feat(session): expose maxOutputTokens on llm adapters

packages/session/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@stello-ai/session",
-  "version": "0.7.1",
+  "version": "0.7.2",
   "description": "Session layer for Stello — conversation topology engine",
   "license": "Apache-2.0",
   "author": "Stello Contributors",

packages/session/src/__tests__/anthropic.test.ts

@@ -162,3 +162,80 @@ describe('createAnthropicAdapter stream()', () => {
     expect(chunks.map((c) => c.delta).join('')).toBe('hello')
   })
 })
+
+describe('createAnthropicAdapter complete() max_tokens', () => {
+  beforeEach(() => {
+    messagesStream.mockReset()
+    messagesCreate.mockReset()
+    messagesCreate.mockResolvedValue({
+      content: [{ type: 'text', text: 'ok' }],
+      usage: { input_tokens: 1, output_tokens: 1 },
+    })
+  })
+
+  it('未配置时回落到内建默认值 4096', async () => {
+    const adapter = createAnthropicAdapter({
+      apiKey: 'k',
+      model: 'm',
+      maxContextTokens: 200_000,
+    })
+    await adapter.complete([{ role: 'user', content: 'hi' }])
+    expect(messagesCreate).toHaveBeenCalledWith(
+      expect.objectContaining({ max_tokens: 4096 }),
+      undefined,
+    )
+  })
+
+  it('options.maxOutputTokens 覆盖内建默认值', async () => {
+    const adapter = createAnthropicAdapter({
+      apiKey: 'k',
+      model: 'm',
+      maxContextTokens: 200_000,
+      maxOutputTokens: 8192,
+    })
+    await adapter.complete([{ role: 'user', content: 'hi' }])
+    expect(messagesCreate).toHaveBeenCalledWith(
+      expect.objectContaining({ max_tokens: 8192 }),
+      undefined,
+    )
+  })
+
+  it('调用方 maxTokens 优先级最高，盖过 options.maxOutputTokens', async () => {
+    const adapter = createAnthropicAdapter({
+      apiKey: 'k',
+      model: 'm',
+      maxContextTokens: 200_000,
+      maxOutputTokens: 8192,
+    })
+    await adapter.complete([{ role: 'user', content: 'hi' }], { maxTokens: 2048 })
+    expect(messagesCreate).toHaveBeenCalledWith(
+      expect.objectContaining({ max_tokens: 2048 }),
+      undefined,
+    )
+  })
+})
+
+describe('createAnthropicAdapter stream() max_tokens', () => {
+  beforeEach(() => {
+    messagesStream.mockReset()
+    messagesCreate.mockReset()
+    messagesStream.mockReturnValue(asyncIterableFrom([]))
+  })
+
+  it('options.maxOutputTokens 用于 stream() 请求', async () => {
+    const adapter = createAnthropicAdapter({
+      apiKey: 'k',
+      model: 'm',
+      maxContextTokens: 200_000,
+      maxOutputTokens: 8192,
+    })
+    if (!adapter.stream) throw new Error('adapter.stream is required')
+    for await (const _ of adapter.stream([{ role: 'user', content: 'hi' }])) {
+      void _
+    }
+    expect(messagesStream).toHaveBeenCalledWith(
+      expect.objectContaining({ max_tokens: 8192 }),
+      undefined,
+    )
+  })
+})

packages/session/src/__tests__/openai-compatible.test.ts

@@ -72,6 +72,40 @@ describe('createOpenAICompatibleAdapter', () => {
     )
   })
 
+  it('options.maxOutputTokens 覆盖内建默认值', async () => {
+    const adapter = createOpenAICompatibleAdapter({
+      apiKey: 'test-key',
+      baseURL: 'https://api.example.com/v1',
+      model: 'test-model',
+      maxContextTokens: 128_000,
+      maxOutputTokens: 8192,
+    })
+
+    await adapter.complete([{ role: 'user', content: 'hello' }])
+
+    expect(createCompletion).toHaveBeenCalledWith(
+      expect.objectContaining({ max_tokens: 8192, stream: false }),
+      undefined,
+    )
+  })
+
+  it('调用方 maxTokens 优先级最高，盖过 options.maxOutputTokens', async () => {
+    const adapter = createOpenAICompatibleAdapter({
+      apiKey: 'test-key',
+      baseURL: 'https://api.example.com/v1',
+      model: 'test-model',
+      maxContextTokens: 128_000,
+      maxOutputTokens: 8192,
+    })
+
+    await adapter.complete([{ role: 'user', content: 'hello' }], { maxTokens: 2048 })
+
+    expect(createCompletion).toHaveBeenCalledWith(
+      expect.objectContaining({ max_tokens: 2048, stream: false }),
+      undefined,
+    )
+  })
+
   it('signal 透传到 SDK request options', async () => {
     const adapter = createOpenAICompatibleAdapter({
       apiKey: 'test-key',

packages/session/src/adapters/anthropic.ts

@@ -13,10 +13,17 @@ import type { LLMAdapter, LLMResult, LLMChunk, Message, ToolCall, LLMCompleteOpt
 export interface AnthropicAdapterOptions {
   apiKey: string
   model: string
-  /** 模型上下文窗口大小（token 数） */
+  /** 模型上下文窗口大小（token 数），用于自动压缩判断 */
   maxContextTokens: number
   /** 自定义 API 端点，兼容 MiniMax 等 Anthropic 协议服务 */
   baseURL?: string
+  /**
+   * 单次请求的输出 token 上限。被写入 Anthropic API 的 `max_tokens`。
+   * 优先级：`completeOptions.maxTokens` > `options.maxOutputTokens` > 4096。
+   * 设过低会让长输出（多个子话题的 tool call args、长 synthesis 等）
+   * 在中途被截断，引发上层 JSON 解析失败。建议按模型上限设置。
+   */
+  maxOutputTokens?: number
 }
 
 /** 将 Stello 内部 Message 转换为 Anthropic MessageParam 格式 */
@@ -138,7 +145,7 @@ export function createAnthropicAdapter(options: AnthropicAdapterOptions): LLMAda
       const response = await client.messages.create(
         {
           model: options.model,
-          max_tokens: completeOptions?.maxTokens ?? 4096,
+          max_tokens: completeOptions?.maxTokens ?? options.maxOutputTokens ?? 4096,
           ...(completeOptions?.temperature !== undefined && { temperature: completeOptions.temperature }),
           ...(system && { system }),
           ...(completeOptions?.tools && completeOptions.tools.length > 0
@@ -172,7 +179,7 @@ export function createAnthropicAdapter(options: AnthropicAdapterOptions): LLMAda
       const stream = client.messages.stream(
         {
           model: options.model,
-          max_tokens: completeOptions?.maxTokens ?? 4096,
+          max_tokens: completeOptions?.maxTokens ?? options.maxOutputTokens ?? 4096,
           ...(completeOptions?.temperature !== undefined && { temperature: completeOptions.temperature }),
           ...(system && { system }),
           ...(completeOptions?.tools && completeOptions.tools.length > 0

packages/session/src/adapters/openai-compatible.ts

@@ -11,11 +11,18 @@ type ChatToolCallDelta = NonNullable<
 export interface OpenAICompatibleOptions {
   apiKey: string
   model: string
-  /** 模型上下文窗口大小（token 数） */
+  /** 模型上下文窗口大小（token 数），用于自动压缩判断 */
   maxContextTokens: number
   baseURL: string
   /** 额外的请求参数（如 MiniMax 的 reasoning_split 等） */
   extraBody?: Record<string, unknown>
+  /**
+   * 单次请求的输出 token 上限。被写入请求的 `max_tokens`。
+   * 优先级：`completeOptions.maxTokens` > `options.maxOutputTokens` > 4096。
+   * 设过低会让长输出（多个子话题的 tool call args、长 synthesis 等）
+   * 在中途被截断，引发上层 JSON 解析失败。
+   */
+  maxOutputTokens?: number
 }
 
 /** 合并连续的 system 消息，兼容只接受单条 system 的提供方。 */
@@ -46,7 +53,7 @@ export function createOpenAICompatibleAdapter(options: OpenAICompatibleOptions):
     const normalizedMessages = mergeConsecutiveSystemMessages(messages)
     return {
       model: options.model,
-      max_tokens: completeOptions?.maxTokens ?? 4096,
+      max_tokens: completeOptions?.maxTokens ?? options.maxOutputTokens ?? 4096,
       ...(completeOptions?.temperature !== undefined && { temperature: completeOptions.temperature }),
       ...(completeOptions?.tools
         ? {