fix deepseek v4 reasoning content problem in thinking mode

[wya-tt]

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

This fixes DeepSeek reasoning message handling when assistant messages include tool calls. The changes is made by Cursor with GPT 5.5, based on the official update of OpenCode.

DeepSeek-compatible models can require reasoning_content to be sent back in provider options on later requests, including when the reasoning content is empty. Previously the transform could drop that field, which breaks follow-up tool-call flows.

The change preserves existing reasoning_content, keeps empty reasoning content instead of omitting it, and defaults reasoning-capable models to use reasoning_content when no explicit interleaved field is configured.

How did you verify your code works?

Added regression tests for DeepSeek provider transforms covering:

• assistant messages with reasoning and tool calls
• empty reasoning_content
• repeated transform calls preserving existing reasoning content
• reasoning models without explicit interleaved config

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[panyw5]

谢谢提交。Bug 抓得对，patch 大方向也对，但作用域太宽。已在 dev (commit f5d2963) 落了一个收窄版。

采用了原 PR 的

• transform.ts 把硬编码的 "openaiCompatible" 换成 sdkKey(model.api.npm) ?? "openaiCompatible"
• transform.ts 总是写 [field]（即使为空字符串）+ 跨 transform 用 || 链保留已有 reasoning_content——核心 bug fix
• 测试 fixture 重构（共享 deepseekModel）+ 空 reasoning / 重复 transform 两个回归测试

没有采用的

• provider.ts 把 interleaved 兜底到所有 model.reasoning === true 的模型——会把 DeepSeek 专属字段套到 OpenAI o1/o3、Claude reasoning、Gemini thinking 上，污染它们的 providerOptions
• transform.ts 新增的 if (model.capabilities.reasoning) { … } 分支——同上，且会硬编码 reasoning_content 字段名，无视 interleaved.field 配置

改用 upstream 的设计

provider.ts 的 interleaved 默认值改为 anomalyco#24630（commit 738b3065d）的写法，三重门控：

interleaved:
  model.interleaved ??
  existingModel?.capabilities.interleaved ??
  (!existingModel && apiNpm === "@ai-sdk/openai-compatible" && apiID.toLowerCase().includes("deepseek")
    ? { field: "reasoning_content" }
    : false),

只对全新 openai-compatible + DeepSeek 命名的模型默认开启，user/source config 仍然优先。fromModelsDevModel 那处不动——models.dev 已经显式给 DeepSeek 配了 interleaved。

额外加的

OpenAI o1-mini（reasoning=true, interleaved=false）transform 后不应有 providerOptions.openai.reasoning_content 的反向 guard，防止之后再有人想往这个方向扩。

[panyw5]

Close pull request. Thanks.