ZMLLMGW

An OpenCode plugin that exposes a configurable language-model gateway and optionally routes models through a secondary compatible endpoint.

Configuration

Create ~/.config/opencode/zmllmgw/config.yaml, or set ZMLLMGW_CONFIG to a configuration file:

providers:
  gateway:
    endpoint: https://gateway.example
  devmate:
    endpoint: https://secondary.example/api
  admin:
    endpoint: https://account.example

models:
  - gatewayId: example-model
    api: openai-completions
    tools: true
    autoContinue: true
    limit:
      context: 200000
      output: 32000

autoContinue:
  prompt: |-
    上一次回答因输出长度限制被截断。
    请从上一次回答结束的位置继续，不要重复已经输出的内容，继续完成原任务。
    由于存在输出长度限制，如果需要调用工具，请考虑少量多次调用，
    通过多次较短的 tool call 完成同样的任务，避免单次 tool call 过长导致生成失败。
  maxRounds: 15

Only providers.gateway.endpoint is required. Provider IDs default to zmllmgw, zmdevmate, and zmllmadm; each can be overridden in its provider block. The optional secondary and account providers are enabled only when their endpoints are configured.

Each model may set an optional limit block matching the OpenCode model shape, with context (input context window) and output (max output tokens); input is also accepted and passed through when provided.

When autoContinue: true is set on a model, text-only responses that finish with length, or fail because an incomplete stream has no finish_reason, are continued automatically as visible user prompts, regardless of which configured backend route handles the request. The root autoContinue block is optional; its default prompt and maxRounds of 16 are used when omitted. The feature does not alter the model's output-token limit, does not recover truncated tool calls, and does not pin retries to a backend route.

Add opencode-zmllmgw to the OpenCode plugin list, then use /reload-zmllmgw after changing the configuration.