OpenMAIC
OpenMAIC 未来与实践项目：构建可复现课堂工作室
围绕可回放课程会话、Provider 能力清单、评估、无障碍和治理导出，设计一个 OpenMAIC 实践项目。
你将学会
- Design a replayable classroom package
- Define provider and quality governance
- Plan a reversible roadmap project
开始前需要
- Basic Git and command-line usage
- Comfort reading a project README
You can explain the project, run its documented first step, and decide what to verify before adopting it.
先看结论
- A replayable classroom package can carry provenance from source to accepted lesson.
- Capability manifests make provider fallbacks visible and governable.
- A narrow replay studio with accessibility and safety fixtures is a useful capstone.
The next useful abstraction
OpenMAIC 的下一步不只是增加场景类型，而是让课程包携带学习目标、材料哈希、代理计划、Provider 能力、产物、无障碍检查和审核决定；本文据此设计一个单 Provider、单导出的可回放课堂工作室，并把 trace-only、advisory、enforced 作为渐进式治理路径。
A capability manifest could declare whether a provider supports chat, images, TTS, ASR, search, persistence, and export, along with data residency, latency, and cost limits. The workbench could then explain a fallback instead of silently switching models or media services.
Capstone: the classroom replay studio
Build a small service that accepts one public lesson fixture, runs a pinned OpenMAIC revision, and stores a replay bundle. Add a reviewer view with source hashes, agent events, tool calls, artifact previews, accessibility findings, learner rubric results, and approval state. The first release should support one provider and one export format so every field can be replayed.
Generate a deterministic replay in CI: the same material and prompt should produce a comparable event trace even if model text varies. Compare normalized structure, citations, and accessibility outcomes rather than pretending natural-language bytes are identical.
Research and governance roadmap
Evaluate citation fidelity, learner comprehension, hallucination rate, accessibility, p95 latency, storage, provider cost, and reviewer time. Add adversarial fixtures for prompt injection in documents, malicious links, oversized media, stale sessions, and duplicate exports. Publish failures and rejected lessons with the same care as successful demos.
For adoption, add signed lesson manifests, scoped credentials, retention controls, tenant isolation, and a deletion audit. A future OpenClaw or messaging integration should pass the same fixture and safety suite before it is marked compatible.
A reversible path forward
Start with trace-only replay, then add advisory quality and accessibility checks, then enforce publication gates for a small cohort. Keep the original materials, accepted artifact, and native fallback available. These proposals are a roadmap, not current guarantees; the practical contribution is a narrow fixture, measurable contract, and evidence that another educator can inspect.
如何选择
| 比较维度 | 方案 A | 方案 B |
|---|---|---|
| Best when | You need predictable behavior and easy auditing | You need adaptive optimization and have reliable telemetry |
| Main risk | May leave performance on the table | Can become difficult to explain or debug |
实施步骤
- 1
Choose one public lesson fixture, provider, and export.
- 2
Store hashes, agent events, tool calls, artifacts, checks, and reviewer state.
- 3
Replay in CI and evaluate fidelity, accessibility, cost, latency, and review.
- 4
Roll out trace-only → advisory → enforced publication gates with fallback.
可复制示例
{
"lesson_revision": "<commit>",
"materials": [{ "sha256": "<hash>" }],
"capabilities": { "chat": true, "export": ["html"] },
"events": [],
"accessibility": { "status": "review" },
"approval": { "status": "pending" }
}常见问题
Is a replay studio part of OpenMAIC today?
No. It is a proposed capstone that builds on the repository's workbench, durable sessions, skills, providers, and exports.
How should a team adopt future checks?
Begin with trace-only fixtures, then advisory checks, and enforce publication only after quality, accessibility, security, and rollback evidence is stable.
资料来源
- OpenMAIC README (captured 2026-08-31)来源核查 2026-08-31
- OpenMAIC repository来源核查 2026-08-31