fix: implement Anthropic thinking budget calculation in reasoning logic (#9991)

src/renderer/src/aiCore/prepareParams/parameterBuilder.ts

@@ -9,17 +9,19 @@ import {
   isOpenRouterBuiltInWebSearchModel,
   isReasoningModel,
   isSupportedReasoningEffortModel,
+  isSupportedThinkingTokenClaudeModel,
   isSupportedThinkingTokenModel,
   isWebSearchModel
 } from '@renderer/config/models'
 import { getAssistantSettings, getDefaultModel } from '@renderer/services/AssistantService'
-import type { Assistant, MCPTool, Provider } from '@renderer/types'
+import { type Assistant, type MCPTool, type Provider } from '@renderer/types'
 import type { StreamTextParams } from '@renderer/types/aiCoreTypes'
 import type { ModelMessage } from 'ai'
 import { stepCountIs } from 'ai'
 
 import { setupToolsConfig } from '../utils/mcp'
 import { buildProviderOptions } from '../utils/options'
+import { getAnthropicThinkingBudget } from '../utils/reasoning'
 import { getTemperature, getTopP } from './modelParameters'
 
 const logger = loggerService.withContext('parameterBuilder')
@@ -55,7 +57,7 @@ export async function buildStreamTextParams(
 
   const model = assistant.model || getDefaultModel()
 
-  const { maxTokens } = getAssistantSettings(assistant)
+  let { maxTokens } = getAssistantSettings(assistant)
 
   // 这三个变量透传出来，交给下面启用插件/中间件
   // 也可以在外部构建好再传入buildStreamTextParams
@@ -88,6 +90,16 @@ export async function buildStreamTextParams(
     enableGenerateImage
   })
 
+  // NOTE: ai-sdk会把maxToken和budgetToken加起来
+  if (
+    enableReasoning &&
+    maxTokens !== undefined &&
+    isSupportedThinkingTokenClaudeModel(model) &&
+    (provider.type === 'anthropic' || provider.type === 'aws-bedrock')
+  ) {
+    maxTokens -= getAnthropicThinkingBudget(assistant, model)
+  }
+
   // 构建基础参数
   const params: StreamTextParams = {
     messages: sdkMessages,

src/renderer/src/aiCore/utils/reasoning.ts

@@ -310,6 +310,26 @@ export function getOpenAIReasoningParams(assistant: Assistant, model: Model): Re
   return {}
 }
 
+export function getAnthropicThinkingBudget(assistant: Assistant, model: Model): number {
+  const { maxTokens, reasoning_effort: reasoningEffort } = getAssistantSettings(assistant)
+  if (maxTokens === undefined || reasoningEffort === undefined) {
+    return 0
+  }
+  const effortRatio = EFFORT_RATIO[reasoningEffort]
+
+  const budgetTokens = Math.max(
+    1024,
+    Math.floor(
+      Math.min(
+        (findTokenLimit(model.id)?.max! - findTokenLimit(model.id)?.min!) * effortRatio +
+          findTokenLimit(model.id)?.min!,
+        (maxTokens || DEFAULT_MAX_TOKENS) * effortRatio
+      )
+    )
+  )
+  return budgetTokens
+}
+
 /**
  * 获取 Anthropic 推理参数
  * 从 AnthropicAPIClient 中提取的逻辑
@@ -331,19 +351,7 @@ export function getAnthropicReasoningParams(assistant: Assistant, model: Model):
 
   // Claude 推理参数
   if (isSupportedThinkingTokenClaudeModel(model)) {
-    const { maxTokens } = getAssistantSettings(assistant)
-    const effortRatio = EFFORT_RATIO[reasoningEffort]
-
-    const budgetTokens = Math.max(
-      1024,
-      Math.floor(
-        Math.min(
-          (findTokenLimit(model.id)?.max! - findTokenLimit(model.id)?.min!) * effortRatio +
-            findTokenLimit(model.id)?.min!,
-          (maxTokens || DEFAULT_MAX_TOKENS) * effortRatio
-        )
-      )
-    )
+    const budgetTokens = getAnthropicThinkingBudget(assistant, model)
 
     return {
       thinking: {