diff --git a/packages/session/src/__tests__/openai-compatible.test.ts b/packages/session/src/__tests__/openai-compatible.test.ts
index dfb29c5..4355a38 100644
--- a/packages/session/src/__tests__/openai-compatible.test.ts
+++ b/packages/session/src/__tests__/openai-compatible.test.ts
@@ -106,6 +106,66 @@ describe('createOpenAICompatibleAdapter', () => {
     )
   })
 
+  it('complete() 提取响应中的 reasoning_content', async () => {
+    createCompletion.mockResolvedValueOnce({
+      choices: [{ message: { content: 'answer', reasoning_content: 'thinking...' } }],
+      usage: { prompt_tokens: 10, completion_tokens: 5 },
+    })
+
+    const adapter = createOpenAICompatibleAdapter({
+      apiKey: 'test-key',
+      baseURL: 'https://api.example.com/v1',
+      model: 'test-model',
+      maxContextTokens: 128_000,
+    })
+
+    const result = await adapter.complete([{ role: 'user', content: 'hello' }])
+    expect(result.reasoningContent).toBe('thinking...')
+    expect(result.content).toBe('answer')
+  })
+
+  it('assistant 消息的 reasoningContent 回传为 reasoning_content', async () => {
+    const adapter = createOpenAICompatibleAdapter({
+      apiKey: 'test-key',
+      baseURL: 'https://api.example.com/v1',
+      model: 'test-model',
+      maxContextTokens: 128_000,
+    })
+
+    const messages: Message[] = [
+      { role: 'user', content: 'hello' },
+      {
+        role: 'assistant',
+        content: 'let me think',
+        reasoningContent: 'step 1: ...',
+        toolCalls: [{ id: 'tc_1', name: 'foo', input: {} }],
+      },
+      { role: 'tool', content: 'result', toolCallId: 'tc_1' },
+    ]
+
+    await adapter.complete(messages)
+
+    const sentMessages = createCompletion.mock.calls[0]![0].messages
+    expect(sentMessages[1]).toMatchObject({
+      role: 'assistant',
+      content: 'let me think',
+      reasoning_content: 'step 1: ...',
+      tool_calls: [{ id: 'tc_1', type: 'function', function: { name: 'foo', arguments: '{}' } }],
+    })
+  })
+
+  it('非推理模型响应不含 reasoningContent', async () => {
+    const adapter = createOpenAICompatibleAdapter({
+      apiKey: 'test-key',
+      baseURL: 'https://api.example.com/v1',
+      model: 'test-model',
+      maxContextTokens: 128_000,
+    })
+
+    const result = await adapter.complete([{ role: 'user', content: 'hello' }])
+    expect(result.reasoningContent).toBeUndefined()
+  })
+
   it('signal 透传到 SDK request options', async () => {
     const adapter = createOpenAICompatibleAdapter({
       apiKey: 'test-key',
diff --git a/packages/session/src/adapters/openai-compatible.ts b/packages/session/src/adapters/openai-compatible.ts
index 42d4f27..22ad1ee 100644
--- a/packages/session/src/adapters/openai-compatible.ts
+++ b/packages/session/src/adapters/openai-compatible.ts
@@ -71,6 +71,9 @@ export function createOpenAICompatibleAdapter(options: OpenAICompatibleOptions):
         role: m.role as 'system' | 'user' | 'assistant' | 'tool',
         content: m.content,
         ...(m.role === 'tool' && m.toolCallId ? { tool_call_id: m.toolCallId } : {}),
+        ...(m.role === 'assistant' && m.reasoningContent
+          ? { reasoning_content: m.reasoningContent }
+          : {}),
         ...(m.role === 'assistant' && m.toolCalls && m.toolCalls.length > 0
           ? {
               tool_calls: m.toolCalls.map((toolCall) => ({
@@ -100,9 +103,15 @@ export function createOpenAICompatibleAdapter(options: OpenAICompatibleOptions):
       ) as ChatCompletion
 
       const choice = response.choices[0]
+      // 提取推理模型的思考内容（stepFun/DeepSeek 等使用 reasoning_content 字段）
+      const rawMessage = choice?.message as Record<string, unknown> | undefined
+      const reasoningContent = typeof rawMessage?.reasoning_content === 'string'
+        ? rawMessage.reasoning_content
+        : null
 
       return {
         content: choice?.message?.content ?? null,
+        ...(reasoningContent ? { reasoningContent } : {}),
         toolCalls: (choice?.message?.tool_calls ?? []).flatMap((call) => {
           if (!('function' in call) || !call.function) return []
           return [{
@@ -131,14 +140,19 @@ export function createOpenAICompatibleAdapter(options: OpenAICompatibleOptions):
 
       for await (const chunk of stream) {
         const delta = chunk.choices[0]?.delta?.content ?? ''
+        // 提取推理模型的思考内容增量（stepFun/DeepSeek 等使用 reasoning_content 字段）
+        const rawDelta = chunk.choices[0]?.delta as Record<string, unknown> | undefined
+        const reasoningDelta = typeof rawDelta?.reasoning_content === 'string'
+          ? rawDelta.reasoning_content
+          : undefined
         const toolCallDeltas = (chunk.choices[0]?.delta?.tool_calls ?? []).map((call: ChatToolCallDelta) => ({
           index: call.index ?? 0,
           id: call.id,
           name: call.function?.name,
           input: call.function?.arguments,
         }))
-        if (delta || toolCallDeltas.length > 0) {
-          yield { delta, toolCallDeltas }
+        if (delta || reasoningDelta || toolCallDeltas.length > 0) {
+          yield { delta, ...(reasoningDelta ? { reasoningDelta } : {}), toolCallDeltas }
         }
       }
     },
diff --git a/packages/session/src/create-session.ts b/packages/session/src/create-session.ts
index 66e8396..4f705d2 100644
--- a/packages/session/src/create-session.ts
+++ b/packages/session/src/create-session.ts
@@ -244,6 +244,7 @@ function buildSession(
       const assistantRecord: Message = {
         role: 'assistant',
         content: result.content ?? '',
+        ...(result.reasoningContent ? { reasoningContent: result.reasoningContent } : {}),
         ...(result.toolCalls && result.toolCalls.length > 0 ? { toolCalls: result.toolCalls } : {}),
         timestamp: new Date().toISOString(),
       }
@@ -254,6 +255,7 @@ function buildSession(
 
       return {
         content: result.content,
+        reasoningContent: result.reasoningContent,
         toolCalls: result.toolCalls,
         usage: result.usage,
       }
@@ -314,11 +316,13 @@ function buildSession(
         let result: SendResult
         if (options.llm.stream) {
           let accumulated = ''
+          let accumulatedReasoning = ''
           const toolCallsByIndex = new Map<number, { id?: string; name?: string; input: string }>()
           // adapter 在 abort 时抛 AbortError，这里直接向上传播给 result promise；
           // 下方 L3 写入分支不会执行（policy: drop entirely），与非流式 send() 对称。
           for await (const chunk of options.llm.stream(promptMessages, { tools, signal: sendOptions?.signal })) {
             accumulated += chunk.delta
+            if (chunk.reasoningDelta) accumulatedReasoning += chunk.reasoningDelta
             push(chunk.delta)
             for (const delta of chunk.toolCallDeltas ?? []) {
               const current = toolCallsByIndex.get(delta.index) ?? { input: '' }
@@ -333,7 +337,11 @@ function buildSession(
             name: call.name ?? 'unknown_tool',
             input: call.input ? JSON.parse(call.input) as Record<string, unknown> : {},
           }))
-          result = { content: accumulated, toolCalls }
+          result = {
+            content: accumulated,
+            ...(accumulatedReasoning ? { reasoningContent: accumulatedReasoning } : {}),
+            toolCalls,
+          }
         } else {
           result = await options.llm.complete(promptMessages, { tools, signal: sendOptions?.signal })
           if (result.content) {
@@ -344,6 +352,7 @@ function buildSession(
         const assistantRecord: Message = {
           role: 'assistant',
           content: result.content ?? '',
+          ...(result.reasoningContent ? { reasoningContent: result.reasoningContent } : {}),
           ...(result.toolCalls && result.toolCalls.length > 0 ? { toolCalls: result.toolCalls } : {}),
           timestamp: new Date().toISOString(),
         }
@@ -359,6 +368,7 @@ function buildSession(
 
         return {
           content: result.content,
+          reasoningContent: result.reasoningContent,
           toolCalls: result.toolCalls,
           usage: result.usage,
         }
diff --git a/packages/session/src/types/functions.ts b/packages/session/src/types/functions.ts
index 3d538c2..f53d94f 100644
--- a/packages/session/src/types/functions.ts
+++ b/packages/session/src/types/functions.ts
@@ -49,6 +49,8 @@ export interface LoadSessionOptions {
 export interface SendResult {
   /** LLM 文本响应 */
   content: string | null
+  /** 推理模型的思考内容，多轮对话时需回传给 API */
+  reasoningContent?: string | null
   /** LLM 返回的工具调用（由上层决定是否执行） */
   toolCalls?: ToolCall[]
   /** token 用量统计 */
diff --git a/packages/session/src/types/llm.ts b/packages/session/src/types/llm.ts
index a2d9f87..0449182 100644
--- a/packages/session/src/types/llm.ts
+++ b/packages/session/src/types/llm.ts
@@ -2,6 +2,8 @@
 export interface Message {
   role: 'system' | 'user' | 'assistant' | 'tool'
   content: string
+  /** 推理模型的思考内容（stepFun/DeepSeek 等），仅 role=assistant 时有效 */
+  reasoningContent?: string
   /** assistant 发起的工具调用列表，仅 role=assistant 时有效 */
   toolCalls?: ToolCall[]
   /** 关联的工具调用 ID，仅 role=tool 时有效 */
@@ -35,6 +37,8 @@ export interface LLMCompleteOptions {
 /** LLM 完成后的返回结果 */
 export interface LLMResult {
   content: string | null
+  /** 推理模型的思考内容，多轮对话时需回传给 API */
+  reasoningContent?: string | null
   toolCalls?: ToolCall[]
   usage?: {
     promptTokens: number
@@ -46,6 +50,8 @@ export interface LLMResult {
 export interface LLMChunk {
   /** 文本增量片段 */
   delta: string
+  /** 推理内容增量片段（stepFun/DeepSeek 等推理模型） */
+  reasoningDelta?: string
   /** 工具调用增量片段（用于流式拼接 tool call） */
   toolCallDeltas?: Array<{
     index: number