架构与运行逻辑¶

IndustrialAgentFlow 的主线由四层组成： 1. 数据入口层
负责 YAML 导入、点位绑定和快照上下文构建。 2. Agent 层
Parser 负责字段映射，Generator 负责草案生成，Critic 负责失败修正指令。 3. 确定性校验层
执行模板校验与评分阈值检查，决定是否允许发布。 4. 执行评测层
运行预测优化仿真和批量评测。

主链路时序¶

sequenceDiagram
    participant User as User
    participant API as API
    participant Loader as CatalogLoader
    participant Repo as CatalogBundleRepo
    participant Parser as ParserAgent
    participant Generator as GeneratorAgent
    participant Gate as Validator+QualityGate
    participant Critic as CriticAgent
    participant TplRepo as TemplateRepo
    participant Pipeline as SimulationPipeline

    User->>API: POST /v1/catalogs/import
    API->>Loader: load(yaml, mode)
    Loader-->>API: CatalogBundle
    API->>Repo: put(bundle)
    Repo-->>API: ok

    User->>API: POST /v1/agentic/run-from-catalog
    API->>Repo: get(catalog_id)
    Repo-->>API: CatalogBundle
    API->>Parser: build parser result from bindings
    API->>Generator: generate draft
    API->>Gate: validate + quality-check
    alt gate pass
        API-->>User: AgenticRunReport(approved)
    else gate fail
        API->>Critic: review(failed_draft, errors)
        Critic-->>API: correction instruction
        API->>Generator: regenerate with instruction
        API->>Gate: validate + quality-check
    end

    User->>API: POST /v1/templates/publish
    API->>TplRepo: publish(template)

    User->>API: POST /v1/pipeline/simulate
    API->>Pipeline: run(context, template)
    Pipeline-->>User: PipelineResult

关键设计点¶

LLM 不直接决定发布，发布由确定性检查决定。
点位输入支持大规模 legacy YAML，不要求逐点手填。
存储支持内存和文件后端，开发和轻量部署都可用。
LLM 失败策略可配置，默认自动降级，支持严格失败。