auto-research

claude-auto-research

Claude Code plugin for AI/ML research papers targeting CCF-A venues (ICML, ICLR, NeurIPS, CVPR, etc.) with a live experiment dashboard.

Skills

• /auto-research:pipeline — Full research pipeline (Phase 0–12: ideation → literature review → pilot → experiments → analysis → writing → rebuttal)
• /auto-research:ideation — Standalone ideation (Phase 1–2: literature review, idea generation, proposal)
• /auto-research:lab — Standalone lab (Phase 3–8: pilot and full experiments)
• /auto-research:reviewer — Standalone reviewer (7 modes: E=idea debate, F=code review, A=experiment design, B=pilot verdict, C=paper peer review, D=rebuttal strategy, G=paper consistency check)
• /auto-research:dashboard — Live experiment dashboard with 7 per-phase tabs (Setup/Ideation/Pilot/Experiments/Analysis/Writing/Rebuttal); experiment design tables with todo→pending→running→done cell states; click a done cell to open its W&B run directly
• /auto-research:dashboard-update — Update dashboard metadata for current project (human-readable table/method labels and insights)

Prerequisites

Before first use, install and authenticate the following:

Tool	Purpose	Install / Auth
wandb login	Experiment tracking	wandb.ai/settings
huggingface-cli login	Checkpoint storage	hf.co/settings/tokens
gnvitop	GPU availability check	Ask your sysadmin
tmux	Agent progress display	apt install tmux / brew install tmux
gh auth login	GitHub CLI (git ops)	cli.github.com
uv	Python package runner	curl -LsSf https://astral.sh/uv/install.sh | sh
supervisor	Local GPU experiment dispatch	See skills/pipeline/shared/supervisor-setup.md
SSH	Access to compute machines	Configure ~/.ssh/config
TELEGRAM_BOT_TOKEN + TELEGRAM_CHAT_ID	Notifications	Set up via @BotFather; see skills/pipeline/shared/notifications.md

Installation

# 1. Clone the repo
git clone https://github.com/Linwei94/claude-auto-research.git

# 2. Install plugin from local directory
claude plugin install ./claude-auto-research

# 3. Restart Claude Code

Skills will appear in the / menu after restart.

Usage

Type / in Claude Code to trigger either skill, or just describe what you want:

• "帮我做个文献综述" → triggers /auto-research:pipeline
• "看看实验进度" → triggers /auto-research:dashboard
• "open dashboard" → triggers /auto-research:dashboard
• "standalone literature review" → triggers /auto-research:ideation

Architecture

Agent Team

• Pipeline Lead — coordinates all phases (0–12)
• Ideation Agent — Phase 1–2: literature review, idea generation, proposal
• Lab Agent — Phase 3–8: experiments (3-tier orchestrator)
• Reviewer Agent — quality gates at Phases 2 (Mode E: idea), 5 (Mode B: pilot), 11 (Mode C: paper), 12 (Mode D: rebuttal). Phase 7 experiment design debate is spawned internally by Lab Agent (not by Reviewer).

Lab Agent 3-Tier Design

Lab Agent (orchestrator)
├── Environment Sub-agents × M (one per machine, Sonnet)
│   └── rsync → conda → GPU check → dry-run → write progress/env_<HOST>.json
└── Execution Sub-agents × N (one per experiment, Haiku)
    ├── exec_local.md  — SSH + nohup (xuchang-lab*)
    ├── exec_c500.md   — sco acp jobs create (C500 MetaX cluster)
    └── exec_gadi.md   — PBS/qsub (Gadi NCI cluster)

Phases

Phase	Name	Owner
0	Setup	Pipeline Lead
1–2	Literature Review + Ideation	Ideation Agent
3–5	Pilot Experiments	Lab Agent
6–8	Full Experiments	Lab Agent
9	Result Analysis	Lab Agent + Reviewer
10–11	Paper Writing + Internal Review	Pipeline Lead
12	Post-Submission	Pipeline Lead