运行时流程图

服务启动

sequenceDiagram
    participant M as main
    participant CLI as cmd.NewCLI
    participant S as server.Serve
    participant SC as Scheduler
    participant G as discover.GPUDevices
    participant H as http.Server
    M->>CLI: ExecuteContext
    CLI->>S: RunServer
    S->>S: fixBlobs + prune
    S->>S: GenerateRoutes
    S->>SC: InitScheduler / Run
    S->>G: GPUDevices
    S->>S: choose default context
    S->>H: Serve(listener)

Chat 请求

flowchart LR
    A[JSON request] --> B[bind + validate]
    B --> C[parse model ref]
    C --> D[GetModel]
    D --> E[merge options]
    E --> F[Scheduler.getRunner]
    F --> G[prompt/template]
    G --> H[runner Chat/Completion]
    H --> I[thinking/tool/structured parser]
    I --> J[NDJSON response]

Agent 请求

flowchart TB
    U[user message] --> M[Session.Run]
    M --> L[model step]
    L --> Q{tool calls?}
    Q -- no --> F[finish]
    Q -- yes --> A[approval]
    A --> X[execute batch]
    X --> T[append tool results]
    T --> C{over budget?}
    C -- no --> L
    C -- yes --> S[compact history]
    S --> L

模型拉取

flowchart TB
    R[model name] --> P[remote manifest]
    P --> D{local blob exists?}
    D -- yes --> N[next layer]
    D -- no --> DL[download chunks]
    DL --> V[verify digest/size]
    V --> N
    N --> Z[write manifest]
    Z --> READY[GetModel can resolve]