diff --git a/src-tauri/capabilities/default.json b/src-tauri/capabilities/default.json
new file mode 100644
index 0000000..df30d26
--- /dev/null
+++ b/src-tauri/capabilities/default.json
@@ -0,0 +1,12 @@
+{
+  "$schema": "../gen/schemas/desktop-schema.json",
+  "identifier": "default",
+  "description": "u-ppt 主窗口默认能力：IPC 命令调用 + AI 流式事件订阅",
+  "windows": ["main"],
+  "permissions": [
+    "core:default",
+    "core:event:default",
+    "core:event:allow-listen",
+    "core:event:allow-unlisten"
+  ]
+}
diff --git a/src-tauri/src/commands.rs b/src-tauri/src/commands.rs
index 26a97e5..0ad4196 100644
--- a/src-tauri/src/commands.rs
+++ b/src-tauri/src/commands.rs
@@ -205,8 +205,11 @@ pub async fn ai_proxy_stream(
         .map_err(|e| format!("请求失败: {}", e))?;
 
     if !resp.status().is_success() {
+        let status = resp.status().as_u16();
         let text = resp.text().await.unwrap_or_default();
-        return Ok(text);
+        // 错误体带状态码前缀，前端按「接口返回 N」格式提示
+        let _ = app.emit("ai-error", &text);
+        return Err(format!("__HTTP_{}__{}", status, text));
     }
 
     // 流式读取 SSE
diff --git a/src/components/ai/AiPanel.vue b/src/components/ai/AiPanel.vue
index a05d34b..e63ec48 100644
--- a/src/components/ai/AiPanel.vue
+++ b/src/components/ai/AiPanel.vue
@@ -3,74 +3,14 @@
   发送/停止/生成整套/润色本页/流式渲染/操作应用
   ===================================================================== -->
 <script setup lang="ts">
-import { ref, nextTick, computed, watch } from 'vue'
+import { ref, nextTick, computed, watch, onUnmounted } from 'vue'
 import type { ChatMessage, AiOp, Slide } from '../../core/types'
 import { store } from '../../core/store'
 import { generate, polish, chat, beautifyPage, isConfigured } from '../../core/ai'
 import { elementTypes } from '../../core/sample'
+import { renderMd } from '../../core/markdown'
 import OutlinePanel from './OutlinePanel.vue'
 
-/* ---------- 消息内容渲染：提取 JSON 代码块并格式化 ---------- */
-
-function escapeHtml(s: string): string {
-  return s.replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;')
-}
-
-/** 尝试把 JSON 字符串格式化（缩进），失败则返回原文 */
-function tryFormatJson(raw: string): string {
-  try { return JSON.stringify(JSON.parse(raw), null, 2) } catch (e) { return raw }
-}
-
-/** 对格式化后的 JSON 做轻量语法着色（不引依赖，纯正则） */
-function highlightJson(code: string): string {
-  return escapeHtml(code)
-    .replace(/("(?:\\.|[^"\\])*"\s*:)/g, '<span class="jk">$1</span>')
-    .replace(/:\s*("(?:\\.|[^"\\])*")/g, ': <span class="js">$1</span>')
-    .replace(/:\s*(-?\d+\.?\d*)/g, ': <span class="jn">$1</span>')
-    .replace(/:\s*(true|false|null)/g, ': <span class="jb">$1</span>')
-}
-
-interface RenderedPart {
-  type: 'text' | 'code'
-  html: string
-  lang?: string
-}
-
-/** 把消息内容拆分为文本段 + JSON 代码块段 */
-function renderMessageContent(content: string): RenderedPart[] {
-  if (!content) return []
-  const parts: RenderedPart[] = []
-  // 匹配 ```json ... ``` 或 ``` ... ``` 代码块
-  const codeBlockRe = /```(\w*)\n?([\s\S]*?)```/g
-  let lastIdx = 0
-  let m: RegExpExecArray | null
-  while ((m = codeBlockRe.exec(content)) !== null) {
-    // 代码块前的文本
-    if (m.index > lastIdx) {
-      const text = content.slice(lastIdx, m.index)
-      if (text.trim()) parts.push({ type: 'text', html: escapeHtml(text) })
-    }
-    const lang = m[1] || ''
-    const code = m[2]
-    const trimmed = code.trim()
-    // JSON 代码块 → 格式化 + 着色
-    if (!lang || lang === 'json') {
-      parts.push({ type: 'code', lang: 'json', html: highlightJson(tryFormatJson(trimmed)) })
-    } else {
-      parts.push({ type: 'code', lang, html: escapeHtml(trimmed) })
-    }
-    lastIdx = m.index + m[0].length
-  }
-  // 尾部文本
-  if (lastIdx < content.length) {
-    const text = content.slice(lastIdx)
-    if (text.trim()) parts.push({ type: 'text', html: escapeHtml(text) })
-  }
-  // 没有代码块 → 返回整个为文本段
-  if (!parts.length) parts.push({ type: 'text', html: escapeHtml(content) })
-  return parts
-}
-
 const emit = defineEmits<{
   (e: 'busy-change', busy: boolean): void
   (e: 'toast', msg: string): void
@@ -131,6 +71,8 @@ function persistChat() {
 function setBusy(b: boolean) {
   busy.value = b
   emit('busy-change', b)
+  // 停止/完成后焦点回输入框，继续对话无需重新点击
+  if (!b) nextTick(() => inputEl.value?.focus())
 }
 
 function scrollBottom() {
@@ -159,10 +101,20 @@ function addPersisted(role: ChatMessage['role'], text: string) {
   scrollBottom()
 }
 
+/** 在跑的流式气泡集（组件卸载时统一停 rAF） */
+const activeStreams = new Set<StreamCtrl>()
+
 /** 创建一条流式气泡（返回控制器对象） */
 interface StreamCtrl {
   msg: RenderMsg
   started: boolean
+  /** 打字机缓冲：网络 delta 先入队，rAF 匀速释放渲染（参考 devflow 流式体验） */
+  buffer: string
+  shown: string
+  rafId: number | null
+  flushPending: boolean
+  /** 冲刷完成回调（持久化等收尾动作） */
+  settled?: (() => void) | null
 }
 function streamBubble(placeholder?: string): StreamCtrl {
   const msg: RenderMsg = {
@@ -173,27 +125,74 @@ function streamBubble(placeholder?: string): StreamCtrl {
   }
   renderMsgs.value.push(msg)
   scrollBottom()
-  return { msg, started: !placeholder }
+  const ctrl: StreamCtrl = { msg, started: !placeholder, buffer: '', shown: '', rafId: null, flushPending: false }
+  activeStreams.add(ctrl)
+  return ctrl
 }
+
+/** rAF 匀速释放缓冲：每帧最多放 CHARS_PER_FRAME 字符，网络抖动时也不突兀蹦字 */
+const CHARS_PER_FRAME = 4
+
+function pumpTypewriter(s: StreamCtrl) {
+  if (s.rafId != null) return
+  s.rafId = requestAnimationFrame(() => {
+    s.rafId = null
+    if (!s.buffer.length) {
+      // 缓冲排空：流已结束则收尾，否则待下一批 delta
+      if (s.flushPending) {
+        s.msg.streaming = false
+        s.flushPending = false
+        s.settled?.()
+        s.settled = null
+      }
+      return
+    }
+    // 每帧固定放 4 字符 ≈ 240 字符/秒，网络抖动时也保持匀速观感
+    const take = Math.min(s.buffer.length, CHARS_PER_FRAME)
+    s.shown += s.buffer.slice(0, take)
+    s.buffer = s.buffer.slice(take)
+    s.msg.content = s.shown
+    scrollBottom()
+    if (s.buffer.length || s.flushPending) pumpTypewriter(s)
+  })
+}
+
 function streamOnVisible(s: StreamCtrl) {
   return (t: string) => {
-    if (!s.started) { s.msg.content = ''; s.started = true }
-    s.msg.content += t
-    scrollBottom()
+    if (!s.started) { s.msg.content = ''; s.shown = ''; s.started = true }
+    s.buffer += t
+    pumpTypewriter(s)
   }
 }
-function streamDone(s: StreamCtrl) {
-  s.msg.streaming = false
-  if (!s.msg.content) s.msg.content = ''
+/** 流结束：标记待冲刷（缓冲剩余字符继续匀速放完再收尾），可选收尾回调 */
+function streamFlush(s: StreamCtrl, onSettled?: () => void) {
+  s.settled = onSettled || null
+  s.flushPending = true
+  if (!s.buffer.length) {
+    // 缓冲已空：立即收尾
+    s.msg.streaming = false
+    s.flushPending = false
+    s.settled?.()
+    s.settled = null
+    return
+  }
+  pumpTypewriter(s)
+}
+function streamDone(s: StreamCtrl, onSettled?: () => void) {
+  streamFlush(s, onSettled)
 }
 function streamSetText(s: StreamCtrl, txt: string) {
+  // 外部直接设置最终文本（跳过打字机，用于错误/提示）
+  if (s.rafId != null) { cancelAnimationFrame(s.rafId); s.rafId = null }
+  s.buffer = ''
+  s.shown = txt
   s.msg.content = txt
   s.msg.streaming = false
+  s.flushPending = false
 }
 function streamError(s: StreamCtrl, msg: string) {
+  streamSetText(s, '⚠ ' + msg)
   s.msg.error = true
-  s.msg.content = '⚠ ' + msg
-  s.msg.streaming = false
 }
 function streamTag(s: StreamCtrl, txt: string) {
   if (txt) s.msg.tag = txt
@@ -215,6 +214,11 @@ function applyOp(op: AiOp, lockedIdx: number): string {
     let t = op.target != null ? op.target : lockedIdx
     t = Math.max(0, Math.min(t, store.getCount() - 1))
     store.replaceSlide(t, slides[0])
+    // 目标页与发送时页不同：跳转过去让用户直接看到改动
+    if (t !== lockedIdx) {
+      store.setCurrentIndex(t)
+      return '已更新第 ' + (t + 1) + ' 页并跳转'
+    }
     return '已更新第 ' + (t + 1) + ' 页'
   }
   return ''
@@ -230,10 +234,7 @@ function persistStream(s: StreamCtrl) {
 function onSend() {
   if (busy.value) return
   const text = inputText.value.trim()
-  if (!text) {
-    if (!isConfigured()) { emit('open-settings'); return }
-    return
-  }
+  if (!text) return
   if (!isConfigured()) { toast('请先配置 API Key'); emit('open-settings'); return }
   inputText.value = ''
   runChat(text)
@@ -241,7 +242,11 @@ function onSend() {
 
 async function runChat(input: string) {
   addPersisted('user', input)
-  const history = chatLog.value.slice(-10).map(m => ({ role: m.role, content: m.content }))
+  // history 发送前剥离系统追加的 [✓ ...] 标记行，避免污染 AI 上下文
+  const history = chatLog.value.slice(-10).map(m => ({
+    role: m.role,
+    content: m.content.replace(/\n?\[✓[^\]]*\]$/g, '')
+  }))
   const idx0 = store.getCurrentIndex()
   const stream = streamBubble()
 
@@ -256,18 +261,24 @@ async function runChat(input: string) {
       signal: abortCtrl.signal,
       selectedElement: selectedEl.value
     })
-    streamDone(stream)
-    if (!r.reply) {
-      streamSetText(stream, '(已完成)')
-    }
+    // op 应用与 tag 立即执行；持久化等打字机冲刷完成（内容已齐）再落
     if (r.op && r.op.action !== 'answer' && r.op.slides.length) {
       const applied = applyOp(r.op, idx0)
       streamTag(stream, applied)
     }
-    persistStream(stream)
+    streamDone(stream, () => {
+      if (!stream.msg.content) streamSetText(stream, '(已完成)')
+      persistStream(stream)
+    })
   } catch (e: any) {
-    if (e?.name === 'AbortError') streamSetText(stream, '（已停止）')
-    else streamError(stream, e?.message || String(e))
+    if (e?.name === 'AbortError') {
+      // 停止：放弃缓冲中未显示的部分之外，把已显示内容立即固化
+      if (stream.rafId != null) { cancelAnimationFrame(stream.rafId); stream.rafId = null }
+      stream.msg.streaming = false
+      stream.msg.content = stream.shown ? stream.shown + '\n（已停止）' : '（已停止）'
+    } else {
+      streamError(stream, e?.message || String(e))
+    }
     persistStream(stream)
   } finally {
     setBusy(false); abortCtrl = null
@@ -392,6 +403,14 @@ defineExpose({ isBusy, focus })
 /* ---------- 初始化：从持久化记录重建 ---------- */
 rebuildFromChatLog()
 
+/* 组件卸载：停掉所有在跑的打字机 rAF */
+onUnmounted(() => {
+  for (const s of activeStreams) {
+    if (s.rafId != null) cancelAnimationFrame(s.rafId)
+  }
+  activeStreams.clear()
+})
+
 /* ---------- 会话切换：chatId 变化时重新加载对话 ---------- */
 watch(currentChatId, () => {
   chatLog.value = loadChat()
@@ -427,10 +446,8 @@ watch(currentChatId, () => {
       <div v-for="m in renderMsgs" :key="m.key" class="msg" :class="m.role">
         <div class="bubble" :class="{ error: m.error }">
           <span v-if="m.tag" class="diff-tag">✓ {{ m.tag }}</span>
-          <template v-for="(p, pi) in renderMessageContent(m.content)" :key="pi">
-            <pre v-if="p.type === 'code'" class="msg-code" :data-lang="p.lang"><code v-html="p.html"></code></pre>
-            <span v-else class="stream-text" v-html="p.html"></span>
-          </template>
+          <!-- Markdown 整段渲染（含代码块/列表/加粗等，经 DOMPurify 消毒） -->
+          <div v-if="m.content" class="md-body" v-html="renderMd(m.content)"></div>
           <span v-if="m.streaming" class="cursor"></span>
         </div>
       </div>
@@ -443,7 +460,6 @@ watch(currentChatId, () => {
         id="chatInput"
         rows="3"
         placeholder="输入指令，回车发送（Shift+Enter 换行）"
-        :disabled="busy"
         @keydown="onInputKeydown"
       ></textarea>
       <div class="btns">
diff --git a/src/core/ai.ts b/src/core/ai.ts
index 31828b4..340b05b 100644
--- a/src/core/ai.ts
+++ b/src/core/ai.ts
@@ -10,6 +10,7 @@ import type { AiOp, ChartItem, Deck, Slide, SlideElement, ElementStyle, RichLine
 import { elementTypes, uid } from './sample'
 import { store } from './store'
 import { normSegments } from './richtext'
+import { isTauri, aiProxy, aiProxyStream } from './bridge'
 
 const SEP = '%%PPT_JSON%%'       // 对话模式中，自然语言回复与结构化操作的分隔标记
 const VALID_TYPES = ['title', 'text', 'list', 'stat', 'quote', 'image', 'shape', 'chart', 'card', 'table', 'code', 'formula'] as const
@@ -103,20 +104,36 @@ const SYS_CHAT =
  * ============================================================ */
 function deckContext(currentIdx: number, selectedEl?: SlideElement | null): string {
   const deck: Deck = store.getDeck()
+  /** 图片/超长 content 脱敏：base64 会撑爆上下文，替换为占位说明 */
+  const sanitizeEl = (el: SlideElement): SlideElement => {
+    if (el.type === 'image') {
+      const src = el.content || ''
+      const desc = src.startsWith('data:')
+        ? `[图片 base64 ${Math.round(src.length / 1024)}KB]`
+        : (src ? `[图片URL: ${src.slice(0, 80)}${src.length > 80 ? '…' : ''}]` : '[空图片]')
+      return { ...el, content: desc }
+    }
+    if (el.content && el.content.length > 2000) {
+      return { ...el, content: el.content.slice(0, 2000) + `…[截断，共${el.content.length}字符]` }
+    }
+    return el
+  }
   const lines = ['当前主题: ' + deck.theme + '，共 ' + deck.slides.length + ' 页。']
   deck.slides.forEach((s, i) => {
     const types = s.elements.map(e => e.type).join('/')
     const head = (s.elements[0] && s.elements[0].type === 'title') ? (' 标题:"' + (s.elements[0].content || '') + '"') : ''
     lines.push('第' + (i + 1) + '页 [' + types + ']' + head + (i === currentIdx ? '  ← 当前页' : ''))
   })
-  lines.push('\n当前页(第' + (currentIdx + 1) + '页)完整JSON:\n' + JSON.stringify(store.currentSlide.value))
+  const cur = store.currentSlide.value
+  const curSanitized = { ...cur, elements: cur.elements.map(sanitizeEl) }
+  lines.push('\n当前页(第' + (currentIdx + 1) + '页)完整JSON:\n' + JSON.stringify(curSanitized))
   // 选中元素上下文：让 AI 知道用户正在编辑哪个元素，对话直接围绕它
   if (selectedEl) {
     const typeLabel = selectedEl.type
-    const preview = (selectedEl.content || '').slice(0, 100)
+    const preview = (selectedEl.content || '').replace(/\n/g, ' ').slice(0, 100)
     lines.push('\n【用户当前选中的元素】（第' + (currentIdx + 1) + '页）')
     lines.push('类型: ' + typeLabel + '，内容预览: "' + preview + '"')
-    lines.push('完整JSON: ' + JSON.stringify(selectedEl))
+    lines.push('完整JSON: ' + JSON.stringify(sanitizeEl(selectedEl)))
     lines.push('用户接下来的对话默认针对此元素，除非明确说整页/整套。')
   }
   return lines.join('\n')
@@ -145,9 +162,55 @@ function apiUrl(cfg: { proxy: string; base: string }): string {
   return (cfg.proxy || cfg.base || '').replace(/\/+$/, '')
 }
 
+/**
+ * 桌面流式请求：Rust ai_proxy_stream 转发 SSE 原始 chunk，
+ * 事件桥逐 chunk 喂 SSE 解析器（与浏览器路径共用 createSSESink，保留打字机效果与 SEP 协议）。
+ * 返回 null 表示桌面桥不可用（回退浏览器 fetch 路径）。
+ */
+async function desktopStream(
+  url: string, apiKey: string, body: Record<string, unknown>,
+  extractDelta: (obj: any) => string | null, opts: StreamOpts
+): Promise<{ json: any; reply: string; op: any } | null> {
+  const sink = createSSESink(extractDelta, opts)
+  const full = await aiProxyStream(url, apiKey, JSON.stringify(body), (chunk) => sink.push(chunk))
+  if (!full) return null
+  if (full.status === 0 && full.error) throw new Error('桌面代理请求失败：' + full.error)
+  // 非 200：按 consumeStream 相同格式报错（含智谱错误码翻译）
+  if (full.status >= 400) {
+    let msg = '接口返回 ' + full.status
+    const hint = ERROR_HINTS_BY_STATUS[full.status]
+    try {
+      const err = JSON.parse(full.body).error
+      if (err) {
+        msg += ' [' + err.code + '] ' + (err.message || '')
+        if (err.code && ERROR_HINTS[err.code]) msg += '\n\n💡 ' + ERROR_HINTS[err.code]
+      } else msg += ' ' + (full.body || '').slice(0, 200)
+    } catch { msg += ' ' + (full.body || '').slice(0, 200) }
+    if (hint && !msg.includes('💡')) msg += '\n\n💡 ' + hint
+    throw new Error(msg)
+  }
+  return sink.finish()
+}
+
+/**
+ * 统一 POST 入口：
+ * - 桌面（Tauri）→ Rust 侧 reqwest 代理（无 CORS）
+ * - Web → 浏览器 fetch（支持 CORS 的网关如智谱可直连）
+ * 桌面流式请求的 SSE 由 bridge 内部转为 onVisible 增量回调
+ */
 async function postJSON(url: string, headers: Record<string, string>, body: unknown, signal?: AbortSignal): Promise<Response> {
+  const bodyStr = JSON.stringify(body)
+  // 桌面代理路径：headers 中取鉴权（Authorization Bearer 或 x-api-key）
+  if (isTauri()) {
+    const apiKey = headers['Authorization']?.replace(/^Bearer\s+/i, '') || headers['x-api-key'] || ''
+    const proxied = await aiProxy(url, apiKey, bodyStr)
+    if (proxied) {
+      if (proxied.status === 0 && proxied.error) throw new Error('桌面代理请求失败：' + proxied.error)
+      return new Response(proxied.body, { status: proxied.status })
+    }
+  }
   try {
-    return await fetch(url, { method: 'POST', headers, body: JSON.stringify(body), signal })
+    return await fetch(url, { method: 'POST', headers, body: bodyStr, signal })
   } catch (e: any) {
     if (e.name === 'AbortError') throw e
     throw new Error('请求失败（可能是 CORS 跨域拦截）。可在设置中配置"代理 URL"。\n' + e.message)
@@ -159,6 +222,16 @@ async function runOpenAI(messages: Message[], opts: StreamOpts, cfg: ReturnType<
   const url = apiUrl(cfg) + '/chat/completions'
   const body: Record<string, unknown> = { model: cfg.model || 'glm-4.6', messages, stream: true, temperature: 0.75 }
   if (opts.jsonMode) body.response_format = { type: 'json_object' }
+
+  // 桌面流式：Rust 事件桥推送 SSE 增量（保留打字机效果），完成后一次性解析
+  if (isTauri() && opts.onVisible) {
+    const extract = (obj: any) => {
+      const ch = obj.choices && obj.choices[0]
+      return (ch && ch.delta && ch.delta.content) || null
+    }
+    const r = await desktopStream(url, cfg.key, body, extract, opts)
+    if (r) return r
+  }
   const resp = await postJSON(url,
     { 'Content-Type': 'application/json', 'Authorization': 'Bearer ' + cfg.key }, body, opts.signal)
   return consumeStream(resp, (obj: any) => {
@@ -184,6 +257,16 @@ async function runAnthropic(messages: Message[], opts: StreamOpts, cfg: ReturnTy
     temperature: 0.75
   }
   if (sysParts.length) body.system = sysParts.join('\n\n')
+
+  // 桌面流式（同 OpenAI 路径）
+  if (isTauri() && opts.onVisible) {
+    const extract = (obj: any) => {
+      if (obj.type === 'content_block_delta' && obj.delta) return obj.delta.text || null
+      return null
+    }
+    const r = await desktopStream(url, cfg.key, body, extract, opts)
+    if (r) return r
+  }
   const headers = {
     'Content-Type': 'application/json',
     'x-api-key': cfg.key,
@@ -197,7 +280,7 @@ async function runAnthropic(messages: Message[], opts: StreamOpts, cfg: ReturnTy
   }, opts)
 }
 
-// 通用 SSE 消费
+// 通用 SSE 消费（浏览器 fetch 流）
 async function consumeStream(resp: Response, extractDelta: (obj: any) => string | null, opts: StreamOpts): Promise<{ json: any; reply: string; op: any }> {
   if (!resp.ok) {
     let t = ''; try { t = await resp.text() } catch (e) {}
@@ -215,6 +298,20 @@ async function consumeStream(resp: Response, extractDelta: (obj: any) => string
   }
   const reader = resp.body!.getReader()
   const dec = new TextDecoder('utf-8')
+  const sseSink = createSSESink(extractDelta, opts)
+  for (;;) {
+    const chunk = await reader.read()
+    if (chunk.done) break
+    sseSink.push(dec.decode(chunk.value, { stream: true }))
+  }
+  return sseSink.finish()
+}
+
+/**
+ * SSE 解析核心：数据源无关（fetch 流 / Tauri 事件流通用）。
+ * push() 喂原始 chunk（可能含多行/半行），finish() 返回与 consumeStream 相同结构。
+ */
+function createSSESink(extractDelta: (obj: any) => string | null, opts: StreamOpts) {
   let sseBuf = ''
   let full = ''
   let pending = ''
@@ -241,39 +338,41 @@ async function consumeStream(resp: Response, extractDelta: (obj: any) => string
   }
   function emit(text: string) { if (opts.onVisible) opts.onVisible(text) }
 
-  for (;;) {
-    const chunk = await reader.read()
-    if (chunk.done) break
-    sseBuf += dec.decode(chunk.value, { stream: true })
-    const lines = sseBuf.split('\n')
-    sseBuf = lines.pop()!
-    for (let k = 0; k < lines.length; k++) {
-      const line = lines[k].trim()
-      if (!line || line.indexOf('data:') !== 0) continue
-      const payload = line.slice(5).trim()
-      if (!payload || payload === '[DONE]') continue
-      let obj: any; try { obj = JSON.parse(payload) } catch (e) { continue }
-      const delta = extractDelta(obj)
-      if (delta != null) feed(delta)
-    }
-  }
-  // 末帧残留 data: 行补解析
-  const tail = sseBuf.trim()
-  if (tail.indexOf('data:') === 0) {
-    const tp = tail.slice(5).trim()
-    if (tp && tp !== '[DONE]') {
-      let to: any; try { to = JSON.parse(tp) } catch (e) { to = null }
-      if (to) { const td = extractDelta(to); if (td != null) feed(td) }
-    }
-  }
-  if (!opts.jsonMode && !sepMode && pending) emit(pending)
-
-  if (opts.jsonMode) return { json: tryParse(full), reply: '', op: null }
-  const parts = full.split(SEP)
   return {
-    json: null,
-    reply: (parts[0] || '').trim(),
-    op: parts.length > 1 ? tryParse(parts.slice(1).join(SEP)) : null
+    /** 喂一个网络 chunk（SSE 帧文本，可跨界） */
+    push(chunk: string) {
+      sseBuf += chunk.replace(/\r/g, '')
+      const lines = sseBuf.split('\n')
+      sseBuf = lines.pop() || ''
+      for (const line of lines) {
+        const l = line.trim()
+        if (!l || l.indexOf('data:') !== 0) continue
+        const payload = l.slice(5).trim()
+        if (!payload || payload === '[DONE]') continue
+        let obj: any; try { obj = JSON.parse(payload) } catch (e) { continue }
+        const delta = extractDelta(obj)
+        if (delta != null) feed(delta)
+      }
+    },
+    /** 流结束：解析残留行并汇总 */
+    finish(): { json: any; reply: string; op: any } {
+      const tail = sseBuf.trim()
+      if (tail.indexOf('data:') === 0) {
+        const tp = tail.slice(5).trim()
+        if (tp && tp !== '[DONE]') {
+          let to: any; try { to = JSON.parse(tp) } catch (e) { to = null }
+          if (to) { const td = extractDelta(to); if (td != null) feed(td) }
+        }
+      }
+      if (!opts.jsonMode && !sepMode && pending) emit(pending)
+      if (opts.jsonMode) return { json: tryParse(full), reply: '', op: null }
+      const parts = full.split(SEP)
+      return {
+        json: null,
+        reply: (parts[0] || '').trim(),
+        op: parts.length > 1 ? tryParse(parts.slice(1).join(SEP)) : null
+      }
+    }
   }
 }
 
@@ -555,16 +654,18 @@ export async function generateImage(opts: { prompt: string; signal?: AbortSignal
   const cfg = store.getCfg()
   const imgBase = (cfg.imgBase || cfg.base || '').replace(/\/+$/, '')
   const imgKey = cfg.imgKey || cfg.key
-  const imgModel = cfg.imgModel || 'dall-e-3'
+  // 图像模型默认值跟随服务商：智谱用 cogview，其余走 OpenAI 兼容默认
+  const imgModel = cfg.imgModel || (imgBase.includes('bigmodel.cn') ? 'cogview-3-plus' : 'dall-e-3')
   if (!imgKey) throw new Error('未配置 API Key，无法生成图片。')
   const url = imgBase + '/images/generations'
-  const body = {
+  const body: Record<string, unknown> = {
     model: imgModel,
     prompt: opts.prompt,
     n: 1,
-    size: '1024x1024',
-    response_format: 'b64_json'
+    size: '1024x1024'
   }
+  // 智谱 CogView 不支持 response_format 参数（会报错），仅对 OpenAI 兼容网关传 b64_json
+  if (!imgBase.includes('bigmodel.cn')) body.response_format = 'b64_json'
   let resp: Response
   try {
     resp = await fetch(url, {
@@ -588,6 +689,25 @@ export async function generateImage(opts: { prompt: string; signal?: AbortSignal
     ? 'data:image/png;base64,' + item.b64_json
     : (item.url || '')
   if (!dataUrl) throw new Error('图片生成返回无图像数据。')
+  // 返回的是临时 URL 而非 base64：链接会过期导致日后裂图，先抓回本地存 data URL
+  if (/^https?:/i.test(dataUrl)) {
+    try {
+      const imgResp = await fetch(dataUrl, { signal: opts.signal })
+      if (imgResp.ok) {
+        const blob = await imgResp.blob()
+        const b64 = await new Promise<string>((resolve, reject) => {
+          const r = new FileReader()
+          r.onload = () => resolve(r.result as string)
+          r.onerror = () => reject(r.error)
+          r.readAsDataURL(blob)
+        })
+        return { url: b64, revisedPrompt: item.revised_prompt }
+      }
+    } catch (e: any) {
+      if (e?.name === 'AbortError') throw e
+      // 抓取失败（常见为 CORS）——退回原 URL，仍可显示但有过期风险
+    }
+  }
   return { url: dataUrl, revisedPrompt: item.revised_prompt }
 }
 
@@ -597,6 +717,17 @@ export function isImageConfigured(): boolean {
   return !!(cfg.imgKey || cfg.key)
 }
 
+/** 检查写入图片后是否超 localStorage 上限（约 5MB 字符），返回错误文案或 null */
+export function checkImageQuota(dataUrl: string): string | null {
+  const QUOTA_CHARS = 4_500_000
+  let deckChars = 0
+  try { deckChars = JSON.stringify(store.getDeck()).length } catch (e) { /* ignore */ }
+  if (deckChars + dataUrl.length > QUOTA_CHARS) {
+    return '图片写入后将超本地存储上限（约 5MB），已取消。请删除部分旧图，或导出 JSON 备份后清理文库'
+  }
+  return null
+}
+
 /* ---------- 4. AI 主题/配色建议 ---------- */
 
 export async function suggestTheme(opts: { topic: string; signal?: AbortSignal }): Promise<ThemeSuggestion> {
diff --git a/src/core/bridge.ts b/src/core/bridge.ts
index 5d72c51..42c1101 100644
--- a/src/core/bridge.ts
+++ b/src/core/bridge.ts
@@ -1,44 +1,71 @@
 /* =====================================================================
- * bridge.ts — Tauri IPC 安全调用桥接
- *
- * 安全检测 __TAURI__ 环境，动态 import @tauri-apps/api/core。
- * Web 模式下返回 mock，组件无需关心环境。
+ * bridge.ts — Tauri IPC 桥接（AI 代理专用）
+ * 桌面版（Tauri WebView）直连 AI 网关受 CORS 限制，走 Rust 侧 reqwest 代理；
+ * Web 版返回 null，调用方回退浏览器 fetch（智谱等支持 CORS 的网关直连可用）
  * ===================================================================== */
 
-/** 动态获取 invoke 函数 */
-let _invoke: ((cmd: string, args?: Record<string, unknown>) => Promise<unknown>) | null = null
+/** 是否运行在 Tauri 桌面环境 */
+export function isTauri(): boolean {
+  return typeof window !== 'undefined' && ('__TAURI__' in window || '__TAURI_INTERNALS__' in window)
+}
 
-async function ensureInvoke() {
-  if (_invoke) return true
-  if (typeof window === 'undefined' || !('__TAURI__' in window)) return false
+let _invoke: ((cmd: string, args?: Record<string, unknown>) => Promise<unknown>) | null = null
+let _invokeTried = false
+
+/** 懒加载 Tauri invoke（动态 import，Web 版不会打进包） */
+async function getInvoke() {
+  if (_invokeTried) return _invoke
+  _invokeTried = true
+  if (!isTauri()) return null
   try {
     const mod = await import('@tauri-apps/api/core')
     _invoke = mod.invoke
-    return true
-  } catch {
-    return false
-  }
+  } catch { _invoke = null }
+  return _invoke
 }
 
-/**
- * 安全调用 Tauri IPC 命令。
- * Web 模式下返回 undefined（调用方需处理）。
- */
-export async function invoke<T = unknown>(cmd: string, args?: Record<string, unknown>): Promise<T | undefined> {
-  const ok = await ensureInvoke()
-  if (!ok || !_invoke) return undefined
-  return _invoke(cmd, args) as Promise<T>
+export interface ProxyResp {
+  status: number
+  body: string
+  error?: string
 }
 
-/** 注册 Tauri 事件监听（用于 Rust → 前端推送） */
-export async function listen<T>(event: string, handler: (payload: T) => void): Promise<() => void> {
-  if (typeof window === 'undefined' || !('__TAURI__' in window)) {
-    return () => {} // 空注销函数
-  }
+/** 非流式代理：Rust 侧 fetch，返回 {status, body}；非桌面环境返回 null */
+export async function aiProxy(url: string, apiKey: string, body: string): Promise<ProxyResp | null> {
+  const invoke = await getInvoke()
+  if (!invoke) return null
   try {
-    const { listen } = await import('@tauri-apps/api/event')
-    return listen(event, (e: { payload: T }) => handler(e.payload))
-  } catch {
-    return () => {}
+    const r = await invoke('ai_proxy', { url, apiKey, body }) as ProxyResp
+    return r
+  } catch (e: any) {
+    return { status: 0, body: '', error: String(e?.message || e) }
+  }
+}
+
+/** 流式代理：Rust 侧 SSE，原始 chunk（含 "data: {...}" 帧）通过 onChunk 推送；
+ *  SSE 解析由 ai.ts 的 createSSESink 统一处理（与浏览器路径共用）。非桌面返回 null */
+export async function aiProxyStream(
+  url: string, apiKey: string, body: string,
+  onChunk: (raw: string) => void
+): Promise<ProxyResp | null> {
+  const invoke = await getInvoke()
+  if (!invoke) return null
+  try {
+    const eventMod = await import('@tauri-apps/api/event')
+    const unlisten = await eventMod.listen<string>('ai-delta', (ev) => {
+      onChunk(ev.payload || '')
+    })
+    try {
+      const full = await invoke('ai_proxy_stream', { url, apiKey, body }) as unknown as string
+      return { status: 200, body: full }
+    } finally {
+      unlisten()
+    }
+  } catch (e: any) {
+    // Rust 侧错误协议：__HTTP_<status>__<body>
+    const raw = String(e?.message || e)
+    const m = raw.match(/^__HTTP_(\d+)__/)
+    if (m) return { status: Number(m[1]), body: raw.slice(m[0].length) }
+    return { status: 0, body: '', error: raw }
   }
 }
diff --git a/src/core/markdown.ts b/src/core/markdown.ts
new file mode 100644
index 0000000..9b8fc42
--- /dev/null
+++ b/src/core/markdown.ts
@@ -0,0 +1,66 @@
+/* =====================================================================
+ * markdown.ts — AI 消息 Markdown 渲染（参考 devflow useMarkdown 设计）
+ * - marked 静态导入（u-ppt 包体敏感度低，无需动态加载）
+ * - DOMPurify.sanitize 防 XSS，可安全 v-html
+ * - 历史消息整段缓存（文本不变命中跳过 parse+sanitize）
+ * - JSON 代码块轻量着色（保留原有正则方案，不引 hljs 控包体）
+ * ===================================================================== */
+import { marked } from 'marked'
+import DOMPurify from 'dompurify'
+
+marked.setOptions({ gfm: true, breaks: true })
+
+/* ---------- JSON 轻量着色（与原 AiPanel 实现一致） ---------- */
+
+function escapeHtml(s: string): string {
+  return s.replace(/&/g, '&amp;').replace(/</g, '&lt;').replace(/>/g, '&gt;')
+}
+
+function tryFormatJson(raw: string): string {
+  try { return JSON.stringify(JSON.parse(raw), null, 2) } catch (e) { return raw }
+}
+
+/** JSON 代码着色（纯正则，无依赖） */
+function highlightJson(code: string): string {
+  return escapeHtml(code)
+    .replace(/("(?:\\.|[^"\\])*"\s*:)/g, '<span class="jk">$1</span>')
+    .replace(/:\s*("(?:\\.|[^"\\])*")/g, ': <span class="js">$1</span>')
+    .replace(/:\s*(-?\d+\.?\d*)/g, ': <span class="jn">$1</span>')
+    .replace(/:\s*(true|false|null)/g, ': <span class="jb">$1</span>')
+}
+
+/* ---------- code renderer：json 格式化+着色，其余转义 ---------- */
+const renderer = new marked.Renderer()
+renderer.code = function ({ text, lang }: { text: string; lang?: string }) {
+  const language = (lang || '').toLowerCase()
+  if (!language || language === 'json') {
+    return `<pre class="msg-code" data-lang="json"><code>${highlightJson(tryFormatJson(text.trim()))}</code></pre>`
+  }
+  return `<pre class="msg-code" data-lang="${escapeHtml(language)}"><code>${escapeHtml(text)}</code></pre>`
+}
+marked.use({ renderer })
+
+/* ---------- 缓存 + 主入口 ---------- */
+const _mdCache = new Map<string, string>()
+const MD_CACHE_LIMIT = 300
+
+/** 渲染 Markdown → 消毒 HTML（可安全 v-html）。失败降级纯文本转义。 */
+export function renderMd(text: string): string {
+  const hit = _mdCache.get(text)
+  if (hit != null) return hit
+  let html: string
+  try {
+    html = DOMPurify.sanitize(marked.parse(text) as string, {
+      ADD_ATTR: ['data-lang'],
+    })
+  } catch (e) {
+    html = '<p>' + escapeHtml(text).replace(/\n/g, '<br>') + '</p>'
+  }
+  if (_mdCache.size >= MD_CACHE_LIMIT) {
+    // 简单淘汰：清掉最早的键（Map 保持插入序）
+    const firstKey = _mdCache.keys().next().value
+    if (firstKey !== undefined) _mdCache.delete(firstKey)
+  }
+  _mdCache.set(text, html)
+  return html
+}
diff --git a/src/styles/chat.css b/src/styles/chat.css
index ba04021..254de60 100644
--- a/src/styles/chat.css
+++ b/src/styles/chat.css
@@ -73,6 +73,31 @@
 
 .msg .bubble .stream-text { white-space: pre-wrap; }
 
+/* Markdown 正文排版（marked 渲染后的气泡内容） */
+.msg .bubble .md-body > *:first-child { margin-top: 0; }
+.msg .bubble .md-body > *:last-child { margin-bottom: 0; }
+.msg .bubble .md-body p { margin: 6px 0; }
+.msg .bubble .md-body h1, .msg .bubble .md-body h2, .msg .bubble .md-body h3,
+.msg .bubble .md-body h4 { margin: 10px 0 6px; font-size: 14px; font-weight: 600; }
+.msg .bubble .md-body h1 { font-size: 16px; }
+.msg .bubble .md-body h2 { font-size: 15px; }
+.msg .bubble .md-body ul, .msg .bubble .md-body ol { margin: 6px 0; padding-left: 1.4em; }
+.msg .bubble .md-body li { margin: 3px 0; }
+.msg .bubble .md-body strong { font-weight: 600; }
+.msg .bubble .md-body code {
+  font-family: 'Consolas','Monaco','Courier New',monospace;
+  font-size: 12px; padding: 1px 5px; border-radius: 4px;
+  background: rgba(15,23,42,.08);
+}
+.msg.user .md-body code { background: rgba(255,255,255,.2); }
+.msg .bubble .md-body blockquote {
+  margin: 6px 0; padding: 2px 10px; border-left: 3px solid var(--ui-border);
+  color: var(--ui-muted);
+}
+.msg .bubble .md-body table { border-collapse: collapse; margin: 6px 0; font-size: 12px; }
+.msg .bubble .md-body th, .msg .bubble .md-body td { border: 1px solid var(--ui-border); padding: 4px 8px; }
+.msg .bubble .md-body hr { border: none; border-top: 1px solid var(--ui-border); margin: 8px 0; }
+
 /* 消息中的代码块 */
 .msg .msg-code {
   margin: 6px 0; padding: 10px 12px;
@@ -89,12 +114,18 @@
 .msg .msg-code .jn { color: #d97706; }                          /* number */
 .msg .msg-code .jb { color: #e11d48; font-weight: 600; }        /* boolean/null */
 
+/* 流式打字机光标（参考 devflow：细竖线 + 柔和步进闪烁） */
 .cursor {
-  display: inline-block; width: 7px; height: 1em;
-  background: currentColor; vertical-align: -2px; margin-left: 2px;
-  animation: blink 1s steps(2, start) infinite;
+  display: inline-block;
+  width: 2px; height: 14px;
+  background: var(--ui-primary, #4f46e5);
+  margin-left: 2px; vertical-align: text-bottom;
+  animation: blink 0.8s step-end infinite;
+}
+@keyframes blink {
+  0%, 100% { opacity: 1; }
+  50% { opacity: 0; }
 }
-@keyframes blink { 50% { opacity: 0; } }
 
 /* 输入栏 */
 .chat-input-bar {
diff --git a/vite.config.ts b/vite.config.ts
index f137c19..9cfeda0 100644
--- a/vite.config.ts
+++ b/vite.config.ts
@@ -6,9 +6,19 @@ export default defineConfig(({ mode }) => ({
   build: {
     chunkSizeWarningLimit: 800,
   },
-  // npm run dev（mode=web）→ 浏览器模式，端口 8080
+  // npm run dev:web（mode=web）→ 浏览器模式，端口 8080
   // npm run tauri:dev → 默认 mode，端口 5173
   server: mode === 'web'
-    ? { host: '127.0.0.1', port: 8080, open: true }
-    : { host: '127.0.0.1', port: 5173, strictPort: true },
+    ? {
+        host: '127.0.0.1', port: 8080, open: true,
+        // 与 tauri dev 共存时防止 watch Cargo 产物（EBUSY 崩 watcher）
+        watch: { ignored: ['**/src-tauri/target/**'] },
+      }
+    : {
+        host: '127.0.0.1', port: 5173, strictPort: true,
+        // 不 watch src-tauri/target：Cargo 编译期间写 dll 会触发 EBUSY 崩掉 watcher
+        watch: { ignored: ['**/src-tauri/target/**'] },
+      },
+  // 桌面模式 HMR 由 tauri-cli 管理，无需自动打开浏览器
+  clearScreen: false,
 }))