feat: Add reasoning effort control for Claude 3.7 (#2540)

* feat: Add reasoning effort control for Anthropic models with Anthropic Provider and OpenAI Provider

- Add reasoning effort settings with low/medium/high options
- Implement reasoning effort for Claude 3.7 Sonnet models
- Update localization tips for reasoning effort
- Enhance provider handling of reasoning effort parameters

* fix: Extract o1-mini and o1-preview

* fix: Add OpenAI o-series model to ReasoningModel

* fix: Improve OpenAI o-series model detection

* style: Reduce font size

* fix: Add default token handling using DEFAULT_MAX_TOKENS

* fix: Add beta parameter for Anthropic reasoning models

package.json

@@ -81,7 +81,7 @@
     "webdav": "4.11.4"
   },
   "devDependencies": {
-    "@anthropic-ai/sdk": "^0.24.3",
+    "@anthropic-ai/sdk": "^0.38.0",
     "@electron-toolkit/eslint-config-prettier": "^2.0.0",
     "@electron-toolkit/eslint-config-ts": "^1.0.1",
     "@electron-toolkit/tsconfig": "^1.0.1",

src/renderer/src/config/models.ts

@@ -1839,6 +1839,10 @@ export function isVisionModel(model: Model): boolean {
   return VISION_REGEX.test(model.id) || model.type?.includes('vision') || false
 }
 
+export function isOpenAIoSeries(model: Model): boolean {
+  return ['o1', 'o1-2024-12-17'].includes(model.id) || model.id.includes('o3')
+}
+
 export function isReasoningModel(model: Model): boolean {
   if (!model) {
     return false
@@ -1848,6 +1852,10 @@ export function isReasoningModel(model: Model): boolean {
     return REASONING_REGEX.test(model.name) || model.type?.includes('reasoning') || false
   }
 
+  if (model.id.includes('claude-3-7-sonnet') || model.id.includes('claude-3.7-sonnet') || isOpenAIoSeries(model)) {
+    return true
+  }
+
   return REASONING_REGEX.test(model.id) || model.type?.includes('reasoning') || false
 }
 

src/renderer/src/i18n/locales/en-us.json

@@ -52,7 +52,7 @@
       "settings.reasoning_effort.low": "low",
       "settings.reasoning_effort.medium": "medium",
       "settings.reasoning_effort.off": "off",
-      "settings.reasoning_effort.tip": "Only supports reasoning models",
+      "settings.reasoning_effort.tip": "Only supports OpenAI o-series and Anthropic reasoning models",
       "title": "Assistants"
     },
     "auth": {

src/renderer/src/i18n/locales/ja-jp.json

@@ -52,7 +52,7 @@
       "settings.reasoning_effort.low": "短い",
       "settings.reasoning_effort.medium": "中程度",
       "settings.reasoning_effort.off": "オフ",
-      "settings.reasoning_effort.tip": "この設定は推論モデルのみサポートしています",
+      "settings.reasoning_effort.tip": "OpenAIのoシリーズとAnthropicの推論モデルのみサポートしています",
       "title": "アシスタント"
     },
     "auth": {

src/renderer/src/i18n/locales/ru-ru.json

@@ -52,7 +52,7 @@
       "settings.reasoning_effort.low": "Короткая",
       "settings.reasoning_effort.medium": "Средняя",
       "settings.reasoning_effort.off": "Выключено",
-      "settings.reasoning_effort.tip": "Эта настройка поддерживается только моделями с рассуждением",
+      "settings.reasoning_effort.tip": "Поддерживается только моделями с рассуждением OpenAI o-series и Anthropic",
       "title": "Ассистенты"
     },
     "auth": {

src/renderer/src/i18n/locales/zh-cn.json

@@ -52,7 +52,7 @@
       "settings.reasoning_effort.low": "短",
       "settings.reasoning_effort.medium": "中",
       "settings.reasoning_effort.off": "关",
-      "settings.reasoning_effort.tip": "该设置仅支持推理模型",
+      "settings.reasoning_effort.tip": "仅支持 OpenAI o-series 和 Anthropic 推理模型",
       "title": "助手"
     },
     "auth": {

src/renderer/src/i18n/locales/zh-tw.json

@@ -52,7 +52,7 @@
       "settings.reasoning_effort.low": "短",
       "settings.reasoning_effort.medium": "中",
       "settings.reasoning_effort.off": "關",
-      "settings.reasoning_effort.tip": "該設置僅支持推理模型",
+      "settings.reasoning_effort.tip": "僅支持OpenAI o系列和Anthropic推理模型",
       "title": "助手"
     },
     "auth": {

src/renderer/src/pages/home/Tabs/SettingsTab.tsx

@@ -32,7 +32,7 @@ import {
 } from '@renderer/store/settings'
 import { Assistant, AssistantSettings, ThemeMode, TranslateLanguageVarious } from '@renderer/types'
 import { modalConfirm } from '@renderer/utils'
-import { Col, InputNumber, Row, Select, Slider, Switch, Tooltip } from 'antd'
+import { Col, InputNumber, Row, Segmented, Select, Slider, Switch, Tooltip } from 'antd'
 import { FC, useEffect, useState } from 'react'
 import { useTranslation } from 'react-i18next'
 import styled from 'styled-components'
@@ -51,6 +51,7 @@ const SettingsTab: FC<Props> = (props) => {
   const [maxTokens, setMaxTokens] = useState(assistant?.settings?.maxTokens ?? 0)
   const [fontSizeValue, setFontSizeValue] = useState(fontSize)
   const [streamOutput, setStreamOutput] = useState(assistant?.settings?.streamOutput ?? true)
+  const [reasoningEffort, setReasoningEffort] = useState(assistant?.settings?.reasoning_effort)
   const { t } = useTranslation()
 
   const dispatch = useAppDispatch()
@@ -96,9 +97,14 @@ const SettingsTab: FC<Props> = (props) => {
     }
   }
 
+  const onReasoningEffortChange = (value) => {
+    updateAssistantSettings({ reasoning_effort: value })
+  }
+
   const onReset = () => {
     setTemperature(DEFAULT_TEMPERATURE)
     setContextCount(DEFAULT_CONTEXTCOUNT)
+    setReasoningEffort(undefined)
     updateAssistant({
       ...assistant,
       settings: {
@@ -109,6 +115,7 @@ const SettingsTab: FC<Props> = (props) => {
         maxTokens: DEFAULT_MAX_TOKENS,
         streamOutput: true,
         hideMessages: false,
+        reasoning_effort: undefined,
         customParameters: []
       }
     })
@@ -120,6 +127,7 @@ const SettingsTab: FC<Props> = (props) => {
     setEnableMaxTokens(assistant?.settings?.enableMaxTokens ?? false)
     setMaxTokens(assistant?.settings?.maxTokens ?? DEFAULT_MAX_TOKENS)
     setStreamOutput(assistant?.settings?.streamOutput ?? true)
+    setReasoningEffort(assistant?.settings?.reasoning_effort)
   }, [assistant])
 
   return (
@@ -223,6 +231,45 @@ const SettingsTab: FC<Props> = (props) => {
             </Col>
           </Row>
         )}
+        <SettingDivider />
+        <Row align="middle">
+          <Label>{t('assistants.settings.reasoning_effort')}</Label>
+          <Tooltip title={t('assistants.settings.reasoning_effort.tip')}>
+            <QuestionIcon />
+          </Tooltip>
+        </Row>
+        <Row align="middle" gutter={10}>
+          <Col span={24}>
+            <SegmentedContainer>
+              <Segmented<'low' | 'medium' | 'high' | undefined>
+                value={reasoningEffort}
+                onChange={(value) => {
+                  setReasoningEffort(value)
+                  onReasoningEffortChange(value)
+                }}
+                options={[
+                  {
+                    value: 'low',
+                    label: t('assistants.settings.reasoning_effort.low')
+                  },
+                  {
+                    value: 'medium',
+                    label: t('assistants.settings.reasoning_effort.medium')
+                  },
+                  {
+                    value: 'high',
+                    label: t('assistants.settings.reasoning_effort.high')
+                  },
+                  {
+                    value: undefined,
+                    label: t('assistants.settings.reasoning_effort.off')
+                  }
+                ]}
+                block
+              />
+            </SegmentedContainer>
+          </Col>
+        </Row>
       </SettingGroup>
       <SettingGroup>
         <SettingSubtitle style={{ marginTop: 0 }}>{t('settings.messages.title')}</SettingSubtitle>
@@ -485,4 +532,24 @@ export const SettingGroup = styled.div<{ theme?: ThemeMode }>`
   margin-bottom: 10px;
 `
 
+// Define the styled component with hover state styling
+const SegmentedContainer = styled.div`
+  .ant-segmented-item {
+    font-size: 12px;
+  }
+  .ant-segmented-item-selected {
+    background-color: var(--color-primary) !important;
+    color: white !important;
+  }
+
+  .ant-segmented-item:hover:not(.ant-segmented-item-selected) {
+    background-color: var(--color-primary-bg) !important;
+    color: var(--color-primary) !important;
+  }
+
+  .ant-segmented-thumb {
+    background-color: var(--color-primary) !important;
+  }
+`
+
 export default SettingsTab

src/renderer/src/pages/settings/AssistantSettings/AssistantModelSettings.tsx

@@ -154,6 +154,7 @@ const AssistantModelSettings: FC<Props> = ({ assistant, updateAssistant, updateA
     setMaxTokens(0)
     setStreamOutput(true)
     setTopP(1)
+    setReasoningEffort(undefined)
     setCustomParameters([])
     updateAssistantSettings({
       temperature: DEFAULT_TEMPERATURE,
@@ -162,6 +163,7 @@ const AssistantModelSettings: FC<Props> = ({ assistant, updateAssistant, updateA
       maxTokens: 0,
       streamOutput: true,
       topP: 1,
+      reasoning_effort: undefined,
       customParameters: []
     })
   }

src/renderer/src/providers/AnthropicProvider.ts

@@ -1,6 +1,7 @@
 import Anthropic from '@anthropic-ai/sdk'
 import { MessageCreateParamsNonStreaming, MessageParam } from '@anthropic-ai/sdk/resources'
 import { DEFAULT_MAX_TOKENS } from '@renderer/config/constant'
+import { isReasoningModel } from '@renderer/config/models'
 import { getStoreSetting } from '@renderer/hooks/useSettings'
 import i18n from '@renderer/i18n'
 import { getAssistantSettings, getDefaultModel, getTopNamingModel } from '@renderer/services/AssistantService'
@@ -18,7 +19,11 @@ export default class AnthropicProvider extends BaseProvider {
 
   constructor(provider: Provider) {
     super(provider)
-    this.sdk = new Anthropic({ apiKey: this.apiKey, baseURL: this.getBaseURL() })
+    this.sdk = new Anthropic({
+      apiKey: this.apiKey,
+      baseURL: this.getBaseURL(),
+      dangerouslyAllowBrowser: true
+    })
   }
 
   public getBaseURL(): string {
@@ -60,6 +65,47 @@ export default class AnthropicProvider extends BaseProvider {
     }
   }
 
+  private getTemperature(assistant: Assistant, model: Model) {
+    if (isReasoningModel(model)) return undefined
+
+    return assistant?.settings?.temperature
+  }
+
+  private getTopP(assistant: Assistant, model: Model) {
+    if (isReasoningModel(model)) return undefined
+
+    return assistant?.settings?.topP
+  }
+
+  private getReasoningEffort(assistant: Assistant, model: Model) {
+    if (isReasoningModel(model)) {
+      const effort_ratio =
+        assistant?.settings?.reasoning_effort === 'high'
+          ? 0.8
+          : assistant?.settings?.reasoning_effort === 'medium'
+            ? 0.5
+            : assistant?.settings?.reasoning_effort === 'low'
+              ? 0.2
+              : undefined
+      if (!effort_ratio)
+        return {
+          type: 'disabled'
+        }
+
+      if (model.id.includes('claude-3.7-sonnet') || model.id.includes('claude-3-7-sonnet')) {
+        return {
+          type: 'enabled',
+          budget_tokens: Math.max(
+            Math.min((assistant?.settings?.maxTokens || DEFAULT_MAX_TOKENS) * effort_ratio, 32000),
+            1024
+          )
+        }
+      }
+    }
+
+    return undefined
+  }
+
   public async completions({ messages, assistant, onChunk, onFilterMessages }: CompletionsParams) {
     const defaultModel = getDefaultModel()
     const model = assistant.model || defaultModel
@@ -84,20 +130,43 @@ export default class AnthropicProvider extends BaseProvider {
       model: model.id,
       messages: userMessages,
       max_tokens: maxTokens || DEFAULT_MAX_TOKENS,
-      temperature: assistant?.settings?.temperature,
+      temperature: this.getTemperature(assistant, model),
-      top_p: assistant?.settings?.topP,
+      top_p: this.getTopP(assistant, model),
       system: assistant.prompt,
       ...this.getCustomParameters(assistant)
     }
 
+    if (isReasoningModel(model)) {
+      ;(body as any).thinking = this.getReasoningEffort(assistant, model)
+      ;(body as any).betas = ['output-128k-2025-02-19']
+    }
+
     let time_first_token_millsec = 0
+    let time_first_content_millsec = 0
     const start_time_millsec = new Date().getTime()
 
     if (!streamOutput) {
       const message = await this.sdk.messages.create({ ...body, stream: false })
       const time_completion_millsec = new Date().getTime() - start_time_millsec
+
+      let text = ''
+      let reasoning_content = ''
+
+      if (message.content && message.content.length > 0) {
+        const thinkingBlock = message.content.find((block) => block.type === 'thinking')
+        const textBlock = message.content.find((block) => block.type === 'text')
+
+        if (thinkingBlock && 'thinking' in thinkingBlock) {
+          reasoning_content = thinkingBlock.thinking
+        }
+
+        if (textBlock && 'text' in textBlock) {
+          text = textBlock.text
+        }
+      }
       return onChunk({
-        text: message.content[0].type === 'text' ? message.content[0].text : '',
+        text,
+        reasoning_content,
         usage: message.usage,
         metrics: {
           completion_tokens: message.usage.output_tokens,
@@ -113,6 +182,7 @@ export default class AnthropicProvider extends BaseProvider {
     const { signal } = abortController
 
     return new Promise<void>((resolve, reject) => {
+      let hasThinkingContent = false
       const stream = this.sdk.messages
         .stream({ ...body, stream: true }, { signal })
         .on('text', (text) => {
@@ -123,9 +193,34 @@ export default class AnthropicProvider extends BaseProvider {
           if (time_first_token_millsec == 0) {
             time_first_token_millsec = new Date().getTime() - start_time_millsec
           }
+
+          if (hasThinkingContent && time_first_content_millsec === 0) {
+            time_first_content_millsec = new Date().getTime()
+          }
+
+          const time_thinking_millsec = time_first_content_millsec ? time_first_content_millsec - start_time_millsec : 0
+
           const time_completion_millsec = new Date().getTime() - start_time_millsec
           onChunk({
             text,
+            metrics: {
+              completion_tokens: undefined,
+              time_completion_millsec,
+              time_first_token_millsec,
+              time_thinking_millsec
+            }
+          })
+        })
+        .on('thinking', (thinking) => {
+          hasThinkingContent = true
+          if (time_first_token_millsec == 0) {
+            time_first_token_millsec = new Date().getTime() - start_time_millsec
+          }
+
+          const time_completion_millsec = new Date().getTime() - start_time_millsec
+          onChunk({
+            reasoning_content: thinking,
+            text: '',
             metrics: {
               completion_tokens: undefined,
               time_completion_millsec,
@@ -134,6 +229,8 @@ export default class AnthropicProvider extends BaseProvider {
           })
         })
         .on('finalMessage', (message) => {
+          const time_completion_millsec = new Date().getTime() - start_time_millsec
+          const time_thinking_millsec = time_first_content_millsec ? time_first_content_millsec - start_time_millsec : 0
           onChunk({
             text: '',
             usage: {
@@ -143,8 +240,9 @@ export default class AnthropicProvider extends BaseProvider {
             },
             metrics: {
               completion_tokens: message.usage.output_tokens,
-              time_completion_millsec: new Date().getTime() - start_time_millsec,
+              time_completion_millsec,
-              time_first_token_millsec
+              time_first_token_millsec,
+              time_thinking_millsec
             }
           })
           resolve()

src/renderer/src/providers/OpenAIProvider.ts

@@ -1,4 +1,11 @@
-import { getOpenAIWebSearchParams, isReasoningModel, isSupportedModel, isVisionModel } from '@renderer/config/models'
+import { DEFAULT_MAX_TOKENS } from '@renderer/config/constant'
+import {
+  getOpenAIWebSearchParams,
+  isOpenAIoSeries,
+  isReasoningModel,
+  isSupportedModel,
+  isVisionModel
+} from '@renderer/config/models'
 import { getStoreSetting } from '@renderer/hooks/useSettings'
 import i18n from '@renderer/i18n'
 import { getAssistantSettings, getDefaultModel, getTopNamingModel } from '@renderer/services/AssistantService'
@@ -156,9 +163,46 @@ export default class OpenAIProvider extends BaseProvider {
     }
 
     if (isReasoningModel(model)) {
-      return {
+      if (model.provider === 'openrouter') {
-        reasoning_effort: assistant?.settings?.reasoning_effort
+        return {
+          reasoning: {
+            effort: assistant?.settings?.reasoning_effort
+          }
+        }
       }
+
+      if (isOpenAIoSeries(model)) {
+        return {
+          reasoning_effort: assistant?.settings?.reasoning_effort
+        }
+      }
+
+      const effort_ratio =
+        assistant?.settings?.reasoning_effort === 'high'
+          ? 0.8
+          : assistant?.settings?.reasoning_effort === 'medium'
+            ? 0.5
+            : assistant?.settings?.reasoning_effort === 'low'
+              ? 0.2
+              : undefined
+
+      if (model.id.includes('claude-3.7-sonnet') || model.id.includes('claude-3-7-sonnet')) {
+        if (!effort_ratio) {
+          return {
+            type: 'disabled'
+          }
+        }
+        return {
+          thinking: {
+            budget_tokens: Math.max(
+              Math.min((assistant?.settings?.maxTokens || DEFAULT_MAX_TOKENS) * effort_ratio, 32000),
+              1024
+            )
+          }
+        }
+      }
+
+      return {}
     }
 
     return {}
@@ -175,7 +219,7 @@ export default class OpenAIProvider extends BaseProvider {
 
     let systemMessage = assistant.prompt ? { role: 'system', content: assistant.prompt } : undefined
 
-    if (['o1', 'o1-2024-12-17'].includes(model.id) || model.id.startsWith('o3')) {
+    if (isOpenAIoSeries(model)) {
       systemMessage = {
         role: 'developer',
         content: `Formatting re-enabled${systemMessage ? '\n' + systemMessage.content : ''}`
@@ -212,6 +256,7 @@ export default class OpenAIProvider extends BaseProvider {
       delta: OpenAI.Chat.Completions.ChatCompletionChunk.Choice.Delta & {
         reasoning_content?: string
         reasoning?: string
+        thinking?: string
       }
     ) => {
       if (!delta?.content) return false
@@ -226,7 +271,7 @@ export default class OpenAIProvider extends BaseProvider {
       }
 
       // 如果有reasoning_content或reasoning，说明是在思考中
-      if (delta?.reasoning_content || delta?.reasoning) {
+      if (delta?.reasoning_content || delta?.reasoning || delta?.thinking) {
         hasReasoningContent = true
       }
 

yarn.lock

@@ -110,9 +110,9 @@ __metadata:
   languageName: node
   linkType: hard
 
-"@anthropic-ai/sdk@npm:^0.24.3":
+"@anthropic-ai/sdk@npm:^0.38.0":
-  version: 0.24.3
+  version: 0.38.0
-  resolution: "@anthropic-ai/sdk@npm:0.24.3"
+  resolution: "@anthropic-ai/sdk@npm:0.38.0"
   dependencies:
     "@types/node": "npm:^18.11.18"
     "@types/node-fetch": "npm:^2.6.4"
@@ -121,8 +121,7 @@ __metadata:
     form-data-encoder: "npm:1.7.2"
     formdata-node: "npm:^4.3.2"
     node-fetch: "npm:^2.6.7"
-    web-streams-polyfill: "npm:^3.2.1"
+  checksum: 10c0/accd003cbe314d32d4d36f5fd7fd743c32e2a896c9ea57190966eda20b8c46e00f542bf03ec3603d1274a7ac18e902bed4158ff5980e4e248a6d5c75e3fd891a
-  checksum: 10c0/1c73c3df9637522da548d2cddfaf89513dac935c5cdb7c0b3db1c427c069a0de76df935bd189e477822063e9f944360e2d059827d5be4dca33bd388c61e97a30
   languageName: node
   linkType: hard
 
@@ -2996,7 +2995,7 @@ __metadata:
   version: 0.0.0-use.local
   resolution: "CherryStudio@workspace:."
   dependencies:
-    "@anthropic-ai/sdk": "npm:^0.24.3"
+    "@anthropic-ai/sdk": "npm:^0.38.0"
     "@electron-toolkit/eslint-config-prettier": "npm:^2.0.0"
     "@electron-toolkit/eslint-config-ts": "npm:^1.0.1"
     "@electron-toolkit/preload": "npm:^3.0.0"
@@ -14363,7 +14362,7 @@ __metadata:
   languageName: node
   linkType: hard
 
-"web-streams-polyfill@npm:^3.0.3, web-streams-polyfill@npm:^3.2.1":
+"web-streams-polyfill@npm:^3.0.3":
   version: 3.3.3
   resolution: "web-streams-polyfill@npm:3.3.3"
   checksum: 10c0/64e855c47f6c8330b5436147db1c75cb7e7474d924166800e8e2aab5eb6c76aac4981a84261dd2982b3e754490900b99791c80ae1407a9fa0dcff74f82ea3a7f