diff --git a/frontends/desktop/static/app.js b/frontends/desktop/static/app.js
index 737303d4..c8fa0114 100644
--- a/frontends/desktop/static/app.js
+++ b/frontends/desktop/static/app.js
@@ -329,7 +329,7 @@ const I18N = {
     'collab.timeMin': '{n} 分钟前',
     'collab.timeHr': '{n} 小时前',
     'collab.timeDay': '{n} 天前',
-    'page.token.title': 'Token 统计', 'page.token.sub': '每会话与累计的 token 用量及估算成本',
+    'page.token.title': 'Token 统计', 'page.token.sub': '每会话与累计的 token 用量及缓存率',
     'status.connecting': '连接中…', 'status.ready': '就绪', 'status.running': '运行中',
     'status.disconnected': '未连接', 'status.stopped': '已停止', 'status.idle': '空闲',
     'conv.emptyList': '暂无会话，点「＋ 新对话」开始', 'conv.defaultTitle': '新对话',
@@ -385,7 +385,7 @@ const I18N = {
     'proc.imbotWechat': 'imbot · 微信', 'proc.imbotDing': 'imbot · 钉钉', 'proc.scheduler': '定时任务调度',
     'cm.scheduling': '调度中', 'cm.running': '执行中', 'cm.idleSt': '空闲',
     'cm.master': '已派 3 子任务', 'cm.w1': '子任务：抓取数据', 'cm.w2': '子任务：复核结果', 'cm.sub': '等待派单',
-    'tok.total': '累计 token', 'tok.cost': '估算成本', 'tok.today': '今日 token', 'tok.tabAll': '聊天', 'tok.tabConductor': 'Conductor', 'tok.condTotal': 'Conductor 累计', 'tok.condCurrent': 'Conductor 本次', 'tok.condTip': 'Conductor 消耗的 token 不计入聊天累计 token 中', 'tok.disclaimer': '不同 API 网站的计费价格可能会有差异，请以实际网站为准。', 'tok.chartToggle': '趋势图',
+    'tok.total': '累计 token', 'tok.cost': '缓存率', 'tok.today': '今日 token', 'tok.tabAll': '聊天', 'tok.tabConductor': 'Conductor', 'tok.condTotal': 'Conductor 累计', 'tok.condCurrent': 'Conductor 本次', 'tok.condTip': 'Conductor 消耗的 token 不计入聊天累计 token 中', 'tok.disclaimer': '不同 API 网站的计费价格可能会有差异，请以实际网站为准。', 'tok.chartToggle': '趋势图',
     'tok.colSession': '会话', 'tok.colIn': '输入', 'tok.colOut': '输出', 'tok.colCacheW': '缓存写入', 'tok.colCache': '缓存读取', 'tok.colCost': '成本',
     'tok.from': '从', 'tok.to': '到', 'tok.reset': '重置', 'tok.noData': '暂无记录', 'tok.deleted': '此会话已删除',
     'tok.pricingUnknown': '⚠ 此模型计费规则尚未明确，按默认估算',
@@ -481,7 +481,7 @@ const I18N = {
     'collab.timeMin': '{n}m ago',
     'collab.timeHr': '{n}h ago',
     'collab.timeDay': '{n}d ago',
-    'page.token.title': 'Token usage', 'page.token.sub': 'Per-session and total token usage & estimated cost',
+    'page.token.title': 'Token usage', 'page.token.sub': 'Per-session and total token usage & cache rate',
     'status.connecting': 'Connecting…', 'status.ready': 'Ready', 'status.running': 'Running',
     'status.disconnected': 'Disconnected', 'status.stopped': 'Stopped', 'status.idle': 'Idle',
     'conv.emptyList': 'No chats yet — click “＋ New chat”', 'conv.defaultTitle': 'New chat',
@@ -537,7 +537,7 @@ const I18N = {
     'proc.imbotWechat': 'imbot · WeChat', 'proc.imbotDing': 'imbot · DingTalk', 'proc.scheduler': 'Scheduler',
     'cm.scheduling': 'Scheduling', 'cm.running': 'Running', 'cm.idleSt': 'Idle',
     'cm.master': 'Dispatched 3 subtasks', 'cm.w1': 'Subtask: fetch data', 'cm.w2': 'Subtask: review results', 'cm.sub': 'Waiting for tasks',
-    'tok.total': 'Total tokens', 'tok.cost': 'Est. cost', 'tok.today': 'Today tokens', 'tok.tabAll': 'Chat', 'tok.tabConductor': 'Conductor', 'tok.condTotal': 'Conductor Total', 'tok.condCurrent': 'Conductor Current', 'tok.condTip': 'Conductor tokens are not included in chat totals', 'tok.disclaimer': 'Pricing may vary by API provider. Please refer to the actual website.', 'tok.chartToggle': 'Trend',
+    'tok.total': 'Total tokens', 'tok.cost': 'Cache rate', 'tok.today': 'Today tokens', 'tok.tabAll': 'Chat', 'tok.tabConductor': 'Conductor', 'tok.condTotal': 'Conductor Total', 'tok.condCurrent': 'Conductor Current', 'tok.condTip': 'Conductor tokens are not included in chat totals', 'tok.disclaimer': 'Pricing may vary by API provider. Please refer to the actual website.', 'tok.chartToggle': 'Trend',
     'tok.colSession': 'Session', 'tok.colIn': 'Input', 'tok.colOut': 'Output', 'tok.colCacheW': 'Cache write', 'tok.colCache': 'Cache read', 'tok.colCost': 'Cost',
     'tok.from': 'From', 'tok.to': 'To', 'tok.reset': 'Reset', 'tok.noData': 'No records', 'tok.deleted': 'Session deleted',
     'tok.pricingUnknown': '⚠ Pricing not confirmed, using defaults',
@@ -3693,12 +3693,13 @@ function tokGetFiltered() {
 }
 
 function tokRenderStats(filtered, all) {
-  let total=0, cost=0;
-  filtered.forEach(r=>{total+=(r.input||0)+(r.output||0); cost+=parseFloat(estCost(r.input||0,r.output||0,r.model,r.cacheRead||0,r.cacheCreate||0));});
+  let total=0, totalInput=0, totalCacheRead=0, totalCacheCreate=0;
+  filtered.forEach(r=>{total+=(r.input||0)+(r.output||0)+(r.cacheRead||0)+(r.cacheCreate||0); totalInput+=(r.input||0); totalCacheRead+=(r.cacheRead||0); totalCacheCreate+=(r.cacheCreate||0);});
   if(tokTotalN) tokTotalN.textContent=fmtTok(total);
-  if(tokCostN) tokCostN.textContent='¥ '+cost.toFixed(1);
+  const cacheBase = totalInput + totalCacheRead + totalCacheCreate;
+  if(tokCostN) tokCostN.textContent= cacheBase > 0 ? (totalCacheRead / cacheBase * 100).toFixed(1) + '%' : '0%';
   const todayStart=new Date(); todayStart.setHours(0,0,0,0); const todayTs=todayStart.getTime()/1000;
-  let todayT=0; all.filter(r=>r.ts>=todayTs).forEach(r=>{todayT+=(r.input||0)+(r.output||0);});
+  let todayT=0; all.filter(r=>r.ts>=todayTs).forEach(r=>{todayT+=(r.input||0)+(r.output||0)+(r.cacheRead||0)+(r.cacheCreate||0);});
   if(tokTodayN) tokTodayN.textContent=fmtTok(todayT);
 }
 
@@ -3721,15 +3722,18 @@ function tokRenderTable(records) {
   if(_tokPage>=totalPages)_tokPage=totalPages-1;
   const pageItems=sorted.slice(_tokPage*TOK_PER_PAGE,(_tokPage+1)*TOK_PER_PAGE);
   for(const s of pageItems){
-    let sc=0; s.prompts.forEach(p=>{sc+=parseFloat(estCost(p.input||0,p.output||0,p.model,p.cacheRead||0,p.cacheCreate||0));});
+    const sCacheBase = s.input + s.cacheRead + s.cacheCreate;
+    const sCacheRate = sCacheBase > 0 ? (s.cacheRead / sCacheBase * 100).toFixed(1) + '%' : '0%';
     const tr=document.createElement('tr'); tr.className='tok-row-session';
-    tr.innerHTML=`<td>${escapeHtml(s.title)}${s.deleted?'<span class="tok-deleted">'+t('tok.deleted')+'</span>':''}</td><td>${fmtTok(s.input)}</td><td>${fmtTok(s.output)}</td><td>${fmtTok(s.cacheCreate)}</td><td>${fmtTok(s.cacheRead)}</td><td>¥${sc.toFixed(2)}</td>`;
+    tr.innerHTML=`<td>${escapeHtml(s.title)}${s.deleted?'<span class="tok-deleted">'+t('tok.deleted')+'</span>':''}</td><td>${fmtTok(s.input)}</td><td>${fmtTok(s.output)}</td><td>${fmtTok(s.cacheCreate)}</td><td>${fmtTok(s.cacheRead)}</td><td>${sCacheRate}</td>`;
     tokTbody.appendChild(tr);
     const details=[]; s.prompts.sort((a,b)=>b.ts-a.ts);
     for(const p of s.prompts){
       const dr=document.createElement('tr'); dr.className='tok-detail'; dr.hidden=true;
-      const modelHtml = p.model ? ` · <span class="tok-model-tip" data-tip="${escapeHtml(modelPriceTip(p.model))}">${escapeHtml(p.model)}</span>` : '';
-      dr.innerHTML=`<td>${fmtTime(p.ts)}${modelHtml}</td><td>${fmtTok(p.input||0)}</td><td>${fmtTok(p.output||0)}</td><td>${fmtTok(p.cacheCreate||0)}</td><td>${fmtTok(p.cacheRead||0)}</td><td>¥${estCost(p.input||0,p.output||0,p.model,p.cacheRead||0,p.cacheCreate||0)}</td>`;
+      const modelHtml = p.model ? ` · <span class="tok-model-tip">${escapeHtml(p.model)}</span>` : '';
+      const pCacheBase = (p.input||0) + (p.cacheRead||0) + (p.cacheCreate||0);
+      const pCacheRate = pCacheBase > 0 ? ((p.cacheRead||0) / pCacheBase * 100).toFixed(1) + '%' : '0%';
+      dr.innerHTML=`<td>${fmtTime(p.ts)}${modelHtml}</td><td>${fmtTok(p.input||0)}</td><td>${fmtTok(p.output||0)}</td><td>${fmtTok(p.cacheCreate||0)}</td><td>${fmtTok(p.cacheRead||0)}</td><td>${pCacheRate}</td>`;
       tokTbody.appendChild(dr); details.push(dr);
     }
     tr.addEventListener('click',()=>{const o=tr.classList.toggle('open');details.forEach(d=>d.hidden=!o);});
@@ -3795,7 +3799,11 @@ async function loadConductorTokens() {
   if (!tokTbody) return;
   const tip = t('tok.condTip');
   const _ci = `<svg width="14" height="14" ${CONDUCTOR_SVG_ATTRS} style="vertical-align:-2px;margin-right:4px">${CONDUCTOR_SVG_INNER}</svg>`;
-  tokTbody.innerHTML = `<tr class="tok-row-conductor" title="${tip}"><td>${_ci}${t('tok.condTotal')}</td><td>${fmtTok(hIn)}</td><td>${fmtTok(hOut)}</td><td>${fmtTok(hCc)}</td><td>${fmtTok(hCr)}</td><td>¥${hCost.toFixed(2)}</td></tr><tr class="tok-row-conductor" title="${tip}"><td>${_ci}${t('tok.condCurrent')}</td><td>${fmtTok(curIn)}</td><td>${fmtTok(curOut)}</td><td>${fmtTok(curCc)}</td><td>${fmtTok(curCr)}</td><td>¥${curCost.toFixed(2)}</td></tr>`;
+  const hCacheBase = hIn + hCr + hCc;
+  const hCacheRate = hCacheBase > 0 ? (hCr / hCacheBase * 100).toFixed(1) + '%' : '0%';
+  const curCacheBase = curIn + curCr + curCc;
+  const curCacheRate = curCacheBase > 0 ? (curCr / curCacheBase * 100).toFixed(1) + '%' : '0%';
+  tokTbody.innerHTML = `<tr class="tok-row-conductor" title="${tip}"><td>${_ci}${t('tok.condTotal')}</td><td>${fmtTok(hIn)}</td><td>${fmtTok(hOut)}</td><td>${fmtTok(hCc)}</td><td>${fmtTok(hCr)}</td><td>${hCacheRate}</td></tr><tr class="tok-row-conductor" title="${tip}"><td>${_ci}${t('tok.condCurrent')}</td><td>${fmtTok(curIn)}</td><td>${fmtTok(curOut)}</td><td>${fmtTok(curCc)}</td><td>${fmtTok(curCr)}</td><td>${curCacheRate}</td></tr>`;
   const pager = document.getElementById('tok-pager');
   if (pager) pager.innerHTML = '';
 }
@@ -3824,19 +3832,18 @@ function renderTokChart() {
   const daily = {};
   for (const r of history) {
     const day = new Date(r.ts * 1000).toLocaleDateString('sv');
-    const cost = parseFloat(estCost(r.input||0, r.output||0, r.model||'', r.cacheRead||0, r.cacheCreate||0));
-    daily[day] = (daily[day] || 0) + cost;
+    daily[day] = (daily[day] || 0) + (r.input || 0) + (r.output || 0) + (r.cacheRead || 0) + (r.cacheCreate || 0);
   }
   const rawDays = Object.keys(daily).sort();
   if (!rawDays.length) { tokChartEl.innerHTML = ''; return; }
-  // Fill gaps: include days with 0 cost between first and last
+  // Fill gaps: include days with 0 between first and last
   const days = [];
   const d0 = new Date(rawDays[0]), d1 = new Date(rawDays[rawDays.length - 1]);
   for (let d = new Date(d0); d <= d1; d.setDate(d.getDate() + 1)) {
     days.push(d.toLocaleDateString('sv'));
   }
   const vals = days.map(d => daily[d] || 0);
-  const maxVal = Math.max(...vals, 0.01);
+  const maxVal = Math.max(...vals, 1);
   const W = 600, H = 260, PL = 48, PR = 30, PT = 14, PB = 28;
   const cw = W - PL - PR, ch = H - PT - PB;
   const step = days.length > 1 ? cw / (days.length - 1) : cw;
@@ -3844,14 +3851,14 @@ function renderTokChart() {
   const polyline = pts.map(p => p.join(',')).join(' ');
   const yLines = [0, 0.25, 0.5, 0.75, 1].map(f => {
     const y = PT + ch - f * ch;
-    const label = '¥' + (maxVal * f).toFixed(1);
+    const label = fmtTok(maxVal * f);
     return `<line x1="${PL}" x2="${W-PR}" y1="${y}" y2="${y}" stroke="var(--line-soft)" stroke-width="0.5"/><text x="${PL-4}" y="${y+3}" text-anchor="end" font-size="9" fill="var(--muted)">${label}</text>`;
   }).join('');
   const xLabels = days.map((d, i) => {
     if (days.length > 14 && i % Math.ceil(days.length / 7) !== 0 && i !== days.length - 1) return '';
     return `<text x="${pts[i][0]}" y="${H-4}" text-anchor="middle" font-size="9" fill="var(--muted)">${d.slice(5)}</text>`;
   }).join('');
-  const dots = pts.map((p, i) => `<circle cx="${p[0]}" cy="${p[1]}" r="3" fill="var(--blue)"><title>${days[i]}: ¥${vals[i].toFixed(2)}</title></circle>`).join('');
+  const dots = pts.map((p, i) => `<circle cx="${p[0]}" cy="${p[1]}" r="3" fill="var(--blue)"><title>${days[i]}: ${fmtTok(vals[i])}</title></circle>`).join('');
   tokChartEl.innerHTML = `<svg viewBox="0 0 ${W} ${H}" preserveAspectRatio="none">${yLines}${xLabels}<polyline points="${polyline}" fill="none" stroke="var(--blue)" stroke-width="1.5"/>${dots}</svg>`;
 }
 
diff --git a/frontends/desktop/static/index.html b/frontends/desktop/static/index.html
index c7a14cab..45cf9e6c 100644
--- a/frontends/desktop/static/index.html
+++ b/frontends/desktop/static/index.html
@@ -4,7 +4,7 @@
 <meta charset="UTF-8">
 <title>GenericAgent</title>
 <link rel="stylesheet" href="assets/fonts/fonts.css">
-<link rel="stylesheet" href="styles.css?v=117">
+<link rel="stylesheet" href="styles.css?v=118">
 <!-- KaTeX: LaTeX 公式渲染 -->
 <link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.16.11/dist/katex.min.css" crossorigin="anonymous">
 <!-- highlight.js: 代码高亮 -->
@@ -329,7 +329,7 @@
             </div>
             <div class="stat-row">
               <div class="stat"><div class="s-n" id="tok-total-n">0</div><div class="s-l" data-i18n="tok.total"></div></div>
-              <div class="stat"><div class="s-n" id="tok-cost-n">¥ 0</div><div class="s-l" data-i18n="tok.cost"></div></div>
+              <div class="stat"><div class="s-n" id="tok-cost-n">0%</div><div class="s-l" data-i18n="tok.cost"></div></div>
               <div class="stat"><div class="s-n" id="tok-today-n">0</div><div class="s-l" data-i18n="tok.today"></div></div>
             </div>
             <div class="tok-chart-wrap" id="tok-chart-wrap">
@@ -588,6 +588,6 @@
 <!-- highlight.js -->
 <script src="https://cdn.jsdelivr.net/gh/highlightjs/cdn-release@11.9.0/build/highlight.min.js" crossorigin="anonymous"></script>
 <script src="phosphor-icons.js"></script>
-<script src="app.js?v=159"></script>
+<script src="app.js?v=160"></script>
 </body>
 </html>
diff --git a/frontends/desktop/static/styles.css b/frontends/desktop/static/styles.css
index 20c31f5a..0aa5a554 100644
--- a/frontends/desktop/static/styles.css
+++ b/frontends/desktop/static/styles.css
@@ -955,18 +955,8 @@ span.flatpickr-weekday{ color:var(--muted) !important; font-size:11px; }
 .tok-detail td:not(:first-child){ text-align:right; }
 .tok-detail[hidden]{ display:none; }
 .tok-model-tip{
-  position:relative; cursor:help; color:var(--accent);
-  border-bottom:1px dashed var(--muted2);
+  color:var(--accent);
 }
-.tok-model-tip::after{
-  content:attr(data-tip); position:absolute; bottom:calc(100% + 6px); left:50%;
-  transform:translateX(-50%); white-space:pre-line; font-size:11px; line-height:1.5;
-  background:var(--card); color:var(--txt); border:1px solid var(--line);
-  border-radius:4px; padding:8px 10px; box-shadow:var(--shadow-pop);
-  pointer-events:none; opacity:0; transition:opacity .15s; z-index:100;
-  min-width:160px; text-align:left;
-}
-.tok-model-tip:hover::after{ opacity:1; }
 .tok-pager{ display:flex; gap:4px; justify-content:center; margin-top:12px; }
 .tok-pager button{
   min-width:28px; height:28px; border:1px solid var(--line); border-radius:4px;
diff --git a/frontends/desktop_bridge.py b/frontends/desktop_bridge.py
index 3b2c119a..0576a428 100644
--- a/frontends/desktop_bridge.py
+++ b/frontends/desktop_bridge.py
@@ -88,6 +88,7 @@ class Session:
     untitled: bool = True
     plan_scan_baseline: int = 0
     plan_path: str = ""
+    llm_history: Optional[List[dict]] = None
 
 
 def _load_plan_baseline(item: dict, msgs: list) -> int:
@@ -129,12 +130,19 @@ class AgentManager:
             arr = []
             with self.lock:
                 for s in self.sessions.values():
+                    llm_hist = None
+                    if s.agent and hasattr(s.agent, 'llmclient'):
+                        try: llm_hist = s.agent.llmclient.backend.history
+                        except Exception: pass
+                    if llm_hist is None:
+                        llm_hist = s.llm_history
                     arr.append({"id": s.id, "title": s.title, "cwd": s.cwd,
                                 "created_at": s.created_at, "updated_at": s.updated_at,
                                 "messages": s.messages, "msg_seq": s.msg_seq,
                                 "pinned": s.pinned, "untitled": s.untitled,
                                 "plan_scan_baseline": s.plan_scan_baseline,
-                                "plan_path": s.plan_path or ""})
+                                "plan_path": s.plan_path or "",
+                                "llm_history": llm_hist})
             self._sessions_file.write_text(json.dumps(arr, ensure_ascii=False, default=str), encoding="utf-8")
         except Exception as e:
             print(f"[bridge] persist sessions failed: {e}", file=sys.stderr)
@@ -157,7 +165,8 @@ class AgentManager:
                                plan_scan_baseline=_load_plan_baseline(item, msgs),
                                plan_path=_sanitize_desktop_plan_path(
                                    item["id"], item.get("plan_path") or ""),
-                               status="idle", agent=None)
+                               status="idle", agent=None,
+                               llm_history=item.get("llm_history"))
                 self.sessions[sess.id] = sess
             if self.sessions:
                 self.active_session_id = max(self.sessions.values(), key=lambda s: s.updated_at).id
@@ -512,6 +521,8 @@ class AgentManager:
                 import plan_state
                 plan_state.sync_plan_path_from_text(sess, full, sess.cwd or self.ga_root)
                 self.add_message(sess, "assistant", full)
+                try: sess.llm_history = json.loads(json.dumps(agent.llmclient.backend.history, ensure_ascii=False, default=str))
+                except Exception: pass
                 sess.status = "idle"
                 sess.last_error = ""
             emit_session_state(sess, "idle")
@@ -584,23 +595,29 @@ class AgentManager:
             if sess.agent is not None:
                 return {"ok": True, "sessionId": sid, "restored": False, "reason": "agent already alive"}
         agent = self.make_agent(sess)
-        history = []
-        for m in sess.messages:
-            role = m.get("role")
-            content = m.get("content", "")
-            if role == "user":
-                history.append({"role": "user", "content": [{"type": "text", "text": content}]})
-            elif role == "assistant":
-                history.append({"role": "assistant", "content": [{"type": "text", "text": content}]})
-        if history:
+        if sess.llm_history:
             try:
-                agent.llmclient.backend.history = history
+                agent.llmclient.backend.history = sess.llm_history
             except Exception as e:
-                print(f"[bridge] inject history failed: {e}", file=sys.stderr)
+                print(f"[bridge] restore llm_history failed: {e}", file=sys.stderr)
+        else:
+            history = []
+            for m in sess.messages:
+                role = m.get("role")
+                content = m.get("content", "")
+                if role == "user":
+                    history.append({"role": "user", "content": [{"type": "text", "text": content}]})
+                elif role == "assistant":
+                    history.append({"role": "assistant", "content": [{"type": "text", "text": content}]})
+            if history:
+                try:
+                    agent.llmclient.backend.history = history
+                except Exception as e:
+                    print(f"[bridge] inject history failed: {e}", file=sys.stderr)
         with self.lock:
             sess.agent = agent
             sess.status = "idle"
-        return {"ok": True, "sessionId": sid, "restored": True, "messageCount": len(history)}
+        return {"ok": True, "sessionId": sid, "restored": True, "messageCount": len(sess.llm_history or sess.messages)}
 
 
 import base64