新增: 待发箱抽屉与对话体验前端批

- 待发箱抽屉: ChatInput ⏰ 入口 + OutboxDrawer + useAiOutbox IPC 封装
- U2 进度语义三层: 内容层(最近动作) + 静默三档降速脉冲, 永不自动判死
- 流式心跳自愈: 竞态窗口旧任务迟到完成事件误收敛的自愈判定(949 前端兜底)
- 953 截断提示: 末条气泡渲染截断标记 + 继续输出按钮, 事件契约透传

src/App.vue

@@ -136,10 +136,32 @@ async function handleCancelApproval() {
   showApprovalDialog.value = false
 }
 
-// 点击菜单自动退出最大化 → 切为分栏模式
+// ── 路由 ↔ 最大化态联动(P0 修复 + P1 返回记忆)──
+// P0:统一走 setMaximized(落盘),不再直写 state 绕过 persistUiState
+//     (原内存/磁盘分歧:退出最大化未落盘,重启后读到旧值 true 闪回)
+// P1:离开 AI 域时记忆全宽态,回到 AI 域自动恢复 —— 解决「最大化态点设置/任务被迫退出,
+//     回来无法自动恢复全宽」的心流断裂(用户不必手动再点最大化)
+const AI_DOMAIN_PATHS = new Set(['/', '/ai-home', '/ai-detached'])
+const restoreMaximizedOnReturn = ref(false)
+
 watch(() => route.path, (path) => {
-  if (!['/', '/ai-home', '/ai-detached'].includes(path) && aiStore.state.maximized) {
-    aiStore.state.maximized = false
+  // 分离窗口不参与主面板最大化联动:分离窗是独立 JS context,state 与主窗无关;
+  // 且其 path(/ai-detached、/file-explorer-detached、/approval-popup)不在 AI 域内,
+  // 若不守卫会因分离窗默认 maximized=true 误触发 setMaximized(false) 并落盘,
+  // 覆盖主窗口的 df-ai-ui 布局
+  if (isDetached.value) return
+  if (AI_DOMAIN_PATHS.has(path)) {
+    // 回到 AI 域:若离开前是全宽则恢复(面板已关闭则不恢复,避免 <main> 被 v-show 隐藏致其他页不可交互)
+    if (restoreMaximizedOnReturn.value) {
+      restoreMaximizedOnReturn.value = false
+      if (aiStore.state.panelOpen) aiStore.setMaximized(true)
+    }
+    return
+  }
+  // 离开 AI 域:记忆全宽态后退出最大化(<main> 恢复可交互)
+  if (aiStore.state.maximized) {
+    restoreMaximizedOnReturn.value = true
+    aiStore.setMaximized(false)
   }
 })
 

src/api/types.ts

@@ -463,6 +463,12 @@ export type AiChatEvent = ({
    * undefined/false=真实值。前端 MessageList 据此标注『估算』区分 estimated vs real。
    */
   is_estimated?: boolean
+  /**
+   * 看板 953: 最终回复是否因 provider max_tokens 上限被截断(finish_reason=length)。
+   * true=正常收敛但内容被腰斩,MessageList 末条 assistant 气泡渲染截断提示条 + 继续输出按钮;
+   * undefined/false=完整(旧后端事件无此字段,按 false 兜底,向后兼容)。
+   */
+  truncated_by_max_tokens?: boolean
   /**
    * incomplete 场景细分(达上限自动收尾 vs 网络中断,前端据此区分提示文案):
    * "max_rounds"=达最大轮次后端自动收尾;其他/缺省=网络中断等异常。
@@ -647,6 +653,13 @@ export interface AiMessage {
   prompt_cache_miss_tokens?: number
   /** 思考 token(deepseek-reasoner/o1 reasoning_tokens)。老消息 undefined。 */
   reasoning_tokens?: number
+  /**
+   * 看板 953: 该 assistant 回复因 provider max_tokens 上限被截断(live 态 AiCompleted 事件透传标记)。
+   * 末条 assistant 带此标记时 MessageList 渲染截断提示条 + 继续输出按钮;用户再发消息/
+   * 继续输出后新气泡顶掉末条位置,标记自然失效,无需清除逻辑。仅 live 内存标记不入库
+   * (reload 后历史消息不带,截断提示属当轮即时性信息,可接受口径)。
+   */
+  truncatedByMaxTokens?: boolean
   timestamp: number
 }
 

src/components/AiChat.vue

@@ -81,7 +81,7 @@
             <span class="ai-queue-title">{{ $t('aiChat.queueTitle', { n: activeQueue.length }) }}</span>
             <div class="ai-queue-actions">
               <button v-if="queueTimedOut" class="ai-queue-force" @click="handleForceSend">{{ $t('ai.forceSendBtn') }}</button>
-              <button class="ai-queue-clear" @click="store.clearQueue(store.state.activeConversationId)">{{ $t('aiChat.clearQueue') }}</button>
+              <button class="ai-queue-clear" @click="handleClearQueue">{{ $t('aiChat.clearQueue') }}</button>
             </div>
           </div>
           <div class="ai-queue-list">
@@ -101,7 +101,7 @@
               </template>
               <template v-else>
                 <span v-if="item.skill" class="ai-queue-skill">/{{ item.skill }}</span>
-                <span class="ai-queue-text">{{ item.text }}</span>
+                <span class="ai-queue-text" :title="item.text">{{ item.text }}</span>
                 <!-- V7 排队时钟:该项已等待时长 m:ss(nowTick 每秒推进,入发送随队列清空即停) -->
                 <span class="ai-queue-wait">{{ queueItemWait(item) }}</span>
                 <!-- I4 队列回填:取出该项填回输入框重新编辑(从队列移除) -->
@@ -178,7 +178,7 @@ import { useToast } from '../composables/useToast'
 import { useProjectStore } from '../stores/project'
 import { collectTurnStats, turnMessages } from '../composables/ai/useTurnTrace'
 import { useAiContext } from '../composables/ai/useAiContext'
-import { emitReattachToMain, registerMainReattachListener } from '@/composables/ai/useAiWindow'
+import { emitReattachToMain, registerMainReattachListener, emitOpenSettings, registerMainOpenSettingsListener } from '@/composables/ai/useAiWindow'
 import type { AiMessage } from '../api/types'
 
 const props = defineProps<{
@@ -206,7 +206,10 @@ const pathAuthBypass = appSettings.useSetting<boolean>('df-ai-path-auth-bypass',
 /** 跳 Settings 配置 provider:嵌入模式 router.push;分离窗口无 Settings 路由,降级 toast 提示 */
 function goToSettings() {
   if (props.detached) {
-    showToast(t('aiChat.goToSettingsHint'), 'warning')
+    // P1(2026-10-10):分离窗跨窗直达设置 —— 广播 SETTINGS_EVENT,主窗口 listen 后显示主窗并导航 /settings。
+    // 原实现仅弹 toast 提示用户「去主窗口设置」,用户在分离窗工作流中改配置须手动找窗口,链路断裂。
+    void emitOpenSettings()
+    showToast(t('aiChat.goToSettingsOpened'), 'info')
     return
   }
   router.push('/settings')
@@ -535,6 +538,16 @@ async function handleForceSend() {
   await store.tryForceSend(confirmDialog, store.state.activeConversationId)
 }
 
+/** UI-4(看板 943)清空队列:二次确认后全清当前活跃会话排队消息。
+ *  一键误触会静默丢掉全部排队消息且不可恢复,复用本组件 ConfirmDialog 先例
+ *  (同 tryForceSend 的 confirmFn 注入形态),不引 window.confirm(仓风格)。 */
+async function handleClearQueue() {
+  const convId = store.state.activeConversationId
+  if (!convId) return
+  const confirmed = await confirmDialog(t('aiChat.clearQueueConfirm'))
+  if (confirmed) store.clearQueue(convId)
+}
+
 // ── 队列消息编辑(只编 text)+ 立即发送(插队)──
 
 /** 当前正在编辑的队列项 index;null=非编辑态 */
@@ -692,6 +705,18 @@ onMounted(async () => {
   // V11b: 主窗口注册跨窗收回分离窗监听(label 判定在内部,分离窗调用为 no-op);
   // reattachPanel 由主窗 store 执行(关全部 ai-detached-* 窗 + 恢复主面板)。
   registerMainReattachListener(() => store.reattachPanel())
+  // P1(2026-10-10):主窗口承接分离窗「去设置」广播(label 判定在内部,分离窗调用为 no-op)
+  registerMainOpenSettingsListener(async () => {
+    try {
+      const { WebviewWindow } = await import('@tauri-apps/api/webviewWindow')
+      const mainWin = await WebviewWindow.getByLabel('main')
+      if (mainWin) {
+        try { await mainWin.setFocus() } catch { /* setFocus 失败不阻塞导航 */ }
+        try { await mainWin.show() } catch { /* show 未授权,忽略 */ }
+      }
+    } catch { /* 忽略 */ }
+    router.push('/settings')
+  })
   // P1-e-d: 审批失败 toast(useAiApproval.approveToolCall 审批 IPC 未送达经事件总线广播,
   // 本组件 listen 弹本地 toast。主窗口/分离窗口各挂一份 AiChat,各自消费)
   _unlistenAiToast = await listen<{ msg: string; type: 'info' | 'error' | 'warning' | 'success' }>('ai-toast', (e) => {

src/components/ToolCard.vue

@@ -363,8 +363,22 @@ const statusTitle = computed(() => {
   return ''
 })
 
-/** 头部摘要 hover 回看完整路径(摘要统一缩略尾段,全路径走 title;路径类工具才有值) */
-const summaryTitle = computed(() => argString(props.tc.args, 'path'))
+/** 头部摘要 hover 回看关键参数全值(看板 945,按工具类型取参):
+ *  run_command→command(命令行是唯一关键参数);其余 path 优先(覆盖 read_file/grep/
+ *  search_files 等路径类),path 缺省回退首个非空字符串标量参数(hover 总有内容可看)。
+ *  摘要统一缩略尾段,全值走 title。 */
+const summaryTitle = computed(() => {
+  const args = props.tc.args
+  if (props.tc.name === 'run_command') return argString(args, 'command')
+  const byPath = argString(args, 'path')
+  if (byPath) return byPath
+  if (args && typeof args === 'object' && !Array.isArray(args)) {
+    for (const v of Object.values(args as Record<string, unknown>)) {
+      if (typeof v === 'string' && v) return v
+    }
+  }
+  return ''
+})
 
 /**
  * 风险等级(high/medium/low):审批挂起时卡片左侧着色条 + 风险徽章数据源。

src/components/ai/ChatInput.vue

@@ -66,6 +66,17 @@
       >
         <svg aria-hidden="true" width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><path d="M13 2H6a2 2 0 00-2 2v16a2 2 0 002 2h12a2 2 0 002-2V9z"/><polyline points="13 2 13 9 20 9"/></svg>
       </button>
+      <!-- 定时/待发箱入口(设计 2026-10-10 §3.4):icon-only ⏰ 开 OutboxDrawer
+           (组件内 Teleport 到 body,不占输入区布局;抽屉组件标签在模板尾部) -->
+      <button
+        type="button"
+        class="ai-img-pick-btn"
+        :title="$t('aiChat.outboxEntry')"
+        :aria-label="$t('aiChat.outboxEntry')"
+        @click="outboxOpen = true"
+      >
+        <svg aria-hidden="true" width="16" height="16" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><circle cx="12" cy="12" r="10"/><polyline points="12 6 12 12 16 14"/></svg>
+      </button>
       <!-- L2 状态机停止按钮三态(批2 1c):conv_state 派生 generating/stopping/error 三态,
            idle(或 conv_state 未收到回退旧 streaming=false)显发送按钮。双轨过渡不强制全替旧 bool。 -->
       <button
@@ -119,6 +130,8 @@
       @dismiss="dismissEnrichment"
       @update:enrichment-expanded="enrichmentExpanded = $event"
     />
+    <!-- 定时/待发箱抽屉(⏰ 入口;Teleport 到 body 渲染,此处仅挂组件) -->
+    <OutboxDrawer :open="outboxOpen" @close="outboxOpen = false" />
     <div class="ai-input-hint">{{ $t('aiChat.inputHint') }}</div>
   </div>
 </template>
@@ -148,6 +161,7 @@ import SkillMention from './SkillMention.vue'
 import ImageInput from './ImageInput.vue'
 import MentionPopover from './MentionPopover.vue'
 import EnrichmentPanel from './EnrichmentPanel.vue'
+import OutboxDrawer from './OutboxDrawer.vue'
 import { taskApi } from '../../api/task'
 import type { AiMessage, ContentPart, SkillInfo, ProjectRecord, TaskRecord, IdeaRecord, MentionSpan } from '../../api/types'
 import type { MentionItem } from './MentionItem'
@@ -173,6 +187,8 @@ const { t } = useI18n()
 
 const inputText = ref('')
 const inputEl = ref<HTMLTextAreaElement>()
+// 定时/待发箱抽屉开合(⏰ 按钮开,抽屉 close 事件关;列表数据在 useAiOutbox 单例缓存)
+const outboxOpen = ref(false)
 
 // ── F-260614-05 Phase 2b: 图片输入(粘贴/拖拽/选择) ──
 // 待发送图片列表(url 或 base64 data URI + media_type),handleSend 时构造 ContentPart[]。

src/components/ai/MessageList.vue

@@ -815,6 +815,23 @@ async function onInstallObscura(): Promise<void> {
   }
 }
 
+/**
+ * 看板 953:截断气泡「继续输出」按钮。回复被 provider max_tokens 上限腰斩时,注入续写指令
+ * 走既有发送通道(复用 store.sendMessage,对齐 onInstallObscura 模式),模型从上条回复的
+ * 中断处衔接输出,不再整段重发。streaming 中按钮 disabled(模板绑定);失败 toast 对齐
+ * handleRetry 错误处理模式。
+ */
+async function onContinueAfterTruncation(): Promise<void> {
+  try {
+    await store.sendMessage(t('ai.continueOutputPrompt'))
+    await nextTick()
+    scrollToBottom()
+  } catch (e) {
+    const msg = e instanceof Error ? e.message : String(e)
+    emit('toast', { msg: t('aiChat.sendFailed', { msg }), type: 'error' })
+  }
+}
+
 // ── F-15 阶段2: 消息分段(归档段折叠) ──
 // ChatMessage.status 渲染语义(types.ts AiMessage.status 字面量联合定义):
 //   null | undefined | 'active'        → 正常消息(现状)
@@ -1835,6 +1852,20 @@ defineExpose({
             <span class="ai-cursor"></span>
           </div>
 
+          <!-- 看板 953:max_tokens 截断提示条(仅末条 assistant 带标记时,气泡尾部弱提示)。
+               点击继续输出=注入续写指令复用 sendMessage,衔接勿重复;streaming 中禁用防叠发。 -->
+          <div v-if="isLastAi(item.msg) && item.msg.truncatedByMaxTokens" class="ai-msg-truncated">
+            <span class="ai-msg-truncated-text">{{ $t('ai.truncatedByMaxTokens') }}</span>
+            <button
+              class="ai-msg-act"
+              :disabled="store.state.streaming"
+              @click.stop="onContinueAfterTruncation"
+            >
+              <svg aria-hidden="true" width="11" height="11" viewBox="0 0 24 24" fill="none" stroke="currentColor" stroke-width="2" stroke-linecap="round" stroke-linejoin="round"><polyline points="23 4 23 10 17 10"/><path d="M20.49 15a9 9 0 11-2.12-9.36L23 10"/></svg>
+              <span>{{ $t('ai.continueOutput') }}</span>
+            </button>
+          </div>
+
           <!-- UX-03 / 错误气泡操作入口(重试 / 去设置) -->
           <div v-if="item.msg.isError" class="ai-msg-actions ai-msg-actions--error">
             <span v-if="!canRetry(item.msg)" class="ai-msg-not-retryable">{{ $t('ai.notRetryable') }}</span>
@@ -2458,6 +2489,23 @@ defineExpose({
   color: var(--df-accent);
 }
 
+/* 看板 953:max_tokens 截断弱提示条(末条 assistant 气泡尾部,warning 弱化不着色抢焦) */
+.ai-msg-truncated {
+  display: flex;
+  align-items: center;
+  gap: 6px;
+  margin-top: 4px;
+  padding: 3px 8px;
+  border-radius: var(--df-radius-sm);
+  background: color-mix(in srgb, var(--df-warning, #ffc107) 10%, transparent);
+  width: fit-content;
+}
+.ai-msg-truncated-text {
+  font-size: 10.5px;
+  line-height: 1.4;
+  color: var(--df-text-dim);
+}
+
 /* -- Streaming Cursor -- */
 .ai-cursor {
   display: inline-block;

src/components/ai/OutboxDrawer.vue

@@ -0,0 +1,405 @@
+<template>
+  <!-- 定时/待发列表抽屉(设计 2026-10-10 §3.4)。Teleport 到 body:入口按钮在 ChatInput
+       输入区,抽屉不能被 .ai-input-wrap 的 flex/overflow 布局裁剪,固定定位挂 body 最稳。
+       数据走 useAiOutbox(后端命令未接线时静默降级为空列表,抽屉可用不崩)。 -->
+  <Teleport to="body">
+    <div v-if="open" class="outbox-overlay" @click.self="emit('close')">
+      <div class="outbox-panel" role="dialog" aria-modal="true" :aria-label="t('aiChat.outboxTitle')">
+        <div class="outbox-header">
+          <h3 class="outbox-title">{{ t('aiChat.outboxTitle') }}</h3>
+          <button type="button" class="outbox-close" :aria-label="t('common.close')" @click="emit('close')">✕</button>
+        </div>
+
+        <div class="outbox-list">
+          <div v-if="loading && items.length === 0" class="outbox-empty">{{ t('common.loading') }}</div>
+          <div v-else-if="visibleItems.length === 0" class="outbox-empty">{{ t('aiChat.outboxEmpty') }}</div>
+          <div v-for="item in visibleItems" :key="item.id" class="outbox-row">
+            <div class="outbox-row-main">
+              <span class="outbox-time">{{ timeLabel(item) }}</span>
+              <span class="outbox-badge" :class="'outbox-badge--' + item.status">{{ statusLabel(item.status) }}</span>
+              <span class="outbox-text" :title="item.text">{{ summarize(item.text) }}</span>
+            </div>
+            <!-- 行操作仅活跃项(pending/sending/failed)展示;终态行只读留档 -->
+            <div v-if="isActive(item)" class="outbox-row-actions">
+              <template v-if="editingId === item.id">
+                <input
+                  v-model="editDue"
+                  type="datetime-local"
+                  class="outbox-dt"
+                  :aria-label="t('aiChat.outboxTimeLabel')"
+                />
+                <button type="button" class="outbox-act outbox-act--primary" :disabled="acting" @click="confirmReschedule(item)">{{ t('common.save') }}</button>
+                <button type="button" class="outbox-act" @click="editingId = null">{{ t('common.cancel') }}</button>
+              </template>
+              <template v-else>
+                <button type="button" class="outbox-act" :disabled="acting" @click="startReschedule(item)">{{ t('aiChat.outboxReschedule') }}</button>
+                <button type="button" class="outbox-act" :disabled="acting" @click="onSendNow(item)">{{ t('aiChat.outboxSendNow') }}</button>
+                <button type="button" class="outbox-act outbox-act--danger" :disabled="acting" @click="onCancel(item)">{{ t('aiChat.outboxCancel') }}</button>
+              </template>
+            </div>
+          </div>
+        </div>
+
+        <div class="outbox-new">
+          <div class="outbox-new-title">{{ t('aiChat.outboxNewTitle') }}</div>
+          <textarea
+            v-model="newText"
+            class="outbox-new-text"
+            rows="2"
+            :placeholder="t('aiChat.outboxNewPlaceholder')"
+          ></textarea>
+          <div class="outbox-new-row">
+            <input
+              v-model="newDue"
+              type="datetime-local"
+              class="outbox-dt"
+              :aria-label="t('aiChat.outboxTimeLabel')"
+            />
+            <button
+              type="button"
+              class="outbox-act outbox-act--primary"
+              :disabled="acting || !canCreate"
+              @click="onCreate"
+            >{{ t('aiChat.outboxCreate') }}</button>
+          </div>
+          <div v-if="notice" class="outbox-notice">{{ notice }}</div>
+        </div>
+      </div>
+    </div>
+  </Teleport>
+</template>
+
+<script setup lang="ts">
+// OutboxDrawer:定时/待发消息统一列表(跨会话),支持 改期/立即发送/取消 + 底部新建。
+// 挂载点在 ChatInput.vue(⏰ 入口按钮旁,Teleport 自脱离输入区布局);
+// 后端 ai_outbox_* 命令未接线时 invoke 静默降级(useAiOutbox 内),抽屉空态可用。
+import { ref, computed, watch } from 'vue'
+import { useI18n } from 'vue-i18n'
+import { useAiOutbox, type AiOutboxItem } from '../../composables/ai/useAiOutbox'
+import { useAiStore } from '../../stores/ai'
+
+const props = defineProps<{
+  /** 抽屉开合(ChatInput ⏰ 按钮控制;开时拉取列表,关时仅隐藏不销毁缓存) */
+  open: boolean
+}>()
+
+const emit = defineEmits<{
+  (e: 'close'): void
+}>()
+
+const { t } = useI18n()
+const store = useAiStore()
+const { items, loading, refresh, enqueue, cancel, reschedule, sendNow } = useAiOutbox()
+
+// 进入抽屉拉取(操作后由 composable 自动失效重拉,此处只管开合时机)
+watch(() => props.open, (v) => {
+  if (v) void refresh()
+})
+
+// ── 列表展示 ──
+
+// 排序:活跃在前(sending→pending→failed),终态沉底;同级按 due_at 升序(尽快项 due null 排最前)
+const STATUS_ORDER: Record<string, number> = { sending: 0, pending: 1, failed: 2, sent: 3, canceled: 4 }
+const visibleItems = computed<AiOutboxItem[]>(() => {
+  const arr = [...items.value]
+  arr.sort((a, b) => {
+    const d = (STATUS_ORDER[a.status] ?? 9) - (STATUS_ORDER[b.status] ?? 9)
+    if (d !== 0) return d
+    return (a.due_at ?? 0) - (b.due_at ?? 0)
+  })
+  return arr
+})
+
+/** 活跃项(有行操作):pending/sending/failed;终态只读 */
+function isActive(item: AiOutboxItem): boolean {
+  return item.status === 'pending' || item.status === 'sending' || item.status === 'failed'
+}
+
+const STATUS_KEYS: Record<string, string> = {
+  pending: 'aiChat.outboxStatusPending',
+  sending: 'aiChat.outboxStatusSending',
+  sent: 'aiChat.outboxStatusSent',
+  failed: 'aiChat.outboxStatusFailed',
+  canceled: 'aiChat.outboxStatusCanceled',
+}
+function statusLabel(s: string): string {
+  return t(STATUS_KEYS[s] ?? 'aiChat.outboxStatusPending')
+}
+
+/** 时间列:队列项(无 due_at)显「尽快」;定时项本地时间 */
+function timeLabel(item: AiOutboxItem): string {
+  if (item.due_at == null) return t('aiChat.outboxQueueAsap')
+  return new Date(item.due_at).toLocaleString()
+}
+
+/** 文本摘要(首行 + 截断,完整内容 title 悬浮) */
+function summarize(text: string): string {
+  const line = text.split('\n')[0] ?? ''
+  return line.length > 48 ? line.slice(0, 48) + '…' : line
+}
+
+// ── 行操作 ──
+
+const acting = ref(false)
+const notice = ref('')
+
+// 改期:行内 datetime 输入,预填当前 due(无 due 用现在+30min)
+const editingId = ref<string | null>(null)
+const editDue = ref('')
+
+function toLocalInput(ms: number): string {
+  const d = new Date(ms)
+  const pad = (n: number) => String(n).padStart(2, '0')
+  return `${d.getFullYear()}-${pad(d.getMonth() + 1)}-${pad(d.getDate())}T${pad(d.getHours())}:${pad(d.getMinutes())}`
+}
+
+function startReschedule(item: AiOutboxItem): void {
+  editingId.value = item.id
+  editDue.value = toLocalInput(item.due_at ?? Date.now() + 30 * 60 * 1000)
+  notice.value = ''
+}
+
+async function confirmReschedule(item: AiOutboxItem): Promise<void> {
+  const due = editDue.value ? new Date(editDue.value).getTime() : NaN
+  if (Number.isNaN(due)) return
+  acting.value = true
+  const ok = await reschedule(item.id, due)
+  acting.value = false
+  if (ok) {
+    editingId.value = null
+  } else {
+    notice.value = t('aiChat.outboxActionFail')
+  }
+}
+
+async function onSendNow(item: AiOutboxItem): Promise<void> {
+  acting.value = true
+  const ok = await sendNow(item.id)
+  acting.value = false
+  if (!ok) notice.value = t('aiChat.outboxActionFail')
+}
+
+async function onCancel(item: AiOutboxItem): Promise<void> {
+  acting.value = true
+  const ok = await cancel(item.id)
+  acting.value = false
+  if (!ok) notice.value = t('aiChat.outboxActionFail')
+}
+
+// ── 新建定时消息(目标会话=当前活跃会话;图片/技能入口待后续,纯文本先行) ──
+
+const newText = ref('')
+const newDue = ref('')
+
+const canCreate = computed(() => newText.value.trim().length > 0 && newDue.value !== '')
+
+async function onCreate(): Promise<void> {
+  const conv = store.state.activeConversationId
+  if (!conv) {
+    notice.value = t('aiChat.outboxNeedConv')
+    return
+  }
+  const due = newDue.value ? new Date(newDue.value).getTime() : NaN
+  if (Number.isNaN(due) || due <= Date.now()) {
+    notice.value = t('aiChat.outboxDuePast')
+    return
+  }
+  acting.value = true
+  const id = await enqueue(newText.value.trim(), due, { conversationId: conv })
+  acting.value = false
+  if (!id) {
+    notice.value = t('aiChat.outboxActionFail')
+    return
+  }
+  // 成功:清表单与提示(列表由 composable 自动刷新)
+  notice.value = ''
+  newText.value = ''
+  newDue.value = ''
+}
+</script>
+
+<style scoped>
+.outbox-overlay {
+  position: fixed;
+  inset: 0;
+  z-index: 1000;
+  background: rgba(0, 0, 0, 0.35);
+  display: flex;
+  justify-content: flex-end;
+}
+.outbox-panel {
+  width: min(420px, 92vw);
+  height: 100%;
+  background: var(--df-bg-card);
+  border-left: 0.5px solid var(--df-border);
+  display: flex;
+  flex-direction: column;
+  box-shadow: -8px 0 24px rgba(0, 0, 0, 0.15);
+}
+.outbox-header {
+  display: flex;
+  align-items: center;
+  justify-content: space-between;
+  padding: 12px 14px;
+  border-bottom: 0.5px solid var(--df-border);
+}
+.outbox-title {
+  margin: 0;
+  font-size: 14px;
+  font-weight: 600;
+  color: var(--df-text);
+}
+.outbox-close {
+  border: none;
+  background: transparent;
+  color: var(--df-text-dim);
+  cursor: pointer;
+  font-size: 14px;
+  padding: 4px 6px;
+  border-radius: var(--df-radius);
+}
+.outbox-close:hover {
+  background: var(--df-bg-raised);
+  color: var(--df-text);
+}
+.outbox-list {
+  flex: 1;
+  overflow-y: auto;
+  padding: 8px 10px;
+}
+.outbox-empty {
+  padding: 32px 0;
+  text-align: center;
+  font-size: 12px;
+  color: var(--df-text-dim);
+}
+.outbox-row {
+  display: flex;
+  flex-direction: column;
+  gap: 6px;
+  padding: 8px 6px;
+  border-bottom: 0.5px solid var(--df-border);
+}
+.outbox-row-main {
+  display: flex;
+  align-items: baseline;
+  gap: 8px;
+  min-width: 0;
+}
+.outbox-time {
+  flex-shrink: 0;
+  font-size: 11px;
+  color: var(--df-text-dim);
+  font-variant-numeric: tabular-nums;
+}
+.outbox-badge {
+  flex-shrink: 0;
+  font-size: 10px;
+  padding: 1px 6px;
+  border-radius: 999px;
+  border: 0.5px solid var(--df-border);
+  color: var(--df-text-dim);
+}
+.outbox-badge--pending {
+  color: var(--df-accent);
+  border-color: color-mix(in srgb, var(--df-accent) 40%, transparent);
+}
+.outbox-badge--sending {
+  color: var(--df-warning);
+  border-color: color-mix(in srgb, var(--df-warning) 40%, transparent);
+}
+.outbox-badge--failed {
+  color: #e5484d;
+  border-color: color-mix(in srgb, #e5484d 40%, transparent);
+}
+.outbox-text {
+  font-size: 12px;
+  color: var(--df-text);
+  overflow: hidden;
+  text-overflow: ellipsis;
+  white-space: nowrap;
+}
+.outbox-row-actions {
+  display: flex;
+  align-items: center;
+  gap: 6px;
+  flex-wrap: wrap;
+}
+.outbox-act {
+  border: 0.5px solid var(--df-border);
+  background: transparent;
+  color: var(--df-text-dim);
+  font-size: 11px;
+  padding: 3px 8px;
+  border-radius: var(--df-radius);
+  cursor: pointer;
+  transition: background 0.15s, color 0.15s;
+}
+.outbox-act:hover:not(:disabled) {
+  background: var(--df-bg-raised);
+  color: var(--df-text);
+}
+.outbox-act:disabled {
+  opacity: 0.4;
+  cursor: not-allowed;
+}
+.outbox-act--primary {
+  background: var(--df-accent);
+  border-color: var(--df-accent);
+  color: #fff;
+}
+.outbox-act--primary:hover:not(:disabled) {
+  background: var(--df-accent-hover);
+  color: #fff;
+}
+.outbox-act--danger:hover:not(:disabled) {
+  color: #e5484d;
+  border-color: color-mix(in srgb, #e5484d 40%, transparent);
+}
+.outbox-dt {
+  border: 0.5px solid var(--df-border);
+  background: var(--df-bg-raised);
+  color: var(--df-text);
+  font-size: 11px;
+  padding: 3px 6px;
+  border-radius: var(--df-radius);
+  color-scheme: inherit;
+}
+.outbox-new {
+  border-top: 0.5px solid var(--df-border);
+  padding: 10px 12px;
+  display: flex;
+  flex-direction: column;
+  gap: 8px;
+}
+.outbox-new-title {
+  font-size: 12px;
+  font-weight: 600;
+  color: var(--df-text);
+}
+.outbox-new-text {
+  width: 100%;
+  box-sizing: border-box;
+  border: 0.5px solid var(--df-border);
+  border-radius: var(--df-radius);
+  background: var(--df-bg-raised);
+  color: var(--df-text);
+  font-size: 12px;
+  font-family: var(--df-font-sans);
+  padding: 6px 8px;
+  resize: vertical;
+}
+.outbox-new-text::placeholder {
+  color: var(--df-text-dim);
+}
+.outbox-new-row {
+  display: flex;
+  align-items: center;
+  gap: 8px;
+  justify-content: space-between;
+}
+.outbox-notice {
+  font-size: 11px;
+  color: var(--df-warning);
+}
+</style>

src/components/ai/TaskParkedCard.vue

@@ -2,7 +2,10 @@
   <!-- 无人值守安全停靠卡(设计 §四):后端非正常终局(六路,stop_kind 词表见 i18n)emit AiTaskParked(已保存现场并复位
        generating,非死等),useAiEvents/useAiLifecycleEvents 登记 pendingTaskParked(per-conv Map)
        后本卡渲染。复用 HelpRequiredCard 模式:per-conv 守卫(挂起会话===当前展示会话)+ 行为后摘卡。
-       卡面 = reason 人话 + stop_kind 标签;操作两枚:继续(ai_continue_loop 重新授权续跑)/
+       卡面话术按 U5 异常态公式组织:状态(title+stop_kind 标签)+真实原因(reason)+已保留什么
+       (preserved 常驻行,前端兜底后端 reason 措辞不齐)+出路(继续/停止归档两枚;cancelled 只显
+       停止归档,第二条出路由 reason 文案"重发消息可继续"承担)。
+       操作两枚:继续(ai_continue_loop 重新授权续跑)/
        停止归档(ai_chat_stop 既有停止路径,归档会话走侧栏既有入口,此处不代做)。
        cancelled 例外只显停止归档(用户主动取消后不再诱导续跑)。 -->
   <div v-if="parkedActive" class="ai-task-parked">
@@ -11,6 +14,8 @@
       <span class="ai-task-parked-title">{{ t('ai.taskParked.title') }}</span>
     </div>
     <span class="ai-task-parked-reason">{{ parkedActive.reason }}</span>
+    <!-- U5 公式「已保留什么」:常驻一行而非依赖后端 reason 措辞,保证任意 stop_kind 公式完整 -->
+    <span class="ai-task-parked-preserved">{{ t('ai.taskParked.preserved') }}</span>
     <div class="ai-task-parked-actions">
       <!-- cancelled(审批挂起中被用户主动取消)不显「继续」:用户刚点过停止,首位按钮再诱导续跑与意图相逆 -->
       <button v-if="parkedActive.stopKind !== 'cancelled'" class="ai-task-parked-btn" :disabled="busy" @click="handleContinue">{{ t('aiChat.continueLoop') }}</button>
@@ -136,6 +141,12 @@ async function handleStopArchive(): Promise<void> {
   word-break: break-word;
   white-space: pre-wrap;
 }
+/* preserved 行与 reason 同色系但更弱一级:保证信息在,不与真实原因抢视觉权重 */
+.ai-task-parked-preserved {
+  font-size: 10.5px;
+  color: var(--df-text-dim);
+  opacity: 0.8;
+}
 .ai-task-parked-actions {
   display: flex;
   align-items: center;

src/components/ai/TopBar.vue

@@ -156,13 +156,22 @@
     <div class="ai-bottom-tools-right">
     <!-- 无人值守任务进度(AiTaskProgress,设计 §4.3):当前会话有进度快照时显示,
          静默期(子代理长跑/大工具执行/重入退避)凭最后快照可见「在干活」;终局即清。
-         max_rounds=0(不限轮次)不显分母进度条,只显轮次与 token。
+         max_rounds=0(不限轮次)不显分母进度条,只显轮次与 token(防假进度规则2:无分母禁画百分比)。
+         U2 三层(设计 §三):活性=脉冲点(静默降速,永不伪装停止);位置=轮次/token 数字;
+         内容=阶段标签(最近工具映射)+已完成步数,静默档附加判断材料(不自动判死,决策留给用户)。
          A8 双进度矛盾:输入区 agentic 实时进度条可见时(streaming 且 agentRound>0)
          隐藏本快照徽章——两处同屏且轮数口径不同(实时 vs 每5轮快照)互相矛盾;
          快照徽章只服务输入区不可见场景(托盘/单轮大工具静默期)。 -->
-    <div v-if="taskProgress && !agenticProgressVisible" class="ai-task-progress">
+    <div
+      v-if="taskProgress && !agenticProgressVisible"
+      class="ai-task-progress"
+      :class="silenceTier !== 'normal' ? `ai-task-progress--${silenceTier}` : ''"
+      :title="taskProgressHint"
+    >
       <span class="ai-task-progress-dot"></span>
       <span class="ai-task-progress-text">{{ taskProgressText }}</span>
+      <span v-if="stageText" class="ai-task-progress-stage">{{ stageText }}</span>
+      <span v-if="silenceSuffix" class="ai-task-progress-silence">{{ silenceSuffix }}</span>
       <span
         v-if="taskProgress.maxRounds > 0"
         class="ai-task-progress-bar"
@@ -589,6 +598,106 @@ const taskProgressPct = computed(() => {
   return Math.min(100, Math.round((p.round / p.maxRounds) * 100))
 })
 
+// ── U2 进度语义三层补齐(设计 §三):内容层(最近动作)+静默期三档(§3.4) ──
+// 规则1 兑现:下方定时器只算「距上次信号多久」,永不推进任何计数(轮次/步数只由事件写)。
+// 规则4 兑现:活性=脉冲节奏(静默降速),位置=数字,分开渲染;永不自动判死(重入退避最长
+// 15min 是合法等待),警示档只给判断材料,「疑似卡死可停止」的决策留给用户。
+
+// 内容层数据源:工具卡挂在 assistant 消息上(useAiToolEvents),从消息流纯派生——零新事件
+// 监听零协议改动。steps=当前消息流工具调用总数(压缩仅标记不清消息,计数不回退;清空上下文
+// 重置属用户显式重开,语义自洽);last 优先取 running 卡(当前动作),无 running 取最后一张。
+const toolActivity = computed(() => {
+  let steps = 0
+  let lastName: string | null = null
+  let runningName: string | null = null
+  for (const m of store.state.messages) {
+    const tcs = m.toolCalls
+    if (!tcs?.length) continue
+    steps += tcs.length
+    for (const tc of tcs) {
+      lastName = tc.name
+      if (tc.status === 'running') runningName = tc.name
+    }
+  }
+  return { steps, last: runningName ?? lastName }
+})
+
+// 阶段标签映射(设计 §3.3 同源词表桌面份):工具名→语义标签;未命中直显工具名
+const STAGE_KEYS: Record<string, string> = {
+  read_file: 'read', read_symbol: 'read', search_files: 'read', list_directory: 'read',
+  write_file: 'edit', patch_file: 'edit',
+  run_command: 'run',
+  spawn_agent: 'agent',
+}
+const STAGE_PREFIXES: [string, string][] = [['knowledge_', 'knowledge'], ['idea_', 'knowledge'], ['task_', 'knowledge']]
+function stageKeyOf(name: string): string | null {
+  if (STAGE_KEYS[name]) return STAGE_KEYS[name]
+  for (const [p, k] of STAGE_PREFIXES) if (name.startsWith(p)) return k
+  return null
+}
+const stageText = computed(() => {
+  const act = toolActivity.value
+  if (!act.last) return ''
+  const key = stageKeyOf(act.last)
+  const parts = [key ? t(`ai.taskProgress.stage.${key}`) : act.last]
+  if (act.steps > 0) parts.push(t('ai.taskProgress.steps', { n: act.steps }))
+  return parts.join(' · ')
+})
+
+// 静默期计时:基准=最近一次真实信号(AiTaskProgress.updatedAt 或消息流指纹变化时刻)。
+// 指纹 watch 只在信号真实变化时刷新(规则5:同值重刷不产生新活性);文本增量/工具卡增减/
+// 轮次推进都会被捕获。
+const lastActivityAt = ref(Date.now())
+watch(() => {
+  const msgs = store.state.messages
+  const last = msgs[msgs.length - 1]
+  return [msgs.length, last?.id, last?.content?.length, last?.toolCalls?.length, store.state.agentRound]
+}, () => { lastActivityAt.value = Date.now() })
+// 切会话用已知时间锚点重置(快照 updatedAt/末条消息时间),防「刚切回就看不到本就持续的静默」
+watch(() => store.state.activeConversationId, () => {
+  const msgs = store.state.messages
+  const last = msgs[msgs.length - 1]
+  lastActivityAt.value = Math.max(taskProgress.value?.updatedAt ?? 0, last?.timestamp ?? 0)
+})
+// 15s 心跳只为刷新「距上次信号」展示(§3.4 档位阈值 2/10min,15s 粒度足够;规则1:定时器
+// 不碰任何计数)
+const nowTick = ref(Date.now())
+let silenceTimer: ReturnType<typeof setInterval> | null = null
+onMounted(() => {
+  silenceTimer = setInterval(() => { nowTick.value = Date.now() }, 15_000)
+})
+onBeforeUnmount(() => {
+  if (silenceTimer) { clearInterval(silenceTimer); silenceTimer = null }
+})
+const silenceSecs = computed(() => {
+  const p = taskProgress.value
+  if (!p) return 0
+  return Math.max(0, Math.floor((nowTick.value - Math.max(p.updatedAt, lastActivityAt.value)) / 1000))
+})
+// §3.4 三档:<2min 正常 / 2-10min 长执行(脉冲降速) / >10min 警示(给判断材料不自动判死)
+const silenceTier = computed<'normal' | 'long' | 'warn'>(() => {
+  const s = silenceSecs.value
+  if (s >= 600) return 'warn'
+  if (s >= 120) return 'long'
+  return 'normal'
+})
+const silenceSuffix = computed(() => {
+  if (!taskProgress.value) return ''
+  if (silenceTier.value === 'warn') return t('ai.taskProgress.silentShort', { n: Math.floor(silenceSecs.value / 60) })
+  if (silenceTier.value === 'long') return t('ai.taskProgress.longRun')
+  return ''
+})
+// 悬浮提示:警示档给完整判断材料(距上次活动 n 分钟+最后动作+出路),长执行档给成因提示,
+// 正常档不弹(位置信息正文已有,避免无意义提示)
+const taskProgressHint = computed(() => {
+  if (!taskProgress.value) return ''
+  if (silenceTier.value === 'warn') {
+    return t('ai.taskProgress.silentWarn', { n: Math.floor(silenceSecs.value / 60), tool: toolActivity.value.last ?? '-' })
+  }
+  if (silenceTier.value === 'long') return t('ai.taskProgress.longRunHint')
+  return ''
+})
+
 // 🎯 对话目标面板(对话透明化 L1:Goal visibility)
 const goals = ref<GoalEntry[]>([])
 const goalsExpanded = ref(false)
@@ -1077,6 +1186,20 @@ onBeforeUnmount(() => {
   0%, 100% { opacity: 1; }
   50% { opacity: 0.35; }
 }
+/* U2 静默期档位(设计 §3.4):长执行/警示=脉冲降速——活性层表达「慢」而非伪装「停」,
+   不自动判死;内容层/静默档文字用弱化色,视觉从属位置层数字 */
+.ai-task-progress--long .ai-task-progress-dot,
+.ai-task-progress--warn .ai-task-progress-dot {
+  animation-duration: 3.2s;
+}
+.ai-task-progress-stage {
+  font-size: 11px;
+  color: var(--df-text-dim);
+}
+.ai-task-progress-silence {
+  font-size: 11px;
+  color: color-mix(in srgb, var(--df-warning) 70%, var(--df-text-secondary));
+}
 /* ═══ Provider Bar (header 中间区;div→button 后归零默认样式) ═══ */
 .ai-provider-bar {
   display: inline-flex;

src/components/layout/AppLayout.vue

@@ -17,7 +17,7 @@
 
     <!-- ═══ AI Chat Panel ═══ -->
     <transition name="ai-slide">
-      <div v-if="aiStore.state.panelOpen && !aiStore.state.detached" class="df-ai-panel" :style="panelStyle">
+      <div v-if="aiStore.state.panelOpen && !aiStore.state.detached" class="df-ai-panel" :class="{ 'ai-panel--resizing': resizing }" :style="panelStyle">
         <div v-show="!aiStore.state.maximized" class="ai-resize-handle" :class="{ active: resizing }" @mousedown="startResize"></div>
         <slot name="ai-panel" />
       </div>
@@ -165,6 +165,13 @@ function startResize(e: MouseEvent) {
   position: relative;
   height: 100vh;
   overflow: hidden;
+  /* P2 修复:最大化↔分栏切换原为瞬时跳变,加宽度/弹性过渡消除视觉闪跳。
+     仅过渡 flex-grow/width/max-width,不动 opacity(避免与 ai-slide 叠加抖动) */
+  transition: max-width 0.25s var(--df-ease), width 0.25s var(--df-ease);
+}
+/* 拖拽调整宽度期间禁用过渡:否则每像素跟随都被 0.25s 缓动拖尾,手感发飘 */
+.df-ai-panel.ai-panel--resizing {
+  transition: none;
 }
 
 /* — AI Resize Handle — */

src/composables/ai/streamSelfHeal.ts

@@ -0,0 +1,16 @@
+//! 心跳自愈判定纯函数(看板 949 前端兜底)
+//!
+//! 场景:force_send 竞态窗口下,旧任务迟到的 AiCompleted 可能把前端流式态误收敛为
+//! false(迟到仲裁查后端 generating 时新 loop 尚未占用,误判真收尾);后端新 loop 仍在
+//! 生成,delta 照流但发送按钮失态。心跳(后端 30s 一拍)到达时用本判定决定是否自愈。
+//!
+//! 互补不打架:仅「后端在生成 且 前端流式态丢失」才自愈;迟到仲裁只拦「后端 true 且
+//! 前端 true」,两机制判定域不相交,不会同时动作。
+//!
+//! 独立零依赖模块:node --test 单测直接 import(useAiStreamingEvents.ts 带 @/ 别名与
+//! Tauri 依赖,单测环境无法加载),useAiStreamingEvents.ts 消费本判定。
+
+/** 自愈判定:后端生成中(backendGenerating=true)且前端流式态丢失(frontendStreaming=false)→ true */
+export function shouldSelfHealStreaming(backendGenerating: boolean, frontendStreaming: boolean): boolean {
+  return backendGenerating && !frontendStreaming
+}

src/composables/ai/useAiLifecycleEvents.ts

@@ -202,6 +202,19 @@ export async function handleLifecycleEvent(event: AiChatEvent): Promise<boolean>
           timestamp: Date.now(),
         })
       }
+      // 看板 953:max_tokens 截断标记(主对话气泡截断提示条的数据源)。后端正常收敛出口且
+      // 最终回复被 provider 上限腰斩时为 true;旧事件无字段(undefined)不标,向后兼容。
+      // 只标末条 assistant(倒序找,跳过错误气泡):用户再发消息/继续输出后新气泡顶掉末条
+      // 位置,提示条随 isLastAi 判定自然消失,无需清除逻辑。
+      if (event.truncated_by_max_tokens) {
+        for (let i = state.messages.length - 1; i >= 0; i--) {
+          const m = state.messages[i]
+          if (m.role === 'assistant' && !m.isError) {
+            m.truncatedByMaxTokens = true
+            break
+          }
+        }
+      }
       // 直接从事件更新 pinned_goals,不等 loadConversations 异步刷新
       if (event.pinned_goals && state.activeConversationId) {
         const conv = state.conversations.find(c => c.id === state.activeConversationId)

src/composables/ai/useAiOutbox.ts

@@ -0,0 +1,131 @@
+//! 待发箱(outbox)视图 — 定时消息/待发队列的 IPC 封装 + 本地缓存
+//!
+//! 对应后端 `src-tauri/src/commands/ai/outbox.rs` 的 ai_outbox_* 命令族。
+//! ⚠️ 后端命令已实现但未注册进 generate_handler(接线清单见交付说明):接线前所有
+//! invoke 均失败,本模块**静默降级**(console.warn + 空列表/false 返回),保证抽屉
+//! 可用不崩;接线后行为自动完整,前端零改动。
+//!
+//! 加载/失效时机:进入抽屉 refresh()(OutboxDrawer watch open)+ 每次操作成功后
+//! 自动失效重拉 + 监听后端 ai-outbox-changed 事件增量刷新(他窗口操作也能同步)。
+//!
+//! 设计见 docs/02-架构设计/专项设计/定时消息与待发箱持久化-设计-2026-10-10.md §3.4。
+
+import { ref } from 'vue'
+import { invoke } from '@tauri-apps/api/core'
+import { listen } from '@tauri-apps/api/event'
+
+/** 与后端 AiOutboxItem(Tauri DTO)同形,snake_case 原样透传 */
+export interface AiOutboxItem {
+  id: string
+  conversation_id: string
+  kind: 'queue' | 'scheduled'
+  text: string
+  skill: string | null
+  model_override: string | null
+  /** epoch ms;null=尽快队列项 */
+  due_at: number | null
+  status: 'pending' | 'sending' | 'sent' | 'failed' | 'canceled'
+  attempts: number
+  last_error: string | null
+  created_at: number
+  updated_at: number
+  sent_at: number | null
+}
+
+// 模块级单例缓存:抽屉与入口按钮(角标计数等未来消费方)共享同一份视图,
+// 操作后统一失效重拉,避免多实例缓存漂移。主/分离窗口各自 JS context 天然隔离。
+const items = ref<AiOutboxItem[]>([])
+const loading = ref(false)
+
+/** invoke 静默降级封装:命令未接线/后端错误时 warn 并返回 fallback,调用方 UI 不崩 */
+async function safeInvoke<T>(cmd: string, args: Record<string, unknown>, fallback: T, label: string): Promise<T> {
+  try {
+    return await invoke<T>(cmd, args)
+  } catch (e) {
+    console.warn(`[AI outbox] ${label}失败(命令未接线或后端错误):`, e)
+    return fallback
+  }
+}
+
+/** 拉取列表写入缓存。convId 缺省=跨会话全量(抽屉口径);失败降级空列表 */
+async function refresh(convId?: string): Promise<void> {
+  loading.value = true
+  try {
+    const list = await invoke<AiOutboxItem[]>('ai_outbox_list', { convId: convId ?? null })
+    items.value = Array.isArray(list) ? list : []
+  } catch (e) {
+    console.warn('[AI outbox] 列表拉取失败(命令未接线或后端错误):', e)
+    items.value = []
+  } finally {
+    loading.value = false
+  }
+}
+
+/**
+ * 新建定时/待发项。dueAt=epoch ms(定时项,须晚于当前时刻);缺省=尽快队列项。
+ * conversationId 必传(后端不猜 active);返回新项 id,失败 null(错误已 warn)。
+ */
+async function enqueue(
+  text: string,
+  dueAt?: number,
+  opts?: { skill?: string; modelOverride?: string; conversationId?: string },
+): Promise<string | null> {
+  const id = await safeInvoke<string | null>(
+    'ai_outbox_enqueue',
+    {
+      text,
+      dueAt: dueAt ?? null,
+      skill: opts?.skill ?? null,
+      modelOverride: opts?.modelOverride ?? null,
+      conversationId: opts?.conversationId ?? null,
+    },
+    null,
+    '入队',
+  )
+  if (id) void refresh()
+  return id
+}
+
+/** 取消待发项。成功 true 并刷新;未接线/后端拒绝(终态等)false */
+async function cancel(id: string): Promise<boolean> {
+  const ok = await safeInvoke<boolean>('ai_outbox_cancel', { id }, false, '取消')
+  if (ok) void refresh()
+  return ok
+}
+
+/** 改期(dueAt=epoch ms;后端允许过去时刻=立即派发语义)。成功 true 并刷新 */
+async function reschedule(id: string, dueAt: number): Promise<boolean> {
+  const ok = await safeInvoke<boolean>('ai_outbox_reschedule', { id, dueAt }, false, '改期')
+  if (ok) void refresh()
+  return ok
+}
+
+/** 编辑待发文本。终态项后端拒绝(返 false) */
+async function updateText(id: string, text: string): Promise<boolean> {
+  const ok = await safeInvoke<boolean>('ai_outbox_update_text', { id, text }, false, '编辑')
+  if (ok) void refresh()
+  return ok
+}
+
+/** 立即派发一次(后端与调度器同一条派发链;不受 df-outbox-enabled 开关约束)。成功 true 并刷新 */
+async function sendNow(id: string): Promise<boolean> {
+  const ok = await safeInvoke<boolean>('ai_outbox_send_now', { id }, false, '立即发送')
+  if (ok) void refresh()
+  return ok
+}
+
+// 后端变更事件驱动刷新:任何入队/派发/失败/改期后各窗口视图同步。
+// 幂等注册(模块级一次);后端未接线时不会有该事件,注册本身零开销。
+let changeListenerReady = false
+function ensureChangeListener(): void {
+  if (changeListenerReady) return
+  changeListenerReady = true
+  listen('ai-outbox-changed', () => { void refresh() }).catch(() => {
+    changeListenerReady = false // 注册失败(非 Tauri 环境等)允许下次重试
+  })
+}
+
+export function useAiOutbox() {
+  ensureChangeListener()
+  return { items, loading, refresh, enqueue, cancel, reschedule, updateText, sendNow }
+}

src/composables/ai/useAiPanel.ts

@@ -94,6 +94,16 @@ export function setSidebarWidth(px: number) {
   state.sidebarWidth = clampSidebarWidth(px)
 }
 
+/** 设置最大化态 — 单一写入入口(所有 maximized 赋值收敛于此,杜绝绕过 persistUiState 的分歧)。
+ *  修复 P0:App.vue 路由 watch 直接置 state.maximized=false 未走 persistUiState,
+ *  导致内存态与磁盘态(df-ai-ui)不一致,重启后旧值闪回。本函数保证任何调用方改最大化
+ *  态都会落盘。
+ *  参数 persist:默认 true(立即落盘)。高频/连续赋值场景可调 false,由调用方收口落盘。 */
+export function setMaximized(v: boolean, persist = true) {
+  state.maximized = v
+  if (persist) persistUiState()
+}
+
 /** 切换面板展开/收起(收起时同步退出最大化,避免 v-show 卡死 main 交互) */
 function togglePanel() {
   state.panelOpen = !state.panelOpen
@@ -104,8 +114,7 @@ function togglePanel() {
 
 /** 最大化/还原侧栏宽度 */
 function toggleMaximize() {
-  state.maximized = !state.maximized
-  persistUiState()
+  setMaximized(!state.maximized)
 }
 
 /**
@@ -225,6 +234,7 @@ export function useAiPanel() {
     clearChat,
     // 侧栏宽度拖拽
     setSidebarWidth,
+    setMaximized,
     persistUiState,
   }
 }

src/composables/ai/useAiStreamingEvents.ts

@@ -10,9 +10,12 @@ import { emit } from '@tauri-apps/api/event'
 import { useAppSettingsStore } from '@/stores/appSettings'
 import { state } from '@/stores/ai'
 import { t } from '@/i18n/i18n-helpers'
-import { nextMsgId, findToolCall, startApprovalTimer, clearToolSlowTimer, getLastDelta, setLastDelta, flushCurrentText, addPendingApproval, setConvRound } from './aiShared'
+import { nextMsgId, findToolCall, startApprovalTimer, clearToolSlowTimer, getLastDelta, setLastDelta, flushCurrentText, addPendingApproval, setConvRound, getConvStreamState } from './aiShared'
 import { resetTextIdleTimer, clearTextIdleTimer, pendingDirAuths, pushPendingDirAuth } from './useAiPendingState'
-import { clearStreamWatchdog } from './useAiStream'
+import { clearStreamWatchdog, resetStreamWatchdog } from './useAiStream'
+import { setStreaming } from './streamingGuard'
+import { aiApi } from '@/api/ai'
+import { shouldSelfHealStreaming } from './streamSelfHeal'
 import type { AiChatEvent, AiMessage, MessageId } from '@/api/types'
 
 const appSettings = useAppSettingsStore()
@@ -23,6 +26,32 @@ function isUnifiedApproval(): boolean {
   return appSettings.get<boolean>('df-ai-unified-approval', true)
 }
 
+/**
+ * 心跳自愈(火后不理,看板 949 兜底):前端流式态丢失而后端仍在生成时恢复 streaming + 重启看门狗。
+ *
+ * 链路:force_send 竞态让旧任务迟到 AiCompleted 误杀前端流式态(仲裁查后端时新 loop 尚未
+ * 占用 generating),后端内容照流但按钮失态;心跳 30s 一拍,到达即先零开销短路(前端仍 true
+ * 不查态),疑似丢失才 IPC 查后端真值,「后端 true 且前端 false」才动作 —— 与迟到仲裁
+ * (后端 true 且前端 true 才拦截)判定域互补,两机制不同时动作。
+ * IPC 返回后重读前端流式态再决策(等待窗口内可能有新事件改写,防陈旧快照反向覆盖);
+ * 查态失败静默放弃,等下一拍心跳再试。
+ */
+function selfHealStreamingOnHeartbeat(convId?: string | null): void {
+  const target = convId || state.activeConversationId
+  // 仅当前视图会话:state.streaming 是当前视图流式态,后台会话事件在 useAiEvents 路由层已过滤
+  if (!target || target !== state.activeConversationId) return
+  // 疑似丢失才发起 IPC(前端仍 streaming=true 是常态,零开销短路)
+  if (!shouldSelfHealStreaming(true, getConvStreamState(target)?.streaming ?? false)) return
+  void aiApi.isGenerating(target)
+    .then(backendGenerating => {
+      if (!shouldSelfHealStreaming(backendGenerating, getConvStreamState(target)?.streaming ?? false)) return
+      setStreaming(true, { convId: target, reason: 'AiHeartbeat 自愈' })
+      resetStreamWatchdog(target)
+      console.warn(`[AI-Streaming] 心跳自愈:后端生成中而前端流式态丢失,已恢复(conv=${target.slice(0, 8)})`)
+    })
+    .catch(() => { /* 查态失败:本拍不动作,等下一拍心跳 */ })
+}
+
 /** streaming 域:流式文本累积/新轮/心跳/重试。返回 true=已处理。 */
 export function handleStreamingEvent(event: AiChatEvent): boolean {
   switch (event.type) {
@@ -72,9 +101,13 @@ export function handleStreamingEvent(event: AiChatEvent): boolean {
       return true
     }
 
-    case 'AiHeartbeat':
+    case 'AiHeartbeat': {
       // 心跳仅维持看门狗(已在上方 reset),显式 case 防 switch 穿透
+      // 看板 949 自愈兜底:后端在生成而前端流式态丢失(force_send 竞态误杀)时恢复,
+      // 自愈延迟≤一拍心跳(后端 30s)
+      selfHealStreamingOnHeartbeat(event.conversation_id)
       return true
+    }
 
     case 'AiStreamRetry': {
       // 重试中——首条重试推过渡气泡(isRetry 标记,MessageList 渲染为琥珀细条而非错误红卡),

src/composables/ai/useAiWindow.ts

@@ -98,6 +98,44 @@ export function registerMainReattachListener(handler: () => Promise<void>): void
   })()
 }
 
+// ── P1(2026-10-10)分离窗口「去设置」跨窗直达:分离窗无 Settings 路由,原实现仅 toast
+// 提示用户「去主窗口设置」,链路断裂(分离窗长时间工作场景下改配置须手动找窗口)。
+// 现走事件广播(对齐 REATTACH_EVENT 模式):分离窗 emit → 主窗口 listen → 显示/聚焦
+// 主窗 + router.push('/settings')。
+const SETTINGS_EVENT = 'ai-open-settings'
+
+/** 发起方(分离窗)侧:广播「打开设置」事件,主窗口消费 */
+export async function emitOpenSettings(): Promise<void> {
+  try {
+    const { emit } = await import('@tauri-apps/api/event')
+    await emit(SETTINGS_EVENT, {})
+  } catch (e) {
+    console.error('[AI] 跨窗打开设置事件发送失败:', e)
+  }
+}
+
+/** 承接方(主窗口)侧:注册「打开设置」监听。handler 由 AiChat 注入(显示主窗 + 导航) */
+let _settingsListenerRegistered = false
+export function registerMainOpenSettingsListener(handler: () => Promise<void>): void {
+  if (_settingsListenerRegistered) return
+  _settingsListenerRegistered = true
+  void (async () => {
+    try {
+      const [{ listen }, { getCurrentWebviewWindow }] = await Promise.all([
+        import('@tauri-apps/api/event'),
+        import('@tauri-apps/api/webviewWindow'),
+      ])
+      if (getCurrentWebviewWindow().label !== 'main') return
+      await listen<Record<string, never>>(SETTINGS_EVENT, async () => {
+        await handler()
+      })
+    } catch (e) {
+      console.error('[AI] 主窗口设置监听注册失败:', e)
+      _settingsListenerRegistered = false
+    }
+  })()
+}
+
 /** 承接方(分离窗)侧:注册跨窗切会话监听,切换完成后聚焦本窗口(点击发生在主窗口,
  *  焦点须跟到承接会话的窗口,否则用户感知「点了没反应」)。
  *  handler 由 useAiConversations 注入:切会话逻辑归会话域,避免本模块反向依赖造成模块环。 */

src/i18n/en/ai.ts

@@ -37,6 +37,12 @@ export default {
     responseIncomplete: '⚠ The response is incomplete due to a network interruption. The above is the partial content received. You can resend to get a full reply.',
     // Backend auto-finalized after reaching the max rounds (not a network interruption; reuses the incomplete bubble with distinct semantics)
     responseIncompleteMaxRounds: 'Maximum rounds reached. The AI has finalized this round with the results so far; continue the conversation to have it refine and complete the answer.',
+    // Kanban 953: weak hint bar at the bubble tail when a reply converged normally but was cut off by the provider max_tokens limit (finish_reason=length)
+    truncatedByMaxTokens: '⚠ Output truncated by the length limit; content may be incomplete',
+    // Continue button on the truncation hint bar
+    continueOutput: 'Continue',
+    // Continuation instruction injected by the Continue button (sent as a new user message; resume seamlessly, no repetition)
+    continueOutputPrompt: 'Continue your previous reply from where it stopped. Resume seamlessly; do not repeat content already output.',
     queueFull: 'Send queue is full (max {limit} items)',
     forceSendConfirm: 'AI generation appears stuck. Force-stop and send anyway? (This may interrupt the current response)',
     forceSendBtn: 'Force Send',
@@ -139,16 +145,35 @@ export default {
         midstream: 'Stream interrupted',
       },
       unknownKind: 'Parked',
+      // U5 copy formula element 3 "what is preserved": always-on frontend fallback —
+      // backend park reasons vary in wording (non-timeout fatal / external_change omit it),
+      // one line here keeps the formula intact for every stop_kind
+      preserved: 'Conversation context and completed progress are fully preserved, ready to resume',
       stopArchive: 'Stop & Archive',
       stoppedToast: 'Task stopped, will not resume',
       archiveFailed: 'Stop failed: {msg}',
     },
     // Unattended task progress (AiTaskProgress, design §4.3): TopBar lightweight progress,
-    // visible as "still working" during silent periods
+    // visible as "still working" during silent periods.
+    // U2 additions: stage = content-layer phase labels (§3.3 tool-to-semantic mapping,
+    // unmatched tools show the raw name instead); longRun/silent = §3.4 silence tiers,
+    // judgement material only, never an automatic stuck verdict.
     taskProgress: {
       running: 'Running · round {round}/{max} · {tokens} tokens',
       unlimited: 'Running · round {round} · {tokens} tokens',
       barHint: 'Round {round}/{max}',
+      steps: '{n} steps done',
+      stage: {
+        read: 'Reading & searching',
+        edit: 'Editing & writing',
+        run: 'Running & verifying',
+        agent: 'Sub-agent working',
+        knowledge: 'Knowledge & task ops',
+      },
+      longRun: 'Slow step',
+      longRunHint: 'Current step is taking a while (large tool or sub-agent running)',
+      silentShort: '{n} min no activity',
+      silentWarn: 'No activity for {n} min (last action: {tool}). Still running — check the trace, or stop if it looks stuck.',
     },
     // Unattended boundary config card (P1 simplified: single authorization tier,
     // only the two budget fields remain user-tunable)

src/i18n/en/aiChat.ts

@@ -113,6 +113,7 @@ export default {
     examplePrompt4: 'Generate a project idea',
     goToSettings: 'Open settings',
     goToSettingsHint: 'The detached window cannot open settings directly. Configure AI providers on the main window’s Settings page.',
+    goToSettingsOpened: 'Settings opened in the main window',
     goToTaskHint: 'The detached window cannot open tasks directly. View task details in the main window.',
 
     // ── Provider switch feedback ──
@@ -125,6 +126,8 @@ export default {
     // ── Pending queue ──
     queueTitle: 'Queued {n}/10',
     clearQueue: 'Clear',
+    // UI-4 (board 943): clear-queue confirm (one click wipes all queued items, unrecoverable)
+    clearQueueConfirm: 'Clear all queued messages? This cannot be undone',
     // queue item edit + send now
     queueEdit: 'Edit',
     queueSave: 'Save',
@@ -339,5 +342,26 @@ export default {
     editEmptyBlocked: 'Edit content is empty, submission cancelled',
     skillNoMatchQuery: 'No skill matches "{q}", send blocked',
     backfillOverwriteConfirm: 'The input box has an unsent draft; backfilling will overwrite it. Continue?',
+
+    // ── Outbox / scheduled messages (OutboxDrawer + ChatInput clock entry, design 2026-10-10 §3.4) ──
+    outboxEntry: 'Scheduled messages',
+    outboxTitle: 'Scheduled & Outbox',
+    outboxEmpty: 'No scheduled or queued messages',
+    outboxNewTitle: 'New scheduled message',
+    outboxNewPlaceholder: 'Content to send on schedule',
+    outboxTimeLabel: 'Send time',
+    outboxCreate: 'Schedule',
+    outboxReschedule: 'Reschedule',
+    outboxSendNow: 'Send now',
+    outboxCancel: 'Cancel',
+    outboxQueueAsap: 'ASAP',
+    outboxStatusPending: 'Pending',
+    outboxStatusSending: 'Sending',
+    outboxStatusSent: 'Sent',
+    outboxStatusFailed: 'Failed',
+    outboxStatusCanceled: 'Canceled',
+    outboxActionFail: 'Action failed, please retry',
+    outboxDuePast: 'Scheduled time must be in the future',
+    outboxNeedConv: 'No target conversation; create one first',
   },
 }

src/i18n/zh-CN/ai.ts

@@ -37,6 +37,12 @@ export default {
     responseIncomplete: '⚠ 响应因网络中断不完整，以上为已接收的部分内容。可重新发送以获取完整回复。',
     // 达最大轮次后端自动收尾(非网络中断,复用 incomplete 气泡但语义独立)
     responseIncompleteMaxRounds: '已达最大轮次，AI 已自动收尾本轮结果；可继续对话让 AI 补充完善',
+    // 看板 953: 主对话回复正常收敛但被 provider max_tokens 上限截断(finish_reason=length)的气泡尾部弱提示条
+    truncatedByMaxTokens: '⚠ 输出因长度上限截断，内容可能不完整',
+    // 截断提示条的继续输出按钮
+    continueOutput: '继续输出',
+    // 「继续输出」注入的续写指令(作为新 user 消息发送,要求衔接勿重复)
+    continueOutputPrompt: '从上一条回复的中断处继续输出，直接衔接，勿重复已输出的内容。',
     queueFull: '待发送队列已满（最多 {limit} 条）',
     forceSendConfirm: 'AI 生成似乎卡住了。是否强制结束当前生成并发送消息？（可能中断正在生成的回复）',
     forceSendBtn: '强制发送',
@@ -136,15 +142,32 @@ export default {
         midstream: '响应流中断',
       },
       unknownKind: '任务停靠',
+      // U5 异常态话术公式第 3 要素「已保留什么」:前端常驻兜底——后端各停靠路径 reason 措辞不齐
+      // (fatal 非超时类/external_change 未带保留信息),卡面统一保证公式对全部 stop_kind 成立
+      preserved: '对话上下文与已完成进度已完整保留,随时可继续',
       stopArchive: '停止归档',
       stoppedToast: '任务已停止,不再续跑',
       archiveFailed: '停止失败:{msg}',
     },
     // 无人值守任务进度(AiTaskProgress,设计 §4.3):TopBar 轻量进度条,静默期可见「在干活」
+    // U2 三层补充:stage=内容层阶段标签(§3.3 工具→语义映射,未命中工具直显工具名不走此表);
+    // longRun/silent=静默期三档(§3.4),只给判断材料不自动判死
     taskProgress: {
       running: '进行中 · 第 {round}/{max} 轮 · {tokens} tokens',
       unlimited: '进行中 · 第 {round} 轮 · {tokens} tokens',
       barHint: '第 {round}/{max} 轮',
+      steps: '已完成 {n} 步',
+      stage: {
+        read: '阅读检索中',
+        edit: '编辑写入中',
+        run: '执行验证中',
+        agent: '子代理执行中',
+        knowledge: '知识与任务操作中',
+      },
+      longRun: '耗时较久',
+      longRunHint: '当前步骤耗时较久（大工具/子代理运行中）',
+      silentShort: '{n} 分钟无活动',
+      silentWarn: '距上次活动 {n} 分钟（最后动作：{tool}）。仍在运行，可查看轨迹；疑似卡死可停止。',
     },
     // 无人值守边界配置卡(P1 总装简化:授权面收敛单一档,可调项只剩预算两项)
     taskBoundary: {

src/i18n/zh-CN/aiChat.ts

@@ -113,6 +113,7 @@ export default {
     examplePrompt4: '生成一个项目灵感',
     goToSettings: '去设置',
     goToSettingsHint: '分离窗口无法直接打开设置，请在主窗口的设置页配置 AI 提供商',
+    goToSettingsOpened: '已在主窗口打开设置页',
     goToTaskHint: '分离窗口无法直接打开任务，请在主窗口查看任务详情',
 
     // ── Provider 切换反馈──
@@ -125,6 +126,8 @@ export default {
     // ── 待发送队列 ──
     queueTitle: '待发送 {n}/10',
     clearQueue: '清空',
+    // UI-4(看板 943):清空队列二次确认(一键全清不可恢复)
+    clearQueueConfirm: '确定清空全部待发送消息？清空后不可恢复',
     // 队列项编辑 + 立即发送
     queueEdit: '编辑',
     queueSave: '保存',
@@ -340,5 +343,26 @@ export default {
     editEmptyBlocked: '编辑内容为空,已取消提交',
     skillNoMatchQuery: '未找到技能「{q}」,已阻止发送',
     backfillOverwriteConfirm: '输入框已有未发送草稿,回填将覆盖,确定继续?',
+
+    // ── 待发箱/定时消息(OutboxDrawer + ChatInput ⏰ 入口,设计 2026-10-10 §3.4)──
+    outboxEntry: '定时消息',
+    outboxTitle: '定时与待发',
+    outboxEmpty: '暂无定时或待发消息',
+    outboxNewTitle: '新建定时消息',
+    outboxNewPlaceholder: '输入要定时发送的内容',
+    outboxTimeLabel: '发送时间',
+    outboxCreate: '定时发送',
+    outboxReschedule: '改期',
+    outboxSendNow: '立即发送',
+    outboxCancel: '取消',
+    outboxQueueAsap: '尽快',
+    outboxStatusPending: '待发',
+    outboxStatusSending: '发送中',
+    outboxStatusSent: '已发送',
+    outboxStatusFailed: '失败',
+    outboxStatusCanceled: '已取消',
+    outboxActionFail: '操作失败,请重试',
+    outboxDuePast: '定时时间必须晚于当前时刻',
+    outboxNeedConv: '无目标会话,请先创建会话',
   },
 }