mm-agent
v0.0.1
Published
Reserved placeholder for mm-agent — a thin flow that turns a mathematical modeling problem into gated artifacts, reproducible computations and a compiled PDF report. OpenCode is the first adapter. This package only reserves the name on npm; the real imple
Maintainers
Readme
mm-agent
A thin flow that turns a mathematical-modeling problem into gated artifacts, reproducible computations, and a compiled PDF report. OpenCode is the first adapter.
Status
This [email protected] package only reserves the mm-agent name on npm. It ships no usable functionality — index.js is empty. Do not depend on it for behavior.
The real implementation will be released as [email protected] under this same name once the first vertical chain is verified end-to-end. When that happens, npm install mm-agent will fetch the real package automatically (npm resolves to the highest published version).
What mm-agent will be
mm-agent is a harness that runs the four-stage mathematical-modeling pipeline on top of an AI coding assistant:
problem input -> four-stage workflow -> compilable LaTeX -> PDF reportThe four stages:
- Problem Analysis — decompose the problem into explicit tasks and a legal dependency DAG.
- Mathematical Modeling — select and justify methods, retrieve evidence from a method library.
- Computational Solving — execute each DAG task and leave reproducible compute evidence.
- Solution Reporting — write the report from accepted facts and actually compile it to a non-empty PDF.
Each stage produces gated artifacts — nothing advances until a Critic pass and a deterministic Gate validate the candidate. Restoration only reads on-disk case facts.
Architecture
mm-agent is built around a Core / Adapter split:
- Canonical Core is host-agnostic: the thin flow, the case protocol, the gate, the artifact rules.
- Adapters map a specific host's session, delegation, and tool capabilities onto the Core.
The first adapter targets OpenCode via its Plugin API (custom tools, a config hook for hidden agents, and a task hook for fresh child sessions). It injects five hidden agents — mm-analyst, mm-modeler, mm-solver, mm-writer, mm-critic — and six model-visible tools.
Future adapters may target other AI coding hosts. Each adapter requires real engineering; copying a Skill file to another host does not make the tools exist there.
Roadmap
| Version | Status | Contents |
|---|---|---|
| 0.0.1 | this release | Name reservation only. Empty. |
| 1.0.0 | In development | OpenCode adapter, full four-stage flow, gated artifacts, compiled PDF, end-to-end vertical chain verified via npm pack + external-project install. |
Why a placeholder
TypeScript + Plugin-API code is published under the real package.json once the first vertical chain (Analyst → Modeler → Solver → Writer → Critic → non-empty PDF) passes end-to-end acceptance from an external .tgz install. Until then, this empty release holds the name so the eventual real release can ship as the canonical mm-agent rather than a scoped fallback.
Source
关于
mm-agent 面向数学建模竞赛(如 CUMCM / MCM)的端到端流程:赛题输入 -> 四阶段工作流 -> 可编译 LaTeX -> PDF 论文。当前 0.0.1 仅用于在 npm 上保留包名,不含任何可用功能;首条纵向链端到端验证通过后将以 1.0.0 正式发布。OpenCode 是首个 Adapter。
License
MIT © Zhiyan Cheng
