LightChuan 是一个面向科研、项目开发、学术竞赛和专业竞赛的大型 Open Agent Skills 技能包。它把 Claude Code、Codex CLI、Gemini CLI、Grok CLI 和 DeepSeek 组织成一套本地可安装、可验证、可复用的多模型协同工作系统,覆盖搜索、规划、论文、图表、记忆、测试、审查和多模型 council 决策。
它不是为了做演示而设计,而是为了真正处理复杂任务:数学建模、Kaggle/天池、ICPC/ACM、AI/大模型竞赛、计算机设计大赛、科研论文、软件项目、数据分析和多学科交叉交付。
LightChuan 能显著增强工作流,但不承诺奖项、论文录用、榜单名次或无需人工审查的绝对正确性。
很多 agent skill 只解决一个局部问题。LightChuan 的目标是把多个高强度工作流做成一套能协同的技能系统。
核心产品是 skills 本身:每个子技能都能独立使用,组合起来又能形成完整的闭环作战流。
- 9 个可安装的
SKILL.md子技能,每个都能单独调用,也能被 orchestrator 统一编排; - 5 个面向 Claude Code / Codex 风格工作流的 slash commands;
- 面向搜索、图谱、图表、数据分析、论文、质量门、安装和证据回放的确定性脚本;
- 支持 Claude、Codex、Gemini、Grok、DeepSeek 的动态多 CLI 协同;
- 支持长期项目的后台记忆、checkpoint、矛盾检测和 resident action;
- 所有运行证据默认写入
.lightchuan/; - 内置九技能 golden gauntlet、质量 benchmark、需求审计、参考覆盖审计和外部结构对比。
安装后,在 Claude Code、Codex 或其他支持的 agent 里进入项目目录,运行:
/lightchuan start --task "Kaggle competition: EDA, baseline, validation, figures, and report" --workspace .
如果只想调用某一个核心技能,可以直接使用四个聚焦命令:
/lightchuan-search latest baselines and contest rules for this problem
/lightchuan-plan diagnose this repo and propose the strongest innovation plan
/lightchuan-paper inspect this manuscript and produce a CVPR/SCI-style revision plan
/lightchuan-figure plot --data results.csv --kind line --x epoch --y score --out .lightchuan/figures
运行完整质量门:
python scripts\lc_quality_benchmark.py --with-probe --strict示例论文展示了 LightChuan 从检索、选题、规划、实验、科研图表到 IEEEtran PDF 导出的完整工作流。
阅读完整 IEEE PDF 论文 · Raw PDF · LaTeX 源码
点击论文首页预览可打开完整 PDF。检索记录、图表、LaTeX 和可复现命令见 IEEE demo 产物页。
这四个技能可以单独调用,适合用户只需要某一类能力的场景。
| Slash command | Skill | 适合场景 | 典型产物 |
|---|---|---|---|
/lightchuan-search |
lightchuan-search |
文献、赛题规则、baseline、来源核验、Tavily/Firecrawl/Valyu/Grok 搜索流 | search ledger、网页抽取、证据摘要 |
/lightchuan-plan |
lightchuan-project-innovation-plan |
项目诊断、idea 创新、Graphify-style 项目映射、执行计划 | 项目图谱、任务包、时间线、风险、验证计划 |
/lightchuan-paper |
lightchuan-paper-writing |
中英文论文、SCI 一区、CCF-A、CVPR、ICML、ICLR、NeurIPS、竞赛论文 | target contract、source map、claim register、LaTeX/docx/PDF 计划 |
/lightchuan-figure |
lightchuan-figure-studio |
顶刊图表、Matplotlib/Seaborn、TikZ/PGFPlots、VLM 图表审查、敏感性分析图 | PNG/PDF/SVG/PGFPlots、统计标注、审查报告 |
这些技能通常通过 /lightchuan start、后台服务、checkpoint 或质量门自动参与。
| Skill | 适合场景 | 典型产物 |
|---|---|---|
lightchuan-memory-graph |
持久上下文、checkpoint、矛盾检测、Graphify-style 记忆 | .lightchuan/memory/graph.json、latest.md、vector index |
lightchuan-multi-cli-council |
Claude/Codex/Gemini/Grok/DeepSeek 多模型讨论、动态分工、加权共识 | worker prompts、status files、consensus.md |
lightchuan-code-review-ui-perf |
代码审查、安全、性能、架构、前端审美和响应式体验 | resident actions、review prompts、scorecards |
lightchuan-data-test-lab |
EDA、泄漏检查、验证集、交叉验证、ablation、Kaggle/天池模拟 | eda.md、baseline_metrics.json、OOF predictions |
lightchuan-orchestrator |
统一入口、模式切换、后台服务、dashboard、共识投票 | route JSON、sessions、dashboard、quality reports |
flowchart LR
User["用户 / Agent"] --> Slash["/lightchuan 与聚焦 slash commands"]
Slash --> Orchestrator["LightChuan Orchestrator"]
Slash --> Manual["Search / Plan / Paper / Figure"]
Orchestrator --> Router["动态任务路由"]
Router --> Council["Claude / Codex / Gemini / Grok / DeepSeek"]
Orchestrator --> Resident["常驻记忆与 action"]
Manual --> Evidence["证据产物"]
Resident --> Evidence
Council --> Consensus["加权共识"]
Evidence --> Quality["质量 benchmark"]
Quality --> Readme["README 级证据"]
LightChuan 不把模型角色硬编码死。Orchestrator 会根据任务类型、模型可用性、人工 override、成本约束和证据风险动态分工。常见路由如下:
- Claude:写作、论文结构、规划、LaTeX;
- Codex:代码、notebook、测试、图表、原型;
- Gemini:严谨性、长上下文批判、数学与统计审查;
- Grok:实时搜索、赛题规则、趋势、最新 baseline;
- DeepSeek:低成本独立复核、矛盾检查和二次验证。
LightChuan 可安装到这些本地 skill 和 command 目录:
| Agent | Skills 目录 | Commands 目录 |
|---|---|---|
| Claude Code | %USERPROFILE%\.claude\skills |
%USERPROFILE%\.claude\commands |
| Codex | %USERPROFILE%\.codex\skills |
%USERPROFILE%\.codex\commands |
| Gemini CLI | %USERPROFILE%\.gemini\skills |
%USERPROFILE%\.gemini\commands |
| Grok CLI | %USERPROFILE%\.grok\skills |
%USERPROFILE%\.grok\commands |
也可以直接在仓库中使用 scripts/lightchuan.ps1 或 scripts/lightchuan.sh。
必需:
- Python 3.10 或更新版本;
- Windows PowerShell,或 Linux/macOS/WSL 下的 Bash;
- 至少一个本地 agent CLI,例如 Claude Code、Codex CLI、Gemini CLI 或 Grok CLI。
推荐:
- 本地已登录四个 CLI:
claude、codex、gemini、grok; - 用于多终端启动的
tmux或 GNUparallel; - Git;
- 用于论文编译的 LaTeX 环境;
- 常用科研包:
pandas、numpy、matplotlib、seaborn、scikit-learn、scipy。
可选 API key:
| Key | 用途 |
|---|---|
DEEPSEEK_API_KEY |
低成本独立审查与验证 |
TAVILY_API_KEY |
搜索 |
FIRECRAWL_API_KEY |
网页抽取与 crawling |
VALYU_API_KEY |
检索与搜索 |
OPENAI_API_KEY |
可选 VLM 或模型 API 审查 |
ANTHROPIC_API_KEY |
可选 Claude API 路径 |
GEMINI_API_KEY 或 GOOGLE_API_KEY |
可选 Gemini API 审查 |
如果你的 CLI 已经登录,直接模型 API key 通常只在 API 专用路径里需要。
克隆仓库后运行:
python scripts\lc_doctor.py --json
python scripts\install_lightchuan.py --agent all --force只安装到某一个 agent:
python scripts\install_lightchuan.py --agent claude --force
python scripts\install_lightchuan.py --agent codex --force安装器会复制:
- 所有
lightchuan-*skill 目录; - 共享的
scripts/、references/、assets/、docs/、commands/; commands/*.md下的所有 slash command;- 仓库 README 文档:
README.md和README.en.md。
安装后请重启 Claude Code 或目标 agent,让新 skills 和 commands 重新加载。
当你希望 LightChuan 先生成 mission-control pack,再进入更重的工作流时,用 quickstart:
python scripts\lc_quickstart.py --recipe kaggle-ai --goal "leakage-safe baseline, validation, figures, and report" --workspace . --run-level light查看可用 recipes:
python scripts\lc_recipes.py list
python scripts\lc_recipes.py show math-modeling
python scripts\lc_recipes.py show research-paper
python scripts\lc_recipes.py show design-competition然后在 agent 里调用匹配的 slash command:
/lightchuan start --task "Kaggle/Tianchi competition: build EDA, validation, figures, and report" --workspace .
当任务很宽泛、还没完全拆清楚时,用 wizard 让 LightChuan 自动推断模式、recipe、证据计划和第一批命令:
python scripts\lc_wizard.py --goal "math modeling contest with optimization, sensitivity analysis, figures, and paper" --workspace . --run-level light --non-interactive多次运行后,可以生成本地导航页:
python scripts\lc_home.py --base .lightchuan --out .lightchuan\home.md使用环境变量或 LightChuan 本地配置助手。不要提交任何 key。
python scripts\lc_config.py set DEEPSEEK_API_KEY --stdin
python scripts\lc_config.py set TAVILY_API_KEY --stdin
python scripts\lc_config.py set FIRECRAWL_API_KEY --stdin
python scripts\lc_config.py set VALYU_API_KEY --stdin
python scripts\lc_config.py list
python scripts\lc_config.py path搜索 API 可能存在额度或付费计划。LightChuan 默认控制结果规模,并在某个引擎不可用时 fallback 到可用引擎。
/lightchuan start --task "math modeling contest: optimization, sensitivity analysis, figures, and paper" --workspace .
PowerShell 等价命令:
.\scripts\lightchuan.ps1 start --task "math modeling contest: optimization, sensitivity, figures, paper" --workspace ./lightchuan-search "latest transformer tabular competition baselines and leakage risks"
PowerShell:
.\scripts\lightchuan.ps1 search --query "contest rules baseline" --mode competition/lightchuan-plan "diagnose this repo, map the project graph, rank ideas, and produce a file-level execution plan"
PowerShell:
.\scripts\lightchuan.ps1 graph build --workspace .
.\scripts\lightchuan.ps1 skill-pack create --skill lightchuan-project-innovation-plan --goal "innovation and execution plan"/lightchuan-paper "inspect this manuscript and prepare a top-tier revision plan"
PowerShell:
.\scripts\lightchuan.ps1 paper inspect --root paper
.\scripts\lightchuan.ps1 paper compile --root paper --main main.tex/lightchuan-figure plot --data results.csv --kind line --x epoch --y score --out .lightchuan/figures
PowerShell:
.\scripts\lightchuan.ps1 figure plot --data results.csv --kind line --x epoch --y score --out .lightchuan\figures
.\scripts\lightchuan.ps1 vlm-review --path .lightchuan\figures --provider auto --detail low --max-images 6.\scripts\lightchuan.ps1 data-lab --data data\train.csv --target target --out .lightchuan\data_lab
.\scripts\lightchuan.ps1 template create --type all --out .lightchuan\templates.\scripts\lightchuan.ps1 service start --workspace . --name my-project
.\scripts\lightchuan.ps1 service status --name my-project
.\scripts\lightchuan.ps1 service stop --name my-projectWindows 下 install-onlogon 会优先尝试计划任务;如果策略阻止,则 fallback 到用户 Startup-folder 命令:
.\scripts\lightchuan.ps1 service install-onlogon --workspace . --name my-project当 prompt 出现竞赛、赛题、Problem、建模、Kaggle、天池、ICPC、设计大赛、人工智能大赛、大模型竞赛、多学科竞赛等关键词时,LightChuan 会自动切换到接近的内部模式。
| 模式 | 触发示例 | 内部重点 |
|---|---|---|
| 数学建模 | MCM、ICM、COMAP、建模、optimization、sensitivity | 假设、变量、优化、敏感性分析、论文资产 |
| Kaggle / AI | Kaggle、天池、leaderboard、submission、ensemble、LLM contest | 指标解析、泄漏审计、验证集、baseline、提交模拟 |
| ICPC / ACM | ICPC、ACM、Problem、constraints、complexity、stress test | 证明、复杂度、暴力 oracle、压力测试 |
| 设计 / 多学科 | 设计大赛、计算机设计大赛、prototype、UI、architecture | 原型、UI、系统架构、跨领域证据 |
| 科研论文 | CVPR、ICML、ICLR、NeurIPS、SCI、CCF、Overleaf、LaTeX | source map、claims、figures、venue template、review loop |
大多数输出会写到当前项目的 .lightchuan/:
| 路径 | 含义 |
|---|---|
.lightchuan/sessions/ |
worker prompts、模型输出、状态、共识 |
.lightchuan/memory/ |
checkpoints、项目图谱、vector index |
.lightchuan/resident/actions/ |
排队的代码、数据、论文和矛盾审查 action |
.lightchuan/search/ |
search ledger 和标准化来源证据 |
.lightchuan/data_lab/ |
EDA 报告、metrics、OOF predictions |
.lightchuan/figures/ |
生成图表和 vector exports |
.lightchuan/quality/ |
release-readiness scorecards |
.lightchuan/showcase/ |
可复现实战演示和证据包 |
这些 optional local evidence viewers 用于在写报告、README 或 release note 前查看本地证据:
python scripts\lc_dashboard_export.py --base .lightchuan --out .lightchuan\dashboard.html
python scripts\lc_session_replay.py --base .lightchuan --out .lightchuan\session_replay.html在信任一次 release 或安装前运行:
python scripts\lc_validate.py
python scripts\lc_skill_audit.py --strict
python scripts\lc_requirements_audit.py --strict
python scripts\lc_reference_coverage.py --strict
python scripts\lc_skill_benchmark.py --strict
python scripts\lc_external_skill_compare.py --strict
python scripts\lc_golden_gauntlet.py --strict
python scripts\lc_resident_probe.py --strict
python scripts\lc_ecosystem_scan.py --strict
python scripts\lc_e2e_paper_demo.py --strict
python scripts\lc_ieee_paper_demo.py --min-pages 10 --max-pages 20 --strict
python scripts\lc_quality_benchmark.py --with-probe --strictGolden gauntlet 是核心功能级证明。它会种子化一个本地项目,并在同一次可复现运行中覆盖九个子技能,报告:
score:端到端通过率;functional_score:每个技能是否完成真任务;rubric_score:产物是否满足质量 rubric;artifact_count:生成的证据产物数量。
当前验证快照,版本 2.0.0-beta.21:
| Gate | Result |
|---|---|
| Source validation | pass |
| Golden gauntlet | 100.0 |
| Quality benchmark with model probe | S / 100.0 |
| Ecosystem scan | skills.sh / GitHub skill lessons coverage |
| E2E paper demo | search, topic, plan, experiment, figure, LaTeX, PDF workflow |
| IEEE paper demo | 10-page IEEEtran PDF with multi-figure artifact pack |
| Installed Codex validation | pass |
| Secret scan | no committed API keys found |
LightChuan/
lightchuan-*/SKILL.md # 9 个 Open Agent Skills
commands/ # slash commands
scripts/ # deterministic orchestration and helper tools
references/ # protocols, playbooks, examples, comparison notes
assets/ # package metadata and recipe library
docs/ # architecture, roadmap, publishing gate
每个 skill 都链接到:
references/skill-playbooks/<skill>.md;references/skill-specializations/<skill>.md;references/skill-examples/<skill>.md;scripts/lc_skill_pack.py create --skill <skill>。
LightChuan 学习强 Agent Skills 和 agent-tool 项目的结构,再改造成适合本地跨 CLI 协同的技能包:
- Open Agent Skills / agentskills.io:可移植
SKILL.md、progressive disclosure、scripts/references/assets 分层; - ComposioHQ 与 VoltAgent:可发现性、工作流清晰度、多 agent 协调;
- K-Dense scientific skills:科研覆盖面与科研工具链;
- PaperSpine 与 scientific writing skills:argument spine、claim register、证据驱动修订;
- Graphify:图谱优先的项目理解;
- GadaaLabs / Claude Code on Steroids、oh-my-claudecode、CLI swarm patterns:多终端编排;
- ds-skills 与 Kaggle-style suites:泄漏检查、验证纪律、leaderboard simulation。
LightChuan 不复制这些项目,而是把它们作为质量标尺,用来约束 skill 结构、README 清晰度和可复现工作流。
LightChuan 适合把复杂任务拆成可执行、可审查、可复现的工作流,尤其是科研写作、项目开发、数据分析和竞赛交付。它会生成证据记录、计划、图表、代码、论文草稿和质量报告,帮助用户更快地组织高强度工作。
重要结论仍需要人工判断。竞赛成绩、论文录用、工程安全、数据合规和最终提交应由使用者结合领域专家意见确认。README 中的示例和质量门用于说明工作流能力,不代表任何奖项、榜单名次或论文录用保证。
- 更多 venue-specific 论文和竞赛模板;
- 比静态 HTML export 更强的实时 dashboard;
- 更多真实项目案例和经过检查的 README 证据摘要;
- 更强的 vector memory backend 和 semantic graph edges;
- 当 provider key 可用时,扩展 live VLM review 测试。
如果 LightChuan 对你有帮助,欢迎通过微信扫码打赏鼓励。建议、问题、使用反馈或合作想法可以发邮件到 1833058953@qq.com。
LightChuan 使用 MIT License。
你可以自由使用、复制、修改、分发、再授权和商业使用本项目,但需要保留版权与 license 声明。项目按 "as is" 提供,不附带任何担保。
