Pass reasoning_content for deepseek models

[vinaysb]

Hello, I have been trying to make queries via your sdk to the deepseek-v4-pro and flash models but the official provider seems to be requiring the presence of "reasoning_content" after every tool call or the api call is rejected with a provider returned error message.

the current format seems to be,

{ role: "user", content: "What time is it in London? Call the get_time tool." },
{
  role: "assistant",
  content: msg1.content,
  tool_calls: msg1.tool_calls,
  reasoning: msg1.reasoning,
  reasoning_details: msg1.reasoning_details,
},
{
  role: "tool",
  tool_call_id: toolCall.id,
  content: "The current time in London is 21:45 UTC.",
}

but it needs to be as follows for deepseek-v4

{ role: "user", content: "What time is it in London? Call the get_time tool." },
{
  role: "assistant",
  content: msg1.content,
  tool_calls: msg1.tool_calls,
  reasoning_content: msg1.reasoning ?? "",
},
{
  role: "tool",
  tool_call_id: toolCall.id,
  content: "The current time in London is 21:45 UTC.",
}

[lianzhiweivincent-cloud]

This looks like the same class of bug that shows up in several OpenAI-compatible DeepSeek/Xiaomi-style routes: the provider exposes reasoning as transport metadata, but the next request requires it to be replayed as provider-specific message state.

For DeepSeek v4/reasoner, I would avoid mapping this only to the AI SDK's generic reasoning / reasoning_details fields. The safer adapter behavior is probably:

• preserve the raw provider reasoning text on the assistant message when tool_calls are present
• replay it as reasoning_content on the next request for DeepSeek-compatible routes
• keep normal OpenAI routes clean, since they should not receive that field
• add a regression with: user -> assistant(tool_calls + reasoning_content) -> tool -> next user request

That test catches the exact failure mode where the first turn succeeds, but the next turn gets rejected because the assistant/tool-call message lost its thinking state.

Disclosure: I am working on Black Eagle AI, a small China-provider relay for overseas developers. We have been tracking this reasoning_content replay issue across DeepSeek/MiMo-compatible routes, so happy to share route notes or compare behavior against a small external endpoint if useful.