fix(streaming): scale the overall stream cap with the user request timeout

The desktop injects CLAUDE_STREAM_MAX_DURATION_MS=600000, a wall-clock cap
that no incoming chunk resets. Slow local models (LM Studio / qwen3) and very
large generations legitimately keep streaming thinking_delta past ten minutes,
so the stream was killed mid-response even though it was alive and producing
content (#1307, #1275). Raising "请求超时" could not help: the cap was
hardcoded, and the first-token budget it was meant to raise was silently
bounded by that same 600s.

Derive the cap from the user's request timeout, never below the existing 600s
default so the #766 trickle protection is not tightened when a short first-byte
budget is configured. Push the derived value through the per-turn env hot
update as well — otherwise a session that is already running keeps its spawn
time 600s no matter what the user configures, which is exactly the retry path
the reports describe.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

desktop/src/i18n/locales/en.ts

@@ -1787,7 +1787,7 @@ Row 9, all 8 cells: continuing from straight down, turning left through lower-le
   '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 30-1800 seconds; slow providers may need several minutes before the first streamed byte.',
+  '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. Raising it also extends the overall time limit for a single streamed response (never below 600 seconds).',
   'settings.general.networkTimeoutUnit': 'sec',
   'settings.general.networkTimeoutDecrease': 'Decrease by 30 seconds',
   'settings.general.networkTimeoutIncrease': 'Increase by 30 seconds',

desktop/src/i18n/locales/jp.ts

@@ -1789,7 +1789,7 @@ 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。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.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。値を大きくすると、1 回のストリーミング応答の総時間上限も広がります（最低 600 秒）。',
   'settings.general.networkTimeoutUnit': '秒',
   'settings.general.networkTimeoutDecrease': '30 秒減らす',
   'settings.general.networkTimeoutIncrease': '30 秒増やす',

desktop/src/i18n/locales/kr.ts

@@ -1789,7 +1789,7 @@ 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. 30~1800초를 지원합니다. 큰 컨텍스트에서는 첫 스트리밍 바이트까지 몇 분이 걸릴 수 있습니다.',
+  '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. 값을 늘리면 단일 스트리밍 응답의 총 시간 상한도 함께 늘어납니다(최소 600초).',
   'settings.general.networkTimeoutUnit': '초',
   'settings.general.networkTimeoutDecrease': '30초 줄이기',
   'settings.general.networkTimeoutIncrease': '30초 늘리기',

desktop/src/i18n/locales/zh-TW.ts

@@ -1788,7 +1788,7 @@ 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 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.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。調大後，單次流式回應的總時長上限同步放寬（不低於 600 秒）。',
   'settings.general.networkTimeoutUnit': '秒',
   'settings.general.networkTimeoutDecrease': '減少 30 秒',
   'settings.general.networkTimeoutIncrease': '增加 30 秒',

desktop/src/i18n/locales/zh.ts

@@ -1788,7 +1788,7 @@ 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 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.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。调大后，单次流式回复的总时长上限同步放宽（不低于 600 秒）。',
   'settings.general.networkTimeoutUnit': '秒',
   'settings.general.networkTimeoutDecrease': '减少 30 秒',
   'settings.general.networkTimeoutIncrease': '增加 30 秒',

src/server/__tests__/conversation-service.test.ts

@@ -184,11 +184,13 @@ describe('ConversationService', () => {
     sessionId: string,
     sent: string[],
     networkDerivedFirstTokenTimeout = true,
+    networkDerivedStreamMaxDuration = true,
   ) {
     const session = {
       outputCallbacks: [],
       networkRoutingFingerprint: '',
       networkDerivedFirstTokenTimeout,
+      networkDerivedStreamMaxDuration,
       sdkSocket: {
         send(line: string) {
           sent.push(line)
@@ -473,6 +475,51 @@ describe('ConversationService', () => {
     }
   })
 
+  test('buildChildEnv raises the overall stream cap with the user request timeout so an active long response is not killed (#1307)', async () => {
+    const prev = process.env.CLAUDE_STREAM_MAX_DURATION_MS
+    delete process.env.CLAUDE_STREAM_MAX_DURATION_MS
+    await fs.writeFile(
+      path.join(tmpDir, 'settings.json'),
+      JSON.stringify({ network: { aiRequestTimeoutMs: 1_800_000 } }),
+      'utf-8',
+    )
+    try {
+      const service = new ConversationService() as any
+      const env = (await service.buildChildEnv('/tmp')) as Record<string, string>
+
+      // The overall cap is NOT reset by incoming chunks, so a local model that
+      // keeps streaming thinking_delta events past it is killed mid-response.
+      // Raising "请求超时" must therefore extend the cap too — otherwise the
+      // user's timeout setting is silently capped at 600s (#1307).
+      expect(env.CLAUDE_STREAM_MAX_DURATION_MS).toBe('1800000')
+    } finally {
+      if (prev === undefined) delete process.env.CLAUDE_STREAM_MAX_DURATION_MS
+      else process.env.CLAUDE_STREAM_MAX_DURATION_MS = prev
+    }
+  })
+
+  test('buildChildEnv keeps the overall stream cap floor for a short request timeout (#766)', async () => {
+    const prev = process.env.CLAUDE_STREAM_MAX_DURATION_MS
+    delete process.env.CLAUDE_STREAM_MAX_DURATION_MS
+    await fs.writeFile(
+      path.join(tmpDir, 'settings.json'),
+      JSON.stringify({ network: { aiRequestTimeoutMs: 30_000 } }),
+      'utf-8',
+    )
+    try {
+      const service = new ConversationService() as any
+      const env = (await service.buildChildEnv('/tmp')) as Record<string, string>
+
+      // Shrinking the cap for a short first-byte budget would re-open #766:
+      // a stream that trickles content deltas just under the idle window must
+      // still be freed after a fixed overall duration.
+      expect(env.CLAUDE_STREAM_MAX_DURATION_MS).toBe('600000')
+    } finally {
+      if (prev === undefined) delete process.env.CLAUDE_STREAM_MAX_DURATION_MS
+      else process.env.CLAUDE_STREAM_MAX_DURATION_MS = prev
+    }
+  })
+
   test('buildChildEnv lets caller env override the first-token watchdog (#826)', async () => {
     const prev = process.env.CLAUDE_STREAM_FIRST_TOKEN_TIMEOUT_MS
     process.env.CLAUDE_STREAM_FIRST_TOKEN_TIMEOUT_MS = '900000'
@@ -777,10 +824,51 @@ describe('ConversationService', () => {
       all_proxy: 'http://127.0.0.1:17892',
       API_TIMEOUT_MS: '180000',
       CLAUDE_STREAM_FIRST_TOKEN_TIMEOUT_MS: '180000',
+      // The floor still holds on the hot-update path: lowering the request
+      // timeout must not shrink the overall cap below its 600s default.
+      CLAUDE_STREAM_MAX_DURATION_MS: '600000',
     })
     expect(JSON.parse(sent[1]!).type).toBe('user')
   })
 
+  test('sendMessage hot-applies a raised request timeout to the stream cap so a running session is not stuck at 600s (#1307)', async () => {
+    await fs.writeFile(
+      path.join(tmpDir, 'settings.json'),
+      JSON.stringify({
+        network: {
+          aiRequestTimeoutMs: 600_000,
+          proxy: { mode: 'direct', url: '' },
+        },
+      }),
+      'utf-8',
+    )
+    const service = new ConversationService() as any
+    const sent: string[] = []
+    const session = installNetworkTestSession(service, 'raised-timeout', sent)
+    await service.refreshNetworkEnvironmentBeforeTurn('raised-timeout', session)
+
+    await fs.writeFile(
+      path.join(tmpDir, 'settings.json'),
+      JSON.stringify({
+        network: {
+          aiRequestTimeoutMs: 1_800_000,
+          proxy: { mode: 'direct', url: '' },
+        },
+      }),
+      'utf-8',
+    )
+
+    expect(await service.sendMessage('raised-timeout', 'retry the long prompt')).toBe(true)
+    expect(sent).toHaveLength(2)
+    const update = JSON.parse(sent[0]!)
+    expect(update.type).toBe('update_environment_variables')
+    // The reported path is: user hits the 600s error, raises the timeout, and
+    // retries in the SAME conversation. The live CLI re-reads this per request,
+    // so it has to be pushed down — otherwise the retry dies at 600s again.
+    expect(update.variables.CLAUDE_STREAM_MAX_DURATION_MS).toBe('1800000')
+    expect(JSON.parse(sent[1]!).type).toBe('user')
+  })
+
   test('sendMessage does not resend network env when the system bridge fingerprint is unchanged', async () => {
     const originalBridgeUrl = process.env.CC_HAHA_SYSTEM_PROXY_URL
     process.env.CC_HAHA_SYSTEM_PROXY_URL = 'http://127.0.0.1:17893'

src/server/services/conversationService.ts

@@ -58,6 +58,7 @@ import { attributionHeaderEnvForModel } from './attributionHeaderPolicy.js'
 import {
   buildNetworkEnvironment,
   loadNetworkSettings,
+  resolveStreamMaxDurationMs,
   SYSTEM_PROXY_URL_ENV,
   type NetworkSettings,
 } from './networkSettings.js'
@@ -200,6 +201,7 @@ type SessionProcess = {
   permissionMode: string
   networkRoutingFingerprint: string
   networkDerivedFirstTokenTimeout: boolean
+  networkDerivedStreamMaxDuration: boolean
   sdkToken: string
   sdkSocket: { send(data: string): void } | null
   sdkAttached: Promise<void>
@@ -426,7 +428,10 @@ export class ConversationService {
     // chdir 后落到正确目录。
     //
     const networkSettings = await loadNetworkSettings()
-    const networkRuntimeMetadata = { firstTokenTimeoutDerived: false }
+    const networkRuntimeMetadata = {
+      firstTokenTimeoutDerived: false,
+      streamMaxDurationDerived: false,
+    }
     const childEnv = await this.buildChildEnv(
       launchWorkDir,
       sdkUrl,
@@ -472,6 +477,7 @@ export class ConversationService {
       permissionMode: options?.permissionMode || 'default',
       networkRoutingFingerprint: networkRoutingFingerprint(networkSettings, childEnv),
       networkDerivedFirstTokenTimeout: networkRuntimeMetadata.firstTokenTimeoutDerived,
+      networkDerivedStreamMaxDuration: networkRuntimeMetadata.streamMaxDurationDerived,
       sdkToken: this.getSdkTokenFromUrl(sdkUrl),
       sdkSocket: null,
       seenSdkMessageUuids: new Set<string>(),
@@ -662,6 +668,14 @@ export class ConversationService {
     if (session.networkDerivedFirstTokenTimeout) {
       variables.CLAUDE_STREAM_FIRST_TOKEN_TIMEOUT_MS = networkEnv.API_TIMEOUT_MS
     }
+    // The CLI re-reads this per request, so the fix for #1307 must reach a
+    // session that is already running — otherwise the user sees the 600s error,
+    // raises the timeout, retries in the same conversation and hits it again.
+    if (session.networkDerivedStreamMaxDuration) {
+      variables.CLAUDE_STREAM_MAX_DURATION_MS = String(
+        resolveStreamMaxDurationMs(networkEnv.API_TIMEOUT_MS),
+      )
+    }
 
     const sent = this.sendSdkMessage(sessionId, {
       type: 'update_environment_variables',
@@ -1522,7 +1536,10 @@ export class ConversationService {
     sdkUrl?: string,
     options?: SessionStartOptions,
     networkSettingsOverride?: NetworkSettings,
-    networkRuntimeMetadata?: { firstTokenTimeoutDerived: boolean },
+    networkRuntimeMetadata?: {
+      firstTokenTimeoutDerived: boolean
+      streamMaxDurationDerived: boolean
+    },
   ): Promise<Record<string, string>> {
     // Provider isolation: when Desktop has its own provider config/index,
     // strip inherited provider env vars so the child CLI reads fresh values
@@ -1568,6 +1585,8 @@ export class ConversationService {
     if (networkRuntimeMetadata) {
       networkRuntimeMetadata.firstTokenTimeoutDerived =
         !cleanEnv.CLAUDE_STREAM_FIRST_TOKEN_TIMEOUT_MS
+      networkRuntimeMetadata.streamMaxDurationDerived =
+        !cleanEnv.CLAUDE_STREAM_MAX_DURATION_MS
     }
     delete cleanEnv.CLAUDE_CODE_OAUTH_TOKEN
     if (options?.resumeInterruptedTurn === false) {
@@ -1603,6 +1622,14 @@ export class ConversationService {
       networkSettingsOverride ?? await loadNetworkSettings(),
       cleanEnv,
     )
+    // The overall-duration cap has to scale with the user's "请求超时" or raising
+    // that setting can never extend a long response — the cap is a wall-clock
+    // budget that no chunk resets, so a slow local model that legitimately
+    // thinks past it is killed mid-stream (#1307). The floor keeps the existing
+    // 600s protection from being TIGHTENED when the user configures a short
+    // first-byte budget (a lower cap would kill legitimate long responses even
+    // earlier); the per-turn hot update below mirrors this for live turns.
+    const streamMaxDurationMs = resolveStreamMaxDurationMs(networkEnv.API_TIMEOUT_MS)
     const traceCaptureEnabled = (await readTraceCaptureSettings()).enabled
     if (explicitProviderEnv && options?.model?.trim()) {
       explicitProviderEnv.ANTHROPIC_MODEL = options.model.trim()
@@ -1639,7 +1666,12 @@ export class ConversationService {
       // just under 240s apart keeps it alive forever and the request hangs with
       // no completion (#766: "卡住" with slowly growing tokens). This independent
       // cap frees such a stream after a fixed duration regardless of trickle.
-      CLAUDE_STREAM_MAX_DURATION_MS: cleanEnv.CLAUDE_STREAM_MAX_DURATION_MS || '600000',
+      // It tracks the user's "请求超时" (never below MIN_STREAM_MAX_DURATION_MS)
+      // so that raising the timeout also extends legitimately long responses,
+      // and a provider preset's own CLAUDE_STREAM_MAX_DURATION_MS still wins
+      // via the explicitProviderEnv spread below (#1307).
+      CLAUDE_STREAM_MAX_DURATION_MS:
+        cleanEnv.CLAUDE_STREAM_MAX_DURATION_MS || String(streamMaxDurationMs),
       // Abort a local tool call when its JSON arguments stop making progress.
       // Healthy input_json_delta events reset this budget; the independent full
       // response cap above still bounds a stream that trickles forever.

src/server/services/networkSettings.ts

@@ -27,6 +27,21 @@ export type NetworkSettings = {
 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
+// Floor for the CLI's overall stream-duration cap (CLAUDE_STREAM_MAX_DURATION_MS).
+// That cap is what frees an endlessly-trickling provider stream (#766), but it is
+// a wall-clock budget that no incoming chunk resets. It must therefore never be
+// LOWER than the user's own "请求超时": a local model that legitimately spends
+// longer than that thinking or generating would otherwise be killed mid-response
+// no matter how far the user raises the timeout (#1307). The floor keeps the
+// #766 protection intact when the user configures a very short first-byte budget.
+export const MIN_STREAM_MAX_DURATION_MS = 600_000
+// Shared by the spawn-time child env and the per-turn hot update so the two
+// cannot drift apart.
+export function resolveStreamMaxDurationMs(
+  apiTimeoutMs: string | number | undefined,
+): number {
+  return Math.max(MIN_STREAM_MAX_DURATION_MS, Number(apiTimeoutMs) || 0)
+}
 export const SYSTEM_PROXY_URL_ENV = 'CC_HAHA_SYSTEM_PROXY_URL'
 export const SYSTEM_PROXY_ERROR_ENV = 'CC_HAHA_SYSTEM_PROXY_ERROR'