OpenMAIC 教学:课堂、幻灯片、交互组件与苏格拉底式教学。
安装
# GitHub 源码(首次需按提示配置 allowBuilds 构建授权后重试)
dsh plugin --profile web add github:THU-MAIC/dsh-openmaic
GitHub 来源的插件在安装时会在你的机器上执行构建脚本。请只安装可信来源,并尽量锁定 commit(github:owner/repo#sha)。
README
把 OpenMAIC 带进 DeepSeek Harness。Bring OpenMAIC into DeepSeek Harness.
dsh-openmaic is a DeepSeek Harness plugin that registers four tools and a
Socratic teaching skill:
openmaic_generate: tell your agent "make me a lesson about X", and the plugin submits the requirement to open.maic.chat, waits for the async generation job, and returns a playable classroom link.openmaic_slide: the agent writes one OpenMAIC slide (PPTist-style Slide JSON) and the plugin renders it with OpenMAIC's official renderer (text, shapes, images, tables, charts, formulas, code).openmaic_widget: the agent writes an OpenMAIC-style interactive widget (simulation, game, or code) per the bundled contract; the code streams as it writes, then renders inline as a sandboxed card.openmaic_render: the agent writes an inline HTML teaching fragment (concept card, quiz, walkthrough) and the plugin renders it as a sandboxed card right in the conversation.openmaic-teachskill: turns a session into a Socratic OpenMAIC lesson, teaching by guided questioning and pulling in slides, widgets, and cards as aids.
What it looks like
用户: 帮我做一节量子物理入门课
模型 → openmaic_generate(requirement="量子物理入门课", language="zh-CN")
← "Classroom ID: class-abc123
Classroom URL:
https://open.maic.chat/classroom/class-abc123"
模型: 课堂已经生成好了,点开就能上课:
https://open.maic.chat/classroom/class-abc123
Interactive widget:
用户: 做一个抛体运动模拟器
模型 → 按 openmaic-widget 模板写完整 HTML(流式输出)
→ openmaic_widget(html="<!doctype html>…", widgetType="simulation", title="抛体运动")
← "Rendered the simulation widget …"
对话里就地出现一个可交互的 OpenMAIC 模拟器
Install
dsh plugin --profile web add git+https://github.com/THU-MAIC/dsh-openmaic.git
Then restart dsh web and refresh. The plugin ships its compiled lib/, so a
git install needs no build step.
Config
dsh-openmaic:
baseUrl: https://open.maic.chat
accessCode: "" # invite code; not enforced online yet, leave empty
pollIntervalMs: 5000
maxWaitMs: 600000
| Key | Default | Notes |
|---|---|---|
baseUrl |
https://open.maic.chat |
API base. Point at http://localhost:3000 to develop against a local OpenMAIC. |
accessCode |
"" |
Invite code for open.maic.chat. Not enforced online yet, leave empty; fill it in once enabled. |
pollIntervalMs |
5000 |
Poll interval in ms. Generation is slow, so 60000 is friendlier than the default. |
maxWaitMs |
600000 |
Cap for one job, 10 minutes. |
API flow
- If
accessCodeis set,POST /api/access-code/verifyand replay theopenmaic_accesscookie on later requests. POST /api/generate-classroomwith the requirement, plus only the optional flags you passed. Returns ajobIdandpollUrl.- Poll
GET {pollUrl}until the job issucceededorfailed, ormaxWaitMsruns out. - On success, return
{baseUrl}/classroom/{classroomId}(or the server-providedresult.url).
Scope
openmaic_generate: generate a classroom and return a playable link.openmaic_slide: render one OpenMAIC slide with the official renderer.openmaic_widget: render a simulation / game / code widget the agent writes (a full HTML document). It streams the code while the agent writes it and renders on completion.openmaic_render: render an inline HTML teaching fragment as a sandboxed card.openmaic-teach: Socratic teaching session that uses the tools above as aids.
The slide/widget/render tools do no server-side generation; they render content
the agent authors against the OpenMAIC SDK contracts (@openmaic/dsl,
@openmaic/generation, @openmaic/renderer).
Roadmap
- Wire the remaining widget types (diagram, visualization3d, procedural-skill).
- Action loop back to the model (teaching-agent interactions: highlight/annotate/reveal widget elements).
Development
./scripts/build.sh # links host deps, bundles src/ to lib/ with tsdown
./scripts/test.sh # links host deps, runs the vitest suite
The scripts locate the harness checkout from dsh on PATH; set DSH_CHECKOUT to build against a specific checkout.
License
MIT
链接
同类插件
liustack/modlens★ 1199
为纯文本模型架起视觉桥梁:粘贴图片,输出结构化 JSON 证据(OCR、版面、语义)。
Anionex/dsh-vision-toolkit★ 308
让纯文本模型更好地做视觉任务:带意图的图片问答、长截图 OCR、UI 还原等。
zhaoolee/notes★ 138
将 DSH 对话导出为锤子便签风格 PNG,或在配置的账号工作区中新建和更新 Markdown 便签。
liustack/modsearch★ 85
纯文本 agent 的联网搜索桥:搜索网页与 X,返回结构化 JSON 证据(search/fetch/引用)。
Lum1104/dsh-browser★ 80
Chrome 侧边栏扩展,让 DSH 直接操控你的浏览器,无需视觉能力。
taxueseek/argo★ 69
专为 agent 打造的搜索工具:多语言,覆盖中文/英文/学术/代码/购物/金融/新闻/百科。