@mzzsfy/dsh-llm-pi-gateway

配置(settings.yaml 命名空间 llm-pi-gateway)

llm-pi-gateway:
  providers:
    new-api-claude:                 # 路由名 = LLM provider 名
      displayName: New API Claude   # 可选,配置面/选择器显示名
      api: anthropic-messages       # 或 openai-completions / openai-responses
      baseURL: https://newapi.example.com
      apiKeyEnv: NEW_API_API_KEY    # 凭据引用(credentials 服务或环境变量)
      reasoning: high               # 可选,路由级默认档位(模型需支持)
      sessionMarker:                # 会话标记(请求体通道),默认开启
        enabled: true
        prefix: dsh                 # 派生标记前缀
      metadata:                     # anthropic metadata 模板(可选);user_id 键由标记占用,勿在此声明
        gateway: newapi
        session: '{sessionId}'
      compat:                       # 路由级 compat 覆盖,无 withhold;字段按协议名单校验
        sendSessionAffinityHeaders: true   # 亲和头通道(anthropic 即 x-session-affinity)
      headers:                      # 静态自定义头(粘性兜底通道)
        x-gateway-group: pool-a
      cacheRetention: short         # 可选:none / short / long(none 会抑制亲和头)
      defaultContextWindow: 262144  # 可选,模型未声明时的兜底
      retryPolicy:                  # 可选,注册捕获,进 runtime 重试
        mode: normal
        maxRetries: 2
      models:
        - id: claude-sonnet
          name: Claude Sonnet
          input: [text, image]
          maxTokens: 16384
          # 档位字典:key = 档位,value = wire 拼写;未声明档位 = 不支持;
          # off 无值 = 支持且不发参数;false = 非推理模型
          reasoningEfforts:
            off:
            low: low
            high: high
            max: ultra
          compat:
            sendSessionAffinityHeaders: true
    new-api-gpt:                    # openai 系协议示例(sessionAffinityFormat 仅在此类协议合法)
      api: openai-responses         # 或 openai-completions
      baseURL: https://newapi.example.com/v1
      apiKeyEnv: NEW_API_API_KEY
      compat:
        sessionAffinityFormat: openai   # openai / openai-nosession / openrouter,头集合见注入通道表
      models:
        - id: gpt-4o

路由与模型的全量可配键以本包 settings schema 为准,示例未穷举。

配置修改即时生效(无需重启);解析失败时保留上一份好配置。纯 host 端,无 GUI。