feat(cli): 面板与状态栏接真实数据，/new 对齐草稿语义；host 设置面双通道化

CLI（bun 跑源码即生效）：
- boot：loadBootData 一次性加载模型清单（含 context_window）/个性化默认
  （含压缩设置）/path 授权/会话列表；新会话 unshift 标 current
- /session：Enter 真实加载（timeline.reset + getSessionHistory 渲染历史文本）
- /new：对齐 web /new 页面语义——空对话草稿不立即建会话，首条消息
  createTask 不带 cid 由服务端惰性创建，响应回 id 后插入列表标 current
- /context + 状态栏：token_update 事件（运行期事件流）+ 打开面板时查询
  新端点 GET /api/runtime/sessions/<cid>/token-stats；百分比算法对齐 web
  InputComposer（自动压缩阈值优先于模型窗口）
- /model：Enter = 本地生效 + createTask 带 model_key/run_mode/thinking_mode
  覆盖（会话级即时）+ savePersonalization patch 存默认（effort default 存 null）
- /path：增删后 savePathAuths 全量提交两组（服务端 deny 校验保留）
- /agents /tasks /workflow /rewind：打开时拉真实列表（sub_agents/
  background_commands/workflows/versioning-checkpoints）
- 发送链路统一 sendWithSettings 携带 /model 生效值（含排队队列）

后端：
- 8 端点装饰器 api_login_required → api_login_or_host_token_required
  （personalization×2 / path-authorization×2 / sub_agents / background_commands
  / workflows / versioning-checkpoints），web session 行为不变
- headless_app 用 add_url_rule 共享上述视图函数（URL 与 full 一致）；
  sub_agents/background_commands 支持 query 显式传 conversation_id
  （Bearer 装配的 terminal 无当前对话概念）
- RuntimeService 新增 get_session_token_stats；gateway_api 挂 token-stats 路由
- headless 404 兜底：非 /api/ 的 GET 路径回落告知页（对齐 SPA fallback）

验证：tsc ✅ py_compile ✅ 冒烟 6/6 ✅ headless 装配 54 路由 ✅
localconfig 无头实测读到真实配置 ✅；真实服务端到端待用户重启 8091 实机验收

Co-authored-by: Astrion powered by Kimi-K3 <astrion-agent@users.noreply.github.com>

AGENTS.md

@@ -111,7 +111,8 @@
 - 兼容启动方式：`python web_server.py`
 - 统一入口：`python main.py`（当前实现会默认进入 Web 启动流程）
 - **Headless 启动（CLI 自启动专用）**：`python -m server.headless_app --path <cwd> --port 8091 --thinking-mode`
-  - 只挂运行时蓝图（gateway / tasks / status / approval / usage 五件套），不含 web 站点路由面（无 /login、无静态页、无会话管理）；根路径 `/` 返回简单 HTML 告知页
+  - 只挂运行时蓝图（gateway / tasks / status / approval / usage 五件套），不含 web 站点路由面（无 /login、无静态页、无会话管理）；根路径 `/` 返回简单 HTML 告知页；非 /api/ 的 GET 路径 404 时回落告知页（对齐 SPA fallback 语义）
+  - **headless 额外挂 host 设置面**（`add_url_rule` 直接共享 chat_bp/conversation_bp/workflow_page_bp 的视图函数，URL 与 full 一致）：`/api/personalization` GET/POST、`/api/path-authorization` GET/POST、`/api/sub_agents` GET、`/api/background_commands` GET、`/api/workflows` GET、`/api/conversations/<cid>/versioning/checkpoints` GET——这些端点装饰器已双通道化（`api_login_or_host_token_required`），list 类端点支持 query 显式传 conversation_id（Bearer 装配的 terminal 无当前对话概念）
   - 与 full 形态同进程模型（RuntimeService 单例/同一数据目录/同一端口 8091），只是启动时挂的路由面不同；**绝不能与 full 形态双进程并存**（MultiAgentState/门闸/审批 Map 是进程内单例）
   - CLI（`cli/src/gateway.ts`）探测无服务时 spawn 的就是这个入口；python 解释器经依赖探测（import yaml/flask）从 .venv → homebrew 3.12/3.11 → python3 中选第一个可用的
 

cli/src/App.tsx

@@ -36,6 +36,11 @@ function makeApi(setBlocks: React.Dispatch<React.SetStateAction<Block[]>>, getEl
   };
 
   const api: TimelineApi = {
+    reset() {
+      thinkingId = null;
+      assistantId = null;
+      setBlocks([]);
+    },
     addUser(text) {
       setBlocks((prev) => [...prev, { kind: 'user', id: idCounter++, text }]);
     },
@@ -108,6 +113,35 @@ function patchLastTool(blocks: Block[], patch: Partial<Extract<Block, { kind: 't
   return blocks;
 }
 
+/** 子智能体端点状态 → 面板状态（running/idle 直通，失败类归 error，其余终态归 terminated） */
+function mapAgentStatus(status: unknown): 'running' | 'idle' | 'terminated' | 'error' {
+  const s = String(status ?? '').toLowerCase();
+  if (s === 'running' || s === 'idle') return s;
+  if (s === 'failed' || s === 'error') return 'error';
+  return 'terminated';
+}
+
+/** 后台指令端点状态 → 面板状态（completed→done；failed/timeout/cancelled→error） */
+function mapCommandStatus(status: unknown): 'running' | 'done' | 'error' {
+  const s = String(status ?? '').toLowerCase();
+  if (s === 'completed' || s === 'done' || s === 'succeeded') return 'done';
+  if (s === 'failed' || s === 'timeout' || s === 'cancelled' || s === 'error') return 'error';
+  return 'running';
+}
+
+/** created_at（epoch 秒或 ISO）→ 运行时长粗文案（mm:ss / hh:mm） */
+function relativeElapsed(createdAt: unknown): string {
+  let ts = 0;
+  if (typeof createdAt === 'number') ts = createdAt > 1e12 ? createdAt : createdAt * 1000;
+  else ts = Date.parse(String(createdAt ?? ''));
+  if (!Number.isFinite(ts) || ts <= 0) return '';
+  const sec = Math.max(0, Math.floor((Date.now() - ts) / 1000));
+  const mm = Math.floor(sec / 60);
+  const ss = sec % 60;
+  if (mm < 60) return `${mm}:${String(ss).padStart(2, '0')}`;
+  return `${Math.floor(mm / 60)}:${String(mm % 60).padStart(2, '0')}:${String(ss).padStart(2, '0')}`;
+}
+
 export function App({ boot }: { boot: BootResult }) {
   const renderer = useRenderer();
   const [blocks, setBlocks] = useState<Block[]>([]);
@@ -148,6 +182,103 @@ export function App({ boot }: { boot: BootResult }) {
         );
       }
     },
+    onSessionLoad: (s) => {
+      void runtimeRef.current?.loadSession(s.id);
+    },
+    onContextRefresh: () => {
+      void runtimeRef.current?.refreshTokenStats();
+    },
+    onSaveModelDefaults: (v) => {
+      // patch 语义（服务端 sanitize fallback=existing）；effort 'default' = null（不指定）
+      void boot.gateway
+        .savePersonalization({
+          default_model: v.model,
+          default_run_mode: v.thinking ? 'deep' : 'fast',
+          default_reasoning_effort: v.effort === 'default' ? null : v.effort,
+        })
+        .catch((err) =>
+          apiRef.current?.addSystem(`${t('settings.saveFailed')}${err instanceof Error ? err.message : String(err)}`),
+        );
+    },
+    onSavePathAuths: (auths) => {
+      void boot.gateway
+        .savePathAuths(
+          auths.filter((a) => a.access === 'rw').map((a) => a.path),
+          auths.filter((a) => a.access === 'ro').map((a) => a.path),
+        )
+        .catch((err) =>
+          apiRef.current?.addSystem(`${t('settings.saveFailed')}${err instanceof Error ? err.message : String(err)}`),
+        );
+    },
+    onAgentsRefresh: () => {
+      const cid = runtimeRef.current?.conversationId ?? '';
+      void boot.gateway
+        .listSubAgents(cid)
+        .then((list) => {
+          menuRef.current.setAgents(
+            list.map((a: any) => ({
+              name: String(a.display_name ?? a.agent_name ?? a.task_id ?? '?'),
+              task: String(a.task ?? a.description ?? ''),
+              status: mapAgentStatus(a.status),
+              elapsed: relativeElapsed(a.created_at),
+            })),
+          );
+        })
+        .catch(() => {});
+    },
+    onTasksRefresh: () => {
+      const cid = runtimeRef.current?.conversationId ?? '';
+      void boot.gateway
+        .listBackgroundCommands(cid)
+        .then((list) => {
+          menuRef.current.setTasks(
+            list.map((c: any) => ({
+              id: String(c.command_id ?? ''),
+              command: String(c.command ?? ''),
+              status: mapCommandStatus(c.status),
+              lastLine: typeof c.return_code === 'number' ? `rc=${c.return_code}` : '',
+            })),
+          );
+        })
+        .catch(() => {});
+    },
+    onWorkflowsRefresh: () => {
+      void boot.gateway
+        .listWorkflows()
+        .then((list) => {
+          menuRef.current.setWorkflows(
+            list.map((w: any) => ({
+              name: String(w.name ?? '?'),
+              desc: String(w.description ?? w.desc ?? ''),
+              stages: Number(w.stage_count ?? w.stages ?? 0) || 0,
+            })),
+          );
+        })
+        .catch(() => {});
+    },
+    onCheckpointsRefresh: () => {
+      const cid = runtimeRef.current?.conversationId ?? '';
+      if (!cid) return;
+      void boot.gateway
+        .listCheckpoints(cid)
+        .then(({ items }) => {
+          menuRef.current.setCheckpoints(
+            items.map((c: any) => ({
+              id: String(c.seq ?? c.id ?? ''),
+              when: relativeElapsed(c.created_at ?? c.ts),
+              summary: String(c.summary ?? c.label ?? ''),
+              files: Number(c.files ?? c.file_count ?? 0) || 0,
+            })),
+          );
+        })
+        .catch(() => {});
+    },
+    onNewSession: () => {
+      // /new = web /new 页面语义：空对话草稿，不创建会话；列表全部取消 current
+      runtimeRef.current?.enterNewSession();
+      menuRef.current.setSessions((prev) => prev.map((s) => ({ ...s, current: false })));
+      apiRef.current?.addSystem(t('session.draftHint'));
+    },
   });
   const menuRef = useRef(menu);
   menuRef.current = menu;
@@ -159,6 +290,14 @@ export function App({ boot }: { boot: BootResult }) {
       onRunningChange: setRunning,
       onApprovalRequired: (a) => menuRef.current.openApproval(a),
       onSystemMessage: (text) => apiRef.current?.addSystem(text),
+      onTokenUpdate: (stats) => menuRef.current.setContextStats(stats),
+      onConversationCreated: (id) => {
+        // /new 草稿首发后：真实会话插入列表并标 current
+        menuRef.current.setSessions((prev) => [
+          { id, title: t('session.new'), when: t('time.justNow'), current: true },
+          ...prev.map((s) => ({ ...s, current: false })),
+        ]);
+      },
       tr: t,
     });
     rt.conversationId = boot.conversationId;
@@ -200,12 +339,22 @@ export function App({ boot }: { boot: BootResult }) {
     if (running || queue.length === 0) return;
     const timer = setTimeout(() => {
       const [head, ...rest] = queue;
-      if (head) void runtimeRef.current?.send(head);
+      if (head) sendWithSettings(head);
       setQueue(rest);
     }, 300);
     return () => clearTimeout(timer);
   }, [running, queue]);
 
+  // 发送统一入口：携带 /model 面板的生效值（会话级覆盖，服务端优先级最高）
+  const sendWithSettings = (text: string) => {
+    const sf = menuRef.current.statusFields;
+    void runtimeRef.current?.send(text, {
+      model_key: sf.model || undefined,
+      thinking_mode: sf.thinking,
+      run_mode: sf.thinking ? 'deep' : 'fast',
+    });
+  };
+
   // 引导：输入框有字=直接引导注入当前轮；无字=提升队首为引导
   // TODO(gateway)：接 POST /api/tasks/<id>/runtime_guidance（当前仅本地上屏，服务端注入后续接）
   const handleGuide = () => {
@@ -249,7 +398,7 @@ export function App({ boot }: { boot: BootResult }) {
     if (runningRef.current) {
       setQueue((q) => [...q, text]);
     } else {
-      void runtimeRef.current?.send(text);
+      sendWithSettings(text);
     }
   };
 

cli/src/boot.tsx

@@ -6,7 +6,45 @@ import { useEffect, useRef, useState } from 'react';
 import { RGBA, TextAttributes } from '@opentui/core';
 import { GatewayClient, type WorkspaceItem } from './gateway';
 import { t } from './i18n';
-import { WORKSPACE } from './data';
+import { BOOT_STATE, INITIAL_PATH_AUTHS, MODEL_OPTIONS, SESSIONS, WORKSPACE } from './data';
+import { loadCustomModels, loadPathAuths, loadPersonalizationDefaults } from './localconfig';
+
+/** ISO 时间 → 相对时间短文案（会话列表 when 列） */
+function relativeTime(iso: unknown): string {
+  const ts = Date.parse(String(iso ?? ''));
+  if (!Number.isFinite(ts)) return '';
+  const diffMin = Math.floor((Date.now() - ts) / 60000);
+  if (diffMin < 1) return t('time.justNow');
+  if (diffMin < 60) return `${diffMin} ${t('time.minutesAgo')}`;
+  const diffHour = Math.floor(diffMin / 60);
+  if (diffHour < 24) return `${diffHour} ${t('time.hoursAgo')}`;
+  return `${Math.floor(diffHour / 24)} ${t('time.daysAgo')}`;
+}
+
+/** 启动数据加载：本地配置（模型清单/个性化默认/路径授权）+ Gateway 会话列表。
+ *  本地配置与读 token 文件同一安全语义（host 本机单人）；会话列表失败不阻塞启动。 */
+async function loadBootData(gw: GatewayClient): Promise<void> {
+  if (MODEL_OPTIONS.length === 0) MODEL_OPTIONS.push(...loadCustomModels());
+  const prefs = loadPersonalizationDefaults();
+  BOOT_STATE.model = prefs.model || MODEL_OPTIONS[0]?.name || '';
+  BOOT_STATE.thinking = prefs.thinking;
+  BOOT_STATE.effort = prefs.effort;
+  BOOT_STATE.workMode = prefs.workMode;
+  BOOT_STATE.permMode = prefs.permMode;
+  BOOT_STATE.autoDeepCompress = prefs.autoDeepCompress;
+  BOOT_STATE.deepCompressLimit = prefs.deepCompressLimit;
+  if (INITIAL_PATH_AUTHS.length === 0) INITIAL_PATH_AUTHS.push(...loadPathAuths());
+  try {
+    const list = await gw.listSessions(20);
+    for (const c of list) {
+      const id = String(c?.id ?? c?.conversation_id ?? '');
+      if (!id) continue;
+      SESSIONS.push({ id, title: String(c?.title || t('session.untitled')), when: relativeTime(c?.updated_at) });
+    }
+  } catch {
+    // 列表失败不阻塞启动（/session 面板将只显示当前新会话）
+  }
+}
 
 const FG = RGBA.defaultForeground();
 const DIM = TextAttributes.DIM;
@@ -45,7 +83,11 @@ export function BootFlow({ cwd, onReady, onExit }: { cwd: string; onReady: (r: B
     WORKSPACE.path = workspace.path;
     setStage('creating');
     try {
+      await loadBootData(gw);
       const session = await gw.createSession();
+      // 新会话置顶并标 current（menuState 初值读 SESSIONS）
+      for (const s of SESSIONS) s.current = false;
+      SESSIONS.unshift({ id: session.conversation_id, title: t('session.new'), when: t('time.justNow'), current: true });
       onReady({ gateway: gw, conversationId: session.conversation_id, workspace });
     } catch (err) {
       setError(err instanceof Error ? err.message : String(err));

cli/src/data.ts

@@ -26,6 +26,9 @@ export const BOOT_STATE = {
   execEnv: 'direct',
   network: 'restricted',
   contextUsage: '—',
+  /** 自动深度压缩开关与触发阈值（上下文百分比分母，对齐 web InputComposer 算法） */
+  autoDeepCompress: false,
+  deepCompressLimit: 150000,
 };
 
 // ── 工作区（/session 面板顶部展示；boot 确定后写入） ──
@@ -36,10 +39,12 @@ export interface WorkspaceInfo {
 }
 export const WORKSPACE: WorkspaceInfo = { id: '', name: '', path: '' };
 
-// ── 模型（boot 从 /api/v1/models 加载后填充） ──
+// ── 模型（boot 从本地 custom_models.json 加载后填充） ──
 export interface ModelOption {
   name: string;
   meta: string;
+  /** 上下文窗口上限（token；配置缺省为 null，百分比显示回退压缩阈值） */
+  contextWindow: number | null;
 }
 export const MODEL_OPTIONS: ModelOption[] = [];
 
@@ -77,26 +82,49 @@ export interface PathAuth {
 export const PATH_ACCESS_LABEL: Record<PathAccess, string> = { rw: t('path.access.rw'), ro: t('path.access.ro') };
 export const INITIAL_PATH_AUTHS: PathAuth[] = [];
 
-// ── 上下文统计（/context 面板；后续接会话 token 统计端点） ──
+// ── 上下文统计（/context 面板与状态栏共用；token_update 事件/查询端点刷新，数字型原始值） ──
 export interface ContextStats {
-  used: string;
+  currentTokens: number;
-  total: string;
+  totalInput: number;
-  percent: number;
+  totalOutput: number;
-  totalInput: string;
+  cacheInput: number;
-  totalOutput: string;
+  cacheExemptInput: number;
-  cacheInput: string;
-  cacheHitRate: string;
 }
-export const CONTEXT_STATS: ContextStats = {
+export const EMPTY_CONTEXT_STATS: ContextStats = {
-  used: '—',
+  currentTokens: 0,
-  total: '—',
+  totalInput: 0,
-  percent: 0,
+  totalOutput: 0,
-  totalInput: '—',
+  cacheInput: 0,
-  totalOutput: '—',
+  cacheExemptInput: 0,
-  cacheInput: '—',
-  cacheHitRate: '—',
 };
 
+/** token 数紧凑格式（对齐 web InputComposer：≥1000 转 k，<100k 保留一位小数） */
+export function formatCompactTokens(value: number): string {
+  const n = Math.max(0, Math.round(value || 0));
+  if (n >= 1000) {
+    const raw = n / 1000;
+    const text = raw >= 100 ? String(Math.round(raw)) : raw.toFixed(1).replace(/\.0$/, '');
+    return `${text}k`;
+  }
+  return String(n);
+}
+
+/** 上下文百分比分母（对齐 web：开自动深度压缩用触发阈值，否则模型上下文窗口） */
+export function contextUsageLimitFor(modelName: string): number {
+  if (BOOT_STATE.autoDeepCompress) return BOOT_STATE.deepCompressLimit;
+  const m = MODEL_OPTIONS.find((o) => o.name === modelName);
+  return m?.contextWindow ?? 0;
+}
+
+/** 状态栏上下文用量文案：`45% 90.2k/200k`（limit 未知时 `90.2k/—`） */
+export function formatContextUsage(stats: ContextStats, modelName: string): string {
+  const cur = formatCompactTokens(stats.currentTokens);
+  const limit = contextUsageLimitFor(modelName);
+  if (limit <= 0) return `${cur}/—`;
+  const pct = Math.max(0, Math.min(100, Math.round((stats.currentTokens / limit) * 100)));
+  return `${pct}% ${cur}/${formatCompactTokens(limit)}`;
+}
+
 // ── 边界四选项（/mode /permission /env /network 共用 Selector 面板；切换后调对应 API） ──
 export interface BoundaryOption {
   value: string;

cli/src/gateway.ts

@@ -190,7 +190,8 @@ export class GatewayClient {
 
   async listSessions(limit = 20): Promise<any[]> {
     const res = await this.request<any>(`/api/runtime/sessions?limit=${limit}`, { method: 'GET' });
-    return res.data?.sessions ?? res.sessions ?? res.data?.items ?? res.items ?? [];
+    // 顶层展开风格的实际键是 conversations（8093 实测），兼容 sessions/items 包装
+    return res.data?.sessions ?? res.sessions ?? res.conversations ?? res.data?.items ?? res.items ?? [];
   }
 
   async getSessionHistory(conversationId: string): Promise<any> {
@@ -200,8 +201,72 @@ export class GatewayClient {
     return res.data ?? res.conversation ?? res;
   }
 
+  /** 会话 token 统计（/context 面板；顶层展开风格 res.stats） */
+  async getTokenStats(conversationId: string): Promise<any> {
+    const res = await this.request<any>(`/api/runtime/sessions/${encodeURIComponent(conversationId)}/token-stats`, {
+      method: 'GET',
+    });
+    return res.stats ?? res.data?.stats ?? res.data ?? {};
+  }
+
+  // ── host 全局设置（写操作；端点为 chat_bp 共享视图函数，装饰器双通道化后 Bearer 可达） ──
+
+  /** 保存个性化默认（patch 语义：服务端 sanitize fallback=existing，未传字段不动） */
+  async savePersonalization(patch: Record<string, unknown>): Promise<void> {
+    await this.request<any>('/api/personalization', { method: 'POST', body: patch });
+  }
+
+  /** 路径授权全量保存（端点语义为两组全量提交；返回服务端落定后的两组列表） */
+  async savePathAuths(writable: string[], readableExtra: string[]): Promise<{ writable: string[]; readableExtra: string[] }> {
+    const res = await this.request<any>('/api/path-authorization', {
+      method: 'POST',
+      body: { writable_paths: writable, readable_extra_paths: readableExtra },
+    });
+    return {
+      writable: res.writable_paths ?? res.data?.writable_paths ?? writable,
+      readableExtra: res.readable_extra_paths ?? res.data?.readable_extra_paths ?? readableExtra,
+    };
+  }
+
+  /** 子智能体列表（/agents 面板；conversation_id 经 query 显式指定） */
+  async listSubAgents(conversationId: string): Promise<any[]> {
+    const q = conversationId ? `?conversation_id=${encodeURIComponent(conversationId)}` : '';
+    const res = await this.request<any>(`/api/sub_agents${q}`, { method: 'GET' });
+    return res.data ?? [];
+  }
+
+  /** 后台指令列表（/tasks 面板；conversation_id 经 query 显式指定） */
+  async listBackgroundCommands(conversationId: string): Promise<any[]> {
+    const q = conversationId ? `?conversation_id=${encodeURIComponent(conversationId)}` : '';
+    const res = await this.request<any>(`/api/background_commands${q}`, { method: 'GET' });
+    return res.data ?? [];
+  }
+
+  /** 工作流库列表（/workflow 面板；注意该端点顶层展开为 {workflows}，无 success 包装） */
+  async listWorkflows(): Promise<any[]> {
+    const res = await this.request<any>('/api/workflows', { method: 'GET' });
+    return res.workflows ?? res.data?.workflows ?? [];
+  }
+
+  /** 版本回溯检查点（/rewind 面板；返回 data.items，含 enabled 标记） */
+  async listCheckpoints(conversationId: string): Promise<{ items: any[]; enabled: boolean }> {
+    const res = await this.request<any>(
+      `/api/conversations/${encodeURIComponent(conversationId)}/versioning/checkpoints`,
+      { method: 'GET' },
+    );
+    const data = res.data ?? {};
+    return { items: Array.isArray(data.items) ? data.items : [], enabled: data.enabled !== false };
+  }
+
   // ── Run（任务） ──
-  async createTask(payload: { message: string; conversation_id?: string }): Promise<{ task_id: string; conversation_id?: string }> {
+  async createTask(payload: {
+    message: string;
+    conversation_id?: string;
+    /** 会话级覆盖（/model 面板生效值）；不传则服务端用对话/个性化默认 */
+    model_key?: string;
+    run_mode?: string;
+    thinking_mode?: boolean;
+  }): Promise<{ task_id: string; conversation_id?: string }> {
     const res = await this.request<{ data: any }>('/api/tasks', { method: 'POST', body: payload });
     return res.data;
   }

cli/src/i18n/en-US.ts

@@ -95,4 +95,26 @@ export default {
   'tool.write_file': 'Write file',
   'tool.edit_file': 'Edit file',
   'tool.read_file': 'Read file',
+
+  // ── Session list / loading ──
+  'session.new': 'New chat',
+  'session.untitled': 'Untitled chat',
+  'session.draftHint': 'New chat draft (starts on first message)',
+  'session.loaded': 'Loaded conversation: ',
+  'session.loadFailed': 'Failed to load conversation: ',
+  'time.justNow': 'just now',
+  'time.minutesAgo': 'm ago',
+  'time.hoursAgo': 'h ago',
+  'time.daysAgo': 'd ago',
+
+  // ── /context panel ──
+  'context.title': 'Context usage   Esc to close',
+  'context.current': 'Current ',
+  'context.totalInput': 'Input   ',
+  'context.totalOutput': 'Output  ',
+  'context.cacheInput': 'Cached  ',
+  'context.cacheHitRate': 'Hit rate',
+
+  // ── Settings persistence ──
+  'settings.saveFailed': 'Failed to save settings: ',
 } as const;

cli/src/i18n/zh-CN.ts

@@ -95,4 +95,26 @@ export default {
   'tool.write_file': '写入文件',
   'tool.edit_file': '编辑文件',
   'tool.read_file': '读取文件',
+
+  // ── 会话列表 / 加载 ──
+  'session.new': '新对话',
+  'session.untitled': '未命名对话',
+  'session.draftHint': '已进入新对话（发送消息后开始）',
+  'session.loaded': '已加载对话：',
+  'session.loadFailed': '加载对话失败：',
+  'time.justNow': '刚刚',
+  'time.minutesAgo': '分钟前',
+  'time.hoursAgo': '小时前',
+  'time.daysAgo': '天前',
+
+  // ── /context 面板 ──
+  'context.title': '上下文用量   Esc 关闭',
+  'context.current': '当前上下文',
+  'context.totalInput': '累计输入　',
+  'context.totalOutput': '累计输出　',
+  'context.cacheInput': '缓存输入　',
+  'context.cacheHitRate': '缓存命中率',
+
+  // ── 设置保存 ──
+  'settings.saveFailed': '保存设置失败：',
 } as const;

cli/src/localconfig.ts

@@ -30,7 +30,7 @@ function resolveDeployConfig(name: string): any | null {
   return null;
 }
 
-/** 模型清单（custom_models.json；visible=false 的条目不上架） */
+/** 模型清单（custom_models.json；visible=false 的条目不上架；context_window 透传供上下文百分比） */
 export function loadCustomModels(): ModelOption[] {
   const cfg = resolveDeployConfig('custom_models.json');
   const items = Array.isArray(cfg?.models) ? cfg.models : [];
@@ -39,6 +39,7 @@ export function loadCustomModels(): ModelOption[] {
     .map((m: any) => ({
       name: String(m.model_name),
       meta: String(m.model_description || m.description || ''),
+      contextWindow: typeof m.context_window === 'number' && m.context_window > 0 ? m.context_window : null,
     }));
 }
 
@@ -50,14 +51,18 @@ export interface BootPreferences {
   effort: EffortLevel;
   workMode: string;
   permMode: string;
+  /** 自动深度压缩开关与触发阈值（上下文百分比分母，对齐 web 端算法） */
+  autoDeepCompress: boolean;
+  deepCompressLimit: number;
 }
 
 const EFFORT_VALUES: EffortLevel[] = ['default', 'low', 'medium', 'high', 'xhigh', 'max'];
 
-/** 个性化默认（personalization.json：默认模型/运行模式/推理强度/工作模式/权限模式） */
+/** 个性化默认（personalization.json：默认模型/运行模式/推理强度/工作模式/权限模式/压缩设置） */
 export function loadPersonalizationDefaults(): BootPreferences {
   const p = readJson(resolve(HOST_DATA_DIR, 'personalization.json')) ?? {};
   const effortRaw = String(p.default_reasoning_effort ?? 'default');
+  const limitRaw = Number(p.deep_compress_trigger_tokens);
   return {
     model: String(p.default_model ?? ''),
     runMode: String(p.default_run_mode ?? 'fast'),
@@ -65,6 +70,8 @@ export function loadPersonalizationDefaults(): BootPreferences {
     effort: (EFFORT_VALUES as string[]).includes(effortRaw) ? (effortRaw as EffortLevel) : 'default',
     workMode: String(p.default_work_mode ?? 'ask'),
     permMode: String(p.default_permission_mode ?? 'unrestricted'),
+    autoDeepCompress: !!p.auto_deep_compress_enabled,
+    deepCompressLimit: Number.isFinite(limitRaw) && limitRaw > 0 ? limitRaw : 150000,
   };
 }
 

cli/src/menu.tsx

@@ -20,15 +20,16 @@ import { padEndWidth } from './width';
 import {
   BOUNDARY_PANELS,
   MODEL_OPTIONS,
-  WORKFLOWS,
-  CHECKPOINTS,
   type AgentMock,
+  type CheckpointMock,
+  type ContextStats,
   type EffortLevel,
   type PathAccess,
   type PathAuth,
   type PendingApprovalMock,
   type SessionMock,
   type TaskMock,
+  type WorkflowMock,
 } from './data';
 
 const FG = RGBA.defaultForeground();
@@ -63,6 +64,12 @@ export interface SlashMenuProps {
   pathAuths: PathAuth[];
   pendingApproval: PendingApprovalMock | null;
   activeWorkflow: string | null;
+  /** 上下文统计（/context 面板；token_update 事件与打开时查询双源刷新） */
+  contextStats: ContextStats;
+  /** 工作流库（/workflow；打开时经 API 刷新） */
+  workflows: WorkflowMock[];
+  /** 版本回溯检查点（/rewind；打开时经 API 刷新） */
+  checkpoints: CheckpointMock[];
   /** 当前对话标题（/rename /delete /export 面板显示用） */
   currentSessionTitle: string;
 }
@@ -164,7 +171,7 @@ export function SlashMenu(props: SlashMenuProps) {
     case 'path':
       return <PathsPanel sel={props.sel} width={props.width} pathAuths={props.pathAuths} group={props.pathGroup} />;
     case 'context':
-      return <ContextPanel width={props.width} />;
+      return <ContextPanel width={props.width} stats={props.contextStats} model={props.model} />;
     case 'rename':
       return <InputPanel title="重命名对话" hint={`当前  ${props.currentSessionTitle}`} width={props.width} />;
     case 'export':
@@ -176,9 +183,9 @@ export function SlashMenu(props: SlashMenuProps) {
     case 'approvals':
       return <ApprovalsPanel sel={props.sel} width={props.width} pending={props.pendingApproval} />;
     case 'workflow':
-      return <WorkflowPanel sel={props.sel} width={props.width} workflows={WORKFLOWS} active={props.activeWorkflow} />;
+      return <WorkflowPanel sel={props.sel} width={props.width} workflows={props.workflows} active={props.activeWorkflow} />;
     case 'rewind':
-      return <RewindPanel sel={props.sel} width={props.width} checkpoints={CHECKPOINTS} />;
+      return <RewindPanel sel={props.sel} width={props.width} checkpoints={props.checkpoints} />;
     case 'help':
       return <HelpPanel sel={props.sel} width={props.width} />;
     case 'agents':
@@ -212,9 +219,9 @@ export function panelItemCount(props: SlashMenuProps): number {
     case 'approvals':
       return 0; // 单条待审批，无 ↑↓ 列表；←→ 选择操作
     case 'workflow':
-      return WORKFLOWS.length;
+      return props.workflows.length;
     case 'rewind':
-      return CHECKPOINTS.length;
+      return props.checkpoints.length;
     case 'help':
       return helpMaxOffset() + 1; // sel 在此面板作为滚动偏移
     case 'context':

cli/src/menuState.ts

@@ -10,24 +10,28 @@ import { panelItemCount, type BoundaryKind, type SlashMenuProps } from './menu';
 import type { ModelStep } from './panels/model';
 import { APPROVAL_ACTIONS } from './panels/approvals';
 import {
+  BOOT_STATE,
   BOUNDARY_PANELS,
-  CHECKPOINTS,
   EFFORT_LEVELS,
   EFFORT_META,
+  EMPTY_CONTEXT_STATS,
   INITIAL_AGENTS,
   INITIAL_PATH_AUTHS,
   INITIAL_TASKS,
   MODEL_OPTIONS,
   PATH_ACCESS_LABEL,
   SESSIONS,
-  WORKFLOWS,
+  formatContextUsage,
   type AgentMock,
+  type CheckpointMock,
+  type ContextStats,
   type EffortLevel,
   type PathAccess,
   type PathAuth,
   type PendingApprovalMock,
   type SessionMock,
   type TaskMock,
+  type WorkflowMock,
 } from './data';
 import type { TimelineApi } from './timeline';
 
@@ -58,6 +62,15 @@ export function useSlashMenu({
   width,
   elapsed,
   onApprovalAction,
+  onSessionLoad,
+  onContextRefresh,
+  onSaveModelDefaults,
+  onSavePathAuths,
+  onAgentsRefresh,
+  onTasksRefresh,
+  onWorkflowsRefresh,
+  onCheckpointsRefresh,
+  onNewSession,
 }: {
   textareaRef: React.RefObject<TextareaRenderable | null>;
   apiRef: React.RefObject<TimelineApi | null>;
@@ -66,24 +79,40 @@ export function useSlashMenu({
   elapsed: number;
   /** 审批裁决回调（正式版由 runtime 注入，调 Gateway decision 端点）；缺省只上屏系统消息 */
   onApprovalAction?: (action: 'run' | 'reject' | 'unrestricted', approval: PendingApprovalMock) => void;
+  /** /session Enter：加载选中对话（runtime.loadSession；缺省只上屏系统消息） */
+  onSessionLoad?: (session: SessionMock) => void;
+  /** /context 打开：触发一次 token 统计查询（gw.getTokenStats；结果经 setContextStats 回流） */
+  onContextRefresh?: () => void;
+  /** /model Enter：保存为默认（gw.savePersonalization patch）；失败由注入方上屏提示 */
+  onSaveModelDefaults?: (v: { model: string; thinking: boolean; effort: EffortLevel }) => void;
+  /** /path 增删后：全量保存两组授权（gw.savePathAuths）；失败由注入方上屏提示 */
+  onSavePathAuths?: (auths: PathAuth[]) => void;
+  /** /agents /tasks 打开：拉真实列表（gw.listSubAgents/listBackgroundCommands，结果经 setter 回流） */
+  onAgentsRefresh?: () => void;
+  onTasksRefresh?: () => void;
+  /** /workflow /rewind 打开：拉真实列表（gw.listWorkflows/listCheckpoints，结果经 setter 回流） */
+  onWorkflowsRefresh?: () => void;
+  onCheckpointsRefresh?: () => void;
+  /** /new：进入空对话草稿（runtime.enterNewSession；对齐 web /new 页面语义） */
+  onNewSession?: () => void;
 }) {
   // / 菜单：menuStack=null 关闭；['commands'] 命令层；级联 push（如 ['commands','model']）
   const [menuStack, setMenuStack] = useState<PanelKind[] | null>(null);
   const [query, setQuery] = useState('');
   const [sel, setSel] = useState(0);
-  // /model 两级流程
+  // /model 两级流程（初值 = boot 加载的真实快照 BOOT_STATE）
-  const [model, setModel] = useState('kimi-k2.6');
+  const [model, setModel] = useState(BOOT_STATE.model);
-  const [pendingModel, setPendingModel] = useState('kimi-k2.6');
+  const [pendingModel, setPendingModel] = useState(BOOT_STATE.model);
   const [modelStep, setModelStep] = useState<ModelStep>('model');
-  const [thinking, setThinking] = useState(true);
+  const [thinking, setThinking] = useState(BOOT_STATE.thinking);
-  const [effort, setEffort] = useState<EffortLevel>('high');
+  const [effort, setEffort] = useState<EffortLevel>(BOOT_STATE.effort);
   // 面板内编辑中的强度（Enter 才提交到 effort；Esc 放弃不影响已生效值）
-  const [pendingEffort, setPendingEffort] = useState<EffortLevel>('high');
+  const [pendingEffort, setPendingEffort] = useState<EffortLevel>(BOOT_STATE.effort);
-  // 边界（value 与 mock.BOUNDARY_PANELS 对齐）
+  // 边界（value 与 BOUNDARY_PANELS 对齐）
-  const [workMode, setWorkMode] = useState('ask');
+  const [workMode, setWorkMode] = useState(BOOT_STATE.workMode);
-  const [permMode, setPermMode] = useState('unrestricted');
+  const [permMode, setPermMode] = useState(BOOT_STATE.permMode);
-  const [execEnv, setExecEnv] = useState('direct');
+  const [execEnv, setExecEnv] = useState(BOOT_STATE.execEnv);
-  const [network, setNetwork] = useState('restricted');
+  const [network, setNetwork] = useState(BOOT_STATE.network);
   // 路径授权（pathGroup = 当前显示的分组；←→ 切换显示组，不做行内权限切换）
   const [pathAuths, setPathAuths] = useState<PathAuth[]>(INITIAL_PATH_AUTHS.map((p) => ({ ...p })));
   const [pathGroup, setPathGroup] = useState<PathAccess>('rw');
@@ -94,17 +123,22 @@ export function useSlashMenu({
   const [sessions, setSessions] = useState<SessionMock[]>(SESSIONS.map((s) => ({ ...s })));
   const [pendingApproval, setPendingApproval] = useState<PendingApprovalMock | null>(null);
   const [activeWorkflow, setActiveWorkflow] = useState<string | null>(null);
+  // 上下文统计（/context 面板 + 状态栏；token_update 事件与打开时查询双源刷新）
+  const [contextStats, setContextStats] = useState<ContextStats>({ ...EMPTY_CONTEXT_STATS });
+  // 工作流库与检查点（/workflow /rewind；打开时经 API 刷新）
+  const [workflows, setWorkflows] = useState<WorkflowMock[]>([]);
+  const [checkpoints, setCheckpoints] = useState<CheckpointMock[]>([]);
 
   // useKeyboard 回调只注册一次，全部通过 ref 读最新状态，避免过期闭包
   const stateRef = useRef({
     menuStack, query, sel, model, pendingModel, modelStep, thinking, effort, pendingEffort,
     workMode, permMode, execEnv, network, pathAuths, pathGroup, agents, tasks,
-    sessions, pendingApproval, activeWorkflow,
+    sessions, pendingApproval, activeWorkflow, contextStats, workflows, checkpoints,
   });
   stateRef.current = {
     menuStack, query, sel, model, pendingModel, modelStep, thinking, effort, pendingEffort,
     workMode, permMode, execEnv, network, pathAuths, pathGroup, agents, tasks,
-    sessions, pendingApproval, activeWorkflow,
+    sessions, pendingApproval, activeWorkflow, contextStats, workflows, checkpoints,
   };
 
   const currentSessionTitle = sessions.find((s) => s.current)?.title ?? '';
@@ -129,6 +163,9 @@ export function useSlashMenu({
         pathAuths,
         pendingApproval,
         activeWorkflow,
+        contextStats,
+        workflows,
+        checkpoints,
         currentSessionTitle,
       }
     : null;
@@ -158,6 +195,9 @@ export function useSlashMenu({
       pathAuths: stateRef.current.pathAuths,
       pendingApproval: stateRef.current.pendingApproval,
       activeWorkflow: stateRef.current.activeWorkflow,
+      contextStats: stateRef.current.contextStats,
+      workflows: stateRef.current.workflows,
+      checkpoints: stateRef.current.checkpoints,
       currentSessionTitle: stateRef.current.sessions.find((s) => s.current)?.title ?? '',
     } satisfies SlashMenuProps);
 
@@ -194,7 +234,8 @@ export function useSlashMenu({
         sys('/compact  压缩上下文（CLI 尚未接入）');
         break;
       case 'new':
-        sys('已开始新对话（CLI 尚未接入）');
+        // 对齐 web /new 页面：进入空对话草稿（不立即建会话），首条消息惰性创建
+        onNewSession?.();
         break;
       default:
         sys(`/${cmd.name}  ${cmd.desc}（CLI 尚未接入）`);
@@ -219,8 +260,23 @@ export function useSlashMenu({
       case 'session':
         setSel(Math.max(0, st.sessions.findIndex((s) => s.current)));
         break;
+      case 'context':
+        // 打开时触发一次统计查询（结果经 setContextStats 回流；事件流为运行期另一源）
+        onContextRefresh?.();
+        break;
+      case 'agents':
+        onAgentsRefresh?.();
+        break;
+      case 'tasks':
+        onTasksRefresh?.();
+        break;
       case 'workflow':
-        setSel(Math.max(0, WORKFLOWS.findIndex((w) => w.name === st.activeWorkflow)));
+        onWorkflowsRefresh?.();
+        setSel(Math.max(0, st.workflows.findIndex((w) => w.name === st.activeWorkflow)));
+        break;
+      case 'rewind':
+        onCheckpointsRefresh?.();
+        setSel(0);
         break;
       case 'approvals':
         // 单条待审批：光标落在第一个操作「运行」上
@@ -262,8 +318,8 @@ export function useSlashMenu({
     if (top === 'session') {
       const s = st.sessions[st.sel];
       if (s && !s.current) {
-        setSessions((prev) => prev.map((x) => ({ ...x, current: x.title === s.title })));
+        setSessions((prev) => prev.map((x) => ({ ...x, current: x.id === s.id })));
-        sys(`已切换到对话「${s.title}」（CLI 尚未接入）`);
+        onSessionLoad?.(s);
       }
       closeMenu({ clearText: true });
       return;
@@ -332,7 +388,7 @@ export function useSlashMenu({
     }
 
     if (top === 'workflow') {
-      const w = WORKFLOWS[st.sel];
+      const w = st.workflows[st.sel];
       if (!w) return;
       if (st.activeWorkflow === w.name) {
         setActiveWorkflow(null);
@@ -346,7 +402,7 @@ export function useSlashMenu({
     }
 
     if (top === 'rewind') {
-      const c = CHECKPOINTS[st.sel];
+      const c = st.checkpoints[st.sel];
       if (c) sys(`已回溯到检查点「${c.summary}」（${c.when}；CLI 尚未接入）`);
       closeMenu({ clearText: true });
       return;
@@ -376,6 +432,8 @@ export function useSlashMenu({
       setThinking(toThinking);
       setEffort(st.pendingEffort);
       if (before !== after) sys(`已切换：${before} → ${after}`);
+      // 同步保存为个性化默认（影响新对话；当前会话经 createTask 覆盖参数即时生效）
+      onSaveModelDefaults?.({ model: st.pendingModel, thinking: toThinking, effort: st.pendingEffort });
       closeMenu({ clearText: true });
       return;
     }
@@ -402,10 +460,12 @@ export function useSlashMenu({
       const text = (textareaRef.current?.plainText ?? '').trim();
       if (text) {
         const access = st.pathGroup;
-        setPathAuths((prev) => [...prev, { path: text, access }]);
+        const next = [...st.pathAuths, { path: text, access }];
+        setPathAuths(next);
         setInputText('');
         setSel(st.pathAuths.filter((p) => p.access === access).length); // 选中新添加的行
         sys(`已添加路径授权：${text}（${PATH_ACCESS_LABEL[access]}）`);
+        onSavePathAuths?.(next);
       }
       return;
     }
@@ -530,8 +590,10 @@ export function useSlashMenu({
         key.preventDefault();
         // sel 是组内下标，换算成全量数组下标再删
         const target = groupRows[st.sel];
-        setPathAuths((prev) => prev.filter((p) => p !== target));
+        const next = st.pathAuths.filter((p) => p !== target);
+        setPathAuths(next);
         setSel((s) => Math.max(0, Math.min(s, groupRows.length - 2)));
+        onSavePathAuths?.(next);
         return true;
       }
       return false;
@@ -609,7 +671,7 @@ export function useSlashMenu({
     workMode: BOUNDARY_PANELS.mode.options.find((o) => o.value === workMode)?.label ?? workMode,
     permMode: BOUNDARY_PANELS.permission.options.find((o) => o.value === permMode)?.label ?? permMode,
     execEnv: BOUNDARY_PANELS.env.options.find((o) => o.value === execEnv)?.label ?? execEnv,
-    contextUsage: '—',
+    contextUsage: formatContextUsage(contextStats, model),
   };
 
   return {
@@ -621,5 +683,11 @@ export function useSlashMenu({
     closeMenu,
     openApproval,
     resetMenu,
+    setContextStats,
+    setAgents,
+    setTasks,
+    setWorkflows,
+    setCheckpoints,
+    setSessions,
   };
 }

cli/src/panels/context.tsx

@@ -1,27 +1,41 @@
 // /context 面板：交互区只读展示上下文用量统计
+// 数据：ContextStats 数字原始值（token_update 事件 + 打开时查询双源刷新）；
+// 百分比算法对齐 web InputComposer（开自动深度压缩用触发阈值，否则模型上下文窗口）。
 import { RGBA, TextAttributes } from '@opentui/core';
 import { T } from '../components';
 import { PanelFrame } from './frame';
-import { CONTEXT_STATS } from '../data';
+import { contextUsageLimitFor, formatCompactTokens, type ContextStats } from '../data';
+import { t } from '../i18n';
 
 const FG = RGBA.defaultForeground();
 const DIM = TextAttributes.DIM;
 
-export function ContextPanel({ width }: { width: number }) {
+export function ContextPanel({ width, stats, model }: { width: number; stats: ContextStats; model: string }) {
-  // 标签列对齐：最长 5 个汉字，短的补全角空格
+  const limit = contextUsageLimitFor(model);
+  const pct =
+    limit > 0 ? Math.max(0, Math.min(100, Math.round((stats.currentTokens / limit) * 100))) : 0;
+  const hitRate =
+    stats.totalInput > 0 ? `${Math.round((stats.cacheInput / stats.totalInput) * 100)}%` : '—';
+  const currentText =
+    limit > 0
+      ? `${formatCompactTokens(stats.currentTokens)} / ${formatCompactTokens(limit)}（${pct}%）`
+      : `${formatCompactTokens(stats.currentTokens)} / —`;
+
   const rows: Array<[string, string]> = [
-    ['当前上下文', `${CONTEXT_STATS.used} / ${CONTEXT_STATS.total}（${CONTEXT_STATS.percent}%）`],
+    [t('context.current'), currentText],
-    ['累计输入　', CONTEXT_STATS.totalInput],
+    [t('context.totalInput'), formatCompactTokens(stats.totalInput)],
-    ['累计输出　', CONTEXT_STATS.totalOutput],
+    [t('context.totalOutput'), formatCompactTokens(stats.totalOutput)],
-    ['缓存输入　', CONTEXT_STATS.cacheInput],
+    [t('context.cacheInput'), formatCompactTokens(stats.cacheInput)],
-    ['缓存命中率', CONTEXT_STATS.cacheHitRate],
+    [t('context.cacheHitRate'), hitRate],
   ];
   return (
-    <PanelFrame title="上下文用量   Esc 关闭" width={width}>
+    <PanelFrame title={t('context.title')} width={width}>
       {rows.map(([label, value]) => (
         <box key={label} flexDirection="row">
           <T fg={FG}>{`  ${label}  `}</T>
-          <T fg={FG} attributes={DIM}>{value}</T>
+          <T fg={FG} attributes={DIM}>
+            {value}
+          </T>
         </box>
       ))}
     </PanelFrame>

cli/src/runtime.ts

@@ -2,7 +2,7 @@
 // 协议依据 docs/runtime_protocol.md：run.start(POST /api/tasks) → run.events(GET ?from=offset)；
 // 事件窗口缺口（offset < window_start）按 §5.2 提示并对齐续读（CLI 不做快照对账）。
 import type { GatewayClient, TaskPollResult } from './gateway';
-import type { PendingApprovalMock } from './data';
+import type { ContextStats, PendingApprovalMock } from './data';
 import { CPS_INSTANT, type TimelineApi } from './timeline';
 
 const POLL_INTERVAL_MS = 500;
@@ -25,6 +25,10 @@ export interface RuntimeCallbacks {
   onApprovalRequired(approval: PendingApprovalMock): void;
   /** 系统消息（审批回执/错误/窗口缺口提示等），经 menuState 的 sys 通道上屏 */
   onSystemMessage(text: string): void;
+  /** token 统计更新（token_update 事件流）：刷新状态栏与 /context 面板 */
+  onTokenUpdate?(stats: ContextStats): void;
+  /** 首条消息触发服务端惰性建会话后回调（/new 草稿 → 真实会话，列表插入用） */
+  onConversationCreated?(id: string): void;
   /** 审批操作翻译函数（i18n t()） */
   tr(key: string): string;
 }
@@ -46,18 +50,37 @@ export class ChatRuntime {
     this.destroyed = true;
   }
 
-  /** 发送一条用户消息并启动事件轮询（空闲时调用；运行中走排队由上层管理） */
+  /** /new（对齐 web 端 /new 页面语义）：进入空对话草稿状态——不立即创建会话，
-  async send(text: string): Promise<void> {
+   *  清空时间线并置空 conversationId；下一条 send 不带 cid，服务端惰性创建。 */
+  enterNewSession(): void {
+    this.conversationId = '';
+    this.taskId = '';
+    this.offset = 0;
+    this.cb.api.reset();
+  }
+
+  /** 发送一条用户消息并启动事件轮询（空闲时调用；运行中走排队由上层管理）。
+   *  opts 为会话级模型/模式覆盖（/model 面板生效值），缺省由服务端用对话/个性化默认。 */
+  async send(text: string, opts?: { model_key?: string; run_mode?: string; thinking_mode?: boolean }): Promise<void> {
     this.cb.api.addUser(text);
+    const wasDraft = !this.conversationId;
     let accepted: { task_id: string; conversation_id?: string };
     try {
-      accepted = await this.gw.createTask({ message: text, conversation_id: this.conversationId || undefined });
+      accepted = await this.gw.createTask({
+        message: text,
+        conversation_id: this.conversationId || undefined,
+        model_key: opts?.model_key,
+        run_mode: opts?.run_mode,
+        thinking_mode: opts?.thinking_mode,
+      });
     } catch (err) {
       this.cb.onSystemMessage(`${this.cb.tr('runtime.sendFailed')}${err instanceof Error ? err.message : String(err)}`);
       return;
     }
     this.taskId = accepted.task_id;
     if (accepted.conversation_id) this.conversationId = accepted.conversation_id;
+    // /new 草稿的首条消息：服务端已建会话，通知上层把新会话插入列表
+    if (wasDraft && accepted.conversation_id) this.cb.onConversationCreated?.(accepted.conversation_id);
     this.offset = 0;
     this.setRunning(true);
     void this.pollLoop();
@@ -72,6 +95,50 @@ export class ChatRuntime {
     }
   }
 
+  /** 加载既有对话（/session Enter）：清空时间线 → 拉历史 → 按序渲染 user/assistant 文本。
+   *  历史中的工具/思考块不重建（保持轻量）；切换后发送消息即挂到该对话。 */
+  async loadSession(conversationId: string): Promise<void> {
+    this.conversationId = conversationId;
+    this.taskId = '';
+    this.offset = 0;
+    this.cb.api.reset();
+    try {
+      const res = await this.gw.getSessionHistory(conversationId);
+      const messages = Array.isArray(res?.messages) ? res.messages : Array.isArray(res) ? res : [];
+      for (const m of messages) {
+        const role = m?.role;
+        const text = extractText(m?.content);
+        if (!text) continue;
+        if (role === 'user') {
+          this.cb.api.addUser(text);
+        } else if (role === 'assistant') {
+          this.cb.api.startAssistant();
+          this.cb.api.appendAssistant(text);
+        }
+      }
+      this.cb.onSystemMessage(`${this.cb.tr('session.loaded')}${conversationId}`);
+    } catch (err) {
+      this.cb.onSystemMessage(`${this.cb.tr('session.loadFailed')}${err instanceof Error ? err.message : String(err)}`);
+    }
+  }
+
+  /** 闲时拉一次会话 token 统计（/context 面板打开时调用；失败静默，面板保留上次值） */
+  async refreshTokenStats(): Promise<void> {
+    if (!this.conversationId) return;
+    try {
+      const stats = await this.gw.getTokenStats(this.conversationId);
+      this.cb.onTokenUpdate?.({
+        currentTokens: Number(stats.current_context_tokens ?? 0),
+        totalInput: Number(stats.total_input_tokens ?? 0),
+        totalOutput: Number(stats.total_output_tokens ?? 0),
+        cacheInput: Number(stats.total_cached_input_tokens ?? 0),
+        cacheExemptInput: Number(stats.cache_exempt_input_tokens ?? 0),
+      });
+    } catch {
+      // 静默：新对话无统计属正常
+    }
+  }
+
   private setRunning(v: boolean): void {
     if (this.running === v) return;
     this.running = v;
@@ -165,6 +232,16 @@ export class ChatRuntime {
       case 'task_stopped':
         api.addSystem(this.cb.tr('runtime.stopped'));
         break;
+      case 'token_update':
+        // 运行期 token 统计推送（任务期间 context_manager 回调切入事件流）
+        this.cb.onTokenUpdate?.({
+          currentTokens: Number(data.current_context_tokens ?? 0),
+          totalInput: Number(data.cumulative_input_tokens ?? 0),
+          totalOutput: Number(data.cumulative_output_tokens ?? 0),
+          cacheInput: Number(data.cumulative_cached_input_tokens ?? 0),
+          cacheExemptInput: Number(data.cache_exempt_input_tokens ?? 0),
+        });
+        break;
       case 'error':
         api.addSystem(`${this.cb.tr('runtime.error')}${String(data.message ?? data.error ?? '')}`);
         break;
@@ -186,6 +263,18 @@ function isTerminalStatus(status: string): boolean {
   return ['succeeded', 'failed', 'stopped', 'canceled', 'completed', 'done', 'error'].includes((status || '').toLowerCase());
 }
 
+/** 历史消息 content → 纯文本（string 直返；多段数组拼接 text 段；其余空） */
+function extractText(content: any): string {
+  if (typeof content === 'string') return content;
+  if (Array.isArray(content)) {
+    return content
+      .map((seg) => (typeof seg === 'string' ? seg : typeof seg?.text === 'string' ? seg.text : ''))
+      .filter(Boolean)
+      .join('\n');
+  }
+  return '';
+}
+
 /** 工具参数逐行展示（审批/工具块共用）：每个参数一行，多行值展开原样 */
 function toolParamLines(toolName: string, args: any): string[] {
   if (!args || typeof args !== 'object') return [];

cli/src/timeline.ts

@@ -25,6 +25,8 @@ export type Block =
   | { kind: 'assistant'; id: number; full: string; revealStart: number; cps: number };
 
 export interface TimelineApi {
+  /** 清空全部块（/session 切换对话时用） */
+  reset(): void;
   addUser(text: string): void;
   addGuide(text: string): void;
   addSystem(text: string): void;

server/chat/permission.py

@@ -45,6 +45,7 @@ from core.web_terminal import WebTerminal
 from config.model_profiles import get_model_context_window
 
 from server.auth_helpers import api_login_required, resolve_admin_policy, get_current_user_record, get_current_username
+from server.gateway_auth import api_login_or_host_token_required
 from server.context import with_terminal, get_gui_manager, get_upload_guard, build_upload_error_response, ensure_conversation_loaded, get_or_create_usage_tracker, get_user_resources
 from server.security import rate_limited
 from server.utils_common import debug_log
@@ -489,7 +490,7 @@ def update_work_mode(terminal: WebTerminal, workspace: UserWorkspace, username:
     })
 
 @chat_bp.route('/api/path-authorization', methods=['GET'])
-@api_login_required
+@api_login_or_host_token_required
 @with_terminal
 def get_path_authorization(terminal: WebTerminal, workspace: UserWorkspace, username: str):
     is_host = bool(getattr(terminal, "_is_host_mode", lambda: False)())
@@ -506,7 +507,7 @@ def get_path_authorization(terminal: WebTerminal, workspace: UserWorkspace, user
     })
 
 @chat_bp.route('/api/path-authorization', methods=['POST'])
-@api_login_required
+@api_login_or_host_token_required
 @with_terminal
 @rate_limited("path_authorization_update", 20, 60, scope="user")
 def update_path_authorization(terminal: WebTerminal, workspace: UserWorkspace, username: str):

server/chat/settings.py

@@ -38,6 +38,7 @@ from core.tool_loading import build_registry_payload
 from config.model_profiles import get_model_context_window
 
 from server.auth_helpers import api_login_required, resolve_admin_policy, get_current_user_record, get_current_username
+from server.gateway_auth import api_login_or_host_token_required
 from server.context import with_terminal, get_gui_manager, get_upload_guard, build_upload_error_response, ensure_conversation_loaded, get_or_create_usage_tracker
 from server.security import rate_limited
 from server.utils_common import debug_log
@@ -268,7 +269,7 @@ def update_model(terminal: WebTerminal, workspace: UserWorkspace, username: str)
         return jsonify({"success": False, "error": str(exc), "message": str(exc)}), code
 
 @chat_bp.route('/api/personalization', methods=['GET'])
-@api_login_required
+@api_login_or_host_token_required
 @with_terminal
 def get_personalization_settings(terminal: WebTerminal, workspace: UserWorkspace, username: str):
     """获取个性化配置"""
@@ -317,7 +318,7 @@ def get_personalization_settings(terminal: WebTerminal, workspace: UserWorkspace
         return jsonify({"success": False, "error": str(exc)}), 500
 
 @chat_bp.route('/api/personalization', methods=['POST'])
-@api_login_required
+@api_login_or_host_token_required
 @with_terminal
 @rate_limited("personalization_update", 20, 300, scope="user")
 def update_personalization_settings(terminal: WebTerminal, workspace: UserWorkspace, username: str):

server/conversation.py

@@ -71,6 +71,7 @@ from utils.conversation_manager import ConversationManager
 from utils.api_client import APIClient
 
 from .auth_helpers import api_login_required, resolve_admin_policy, get_current_user_record, get_current_username
+from .gateway_auth import api_login_or_host_token_required
 from .context import with_terminal, get_gui_manager, get_upload_guard, build_upload_error_response, ensure_conversation_loaded, reset_system_state, get_user_resources, get_or_create_usage_tracker
 from .utils_common import (
     build_review_lines,
@@ -1307,7 +1308,7 @@ def update_conversation_versioning(conversation_id, terminal: WebTerminal, works
 
 
 @conversation_bp.route('/api/conversations/<conversation_id>/versioning/checkpoints', methods=['GET'])
-@api_login_required
+@api_login_or_host_token_required
 @with_terminal
 def list_conversation_versioning_checkpoints(conversation_id, terminal: WebTerminal, workspace: UserWorkspace, username: str):
     try:
@@ -1820,7 +1821,7 @@ def cancel_conversation_compression(conversation_id, terminal: WebTerminal, work
 
 
 @conversation_bp.route('/api/sub_agents', methods=['GET'])
-@api_login_required
+@api_login_or_host_token_required
 @with_terminal
 def list_sub_agents(terminal: WebTerminal, workspace: UserWorkspace, username: str):
     """返回当前对话的子智能体任务列表。"""
@@ -1833,7 +1834,8 @@ def list_sub_agents(terminal: WebTerminal, workspace: UserWorkspace, username: s
             manager._load_state()
         except Exception:
             pass
-        conversation_id = terminal.context_manager.current_conversation_id
+        # CLI 等 Bearer 客户端经 query 显式指定对话（其装配的 terminal 无当前对话概念）
+        conversation_id = (request.args.get("conversation_id") or "").strip() or terminal.context_manager.current_conversation_id
         data = manager.get_overview(conversation_id=conversation_id)
 
         # 传统模式子智能体列表必须排除多智能体任务，避免 /new 等无当前对话场景
@@ -2067,7 +2069,7 @@ def terminate_sub_agent(task_id: str, terminal: WebTerminal, workspace: UserWork
 
 
 @conversation_bp.route('/api/background_commands', methods=['GET'])
-@api_login_required
+@api_login_or_host_token_required
 @with_terminal
 def list_background_commands(terminal: WebTerminal, workspace: UserWorkspace, username: str):
     """返回当前对话的后台 run_command 列表。"""
@@ -2075,7 +2077,8 @@ def list_background_commands(terminal: WebTerminal, workspace: UserWorkspace, us
     if not manager:
         return jsonify({"success": True, "data": []})
     try:
-        conversation_id = terminal.context_manager.current_conversation_id
+        # CLI 等 Bearer 客户端经 query 显式指定对话（其装配的 terminal 无当前对话概念）
+        conversation_id = (request.args.get("conversation_id") or "").strip() or terminal.context_manager.current_conversation_id
         limit_raw = request.args.get("limit", "200")
         try:
             limit_num = max(1, min(int(limit_raw), 1000))

server/gateway_api.py

@@ -108,4 +108,24 @@ def get_runtime_session_history(conversation_id: str):
     return jsonify({"success": True, "conversation": result})
 
 
+@gateway_bp.route("/api/runtime/sessions/<conversation_id>/token-stats", methods=["GET"])
+@api_login_or_host_token_required
+def get_runtime_session_token_stats(conversation_id: str):
+    """session.token_stats：会话 token 统计（累计输入/输出/缓存/当前上下文）。"""
+    workspace_id = _resolve_workspace_id(request.args.get("workspace_id", ""))
+    username = str(session.get("username") or "")
+    principal = principal_from_session_snapshot(dict(session), workspace_id, username)
+    try:
+        stats = runtime_service.get_session_token_stats(
+            username, workspace_id, conversation_id, principal
+        )
+        return jsonify({"success": True, "stats": stats})
+    except PermissionError as exc:
+        return jsonify({"success": False, "error": str(exc)}), 403
+    except ValueError as exc:
+        return jsonify({"success": False, "error": str(exc)}), 400
+    except RuntimeError as exc:
+        return jsonify({"success": False, "error": str(exc)}), 503
+
+
 __all__ = ["gateway_bp"]

server/headless_app.py

@@ -27,7 +27,7 @@ import secrets
 from datetime import timedelta
 from pathlib import Path
 
-from flask import Flask
+from flask import Flask, jsonify, request
 
 from config import (
     DEFAULT_PROJECT_PATH,
@@ -151,10 +151,48 @@ def create_headless_app() -> Flask:
     for _name, bp in HEADLESS_BLUEPRINTS:
         app.register_blueprint(bp)
 
+    # host 全局设置面（personalization / path-authorization）：视图函数定义在 chat_bp 模块内，
+    # 装饰器已双通道化（web session 或 host Bearer 均可）；此处直接注册到 headless app，
+    # URL 与 full 形态一致（Blueprint.route 只登记不包装，函数对象可安全复用）。
+    from server.chat.settings import (
+        get_personalization_settings,
+        update_personalization_settings,
+    )
+    from server.chat.permission import (
+        get_path_authorization,
+        update_path_authorization,
+    )
+    from server.conversation import list_background_commands, list_sub_agents, list_conversation_versioning_checkpoints
+    from server.workflow_page import api_list_workflows
+
+    app.add_url_rule('/api/personalization', view_func=get_personalization_settings, methods=['GET'])
+    app.add_url_rule('/api/personalization', view_func=update_personalization_settings, methods=['POST'])
+    app.add_url_rule('/api/path-authorization', view_func=get_path_authorization, methods=['GET'])
+    app.add_url_rule('/api/path-authorization', view_func=update_path_authorization, methods=['POST'])
+    # 子智能体/后台指令列表（/agents /tasks 面板数据源；conversation_id 走 query 显式指定）
+    app.add_url_rule('/api/sub_agents', view_func=list_sub_agents, methods=['GET'])
+    app.add_url_rule('/api/background_commands', view_func=list_background_commands, methods=['GET'])
+    # 工作流库列表（/workflow）与版本回溯检查点（/rewind）
+    app.add_url_rule('/api/workflows', view_func=api_list_workflows, methods=['GET'])
+    app.add_url_rule(
+        '/api/conversations/<conversation_id>/versioning/checkpoints',
+        view_func=list_conversation_versioning_checkpoints,
+        methods=['GET'],
+    )
+
     @app.route('/')
     def headless_landing():
         return _HEADLESS_LANDING_HTML
 
+    @app.errorhandler(404)
+    def headless_not_found(_err):
+        # 非 API 的 GET 路径（/new、/<conv_id> 等前端路由）统一回落告知页——
+        # 对齐 full 形态的 SPA fallback 行为，避免浏览器打开收藏链接看到裸 404；
+        # API 路径保持 JSON 404。
+        if request.method == 'GET' and not request.path.startswith('/api/'):
+            return _HEADLESS_LANDING_HTML
+        return jsonify({"success": False, "error": "not found"}), 404
+
     return app
 
 

server/runtime/service.py

@@ -240,6 +240,27 @@ class RuntimeService:
         manager = cm._get_conversation_manager_for_id(conversation_id)
         return manager.load_conversation(conversation_id)
 
+    def get_session_token_stats(
+        self,
+        username: str,
+        workspace_id: str,
+        conversation_id: str,
+        principal: Optional[TrustedPrincipal] = None,
+    ) -> Dict[str, Any]:
+        """会话 token 统计查询（session.token_stats）。
+
+        字段由 context_manager 定义（total_input_tokens/total_output_tokens/
+        total_cached_input_tokens/cache_exempt_input_tokens/current_context_tokens 等），
+        本服务只做资源装配与转发。对话不存在或无统计时返回空 dict。
+        """
+        if not str(conversation_id or "").strip():
+            raise ValueError("runtime_context: conversation_id 不能为空")
+        terminal, _workspace = self._resources_for_query(username, workspace_id, principal)
+        cm = getattr(terminal, "context_manager", None)
+        if cm is None:
+            raise RuntimeError(tr("tasks.system_not_initialized"))
+        return cm.get_conversation_token_statistics(conversation_id) or {}
+
     def create_session(
         self,
         username: str,

server/workflow_page.py

@@ -15,6 +15,7 @@ from modules.workflow_manager import (
     save_workflow,
 )
 from server.auth_helpers import api_login_required, login_required
+from server.gateway_auth import api_login_or_host_token_required
 from server.context import with_terminal
 from modules.i18n import tr
 
@@ -39,7 +40,7 @@ def workflow_editor_page(name: str):
 
 
 @workflow_page_bp.route("/api/workflows", methods=["GET"])
-@api_login_required
+@api_login_or_host_token_required
 @with_terminal
 def api_list_workflows(terminal, workspace, username):
     """工作流列表（内置 + 用户库双源合并，仅元信息）。"""