Fix custom parameters placement for Vercel AI Gateway

For AI Gateway provider, custom parameters are now placed at the body level
instead of being nested inside providerOptions.gateway. This fixes the issue
where parameters like 'tools' were being incorrectly added to
providerOptions.gateway when they should be at the same level as providerOptions.

Fixes #4197

Co-authored-by: DeJeune <67425183+DeJeune@users.noreply.github.com>

CLAUDE.md

@@ -12,7 +12,15 @@ This file provides guidance to AI coding assistants when working with code in th
 - **Always propose before executing**: Before making any changes, clearly explain your planned approach and wait for explicit user approval to ensure alignment and prevent unwanted modifications.
 - **Lint, test, and format before completion**: Coding tasks are only complete after running `yarn lint`, `yarn test`, and `yarn format` successfully.
 - **Write conventional commits**: Commit small, focused changes using Conventional Commit messages (e.g., `feat:`, `fix:`, `refactor:`, `docs:`).
-- **Follow PR template**: When submitting pull requests, follow the template in `.github/pull_request_template.md` to ensure complete context and documentation.
+
+## Pull Request Workflow (CRITICAL)
+
+When creating a Pull Request, you MUST:
+
+1. **Read the PR template first**: Always read `.github/pull_request_template.md` before creating the PR
+2. **Follow ALL template sections**: Structure the `--body` parameter to include every section from the template
+3. **Never skip sections**: Include all sections even if marking them as N/A or "None"
+4. **Use proper formatting**: Match the template's markdown structure exactly (headings, checkboxes, code blocks)
 
 ## Development Commands
 

electron-builder.yml

@@ -134,9 +134,9 @@ artifactBuildCompleted: scripts/artifact-build-completed.js
 releaseInfo:
   releaseNotes: |
     <!--LANG:en-->
-    A New Era of Intelligence with Cherry Studio 1.7.0
+    A New Era of Intelligence with Cherry Studio 1.7.1
 
-    Today we're releasing Cherry Studio 1.7.0 — our most ambitious update yet, introducing Agent: autonomous AI that thinks, plans, and acts.
+    Today we're releasing Cherry Studio 1.7.1 — our most ambitious update yet, introducing Agent: autonomous AI that thinks, plans, and acts.
 
     For years, AI assistants have been reactive — waiting for your commands, responding to your questions. With Agent, we're changing that. Now, AI can truly work alongside you: understanding complex goals, breaking them into steps, and executing them independently.
 
@@ -187,9 +187,9 @@ releaseInfo:
     The Agent Era is here. We can't wait to see what you'll create.
 
     <!--LANG:zh-CN-->
-    Cherry Studio 1.7.0：开启智能新纪元
+    Cherry Studio 1.7.1：开启智能新纪元
 
-    今天，我们正式发布 Cherry Studio 1.7.0 —— 迄今最具雄心的版本，带来全新的 Agent：能够自主思考、规划和行动的 AI。
+    今天，我们正式发布 Cherry Studio 1.7.1 —— 迄今最具雄心的版本，带来全新的 Agent：能够自主思考、规划和行动的 AI。
 
     多年来，AI 助手一直是被动的——等待你的指令，回应你的问题。Agent 改变了这一切。现在，AI 能够真正与你并肩工作：理解复杂目标，将其拆解为步骤，并独立执行。
 

package.json

@@ -1,6 +1,6 @@
 {
   "name": "CherryStudio",
-  "version": "1.7.0",
+  "version": "1.7.1",
   "private": true,
   "description": "A powerful AI assistant for producer.",
   "main": "./out/main/index.js",

packages/ai-sdk-provider/src/cherryin-provider.ts

@@ -69,6 +69,7 @@ export interface CherryInProviderSettings {
   headers?: HeadersInput
   /**
    * Optional endpoint type to distinguish different endpoint behaviors.
+   * "image-generation" is also openai endpoint, but specifically for image generation.
    */
   endpointType?: 'openai' | 'openai-response' | 'anthropic' | 'gemini' | 'image-generation' | 'jina-rerank'
 }

src/main/services/agents/services/claudecode/index.ts

@@ -1,6 +1,7 @@
 // src/main/services/agents/services/claudecode/index.ts
 import { EventEmitter } from 'node:events'
 import { createRequire } from 'node:module'
+import path from 'node:path'
 
 import type {
   CanUseTool,
@@ -121,7 +122,11 @@ class ClaudeCodeService implements AgentServiceInterface {
       // TODO: support set small model in UI
       ANTHROPIC_DEFAULT_HAIKU_MODEL: modelInfo.modelId,
       ELECTRON_RUN_AS_NODE: '1',
-      ELECTRON_NO_ATTACH_CONSOLE: '1'
+      ELECTRON_NO_ATTACH_CONSOLE: '1',
+      // Set CLAUDE_CONFIG_DIR to app's userData directory to avoid path encoding issues
+      // on Windows when the username contains non-ASCII characters (e.g., Chinese characters)
+      // This prevents the SDK from using the user's home directory which may have encoding problems
+      CLAUDE_CONFIG_DIR: path.join(app.getPath('userData'), '.claude')
     }
 
     const errorChunks: string[] = []

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

@@ -107,7 +107,7 @@ export async function buildStreamTextParams(
     searchWithTime: store.getState().websearch.searchWithTime
   }
 
-  const { providerOptions, standardParams } = buildProviderOptions(assistant, model, provider, {
+  const { providerOptions, standardParams, bodyParams } = buildProviderOptions(assistant, model, provider, {
     enableReasoning,
     enableWebSearch,
     enableGenerateImage
@@ -185,6 +185,7 @@ export async function buildStreamTextParams(
   // Note: standardParams (topK, frequencyPenalty, presencePenalty, stopSequences, seed)
   // are extracted from custom parameters and passed directly to streamText()
   // instead of being placed in providerOptions
+  // Note: bodyParams are custom parameters for AI Gateway that should be at body level
   const params: StreamTextParams = {
     messages: sdkMessages,
     maxOutputTokens: getMaxTokens(assistant, model),
@@ -192,6 +193,8 @@ export async function buildStreamTextParams(
     topP: getTopP(assistant, model),
     // Include AI SDK standard params extracted from custom parameters
     ...standardParams,
+    // Include body-level params for AI Gateway custom parameters
+    ...bodyParams,
     abortSignal: options.requestOptions?.signal,
     headers,
     providerOptions,

src/renderer/src/aiCore/trace/AiSdkSpanAdapter.ts

@@ -245,8 +245,8 @@ export class AiSdkSpanAdapter {
       'gen_ai.usage.output_tokens'
     ]
 
-    const completionTokens = attributes[inputsTokenKeys.find((key) => attributes[key]) || '']
-    const promptTokens = attributes[outputTokenKeys.find((key) => attributes[key]) || '']
+    const promptTokens = attributes[inputsTokenKeys.find((key) => attributes[key]) || '']
+    const completionTokens = attributes[outputTokenKeys.find((key) => attributes[key]) || '']
 
     if (completionTokens !== undefined || promptTokens !== undefined) {
       const usage: TokenUsage = {

src/renderer/src/aiCore/trace/__tests__/AiSdkSpanAdapter.test.ts

@@ -0,0 +1,53 @@
+import type { Span } from '@opentelemetry/api'
+import { SpanKind, SpanStatusCode } from '@opentelemetry/api'
+import { describe, expect, it, vi } from 'vitest'
+
+import { AiSdkSpanAdapter } from '../AiSdkSpanAdapter'
+
+vi.mock('@logger', () => ({
+  loggerService: {
+    withContext: () => ({
+      debug: vi.fn(),
+      error: vi.fn(),
+      info: vi.fn(),
+      warn: vi.fn()
+    })
+  }
+}))
+
+describe('AiSdkSpanAdapter', () => {
+  const createMockSpan = (attributes: Record<string, unknown>): Span => {
+    const span = {
+      spanContext: () => ({
+        traceId: 'trace-id',
+        spanId: 'span-id'
+      }),
+      _attributes: attributes,
+      _events: [],
+      name: 'test span',
+      status: { code: SpanStatusCode.OK },
+      kind: SpanKind.CLIENT,
+      startTime: [0, 0] as [number, number],
+      endTime: [0, 1] as [number, number],
+      ended: true,
+      parentSpanId: '',
+      links: []
+    }
+    return span as unknown as Span
+  }
+
+  it('maps prompt and completion usage tokens to the correct fields', () => {
+    const attributes = {
+      'ai.usage.promptTokens': 321,
+      'ai.usage.completionTokens': 654
+    }
+
+    const span = createMockSpan(attributes)
+    const result = AiSdkSpanAdapter.convertToSpanEntity({ span })
+
+    expect(result.usage).toBeDefined()
+    expect(result.usage?.prompt_tokens).toBe(321)
+    expect(result.usage?.completion_tokens).toBe(654)
+    expect(result.usage?.total_tokens).toBe(975)
+  })
+})

src/renderer/src/aiCore/utils/__tests__/options.test.ts

@@ -37,7 +37,7 @@ vi.mock('@cherrystudio/ai-core/provider', async (importOriginal) => {
     },
     customProviderIdSchema: {
       safeParse: vi.fn((id) => {
-        const customProviders = ['google-vertex', 'google-vertex-anthropic', 'bedrock']
+        const customProviders = ['google-vertex', 'google-vertex-anthropic', 'bedrock', 'ai-gateway']
         if (customProviders.includes(id)) {
           return { success: true, data: id }
         }
@@ -56,7 +56,8 @@ vi.mock('../provider/factory', () => ({
       [SystemProviderIds.anthropic]: 'anthropic',
       [SystemProviderIds.grok]: 'xai',
       [SystemProviderIds.deepseek]: 'deepseek',
-      [SystemProviderIds.openrouter]: 'openrouter'
+      [SystemProviderIds.openrouter]: 'openrouter',
+      [SystemProviderIds['ai-gateway']]: 'ai-gateway'
     }
     return mapping[provider.id] || provider.id
   })
@@ -204,6 +205,8 @@ describe('options utils', () => {
         expect(result.providerOptions).toHaveProperty('openai')
         expect(result.providerOptions.openai).toBeDefined()
         expect(result.standardParams).toBeDefined()
+        expect(result.bodyParams).toBeDefined()
+        expect(result.bodyParams).toEqual({})
       })
 
       it('should include reasoning parameters when enabled', () => {
@@ -696,5 +699,90 @@ describe('options utils', () => {
         })
       })
     })
+
+    describe('AI Gateway provider', () => {
+      const aiGatewayProvider: Provider = {
+        id: SystemProviderIds['ai-gateway'],
+        name: 'AI Gateway',
+        type: 'ai-gateway',
+        apiKey: 'test-key',
+        apiHost: 'https://ai-gateway.vercel.sh/v1/ai',
+        isSystem: true,
+        models: [] as Model[]
+      } as Provider
+
+      const aiGatewayModel: Model = {
+        id: 'openai/gpt-4',
+        name: 'GPT-4',
+        provider: SystemProviderIds['ai-gateway']
+      } as Model
+
+      it('should build basic AI Gateway options with empty bodyParams', () => {
+        const result = buildProviderOptions(mockAssistant, aiGatewayModel, aiGatewayProvider, {
+          enableReasoning: false,
+          enableWebSearch: false,
+          enableGenerateImage: false
+        })
+
+        expect(result.providerOptions).toHaveProperty('gateway')
+        expect(result.providerOptions.gateway).toBeDefined()
+        expect(result.bodyParams).toEqual({})
+      })
+
+      it('should place custom parameters in bodyParams for AI Gateway instead of providerOptions', async () => {
+        const { getCustomParameters } = await import('../reasoning')
+
+        vi.mocked(getCustomParameters).mockReturnValue({
+          tools: [{ id: 'openai.image_generation' }],
+          custom_param: 'custom_value'
+        })
+
+        const result = buildProviderOptions(mockAssistant, aiGatewayModel, aiGatewayProvider, {
+          enableReasoning: false,
+          enableWebSearch: false,
+          enableGenerateImage: false
+        })
+
+        // Custom parameters should be in bodyParams, NOT in providerOptions.gateway
+        expect(result.bodyParams).toHaveProperty('tools')
+        expect(result.bodyParams.tools).toEqual([{ id: 'openai.image_generation' }])
+        expect(result.bodyParams).toHaveProperty('custom_param')
+        expect(result.bodyParams.custom_param).toBe('custom_value')
+
+        // providerOptions.gateway should NOT contain custom parameters
+        expect(result.providerOptions.gateway).not.toHaveProperty('tools')
+        expect(result.providerOptions.gateway).not.toHaveProperty('custom_param')
+      })
+
+      it('should still extract AI SDK standard params from custom parameters for AI Gateway', async () => {
+        const { getCustomParameters } = await import('../reasoning')
+
+        vi.mocked(getCustomParameters).mockReturnValue({
+          topK: 5,
+          frequencyPenalty: 0.5,
+          tools: [{ id: 'openai.image_generation' }]
+        })
+
+        const result = buildProviderOptions(mockAssistant, aiGatewayModel, aiGatewayProvider, {
+          enableReasoning: false,
+          enableWebSearch: false,
+          enableGenerateImage: false
+        })
+
+        // Standard params should be extracted and returned separately
+        expect(result.standardParams).toEqual({
+          topK: 5,
+          frequencyPenalty: 0.5
+        })
+
+        // Custom params (non-standard) should be in bodyParams
+        expect(result.bodyParams).toHaveProperty('tools')
+        expect(result.bodyParams.tools).toEqual([{ id: 'openai.image_generation' }])
+
+        // Neither should be in providerOptions.gateway
+        expect(result.providerOptions.gateway).not.toHaveProperty('topK')
+        expect(result.providerOptions.gateway).not.toHaveProperty('tools')
+      })
+    })
   })
 })

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

@@ -144,7 +144,7 @@ describe('reasoning utils', () => {
       expect(result).toEqual({})
     })
 
-    it('should disable reasoning for OpenRouter when no reasoning effort set', async () => {
+    it('should not override reasoning for OpenRouter when reasoning effort undefined', async () => {
       const { isReasoningModel } = await import('@renderer/config/models')
 
       vi.mocked(isReasoningModel).mockReturnValue(true)
@@ -161,6 +161,29 @@ describe('reasoning utils', () => {
         settings: {}
       } as Assistant
 
+      const result = getReasoningEffort(assistant, model)
+      expect(result).toEqual({})
+    })
+
+    it('should disable reasoning for OpenRouter when reasoning effort explicitly none', async () => {
+      const { isReasoningModel } = await import('@renderer/config/models')
+
+      vi.mocked(isReasoningModel).mockReturnValue(true)
+
+      const model: Model = {
+        id: 'anthropic/claude-sonnet-4',
+        name: 'Claude Sonnet 4',
+        provider: SystemProviderIds.openrouter
+      } as Model
+
+      const assistant: Assistant = {
+        id: 'test',
+        name: 'Test',
+        settings: {
+          reasoning_effort: 'none'
+        }
+      } as Assistant
+
       const result = getReasoningEffort(assistant, model)
       expect(result).toEqual({ reasoning: { enabled: false, exclude: true } })
     })
@@ -269,7 +292,9 @@ describe('reasoning utils', () => {
       const assistant: Assistant = {
         id: 'test',
         name: 'Test',
-        settings: {}
+        settings: {
+          reasoning_effort: 'none'
+        }
       } as Assistant
 
       const result = getReasoningEffort(assistant, model)

src/renderer/src/aiCore/utils/options.ts

@@ -155,6 +155,7 @@ export function buildProviderOptions(
 ): {
   providerOptions: Record<string, Record<string, JSONValue>>
   standardParams: Partial<Record<AiSdkParam, any>>
+  bodyParams: Record<string, any>
 } {
   logger.debug('buildProviderOptions', { assistant, model, actualProvider, capabilities })
   const rawProviderId = getAiSdkProviderId(actualProvider)
@@ -253,12 +254,6 @@ export function buildProviderOptions(
   const customParams = getCustomParameters(assistant)
   const { standardParams, providerParams } = extractAiSdkStandardParams(customParams)
 
-  // 合并 provider 特定的自定义参数到 providerSpecificOptions
-  providerSpecificOptions = {
-    ...providerSpecificOptions,
-    ...providerParams
-  }
-
   let rawProviderKey =
     {
       'google-vertex': 'google',
@@ -273,12 +268,27 @@ export function buildProviderOptions(
     rawProviderKey = { gemini: 'google', ['openai-response']: 'openai' }[actualProvider.type] || actualProvider.type
   }
 
-  // 返回 AI Core SDK 要求的格式：{ 'providerId': providerOptions } 以及提取的标准参数
+  // For AI Gateway, custom parameters should be placed at body level, not inside providerOptions.gateway
+  // See: https://github.com/CherryHQ/cherry-studio/issues/4197
+  let bodyParams: Record<string, any> = {}
+  if (rawProviderKey === 'gateway') {
+    // Custom parameters go to body level for AI Gateway
+    bodyParams = providerParams
+  } else {
+    // For other providers, merge custom parameters into providerSpecificOptions
+    providerSpecificOptions = {
+      ...providerSpecificOptions,
+      ...providerParams
+    }
+  }
+
+  // 返回 AI Core SDK 要求的格式：{ 'providerId': providerOptions } 以及提取的标准参数和 body 参数
   return {
     providerOptions: {
       [rawProviderKey]: providerSpecificOptions
     },
-    standardParams
+    standardParams,
+    bodyParams
   }
 }
 

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

@@ -16,10 +16,8 @@ import {
   isGPT5SeriesModel,
   isGPT51SeriesModel,
   isGrok4FastReasoningModel,
-  isGrokReasoningModel,
   isOpenAIDeepResearchModel,
   isOpenAIModel,
-  isOpenAIReasoningModel,
   isQwenAlwaysThinkModel,
   isQwenReasoningModel,
   isReasoningModel,
@@ -64,30 +62,22 @@ export function getReasoningEffort(assistant: Assistant, model: Model): Reasonin
   }
   const reasoningEffort = assistant?.settings?.reasoning_effort
 
-  // Handle undefined and 'none' reasoningEffort.
-  // TODO: They should be separated.
-  if (!reasoningEffort || reasoningEffort === 'none') {
+  // reasoningEffort is not set, no extra reasoning setting
+  // Generally, for every model which supports reasoning control, the reasoning effort won't be undefined.
+  // It's for some reasoning models that don't support reasoning control, such as deepseek reasoner.
+  if (!reasoningEffort) {
+    return {}
+  }
+
+  // Handle 'none' reasoningEffort. It's explicitly off.
+  if (reasoningEffort === 'none') {
     // openrouter: use reasoning
     if (model.provider === SystemProviderIds.openrouter) {
-      // Don't disable reasoning for Gemini models that support thinking tokens
-      if (isSupportedThinkingTokenGeminiModel(model) && !GEMINI_FLASH_MODEL_REGEX.test(model.id)) {
-        return {}
-      }
       // 'none' is not an available value for effort for now.
       // I think they should resolve this issue soon, so I'll just go ahead and use this value.
       if (isGPT51SeriesModel(model) && reasoningEffort === 'none') {
         return { reasoning: { effort: 'none' } }
       }
-      // Don't disable reasoning for models that require it
-      if (
-        isGrokReasoningModel(model) ||
-        isOpenAIReasoningModel(model) ||
-        isQwenAlwaysThinkModel(model) ||
-        model.id.includes('seed-oss') ||
-        model.id.includes('minimax-m2')
-      ) {
-        return {}
-      }
       return { reasoning: { enabled: false, exclude: true } }
     }
 
@@ -101,11 +91,6 @@ export function getReasoningEffort(assistant: Assistant, model: Model): Reasonin
       return { enable_thinking: false }
     }
 
-    // claude
-    if (isSupportedThinkingTokenClaudeModel(model)) {
-      return {}
-    }
-
     // gemini
     if (isSupportedThinkingTokenGeminiModel(model)) {
       if (GEMINI_FLASH_MODEL_REGEX.test(model.id)) {
@@ -118,8 +103,10 @@ export function getReasoningEffort(assistant: Assistant, model: Model): Reasonin
             }
           }
         }
+      } else {
+        logger.warn(`Model ${model.id} cannot disable reasoning. Fallback to empty reasoning param.`)
+        return {}
       }
-      return {}
     }
 
     // use thinking, doubao, zhipu, etc.
@@ -139,6 +126,7 @@ export function getReasoningEffort(assistant: Assistant, model: Model): Reasonin
       }
     }
 
+    logger.warn(`Model ${model.id} doesn't match any disable reasoning behavior. Fallback to empty reasoning param.`)
     return {}
   }
 
@@ -293,6 +281,7 @@ export function getReasoningEffort(assistant: Assistant, model: Model): Reasonin
   }
 
   // OpenRouter models, use reasoning
+  // FIXME: duplicated openrouter handling. remove one
   if (model.provider === SystemProviderIds.openrouter) {
     if (isSupportedReasoningEffortModel(model) || isSupportedThinkingTokenModel(model)) {
       return {

src/renderer/src/assets/styles/ant.css

@@ -215,6 +215,10 @@
   border-top: none !important;
 }
 
+.ant-collapse-header-text {
+  overflow-x: hidden;
+}
+
 .ant-slider .ant-slider-handle::after {
   box-shadow: 0 1px 4px 0px rgb(128 128 128 / 50%) !important;
 }

src/renderer/src/components/ImageViewer.tsx

@@ -14,7 +14,7 @@ import { convertImageToPng } from '@renderer/utils/image'
 import type { ImageProps as AntImageProps } from 'antd'
 import { Dropdown, Image as AntImage, Space } from 'antd'
 import { Base64 } from 'js-base64'
-import { DownloadIcon } from 'lucide-react'
+import { DownloadIcon, ImageIcon } from 'lucide-react'
 import mime from 'mime'
 import React from 'react'
 import { useTranslation } from 'react-i18next'
@@ -72,15 +72,9 @@ const ImageViewer: React.FC<ImageViewerProps> = ({ src, style, ...props }) => {
 
   const getContextMenuItems = (src: string, size: number = 14) => {
     return [
-      {
-        key: 'copy-image',
-        label: t('common.copy'),
-        icon: <CopyIcon size={size} />,
-        onClick: () => handleCopyImage(src)
-      },
       {
         key: 'copy-url',
-        label: t('preview.copy.src'),
+        label: t('common.copy'),
         icon: <CopyIcon size={size} />,
         onClick: () => {
           navigator.clipboard.writeText(src)
@@ -92,6 +86,12 @@ const ImageViewer: React.FC<ImageViewerProps> = ({ src, style, ...props }) => {
         label: t('common.download'),
         icon: <DownloadIcon size={size} />,
         onClick: () => download(src)
+      },
+      {
+        key: 'copy-image',
+        label: t('preview.copy.image'),
+        icon: <ImageIcon size={size} />,
+        onClick: () => handleCopyImage(src)
       }
     ]
   }

src/renderer/src/config/models/reasoning.ts

@@ -460,16 +460,19 @@ export const isSupportedThinkingTokenZhipuModel = (model: Model): boolean => {
 }
 
 export const isDeepSeekHybridInferenceModel = (model: Model) => {
-  const modelId = getLowerBaseModelName(model.id)
-  // deepseek官方使用chat和reasoner做推理控制，其他provider需要单独判断，id可能会有所差别
-  // openrouter: deepseek/deepseek-chat-v3.1 不知道会不会有其他provider仿照ds官方分出一个同id的作为非思考模式的模型，这里有风险
-  // Matches: "deepseek-v3" followed by ".digit" or "-digit".
-  // Optionally, this can be followed by ".alphanumeric_sequence" or "-alphanumeric_sequence"
-  // until the end of the string.
-  // Examples: deepseek-v3.1, deepseek-v3-1, deepseek-v3.1.2, deepseek-v3.1-alpha
-  // Does NOT match: deepseek-v3.123 (missing separator after '1'), deepseek-v3.x (x isn't a digit)
-  // TODO: move to utils and add test cases
-  return /deepseek-v3(?:\.\d|-\d)(?:(\.|-)\w+)?$/.test(modelId) || modelId.includes('deepseek-chat-v3.1')
+  const { idResult, nameResult } = withModelIdAndNameAsId(model, (model) => {
+    const modelId = getLowerBaseModelName(model.id)
+    // deepseek官方使用chat和reasoner做推理控制，其他provider需要单独判断，id可能会有所差别
+    // openrouter: deepseek/deepseek-chat-v3.1 不知道会不会有其他provider仿照ds官方分出一个同id的作为非思考模式的模型，这里有风险
+    // Matches: "deepseek-v3" followed by ".digit" or "-digit".
+    // Optionally, this can be followed by ".alphanumeric_sequence" or "-alphanumeric_sequence"
+    // until the end of the string.
+    // Examples: deepseek-v3.1, deepseek-v3-1, deepseek-v3.1.2, deepseek-v3.1-alpha
+    // Does NOT match: deepseek-v3.123 (missing separator after '1'), deepseek-v3.x (x isn't a digit)
+    // TODO: move to utils and add test cases
+    return /deepseek-v3(?:\.\d|-\d)(?:(\.|-)\w+)?$/.test(modelId) || modelId.includes('deepseek-chat-v3.1')
+  })
+  return idResult || nameResult
 }
 
 export const isLingReasoningModel = (model?: Model): boolean => {
@@ -523,7 +526,6 @@ export function isReasoningModel(model?: Model): boolean {
       REASONING_REGEX.test(model.name) ||
       isSupportedThinkingTokenDoubaoModel(model) ||
       isDeepSeekHybridInferenceModel(model) ||
-      isDeepSeekHybridInferenceModel({ ...model, id: model.name }) ||
       false
     )
   }

src/renderer/src/hooks/useScrollPosition.ts

@@ -1,5 +1,5 @@
 import { throttle } from 'lodash'
-import { useEffect, useRef } from 'react'
+import { useEffect, useMemo, useRef } from 'react'
 
 import { useTimer } from './useTimer'
 
@@ -12,13 +12,18 @@ import { useTimer } from './useTimer'
  */
 export default function useScrollPosition(key: string, throttleWait?: number) {
   const containerRef = useRef<HTMLDivElement>(null)
-  const scrollKey = `scroll:${key}`
+  const scrollKey = useMemo(() => `scroll:${key}`, [key])
+  const scrollKeyRef = useRef(scrollKey)
   const { setTimeoutTimer } = useTimer()
 
+  useEffect(() => {
+    scrollKeyRef.current = scrollKey
+  }, [scrollKey])
+
   const handleScroll = throttle(() => {
     const position = containerRef.current?.scrollTop ?? 0
     window.requestAnimationFrame(() => {
-      window.keyv.set(scrollKey, position)
+      window.keyv.set(scrollKeyRef.current, position)
     })
   }, throttleWait ?? 100)
 
@@ -28,5 +33,9 @@ export default function useScrollPosition(key: string, throttleWait?: number) {
     setTimeoutTimer('scrollEffect', scroll, 50)
   }, [scrollKey, setTimeoutTimer])
 
+  useEffect(() => {
+    return () => handleScroll.cancel()
+  }, [handleScroll])
+
   return { containerRef, handleScroll }
 }

src/renderer/src/hooks/useTimer.ts

@@ -1,4 +1,4 @@
-import { useEffect, useRef } from 'react'
+import { useCallback, useEffect, useRef } from 'react'
 
 /**
  * 定时器管理 Hook，用于管理 setTimeout 和 setInterval 定时器，支持通过 key 来标识不同的定时器
@@ -43,10 +43,38 @@ export const useTimer = () => {
   const timeoutMapRef = useRef(new Map<string, NodeJS.Timeout>())
   const intervalMapRef = useRef(new Map<string, NodeJS.Timeout>())
 
+  /**
+   * 清除指定 key 的 setTimeout 定时器
+   * @param key - 定时器标识符
+   */
+  const clearTimeoutTimer = useCallback((key: string) => {
+    clearTimeout(timeoutMapRef.current.get(key))
+    timeoutMapRef.current.delete(key)
+  }, [])
+
+  /**
+   * 清除指定 key 的 setInterval 定时器
+   * @param key - 定时器标识符
+   */
+  const clearIntervalTimer = useCallback((key: string) => {
+    clearInterval(intervalMapRef.current.get(key))
+    intervalMapRef.current.delete(key)
+  }, [])
+
+  /**
+   * 清除所有定时器，包括 setTimeout 和 setInterval
+   */
+  const clearAllTimers = useCallback(() => {
+    timeoutMapRef.current.forEach((timer) => clearTimeout(timer))
+    intervalMapRef.current.forEach((timer) => clearInterval(timer))
+    timeoutMapRef.current.clear()
+    intervalMapRef.current.clear()
+  }, [])
+
   // 组件卸载时自动清理所有定时器
   useEffect(() => {
     return () => clearAllTimers()
-  }, [])
+  }, [clearAllTimers])
 
   /**
    * 设置一个 setTimeout 定时器
@@ -65,12 +93,15 @@ export const useTimer = () => {
    * cleanup();
    * ```
    */
-  const setTimeoutTimer = (key: string, ...args: Parameters<typeof setTimeout>) => {
-    clearTimeout(timeoutMapRef.current.get(key))
-    const timer = setTimeout(...args)
-    timeoutMapRef.current.set(key, timer)
-    return () => clearTimeoutTimer(key)
-  }
+  const setTimeoutTimer = useCallback(
+    (key: string, ...args: Parameters<typeof setTimeout>) => {
+      clearTimeout(timeoutMapRef.current.get(key))
+      const timer = setTimeout(...args)
+      timeoutMapRef.current.set(key, timer)
+      return () => clearTimeoutTimer(key)
+    },
+    [clearTimeoutTimer]
+  )
 
   /**
    * 设置一个 setInterval 定时器
@@ -89,56 +120,31 @@ export const useTimer = () => {
    * cleanup();
    * ```
    */
-  const setIntervalTimer = (key: string, ...args: Parameters<typeof setInterval>) => {
-    clearInterval(intervalMapRef.current.get(key))
-    const timer = setInterval(...args)
-    intervalMapRef.current.set(key, timer)
-    return () => clearIntervalTimer(key)
-  }
-
-  /**
-   * 清除指定 key 的 setTimeout 定时器
-   * @param key - 定时器标识符
-   */
-  const clearTimeoutTimer = (key: string) => {
-    clearTimeout(timeoutMapRef.current.get(key))
-    timeoutMapRef.current.delete(key)
-  }
-
-  /**
-   * 清除指定 key 的 setInterval 定时器
-   * @param key - 定时器标识符
-   */
-  const clearIntervalTimer = (key: string) => {
-    clearInterval(intervalMapRef.current.get(key))
-    intervalMapRef.current.delete(key)
-  }
+  const setIntervalTimer = useCallback(
+    (key: string, ...args: Parameters<typeof setInterval>) => {
+      clearInterval(intervalMapRef.current.get(key))
+      const timer = setInterval(...args)
+      intervalMapRef.current.set(key, timer)
+      return () => clearIntervalTimer(key)
+    },
+    [clearIntervalTimer]
+  )
 
   /**
    * 清除所有 setTimeout 定时器
    */
-  const clearAllTimeoutTimers = () => {
+  const clearAllTimeoutTimers = useCallback(() => {
     timeoutMapRef.current.forEach((timer) => clearTimeout(timer))
     timeoutMapRef.current.clear()
-  }
+  }, [])
 
   /**
    * 清除所有 setInterval 定时器
    */
-  const clearAllIntervalTimers = () => {
+  const clearAllIntervalTimers = useCallback(() => {
     intervalMapRef.current.forEach((timer) => clearInterval(timer))
     intervalMapRef.current.clear()
-  }
-
-  /**
-   * 清除所有定时器，包括 setTimeout 和 setInterval
-   */
-  const clearAllTimers = () => {
-    timeoutMapRef.current.forEach((timer) => clearTimeout(timer))
-    intervalMapRef.current.forEach((timer) => clearInterval(timer))
-    timeoutMapRef.current.clear()
-    intervalMapRef.current.clear()
-  }
+  }, [])
 
   return {
     setTimeoutTimer,

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

@@ -280,6 +280,7 @@
         "denied": "Tool request was denied.",
         "timeout": "Tool request timed out before receiving approval."
       },
+      "toolPendingFallback": "Tool",
       "waiting": "Waiting for tool permission decision..."
     },
     "type": {
@@ -1208,7 +1209,7 @@
   "endpoint_type": {
     "anthropic": "Anthropic",
     "gemini": "Gemini",
-    "image-generation": "Image Generation",
+    "image-generation": "Image Generation (OpenAI)",
     "jina-rerank": "Jina Rerank",
     "openai": "OpenAI",
     "openai-response": "OpenAI-Response"
@@ -2509,8 +2510,7 @@
   },
   "preview": {
     "copy": {
-      "image": "Copy as image",
-      "src": "Copy Image Source"
+      "image": "Copy as image"
     },
     "dialog": "Open Dialog",
     "label": "Preview",

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

@@ -280,6 +280,7 @@
         "denied": "工具请求已被拒绝。",
         "timeout": "工具请求在收到批准前超时。"
       },
+      "toolPendingFallback": "工具",
       "waiting": "等待工具权限决定..."
     },
     "type": {
@@ -1208,7 +1209,7 @@
   "endpoint_type": {
     "anthropic": "Anthropic",
     "gemini": "Gemini",
-    "image-generation": "图片生成",
+    "image-generation": "图像生成 (OpenAI)",
     "jina-rerank": "Jina 重排序",
     "openai": "OpenAI",
     "openai-response": "OpenAI-Response"
@@ -2509,8 +2510,7 @@
   },
   "preview": {
     "copy": {
-      "image": "复制为图片",
-      "src": "复制图片源"
+      "image": "复制为图片"
     },
     "dialog": "打开预览窗口",
     "label": "预览",

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

@@ -280,6 +280,7 @@
         "denied": "工具請求已被拒絕。",
         "timeout": "工具請求在收到核准前逾時。"
       },
+      "toolPendingFallback": "工具",
       "waiting": "等待工具權限決定..."
     },
     "type": {
@@ -1208,7 +1209,7 @@
   "endpoint_type": {
     "anthropic": "Anthropic",
     "gemini": "Gemini",
-    "image-generation": "圖片生成",
+    "image-generation": "圖像生成 (OpenAI)",
     "jina-rerank": "Jina Rerank",
     "openai": "OpenAI",
     "openai-response": "OpenAI-Response"
@@ -2509,8 +2510,7 @@
   },
   "preview": {
     "copy": {
-      "image": "複製為圖片",
-      "src": "複製圖片來源"
+      "image": "複製為圖片"
     },
     "dialog": "開啟預覽窗口",
     "label": "預覽",

src/renderer/src/i18n/translate/de-de.json

@@ -280,6 +280,7 @@
         "denied": "Tool-Anfrage wurde abgelehnt.",
         "timeout": "Tool-Anfrage ist abgelaufen, bevor eine Genehmigung eingegangen ist."
       },
+      "toolPendingFallback": "Werkzeug",
       "waiting": "Warten auf Entscheidung über Tool-Berechtigung..."
     },
     "type": {
@@ -1208,7 +1209,7 @@
   "endpoint_type": {
     "anthropic": "Anthropic",
     "gemini": "Gemini",
-    "image-generation": "Bildgenerierung",
+    "image-generation": "Bilderzeugung (OpenAI)",
     "jina-rerank": "Jina Reranking",
     "openai": "OpenAI",
     "openai-response": "OpenAI-Response"
@@ -2509,8 +2510,7 @@
   },
   "preview": {
     "copy": {
-      "image": "Als Bild kopieren",
-      "src": "Bildquelle kopieren"
+      "image": "Als Bild kopieren"
     },
     "dialog": "Vorschaufenster öffnen",
     "label": "Vorschau",

src/renderer/src/i18n/translate/el-gr.json

@@ -280,6 +280,7 @@
         "denied": "Το αίτημα για εργαλείο απορρίφθηκε.",
         "timeout": "Το αίτημα για το εργαλείο έληξε πριν λάβει έγκριση."
       },
+      "toolPendingFallback": "Εργαλείο",
       "waiting": "Αναμονή για απόφαση άδειας εργαλείου..."
     },
     "type": {
@@ -1208,7 +1209,7 @@
   "endpoint_type": {
     "anthropic": "Anthropic",
     "gemini": "Gemini",
-    "image-generation": "Δημιουργία Εικόνας",
+    "image-generation": "Δημιουργία Εικόνων (OpenAI)",
     "jina-rerank": "Επαναταξινόμηση Jina",
     "openai": "OpenAI",
     "openai-response": "Απάντηση OpenAI"
@@ -2509,8 +2510,7 @@
   },
   "preview": {
     "copy": {
-      "image": "Αντιγραφή ως εικόνα",
-      "src": "Αντιγραφή πηγής εικόνας"
+      "image": "Αντιγραφή ως εικόνα"
     },
     "dialog": "Άνοιγμα παραθύρου προεπισκόπησης",
     "label": "Προεπισκόπηση",

src/renderer/src/i18n/translate/es-es.json

@@ -280,6 +280,7 @@
         "denied": "La solicitud de herramienta fue denegada.",
         "timeout": "La solicitud de herramienta expiró antes de recibir la aprobación."
       },
+      "toolPendingFallback": "Herramienta",
       "waiting": "Esperando la decisión de permiso de la herramienta..."
     },
     "type": {
@@ -1208,7 +1209,7 @@
   "endpoint_type": {
     "anthropic": "Anthropic",
     "gemini": "Gemini",
-    "image-generation": "Generación de imágenes",
+    "image-generation": "Generación de Imágenes (OpenAI)",
     "jina-rerank": "Reordenamiento Jina",
     "openai": "OpenAI",
     "openai-response": "Respuesta de OpenAI"
@@ -2509,8 +2510,7 @@
   },
   "preview": {
     "copy": {
-      "image": "Copiar como imagen",
-      "src": "Copia la fuente de la imagen"
+      "image": "Copiar como imagen"
     },
     "dialog": "Abrir la ventana de vista previa",
     "label": "Vista previa",

src/renderer/src/i18n/translate/fr-fr.json

@@ -280,6 +280,7 @@
         "denied": "La demande d'outil a été refusée.",
         "timeout": "La demande d'outil a expiré avant d'obtenir l'approbation."
       },
+      "toolPendingFallback": "Outil",
       "waiting": "En attente de la décision d'autorisation de l'outil..."
     },
     "type": {
@@ -1208,7 +1209,7 @@
   "endpoint_type": {
     "anthropic": "Anthropic",
     "gemini": "Gemini",
-    "image-generation": "Génération d'images",
+    "image-generation": "Génération d'images (OpenAI)",
     "jina-rerank": "Reclassement Jina",
     "openai": "OpenAI",
     "openai-response": "Réponse OpenAI"
@@ -2509,8 +2510,7 @@
   },
   "preview": {
     "copy": {
-      "image": "Copier en tant qu'image",
-      "src": "Copier la source de l'image"
+      "image": "Copier en tant qu'image"
     },
     "dialog": "Ouvrir la fenêtre d'aperçu",
     "label": "Aperçu",

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

@@ -280,6 +280,7 @@
         "denied": "ツールリクエストは拒否されました。",
         "timeout": "ツールリクエストは承認を受ける前にタイムアウトしました。"
       },
+      "toolPendingFallback": "ツール",
       "waiting": "ツールの許可決定を待っています..."
     },
     "type": {
@@ -1208,7 +1209,7 @@
   "endpoint_type": {
     "anthropic": "Anthropic",
     "gemini": "Gemini",
-    "image-generation": "画像生成",
+    "image-generation": "画像生成 (OpenAI)",
     "jina-rerank": "Jina Rerank",
     "openai": "OpenAI",
     "openai-response": "OpenAI-Response"
@@ -2509,8 +2510,7 @@
   },
   "preview": {
     "copy": {
-      "image": "画像としてコピー",
-      "src": "画像ソースをコピー"
+      "image": "画像としてコピー"
     },
     "dialog": "ダイアログを開く",
     "label": "プレビュー",

src/renderer/src/i18n/translate/pt-pt.json

@@ -16,7 +16,7 @@
       "error": {
         "failed": "Falha ao excluir o agente"
       },
-      "title": "删除代理"
+      "title": "Excluir Agente"
     },
     "edit": {
       "title": "Agent Editor"
@@ -111,7 +111,7 @@
           "label": "Modo de permissão",
           "options": {
             "acceptEdits": "Aceitar edições automaticamente",
-            "bypassPermissions": "忽略检查 de permissão",
+            "bypassPermissions": "Ignorar verificações de permissão",
             "default": "Padrão (perguntar antes de continuar)",
             "plan": "Modo de planejamento (plano sujeito a aprovação)"
           },
@@ -150,7 +150,7 @@
         },
         "success": {
           "install": "Plugin instalado com sucesso",
-          "uninstall": "插件 desinstalado com sucesso"
+          "uninstall": "Plugin desinstalado com sucesso"
         },
         "tab": "plug-in",
         "type": {
@@ -280,6 +280,7 @@
         "denied": "Solicitação de ferramenta foi negada.",
         "timeout": "A solicitação da ferramenta expirou antes de receber aprovação."
       },
+      "toolPendingFallback": "Ferramenta",
       "waiting": "Aguardando decisão de permissão da ferramenta..."
     },
     "type": {
@@ -1134,7 +1135,7 @@
     "duplicate": "Duplicar",
     "edit": "Editar",
     "enabled": "Ativado",
-    "error": "错误",
+    "error": "Erro",
     "errors": {
       "create_message": "Falha ao criar mensagem",
       "validation": "Falha na verificação"
@@ -1208,7 +1209,7 @@
   "endpoint_type": {
     "anthropic": "Anthropic",
     "gemini": "Gemini",
-    "image-generation": "Geração de Imagem",
+    "image-generation": "Geração de Imagens (OpenAI)",
     "jina-rerank": "Jina Reordenar",
     "openai": "OpenAI",
     "openai-response": "Resposta OpenAI"
@@ -2509,8 +2510,7 @@
   },
   "preview": {
     "copy": {
-      "image": "Copiar como imagem",
-      "src": "Copiar Origem da Imagem"
+      "image": "Copiar como imagem"
     },
     "dialog": "Abrir janela de pré-visualização",
     "label": "Pré-visualização",

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

@@ -280,6 +280,7 @@
         "denied": "Запрос на инструмент был отклонён.",
         "timeout": "Запрос на инструмент превысил время ожидания до получения подтверждения."
       },
+      "toolPendingFallback": "Инструмент",
       "waiting": "Ожидание решения о разрешении на использование инструмента..."
     },
     "type": {
@@ -1208,7 +1209,7 @@
   "endpoint_type": {
     "anthropic": "Anthropic",
     "gemini": "Gemini",
-    "image-generation": "Изображение",
+    "image-generation": "Генерация изображений (OpenAI)",
     "jina-rerank": "Jina Rerank",
     "openai": "OpenAI",
     "openai-response": "OpenAI-Response"
@@ -2509,8 +2510,7 @@
   },
   "preview": {
     "copy": {
-      "image": "Скопировать как изображение",
-      "src": "Копировать источник изображения"
+      "image": "Скопировать как изображение"
     },
     "dialog": "Открыть диалог",
     "label": "Предварительный просмотр",

src/renderer/src/pages/home/Inputbar/AgentSessionInputbar.tsx

@@ -9,6 +9,7 @@ import { getModel } from '@renderer/hooks/useModel'
 import { useSettings } from '@renderer/hooks/useSettings'
 import { useTextareaResize } from '@renderer/hooks/useTextareaResize'
 import { useTimer } from '@renderer/hooks/useTimer'
+import { CacheService } from '@renderer/services/CacheService'
 import { pauseTrace } from '@renderer/services/SpanManagerService'
 import { estimateUserPromptUsage } from '@renderer/services/TokenService'
 import { useAppDispatch, useAppSelector } from '@renderer/store'
@@ -41,19 +42,10 @@ import { getInputbarConfig } from './registry'
 import { TopicType } from './types'
 
 const logger = loggerService.withContext('AgentSessionInputbar')
-const agentSessionDraftCache = new Map<string, string>()
 
-const readDraftFromCache = (key: string): string => {
-  return agentSessionDraftCache.get(key) ?? ''
-}
+const DRAFT_CACHE_TTL = 24 * 60 * 60 * 1000 // 24 hours
 
-const writeDraftToCache = (key: string, value: string) => {
-  if (!value) {
-    agentSessionDraftCache.delete(key)
-  } else {
-    agentSessionDraftCache.set(key, value)
-  }
-}
+const getAgentDraftCacheKey = (agentId: string) => `agent-session-draft-${agentId}`
 
 type Props = {
   agentId: string
@@ -170,16 +162,15 @@ const AgentSessionInputbarInner: FC<InnerProps> = ({ assistant, agentId, session
   const scope = TopicType.Session
   const config = getInputbarConfig(scope)
 
-  // Use shared hooks for text and textarea management
-  const initialDraft = useMemo(() => readDraftFromCache(agentId), [agentId])
-  const persistDraft = useCallback((next: string) => writeDraftToCache(agentId, next), [agentId])
+  // Use shared hooks for text and textarea management with draft persistence
+  const draftCacheKey = getAgentDraftCacheKey(agentId)
   const {
     text,
     setText,
     isEmpty: inputEmpty
   } = useInputText({
-    initialValue: initialDraft,
-    onChange: persistDraft
+    initialValue: CacheService.get<string>(draftCacheKey) ?? '',
+    onChange: (value) => CacheService.set(draftCacheKey, value, DRAFT_CACHE_TTL)
   })
   const {
     textareaRef,
@@ -431,6 +422,7 @@ const AgentSessionInputbarInner: FC<InnerProps> = ({ assistant, agentId, session
         })
       )
 
+      // Clear text after successful send (draft is cleared automatically via onChange)
       setText('')
       setTimeoutTimer('agentSession_sendMessage', () => setText(''), 500)
     } catch (error) {

src/renderer/src/pages/home/Inputbar/Inputbar.tsx

@@ -14,7 +14,6 @@ import { useInputText } from '@renderer/hooks/useInputText'
 import { useMessageOperations, useTopicLoading } from '@renderer/hooks/useMessageOperations'
 import { useSettings } from '@renderer/hooks/useSettings'
 import { useShortcut } from '@renderer/hooks/useShortcuts'
-import { useSidebarIconShow } from '@renderer/hooks/useSidebarIcon'
 import { useTextareaResize } from '@renderer/hooks/useTextareaResize'
 import { useTimer } from '@renderer/hooks/useTimer'
 import {
@@ -24,6 +23,7 @@ import {
   useInputbarToolsState
 } from '@renderer/pages/home/Inputbar/context/InputbarToolsProvider'
 import { getDefaultTopic } from '@renderer/services/AssistantService'
+import { CacheService } from '@renderer/services/CacheService'
 import { EVENT_NAMES, EventEmitter } from '@renderer/services/EventService'
 import FileManager from '@renderer/services/FileManager'
 import { checkRateLimit, getUserMessage } from '@renderer/services/MessagesService'
@@ -39,7 +39,7 @@ import { getSendMessageShortcutLabel } from '@renderer/utils/input'
 import { documentExts, imageExts, textExts } from '@shared/config/constant'
 import { debounce } from 'lodash'
 import type { FC } from 'react'
-import React, { useCallback, useEffect, useMemo, useRef, useState } from 'react'
+import React, { useCallback, useEffect, useEffectEvent, useMemo, useRef, useState } from 'react'
 import { useTranslation } from 'react-i18next'
 
 import { InputbarCore } from './components/InputbarCore'
@@ -51,6 +51,17 @@ import TokenCount from './TokenCount'
 
 const logger = loggerService.withContext('Inputbar')
 
+const INPUTBAR_DRAFT_CACHE_KEY = 'inputbar-draft'
+const DRAFT_CACHE_TTL = 24 * 60 * 60 * 1000 // 24 hours
+
+const getMentionedModelsCacheKey = (assistantId: string) => `inputbar-mentioned-models-${assistantId}`
+
+const getValidatedCachedModels = (assistantId: string): Model[] => {
+  const cached = CacheService.get<Model[]>(getMentionedModelsCacheKey(assistantId))
+  if (!Array.isArray(cached)) return []
+  return cached.filter((model) => model?.id && model?.name)
+}
+
 interface Props {
   assistant: Assistant
   setActiveTopic: (topic: Topic) => void
@@ -80,16 +91,18 @@ const Inputbar: FC<Props> = ({ assistant: initialAssistant, setActiveTopic, topi
     toggleExpanded: () => {}
   })
 
+  const [initialMentionedModels] = useState(() => getValidatedCachedModels(initialAssistant.id))
+
   const initialState = useMemo(
     () => ({
       files: [] as FileType[],
-      mentionedModels: [] as Model[],
+      mentionedModels: initialMentionedModels,
       selectedKnowledgeBases: initialAssistant.knowledge_bases ?? [],
       isExpanded: false,
       couldAddImageFile: false,
       extensions: [] as string[]
     }),
-    [initialAssistant.knowledge_bases]
+    [initialMentionedModels, initialAssistant.knowledge_bases]
   )
 
   return (
@@ -121,7 +134,10 @@ const InputbarInner: FC<InputbarInnerProps> = ({ assistant: initialAssistant, se
   const { setFiles, setMentionedModels, setSelectedKnowledgeBases } = useInputbarToolsDispatch()
   const { setCouldAddImageFile } = useInputbarToolsInternalDispatch()
 
-  const { text, setText } = useInputText()
+  const { text, setText } = useInputText({
+    initialValue: CacheService.get<string>(INPUTBAR_DRAFT_CACHE_KEY) ?? '',
+    onChange: (value) => CacheService.set(INPUTBAR_DRAFT_CACHE_KEY, value, DRAFT_CACHE_TTL)
+  })
   const {
     textareaRef,
     resize: resizeTextArea,
@@ -133,7 +149,6 @@ const InputbarInner: FC<InputbarInnerProps> = ({ assistant: initialAssistant, se
     minHeight: 30
   })
 
-  const showKnowledgeIcon = useSidebarIconShow('knowledge')
   const { assistant, addTopic, model, setModel, updateAssistant } = useAssistant(initialAssistant.id)
   const { sendMessageShortcut, showInputEstimatedTokens, enableQuickPanelTriggers } = useSettings()
   const [estimateTokenCount, setEstimateTokenCount] = useState(0)
@@ -190,6 +205,15 @@ const InputbarInner: FC<InputbarInnerProps> = ({ assistant: initialAssistant, se
     setCouldAddImageFile(canAddImageFile)
   }, [canAddImageFile, setCouldAddImageFile])
 
+  const onUnmount = useEffectEvent((id: string) => {
+    CacheService.set(getMentionedModelsCacheKey(id), mentionedModels, DRAFT_CACHE_TTL)
+  })
+
+  useEffect(() => {
+    return () => onUnmount(assistant.id)
+    // eslint-disable-next-line react-hooks/exhaustive-deps
+  }, [assistant.id])
+
   const placeholderText = enableQuickPanelTriggers
     ? t('chat.input.placeholder', { key: getSendMessageShortcutLabel(sendMessageShortcut) })
     : t('chat.input.placeholder_without_triggers', {
@@ -381,9 +405,10 @@ const InputbarInner: FC<InputbarInnerProps> = ({ assistant: initialAssistant, se
     focusTextarea
   ])
 
+  // TODO: Just use assistant.knowledge_bases as selectedKnowledgeBases. context state is overdesigned.
   useEffect(() => {
-    setSelectedKnowledgeBases(showKnowledgeIcon ? (assistant.knowledge_bases ?? []) : [])
-  }, [assistant.knowledge_bases, setSelectedKnowledgeBases, showKnowledgeIcon])
+    setSelectedKnowledgeBases(assistant.knowledge_bases ?? [])
+  }, [assistant.knowledge_bases, setSelectedKnowledgeBases])
 
   useEffect(() => {
     // Disable web search if model doesn't support it

src/renderer/src/pages/home/Inputbar/components/InputbarCore.tsx

@@ -156,11 +156,8 @@ export const InputbarCore: FC<InputbarCoreProps> = ({
 
   const setText = useCallback<React.Dispatch<React.SetStateAction<string>>>(
     (value) => {
-      if (typeof value === 'function') {
-        onTextChange(value(textRef.current))
-      } else {
-        onTextChange(value)
-      }
+      const newText = typeof value === 'function' ? value(textRef.current) : value
+      onTextChange(newText)
     },
     [onTextChange]
   )

src/renderer/src/pages/home/Inputbar/tools/knowledgeBaseTool.tsx

@@ -1,5 +1,4 @@
 import { useAssistant } from '@renderer/hooks/useAssistant'
-import { useSidebarIconShow } from '@renderer/hooks/useSidebarIcon'
 import { defineTool, registerTool, TopicType } from '@renderer/pages/home/Inputbar/types'
 import type { KnowledgeBase } from '@renderer/types'
 import { isPromptToolUse, isSupportedToolUse } from '@renderer/utils/mcp-tools'
@@ -30,7 +29,6 @@ const knowledgeBaseTool = defineTool({
   render: function KnowledgeBaseToolRender(context) {
     const { assistant, state, actions, quickPanel } = context
 
-    const knowledgeSidebarEnabled = useSidebarIconShow('knowledge')
     const { updateAssistant } = useAssistant(assistant.id)
 
     const handleSelect = useCallback(
@@ -41,10 +39,6 @@ const knowledgeBaseTool = defineTool({
       [updateAssistant, actions]
     )
 
-    if (!knowledgeSidebarEnabled) {
-      return null
-    }
-
     return (
       <KnowledgeBaseButton
         quickPanel={quickPanel}

src/renderer/src/pages/home/Messages/Blocks/ThinkingBlock.tsx

@@ -102,10 +102,12 @@ const ThinkingBlock: React.FC<Props> = ({ block }) => {
   )
 }
 
+const normalizeThinkingTime = (value?: number) => (typeof value === 'number' && Number.isFinite(value) ? value : 0)
+
 const ThinkingTimeSeconds = memo(
   ({ blockThinkingTime, isThinking }: { blockThinkingTime: number; isThinking: boolean }) => {
     const { t } = useTranslation()
-    const [displayTime, setDisplayTime] = useState(blockThinkingTime)
+    const [displayTime, setDisplayTime] = useState(normalizeThinkingTime(blockThinkingTime))
 
     const timer = useRef<NodeJS.Timeout | null>(null)
 
@@ -121,7 +123,7 @@ const ThinkingTimeSeconds = memo(
           clearInterval(timer.current)
           timer.current = null
         }
-        setDisplayTime(blockThinkingTime)
+        setDisplayTime(normalizeThinkingTime(blockThinkingTime))
       }
 
       return () => {
@@ -132,10 +134,10 @@ const ThinkingTimeSeconds = memo(
       }
     }, [isThinking, blockThinkingTime])
 
-    const thinkingTimeSeconds = useMemo(
-      () => ((displayTime < 1000 ? 100 : displayTime) / 1000).toFixed(1),
-      [displayTime]
-    )
+    const thinkingTimeSeconds = useMemo(() => {
+      const safeTime = normalizeThinkingTime(displayTime)
+      return ((safeTime < 1000 ? 100 : safeTime) / 1000).toFixed(1)
+    }, [displayTime])
 
     return isThinking
       ? t('chat.thinking', {

src/renderer/src/pages/home/Messages/Blocks/__tests__/ThinkingBlock.test.tsx

@@ -255,6 +255,20 @@ describe('ThinkingBlock', () => {
         unmount()
       })
     })
+
+    it('should clamp invalid thinking times to a safe default', () => {
+      const testCases = [undefined, Number.NaN, Number.POSITIVE_INFINITY]
+
+      testCases.forEach((thinking_millsec) => {
+        const block = createThinkingBlock({
+          thinking_millsec: thinking_millsec as any,
+          status: MessageBlockStatus.SUCCESS
+        })
+        const { unmount } = renderThinkingBlock(block)
+        expect(getThinkingTimeText()).toHaveTextContent('0.1s')
+        unmount()
+      })
+    })
   })
 
   describe('collapse behavior', () => {

src/renderer/src/pages/home/Messages/ChatNavigation.tsx

@@ -10,6 +10,7 @@ import { useSettings } from '@renderer/hooks/useSettings'
 import { useTimer } from '@renderer/hooks/useTimer'
 import type { RootState } from '@renderer/store'
 // import { selectCurrentTopicId } from '@renderer/store/newMessage'
+import { scrollIntoView } from '@renderer/utils/dom'
 import { Button, Drawer, Tooltip } from 'antd'
 import type { FC } from 'react'
 import { useCallback, useEffect, useRef, useState } from 'react'
@@ -118,7 +119,8 @@ const ChatNavigation: FC<ChatNavigationProps> = ({ containerId }) => {
   }
 
   const scrollToMessage = (element: HTMLElement) => {
-    element.scrollIntoView({ behavior: 'smooth', block: 'start' })
+    // Use container: 'nearest' to keep scroll within the chat pane (Chromium-only, see #11565, #11567)
+    scrollIntoView(element, { behavior: 'smooth', block: 'start', container: 'nearest' })
   }
 
   const scrollToTop = () => {

src/renderer/src/pages/home/Messages/Message.tsx

@@ -15,6 +15,7 @@ import { estimateMessageUsage } from '@renderer/services/TokenService'
 import type { Assistant, Topic } from '@renderer/types'
 import type { Message, MessageBlock } from '@renderer/types/newMessage'
 import { classNames, cn } from '@renderer/utils'
+import { scrollIntoView } from '@renderer/utils/dom'
 import { isMessageProcessing } from '@renderer/utils/messageUtils/is'
 import { Divider } from 'antd'
 import type { Dispatch, FC, SetStateAction } from 'react'
@@ -79,9 +80,10 @@ const MessageItem: FC<Props> = ({
 
   useEffect(() => {
     if (isEditing && messageContainerRef.current) {
-      messageContainerRef.current.scrollIntoView({
+      scrollIntoView(messageContainerRef.current, {
         behavior: 'smooth',
-        block: 'center'
+        block: 'center',
+        container: 'nearest'
       })
     }
   }, [isEditing])
@@ -124,7 +126,7 @@ const MessageItem: FC<Props> = ({
   const messageHighlightHandler = useCallback(
     (highlight: boolean = true) => {
       if (messageContainerRef.current) {
-        messageContainerRef.current.scrollIntoView({ behavior: 'smooth' })
+        scrollIntoView(messageContainerRef.current, { behavior: 'smooth', block: 'center', container: 'nearest' })
         if (highlight) {
           setTimeoutTimer(
             'messageHighlightHandler',

src/renderer/src/pages/home/Messages/MessageAnchorLine.tsx

@@ -12,6 +12,7 @@ import { newMessagesActions } from '@renderer/store/newMessage'
 // import { updateMessageThunk } from '@renderer/store/thunk/messageThunk'
 import type { Message } from '@renderer/types/newMessage'
 import { isEmoji, removeLeadingEmoji } from '@renderer/utils'
+import { scrollIntoView } from '@renderer/utils/dom'
 import { getMainTextContent } from '@renderer/utils/messageUtils/find'
 import { Avatar } from 'antd'
 import { CircleChevronDown } from 'lucide-react'
@@ -119,7 +120,7 @@ const MessageAnchorLine: FC<MessageLineProps> = ({ messages }) => {
           () => {
             const messageElement = document.getElementById(`message-${message.id}`)
             if (messageElement) {
-              messageElement.scrollIntoView({ behavior: 'auto', block: 'start' })
+              scrollIntoView(messageElement, { behavior: 'auto', block: 'start', container: 'nearest' })
             }
           },
           100
@@ -141,7 +142,7 @@ const MessageAnchorLine: FC<MessageLineProps> = ({ messages }) => {
         return
       }
 
-      messageElement.scrollIntoView({ behavior: 'smooth', block: 'start' })
+      scrollIntoView(messageElement, { behavior: 'smooth', block: 'start', container: 'nearest' })
     },
     [setSelectedMessage]
   )

src/renderer/src/pages/home/Messages/MessageGroup.tsx

@@ -10,6 +10,7 @@ import type { MultiModelMessageStyle } from '@renderer/store/settings'
 import type { Topic } from '@renderer/types'
 import type { Message } from '@renderer/types/newMessage'
 import { classNames } from '@renderer/utils'
+import { scrollIntoView } from '@renderer/utils/dom'
 import { Popover } from 'antd'
 import type { ComponentProps } from 'react'
 import { memo, useCallback, useEffect, useMemo, useState } from 'react'
@@ -73,7 +74,7 @@ const MessageGroup = ({ messages, topic, registerMessageElement }: Props) => {
         () => {
           const messageElement = document.getElementById(`message-${message.id}`)
           if (messageElement) {
-            messageElement.scrollIntoView({ behavior: 'smooth', block: 'start' })
+            scrollIntoView(messageElement, { behavior: 'smooth', block: 'start', container: 'nearest' })
           }
         },
         200
@@ -132,7 +133,7 @@ const MessageGroup = ({ messages, topic, registerMessageElement }: Props) => {
             setSelectedMessage(message)
           } else {
             // 直接滚动
-            element.scrollIntoView({ behavior: 'smooth', block: 'start' })
+            scrollIntoView(element, { behavior: 'smooth', block: 'start', container: 'nearest' })
           }
         }
       }

src/renderer/src/pages/home/Messages/MessageOutline.tsx

@@ -3,6 +3,7 @@ import type { RootState } from '@renderer/store'
 import { messageBlocksSelectors } from '@renderer/store/messageBlock'
 import type { Message } from '@renderer/types/newMessage'
 import { MessageBlockType } from '@renderer/types/newMessage'
+import { scrollIntoView } from '@renderer/utils/dom'
 import type { FC } from 'react'
 import React, { useMemo, useRef } from 'react'
 import { useSelector } from 'react-redux'
@@ -72,10 +73,10 @@ const MessageOutline: FC<MessageOutlineProps> = ({ message }) => {
     const parent = messageOutlineContainerRef.current?.parentElement
     const messageContentContainer = parent?.querySelector('.message-content-container')
     if (messageContentContainer) {
-      const headingElement = messageContentContainer.querySelector(`#${id}`)
+      const headingElement = messageContentContainer.querySelector<HTMLElement>(`#${id}`)
       if (headingElement) {
         const scrollBlock = ['horizontal', 'grid'].includes(message.multiModelMessageStyle ?? '') ? 'nearest' : 'start'
-        headingElement.scrollIntoView({ behavior: 'smooth', block: scrollBlock })
+        scrollIntoView(headingElement, { behavior: 'smooth', block: scrollBlock, container: 'nearest' })
       }
     }
   }

src/renderer/src/pages/home/Messages/Tools/MessageAgentTools/BashTool.tsx

@@ -5,8 +5,6 @@ import { Terminal } from 'lucide-react'
 import { ToolTitle } from './GenericTools'
 import type { BashToolInput as BashToolInputType, BashToolOutput as BashToolOutputType } from './types'
 
-const MAX_TAG_LENGTH = 100
-
 export function BashTool({
   input,
   output
@@ -17,12 +15,10 @@ export function BashTool({
   // 如果有输出，计算输出行数
   const outputLines = output ? output.split('\n').length : 0
 
-  // 处理命令字符串的截断，添加空值检查
+  // 处理命令字符串，添加空值检查
   const command = input?.command ?? ''
-  const needsTruncate = command.length > MAX_TAG_LENGTH
-  const displayCommand = needsTruncate ? `${command.slice(0, MAX_TAG_LENGTH)}...` : command
 
-  const tagContent = <Tag className="whitespace-pre-wrap break-all font-mono">{displayCommand}</Tag>
+  const tagContent = <Tag className="!m-0 max-w-full truncate font-mono">{command}</Tag>
 
   return {
     key: 'tool',
@@ -34,16 +30,12 @@ export function BashTool({
           params={input?.description}
           stats={output ? `${outputLines} ${outputLines === 1 ? 'line' : 'lines'}` : undefined}
         />
-        <div className="mt-1">
-          {needsTruncate ? (
-            <Popover
-              content={<div className="max-w-xl whitespace-pre-wrap break-all font-mono">{command}</div>}
-              trigger="hover">
-              {tagContent}
-            </Popover>
-          ) : (
-            tagContent
-          )}
+        <div className="mt-1 max-w-full">
+          <Popover
+            content={<div className="max-w-xl whitespace-pre-wrap break-all font-mono text-xs">{command}</div>}
+            trigger="hover">
+            {tagContent}
+          </Popover>
         </div>
       </>
     ),

src/renderer/src/pages/home/Messages/Tools/MessageAgentTools/GenericTools.tsx

@@ -18,9 +18,9 @@ export function ToolTitle({
 }) {
   return (
     <div className={`flex items-center gap-1 ${className}`}>
-      {icon}
-      {label && <span className="font-medium text-sm">{label}</span>}
-      {params && <span className="flex-shrink-0 text-muted-foreground text-xs">{params}</span>}
+      {icon && <span className="flex flex-shrink-0">{icon}</span>}
+      {label && <span className="flex-shrink-0 font-medium text-sm">{label}</span>}
+      {params && <span className="min-w-0 truncate text-muted-foreground text-xs">{params}</span>}
       {stats && <span className="flex-shrink-0 text-muted-foreground text-xs">{stats}</span>}
     </div>
   )

src/renderer/src/pages/home/Messages/Tools/MessageAgentTools/index.tsx

@@ -1,7 +1,10 @@
 import { loggerService } from '@logger'
+import { useAppSelector } from '@renderer/store'
+import { selectPendingPermission } from '@renderer/store/toolPermissions'
 import type { NormalToolResponse } from '@renderer/types'
 import type { CollapseProps } from 'antd'
-import { Collapse } from 'antd'
+import { Collapse, Spin } from 'antd'
+import { useTranslation } from 'react-i18next'
 
 // 导出所有类型
 export * from './types'
@@ -83,17 +86,41 @@ function ToolContent({ toolName, input, output }: { toolName: AgentToolsType; in
 // 统一的组件渲染入口
 export function MessageAgentTools({ toolResponse }: { toolResponse: NormalToolResponse }) {
   const { arguments: args, response, tool, status } = toolResponse
-  logger.info('Rendering agent tool response', {
+  logger.debug('Rendering agent tool response', {
     tool: tool,
     arguments: args,
+    status,
     response
   })
 
+  const pendingPermission = useAppSelector((state) =>
+    selectPendingPermission(state.toolPermissions, toolResponse.toolCallId)
+  )
+
   if (status === 'pending') {
-    return <ToolPermissionRequestCard toolResponse={toolResponse} />
+    if (pendingPermission) {
+      return <ToolPermissionRequestCard toolResponse={toolResponse} />
+    }
+    return <ToolPendingIndicator toolName={tool?.name} description={tool?.description} />
   }
 
   return (
     <ToolContent toolName={tool.name as AgentToolsType} input={args as ToolInput} output={response as ToolOutput} />
   )
 }
+
+function ToolPendingIndicator({ toolName, description }: { toolName?: string; description?: string }) {
+  const { t } = useTranslation()
+  const label = toolName || t('agent.toolPermission.toolPendingFallback', 'Tool')
+  const detail = description?.trim() || t('agent.toolPermission.executing')
+
+  return (
+    <div className="flex w-full max-w-xl items-center gap-3 rounded-xl border border-default-200 bg-default-100 px-4 py-3 shadow-sm">
+      <Spin size="small" />
+      <div className="flex flex-col gap-1">
+        <span className="font-semibold text-default-700 text-sm">{label}</span>
+        <span className="text-default-500 text-xs">{detail}</span>
+      </div>
+    </div>
+  )
+}

src/renderer/src/services/ApiService.ts

@@ -12,7 +12,7 @@ import type { FetchChatCompletionParams } from '@renderer/types'
 import type { Assistant, MCPServer, MCPTool, Model, Provider } from '@renderer/types'
 import type { StreamTextParams } from '@renderer/types/aiCoreTypes'
 import { type Chunk, ChunkType } from '@renderer/types/chunk'
-import type { Message } from '@renderer/types/newMessage'
+import type { Message, ResponseError } from '@renderer/types/newMessage'
 import type { SdkModel } from '@renderer/types/sdk'
 import { removeSpecialCharactersForTopicName, uuid } from '@renderer/utils'
 import { abortCompletion, readyToAbort } from '@renderer/utils/abortController'
@@ -476,7 +476,7 @@ export async function checkApi(provider: Provider, model: Model, timeout = 15000
   } else {
     const abortId = uuid()
     const signal = readyToAbort(abortId)
-    let chunkError
+    let streamError: ResponseError | undefined
     const params: StreamTextParams = {
       system: assistant.prompt,
       prompt: 'hi',
@@ -495,19 +495,18 @@ export async function checkApi(provider: Provider, model: Model, timeout = 15000
       callType: 'check',
       onChunk: (chunk: Chunk) => {
         if (chunk.type === ChunkType.ERROR) {
-          chunkError = chunk.error
+          streamError = chunk.error
         } else {
           abortCompletion(abortId)
         }
       }
     }
 
-    // Try streaming check
     try {
       await ai.completions(model.id, params, config)
     } catch (e) {
-      if (!isAbortError(e) && !isAbortError(chunkError)) {
-        throw e
+      if (!isAbortError(e) && !isAbortError(streamError)) {
+        throw streamError ?? e
       }
     }
   }

src/renderer/src/types/index.ts

@@ -239,6 +239,7 @@ export type ModelType = 'text' | 'vision' | 'embedding' | 'reasoning' | 'functio
 
 export type ModelTag = Exclude<ModelType, 'text'> | 'free'
 
+// "image-generation" is also openai endpoint, but specifically for image generation.
 export type EndpointType = 'openai' | 'openai-response' | 'anthropic' | 'gemini' | 'image-generation' | 'jina-rerank'
 
 export type ModelPricing = {

src/renderer/src/types/newMessage.ts

@@ -234,6 +234,7 @@ export interface Response {
   error?: ResponseError
 }
 
+// FIXME: Weak type safety. It may be a specific class instance which inherits Error in runtime.
 export type ResponseError = Record<string, any>
 
 export interface MessageInputBaseParams {

src/renderer/src/utils/dom.ts

@@ -1,3 +1,15 @@
+import { loggerService } from '@logger'
+
+const logger = loggerService.withContext('utils/dom')
+
+interface ChromiumScrollIntoViewOptions extends ScrollIntoViewOptions {
+  /**
+   * @see https://developer.mozilla.org/en-US/docs/Web/API/Element/scrollIntoView#container
+   * @see https://github.com/microsoft/TypeScript/issues/62803
+   */
+  container?: 'all' | 'nearest'
+}
+
 /**
  * Simple wrapper for scrollIntoView with common default options.
  * Provides a unified interface with sensible defaults.
@@ -5,7 +17,12 @@
  * @param element - The target element to scroll into view
  * @param options - Scroll options. If not provided, uses { behavior: 'smooth', block: 'center', inline: 'nearest' }
  */
-export function scrollIntoView(element: HTMLElement, options?: ScrollIntoViewOptions): void {
+export function scrollIntoView(element: HTMLElement, options?: ChromiumScrollIntoViewOptions): void {
+  if (!element) {
+    logger.warn('[scrollIntoView] Unexpected falsy element. Do nothing as fallback.')
+    return
+  }
+
   const defaultOptions: ScrollIntoViewOptions = {
     behavior: 'smooth',
     block: 'center',

src/renderer/src/windows/mini/home/HomeWindow.tsx

@@ -254,6 +254,17 @@ const HomeWindow: FC<{ draggable?: boolean }> = ({ draggable = true }) => {
 
         let blockId: string | null = null
         let thinkingBlockId: string | null = null
+        let thinkingStartTime: number | null = null
+
+        const resolveThinkingDuration = (duration?: number) => {
+          if (typeof duration === 'number' && Number.isFinite(duration)) {
+            return duration
+          }
+          if (thinkingStartTime !== null) {
+            return Math.max(0, performance.now() - thinkingStartTime)
+          }
+          return 0
+        }
 
         setIsLoading(true)
         setIsOutputted(false)
@@ -291,6 +302,7 @@ const HomeWindow: FC<{ draggable?: boolean }> = ({ draggable = true }) => {
               case ChunkType.THINKING_START:
                 {
                   setIsOutputted(true)
+                  thinkingStartTime = performance.now()
                   if (thinkingBlockId) {
                     store.dispatch(
                       updateOneBlock({ id: thinkingBlockId, changes: { status: MessageBlockStatus.STREAMING } })
@@ -315,9 +327,13 @@ const HomeWindow: FC<{ draggable?: boolean }> = ({ draggable = true }) => {
                 {
                   setIsOutputted(true)
                   if (thinkingBlockId) {
+                    if (thinkingStartTime === null) {
+                      thinkingStartTime = performance.now()
+                    }
+                    const thinkingDuration = resolveThinkingDuration(chunk.thinking_millsec)
                     throttledBlockUpdate(thinkingBlockId, {
                       content: chunk.text,
-                      thinking_millsec: chunk.thinking_millsec
+                      thinking_millsec: thinkingDuration
                     })
                   }
                 }
@@ -325,14 +341,17 @@ const HomeWindow: FC<{ draggable?: boolean }> = ({ draggable = true }) => {
               case ChunkType.THINKING_COMPLETE:
                 {
                   if (thinkingBlockId) {
+                    const thinkingDuration = resolveThinkingDuration(chunk.thinking_millsec)
                     cancelThrottledBlockUpdate(thinkingBlockId)
                     store.dispatch(
                       updateOneBlock({
                         id: thinkingBlockId,
-                        changes: { status: MessageBlockStatus.SUCCESS, thinking_millsec: chunk.thinking_millsec }
+                        changes: { status: MessageBlockStatus.SUCCESS, thinking_millsec: thinkingDuration }
                       })
                     )
                   }
+                  thinkingStartTime = null
+                  thinkingBlockId = null
                 }
                 break
               case ChunkType.TEXT_START:
@@ -404,6 +423,8 @@ const HomeWindow: FC<{ draggable?: boolean }> = ({ draggable = true }) => {
                 if (!isAborted) {
                   throw new Error(chunk.error.message)
                 }
+                thinkingStartTime = null
+                thinkingBlockId = null
               }
               //fall through
               case ChunkType.BLOCK_COMPLETE:

src/renderer/src/windows/selection/action/components/ActionUtils.ts

@@ -41,8 +41,19 @@ export const processMessages = async (
 
     let textBlockId: string | null = null
     let thinkingBlockId: string | null = null
+    let thinkingStartTime: number | null = null
     let textBlockContent: string = ''
 
+    const resolveThinkingDuration = (duration?: number) => {
+      if (typeof duration === 'number' && Number.isFinite(duration)) {
+        return duration
+      }
+      if (thinkingStartTime !== null) {
+        return Math.max(0, performance.now() - thinkingStartTime)
+      }
+      return 0
+    }
+
     const assistantMessage = getAssistantMessage({
       assistant,
       topic
@@ -79,6 +90,7 @@ export const processMessages = async (
         switch (chunk.type) {
           case ChunkType.THINKING_START:
             {
+              thinkingStartTime = performance.now()
               if (thinkingBlockId) {
                 store.dispatch(
                   updateOneBlock({ id: thinkingBlockId, changes: { status: MessageBlockStatus.STREAMING } })
@@ -102,9 +114,13 @@ export const processMessages = async (
           case ChunkType.THINKING_DELTA:
             {
               if (thinkingBlockId) {
+                if (thinkingStartTime === null) {
+                  thinkingStartTime = performance.now()
+                }
+                const thinkingDuration = resolveThinkingDuration(chunk.thinking_millsec)
                 throttledBlockUpdate(thinkingBlockId, {
                   content: chunk.text,
-                  thinking_millsec: chunk.thinking_millsec
+                  thinking_millsec: thinkingDuration
                 })
               }
               onStream()
@@ -113,6 +129,7 @@ export const processMessages = async (
           case ChunkType.THINKING_COMPLETE:
             {
               if (thinkingBlockId) {
+                const thinkingDuration = resolveThinkingDuration(chunk.thinking_millsec)
                 cancelThrottledBlockUpdate(thinkingBlockId)
                 store.dispatch(
                   updateOneBlock({
@@ -120,12 +137,13 @@ export const processMessages = async (
                     changes: {
                       content: chunk.text,
                       status: MessageBlockStatus.SUCCESS,
-                      thinking_millsec: chunk.thinking_millsec
+                      thinking_millsec: thinkingDuration
                     }
                   })
                 )
                 thinkingBlockId = null
               }
+              thinkingStartTime = null
             }
             break
           case ChunkType.TEXT_START:
@@ -190,6 +208,7 @@ export const processMessages = async (
           case ChunkType.ERROR:
             {
               const blockId = textBlockId || thinkingBlockId
+              thinkingStartTime = null
               if (blockId) {
                 store.dispatch(
                   updateOneBlock({

src/renderer/src/windows/selection/action/components/__tests__/ActionUtils.test.ts

@@ -284,6 +284,54 @@ describe('processMessages', () => {
     })
   })
 
+  describe('thinking timer fallback', () => {
+    it('should use local timer when thinking_millsec is missing', async () => {
+      const nowValues = [1000, 1500, 2000]
+      let nowIndex = 0
+      const performanceSpy = vi.spyOn(performance, 'now').mockImplementation(() => {
+        const value = nowValues[Math.min(nowIndex, nowValues.length - 1)]
+        nowIndex += 1
+        return value
+      })
+
+      const mockChunks = [
+        { type: ChunkType.THINKING_START },
+        { type: ChunkType.THINKING_DELTA, text: 'Thinking...' },
+        { type: ChunkType.THINKING_COMPLETE, text: 'Done thinking' },
+        { type: ChunkType.TEXT_START },
+        { type: ChunkType.TEXT_COMPLETE, text: 'Final answer' },
+        { type: ChunkType.BLOCK_COMPLETE }
+      ]
+
+      vi.mocked(fetchChatCompletion).mockImplementation(async ({ onChunkReceived }: any) => {
+        for (const chunk of mockChunks) {
+          await onChunkReceived(chunk)
+        }
+      })
+
+      await processMessages(
+        mockAssistant,
+        mockTopic,
+        'test prompt',
+        mockSetAskId,
+        mockOnStream,
+        mockOnFinish,
+        mockOnError
+      )
+
+      const thinkingDeltaCall = vi.mocked(throttledBlockUpdate).mock.calls.find(([id]) => id === 'thinking-block-1')
+      const deltaPayload = thinkingDeltaCall?.[1] as { thinking_millsec?: number } | undefined
+      expect(deltaPayload?.thinking_millsec).toBe(500)
+
+      const thinkingCompleteUpdate = vi
+        .mocked(updateOneBlock)
+        .mock.calls.find(([payload]) => (payload as any)?.changes?.thinking_millsec !== undefined)
+      expect((thinkingCompleteUpdate?.[0] as any)?.changes?.thinking_millsec).toBe(1000)
+
+      performanceSpy.mockRestore()
+    })
+  })
+
   describe('stream with exceptions', () => {
     it('should handle error chunks properly', async () => {
       const mockError = new Error('Stream processing error')