# AutoSci **Read, think, experiment, write, evolve — the AI research agent with memory that compounds across every project.** [![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE) [![Python 3.9+](https://img.shields.io/badge/Python-3.9+-yellow.svg)](https://www.python.org/) [![Claude Code](https://img.shields.io/badge/main-Claude_Code_stable-d97706.svg)](https://docs.anthropic.com/en/docs/claude-code) [![Codex Preview](https://img.shields.io/badge/Codex-preview-111111.svg)](https://developers.openai.com/codex) [![OpenCode Preview](https://img.shields.io/badge/OpenCode-preview-6f42c1.svg)](https://opencode.ai/) [![arXiv](https://img.shields.io/badge/arXiv-2605.31468-b31b1b.svg)](https://arxiv.org/abs/2605.31468) [![Status](https://img.shields.io/badge/status-internal_beta-orange.svg)](#️⃣-status--update)

--- ## ⚠️ Status & Update > **Thanks to everyone who's been trying AutoSci — the community response has been amazing!** AutoSci evolved from our earlier OmegaWiki prototype into what we're building toward: a next-generation research agent that can handle the full scientific lifecycle. We're actively testing and iterating on new features, and more capabilities are on the way. Jump in, break things, and tell us what you think — your feedback and ideas are what's shaping where this goes next. 🙏 > **🌿 Which branch?** [`main`](https://github.com/skyllwt/AutoSci/tree/main) remains the **stable Claude Code version**. [`autosci-codex`](https://github.com/skyllwt/AutoSci/tree/autosci-codex) is the **official Codex Preview**, and [`autosci-opencode`](https://github.com/skyllwt/AutoSci/tree/autosci-opencode) is the **official OpenCode Preview**. These are separate runtime adaptations; the existing Claude Code and Codex versions remain available. The **full system described in our [paper](https://arxiv.org/abs/2605.31468)** — SciMem · SciFlow · SciDAG · SciEvolve — lives on the [`paper`](https://github.com/skyllwt/AutoSci/tree/paper) branch (frozen as tag [`arxiv-v1`](https://github.com/skyllwt/AutoSci/tree/arxiv-v1)). ### OpenCode Preview Try the OpenCode adaptation without changing your Claude Code or Codex checkout: ```bash git clone -b autosci-opencode https://github.com/skyllwt/AutoSci.git cd AutoSci ./setup.sh --lang en opencode # Then ask OpenCode to run the init skill for your research topic. ``` Current boundary: | Area | Status | |---|---| | Local OpenCode skills | Preview supported via generated `.opencode/skills` | | Runtime instructions | Root `AGENTS.md`, generated from the selected language | | Shared skill source | `i18n//skills` remains the bilingual source of truth | | Review LLM | Local `llm-review` MCP configured through generated `opencode.json` | | Daily arXiv | Recommendation-only; standalone OpenAI-compatible API with deterministic fallback and optional email | The OpenCode preview does not replace the Claude Code stable release or the Codex Preview. ### Codex Preview Try the Codex preview without changing your `main` checkout: ```bash git clone -b autosci-codex https://github.com/skyllwt/AutoSci.git cd AutoSci ./setup.sh --lang en codex # Then invoke: $init [your-research-topic] ``` Current boundary: | Area | Status | |---|---| | Local Codex skills | Preview supported via `.agents/skills` | | Runtime scope | Codex-only; Claude Code remains available separately on `main` | | Shared skill source | `i18n//skills` regenerates the active `.agents/skills` tree | | Review LLM | Optional `llm-review` MCP configured through the Codex config example | | Daily arXiv | CI is recommendation-only; no repository writeback or unattended ingest | See [docs/codex-preview.md](docs/codex-preview.md) for the preview notes and known boundaries. --- ## 📄 Paper > ### [AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle](https://arxiv.org/abs/2605.31468) > > [![arXiv](https://img.shields.io/badge/arXiv-2605.31468-b31b1b.svg)](https://arxiv.org/abs/2605.31468) · [📄 **Read on arXiv →**](https://arxiv.org/abs/2605.31468) If you find AutoSci useful in your research, please [cite our paper](#citation). --- ## 📌 Poster & Demo

_{AutoSci poster — click to view full size.}

_{▶ Watch the AutoSci demo on Bilibili}

--- ## 🆕 What's New ### 2026-09-28 - Hiera: guided method iteration Explore alternative methods for an existing experiment with `/hiera-experiment` in Claude Code or `$hiera-experiment` in Codex. The agent helps propose and implement candidates, compare results, and choose the next method change or parameter search. This optional workflow supports local and remote execution, keeps candidate history under `runs/hiera/`, and asks for your review before running experiments.

Quick tutorial: iterate on an existing experiment

**Before starting:** complete `/exp-design ` (Codex: `$exp-design`) and choose an experiment slug from `wiki/experiments/`. Python 3.11 is recommended. Open the corresponding AutoSci checkout in Claude Code or Codex. **1. Start a run.** In Claude Code, send the command and instruction together. In Codex, replace `/hiera-experiment` with `$hiera-experiment`. ```text /hiera-experiment --run-id method-01 --init Use the existing experiment design and local execution, with a budget of 10 evaluations. Show the method hypotheses, metric, direction, and execution configuration for review. Ask me about any missing choices. ``` **2. Build and evaluate a candidate.** Reuse the same experiment slug and run ID. Replace `--init` with the flag below and add the accompanying instruction, one step at a time. Substitute the IDs and draft/spec paths the agent shows you. | Step | Flag | What to tell the agent | |---|---|---| | Propose methods | `--propose` | Propose up to 3 method variants against the baseline and explain what each tests. | | Write a selected method | `--propose` | Implement point `` as candidate ``, using its displayed parent relations. Show the draft code and spec for review; do not run yet. | | Accept reviewed code | `--admit` | Admit the reviewed spec at `` from ``. Do not run yet. | | Approve execution | `--approve` | I reviewed the source, configuration, metric, direction, resources, and execution environment. Approve this contract only. | | Check readiness | `--preflight` | Check candidate `` only. | | Run and compare | `--run` | Run candidate `` at screen depth under the approved local contract. Show its score and comparison with any evaluated baseline. | Confirm the proposed hypotheses and execution settings before proceeding. Evaluate a baseline candidate as well, so method comparisons use actual results. **3. Choose the next step after reviewing results.** - **Change the method:** use `--propose`, ask for changes motivated by the previous results, then repeat the draft, review, admit, preflight, and run steps. The existing approval remains valid while the execution contract stays unchanged. - **Tune parameters:** after reviewing tunable ranges in the contract, use `--tune` with "Tune candidate `` with at most 3 variants at deep depth." This creates and runs parameter variants. - **Repeat evaluations:** use `--loop` with "Run up to 3 deep rounds, with at most 3 bouts per candidate." It evaluates existing candidates; new methods still require the proposal and authoring steps. Use `--status` to inspect progress, then `--finalize` once no work is pending to save the exploratory report. Bring the selected method back to the normal `/exp-run` and `/exp-eval` workflow for formal validation (Codex: `$exp-run` and `$exp-eval`). [Full workflow and remote execution guide](i18n/en/skills/hiera-experiment/README.md) | [Chinese guide](i18n/zh/skills/hiera-experiment/README.md)

### 🛠️ 2026-07-12 · OpenCode Preview adaptation Published a separate OpenCode adaptation while keeping the existing Claude Code stable release and Codex Preview available. The OpenCode branch provides bilingual project skills under `.opencode/skills`, root `AGENTS.md` instructions, generated machine-local configuration, and `llm-review` MCP integration. Its daily-arXiv automation is recommendation-only and uses a standalone OpenAI-compatible API with deterministic fallback, optional best-effort email, and digest artifacts without repository writeback. ### 🛠️ 2026-07-10 · autosci-codex branch adaptation Published a separate Codex-only adaptation while keeping the stable Claude Code release on `main` and the OpenCode Preview on `autosci-opencode`. The Codex branch provides bilingual skills under `.agents/skills`, root `AGENTS.md` instructions, Codex-specific setup and Review LLM configuration, and recommendation-only daily-arXiv CI without Claude Action authentication, automatic ingest, or repository writeback. The `$research` workflow delegates cold-wiki bootstrap to `$init` and proceeds only after bootstrap completes. ### 🛠️ 2026-05-19 · Experiment Overhaul A possible usage process：`/ideate [research-direction-or-topic]`(You can use `--skip-pilot` to decide whether to conduct preliminary experiments) -> `/exp-design `-> For each experimental block,recommended flow: `/exp-run [--env local|remote]` to deploy → `/exp-status` to monitor → `/exp-run --collect` to collect.->`/exp-eval ` ✨ : New Skills `/exp-pilot-run` — Pilot experiment execution: write code, deploy, monitor, collect raw results. `/exp-pilot-eval` — Pilot result evaluation: read results, apply lenient verdict logic These two skills are built into Phase5 of `/ideate` 🛠️ : Modified Skills `/ideate` 5 structured generation paths (A-E) for both Claude and Review LLM. Phase restructuring: Filter & Validation merged into Phase 3, Write Wiki moved to Phase 4. Phase 5: Finish pilot design and workflow invocation Your ideas will follow a clearer path, and a more reasonable screening mechanism will be established through pilot experiments. `/exp-design` A brand-new experimental design process:method candidate generation + 5 experiment block types + iterative ablation loop `/exp-run` Add the code decision gate, code optimization and config check ### 🎨 2026-05-18 · /poster — drafted paper → print-ready conference poster Run the poster skill after paper-draft and paper-compile (`/poster` in Claude Code, `$poster` in Codex) to turn your finished draft into a self-contained 1400×900 HTML poster and a print-quality PNG. Figures, booktabs tables, and math macros are extracted automatically from your LaTeX source; the agent walks you through picking which figures land in which sections and customizing the header (venue, affiliation logo). Export to PDF from your browser's print dialog. Pipeline adapted from [PaperX](https://github.com/yutao1024/PaperX) ([arXiv:2602.03866](https://arxiv.org/abs/2602.03866)).

Example /poster output

### 🎯 2026-05-12 · /discover from a venue — "what should I read first from ICLR 2024?" Use `/discover --venue iclr --year 2024` in Claude Code or `$discover --venue iclr --year 2024` in Codex (or any conference/year) and get a personalized shortlist of papers from that venue, ranked by relevance to what's already in your wiki. Instead of scrolling a 7000-paper proceedings, you see the dozen that actually matter for your research direction, each with a rationale tied to topics and methods you already track. No new API keys, no ingest side-effects on your wiki — just a ranked reading list. Supports NeurIPS, ICLR, ICML, and other venues covered by [Paper Copilot](https://github.com/papercopilot/paperlists). ### 📰 2026-05-09 · Daily arXiv — fresh-paper recommendations, on demand or scheduled Use `/daily-arxiv` in Claude Code or `$daily-arxiv` in Codex for a one-off pass. The GitHub Actions scheduler supports Codex CLI for unattended `inform` recommendations, with legacy Claude Code Action and Review LLM fallbacks; CI `auto-ingest` remains on the legacy Claude Action path until Codex writeback is separately verified. The skill builds an evidence packet from arXiv + Semantic Scholar + DeepXiv, lets the LLM rank candidates against your wiki interests, and delivers a digest by e-mail. Explicit `--mode auto-ingest` calls the ingest skill for high-confidence picks; `inform` mode just notifies. ### 🌐 2026-05-06 · Knowledge Graph Visualization — browser + Obsidian Your research graph now has two ways to explore: - **Web UI** — run `python3 tools/serve.py`, open `http://localhost:8765/#/graph`. Click any node to highlight its neighborhood via BFS, filter by entity type or edge category, double-click to open the full page in the Reader. - **Obsidian** — run `/visualize --obsidian` to generate a color-coded graph config, or `/visualize --canvas` to produce a force-layout Canvas with labeled semantic edges. --- ## What is AutoSci? Scientific research has traditionally been **human-intensive**: researchers coordinate literature, ideas, experiments, manuscripts, and review responses across long project cycles. **AutoSci** is a memory-centric agentic system that automates the full research lifecycle — from paper ingestion to rebuttal — while maintaining structured persistent memory across projects and improving its own procedures over time.

--- ## 🔬 Works Produced with AutoSci The following papers were generated end-to-end using AutoSci — from literature ingestion and idea generation to experiment execution and manuscript writing. | Paper | Domain | PDF | |-------|--------|-----| | Agent-driven iterative optimization of Triton GPU kernels | GPU kernel optimization | [📄 PDF](assets/papers/gpu-kernel-optimization.pdf) | | PTM-aware degrader target nomination via calibrated ternary-complex scoring | Biomedical drug discovery | [📄 PDF](assets/papers/protac-target-nomination.pdf) | | Forced Honesty Dissociates Polite Speech from Motivated Cognition in LLM Attitude Ratings | LLMs as cognitive models | [📄 PDF](assets/papers/llm-positivity-bias-cognitive-models.pdf) | **Have you used AutoSci in your own research?** We'd love to feature your work here — open a PR or drop us a message! --- ## OpenCode Preview Quick Start **Prerequisites:** Python 3.9+, Node.js 18+, and OpenCode ```bash # 1. Clone the OpenCode Preview branch git clone -b autosci-opencode https://github.com/skyllwt/AutoSci.git cd AutoSci # 2. Verify OpenCode opencode --version # 3. Generate the local environment, skills, AGENTS.md, and opencode.json chmod +x setup.sh && ./setup.sh --lang en # 4. Add your papers and optional notes # raw/papers/ raw/notes/ raw/web/ # 5. Start OpenCode and load the init skill opencode ``` The generated `.opencode/` tree and `opencode.json` are machine-local and should not be committed. The English and Chinese sources under `i18n/` remain authoritative. --- ## Codex Preview Quick Start **Prerequisites:** Python 3.9+, Node.js 18+ ```bash # 1. Clone the Codex Preview branch git clone -b autosci-codex https://github.com/skyllwt/AutoSci.git cd AutoSci # 2. Install and sign in to Codex # Follow your Codex/OpenAI setup path, then verify: codex --version # 3. One-click setup chmod +x setup.sh && ./setup.sh # Linux / macOS # Windows (PowerShell): # powershell -ExecutionPolicy Bypass -File .\setup.ps1 # setup creates .venv and syncs the Codex `.agents/skills` tree # 4. Put your own papers in raw/papers/ (.tex or .pdf) # Optional: intent notes in raw/notes/, saved pages in raw/web/ # 5. Build your research memory and start a project codex # Then invoke: $init [your-research-topic] ``` Claude Code users should use the stable `main` branch instead: ```bash git clone -b main https://github.com/skyllwt/AutoSci.git cd AutoSci npm install -g @anthropic-ai/claude-code claude login claude # Then type: /init [your-research-topic] ```

Manual setup (Linux / macOS)

```bash python3 -m venv .venv && source .venv/bin/activate pip install -r requirements.txt cp .env.example .env # Edit to add API keys mkdir -p .agents/skills/shared-references cp -R i18n/en/skills/. .agents/skills/ cp i18n/en/shared-references/*.md .agents/skills/shared-references/ mkdir -p .claude/skills/shared-references cp -R i18n/en/skills/. .claude/skills/ cp i18n/en/shared-references/*.md .claude/skills/shared-references/ cp config/settings.local.json.example .claude/settings.local.json # Claude Code compatibility ```

Manual setup (Windows / PowerShell)

```powershell python -m venv .venv .\.venv\Scripts\Activate.ps1 pip install -r requirements.txt Copy-Item .env.example .env # Edit to add API keys New-Item -ItemType Directory -Force .agents\skills\shared-references | Out-Null Copy-Item i18n\en\skills\* .agents\skills -Recurse -Force Copy-Item i18n\en\shared-references\*.md .agents\skills\shared-references -Force New-Item -ItemType Directory -Force .claude\skills\shared-references | Out-Null Copy-Item i18n\en\skills\* .claude\skills -Recurse -Force Copy-Item i18n\en\shared-references\*.md .claude\skills\shared-references -Force Copy-Item config\settings.local.json.example .claude\settings.local.json # Claude Code compatibility ``` Note: native Windows is supported for the local pipeline. Remote-GPU experiments via `/exp-run --env remote` rely on `ssh`/`rsync`/`screen` and are best run from WSL2 or Linux/macOS.

### API Keys | Key | Required? | How to get | What it enables | |-----|-----------|-----------|-----------------| | Agent runtime auth | **Yes** | Claude Code: `claude login`; Codex: sign in through Codex | Powers the interactive coding-agent skills | | `OPENAI_API_KEY` or `CODEX_ACCESS_TOKEN` | Optional | OpenAI / Codex account | GitHub Actions Codex CLI recommender for daily-arxiv inform mode | | `ANTHROPIC_API_KEY` | Claude Code only (or use a third-party compatible API — see below) | `claude login` (automatic) | Powers Claude Code skills | | `CLAUDE_CODE_OAUTH_TOKEN` | Optional | `claude setup-token` | GitHub Actions legacy Claude Code auth for Pro/Max users and daily-arxiv auto-ingest | | `SEMANTIC_SCHOLAR_API_KEY` | Optional | [semanticscholar.org/product/api](https://www.semanticscholar.org/product/api) (free) | Citation graph, paper search | | `DEEPXIV_TOKEN` | Optional | `setup.sh` auto-registers | Semantic search, TLDR, trending | | `LLM_API_KEY` + `LLM_BASE_URL` + `LLM_MODEL` | Optional | Any OpenAI-compatible API | Cross-model review; `/daily-arxiv` inform recommendations | > **Don't have an Anthropic API key?** You can use Codex, or use Claude Code with any Anthropic-protocol-compatible provider — DeepSeek, Kimi, MiMo, GLM, and more. See the [LLM API Configuration](#llm-api-configuration--大模型-api-配置) section below for Claude Code provider snippets. > **Cross-model review**: AutoSci uses a second LLM as an independent reviewer for ideas, experiments, and paper drafts. Works with **any OpenAI-compatible API** — DeepSeek, OpenAI, Qwen, OpenRouter, SiliconFlow, etc. If not configured, skills still work in single-agent mode. --- ## LLM API Configuration / 大模型 API 配置 AutoSci runs on **Claude Code** or **Codex**. Claude Code speaks the **Anthropic API** protocol: you can use Claude directly, or route Claude Code to any third-party provider that exposes an Anthropic-compatible endpoint by overriding a few environment variables. Codex uses the Codex/OpenAI sign-in path and reads the repo skills from `.agents/skills`. AutoSci 支持 **Claude Code** 与 **Codex**。Claude Code 使用 **Anthropic API** 协议通信：你既可以直接使用 Claude, 也可以通过覆盖几个环境变量, 把 Claude Code 指向任意支持 Anthropic 协议的第三方供应商。Codex 使用 Codex/OpenAI 登录路径，并从 `.agents/skills` 读取 repo skills。 ### Option A — Native Claude / 原生 Claude ```bash claude login # OAuth, no manual config / OAuth 登录,无需手动配置 ``` ### Option B — Third-party Anthropic-compatible API / 第三方 Anthropic 兼容 API Pick a provider below, paste the snippet into `~/.claude/settings.json` (or the project's `.claude/settings.json`), and replace the `<...>` placeholder with your own API key. Model names and extra options follow each provider's official Claude Code docs. 从下方任选一个供应商,把对应配置粘贴到 `~/.claude/settings.json`(或项目的 `.claude/settings.json`),并把 `<...>` 占位符替换为你自己的 API key。模型名与额外选项均来自各供应商官方 Claude Code 文档。

MiMo / DeepSeek / Kimi / GLM 配置示例

#### MiMo (小米) ```json { "env": { "ANTHROPIC_BASE_URL": "https://api.xiaomimimo.com/anthropic", "ANTHROPIC_AUTH_TOKEN": "", "ANTHROPIC_MODEL": "mimo-v2.5", "ANTHROPIC_DEFAULT_SONNET_MODEL": "mimo-v2.5", "ANTHROPIC_DEFAULT_OPUS_MODEL": "mimo-v2.5-pro", "ANTHROPIC_DEFAULT_HAIKU_MODEL": "mimo-v2.5" } } ``` #### DeepSeek ```json { "env": { "ANTHROPIC_BASE_URL": "https://api.deepseek.com/anthropic", "ANTHROPIC_AUTH_TOKEN": "", "ANTHROPIC_MODEL": "deepseek-v4-pro[1m]", "ANTHROPIC_DEFAULT_OPUS_MODEL": "deepseek-v4-pro[1m]", "ANTHROPIC_DEFAULT_SONNET_MODEL": "deepseek-v4-pro[1m]", "ANTHROPIC_DEFAULT_HAIKU_MODEL": "deepseek-v4-flash", "CLAUDE_CODE_SUBAGENT_MODEL": "deepseek-v4-flash", "CLAUDE_CODE_EFFORT_LEVEL": "max" } } ``` #### Kimi (Moonshot) ```json { "env": { "ANTHROPIC_BASE_URL": "https://api.moonshot.ai/anthropic", "ANTHROPIC_AUTH_TOKEN": "", "ANTHROPIC_MODEL": "kimi-k2.5", "ANTHROPIC_DEFAULT_OPUS_MODEL": "kimi-k2.5", "ANTHROPIC_DEFAULT_SONNET_MODEL": "kimi-k2.5", "ANTHROPIC_DEFAULT_HAIKU_MODEL": "kimi-k2.5", "CLAUDE_CODE_SUBAGENT_MODEL": "kimi-k2.5", "ENABLE_TOOL_SEARCH": "false" } } ``` #### GLM (Z.AI) ```json { "env": { "ANTHROPIC_BASE_URL": "https://api.z.ai/api/anthropic", "ANTHROPIC_AUTH_TOKEN": "", "API_TIMEOUT_MS": "3000000" } } ``` > Z.AI applies a default server-side model mapping, so no explicit `ANTHROPIC_MODEL` is needed. > Z.AI 默认在服务端做模型映射,无需显式设置 `ANTHROPIC_MODEL`。

**Skip the Claude Code onboarding** / **跳过 Claude Code 初始引导**: when using a third-party key, create or edit `.claude.json` (`~/.claude.json` on macOS/Linux) and add `{ "hasCompletedOnboarding": true }`. --- ## Skills AutoSci ships with 30+ agent skills spanning the full research lifecycle. - Claude Code: invoke skills as slash commands, for example `/init`. - Codex: invoke skills with `$skill-name` or from `/skills`, for example `$init`.

View all skills

Each skill has the same name in both runtimes. Use the Claude Code slash form inside Claude Code, and the Codex dollar form inside Codex or select the skill from Codex `/skills`. ### Phase 0: Setup | Skill | Claude Code | Codex | What it does | |-------|-------------|-------|-------------| | setup | `/setup` | `$setup` | Interactive API key configuration — checks `.env` state and walks through Semantic Scholar, DeepXiv, and Review LLM setup | | reset | `/reset` | `$reset` | Destructive cleanup — reset wiki state to a clean scaffold by scope (`wiki / raw / log / checkpoints / all`) | ### Phase 1: Knowledge Base | Skill | Claude Code | Codex | What it does | |-------|-------------|-------|-------------| | prefill | `/prefill` | `$prefill` | Seed `wiki/foundations/` with domain background so later ingest runs do not create duplicate concept pages for textbook material | | init | `/init` | `$init` | Bootstrap the wiki from your source files, with optional discovery, then ingest the final paper set serially by default, with optional parallel worktree mode | | ingest | `/ingest` | `$ingest` | Ingest a paper (local path or arXiv URL) — creates pages and builds all cross-references and graph edges | | discover | `/discover` | `$discover` | Build a ranked shortlist of candidate papers (anchor-driven, topic-driven, venue-filtered, or from wiki state) without ingesting | | edit | `/edit` | `$edit` | Add or remove raw sources, or update wiki content, per user request | | ask | `/ask` | `$ask` | Ask the wiki a question — retrieve and synthesize relevant pages, optionally crystallize the answer back into the wiki | | check | `/check` | `$check` | Scan the full wiki to detect health issues and produce a tiered fix-recommendation report | ### Phase 2: Ideation & Experiments | Skill | Claude Code | Codex | What it does | |-------|-------------|-------|-------------| | daily-arxiv | `/daily-arxiv` | `$daily-arxiv` | Run or schedule the daily arXiv recommendation feed; delivers a ranked digest by email with optional auto-ingest for high-confidence picks | | ideate | `/ideate` | `$ideate` | Multi-phase research idea generation: landscape scan → dual-model brainstorm → filter & validation → write to wiki → pilot | | exp-pilot-run | `/exp-pilot-run` | `$exp-pilot-run` | Pilot experiment execution — write code, deploy, monitor, collect raw results as part of the ideation pipeline | | exp-pilot-eval | `/exp-pilot-eval` | `$exp-pilot-eval` | Pilot result evaluation — read results, apply success criteria, update idea page as part of the ideation pipeline | | novelty | `/novelty` | `$novelty` | Multi-source novelty verification via WebSearch + Semantic Scholar + wiki + Review LLM; outputs novelty score and recommendations | | review | `/review` | `$review` | Cross-model review of any research artifact — outputs structured scores, wiki entity mapping, and improvement suggestions | | exp-design | `/exp-design` | `$exp-design` | Idea-driven experiment design with iterative ablation — method candidates → benchmark selection → sensitivity analysis → main experiment | | exp-run | `/exp-run` | `$exp-run` | Full experiment execution pipeline — prepare code → deploy → monitor → collect results | | exp-status | `/exp-status` | `$exp-status` | View the status of all running experiments; optionally auto-collect completed runs and advance the pipeline | | exp-eval | `/exp-eval` | `$exp-eval` | Experiment verdict gate — Review LLM independently judges results and auto-updates the linked idea's status and graph edges | | refine | `/refine` | `$refine` | Multi-round iterative improvement — repeatedly reviews an artifact, parses feedback, applies fixes, and updates wiki until target score | ### Phase 3: Writing & Dissemination | Skill | Claude Code | Codex | What it does | |-------|-------------|-------|-------------| | survey | `/survey` | `$survey` | Generate a Related Work section from wiki knowledge — thematic grouping → narrative structure → LaTeX output | | paper-plan | `/paper-plan` | `$paper-plan` | Compile a paper outline from the idea graph — evidence map → narrative structure → section + figure + citation plan | | paper-draft | `/paper-draft` | `$paper-draft` | Draft a LaTeX paper from `PAPER_PLAN` — write each section from wiki sources, generate figures/tables, verify BibTeX | | paper-compile | `/paper-compile` | `$paper-compile` | LaTeX compile → PDF — latexmk compile + auto-fix + page count / anonymity / font checks + submission checklist | | research | `/research` | `$research` | End-to-end research orchestrator — idea discovery → experiment design → execution → verdict → paper writing with human gates | | rebuttal | `/rebuttal` | `$rebuttal` | Parse review comments → atomize concerns → map to wiki → stress-test with Review LLM → generate rebuttal | | poster | `/poster` | `$poster` | Generate an academic poster from a drafted paper — distill sections into a single-page HTML poster with figures | ### Utilities | Skill | Claude Code | Codex | What it does | |-------|-------------|-------|-------------| | visualize | `/visualize` | `$visualize` | Generate Obsidian graph configs and Canvas knowledge maps; the interactive web graph is served by `tools/serve.py` |

--- ## Contributing We welcome contributions and feedback — especially while we're in active iteration. See [CONTRIBUTING.md](CONTRIBUTING.md). ## Community / 交流群 WeChat Group QR Code

Scan to join the AutoSci WeChat group / 扫码加入微信交流群 ## Citation If you find AutoSci useful in your research, please cite our paper: ```bibtex @misc{qian2026autosci, title={AutoSci: A Memory-Centric Agentic System for the Full Scientific Research Lifecycle}, author={Weitong Qian and Beicheng Xu and Zhongao Xie and Bowen Fan and Guozheng Tang and Jiale Chen and Xinzhe Wu and Mingtian Yang and Chenyang Di and Jiajun Li and Lingching Tung and Peichao Lai and Yifei Xia and Ziyi Guo and Yanwei Xu and Yanzhao Qin and Shaoduo Gan and Xupeng Miao and Bin Cui}, year={2026}, eprint={2605.31468}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2605.31468}, } ``` ## Acknowledgments - **[Claude Code](https://docs.anthropic.com/en/docs/claude-code)** and **[Codex](https://developers.openai.com/codex)** — supported coding-agent runtimes for AutoSci - The `/poster` pipeline is adapted from [PaperX](https://github.com/yutao1024/PaperX) ## License [MIT](LICENSE) — use it, fork it, build on it. ---

**Built for [Claude Code](https://docs.anthropic.com/en/docs/claude-code) and [Codex](https://developers.openai.com/codex)** If this project helps your research, give it a ⭐