OpenAI provider forces Responses API, breaking tool calls with Chat Completions-only proxies #15016

Description

When using the openai provider with a custom baseURL pointing to a middleware/proxy that only supports the Chat Completions API (/v1/chat/completions), tool calls are rendered as raw JSON text output instead of being executed as tools.

Root cause: provider.ts hardcodes sdk.responses(modelID) for the openai provider with no fallback:

// provider.ts ~L153
openai: async () => {
  return {
    autoload: false,
    async getModel(sdk: any, modelID: string) {
      return sdk.responses(modelID)  // hardcoded — no fallback to sdk.chat()
    },
  }
},

When the proxy returns a Chat Completions-format response, the Responses API stream parser cannot match the event format. tool_calls fields are ignored and content falls through to text-delta, producing raw JSON in the chat output.

Note: azure and azure-cognitive-services providers already handle this correctly via options.useCompletionUrls — the same pattern could be applied here.

Steps to Reproduce

1. Configure the openai provider with a baseURL pointing to a Chat Completions-only proxy
2. Start a session with any GPT model
3. Tool calls appear as raw JSON text instead of being invoked

Suggested Fix

Mirror the existing azure provider pattern:

openai: async () => {
  return {
    autoload: false,
    async getModel(sdk: any, modelID: string, options?: Record<string, any>) {
      if (options?.['useCompletionUrls']) {
        return sdk.chat(modelID)
      }
      return sdk.responses(modelID)
    },
    options: ,
  }
},

Users opt in via opencode.json:

{
  "provider": {
    "openai": {
      "options": {
        "baseURL": "https://your-proxy/v1",
        "useCompletionUrls": true
      }
    }
  }
}

OpenCode Version

1.2.6

描述

使用 openai provider 并将 baseURL 指向仅支持 Chat Completions API 的中间层/代理时，工具调用会以原始 JSON 文本的形式输出，而不是被正常执行。

根本原因：provider.ts 中 openai provider 硬编码使用 sdk.responses(modelID)，没有任何 fallback 到 sdk.chat(modelID)。

当代理返回 Chat Completions 格式的响应时，Responses API 的流式解析器无法匹配事件格式，导致 tool_calls 字段被忽略，内容落入 text-delta 路径，最终在聊天界面输出原始 JSON。

注：azure 和 azure-cognitive-services provider 已通过 options.useCompletionUrls 标志正确处理了这一情况，同样的模式可以应用到 openai provider。

复现步骤

1. 使用 openai provider 配置自定义 baseURL，指向仅支持 Chat Completions 的代理
2. 用任意 GPT 模型开启会话
3. 工具调用以原始 JSON 文本形式出现在输出中，而非被正常调用

建议修复

为 openai provider 添加 useCompletionUrls 支持，参考现有的 azure provider 实现（见上方英文部分）。