diff --git a/desktop/src/__tests__/generalSettings.test.tsx b/desktop/src/__tests__/generalSettings.test.tsx
index 9f1c0cd8..57e0b981 100644
--- a/desktop/src/__tests__/generalSettings.test.tsx
+++ b/desktop/src/__tests__/generalSettings.test.tsx
@@ -468,8 +468,8 @@ describe('Settings > General tab', () => {
     const timeoutInput = screen.getByLabelText('AI request timeout')
     const saveButton = screen.getAllByRole('button', { name: 'Save' })[0]!
 
-    fireEvent.change(timeoutInput, { target: { value: '700' } })
-    expect(screen.getByText('Enter a whole number from 5 to 600 seconds.')).toBeInTheDocument()
+    fireEvent.change(timeoutInput, { target: { value: '2000' } })
+    expect(screen.getByText('Enter a whole number from 30 to 1800 seconds.')).toBeInTheDocument()
     expect(saveButton).toBeDisabled()
 
     fireEvent.change(timeoutInput, { target: { value: '90' } })
diff --git a/desktop/src/i18n/locales/en.ts b/desktop/src/i18n/locales/en.ts
index 5ee633d2..528681f9 100644
--- a/desktop/src/i18n/locales/en.ts
+++ b/desktop/src/i18n/locales/en.ts
@@ -1074,7 +1074,7 @@ export const en = {
   'settings.general.networkProxyUrlRequired': 'Enter a proxy URL.',
   'settings.general.networkTimeout': 'AI request timeout',
   'settings.general.networkTimeoutValue': '{seconds}s',
-  'settings.general.networkTimeoutHint': 'Applies to provider requests, streaming first responses, and provider connection tests. Supports 5-600 seconds.',
+  'settings.general.networkTimeoutHint': 'Applies to provider requests, streaming first responses, and provider connection tests. Supports 30-1800 seconds; slow providers may need several minutes before the first streamed byte.',
   'settings.general.networkTimeoutUnit': 'sec',
   'settings.general.networkTimeoutDecrease': 'Decrease by 30 seconds',
   'settings.general.networkTimeoutIncrease': 'Increase by 30 seconds',
diff --git a/desktop/src/i18n/locales/jp.ts b/desktop/src/i18n/locales/jp.ts
index dcf51b5e..a4507e7f 100644
--- a/desktop/src/i18n/locales/jp.ts
+++ b/desktop/src/i18n/locales/jp.ts
@@ -1076,7 +1076,7 @@ export const jp: Record<TranslationKey, string> = {
   'settings.general.networkProxyUrlRequired': 'プロキシ URL を入力してください。',
   'settings.general.networkTimeout': 'AI リクエストのタイムアウト',
   'settings.general.networkTimeoutValue': '{seconds}秒',
-  'settings.general.networkTimeoutHint': 'プロバイダーへのリクエスト、ストリーミングの最初の応答、プロバイダー接続テストに適用されます。5〜600 秒に対応します。',
+  'settings.general.networkTimeoutHint': 'プロバイダーへのリクエスト、ストリーミングの最初の応答、プロバイダー接続テストに適用されます。30〜1800 秒に対応します。大きなコンテキストでは最初のストリーミングバイトまで数分かかる場合があります。',
   'settings.general.networkTimeoutUnit': '秒',
   'settings.general.networkTimeoutDecrease': '30 秒減らす',
   'settings.general.networkTimeoutIncrease': '30 秒増やす',
diff --git a/desktop/src/i18n/locales/kr.ts b/desktop/src/i18n/locales/kr.ts
index 9ace308b..4fd1db08 100644
--- a/desktop/src/i18n/locales/kr.ts
+++ b/desktop/src/i18n/locales/kr.ts
@@ -1076,7 +1076,7 @@ export const kr: Record<TranslationKey, string> = {
   'settings.general.networkProxyUrlRequired': '프록시 URL을 입력하세요.',
   'settings.general.networkTimeout': 'AI 요청 시간 초과',
   'settings.general.networkTimeoutValue': '{seconds}초',
-  'settings.general.networkTimeoutHint': '공급자 요청, 스트리밍 첫 응답, 공급자 연결 테스트에 적용됩니다. 5~600초를 지원합니다.',
+  'settings.general.networkTimeoutHint': '공급자 요청, 스트리밍 첫 응답, 공급자 연결 테스트에 적용됩니다. 30~1800초를 지원합니다. 큰 컨텍스트에서는 첫 스트리밍 바이트까지 몇 분이 걸릴 수 있습니다.',
   'settings.general.networkTimeoutUnit': '초',
   'settings.general.networkTimeoutDecrease': '30초 줄이기',
   'settings.general.networkTimeoutIncrease': '30초 늘리기',
diff --git a/desktop/src/i18n/locales/zh-TW.ts b/desktop/src/i18n/locales/zh-TW.ts
index c69a3249..3c39aa7c 100644
--- a/desktop/src/i18n/locales/zh-TW.ts
+++ b/desktop/src/i18n/locales/zh-TW.ts
@@ -1076,7 +1076,7 @@ export const zh: Record<TranslationKey, string> = {
   'settings.general.networkProxyUrlRequired': '請輸入代理地址。',
   'settings.general.networkTimeout': 'AI 請求超時',
   'settings.general.networkTimeoutValue': '{seconds} 秒',
-  'settings.general.networkTimeoutHint': '用於服務商請求、流式首個響應，以及服務商連線測試。支援 5-600 秒。',
+  'settings.general.networkTimeoutHint': '用於服務商請求、流式首個響應，以及服務商連線測試。支援 30-1800 秒；部分服務商在大上下文下首個流式位元組可能需要等待數分鐘。',
   'settings.general.networkTimeoutUnit': '秒',
   'settings.general.networkTimeoutDecrease': '減少 30 秒',
   'settings.general.networkTimeoutIncrease': '增加 30 秒',
diff --git a/desktop/src/i18n/locales/zh.ts b/desktop/src/i18n/locales/zh.ts
index 7c714b98..396ba4d5 100644
--- a/desktop/src/i18n/locales/zh.ts
+++ b/desktop/src/i18n/locales/zh.ts
@@ -1076,7 +1076,7 @@ export const zh: Record<TranslationKey, string> = {
   'settings.general.networkProxyUrlRequired': '请输入代理地址。',
   'settings.general.networkTimeout': 'AI 请求超时',
   'settings.general.networkTimeoutValue': '{seconds} 秒',
-  'settings.general.networkTimeoutHint': '用于服务商请求、流式首个响应，以及服务商连接测试。支持 5-600 秒。',
+  'settings.general.networkTimeoutHint': '用于服务商请求、流式首个响应，以及服务商连接测试。支持 30-1800 秒；部分服务商在大上下文下首个流式字节可能需要等待数分钟。',
   'settings.general.networkTimeoutUnit': '秒',
   'settings.general.networkTimeoutDecrease': '减少 30 秒',
   'settings.general.networkTimeoutIncrease': '增加 30 秒',
diff --git a/desktop/src/pages/Settings.tsx b/desktop/src/pages/Settings.tsx
index 2e44355d..6d806406 100644
--- a/desktop/src/pages/Settings.tsx
+++ b/desktop/src/pages/Settings.tsx
@@ -57,8 +57,8 @@ import {
 } from '../lib/providerSettingsJson'
 import { copyTextToClipboard } from '../components/chat/clipboard'
 
-const NETWORK_TIMEOUT_MIN_SECONDS = 5
-const NETWORK_TIMEOUT_MAX_SECONDS = 600
+const NETWORK_TIMEOUT_MIN_SECONDS = 30
+const NETWORK_TIMEOUT_MAX_SECONDS = 1800
 const NETWORK_TIMEOUT_STEP_SECONDS = 30
 const SETTINGS_CHECKBOX_INPUT_CLASS = 'settings-checkbox-input peer'
 
diff --git a/desktop/src/stores/settingsStore.test.ts b/desktop/src/stores/settingsStore.test.ts
index 670b7a15..2e156f23 100644
--- a/desktop/src/stores/settingsStore.test.ts
+++ b/desktop/src/stores/settingsStore.test.ts
@@ -174,7 +174,7 @@ describe('settingsStore network persistence', () => {
     window.localStorage.clear()
   })
 
-  it('defaults old user settings to 120s system network settings', async () => {
+  it('defaults old user settings to 600s system network settings', async () => {
     vi.doMock('../api/settings', () => ({
       settingsApi: {
         getUser: vi.fn().mockResolvedValue({}),
@@ -215,7 +215,7 @@ describe('settingsStore network persistence', () => {
     await useSettingsStore.getState().fetchAll()
 
     expect(useSettingsStore.getState().network).toEqual({
-      aiRequestTimeoutMs: 120_000,
+      aiRequestTimeoutMs: 600_000,
       proxy: {
         mode: 'system',
         url: '',
@@ -256,7 +256,7 @@ describe('settingsStore network persistence', () => {
     const { useSettingsStore } = await import('./settingsStore')
 
     await useSettingsStore.getState().setNetwork({
-      aiRequestTimeoutMs: 999_999,
+      aiRequestTimeoutMs: 9_999_999,
       proxy: {
         mode: 'manual',
         url: '  http://127.0.0.1:7890  ',
@@ -264,7 +264,7 @@ describe('settingsStore network persistence', () => {
     })
 
     expect(useSettingsStore.getState().network).toEqual({
-      aiRequestTimeoutMs: 600_000,
+      aiRequestTimeoutMs: 1_800_000,
       proxy: {
         mode: 'manual',
         url: 'http://127.0.0.1:7890',
@@ -272,7 +272,7 @@ describe('settingsStore network persistence', () => {
     })
     expect(updateUser).toHaveBeenCalledWith({
       network: {
-        aiRequestTimeoutMs: 600_000,
+        aiRequestTimeoutMs: 1_800_000,
         proxy: {
           mode: 'manual',
           url: 'http://127.0.0.1:7890',
diff --git a/desktop/src/stores/settingsStore.ts b/desktop/src/stores/settingsStore.ts
index cd33fdcf..46d4b4ba 100644
--- a/desktop/src/stores/settingsStore.ts
+++ b/desktop/src/stores/settingsStore.ts
@@ -149,7 +149,7 @@ const DEFAULT_UPDATE_PROXY_SETTINGS: UpdateProxySettings = {
 }
 
 const DEFAULT_NETWORK_SETTINGS: NetworkSettings = {
-  aiRequestTimeoutMs: 120_000,
+  aiRequestTimeoutMs: 600_000,
   proxy: {
     mode: 'system',
     url: '',
@@ -655,7 +655,7 @@ function normalizeNetworkSettings(
   settings: NetworkSettingsInput | undefined,
 ): NetworkSettings {
   const timeout = typeof settings?.aiRequestTimeoutMs === 'number' && Number.isFinite(settings.aiRequestTimeoutMs)
-    ? Math.min(Math.max(Math.round(settings.aiRequestTimeoutMs), 5_000), 600_000)
+    ? Math.min(Math.max(Math.round(settings.aiRequestTimeoutMs), 30_000), 1_800_000)
     : DEFAULT_NETWORK_SETTINGS.aiRequestTimeoutMs
   const proxyMode = settings?.proxy?.mode === 'manual' ? 'manual' : 'system'
 
diff --git a/src/server/__tests__/conversation-service.test.ts b/src/server/__tests__/conversation-service.test.ts
index 67ce5f7d..6f23aa08 100644
--- a/src/server/__tests__/conversation-service.test.ts
+++ b/src/server/__tests__/conversation-service.test.ts
@@ -727,6 +727,52 @@ describe('ConversationService', () => {
     expect(env.CLAUDE_ENABLE_STREAM_WATCHDOG).toBe('1')
   })
 
+  test('buildChildEnv widens the stream idle window and disables the non-streaming fallback (#766)', async () => {
+    const service = new ConversationService() as any
+    const env = (await service.buildChildEnv(
+      '/tmp',
+      'ws://127.0.0.1:3456/sdk/test-session?token=test-token',
+    )) as Record<string, string>
+
+    // 90s default kills healthy-but-silent third-party streams; 240s keeps the
+    // watchdog useful without aborting slow thinking/prefill phases.
+    expect(env.CLAUDE_STREAM_IDLE_TIMEOUT_MS).toBe('240000')
+    // Non-streaming fallback can never finish for slow providers (first byte
+    // only arrives after FULL generation), so retries must stay streaming.
+    expect(env.CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK).toBe('1')
+  })
+
+  test('buildChildEnv respects caller overrides for stream timeout tuning envs', async () => {
+    const service = new ConversationService() as any
+    const previous = {
+      watchdog: process.env.CLAUDE_ENABLE_STREAM_WATCHDOG,
+      idle: process.env.CLAUDE_STREAM_IDLE_TIMEOUT_MS,
+      fallback: process.env.CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK,
+    }
+    process.env.CLAUDE_ENABLE_STREAM_WATCHDOG = '0'
+    process.env.CLAUDE_STREAM_IDLE_TIMEOUT_MS = '90000'
+    process.env.CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK = '0'
+    try {
+      const env = (await service.buildChildEnv(
+        '/tmp',
+        'ws://127.0.0.1:3456/sdk/test-session?token=test-token',
+      )) as Record<string, string>
+
+      expect(env.CLAUDE_ENABLE_STREAM_WATCHDOG).toBe('0')
+      expect(env.CLAUDE_STREAM_IDLE_TIMEOUT_MS).toBe('90000')
+      expect(env.CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK).toBe('0')
+    } finally {
+      for (const [key, value] of [
+        ['CLAUDE_ENABLE_STREAM_WATCHDOG', previous.watchdog],
+        ['CLAUDE_STREAM_IDLE_TIMEOUT_MS', previous.idle],
+        ['CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK', previous.fallback],
+      ] as const) {
+        if (value === undefined) delete process.env[key]
+        else process.env[key] = value
+      }
+    }
+  })
+
   test('buildSessionCliArgs forwards the selected runtime model and effort to the CLI process', () => {
     const service = new ConversationService() as any
     const args = service.buildSessionCliArgs(
diff --git a/src/server/__tests__/network-settings.test.ts b/src/server/__tests__/network-settings.test.ts
index 69dad9e0..e8618fab 100644
--- a/src/server/__tests__/network-settings.test.ts
+++ b/src/server/__tests__/network-settings.test.ts
@@ -37,7 +37,7 @@ describe('network settings', () => {
   beforeEach(setup)
   afterEach(teardown)
 
-  it('normalizes missing settings to the 120s system-proxy default', () => {
+  it('normalizes missing settings to the 600s system-proxy default', () => {
     expect(normalizeNetworkSettings({})).toEqual({
       aiRequestTimeoutMs: DEFAULT_AI_REQUEST_TIMEOUT_MS,
       proxy: {
@@ -50,7 +50,7 @@ describe('network settings', () => {
   it('clamps AI request timeouts and trims manual proxy URLs', () => {
     expect(normalizeNetworkSettings({
       network: {
-        aiRequestTimeoutMs: 999_999,
+        aiRequestTimeoutMs: 9_999_999,
         proxy: {
           mode: 'manual',
           url: '  http://127.0.0.1:7890  ',
diff --git a/src/server/services/conversationService.ts b/src/server/services/conversationService.ts
index 0945f2bf..de1a54fe 100644
--- a/src/server/services/conversationService.ts
+++ b/src/server/services/conversationService.ts
@@ -1109,6 +1109,18 @@ export class ConversationService {
       CLAUDE_CODE_ENABLE_SDK_FILE_CHECKPOINTING: '1',
       // Desktop must fail stuck provider streams instead of leaving the UI running forever.
       CLAUDE_ENABLE_STREAM_WATCHDOG: cleanEnv.CLAUDE_ENABLE_STREAM_WATCHDOG || '1',
+      // Third-party providers can stay silent for minutes mid-stream (thinking
+      // phases emit no SSE bytes, and many gateways never send pings), so the
+      // CLI's 90s idle default kills healthy streams (#766). 240s still frees
+      // a truly dead connection without shooting slow ones.
+      CLAUDE_STREAM_IDLE_TIMEOUT_MS: cleanEnv.CLAUDE_STREAM_IDLE_TIMEOUT_MS || '240000',
+      // When a stream does get aborted, retry as streaming instead of falling
+      // back to non-streaming: a non-streaming request must wait for the FULL
+      // generation before the first response byte, so slow providers can never
+      // finish inside API_TIMEOUT_MS — the fallback loops 5-minute aborts
+      // forever while the UI shows "running" (#766). It can also double-run
+      // tools (upstream inc-4258).
+      CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK: cleanEnv.CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK || '1',
       CLAUDE_CODE_DIAGNOSTICS_FILE: cliDiagnosticsPath,
       CLAUDE_COWORK_MEMORY_PATH_OVERRIDE: this.resolveDesktopAutoMemoryPath(workDir),
       CALLER_DIR: workDir,
diff --git a/src/server/services/networkSettings.ts b/src/server/services/networkSettings.ts
index b97f70c2..33d653e4 100644
--- a/src/server/services/networkSettings.ts
+++ b/src/server/services/networkSettings.ts
@@ -10,9 +10,15 @@ export type NetworkSettings = {
   }
 }
 
-export const DEFAULT_AI_REQUEST_TIMEOUT_MS = 120_000
-export const MIN_AI_REQUEST_TIMEOUT_MS = 5_000
-export const MAX_AI_REQUEST_TIMEOUT_MS = 600_000
+// API_TIMEOUT_MS is the CLI's *time-to-first-byte* budget for streaming
+// requests: third-party Anthropic-compatible gateways (sensenova, bailian,
+// zhipu, ...) often send no response bytes — not even headers or an SSE ping —
+// until prefill finishes, which can take minutes at large contexts. A short
+// default here aborts those healthy-but-slow requests dead at exactly this
+// timeout (#766), so the default matches the SDK's own 600s.
+export const DEFAULT_AI_REQUEST_TIMEOUT_MS = 600_000
+export const MIN_AI_REQUEST_TIMEOUT_MS = 30_000
+export const MAX_AI_REQUEST_TIMEOUT_MS = 1_800_000
 
 const DEFAULT_NETWORK_SETTINGS: NetworkSettings = {
   aiRequestTimeoutMs: DEFAULT_AI_REQUEST_TIMEOUT_MS,