autor

AutoR

Repository path: ~/repos/AutoR

Positioning

AI handles execution, humans handle direction. Not a one-click paper-producing black box, but a structured research harness.

Architecture

Underlying executor: claude / codex 8-stage pipeline: 01 Literature review → 02 Hypothesis generation → 03 Research design → 04 Implementation → 05 Experiments → 06 Analysis → 07 Writing → 08 Publishing

Output directory: runs/<run_id>/ (prompt, logs, code, data, figures, LaTeX, PDF)

Common commands

# 新 run
cd ~/repos/AutoR && python main.py --goal "研究目标"

# 指定执行器
python main.py --goal "..." --operator claude   # 默认
python main.py --goal "..." --operator codex

# 全自动模式（reviewer agent 代替人工审批）
python main.py --goal "..." --full-auto

# 从已有项目起步
python main.py --goal "..." --project-root /path/to/project

# 从论文语料起步
python main.py --goal "..." --paper-corpus /path/to/papers

# 恢复 run
python main.py --resume-run <run_id>
python main.py --resume-run latest

# 从某 stage 重做
python main.py --resume-run <run_id> --redo-stage 06_analysis

# 回滚 stage（下游全部失效）
python main.py --resume-run <run_id> --rollback-stage 04_implementation

# 指定投稿 venue
python main.py --goal "..." --venue neurips_2025  # 默认
python main.py --goal "..." --venue nature
python main.py --goal "..." --venue jmlr

# 带资源
python main.py --goal "..." --resources paper.pdf data/ code/

# 跳过 intake
python main.py --goal "..." --skip-intake

# 生成方法示意图（需 google-genai + Pillow）
python main.py --goal "..." --research-diagram

Core principles

Human approval by default — waits for your review at the end of each stage
--full-auto Human review can be skipped, but it is recommended to manually approve the first few runs
runs/ directoryis the only product — all intermediate results are inside it
Resumable — resume, redo, rollback at any time

Use cases

Starting from an arXiv paper, fully reproduce + extend
Starting from a code repository, complete experiments + write the paper
Starting from a research question, go through the full literature → experiment → writing workflow
Need reproducible, auditable research artifacts

Relationship to Hermes's existing skills

paper2code: single paper → minimal code. AutoR is a more complete 8-stage pipeline
gap-driven-paper-pipeline: gap → multiple papers → batch code. AutoR goes deeper into experiments + writing
Applicable: when you need complete paper output, need experimental validation, need submission formatting

Notes

No pip install required, runs on pure Python stdlib
Requires installed claude or codex
Stage 07 writing produces LaTeX; with a TeX environment, you can compile PDF directly
Default venue: neurips_2025