │
│   8. /run-experiment → /auto-review-loop                         │
└─────────────────────────────────────────────────────────────────┘

Skills involved: research-lit + idea-creator + novelty-check + research-review + research-refine-pipeline

💡 One-command shortcut: /idea-discovery "your research direction" runs this entire workflow automatically.

🔄 Human-in-the-loop: Each phase presents results and waits for your feedback. Not happy? Tell it what's missing — it refines the prompt and regenerates. Trust the defaults? It auto-proceeds with the top-ranked option. You decide how hands-on to be.

⚙️ Pilot experiment budgets (max hours, timeout, GPU budget) are configurable — see Customization.

📝 Blog post: Claude Code 两月 NeurIPS 指北

Workflow 1.5: Experiment Bridge 🔗

"I have a plan. Now implement it, deploy it, and get me initial results."

Already have an experiment plan (from Workflow 1 or your own)? /experiment-bridge turns it into running code:

📋 Parse the experiment plan (refine-logs/EXPERIMENT_PLAN.md)
💻 Implement experiment scripts (reuse existing code, add proper argparse/logging/seeds)
🔍 GPT-5.4 code review — cross-model review catches logic bugs before wasting GPU hours (code review: true by default)
✅ Sanity check — run the smallest experiment first to catch runtime bugs
🚀 Deploy full experiment suite to GPU via /run-experiment
📊 Collect initial results and update the experiment tracker

┌─────────────────────────────────────────────────────────────────┐
│                Workflow 1.5: Experiment Bridge                    │
│                                                                  │
│   EXPERIMENT_PLAN.md                                             │
│         │                                                        │
│         ▼                                                        │
│   ┌──────────┐     ┌──────────┐     ┌──────────┐               │
│   │ Claude   │────▶│ GPT-5.4  │────▶│ Sanity   │               │
│   │ Code     │     │ xhigh    │     │ Check    │               │
│   │ writes   │     │ reviews  │     │ (1 GPU)  │               │
│   │ code     │     │ code     │     │          │               │
│   └──────────┘     └──────────┘     └──────────┘               │
│                                          │                       │
│                                          ▼                       │
│   ┌──────────┐     ┌──────────┐     ┌──────────┐               │
│   │ Collect  │◀────│ Monitor  │◀────│ Deploy   │               │
│   │ results  │     │ progress │     │ to GPUs  │               │
│   │          │     │ (+ W&B)  │     │          │               │
│   └──────────┘     └──────────┘     └──────────┘               │
│         │                                                        │
│         ▼                                                        │
│   Ready for /auto-review-loop                                    │
└─────────────────────────────────────────────────────────────────┘

Skills involved: experiment-bridge + run-experiment + monitor-experiment

💡 One-command shortcut: /experiment-bridge reads refine-logs/EXPERIMENT_PLAN.md automatically. Or point it to any plan: /experiment-bridge "my_plan.md".

⚙️ CODE_REVIEW, AUTO_DEPLOY, SANITY_FIRST, MAX_PARALLEL_RUNS are configurable — see Customization.

Workflow 2: Auto Research Loop 🔁 (sleep & wake up to results)

"Review my paper, fix what's wrong, repeat until it's good."

GPT-5.4 reviews → identifies weaknesses → suggests experiments → Claude Code writes scripts, deploys to GPU, monitors results, rewrites the paper — all while you sleep. Just add your GPU server config to CLAUDE.md.

┌─────────────────────────────────────────────────────────────┐
│                    Auto Review Loop                          │
│                                                              │
│   /research-review          /auto-review-loop                │
│   (single deep review)      (autonomous loop)                │
│         │                         │                          │
│         ▼                         ▼                          │
│   ┌──────────┐   ┌──────────┐   ┌──────────┐               │
│   │ External  │──▶│ Implement│──▶│ Monitor  │──▶ repeat     │
│   │ LLM      │   │ fixes    │   │ results  │    until       │
│   │ reviews  │   │ & run    │   │          │    score ≥ 6   │
│   └──────────┘   │ experiments│  └──────────┘               │
│                   └──────────┘                               │
│                                                              │
│   When reviewer suggests a new method direction:             │
│   /novelty-check — verify idea isn't already published       │
│                                                              │
│   Supporting skills:                                         │
│   /run-experiment    — deploy to local/remote/vast.ai GPU     │
│   /analyze-results   — interpret experiment outputs          │
│   /monitor-experiment — check progress, collect results      │
└─────────────────────────────────────────────────────────────┘

Skills involved: auto-review-loop + research-review + novelty-check + run-experiment + analyze-results + monitor-experiment

💡 One-command shortcut: /auto-review-loop "your paper topic" runs this entire workflow automatically.

What to pass as argument? A short topic or scope is enough — the skill automatically reads your project's narrative docs (NARRATIVE_REPORT.md), memory files, experiment results, and prior reviews to build the full context for GPT-5.4. Examples:

/auto-review-loop "factorized gap in discrete diffusion LMs" — broad topic, skill finds everything

/auto-review-loop "focus on Section 3-5, our CRF results are weak" — targeted scope with hints

/auto-review-loop — also works: skill reads project files and infers the topic

🎮 Reviewer Difficulty — control how adversarial the reviewer is:

Level	What changes	Use when
`medium` (default)	Standard MCP review — same as before	Normal workflow
`hard`	+ Reviewer Memory (GPT tracks suspicions across rounds) + Debate Protocol (Claude rebuts, GPT rules)	Want tougher feedback
`nightmare`	+ GPT reads repo directly via `codex exec` (Claude can't filter what it sees) + adversarial verification	Preparing for top venue, want maximum stress test

/auto-review-loop "topic" — difficulty: nightmare    # GPT reads your code and verifies claims itself

🛡️ Key safety features:

🔒 MAX_ROUNDS = 4 — prevents infinite loops; stops early if score threshold is met
⏱️ > 4 GPU-hour experiments skipped — won't launch massive jobs; flags them for manual follow-up
🧠 Prefer reframing over new experiments — when both can address a weakness, chooses the cheaper path
🪞 No hiding weaknesses — explicit rule: "Do NOT hide weaknesses to game a positive score"
🔧 Fix before re-review — must actually implement fixes before resubmitting; no empty promises
💾 Compact recovery — persists state (REVIEW_STATE.json) after each round. If the context window fills up and auto-compacts mid-loop, the workflow reads the state file and resumes from where it left off — no human intervention needed

⚙️ MAX_ROUNDS, score threshold, and GPU limits are configurable — see Customization.

📝 Blog post: 开源 | 睡觉 Claude 自动跑实验改文

Workflow 3: Paper Writing Pipeline 📝

"Turn my research narrative into a submission-ready PDF." Requires a local LaTeX environment — see Prerequisites.

┌─────────────────────────────────────────────────────────────┐
│                   Paper Writing Pipeline                      │
│                                                               │
│   /paper-plan      /paper-figure     /paper-write             │
│   (outline)        (plots & tables)  (LaTeX draft)            │
│        │                │                 │                   │
│        ▼                ▼                 ▼                   │
│   ┌──────────┐    ┌──────────┐     ┌──────────┐              │
│   │ Claims-  │───▶│ Generate │────▶│ Section  │──┐           │
│   │ Evidence │    │ figures, │     │ by       │  │           │
│   │ Matrix + │    │ tables,  │     │ section  │  │           │
│   │ Section  │    │ LaTeX    │     │ LaTeX    │  │           │
│   │ Plan     │    │ includes │     │ draft    │  │           │
│   └──────────┘    └──────────┘     └──────────┘  │           │
│        │                                          │           │
│        │         /paper-compile                   │           │
│        │         (build PDF)                      │           │
│        │              │                           │           │
│        ▼              ▼                           ▼           │
│   ┌──────────────────────────────────────────────────┐       │
│   │ NARRATIVE_REPORT.md ──► PAPER_PLAN.md ──► paper/ │       │
│   │    (input)             (outline)      (LaTeX+PDF)│       │
│   └──────────────────────────────────────────────────┘       │
│                                                               │
│   Typical flow:                                               │
│   1. Write NARRATIVE_REPORT.md (from Workflow 2 results)      │
│   2. /paper-plan (claims-evidence matrix + section plan)      │
│   3. /paper-figure (comparison tables, training curves, etc.) │
│   4. /paper-write (section-by-section LaTeX generation)       │
│   5. /paper-compile (build PDF, fix errors, page check)       │
│   6. /auto-paper-improvement-loop (review ×2 + format check)  │
└─────────────────────────────────────────────────────────────┘

Skills involved: paper-plan + paper-figure + paper-write + paper-compile + auto-paper-improvement-loop + (post-acceptance) paper-poster + paper-slides

One-command shortcut: /paper-writing "NARRATIVE_REPORT.md" runs this entire workflow automatically.

Input: A NARRATIVE_REPORT.md describing the research: claims, experiments, results, figures. The more detailed the narrative (especially figure descriptions and quantitative results), the better the output. See templates/NARRATIVE_REPORT_TEMPLATE.md for a complete example.

Output: A submission-ready paper/ directory with LaTeX source, clean .bib (only cited entries), and compiled PDF.

Key features:

📐 Claims-Evidence Matrix — every claim maps to evidence, every experiment supports a claim
📊 Auto figure generation — line plots, bar charts, comparison tables from JSON data
🧹 Clean bib — automated filtering removes uncited entries (948→215 lines in testing). Real BibTeX from DBLP/CrossRef instead of LLM-generated entries
📄 Flexible sections — 5-8 sections depending on paper type (theory papers often need 7)
🔍 GPT-5.4 review — each step optionally reviewed by external LLM
✂️ De-AI polish — removes AI writing patterns (delve, pivotal, landscape...)
🎯 Page verification — pdftotext-based precise check that main body fits page limit

⚠️ Figure generation scope: /paper-figure auto-generates data-driven plots (training curves, bar charts, heatmaps) and comparison tables from JSON/CSV. For architecture diagrams and method figures: illustration: gemini (default) uses Claude→Gemini→Nano Banana Pro for publication-quality diagrams; illustration: mermaid generates Mermaid diagrams for free; illustration: false skips AI figures entirely.

Gemini API setup (for illustration: gemini): Get your API key at Google AI Studio, then set it as an environment variable: export GEMINI_API_KEY="your-key". Or add to your shell profile (~/.zshrc / ~/.bashrc). No other dependencies needed.

Tested end-to-end: Generated a 9-page ICLR 2026 theory paper (7 sections, 29 citations, 4 figures, 2 comparison tables) from a single NARRATIVE_REPORT.md — zero compilation errors, zero undefined references.

Auto Paper Improvement Loop ✨

After Workflow 3 generates the paper, /auto-paper-improvement-loop runs 2 rounds of GPT-5.4 xhigh content review → fix → recompile, plus a final format compliance check, autonomously polishing the paper from rough draft to submission-ready.

Score Progression (Real Test — ICLR 2026 theory paper):

Round	Score	Key Changes
Round 0	4/10 (content)	Baseline
Round 1	6/10 (content)	Fixed assumptions, softened claims, renamed notation
Round 2	7/10 (content)	Added synthetic validation, stronger limitations
Round 3	5→8.5/10 (format)	Removed hero fig, appendix, compressed conclusion, float spacing

Final: 8 pages main body (ICLR limit: 9), 0 overfull hbox, ICLR-compliant. +4.5 points across 3 rounds.

Round 1 fixes (6 items)

CRITICAL — Assumption-model mismatch: A boundedness assumption contradicted the model's distributional family. Replaced with a tail-compatible assumption and added formal truncation bridge.
CRITICAL — Theory-practice gap: Theory assumes idealized encoders, experiments use learned nonlinear encoders. Softened "validate" → "demonstrate practical relevance" and added explicit disclaimer.
MAJOR — Missing quantitative metrics: Added parameter count table (latent vs total) with honest accounting of system cost.
MAJOR — Theorem not self-contained: Added "Interpretation" paragraph listing all dependencies explicitly.
MAJOR — Overclaim in novelty statement: Scoped a broad "first convergence guarantee" to precise conditions under which it holds.
MAJOR — Notation confusion: Renamed a symbol that clashed with another key variable. Added Notation paragraph.

Round 2 fixes (4 items)

MAJOR — Missing theory-aligned experiments: Added a synthetic validation subsection directly testing the two main theoretical predictions under controlled conditions.
MAJOR — Overclaim softening: Replaced strong equivalence claims with appropriately hedged language across all files.
MAJOR — Informal theoretical argument: Formalized an informal justification into a proper proposition with explicit error bounds.
MINOR — Weak limitations: Expanded to explicitly list all assumptions and acknowledge missing standard evaluations.

Round 3 format fixes (8 items)

Removed hero figure block (saved ~0.7 pages)
Compressed conclusion from 15→9 lines
Moved synthetic validation to Appendix A
Moved comparison tables to Appendix B
Fixed overfull hbox (85pt) with \resizebox
Added compact float spacing (\captionsetup, \textfloatsep)
Inlined centered question block in introduction
Tightened itemize environments

Workflow 4: Rebuttal 📝 (reply to reviewers safely)

"Reviews are in. Help me draft a safe, grounded rebuttal."

Got reviews back? /rebuttal parses them, builds a strategy, and drafts a venue-compliant response:

📋 Parse — normalize reviews, validate venue rules (character limit, text-only, etc.)
🔍 Atomize — split each review into issue cards (type, severity, reviewer stance)
🗺️ Strategize — global themes, per-reviewer priorities, character budget, blocked claims
🧪 Evidence sprint — if auto experiment: true, auto-run supplementary experiments via /experiment-bridge
✍️ Draft — global opener + numbered per-reviewer responses + closing for meta-reviewer
🛡️ Safety check — 6 lints: coverage, provenance, commitment, tone, consistency, limit
🔬 GPT-5.4 stress test — internal skeptical review of the draft
📄 Finalize — two outputs: PASTE_READY.txt (exact character count) + REBUTTAL_DRAFT_rich.md (extended version for manual editing)
🔄 Follow-up rounds — delta replies for reviewer discussions, technically escalating

┌─────────────────────────────────────────────────────────────────┐
│                   Workflow 4: Rebuttal                            │
│                                                                  │
│   Reviews arrive                                                 │
│         │                                                        │
│         ▼                                                        │
│   ┌──────────┐     ┌──────────┐     ┌──────────┐               │
│   │ Parse &  │────▶│ Strategy │────▶│ Evidence  │               │
│   │ atomize  │     │ plan     │     │ sprint    │               │
│   │ reviews  │     │          │     │ (optional)│               │
│   └──────────┘     └──────────┘     └──────────┘               │
│                                          │                       │
│                                          ▼                       │
│   ┌──────────┐     ┌──────────┐     ┌──────────┐               │
│   │ Finalize │◀────│ GPT-5.4  │◀────│ Draft    │               │
│   │ 2 versions│    │ stress   │     │ rebuttal │               │
│   │          │     │ test     │     │          │               │
│   └──────────┘     └──────────┘     └──────────┘               │
│         │                                                        │
│         ▼                                                        │
│   PASTE_READY.txt (strict) + RICH.md (extended)                  │
│         │                                                        │
│         ▼                                                        │
│   Follow-up rounds (delta replies, per-reviewer threads)         │
└─────────────────────────────────────────────────────────────────┘

Skills involved: rebuttal

💡 Quick mode: /rebuttal — quick mode: true stops after parsing + strategy (Phase 0-3). See what reviewers want before committing to a full draft.

⚙️ VENUE, AUTO_EXPERIMENT, QUICK_MODE, MAX_STRESS_TEST_ROUNDS are configurable — see Customization.

Three safety gates — rebuttal will NOT finalize if any fails:

🔒 Provenance — every claim maps to paper/review/user-confirmed result. No fabrication.
🔒 Commitment — every promise is user-approved. No overpromising.
🔒 Coverage — every reviewer concern is tracked. Nothing disappears.

📚 Research Wiki — Persistent Research Memory

"Stop re-deriving. Start compounding." — inspired by Karpathy's LLM Wiki

Without the wiki, ARIS is stateless — every /idea-discovery starts from scratch. With the wiki, ARIS accumulates knowledge across the entire research lifecycle: papers read, ideas tested, experiments run, claims verified or invalidated.

The key insight: failed ideas are the most valuable memory. A researcher who knows what doesn't work generates better ideas than one starting from zero.

Setup:

> /research-wiki init     # one-time, creates research-wiki/ in your project

That's it. Once initialized, the wiki works automatically:

When	What happens	Wiki action
`/research-lit` finds papers	Papers auto-ingested	`papers/<slug>.md` created, edges added, query_pack rebuilt
`/idea-creator` runs	Reads wiki first	Failed ideas = banlist, gaps = search seeds, papers = known prior work
`/idea-creator` finishes	ALL ideas written back	Both recommended AND eliminated ideas → `ideas/<id>.md`
`/result-to-claim` judges	Results written back	Experiment page created, claim status updated (supported/invalidated)
3+ ideas fail	Re-ideation suggested	"💡 Consider re-running /idea-creator — the wiki now knows what doesn't work"

Four entity types:

Entity	What it stores	Example
📄 Paper	Structured summary: thesis, method, limitations, reusable ingredients	`paper:chen2025_factorized_gap`
💡 Idea	Hypothesis, status (proposed/failed/succeeded), failure notes, lessons	`idea:001`
🧪 Experiment	Metrics, verdict, hardware, duration	`exp:001`
📋 Claim	Testable statement + evidence status (reported/supported/invalidated)	`claim:C1`

Typed relationships (stored in graph/edges.jsonl):

paper --extends--> paper              idea --inspired_by--> paper
paper --contradicts--> paper          idea --tested_by--> experiment
paper --addresses_gap--> gap          experiment --supports--> claim
paper --supersedes--> paper           experiment --invalidates--> claim

Spiral learning in action:

Round 1: read 15 papers → wiki remembers → idea A → experiment → FAIL
         wiki records: "A fails because OOM at batch>32, loss diverges"

Round 2: /idea-creator reads wiki → sees A failed → generates idea D (avoids A's trap)
         → experiment → PARTIAL SUCCESS
         wiki records: "D works on small models, fails on large"

Round 3: /idea-creator reads wiki → knows A failed + D partial → generates idea F
         (combines D's success with new approach) → experiment → SUCCESS 🎉

Subcommands:

/research-wiki init                              # initialize wiki
/research-wiki ingest "paper title" — arxiv: xxx  # manually add a paper
/research-wiki query "topic"                      # rebuild query_pack.md
/research-wiki update idea:001 — outcome: negative # update entity
/research-wiki lint                               # health check (orphans, contradictions, stale claims)
/research-wiki stats                              # overview (paper/idea/experiment/claim counts)

🔒 Safe by design: All workflow hooks are guarded by if research-wiki/ exists. No wiki = no impact. Zero dependencies (pure Python stdlib). You choose when to enable it.

Workflow M: Meta-Optimize 🧬 (ARIS optimizes itself)

"Analyze my usage patterns and improve your own skills."

Unlike Workflows 1–4 which optimize research artifacts (papers, code, experiments), Workflow M optimizes the harness itself — the SKILL.md instructions, default parameters, and convergence rules that govern how ARIS operates. Inspired by Meta-Harness (Lee et al., 2026).

Setup (one-time, in normal terminal):

mkdir -p .claude .aris/meta tools/meta_opt
cp Auto-claude-code-research-in-sleep/templates/claude-hooks/meta_logging.json .claude/settings.json
cp Auto-claude-code-research-in-sleep/tools/meta_opt/*.sh tools/meta_opt/
chmod +x tools/meta_opt/*.sh
claude   # hooks active immediately

Usage (after 5+ workflow runs):

> /meta-optimize                        # analyze current project
> /meta-optimize "auto-review-loop"     # focus on one skill
> /meta-optimize --global               # analyze trends across ALL projects
> /meta-optimize apply 1                # apply recommended change #1

How it works:

📊 Passive logging — Claude Code hooks silently record every skill invocation, tool call, failure, parameter override, and user prompt. Events are written to both project-level (.aris/meta/events.jsonl) and global (~/.aris/meta/events.jsonl, with a "project" tag) logs. Zero user effort.
🔍 Pattern analysis — /meta-optimize reads the log and identifies:
- Parameters users override most often (bad defaults)
- Tools that fail repeatedly in specific skills (missing error handling)
- Review score plateaus (convergence rules too loose/tight)
- Manual corrections users make (skill gaps)
🩹 Patch proposal — generates minimal diffs to target SKILL.md files with data-backed justifications
🔬 Reviewer gate — GPT-5.4 xhigh reviews each patch: does the evidence support it? could it hurt other users?
✅ User approval — only applied with explicit user consent. All changes are logged and reversible.

┌─────────────────────────────────────────────────────────────────┐
│                  Workflow M: Meta-Optimize                        │
│                                                                  │
│   Normal ARIS usage (W1-W4)                                      │
│         │ (hooks log events passively)                           │
│         ▼                                                        │
│   .aris/meta/events.jsonl                                        │
│         │                                                        │
│         ▼                                                        │
│   ┌──────────┐     ┌──────────┐     ┌──────────┐               │
│   │ Analyze  │────▶│ Propose  │────▶│ GPT-5.4  │               │
│   │ patterns │     │ SKILL.md │     │ reviews  │               │
│   │          │     │ patches  │     │ patch    │               │
│   └──────────┘     └──────────┘     └──────────┘               │
│                                          │                       │
│                                          ▼                       │
│                                    User approves?                 │
│                                     Yes → Apply                  │
│                                     No  → Skip                   │
└─────────────────────────────────────────────────────────────────┘

What gets optimized (harness components):

Component	Example
Skill prompts	Reviewer instructions, quality gates, step descriptions
Default parameters	`difficulty`, `MAX_ROUNDS`, `threshold`
Convergence rules	When to stop the review loop, retry counts
Error handling	Auto-debug patterns, failure recovery steps

What does NOT get optimized: research artifacts (papers, code, experiments) — that's what W1–W4 do.

Skills involved: meta-optimize

💡 This is a maintenance workflow, not part of the W1→W1.5→W2→W3→W4 research pipeline. Run it periodically, like git gc for your research harness.

⚡ Effort Levels

"How hard should ARIS work?" — Every skill accepts — effort: lite | balanced | max | beast.

Level	Tokens	Best for	What changes
`lite`	~0.4x	Quick exploration, budget users	Fewer papers, ideas, rounds. Minimum viable depth
`balanced`	1x	Normal workflow (default)	Current ARIS behavior. Zero change for existing users
`max`	~2.5x	Serious submission prep	More papers, deeper review, more ablations
`beast`	~5-8x	Top-venue final sprint	Every knob to maximum. No budget limit

What NEVER changes regardless of effort:

Codex reasoning: always xhigh (reviewer quality is non-negotiable)
DBLP/CrossRef citations: always on
Reviewer independence: always on
Experiment integrity: always on

# Every skill accepts effort independently
/research-lit "topic" — effort: beast              # 40-50 papers, 15+ queries
/idea-creator "direction" — effort: lite           # 4-6 ideas, quick filter
/auto-review-loop — effort: max                    # 6 rounds, 4-6 fixes/round

# Mix with specific overrides
/auto-review-loop — effort: beast, review_rounds: 3  # beast everything, but cap at 3 rounds

# Full pipeline
/research-pipeline "your topic" — effort: beast    # top-venue sprint mode

Full effort comparison table — click to expand

Skill	Dimension	lite	balanced	max	beast
research-lit	papers	6-8	10-15	18-25	40-50
idea-creator	ideas	4-6	8-12	12-16	20-30
idea-creator	pilots	1-2	2-3	3-4	5-6
novelty-check	claims	2-3	3-4	4-6	all
research-refine	rounds	3	5	7	10+
experiment-plan	experiments	3	5	7	10+
experiment-plan	seeds	1	3	5	5
auto-review-loop	rounds	2	3-4	6	8+
paper-improvement	rounds	1	2	3	5
paper-illustration	iterations	2	3	5	7
rebuttal	stress tests	0-1	1	2	3
experiment-audit	depth	skip	basic	full	line-by-line

📖 Full specification: shared-references/effort-contract.md

🧿 Optional: GPT-5.4 Pro via Oracle

For expert researchers who want the strongest possible reviewer.

Oracle unlocks GPT-5.4 Pro as an ARIS reviewer — the strongest reasoning model available. Pro excels at deep mathematical proof verification, line-by-line code auditing, and complex experimental design critique.

Setup:

# 1. Install Oracle
npm install -g @steipete/oracle

# 2. Add Oracle MCP to Claude Code
claude mcp add oracle -s user -- oracle-mcp

# 3. Restart Claude Code

# 4a. API mode (fast, recommended):
export OPENAI_API_KEY="your-key"

# 4b. Browser mode (free, no API key — log in to ChatGPT in Chrome):
# Just open Chrome → chatgpt.com → log in

Usage — add — reviewer: oracle-pro to any skill:

/research-review "my draft" — reviewer: oracle-pro          # Pro-level paper critique
/proof-checker "paper/" — reviewer: oracle-pro              # deepest mathematical verification
/experiment-audit — reviewer: oracle-pro                    # Pro audits your eval code
/auto-review-loop "scope" — reviewer: oracle-pro            # Pro stress test each round
/idea-creator "direction" — reviewer: oracle-pro            # Pro evaluates your ideas
/rebuttal "paper/ + reviews" — reviewer: oracle-pro         # Pro stress tests your rebuttal

Default is always Codex xhigh. Oracle not installed = zero impact. — reviewer: oracle-pro without Oracle installed = graceful fallback to Codex + warning.

📖 Full specification: shared-references/reviewer-routing.md

🧰 All Skills

🚀 Full Pipeline

Skill	Description	Codex MCP?
🏗️ `research-pipeline`	End-to-end: Workflow 1 → 1.5 → 2 → 3, from research direction to submission	Yes

🔍 Workflow 1: Idea Discovery & Method Refinement

Skill	Description	Codex MCP?
🔭 `idea-discovery`	Pipeline orchestrator — runs all skills below in sequence	Yes
├ 📚 `research-lit`	Multi-source literature search (Zotero + Obsidian + local PDFs + arXiv API + web)	No
├ 💡 `idea-creator`	Brainstorm 8-12 ideas, filter by feasibility, pilot on GPU, rank by signal	Yes
├ 🔍 `novelty-check`	Verify idea novelty against recent literature (multi-source + GPT-5.4 cross-check)	Yes
├ 🔬 `research-review`	Single-round deep review from external LLM (xhigh reasoning)	Yes
└ 🧭 `research-refine-pipeline`	Refine method + plan experiments in one chain	Yes
├ 🔬 `research-refine`	Problem anchor → iterative method refinement (up to 5 rounds, score ≥ 9)	Yes
└ 🧪 `experiment-plan`	Claim-driven experiment roadmap with ablations, budgets, and run order	No

🔗 Workflow 1.5: Experiment Bridge

Skill	Description	Codex MCP?
🔗 `experiment-bridge`	Read experiment plan → implement code → sanity check → deploy to GPU → collect initial results	No
├ 🚀 `run-experiment`	Deploy experiments to local, remote, or Vast.ai GPU (`gpu: local/remote/vast`)	No
├ 👀 `monitor-experiment`	Monitor running experiments, check progress, collect results	No
└ ☁️ `vast-gpu`	Rent, manage, and destroy on-demand GPU instances from Vast.ai	No

🔁 Workflow 2: Auto Research Loop

Skill	Description	Codex MCP?
🔁 `auto-review-loop`	Pipeline orchestrator — autonomous review→fix→re-review (max 4 rounds)	Yes
├ 🔬 `research-review`	Deep review from external LLM (shared with Workflow 1)	Yes
├ 🔍 `novelty-check`	Verify novelty when reviewer suggests new directions	Yes
├ 🚀 `run-experiment`	Deploy experiments to local, remote, or Vast.ai GPU (`gpu: local/remote/vast`)	No
├ 📊 `analyze-results`	Analyze experiment results, compute statistics, generate insights	No
└ 👀 `monitor-experiment`	Monitor running experiments, check progress, collect results	No
🔁 `auto-review-loop-llm`	Same as above, but uses any OpenAI-compatible API via `llm-chat` MCP server	No

📝 Workflow 3: Paper Writing

Skill	Description	Codex MCP?
📝 `paper-writing`	Pipeline orchestrator — runs all skills below in sequence	Yes
├ 📐 `paper-plan`	Claims-evidence matrix, section structure, figure plan, citation scaffolding	Yes
├ 📊 `paper-figure`	Publication-quality matplotlib/seaborn plots + LaTeX comparison tables	Optional
├ 🎨 `paper-illustration`	AI-generated architecture diagrams and method figures via Gemini (when `illustration: true`)	No (needs Gemini API)
├ ✍️ `paper-write`	Section-by-section LaTeX generation (ICLR/NeurIPS/ICML). Anti-hallucination BibTeX via DBLP/CrossRef	Yes
├ 🔨 `paper-compile`	Compile LaTeX to PDF, auto-fix errors, submission readiness checks	No
└ 🔄 `auto-paper-improvement-loop`	2-round content review + format check (4/10 → 8.5/10)	Yes

📝 Workflow 4: Rebuttal

Skill	Description	Codex MCP?
📝 `rebuttal`	Parse reviews → atomize → strategy → draft → safety check → stress test → finalize (2 versions) → follow-up	Yes

🛠️ Standalone / Utility

Skill	Description	Codex MCP?
📄 `arxiv`	Search, download, and summarize arXiv papers. Standalone or `/research-lit` supplement	No
🔎 `semantic-scholar`	Search published venue papers (IEEE, ACM, Springer) via Semantic Scholar API. Citation counts, venue metadata, TLDR	No
📚 `deepxiv`	Progressive paper retrieval via DeepXiv CLI: search, brief, section map, section reads, trending, web search	Yes (`pip install deepxiv-sdk`)
🔎 `exa-search`	AI-powered broad web search via Exa: blogs, docs, news, companies, research papers with content extraction (highlights, text, summaries)	Yes (`pip install exa-py`)
📝 `alphaxiv`	Quick single-paper lookup via AlphaXiv LLM-optimized summaries. Three-tier fallback: overview → full markdown → LaTeX source	No
🎨 `pixel-art`	Generate pixel art SVG illustrations for READMEs, docs, or slides	No
📱 `feishu-notify`	Feishu/Lark push (webhook) or interactive (bidirectional). Off by default	No

⚙️ Setup

Prerequisites

Claude Code installed

(For review skills) Codex CLI installed and configured as MCP server:

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

(For Workflow 3: paper writing) LaTeX environment with latexmk and pdfinfo:

# macOS
brew install --cask mactex    # or: brew install basictex
brew install poppler          # provides pdfinfo

# Ubuntu/Debian
sudo apt install texlive-full latexmk poppler-utils

# Verify
latexmk --version && pdfinfo -v

If you only need Workflow 1 & 2 (idea discovery + auto review), LaTeX is not required.

Install Skills

💡 New Claude Code versions may not auto-create ~/.claude/skills/. If the directory doesn't exist, create it first: mkdir -p ~/.claude/skills/. This works alongside the new plugins system — both skills/ and plugins/ are loaded by Claude Code.

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

# Create skills directory if it doesn't exist
mkdir -p ~/.claude/skills/

# Install all skills globally
cp -r skills/* ~/.claude/skills/

# Or install specific skills
cp -r skills/auto-review-loop ~/.claude/skills/
cp -r skills/research-lit ~/.claude/skills/

Optional: Codex Plugin for Code Review

codex-plugin-cc provides additional Codex capabilities that ARIS auto-detects when installed:

# In Claude Code:
/plugin marketplace add openai/codex-plugin-cc
/plugin install codex@openai-codex
/reload-plugins
/codex:setup

Where ARIS uses the plugin:

Skill	Command	What it does
`/codex:review`	Workflow 1.5	Review experiment code before GPU deployment
`/codex:adversarial-review`	Workflow 1.5	Adversarial code review (find edge cases, bugs)
`/codex:rescue`	Workflow 1.5 + 3	Auto-debug rescue — when experiment or LaTeX compilation fails after 2 attempts, Codex independently diagnoses the root cause before the next retry

All plugin features are optional — if not installed, ARIS falls back to Claude's own diagnosis. The plugin just adds a second pair of eyes.

Note: ARIS's core cross-model review (paper scoring, idea evaluation, rebuttal stress test) still uses Codex MCP, which allows custom prompts. The plugin cannot replace this.

Update Skills

cd Auto-claude-code-research-in-sleep
git pull

# 🧠 Smart update (recommended) — analyzes what's safe to update
bash tools/smart_update.sh          # dry-run: shows what would change
bash tools/smart_update.sh --apply  # apply: adds new + updates safe ones

# Manual options (if you prefer):
# cp -r skills/* ~/.claude/skills/       # Option A: overwrite all
# cp -rn skills/* ~/.claude/skills/      # Option B: only add new, keep yours
# cp -r skills/experiment-bridge ~/.claude/skills/  # Option C: specific skill

💡 Smart update compares your local skills with upstream, detects personal customizations (server paths, API keys, etc.), and only updates skills that are safe to replace. Skills with your personal info are flagged for manual review.

Usage

# Workflow 1: Idea Discovery
> /idea-discovery "your research direction"          # full pipeline
> /research-lit "topic"                              # just literature survey (all sources)
> /research-lit "topic" — sources: zotero, web        # mix and match sources
> /research-lit "topic" — sources: deepxiv            # DeepXiv-only progressive retrieval
> /research-lit "topic" — sources: exa                # Exa AI-powered web search with content extraction
> /research-lit "topic" — arxiv download: true         # also download top arXiv PDFs
> /arxiv "discrete diffusion" — download               # standalone arXiv search + download
> /idea-creator "topic"                              # just brainstorm

# Workflow 2: Auto Research Loop
> /auto-review-loop "your paper topic"               # review → fix → repeat
> /research-review "your paper"                      # single deep review

# Workflow 3: Paper Writing
> /paper-writing "NARRATIVE_REPORT.md"               # full pipeline
> /paper-plan "NARRATIVE_REPORT.md"                  # just outline
> /paper-compile "paper/"                            # just compile

# Full Pipeline
> /research-pipeline "your research direction"       # Workflow 1 → 2 → 3 end-to-end

# Supporting Skills
> /run-experiment train.py --lr 1e-4 --epochs 100
> /analyze-results figures/*.json
> /monitor-experiment server5

🌙 Auto-Allow for Overnight Runs (Optional)

To run the auto-review loop without clicking permission prompts, add to .claude/settings.local.json:

{
  "permissions": {
    "allow": [
      "mcp__codex__codex",
      "mcp__codex__codex-reply",
      "Write",
      "Edit",
      "Skill(auto-review-loop)"
    ]
  }
}

🖥️ GPU Server Setup (For Auto-Experiments)

When GPT-5.4 says "run an ablation study" or "add a baseline comparison", Claude Code automatically writes the experiment script and deploys it to your GPU server. For this to work, Claude Code needs to know your server environment.

Three GPU modes are supported — pick one and add it to your project's CLAUDE.md:

Option A: Remote SSH Server (`gpu: remote`)

## Remote Server
- gpu: remote
- SSH: `ssh my-gpu-server` (key-based auth, no password)
- GPU: 4x A100
- Conda env: `research` (Python 3.10 + PyTorch)
- Activate: `eval "$(/opt/conda/bin/conda shell.bash hook)" && conda activate research`
- Code directory: `/home/user/experiments/`
- Use `screen` for background jobs: `screen -dmS exp0 bash -c '...'`

Claude Code reads this and knows how to SSH in, activate the environment, and launch experiments. GPT-5.4 (the reviewer) only decides what experiments to run — Claude Code figures out how based on your CLAUDE.md.

Option B: Local GPU (`gpu: local`)

If you are already on the GPU server, you can add the following to your CLAUDE.md:

## GPU Environment
- gpu: local
- This machine has direct GPU access (no SSH needed)
- GPU: 4x A100 80GB
- Experiment environment: `YOUR_CONDA_ENV` (Python 3.x + PyTorch)
- Activate before any Python command: `The command to activate your experiment environment` (uv, conda, etc.)
- Code directory: `/home/YOUR_USERNAME/YOUR_CODE_DIRECTORY/`

Option C: Vast.ai On-Demand GPU (`gpu: vast`)

No GPU? Rent one from Vast.ai on demand. ARIS analyzes your training task (model size, dataset, estimated time), searches for the cheapest GPU that fits, and presents options with estimated total cost — not just $/hr. After you pick, it handles everything: rent → setup → run → collect results → destroy.

Prerequisites:

Create a Vast.ai account at https://cloud.vast.ai/ and add billing (credit card or crypto)
Install the vastai CLI (requires Python ≥ 3.10):
```
pip install vastai
```
If your Python is older (check with python --version), use a virtual environment with Python ≥ 3.10 (e.g., conda create, pyenv, uv venv, etc.).
Set your API key — get it from https://cloud.vast.ai/cli/:
```
vastai set api-key YOUR_API_KEY
```
Upload your SSH public key at https://cloud.vast.ai/manage-keys/ — this is required before renting any instance (keys are baked in at creation time). If you don't have one:
```
ssh-keygen -t ed25519 -C "your_email@example.com"
cat ~/.ssh/id_ed25519.pub   # copy this to Vast.ai
```

Verify setup — test that search works:

vastai search offers 'gpu_ram>=24 reliability>0.95' -o 'dph+' --limit 3

Add to CLAUDE.md:

## Vast.ai
- gpu: vast                  # rent on-demand GPU from vast.ai
- auto_destroy: true         # auto-destroy after experiment completes (default)
- max_budget: 5.00           # optional: warn if estimated cost exceeds this

That's it — no GPU model or hardware config needed. When you run /run-experiment, ARIS reads your experiment scripts/plan, estimates VRAM and training time, and presents options like:

| # | GPU       | VRAM  | $/hr  | Est. Hours | Est. Total | Offer ID |
|---|-----------|-------|-------|------------|------------|----------|
| 1 | RTX 4090  | 24 GB | $0.28 | ~4h        | ~$1.12     | 6995713  |  ← best value
| 2 | A100 SXM  | 80 GB | $0.95 | ~2h        | ~$1.90     | 7023456  |  ← fastest

Pick a number and it handles the rest. Use /vast-gpu directly for manual control.

No server at all? The review and rewriting skills still work without GPU access. Only experiment-related fixes will be skipped (flagged for manual follow-up).

📚 Zotero Integration (Optional)

If you use Zotero to manage your paper library, /research-lit can search your collections, read your annotations/highlights, and export BibTeX — all before searching the web.

Recommended: zotero-mcp (1.8k⭐, semantic search, PDF annotations, BibTeX export)

# Install
uv tool install zotero-mcp-server   # or: pip install zotero-mcp-server

# Add to Claude Code (Local API — requires Zotero desktop running)
claude mcp add zotero -s user -- zotero-mcp -e ZOTERO_LOCAL=true

# Or use Web API (works without Zotero running)
claude mcp add zotero -s user -- zotero-mcp \
  -e ZOTERO_API_KEY=your_key -e ZOTERO_USER_ID=your_id

Get your API key at https://www.zotero.org/settings/keys

What it enables in /research-lit:

🔍 Search your Zotero library by topic (including semantic/vector search)
📂 Browse collections and tags
📝 Read your PDF annotations and highlights (what you personally found important)
📄 Export BibTeX for direct use in paper writing

Not using Zotero? No problem — /research-lit automatically skips Zotero and uses local PDFs + web search instead.

📓 Obsidian Integration (Optional)

If you use Obsidian for research notes, /research-lit can search your vault for paper summaries, tagged references, and your own insights.

Recommended: mcpvault (760⭐, no Obsidian app needed, 14 tools, BM25 search)

# Add to Claude Code (point to your vault path)
claude mcp add obsidian-vault -s user -- npx @bitbonsai/mcpvault@latest /path/to/your/vault

Optional complement: obsidian-skills (13.6k⭐, by Obsidian CEO) — teaches Claude to understand Obsidian-specific Markdown (wikilinks, callouts, properties). Copy to your vault:

git clone https://github.com/kepano/obsidian-skills.git
cp -r obsidian-skills/.claude /path/to/your/vault/

What it enables in /research-lit:

🔍 Search your vault for notes on the research topic
🏷️ Find notes by tags (e.g., #paper-review, #diffusion-models)
📝 Read your processed summaries and insights (more valuable than raw papers)
🔗 Follow wikilinks to discover related notes

Not using Obsidian? No problem — /research-lit automatically skips Obsidian and works as before.

💡 Zotero + Obsidian together: Many researchers use Zotero for paper storage and Obsidian for notes. Both integrations work simultaneously — /research-lit checks Zotero first (raw papers + annotations), then Obsidian (your processed notes), then local PDFs, then web search.

arXiv Integration

/research-lit automatically queries the arXiv API for structured metadata (title, abstract, full author list, categories) — richer than web search snippets. No setup required.

By default, only metadata is fetched (no files downloaded). To also download the most relevant PDFs:

/research-lit "topic" — arxiv download: true              # download top 5 PDFs
/research-lit "topic" — arxiv download: true, max download: 10  # download up to 10

For standalone arXiv access, use the dedicated /arxiv skill:

/arxiv "attention mechanism"           # search
/arxiv "2301.07041" — download         # download specific paper

📱 Feishu/Lark Integration (Optional)

Get mobile notifications when experiments finish, reviews score, or checkpoints need your input — without sitting in front of the terminal.

Push Only (group cards)	Interactive (private chat)

Three modes — you choose per-project:

Mode	What happens	You need
Off (default)	Nothing. Pure CLI, no Feishu	Nothing
Push only	Webhook notifications at key events. Mobile push, no reply	Feishu bot webhook URL
Interactive	Full bidirectional. Approve/reject ideas, reply to checkpoints from Feishu	feishu-claude-code running

Push Only Setup (5 min)

Group notifications with rich cards — experiment done, review scored, pipeline complete. Mobile push, no reply needed.

Step 1: Create a Feishu group bot

Open your Feishu group (or create a test group)
Group Settings → Bots → Add Bot → Custom Bot
Name it (e.g., ARIS Notifications), copy the Webhook URL
Security: add custom keyword ARIS (all notifications include this word), or leave unrestricted

Step 2: Create config file

cat > ~/.claude/feishu.json << 'EOF'
{
  "mode": "push",
  "webhook_url": "https://open.feishu.cn/open-apis/bot/v2/hook/YOUR_WEBHOOK_ID"
}
EOF

Step 3: Test it

curl -s -X POST "YOUR_WEBHOOK_URL" \
  -H "Content-Type: application/json" \
  -d '{
    "msg_type": "interactive",
    "card": {
      "header": {"title": {"tag": "plain_text", "content": "🧪 ARIS Test"}, "template": "blue"},
      "elements": [{"tag": "markdown", "content": "Push mode working! 🎉"}]
    }
  }'

You should see a blue card in your group. Skills will now automatically send rich cards at key events:

Event	Card color	Content
Review scored ≥ 6	🟢 Green	Score, verdict, top weaknesses
Review scored < 6	🟠 Orange	Score, verdict, action items
Experiment complete	🟢 Green	Results table, delta vs baseline
Checkpoint waiting	🟡 Yellow	Question, options, context
Error	🔴 Red	Error message, suggested fix
Pipeline done	🟣 Purple	Score progression, deliverables

Interactive Setup (15 min)

Everything Push mode does, plus bidirectional private chat with Claude Code via Feishu. Approve/reject ideas, reply to checkpoints, give custom instructions — all from your phone.

How it works: Push cards go to the group (everyone sees status). Interactive conversations happen in private chat with the bot (you reply, Claude Code acts on it).

Step 1: Complete Push setup above first (you'll keep both)

Step 2: Create a Feishu app on open.feishu.cn

Click Create Enterprise App → name it (e.g., ARIS Claude Bot) → create
Left menu → Add Capabilities → check Bot
Left menu → Permissions → search and enable these 5 permissions:

Permission	Scope	Why
`im:message`	Send & receive messages	Core messaging
`im:message:send_as_bot`	Send as bot	Bot replies
`im:message.group_at_msg:readonly`	Receive group @mentions	Group messages
`im:message.p2p_msg:readonly`	Receive private messages	⚠️ Easy to miss! Without this, the bot connects but never receives your messages
`im:resource`	Access attachments	Images/files

Left menu → Events & Callbacks → select Long Connection mode → add event: im.message.receive_v1 → save

⚠️ Important: The "Long Connection" page may show "未检测到应用连接信息" — this is normal. You need to start the bridge first (Step 3), then come back and save.

Left menu → Version Management → Create Version → fill description → Submit for Review

For personal/test Feishu organizations, approval is usually instant.

Step 3: Deploy the bridge

git clone https://github.com/joewongjc/feishu-claude-code.git
cd feishu-claude-code
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Configure
cp .env.example .env

Edit .env:

FEISHU_APP_ID=cli_your_app_id          # From app credentials page
FEISHU_APP_SECRET=your_app_secret      # From app credentials page
DEFAULT_MODEL=claude-opus-4-6          # ⚠️ Default is sonnet — change to opus for best results
DEFAULT_CWD=/path/to/your/project      # Working directory for Claude Code
PERMISSION_MODE=bypassPermissions      # Or "default" for safer mode

⚠️ Model matters: The default claude-sonnet-4-6 works but may struggle with complex project context. claude-opus-4-6 correctly identified 18 ARIS skills on first try where sonnet could not.

Start the bridge:

python main.py
# Expected output:
# ✅ 连接飞书 WebSocket 长连接（自动重连）...
# [Lark] connected to wss://msg-frontier.feishu.cn/ws/v2?...

For long-running use, put it in a screen session:

screen -dmS feishu-bridge bash -c 'cd /path/to/feishu-claude-code && source .venv/bin/activate && python main.py'

Step 4: Save event config — Go back to Feishu Open Platform → Events & Callbacks → the long connection should now show "已检测到连接" → Save

If you published the app version before the bridge was running, you may need to create a new version (e.g., 1.0.1) and re-publish after saving event config.

Step 5: Test private chat

In Feishu, find the bot in your contacts (search by app name)
Send it a message: 你好
It should reply via Claude Code

If the bot doesn't reply: Send /new to reset the session, then try again. Common issues:

Symptom	Cause	Fix
Bot connects but never receives messages	Missing `im:message.p2p_msg:readonly` permission	Add permission → create new version → publish
Bot replies but doesn't know your project	`DEFAULT_CWD` points to wrong directory	Edit `.env` → restart bridge
Bot replies but seems less capable	Using `claude-sonnet-4-6`	Change to `claude-opus-4-6` in `.env` → restart
Old session has stale context	Session cached from before config change	Send `/new` in chat to start fresh session
"未检测到应用连接信息" when saving events	Bridge not running yet	Start bridge first, then save event config

Step 6: Update ARIS config

cat > ~/.claude/feishu.json << 'EOF'
{
  "mode": "interactive",
  "webhook_url": "https://open.feishu.cn/open-apis/bot/v2/hook/YOUR_WEBHOOK_ID",
  "interactive": {
    "bridge_url": "http://localhost:5000",
    "timeout_seconds": 300
  }
}
EOF

Now skills will:

Push rich cards to the group (status notifications, everyone sees)
Private chat you for decisions (checkpoints, approve/reject, custom instructions)

Which skills send notifications?

Skill	Events	Push	Interactive
`/auto-review-loop`	Review scored (each round), loop complete	Score + verdict	+ wait for continue/stop
`/auto-paper-improvement-loop`	Review scored, all rounds done	Score progression	Score progression
`/run-experiment`	Experiments deployed	GPU assignment + ETA	GPU assignment + ETA
`/vast-gpu`	Instance rented/destroyed	Instance ID + cost	Instance ID + cost
`/monitor-experiment`	Results collected	Results table	Results table
`/idea-discovery`	Phase transitions, final report	Summary at each phase	+ approve/reject at checkpoints
`/research-pipeline`	Stage transitions, pipeline done	Stage summary	+ approve/reject

Not using Feishu? No problem — without ~/.claude/feishu.json, all skills behave exactly as before. Zero overhead, zero side effects.

💡 Alternative IM platforms: The push-only webhook pattern works with any service that accepts incoming webhooks (Slack, Discord, DingTalk, WeChat Work). Just change the webhook_url and card format in feishu-notify/SKILL.md. For bidirectional support, see cc-connect (multi-platform bridge) or clawdbot-feishu.

🎛️ Customization

Skills are plain Markdown files. Fork and customize:

💡 Parameter pass-through: Parameters flow down the call chain automatically. For example, /research-pipeline "topic" — sources: zotero, arxiv download: true passes sources and arxiv download through idea-discovery all the way down to research-lit. This also works for optional sources such as deepxiv and exa: /research-pipeline "topic" — sources: all, deepxiv, exa. You can set any downstream parameter at any level — just add — key: value to your command.
research-pipeline  ──→  idea-discovery      ──→  research-lit
                   ──→  experiment-bridge    ──→  run-experiment
                   ──→  auto-review-loop
                                            ──→  idea-creator
                                            ──→  novelty-check
                                            ──→  research-review

Full Research Pipeline (`research-pipeline`)

Constant	Default	Description	Pass-through
`AUTO_PROCEED`	true	Auto-continue with top-ranked option if user doesn't respond	→ `idea-discovery`
`ARXIV_DOWNLOAD`	false	Download top arXiv PDFs after literature search	→ `idea-discovery` → `research-lit`
`HUMAN_CHECKPOINT`	false	When `true`, pause after each review round for approval	→ `auto-review-loop`
`WANDB`	false	Auto-add W&B logging to experiments	→ `experiment-bridge` → `run-experiment`
`CODE_REVIEW`	true	GPT-5.4 reviews experiment code before deployment	→ `experiment-bridge`
`BASE_REPO`	false	GitHub repo URL to clone as base codebase for experiments	→ `experiment-bridge`
`GPU`	`local`	GPU target: `local`, `remote` (SSH), or `vast` (Vast.ai on-demand rental)	→ `experiment-bridge` → `run-experiment`
`COMPACT`	false	Generate compact summary files for short-context models and session recovery	→ all workflows
`REF_PAPER`	false	Reference paper (PDF path or URL) to base ideas on. Summarized first, then used as context	→ `idea-discovery`
`ILLUSTRATION`	`gemini`	AI illustration: `gemini` (default), `mermaid` (free), or `false` (skip)	→ `paper-writing`

Override inline: /research-pipeline "topic" — auto proceed: false, illustration: mermaid

Auto Review Loop (`auto-review-loop`)

Constant	Default	Description
`MAX_ROUNDS`	4	Maximum review→fix→re-review iterations
`POSITIVE_THRESHOLD`	6/10	Score at which the loop stops (submission-ready)
`> 4 GPU-hour skip`	4h	Experiments exceeding this are flagged for manual follow-up

Idea Discovery (`idea-discovery` / `idea-creator`)

Constant	Default	Description	Pass-through
`PILOT_MAX_HOURS`	2h	Skip any pilot estimated to take longer per GPU	—
`PILOT_TIMEOUT_HOURS`	3h	Hard timeout — kill runaway pilots, collect partial results	—
`MAX_PILOT_IDEAS`	3	Maximum number of ideas to pilot in parallel	—
`MAX_TOTAL_GPU_HOURS`	8h	Total GPU budget across all pilots	—
`AUTO_PROCEED`	true	Auto-continue with top-ranked option if user doesn't respond	—
`ARXIV_DOWNLOAD`	false	Download top arXiv PDFs after literature search	→ `research-lit`

Override inline: /idea-discovery "topic" — pilot budget: 4h per idea, sources: zotero, arxiv download: true

Experiment Bridge (`experiment-bridge`)

Constant	Default	Description
`CODE_REVIEW`	true	GPT-5.4 xhigh reviews code before deployment. Catches logic bugs before wasting GPU hours
`AUTO_DEPLOY`	true	Automatically deploy experiments after implementation + review. Set `false` to manually inspect
`SANITY_FIRST`	true	Run smallest experiment first to catch setup bugs before full deployment
`MAX_PARALLEL_RUNS`	4	Maximum experiments to deploy in parallel (limited by available GPUs)
`WANDB`	false	Auto-add W&B logging. Requires `wandb_project` in CLAUDE.md
`BASE_REPO`	false	GitHub repo URL to clone as base codebase for experiments

Override inline: /experiment-bridge — base repo: https://github.com/org/project

Literature Search (`research-lit`)

Constant	Default	Description
`PAPER_LIBRARY`	`papers/`, `literature/`	Local directories to scan for PDFs before searching online
`MAX_LOCAL_PAPERS`	20	Max local PDFs to scan (first 3 pages each)
`SOURCES`	`all`	Which sources to search: `zotero`, `obsidian`, `local`, `web`, `semantic-scholar`, `deepxiv`, `exa`, or `all`. `semantic-scholar`, `deepxiv`, and `exa` must be explicitly listed
`ARXIV_DOWNLOAD`	false	When `true`, download top relevant arXiv PDFs to PAPER_LIBRARY after search
`ARXIV_MAX_DOWNLOAD`	5	Maximum number of PDFs to download when `ARXIV_DOWNLOAD = true`

Override inline: /research-lit "topic" — sources: zotero, web, /research-lit "topic" — sources: all, deepxiv, /research-lit "topic" — sources: all, exa, /research-lit "topic" — arxiv download: true, max download: 10

Paper Writing (`paper-write`)

Constant	Default	Description
`DBLP_BIBTEX`	true	Fetch real BibTeX from DBLP/CrossRef instead of LLM-generated entries
`TARGET_VENUE`	`ICLR`	Target venue: `ICLR`, `NeurIPS`, `ICML`, `CVPR`, `ACL`, `AAAI`, `ACM`, `IEEE_JOURNAL`, `IEEE_CONF`
`ANONYMOUS`	true	Use anonymous author block for blind review. Note: most IEEE venues are NOT anonymous — set `false` for IEEE
`MAX_PAGES`	9	Page limit. ML conferences: main body excl. refs. IEEE: total pages incl. refs
`ILLUSTRATION`	`gemini`	AI illustration mode: `gemini` (default, needs `GEMINI_API_KEY`), `mermaid` (free), or `false` (skip)

Override inline: /paper-write — target venue: NeurIPS, illustration: mermaid

General (all skills using Codex MCP)

Constant	Default	Description
`REVIEWER_MODEL`	`gpt-5.4`	OpenAI model used via Codex MCP. Also available: `gpt-5.3-codex`, `gpt-5.2-codex`, `o3`. See supported models for full list.

Prompt templates — tailor the review persona and evaluation criteria
allowed-tools — restrict or expand what each skill can do

🔀 Alternative Model Combinations

Don't have Claude / OpenAI API access? You can swap in other models — same cross-model architecture, different providers.

⭐ We strongly recommend Claude + GPT-5.4 (default setup). It's the most tested and reliable combination. Alternative setups work but may require prompt tuning.

	Executor	Reviewer	Need Claude API?	Need OpenAI API?	Guide
Default ⭐	Claude Opus/Sonnet	GPT-5.4 (Codex MCP)	Yes	Yes	Quick Start
Alt A	GLM-5 (Z.ai)	GPT-5.4 (Codex MCP)	No	Yes	Setup below
Alt B	GLM-5 (Z.ai)	MiniMax-M2.7	No	No	MINIMAX_MCP_GUIDE
Alt C	Any CC-compatible	Any OpenAI-compatible	No	No	LLM_API_MIX_MATCH_GUIDE
Alt D	Kimi-K2.5 / Qwen3.5+	GLM-5 / MiniMax-M2.7	No	No	ALI_CODING_PLAN_GUIDE
Alt E 🆓	DeepSeek-V3.1 / Qwen3-Coder	DeepSeek-R1 / Qwen3-235B	No	No	MODELSCOPE_GUIDE
Alt F	Codex CLI (GPT-5.4)	Codex `spawn_agent` (GPT-5.4)	No	Yes	skills-codex/
Alt G 🆕	Codex CLI	Claude Code CLI (`claude-review` MCP)	No*	No*	CODEX_CLAUDE_REVIEW_GUIDE
Alt H 🆕	Antigravity (Claude Opus 4.6 / Gemini 3.1 Pro)	GPT-5.4 (Codex MCP) or any via llm-chat	No	Optional	ANTIGRAVITY_ADAPTATION
Alt I 🆕	Codex CLI	Gemini direct API (`gemini-review` MCP)	No	No	CODEX_GEMINI_REVIEW_GUIDE

Alt C supports tested providers: GLM (Z.ai), Kimi (Moonshot), LongCat (Meituan) as executors; DeepSeek, MiniMax as reviewers. Any OpenAI-compatible API should also work via the generic llm-chat MCP server. Alt D uses Alibaba Coding Plan — one API key for both executor and reviewer, 4 models included (Kimi, Qwen, GLM, MiniMax). Alt E uses ModelScope — free (2000 calls/day), one key, no automation restrictions. Alt G keeps Codex as executor but swaps the reviewer to Claude Code CLI via the local claude-review MCP bridge, with async polling for long paper/review prompts. Alt H uses Google Antigravity as the executor with native SKILL.md support — choose Claude Opus 4.6 (Thinking) or Gemini 3.1 Pro (high) as the execution model. Alt I keeps Codex as executor, adds only a thin skills-codex-gemini-review overlay, and routes the reviewer-aware predefined skills through the local gemini-review MCP bridge with direct Gemini API by default. It is the closest Gemini analogue to the existing Codex+Claude review path, while minimizing skill changes and now also covers poster PNG review via the same bridge. Free-tier availability, rate limits, and data-use terms remain subject to Google's current policy.

* Alt G normally relies on local Codex CLI and Claude Code CLI logins. Direct API keys are optional, not required.

Alt A: GLM + GPT

Only replace the executor (Claude → GLM), keep GPT-5.4 as reviewer via Codex MCP.

npm install -g @anthropic-ai/claude-code
npm install -g @openai/codex
codex setup   # set model to gpt-5.4

Configure ~/.claude/settings.json:

{
    "env": {
        "ANTHROPIC_AUTH_TOKEN": "your_zai_api_key",
        "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic",
        "API_TIMEOUT_MS": "3000000",
        "ANTHROPIC_DEFAULT_HAIKU_MODEL": "glm-4.5-air",
        "ANTHROPIC_DEFAULT_SONNET_MODEL": "glm-4.7",
        "ANTHROPIC_DEFAULT_OPUS_MODEL": "glm-5"
    },
    "mcpServers": {
        "codex": {
            "command": "/opt/homebrew/bin/codex",
            "args": ["mcp-server"]
        }
    }
}

Codex CLI uses your existing OPENAI_API_KEY (from ~/.codex/config.toml or environment) — no extra config needed for the reviewer side.

Alt B: GLM + MiniMax

No Claude or OpenAI API needed. Uses a custom MiniMax MCP server instead of Codex (because MiniMax doesn't support OpenAI's Responses API). Full guide: docs/MINIMAX_MCP_GUIDE.md.

Alt C: Any Executor + Any Reviewer

Mix and match freely using the generic llm-chat MCP server. Supports any OpenAI-compatible API as reviewer. Full guide: docs/LLM_API_MIX_MATCH_GUIDE.md.

Example combinations: GLM + DeepSeek, Kimi + MiniMax, Claude + DeepSeek, LongCat + GLM, etc.

After Setup: Install Skills & Verify

git clone https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep.git
cd Auto-claude-code-research-in-sleep
cp -r skills/* ~/.claude/skills/
claude

⚠️ For non-Claude executors (GLM, Kimi, etc.): Let the model read through the project once to ensure skills are correctly parsed. This is especially important if you've rewritten skills to use a different reviewer MCP (e.g., mcp__llm-chat__chat instead of mcp__codex__codex) — the new executor needs to understand the changed tool call patterns:
Read through this project and verify all skills are working:
/idea-creator, /research-review, /auto-review-loop, /novelty-check,
/idea-discovery, /research-pipeline, /research-lit, /run-experiment,
/analyze-results, /monitor-experiment, /pixel-art

⚠️ Note: Alternative models may behave differently from Claude and GPT-5.4. You may need to tune prompt templates for best results. The core cross-model architecture remains the same.

📋 Roadmap

Done

Human-in-the-loop checkpoints — idea-discovery and research-pipeline pause at key decision points for user approval. Configurable via AUTO_PROCEED (default: auto-continue; set false to always wait)
Alternative model combinations — GLM + GPT, GLM + MiniMax fully documented with setup guides. No Claude or OpenAI API required
Workflow 3: Paper Writing Pipeline — full chain: /paper-plan → /paper-figure → /paper-write → /paper-compile. ICLR/NeurIPS/ICML templates, claims-evidence matrix, publication-quality figures, latexmk auto-fix. Inspired by claude-scholar, Research-Paper-Writing-Skills, baoyu-skills

Show 6 more completed items

Planned

Daemon mode — auto-restart Claude Code session via launchd/systemd for true unattended operation. Currently the orchestration layer requires an active CLI session; state files (REVIEW_STATE.json, AUTO_REVIEW.md) support resuming across sessions, but relaunch is manual (#11)
Reference-style figure generation — read figures from reference PDFs → identify chart type, color scheme, layout → generate same-style figures with your own data. Sub-goal remaining: Data charts (extract color/font style → matplotlib rcParams). Method diagrams ✅ solved by paper-illustration
Workflow execution report — after each workflow (1/1.5/2/3) completes, auto-generate a structured summary: what was done, key decisions made, experiments run, results obtained, scores, and time spent. Output as WORKFLOW_REPORT.md for progress tracking, team reporting, and supervisor updates
Document-based pipeline input — /idea-discovery and /research-pipeline auto-detect RESEARCH_BRIEF.md in project root. Detailed context replaces one-line prompt. Template: templates/RESEARCH_BRIEF_TEMPLATE.md
Auto hyperparameter tuning skill — rewrite auto-hparam-tuning as an ARIS SKILL.md. 5-step cycle: understand project → plan tuning strategy → run experiments → analyze metrics (TensorBoard/W&B) → learn and iterate. Would plug into Workflow 1.5 (/experiment-bridge) or Workflow 2 (/auto-review-loop) when reviewer says "tune hyperparameters"
Plugin format — package ARIS as a Claude Code Plugin for one-click install via /plugin install aris. Skills version continues for cross-platform compatibility (Codex CLI, Cursor, Trae, etc.)

💬 Community

Domain-specific skills welcome! The core skills cover general research workflows, but every field has its own tools and patterns. We welcome PRs that add new skills for your domain — EDA, bioinformatics, robotics, HPC, or anything else. Just add a skills/your-skill/SKILL.md and open a PR. See dse-loop for an example.

Join the WeChat group for discussion on Claude Code + AI-driven research workflows:

📖 Citation

If you use ARIS in your research, please cite:

@misc{yang2026aris,
    author       = {Yang, Ruofeng and Li, Yongcan and Li, Shuai},
    title        = {ARIS: Fully Autonomous Research via Adversarial Multi-Agent Collaboration},
    year         = {2026},
    url          = {https://github.com/wanshuiyin/Auto-claude-code-research-in-sleep},
}

⭐ Star History

🙏 Acknowledgements

ARIS is inspired by:

🧪 AI Scientist (Sakana AI) — Automated research pioneer
📖 AutoResearch (Andrej Karpathy) — End-to-end research automation
🔭 FARS (Analemma) — Fully Automated Research System
🎨 PaperBanana (PKU) — Multi-agent academic illustration framework

This project builds on and integrates with many excellent open-source projects:

Core Infrastructure

Claude Code — Anthropic's CLI for Claude, the execution backbone
Codex CLI — OpenAI's CLI, used as MCP server for cross-model review

Zotero Integration (setup guide)

zotero-mcp — Zotero MCP server with semantic search and PDF annotations
Zotero — Open-source reference manager

Obsidian Integration (setup guide)

mcpvault — Obsidian vault MCP server (no app required)
obsidian-skills — Claude Code skills for Obsidian Markdown by Steph Ango (Obsidian CEO)

Paper Writing Inspiration

claude-scholar — Academic paper writing with Claude
Research-Paper-Writing-Skills — Paper writing skill templates
baoyu-skills — Claude Code skills collection

Feishu/Lark Integration (setup guide)

feishu-claude-code — Bidirectional Feishu ↔ Claude Code bridge
clawdbot-feishu — Feishu bot for Claude
cc-connect — Multi-platform messaging bridge
lark-openapi-mcp — Official Lark MCP server

Community

awesome-agent-skills — Curated list of Claude Code skills (featured)

Special Thanks — Platform Adaptation

ARIS wouldn't run on so many platforms without these contributors:

🤖 @Falling-Flower — adapted all ARIS skills for Codex CLI using spawn_agent
🔧 @No-518 — ongoing maintenance of the Codex skill set, keeping parity with latest updates
🖱️ @YecanLee — wrote the Cursor adaptation guide and local GPU setup docs
🏆 @DefanXue & @Monglitay — first community paper built entirely with ARIS, scored 8/10 at CS conference

Special Thanks — Architecture & Vision

💡 @JingxuanKang — beyond code contributions (training-check, result-to-claim, ablation-planner, watchdog, templates, session recovery), deeply shaped ARIS through discussions on architecture design, compact mode, workflow state management, and the vision of what autonomous research workflows should look like. Many of today's core features — from structured project files to context-aware session recovery — grew out of these conversations.

License

MIT

Name		Name	Last commit message	Last commit date
Latest commit History 616 Commits
assets		assets
community_papers		community_papers
docs		docs
mcp-servers		mcp-servers
skills		skills
templates		templates
tests		tests
tools		tools
.env.example		.env.example
.gitattributes		.gitattributes
.gitignore		.gitignore
AGENT_GUIDE.md		AGENT_GUIDE.md
CONTRIBUTING.md		CONTRIBUTING.md
CONTRIBUTING_CN.md		CONTRIBUTING_CN.md
LICENSE		LICENSE
README.md		README.md
README_CN.md		README_CN.md
xhs_post.md		xhs_post.md

Parameter	Default	What it does
`AUTO_PROCEED`	`true`	Auto-continue at idea selection gate. Set `false` to manually pick which idea to pursue before committing GPU time
`human checkpoint`	`false`	Pause after each review round so you can read the score, give custom modification instructions, skip specific fixes, or stop early
`sources`	`all`	Which literature sources to search: `zotero`, `obsidian`, `local`, `web`, `semantic-scholar`, `deepxiv`, `exa`, or `all`. Note: `semantic-scholar`, `deepxiv`, and `exa` must be explicitly listed — not included in `all`
`arxiv download`	`false`	Download top relevant arXiv PDFs during literature survey. When `false`, only fetches metadata (title, abstract, authors)
`DBLP_BIBTEX`	`true`	Fetch real BibTeX from DBLP/CrossRef instead of LLM-generated entries. Eliminates hallucinated citations. Zero install
`code review`	`true`	GPT-5.4 xhigh reviews experiment code before GPU deployment. Set `false` to skip
`wandb`	`false`	Auto-add W&B logging to experiment scripts. Set `true` + configure `wandb_project` in CLAUDE.md. `/monitor-experiment` pulls training curves from W&B
`illustration`	`gemini`	AI illustration in Workflow 3: `gemini` (default, needs `GEMINI_API_KEY`), `mermaid` (free), or `false` (skip)
`venue`	`ICLR`	Target venue: `ICLR`, `NeurIPS`, `ICML`, `CVPR`, `ACL`, `AAAI`, `ACM`. Determines LaTeX style file and page limit
`base repo`	`false`	GitHub repo URL to clone as base codebase (e.g., `— base repo: https://github.com/org/project`). No code? Build on top of an open-source project
`gpu`	`local`	GPU target: `local` (default), `remote` (SSH server), or `vast` (rent on-demand from Vast.ai — auto-provision, auto-destroy)
`compact`	`false`	Generate compact summary files (`IDEA_CANDIDATES.md`, `findings.md`, `EXPERIMENT_LOG.md`) for short-context models and session recovery
`ref paper`	`false`	Reference paper to build on (PDF path or arXiv URL). Summarized first, then ideas extend/improve it. Combine with `base repo` for paper+code workflows
`effort`	`balanced`	Work intensity: `lite` (0.4x tokens), `balanced` (default), `max` (2.5x), `beast` (5-8x). Controls breadth/depth/iterations. Codex reasoning always `xhigh`. See Effort Levels
`reviewer`	`codex`	Reviewer backend: `codex` (GPT-5.4 xhigh, default), `oracle-pro` (GPT-5.4 Pro via Oracle — strongest reasoning). See Setup →
`difficulty`	`medium`	Reviewer adversarial level: `medium` (default), `hard` (+ memory + debate), `nightmare` (+ GPT reads repo via `codex exec`)

Folders and files

Latest commit

History

Repository files navigation

Auto-claude-code-research-in-sleep (ARIS ⚔️🌙)

🎯 More Than Just a Prompt

🏆 Papers Built with ARIS

📢 What's New

🚀 Quick Start

✨ Features

📈 Score Progression (Real Run)

🏆 Community Showcase — Papers Built with ARIS

🧩 Awesome Community Skills & Extensions

🔄 Workflows

Full Pipeline 🚀

Workflow 1: Idea Discovery & Method Refinement 🔍

Workflow 1.5: Experiment Bridge 🔗

Workflow 2: Auto Research Loop 🔁 (sleep & wake up to results)

Workflow 3: Paper Writing Pipeline 📝

Auto Paper Improvement Loop ✨

Workflow 4: Rebuttal 📝 (reply to reviewers safely)

📚 Research Wiki — Persistent Research Memory

Workflow M: Meta-Optimize 🧬 (ARIS optimizes itself)

⚡ Effort Levels

🧿 Optional: GPT-5.4 Pro via Oracle

🧰 All Skills

🚀 Full Pipeline

🔍 Workflow 1: Idea Discovery & Method Refinement

🔗 Workflow 1.5: Experiment Bridge

🔁 Workflow 2: Auto Research Loop

📝 Workflow 3: Paper Writing

📝 Workflow 4: Rebuttal

🛠️ Standalone / Utility

⚙️ Setup

Prerequisites

Install Skills

Update Skills

Usage

🌙 Auto-Allow for Overnight Runs (Optional)

🖥️ GPU Server Setup (For Auto-Experiments)

Option A: Remote SSH Server (gpu: remote)

Option B: Local GPU (gpu: local)

Option C: Vast.ai On-Demand GPU (gpu: vast)

arXiv Integration

📱 Feishu/Lark Integration (Optional)

Which skills send notifications?

🎛️ Customization

Full Research Pipeline (research-pipeline)

Auto Review Loop (auto-review-loop)

Idea Discovery (idea-discovery / idea-creator)

Experiment Bridge (experiment-bridge)

Literature Search (research-lit)

Paper Writing (paper-write)

General (all skills using Codex MCP)

🔀 Alternative Model Combinations

Alt A: GLM + GPT

Alt B: GLM + MiniMax

Alt C: Any Executor + Any Reviewer

After Setup: Install Skills & Verify

📋 Roadmap

Done

Planned

💬 Community

📖 Citation

⭐ Star History

🙏 Acknowledgements

License

About

Topics

Resources

License

Contributing

Uh oh!

Stars

Watchers

Forks

Releases 16

Packages 0

Uh oh!

Contributors

Option A: Remote SSH Server (`gpu: remote`)

Option B: Local GPU (`gpu: local`)

Option C: Vast.ai On-Demand GPU (`gpu: vast`)

Full Research Pipeline (`research-pipeline`)

Auto Review Loop (`auto-review-loop`)

Idea Discovery (`idea-discovery` / `idea-creator`)

Experiment Bridge (`experiment-bridge`)

Literature Search (`research-lit`)

Paper Writing (`paper-write`)

Packages