chore(release): bump to v0.7.12 — 适配 kimi-code 0.40.1（flag 清单/优先级/危险命令守卫）+ WebUI 竞态修复 + models-dev 7495/212

chore(data): 同步 models-dev 最新快照（7495 模型 / 212 供应商）
feat: 适配 kimi-code 0.40.1 —— flag 清单/优先级语义/危险命令守卫
2026-09-03 10:12:17 +08:00 · 2026-09-03 10:11:10 +08:00 · 2026-09-03 10:10:52 +08:00 · 2026-09-03 10:10:40 +08:00 · 2026-08-29 11:15:40 +08:00 · 2026-08-29 11:14:11 +08:00
33 changed files with 96102 additions and 68092 deletions
@@ -81,11 +81,11 @@ jobs:

            ### 本次更新

-            - **高级设置**：原「子代理设置」升级为「高级设置」，聚合子代理模型指定、实验功能开关与 WebUI 快捷入口
-            - **仪表盘热力图双击弹窗**：双击热力图方块即可查看当日模型用量分布，每种模型展示 Token / 请求次数 / 费用 / 命中率
-            - **子代理模型指定（实验功能）**：为子代理绑定次主力模型，不再默认继承主模型；实验开关改为滑动开关并修复开启态隐形
-            - **兼容 kimi-code 0.33+（v2 引擎）**：`loop_control` 写入新键名 `max_attempts_per_step`，旧配置自动迁移并保留旧键兼容 `KIMI_CODE_LEGACY_FLAG=1`；hooks 候选补充新事件
-            - models.dev 模型数据同步；官网企业版视觉升级；出品公司标识统一
+            - **模型数据同步（models.dev 2026-08-27）**：7343 个模型 / 203 个供应商
+            - **新增 GLM-5.3 / GLM-5.3-Flash 全线接入**：智谱官方、OpenRouter、Cloudflare、DeepInfra、HuggingFace、火山引擎等约 20 条渠道；GLM-5.3-Flash 定价 $0.075 / $0.25（输入/输出，每百万 token）
+            - 新增 coding plan 渠道：`zhipuai-coding-plan/glm-5.3`、`glm-5.3-highspeed`、`glm-5.3-flash`
+            - 新增 Qwen3.8 系列、火山引擎豆包 Seed 2.x 全系、DeepSeek-V4 GA 版、MiniMax-M2.7/M3 等共 74 个模型条目
+            - 价格更新 58 处、能力信息修正 56 处（详见 release-notes-v0.7.9.md）

            ### Downloads by platform

@@ -123,7 +123,7 @@

 ![会话管理](docs/screenshots/sessions.png)

-按工作区隔离浏览 Kimi Code 会话，支持活跃/已归档/全部筛选；流式逐行预览（20MB 字节上限，500 字符折叠）；归档/取消归档/批量删除。
+按工作区隔离浏览 Kimi Code 会话，支持活跃/已归档/全部筛选；流式逐行预览（20MB 字节上限，500 字符折叠）；归档/取消归档/批量删除。归档写入 `state.json` 元数据（与 CLI v2 一致），旧版本物理归档目录 `.kcd-archive/` 仍可识别与恢复。

 ## 架构总览

@@ -239,7 +239,7 @@

 - 按工作区隔离浏览 Kimi Code 会话，支持活跃/已归档/全部筛选
 - 流式逐行预览（20MB 字节上限，500 字符折叠可展开）
- 归档 / 取消归档 / 批量删除
+- 归档 / 取消归档 / 批量删除（归档写入 `state.json` 元数据，与 CLI v2 一致；旧版本物理归档目录 `.kcd-archive/` 仍可识别与恢复）
 - 早期版本的"闪崩"已通过流式读取 + 限制解决

 ### 设置面板
@@ -543,6 +543,12 @@ A: 需要分两步：
 **Q: 切换会覆盖其他供应商吗？**
 A: 不会。Kimi Code 的 `config.toml` 始终写入全部供应商，仅 `default_model` 决定生效项。这与 CLI 原生 `/provider` 行为一致。

+**Q: 在 Kimi Code CLI 里新增的顶层配置节（如 `[task]`/`[swarm]`）会被保存时删掉吗？**
+A: 不会（v0.7.10 起）。保存时以加载配置那一刻的顶层键为基线，CLI 后加的节原样保留；只有在本应用中显式删除的节才会从 `config.toml` 移除。
+
+**Q: 遇到 Kimi Code 新增的供应商类型（`type` 值）会怎样？**
+A: 原样保留，不再改写为 `kimi`。编辑界面会显示原始类型值（标注"未知类型"）；该供应商不支持的操作（如模型发现）会返回明确错误提示。
+
 **Q: 新增/激活的供应商位置怎么变？**
 A: 自动提升到列表最前，UI 立即反映。

@@ -1,7 +1,7 @@
 {
  "name": "kimiswitch",
  "private": true,
-  "version": "0.7.7",
+  "version": "0.7.12",
  "type": "module",
  "scripts": {
    "dev": "vite",
@@ -0,0 +1,8 @@
+## v0.7.10
+
+### 配置写回保护（适配 kimi-code 0.38+ 配置演进）
+
+- **修复：保存不再误删 CLI 新增的顶层配置节**。此前在 Kimi Switch 加载配置后，若 Kimi Code CLI 往 `config.toml` 写入了新的顶层节（如 0.38+ 的 `[task]` / `[swarm]` / `[cron]` / `[tools]` / `[identity]` / `[token_counting]` 等），回到 Kimi Switch 保存时会把这些节**静默删除**。现在以导入时的顶层键为基线：CLI 后加的节原样保留；在界面中显式删除的节仍会被移除；老数据（无基线，如 SQLite 快照恢复）不删任何键
+- **修复：未知供应商类型不再被改写**。`config.toml` 中 Kimi Code 新增的供应商 `type` 值往返读写后不再被改写成 `kimi`，原样保留；编辑器 API 格式下拉框显示原始值（标注"未知类型"）；模型发现等不支持的操作返回明确错误提示，不再误用 `KIMI_API_KEY`
+- SQLite 快照恢复路径同步修复（此前从快照恢复同样会丢失未知类型）
+- 新增 3 个配置往返回归测试（Rust 测试 106 项全绿，前端 tsc / vitest 通过）
@@ -0,0 +1,22 @@
+## v0.7.11
+
+### 适配 kimi-code 0.39.1
+
+- **修复：`[thinking]` 保留开关不再破坏思考配置**。此前在 UI 关闭"保留思考内容"会写入布尔 `keep = false`，而上游两个引擎都只接受字符串——v2 会把整个 `[thinking]` 节校验失败并丢弃。现改为写 `keep = "off"`，存量配置里的布尔值加载时自动归一化
+- **修复：`loop_control` 不再触发上游废弃告警**。默认只写新键 `max_attempts_per_step` 并清理存量旧键 `max_retries_per_step`；仅检测到 `KIMI_CODE_LEGACY_FLAG`（v1 引擎）时才双写
+- **思考强度移除 `max` 档**（上游已自动迁移为 `high`），存量配置显示归一
+- **实验开关同步上游全集**：新增 remote-control；`wait_for` 与 persistence_minidb_readmodel 已在上游转正默认开启，UI 会标注"默认开启"（显式关闭优先）
+
+### 会话归档与上游 v2 对齐
+
+- **归档不再搬动文件**：改为在会话 `state.json` 写入 `archived`/`archivedAt` 元数据（与 CLI v2 的 `setArchived` 逐字段一致），CLI 侧与 Kimi Switch 侧归档状态从此互通
+- 旧版本物理归档的会话（`.kcd-archive/` 目录）仍可识别为已归档并正常恢复
+- 归档不再改写 `session_index.jsonl`（删除会话的行为不变）
+
+### 子代理模型池修复
+
+- **修复：模型池无法添加第二个条目**。从"仅默认模型"的隐式单条目形态添加条目时，默认模型现在会自动物化进池，不再被"默认模型不在模型池中"的校验拦截
+
+### 模型数据同步
+
+- models.dev 快照刷新至 **7482 个模型 / 207 个供应商**
@@ -0,0 +1,16 @@
+# KimiSwitch v0.7.12
+
+## 适配 kimi-code 0.40.1
+
+- **实验 flag 清单更新（9→10）**：新增 `file_history`（轮级文件历史快照）、`search_worker`（kap-server 全局搜索 worker）；`secondary-model`（子代理次主力模型）自 kimi-code 0.40.1 起默认开启，设置页显示"默认开启"徽章
+- **flag 优先级语义修正**：跟随上游新优先级——单 flag 环境变量 > `[experimental]` 显式配置 > 总开关环境变量（仅强制开）> 默认值。总开关 `KIMI_CODE_EXPERIMENTAL_FLAG` 不再锁定全部开关，显式写入的 false/true 优先生效
+- **环境变量探测补齐**：`tower` / `subagent_fork` / `wait_for` / `auto_session_title` 等此前漏探测的环境变量现在会正确显示"被环境变量锁定"状态
+- **新增危险命令守卫开关**：权限设置页可配置 `[permission] dangerous_command_guard`（默认开启）——Auto 模式直接拒绝 `rm -rf` / `shutdown` 等危险命令，其它模式强制询问，关闭后回归旧行为
+
+## 修复
+
+- **WebUI 打开竞态**：首次打开较慢时多次点击"在应用内打开"不再报错——并发调用串行化 + 后到的点击自动聚焦已有窗口；失败路径不再误杀正在使用的 server；前端按钮增加"打开中..."防重入态
+
+## 数据
+
+- models-dev 快照更新至 7495 模型 / 212 供应商
@@ -0,0 +1,13 @@
+## v0.7.8
+
+### 修复：模型用量上下显示不同步
+
+- 修复同一供应商卡片上，上方用量摘要（含 🔄 刷新按钮）刷新后、下方用量进度条仍停留在旧数据的问题（如上方 24% vs 下方 8%）
+- 用量查询状态提升至卡片层统一持有：点刷新、自动轮询后，两处显示同帧同步更新
+- 自动刷新间隔不再只挂在其中一处显示上，每卡片恰好一份定时器，行为不变
+
+### 实验功能开关同步 kimi-code 最新旗标注册表
+
+- 高级设置的实验功能清单对齐 kimi-code 上游 v2 旗标注册表（7 项）
+- 新增开关：**Tower 模式**（多智能体协同，/tower 命令切换）、**子代理 Fork 上下文**（Agent/AgentSwarm 携带主代理会话快照）、**WaitFor 工具**（回合内等待后台任务）、**自动会话标题**（托管 chat_title 工具）
+- 移除上游已废弃的 **ACP v2 协议** 开关（手写进 config.toml 的配置仍会完整保留）
@@ -0,0 +1,10 @@
+## v0.7.9
+
+### 模型数据同步（models.dev 2026-08-27 快照）
+
+- 模型库刷新至 **7343 个模型 / 203 个供应商**，内置快照与官网静态资源同步更新
+- **新增 GLM-5.3 / GLM-5.3-Flash 全线接入**：智谱官方、OpenRouter、Cloudflare Workers AI、DeepInfra、HuggingFace、火山引擎等约 20 条渠道，GLM-5.3-Flash 定价 $0.075/$0.25（输入/输出，每百万 token）
+- 新增 coding plan 渠道：`zhipuai-coding-plan/glm-5.3`、`glm-5.3-highspeed`、`glm-5.3-flash`
+- 新增 Qwen3.8 系列（27B / Flash / 2.4T-A95B）、火山引擎豆包 Seed 2.x 全系（2-0 Pro/Lite/Mini/Code-Preview、2-1 Pro/Turbo）、DeepSeek-V4 GA 版、MiniMax-M2.7/M3 等共 74 个模型条目
+- 价格更新 58 处：Amazon Bedrock GPT-5.6 Sol 降价（5.5→4）、DeepSeek V4 Flash 0731 半价、Kimi-Latest 微调降价等
+- 能力信息修正 56 处：MiniMax-M2/M3 上下文扩展至 200K/1M、Grok 4.x 补充 structured_output、豆包 Seed 补充图像输入等
@@ -1978,7 +1978,7 @@ dependencies = [

 [[package]]
 name = "kimiswitch"
-version = "0.7.7"
+version = "0.7.12"
 dependencies = [
 "anyhow",
 "chrono",
@@ -1,6 +1,6 @@
 [package]
 name = "kimiswitch"
-version = "0.7.7"
+version = "0.7.12"
 description = "Kimi Switch - model config manager"
 authors = ["codingplan.site"]
 edition = "2021"
@@ -244,6 +244,7 @@ fn build_active_config(config: &Config) -> Config {
        providers,
        models,
        raw_other: config.raw_other.clone(),
+        imported_section_keys: config.imported_section_keys.clone(),
    }
 }

@@ -277,6 +278,9 @@ pub async fn list_provider_models(provider: Provider) -> Result<Vec<DiscoveredMo
        ProviderType::Anthropic => fetch_anthropic_models(&base, &api_key).await,
        ProviderType::GoogleGenai => fetch_google_genai_models(&base, &api_key).await,
        ProviderType::Vertexai => Err("Vertex AI model discovery requires GCP project/location configuration and is not yet supported".to_string()),
+        ProviderType::Unknown(s) => Err(format!(
+            "unsupported provider type for model discovery: {s}"
+        )),
    }
 }

@@ -338,8 +342,12 @@ fn resolve_api_key(provider: &Provider) -> Option<String> {
            return Some(key.clone());
        }
    }
-    let env_key = expected_api_key_key(&provider.provider_type);
-    provider.env.get(env_key).filter(|s| !s.is_empty()).cloned()
+    if let Some(env_key) = expected_api_key_key(&provider.provider_type) {
+        if let Some(key) = provider.env.get(env_key).filter(|s| !s.is_empty()) {
+            return Some(key.clone());
+        }
+    }
+    None
 }

 fn resolve_base_url(provider: &Provider) -> String {
@@ -356,13 +364,16 @@ fn resolve_base_url(provider: &Provider) -> String {
        })
 }

-fn expected_api_key_key(provider_type: &ProviderType) -> &'static str {
+/// Well-known env var fallback for the API key per provider type; `None`
+/// for unknown types (no well-known variable name to look up).
+fn expected_api_key_key(provider_type: &ProviderType) -> Option<&'static str> {
    match provider_type {
-        ProviderType::Kimi => "KIMI_API_KEY",
-        ProviderType::Anthropic => "ANTHROPIC_API_KEY",
-        ProviderType::Openai | ProviderType::OpenaiResponses => "OPENAI_API_KEY",
-        ProviderType::GoogleGenai => "GOOGLE_API_KEY",
-        ProviderType::Vertexai => "VERTEXAI_API_KEY",
+        ProviderType::Kimi => Some("KIMI_API_KEY"),
+        ProviderType::Anthropic => Some("ANTHROPIC_API_KEY"),
+        ProviderType::Openai | ProviderType::OpenaiResponses => Some("OPENAI_API_KEY"),
+        ProviderType::GoogleGenai => Some("GOOGLE_API_KEY"),
+        ProviderType::Vertexai => Some("VERTEXAI_API_KEY"),
+        ProviderType::Unknown(_) => None,
    }
 }

@@ -650,9 +661,8 @@ pub async fn query_provider_usage(
        .clone()
        .filter(|s| !s.trim().is_empty())
        .or_else(|| {
-            provider
-                .env
-                .get(expected_api_key_key(&provider.provider_type))
+            expected_api_key_key(&provider.provider_type)
+                .and_then(|env_key| provider.env.get(env_key))
                .cloned()
                .filter(|s| !s.is_empty())
        });
@@ -1004,6 +1014,9 @@ pub struct KimiWebState {
    window: Mutex<Option<tauri::WebviewWindow>>,
    /// The `kimi web` child spawned by this app (None when reusing a server).
    child: Mutex<Option<std::process::Child>>,
+    /// Serializes `open_kimi_web_embedded` invocations: a rapid double-click
+    /// queues here instead of racing the window/server setup.
+    launch_lock: tokio::sync::Mutex<()>,
 }

 const KIMI_WEB_PORT: u16 = 58627;
@@ -1054,7 +1067,13 @@ pub async fn open_kimi_web_embedded(
    app: tauri::AppHandle,
    state: tauri::State<'_, KimiWebState>,
 ) -> Result<(), String> {
-    // Singleton: focus the existing window instead of creating a second one.
+    // Serialize concurrent invocations (rapid double-click): a second call
+    // waits here until the first has finished building the window, then
+    // re-checks the singleton below and focuses it instead.
+    let _guard = state.launch_lock.lock().await;
+
+    // Singleton (re-checked after acquiring the lock): focus the existing
+    // window instead of creating a second one.
    if let Ok(guard) = state.window.lock() {
        if let Some(w) = guard.as_ref() {
            if w.is_visible().unwrap_or(false) {
@@ -1065,8 +1084,21 @@ pub async fn open_kimi_web_embedded(
        }
    }

-    // Reuse an already-running server, otherwise spawn `kimi web --no-open`.
-    if !kimi_web_alive(Duration::from_millis(800)).await {
+    // Reuse the server this app spawned earlier (still running), or a server
+    // already running outside the app; only spawn when neither is present.
+    // `spawned_by_me` gates every failure-path kill, so a reused server is
+    // never terminated.
+    let mut spawned_by_me = false;
+    let mut child_running = false;
+    if let Ok(mut guard) = state.child.lock() {
+        // try_wait: Ok(None) = still running. Exited or unwaitable children
+        // are treated as gone so a fresh server is spawned below.
+        child_running = guard
+            .as_mut()
+            .map(|child| child.try_wait().map(|st| st.is_none()).unwrap_or(false))
+            .unwrap_or(false);
+    }
+    if !child_running && !kimi_web_alive(Duration::from_millis(800)).await {
        let mut cmd = std::process::Command::new("kimi");
        cmd.args(["web", "--no-open", "--port", &KIMI_WEB_PORT.to_string()]);
        #[cfg(windows)]
@@ -1081,6 +1113,7 @@ pub async fn open_kimi_web_embedded(
                format!("failed to start `kimi web`: {e}")
            }
        })?;
+        spawned_by_me = true;
        if let Ok(mut guard) = state.child.lock() {
            *guard = Some(child);
        }
@@ -1088,57 +1121,79 @@ pub async fn open_kimi_web_embedded(

    // Wait for the server to come up (fresh start takes a moment).
    if !kimi_web_alive(Duration::from_secs(8)).await {
-        kill_spawned_kimi_web(&state);
+        if spawned_by_me {
+            kill_spawned_kimi_web(&state);
+        }
        return Err(format!("kimi web did not come up within 8s ({KIMI_WEB_ORIGIN})"));
    }

-    // Create the window on the main thread (required on macOS).
+    // Build the URL, then create the window on the main thread (required on
+    // macOS). The build result is awaited while still holding `launch_lock`,
+    // so a concurrent second invocation blocks here and then lands on the
+    // singleton re-check above once the window exists.
    let url: tauri::Url = match kimi_web_url().parse() {
        Ok(u) => u,
        Err(_) => return Err("invalid kimi web url".to_string()),
    };
-    let app = app.clone();
-    tauri::async_runtime::spawn(async move {
-        let builder_app = app.clone();
-        let _ = app.run_on_main_thread(move || {
-            match tauri::WebviewWindowBuilder::new(
-                &builder_app,
-                KIMI_WEB_WINDOW_LABEL,
-                tauri::WebviewUrl::External(url),
-            )
-            .title("Kimi Code WebUI")
-            .inner_size(1100.0, 750.0)
-            .resizable(true)
-            .center()
-            .build()
-            {
-                Ok(window) => {
-                    // Track the window for the singleton check.
-                    if let Ok(mut guard) = builder_app.state::<KimiWebState>().window.lock() {
-                        *guard = Some(window.clone());
-                    }
-                    // Clean up when the window closes: forget it and stop the
-                    // `kimi web` server we spawned (reused servers are untouched).
-                    let w = window.clone();
-                    window.on_window_event(move |event| match event {
-                        tauri::WindowEvent::Destroyed => {
-                            let app = w.app_handle().clone();
-                            let state = app.state::<KimiWebState>();
-                            if let Ok(mut guard) = state.window.lock() {
-                                *guard = None;
-                            }
-                            kill_spawned_kimi_web(&state);
-                        }
-                        _ => {}
-                    });
-                }
-                Err(_) => {
-                    // Window creation failed (e.g. rapid double-click racing
-                    // the singleton check): don't leak the server we spawned.
-                    kill_spawned_kimi_web(&builder_app.state::<KimiWebState>());
+    let (tx, rx) = tokio::sync::oneshot::channel::<Result<(), String>>();
+    let builder_app = app.clone();
+    app.run_on_main_thread(move || {
+        let result = build_kimi_web_window(&builder_app, &url);
+        let _ = tx.send(result);
+    })
+    .map_err(|e| format!("failed to queue window creation: {e}"))?;
+    match rx.await {
+        Ok(result) => {
+            if let Err(e) = result {
+                // Window creation failed (e.g. label already taken): only stop
+                // the server when this invocation spawned it.
+                if spawned_by_me {
+                    kill_spawned_kimi_web(&state);
                }
+                return Err(e);
            }
-        });
+            Ok(())
+        }
+        // Sender dropped without a result (app shutting down, build never
+        // ran): never kill the server — a later invocation may reuse it.
+        Err(_) => Ok(()),
+    }
+}
+
+/// Build the embedded WebUI window on the main thread and track it for the
+/// singleton check. On success, wires close/destroy events to forget the
+/// window and stop the `kimi web` server this app spawned (reused servers are
+/// untouched).
+fn build_kimi_web_window(app: &tauri::AppHandle, url: &tauri::Url) -> Result<(), String> {
+    let window = tauri::WebviewWindowBuilder::new(
+        app,
+        KIMI_WEB_WINDOW_LABEL,
+        tauri::WebviewUrl::External(url.clone()),
+    )
+    .title("Kimi Code WebUI")
+    .inner_size(1100.0, 750.0)
+    .resizable(true)
+    .center()
+    .build()
+    .map_err(|e| format!("failed to create kimi web window: {e}"))?;
+
+    // Track the window for the singleton check.
+    if let Ok(mut guard) = app.state::<KimiWebState>().window.lock() {
+        *guard = Some(window.clone());
+    }
+    // Clean up when the window closes: forget it and stop the `kimi web`
+    // server we spawned (reused servers are untouched).
+    let w = window.clone();
+    window.on_window_event(move |event| match event {
+        tauri::WindowEvent::Destroyed => {
+            let app = w.app_handle().clone();
+            let state = app.state::<KimiWebState>();
+            if let Ok(mut guard) = state.window.lock() {
+                *guard = None;
+            }
+            kill_spawned_kimi_web(&state);
+        }
+        _ => {}
    });
    Ok(())
 }
@@ -1283,24 +1338,43 @@ pub async fn kimi_oauth_poll(
 // Experimental feature env-var probe (Kimi Code secondary model etc.)
 // ---------------------------------------------------------------------------

-/// Read the Kimi Code experimental-feature environment variables that are
-/// currently set (non-empty) in this process's environment. The frontend
-/// uses the result to mark config.toml `[experimental]` toggles as locked:
-/// env vars outrank the config file in Kimi Code's flag resolution.
+/// Read the Kimi Code environment variables that are currently set (non-empty)
+/// in this process's environment. Every `[experimental]` feature env var is
+/// probed — the frontend uses a per-feature env var to mark the matching
+/// config.toml `[experimental]` toggle as locked, since a single-feature env
+/// var outranks the config file in Kimi Code's flag resolution. The master
+/// `KIMI_CODE_EXPERIMENTAL_FLAG` (0.40.1+: only force-ONs flags without an
+/// explicit config entry; never locks a toggle) and `KIMI_CODE_LEGACY_FLAG`
+/// are probed too — the latter lets AgentSettingsPanel decide whether to
+/// dual-write the v1 engine's loop_control keys.
+///
+/// The flag id → env var mapping mirrors `EXPERIMENTAL_FLAGS` in
+/// src/lib/subagent-settings.ts; keep both lists in sync.
 ///
 /// Truthy values per the CLI: `1` / `true` / `yes` / `on` (case-insensitive);
 /// the raw values are returned and the truthy check happens on the frontend.
 #[tauri::command]
 pub fn get_experimental_env_status() -> HashMap<String, String> {
-    const VARS: [&str; 5] = [
-        "KIMI_CODE_EXPERIMENTAL_FLAG",
-        "KIMI_CODE_EXPERIMENTAL_SECONDARY_MODEL",
-        "KIMI_CODE_EXPERIMENTAL_TOOL_SELECT",
-        "KIMI_CODE_EXPERIMENTAL_ACP_V2",
-        "KIMI_CODE_EXPERIMENTAL_PERSISTENCE_MINIDB_READMODEL",
+    const VARS: [(&str, &str); 12] = [
+        ("secondary-model", "KIMI_CODE_EXPERIMENTAL_SECONDARY_MODEL"),
+        ("tool-select", "KIMI_CODE_EXPERIMENTAL_TOOL_SELECT"),
+        (
+            "persistence_minidb_readmodel",
+            "KIMI_CODE_EXPERIMENTAL_PERSISTENCE_MINIDB_READMODEL",
+        ),
+        ("tower", "KIMI_CODE_EXPERIMENTAL_TOWER"),
+        ("subagent_fork", "KIMI_CODE_EXPERIMENTAL_SUBAGENT_FORK"),
+        ("wait_for", "KIMI_CODE_EXPERIMENTAL_WAIT_FOR"),
+        ("auto_session_title", "KIMI_CODE_EXPERIMENTAL_AUTO_SESSION_TITLE"),
+        ("remote-control", "KIMI_CODE_EXPERIMENTAL_REMOTE_CONTROL"),
+        ("search_worker", "KIMI_CODE_EXPERIMENTAL_SEARCH_WORKER"),
+        ("file_history", "KIMI_CODE_EXPERIMENTAL_FILE_HISTORY"),
+        // Non-flag probes consumed by the frontend:
+        ("master", "KIMI_CODE_EXPERIMENTAL_FLAG"),
+        ("legacy", "KIMI_CODE_LEGACY_FLAG"),
    ];
    let mut out = HashMap::new();
-    for name in VARS {
+    for (_, name) in VARS {
        if let Ok(value) = std::env::var(name) {
            if !value.trim().is_empty() {
                out.insert(name.to_string(), value);
@@ -1234,6 +1234,49 @@ fn read_state_safe(session_dir: &Path) -> (Option<String>, Option<String>, Optio
    (None, None, None, None)
 }

+/// Whether a session is archived per its `state.json` metadata: the
+/// kimi-code v2 engine writes `archived: true` (plus an `archivedAt` epoch-ms
+/// timestamp) when archiving and clears both on restore. Unparseable or
+/// missing state.json counts as not archived (and won't panic).
+fn read_state_archived(session_dir: &Path) -> bool {
+    let sp = session_dir.join("state.json");
+    let Ok(content) = fs::read_to_string(&sp) else { return false };
+    let Ok(obj) = serde_json::from_str::<serde_json::Value>(&content) else { return false };
+    obj.get("archived").and_then(|v| v.as_bool()).unwrap_or(false)
+}
+
+/// Set or clear the archived metadata on a session's `state.json`, preserving
+/// every other field. Mirrors upstream kimi-code v2 `SessionMetadata.setArchived`:
+/// archive writes `{ archived: true, archivedAt: Date.now() }` (epoch ms, same
+/// `Date.now()` format as the CLI), restore writes `{ archived: false }` and
+/// drops the `archivedAt` key; `updatedAt` is untouched in both cases.
+/// The write is atomic (tmp file + rename), same pattern as the plugin store.
+/// Missing or malformed state.json returns a descriptive error, never a panic.
+fn set_session_archived_meta(session_dir: &Path, archived: bool) -> Result<(), String> {
+    let sp = session_dir.join("state.json");
+    let content = fs::read_to_string(&sp).map_err(|e| format!("read state.json: {e}"))?;
+    let mut obj: serde_json::Value =
+        serde_json::from_str(&content).map_err(|e| format!("parse state.json: {e}"))?;
+    let obj = obj
+        .as_object_mut()
+        .ok_or_else(|| "state.json is not a JSON object".to_string())?;
+    obj.insert("archived".into(), serde_json::Value::Bool(archived));
+    if archived {
+        let now_ms = std::time::SystemTime::now()
+            .duration_since(std::time::UNIX_EPOCH)
+            .unwrap_or_default()
+            .as_millis() as u64;
+        obj.insert("archivedAt".into(), serde_json::Value::Number(now_ms.into()));
+    } else {
+        obj.remove("archivedAt");
+    }
+    let out = serde_json::to_string(&obj).map_err(|e| format!("serialize state.json: {e}"))?;
+    let tmp = sp.with_file_name("state.json.tmp");
+    fs::write(&tmp, out).map_err(|e| format!("write state.json: {e}"))?;
+    fs::rename(&tmp, &sp).map_err(|e| format!("commit state.json: {e}"))?;
+    Ok(())
+}
+
 fn humanize_workspace(id: &str) -> String {
    let re = Regex::new(r"^wd_(.+)_[0-9a-fA-F]{8,}$").unwrap();
    if let Some(caps) = re.captures(id) {
@@ -1243,7 +1286,11 @@ fn humanize_workspace(id: &str) -> String {
    }
 }

-fn list_sessions_in_dir(dir: &Path, workspace_id: &str, status: &str) -> Vec<SessionRow> {
+/// List session rows under one directory. `in_archive_dir` marks `.kcd-archive`
+/// (legacy physical archive); a session is additionally archived when its own
+/// `state.json` carries `archived: true` — so CLI-v2 metadata-archived sessions
+/// that still live at the active root are correctly classified.
+fn list_sessions_in_dir(dir: &Path, workspace_id: &str, in_archive_dir: bool) -> Vec<SessionRow> {
    let mut sessions = Vec::new();
    if !dir.exists() { return sessions; }
    let re = session_re();
@@ -1251,6 +1298,8 @@ fn list_sessions_in_dir(dir: &Path, workspace_id: &str, status: &str) -> Vec<Ses
        if !entry.file_type().map(|t| t.is_dir()).unwrap_or(false) { continue; }
        let name = entry.file_name().to_string_lossy().to_string();
        if !re.is_match(&name) { continue; }
+        let archived = in_archive_dir || read_state_archived(&entry.path());
+        let status = if archived { "archived" } else { "active" };
        let (title, work_dir, created_at, updated_at) = read_state_safe(&entry.path());
        let (bytes, files) = file_size_approx(&entry.path());
        let mtime = entry.path().metadata().ok().and_then(|m| m.modified().ok())
@@ -1311,19 +1360,33 @@ fn list_sessions_cmd(home: &Path, status: &str, workspace_filter: Option<String>

        let active_dir = root.join(wid);
        let arch_dir = archive_root.join(wid);
-        let active_all = list_sessions_in_dir(&active_dir, wid, "active");
-        let arch_all = list_sessions_in_dir(&arch_dir, wid, "archived");
-        let active_list = if status == "archived" { vec![] } else { active_all.clone() };
-        let arch_list = if status == "active" { vec![] } else { arch_all.clone() };
+        // Merge both sources, deduping by session id (ids are UUIDs and never
+        // reused; a stale legacy copy under .kcd-archive loses to the entry at
+        // the active root). Status comes from each row, so metadata-archived
+        // sessions surface in the archived filter and the archive-only counts.
+        let mut by_id: HashMap<String, SessionRow> = HashMap::new();
+        for row in list_sessions_in_dir(&active_dir, wid, false) {
+            by_id.insert(row.id.clone(), row);
+        }
+        for row in list_sessions_in_dir(&arch_dir, wid, true) {
+            by_id.entry(row.id.clone()).or_insert(row);
+        }
+        let all: Vec<SessionRow> = by_id.into_values().collect();
+        let active_count = all.iter().filter(|s| s.status == "active").count();
+        let archived_count = all.iter().filter(|s| s.status == "archived").count();
+        let listed: Vec<SessionRow> = match status {
+            "active" => all.iter().filter(|s| s.status == "active").cloned().collect(),
+            "archived" => all.iter().filter(|s| s.status == "archived").cloned().collect(),
+            _ => all,
+        };

-        let empty = active_all.is_empty() && arch_all.is_empty();
+        let empty = active_count == 0 && archived_count == 0;
        workspaces.push(WorkspaceRow {
            id: wid.clone(), name, root: root_path,
            created_at: None, last_opened_at: None,
-            active_count: active_all.len(), archived_count: arch_all.len(), empty,
+            active_count, archived_count, empty,
        });
-        all_sessions.extend(active_list);
-        all_sessions.extend(arch_list);
+        all_sessions.extend(listed);
    }

    all_sessions.sort_by(|a, b| {
@@ -1356,38 +1419,57 @@ fn assert_safe_path(home: &Path, workspace_id: &str, session_id: &str) -> Result
 }

 fn archive_session_cmd(home: &Path, workspace_id: &str, session_id: &str) -> Result<ActionResponse, String> {
-    let src = assert_safe_path(home, workspace_id, session_id)?;
-    let dest = sessions_root(home).join(".kcd-archive").join(workspace_id).join(session_id);
-    if !src.exists() { return Err("session not found".into()); }
-    if let Some(parent) = dest.parent() {
-        fs::create_dir_all(parent).map_err(|e| format!("mkdir: {}", e))?;
+    let dir = assert_safe_path(home, workspace_id, session_id)?;
+    if !dir.exists() { return Err("session not found".into()); }
+    // Metadata archive, matching upstream kimi-code v2: write `archived: true`
+    // + `archivedAt` into state.json, leave the session directory in place, and
+    // do NOT touch session_index.jsonl — archived state is derived from
+    // state.json, and the CLI reconciles its own index mirror. A session
+    // without a readable state.json gets a clear error rather than a panic.
+    let sp = dir.join("state.json");
+    if !sp.is_file() {
+        return Err(format!("session {session_id} has no state.json; cannot archive"));
    }
-    fs::rename(&src, &dest).or_else(|_| {
-        // cross-device fallback
-        fs_extra::dir::copy(&src, dest.parent().unwrap(), &Default::default()).ok();
-        fs::remove_dir_all(&src).ok();
-        Ok::<(), String>(())
-    }).map_err(|e: String| format!("move: {}", e))?;
-    scrub_session_index(home, session_id);
+    set_session_archived_meta(&dir, true)?;
    Ok(ActionResponse {
        ok: true, workspace_id: workspace_id.to_string(),
        session_id: session_id.to_string(), status: Some("archived".into()),
-        path: Some(dest.to_string_lossy().to_string()), deleted: None,
+        path: Some(dir.to_string_lossy().to_string()), deleted: None,
    })
 }

 fn unarchive_session_cmd(home: &Path, workspace_id: &str, session_id: &str) -> Result<ActionResponse, String> {
-    let src = sessions_root(home).join(".kcd-archive").join(workspace_id).join(session_id);
    let dest = assert_safe_path(home, workspace_id, session_id)?;
-    if !src.exists() { return Err("archived session not found".into()); }
-    if let Some(parent) = dest.parent() {
-        fs::create_dir_all(parent).map_err(|e| format!("mkdir: {}", e))?;
+    // 1) Legacy recovery: sessions physically moved to .kcd-archive by older
+    //    KimiSwitch versions are moved back, then their metadata flag is
+    //    cleared per upstream restore semantics (archived:false + archivedAt
+    //    dropped). A missing state.json is fine here — there is no flag.
+    let archived_dir = sessions_root(home).join(".kcd-archive").join(workspace_id).join(session_id);
+    if archived_dir.exists() {
+        if let Some(parent) = dest.parent() {
+            fs::create_dir_all(parent).map_err(|e| format!("mkdir: {}", e))?;
+        }
+        fs::rename(&archived_dir, &dest).or_else(|_| {
+            // cross-device fallback
+            fs_extra::dir::copy(&archived_dir, dest.parent().unwrap(), &Default::default()).ok();
+            fs::remove_dir_all(&archived_dir).ok();
+            Ok::<(), String>(())
+        }).map_err(|e: String| format!("move: {}", e))?;
+        // The move already restored the session; a corrupt state.json must not
+        // turn a successful restore into an error — the session simply lists
+        // as active with its flag intact.
+        if dest.join("state.json").is_file() {
+            let _ = set_session_archived_meta(&dest, false);
+        }
+        return Ok(ActionResponse {
+            ok: true, workspace_id: workspace_id.to_string(),
+            session_id: session_id.to_string(), status: Some("active".into()),
+            path: Some(dest.to_string_lossy().to_string()), deleted: None,
+        });
    }
-    fs::rename(&src, &dest).or_else(|_| {
-        fs_extra::dir::copy(&src, dest.parent().unwrap(), &Default::default()).ok();
-        fs::remove_dir_all(&src).ok();
-        Ok::<(), String>(())
-    }).map_err(|e: String| format!("move: {}", e))?;
+    // 2) Metadata-archived session: still at the active root, just clear the flag.
+    if !dest.exists() { return Err("archived session not found".into()); }
+    set_session_archived_meta(&dest, false)?;
    Ok(ActionResponse {
        ok: true, workspace_id: workspace_id.to_string(),
        session_id: session_id.to_string(), status: Some("active".into()),
@@ -1398,6 +1480,9 @@ fn unarchive_session_cmd(home: &Path, workspace_id: &str, session_id: &str) -> R
 fn delete_session_cmd(home: &Path, workspace_id: &str, session_id: &str, status_hint: Option<&str>) -> Result<ActionResponse, String> {
    let active = assert_safe_path(home, workspace_id, session_id)?;
    let archived = sessions_root(home).join(".kcd-archive").join(workspace_id).join(session_id);
+    // Metadata-archived sessions still live at the active root, so a
+    // status_hint of "archived" with no legacy .kcd-archive copy falls through
+    // to the default arm and removes the in-place directory.
    let target = match status_hint {
        Some("archived") if archived.exists() => archived,
        Some("active") if active.exists() => active,
@@ -1440,11 +1525,11 @@ fn delete_workspace_cmd(home: &Path, workspace_id: &str, _confirm: bool, _force:
    let arch_dir = root.join(".kcd-archive").join(workspace_id);

    if active_dir.exists() {
-        let active_list = list_sessions_in_dir(&active_dir, workspace_id, "active");
+        let active_list = list_sessions_in_dir(&active_dir, workspace_id, false);
        if !active_list.is_empty() { return Err("workspace is not empty; archive/delete sessions first".into()); }
    }
    if arch_dir.exists() {
-        let arch_list = list_sessions_in_dir(&arch_dir, workspace_id, "archived");
+        let arch_list = list_sessions_in_dir(&arch_dir, workspace_id, true);
        if !arch_list.is_empty() { return Err("workspace is not empty; archive/delete sessions first".into()); }
    }

@@ -1546,6 +1631,10 @@ fn get_session_preview_cmd(home: &Path, workspace_id: &str, session_id: &str, st
    let archived = root.join(".kcd-archive").join(workspace_id).join(session_id);
    let session_dir = match status_hint {
        Some("archived") if archived.exists() => archived,
+        // Metadata-archived sessions stay at the active root (kimi-code v2
+        // writes state.json.archived, no physical move) — read them in place.
+        Some("archived") if active.exists() => active,
+        Some("active") if active.exists() => active,
        _ => if active.exists() { active } else if archived.exists() { archived }
            else { return Err("session not found".into()); }
    };
@@ -1666,7 +1755,11 @@ fn get_session_preview_cmd(home: &Path, workspace_id: &str, session_id: &str, st
    Ok(PreviewResult {
        workspace_id: workspace_id.to_string(),
        session_id: session_id.to_string(),
-        status: if session_dir.to_string_lossy().contains(".kcd-archive") { "archived".into() } else { "active".into() },
+        status: if session_dir.to_string_lossy().contains(".kcd-archive") || read_state_archived(&session_dir) {
+            "archived".into()
+        } else {
+            "active".into()
+        },
        title, work_dir, created_at, updated_at,
        message_count: messages.len(),
        truncated,
@@ -1998,3 +2091,219 @@ mod pricing_tests {
        assert_eq!(ch, 0.11);
    }
 }
+
+#[cfg(test)]
+mod session_tests {
+    use super::*;
+
+    const WID: &str = "wd_proj_a1b2c3d4e5f6";
+    const SID: &str = "session_11111111-2222-3333-4444-555555555555";
+
+    fn setup_home() -> (tempfile::TempDir, PathBuf) {
+        let td = tempfile::tempdir().unwrap();
+        let home = td.path().join("home");
+        fs::create_dir_all(home.join("sessions").join(WID)).unwrap();
+        (td, home)
+    }
+
+    fn active_dir(home: &Path) -> PathBuf {
+        home.join("sessions").join(WID)
+    }
+
+    fn legacy_arch_dir(home: &Path) -> PathBuf {
+        home.join("sessions").join(".kcd-archive").join(WID)
+    }
+
+    fn write_state(session_dir: &Path, archived: bool, archived_at: Option<u64>) {
+        let mut obj = serde_json::json!({
+            "id": SID,
+            "version": 2,
+            "cwd": "/tmp/work",
+            "createdAt": 1_700_000_000_000u64,
+            "updatedAt": 1_700_000_000_000u64,
+            "archived": archived,
+            "title": "test session",
+        });
+        let o = obj.as_object_mut().unwrap();
+        match archived_at {
+            Some(at) => { o.insert("archivedAt".into(), serde_json::json!(at)); }
+            None => { o.remove("archivedAt"); }
+        }
+        fs::write(session_dir.join("state.json"), serde_json::to_string(&obj).unwrap()).unwrap();
+    }
+
+    fn read_state_json(session_dir: &Path) -> serde_json::Value {
+        serde_json::from_str(&fs::read_to_string(session_dir.join("state.json")).unwrap()).unwrap()
+    }
+
+    #[test]
+    fn metadata_archive_roundtrip() {
+        let (_td, home) = setup_home();
+        let dir = active_dir(&home).join(SID);
+        fs::create_dir_all(&dir).unwrap();
+        write_state(&dir, false, None);
+
+        // Fresh session lists as active.
+        let res = list_sessions_cmd(&home, "all", None);
+        assert_eq!(res.sessions.len(), 1);
+        assert_eq!(res.sessions[0].status, "active");
+        assert_eq!(res.workspaces[0].active_count, 1);
+        assert_eq!(res.workspaces[0].archived_count, 0);
+
+        // Archive is metadata-only: dir stays put, flag + epoch-ms archivedAt written.
+        let ar = archive_session_cmd(&home, WID, SID).unwrap();
+        assert_eq!(ar.status.as_deref(), Some("archived"));
+        assert!(dir.exists(), "metadata archive must not move the session dir");
+        assert!(!legacy_arch_dir(&home).join(SID).exists(), "nothing moves into .kcd-archive");
+        let st = read_state_json(&dir);
+        assert_eq!(st["archived"], true);
+        let archived_at = st["archivedAt"].as_u64().expect("archivedAt as epoch ms");
+        assert!(archived_at > 0, "archivedAt is a Date.now()-style epoch-ms number");
+
+        // List reflects the new status.
+        let res = list_sessions_cmd(&home, "all", None);
+        assert_eq!(res.sessions[0].status, "archived");
+        assert_eq!(res.workspaces[0].active_count, 0);
+        assert_eq!(res.workspaces[0].archived_count, 1);
+        assert!(list_sessions_cmd(&home, "active", None).sessions.is_empty());
+        assert_eq!(list_sessions_cmd(&home, "archived", None).sessions.len(), 1);
+
+        // Unarchive clears the flag in place (archived:false, archivedAt dropped).
+        let ur = unarchive_session_cmd(&home, WID, SID).unwrap();
+        assert_eq!(ur.status.as_deref(), Some("active"));
+        assert!(dir.exists(), "unarchive must not move the dir either");
+        let st = read_state_json(&dir);
+        assert_eq!(st["archived"], false);
+        assert!(st.get("archivedAt").is_none(), "archivedAt key removed on restore");
+
+        let res = list_sessions_cmd(&home, "all", None);
+        assert_eq!(res.sessions[0].status, "active");
+        assert_eq!(res.workspaces[0].active_count, 1);
+        assert_eq!(res.workspaces[0].archived_count, 0);
+    }
+
+    #[test]
+    fn metadata_archive_preserves_other_state_fields() {
+        let (_td, home) = setup_home();
+        let dir = active_dir(&home).join(SID);
+        fs::create_dir_all(&dir).unwrap();
+        let mut obj = serde_json::json!({
+            "id": SID,
+            "version": 2,
+            "cwd": "/repo",
+            "createdAt": 1,
+            "updatedAt": 2,
+            "archived": false,
+            "title": "keep me",
+            "custom": { "goal": "x" },
+            "agents": { "main": { "type": "main" } },
+        });
+        obj.as_object_mut().unwrap().insert("someFutureField".into(), serde_json::json!([1, 2, 3]));
+        fs::write(dir.join("state.json"), serde_json::to_string(&obj).unwrap()).unwrap();
+
+        archive_session_cmd(&home, WID, SID).unwrap();
+        unarchive_session_cmd(&home, WID, SID).unwrap();
+
+        let st = read_state_json(&dir);
+        assert_eq!(st["title"], "keep me");
+        assert_eq!(st["custom"]["goal"], "x");
+        assert_eq!(st["agents"]["main"]["type"], "main");
+        assert_eq!(st["someFutureField"][0], 1);
+        assert_eq!(st["updatedAt"], 2, "updatedAt untouched, like upstream touchUpdatedAt:false");
+    }
+
+    #[test]
+    fn cli_archived_session_in_active_root_lists_as_archived() {
+        // A session already carrying archived:true in state.json while still
+        // living at the active root (written by kimi-code CLI v2 itself) must
+        // surface in the archived filter / counts.
+        let (_td, home) = setup_home();
+        let dir = active_dir(&home).join(SID);
+        fs::create_dir_all(&dir).unwrap();
+        write_state(&dir, true, Some(1_700_000_000_000u64));
+        fs::write(dir.join("wire.jsonl"), "").unwrap();
+
+        let res = list_sessions_cmd(&home, "all", None);
+        assert_eq!(res.sessions.len(), 1);
+        assert_eq!(res.sessions[0].status, "archived");
+        assert_eq!(res.workspaces[0].active_count, 0);
+        assert_eq!(res.workspaces[0].archived_count, 1);
+        assert!(list_sessions_cmd(&home, "active", None).sessions.is_empty());
+        assert_eq!(list_sessions_cmd(&home, "archived", None).sessions.len(), 1);
+
+        // Preview resolves it in place with status "archived".
+        let pv = get_session_preview_cmd(&home, WID, SID, Some("archived")).unwrap();
+        assert_eq!(pv.status, "archived");
+    }
+
+    #[test]
+    fn legacy_kcd_archive_dir_still_listed_and_restored() {
+        let (_td, home) = setup_home();
+        let arch_dir = legacy_arch_dir(&home).join(SID);
+        fs::create_dir_all(&arch_dir).unwrap();
+        write_state(&arch_dir, false, None);
+
+        // Legacy physical archive still shows as archived.
+        let res = list_sessions_cmd(&home, "all", None);
+        assert_eq!(res.sessions.len(), 1);
+        assert_eq!(res.sessions[0].status, "archived");
+        assert_eq!(res.workspaces[0].active_count, 0);
+        assert_eq!(res.workspaces[0].archived_count, 1);
+        assert!(list_sessions_cmd(&home, "active", None).sessions.is_empty());
+
+        // Unarchive moves it back and clears the flag.
+        let ur = unarchive_session_cmd(&home, WID, SID).unwrap();
+        assert_eq!(ur.status.as_deref(), Some("active"));
+        assert!(!arch_dir.exists(), "legacy .kcd-archive copy moved back");
+        let restored = active_dir(&home).join(SID);
+        assert!(restored.exists());
+        let st = read_state_json(&restored);
+        assert_eq!(st["archived"], false);
+        assert!(st.get("archivedAt").is_none());
+
+        let res = list_sessions_cmd(&home, "all", None);
+        assert_eq!(res.sessions[0].status, "active");
+        assert_eq!(res.workspaces[0].active_count, 1);
+        assert_eq!(res.workspaces[0].archived_count, 0);
+    }
+
+    #[test]
+    fn corrupt_state_json_archive_errors_without_panic() {
+        let (_td, home) = setup_home();
+        let dir = active_dir(&home).join(SID);
+        fs::create_dir_all(&dir).unwrap();
+        fs::write(dir.join("state.json"), "{ this is not json").unwrap();
+
+        let err = archive_session_cmd(&home, WID, SID).unwrap_err();
+        assert!(err.contains("state.json"), "clear error mentioning state.json, got: {err}");
+        assert!(dir.exists(), "failed archive leaves the session dir in place");
+        assert_eq!(fs::read_to_string(dir.join("state.json")).unwrap(), "{ this is not json");
+    }
+
+    #[test]
+    fn missing_state_json_archive_errors_without_panic() {
+        let (_td, home) = setup_home();
+        let dir = active_dir(&home).join(SID);
+        fs::create_dir_all(&dir).unwrap();
+
+        let err = archive_session_cmd(&home, WID, SID).unwrap_err();
+        assert!(err.contains("no state.json"), "clear error, got: {err}");
+        assert!(dir.exists());
+    }
+
+    #[test]
+    fn delete_session_covers_metadata_archived() {
+        // Status hint "archived" without a legacy .kcd-archive copy falls back
+        // to removing the metadata-archived in-place directory.
+        let (_td, home) = setup_home();
+        let dir = active_dir(&home).join(SID);
+        fs::create_dir_all(&dir).unwrap();
+        write_state(&dir, true, Some(1_700_000_000_000u64));
+
+        let del = delete_session_cmd(&home, WID, SID, Some("archived")).unwrap();
+        assert_eq!(del.deleted, Some(true));
+        assert!(!dir.exists(), "metadata-archived session deleted in place");
+        assert!(!legacy_arch_dir(&home).join(SID).exists());
+        assert!(list_sessions_cmd(&home, "all", None).sessions.is_empty());
+    }
+}
@@ -207,6 +207,9 @@ pub fn load_config(agent: &Agent) -> DbResult<Config> {
        providers,
        models,
        raw_other: Value::Null,
+        // The SQLite snapshot stores no import baseline; an empty baseline
+        // makes export's stale-key cleanup a no-op (safe degradation).
+        imported_section_keys: Vec::new(),
    })
 }

@@ -328,12 +331,5 @@ pub fn delete_setting_pub(key: &str) -> DbResult<()> {
 }

 fn provider_type_for_str(s: &str) -> ProviderType {
-    match s {
-        "anthropic" => ProviderType::Anthropic,
-        "openai" => ProviderType::Openai,
-        "openai_responses" => ProviderType::OpenaiResponses,
-        "google-genai" => ProviderType::GoogleGenai,
-        "vertexai" => ProviderType::Vertexai,
-        _ => ProviderType::Kimi,
-    }
+    ProviderType::from_kimi_type(s)
 }
@@ -148,7 +148,7 @@ pub fn kimi_code_to_config(value: &TomlValue) -> Config {
            let provider_type = table
                .get("type")
                .and_then(|v| v.as_str())
-                .map(provider_type_for_kimi_type)
+                .map(ProviderType::from_kimi_type)
                .unwrap_or(ProviderType::Kimi);

            let base_url = table.get("base_url").and_then(|v| v.as_str()).map(|s| s.to_string());
@@ -216,19 +216,23 @@ pub fn kimi_code_to_config(value: &TomlValue) -> Config {
        }
    }

-    let raw_other = {
-        let mut rest = root;
-        rest.remove("default_model");
-        rest.remove("providers");
-        rest.remove("models");
-        toml_value_to_json(&TomlValue::Table(rest))
-    };
+    let mut rest = root;
+    rest.remove("default_model");
+    rest.remove("providers");
+    rest.remove("models");
+    // Baseline of top-level keys seen at import time (table order). Export
+    // only drops baseline keys that later disappeared from raw_other, i.e.
+    // sections the user removed in the UI — sections the CLI adds after
+    // import are not in the baseline and survive the round-trip.
+    let imported_section_keys: Vec<String> = rest.keys().cloned().collect();
+    let raw_other = toml_value_to_json(&TomlValue::Table(rest));

    Config {
        default_model,
        providers,
        models,
        raw_other,
+        imported_section_keys,
    }
 }

@@ -354,10 +358,13 @@ pub fn config_to_kimi_code(config: &Config, existing: Option<&TomlValue>) -> Tom
    // Sync UI-managed top-level sections (thinking, experimental,
    // secondary_model, ...) from config.raw_other back into the root.
    // At import time every top-level key except default_model/providers/
-    // models was collected into raw_other, so raw_other is the authoritative
-    // snapshot of those sections. This both persists UI edits to those
-    // sections and drops sections the user explicitly removed (absent from
-    // raw_other but still present in `existing`).
+    // models was collected into raw_other, and their keys recorded in
+    // `imported_section_keys` (the import-time baseline). Keys present in
+    // raw_other are written back (persisting UI edits); keys that were in
+    // the baseline but are absent from raw_other were removed by the user
+    // and are dropped from `existing`. Top-level keys NOT in the baseline
+    // were added by the CLI after import and are preserved untouched. An
+    // empty baseline (Config not built via import) deletes nothing.
    let mut managed_keys = std::collections::HashSet::new();
    if let TomlValue::Table(extra) =
        json_to_toml(&config.raw_other).unwrap_or(TomlValue::Table(Table::new()))
@@ -377,6 +384,7 @@ pub fn config_to_kimi_code(config: &Config, existing: Option<&TomlValue>) -> Tom
                && *k != "providers"
                && *k != "models"
                && !managed_keys.contains(*k)
+                && config.imported_section_keys.iter().any(|b| b == *k)
        })
        .cloned()
        .collect();
@@ -387,17 +395,6 @@ pub fn config_to_kimi_code(config: &Config, existing: Option<&TomlValue>) -> Tom
    TomlValue::Table(root)
 }

-fn provider_type_for_kimi_type(typ: &str) -> ProviderType {
-    match typ {
-        "anthropic" => ProviderType::Anthropic,
-        "openai" => ProviderType::Openai,
-        "openai_responses" => ProviderType::OpenaiResponses,
-        "google-genai" => ProviderType::GoogleGenai,
-        "vertexai" => ProviderType::Vertexai,
-        _ => ProviderType::Kimi,
-    }
-}
-
 fn toml_value_to_json(value: &TomlValue) -> Value {
    match value {
        TomlValue::String(s) => Value::String(s.clone()),
@@ -572,6 +569,7 @@ api_key = ""
            providers,
            models,
            raw_other: Value::Null,
+            imported_section_keys: Vec::new(),
        };

        let exported = config_to_kimi_code(&config, None);
@@ -705,6 +703,7 @@ default_effort = "low"
            providers,
            models,
            raw_other: Value::Null,
+            imported_section_keys: Vec::new(),
        };

        let exported = config_to_kimi_code(&config, None);
@@ -818,6 +817,7 @@ max_context_size = 1048576
            providers,
            models: IndexMap::new(),
            raw_other: Value::Null,
+            imported_section_keys: Vec::new(),
        };

        let exported = config_to_kimi_code(&config, None);
@@ -858,6 +858,7 @@ max_context_size = 1048576
            providers,
            models: IndexMap::new(),
            raw_other: Value::Null,
+            imported_section_keys: Vec::new(),
        };

        let exported = config_to_kimi_code(&config, None);
@@ -908,6 +909,7 @@ max_context_size = 1048576
            )]),
            models: IndexMap::new(),
            raw_other: Value::Null,
+            imported_section_keys: Vec::new(),
        };
        let exported = config_to_kimi_code(&config, None);
        let provider = exported
@@ -927,6 +929,7 @@ max_context_size = 1048576
            )]),
            models: IndexMap::new(),
            raw_other: Value::Null,
+            imported_section_keys: Vec::new(),
        };
        let exported = config_to_kimi_code(&config, None);
        let provider = exported
@@ -940,4 +943,125 @@ max_context_size = 1048576
            Some("sk-user-key")
        );
    }
+
+    #[test]
+    fn kimi_code_unknown_provider_type_roundtrip() {
+        // A provider `type` string the CLI knows but Kimi Switch does not
+        // must survive import → export verbatim, not be rewritten to "kimi".
+        let toml_str = r#"
+default_model = "p1/custom-model"
+
+[providers.p1]
+type = "custom-xyz"
+api_key = "sk-x"
+
+[models."p1/custom-model"]
+provider = "p1"
+model = "custom-model"
+max_context_size = 128000
+"#;
+        let value: TomlValue = toml_str.parse().unwrap();
+        let config = kimi_code_to_config(&value);
+
+        let provider = config.providers.get("p1").unwrap();
+        assert_eq!(
+            provider.provider_type,
+            ProviderType::Unknown("custom-xyz".to_string())
+        );
+
+        let exported = config_to_kimi_code(&config, Some(&value));
+        let provider = exported
+            .as_table().unwrap()
+            .get("providers").unwrap()
+            .as_table().unwrap()
+            .get("p1").unwrap()
+            .as_table().unwrap();
+        assert_eq!(
+            provider.get("type").and_then(|v| v.as_str()),
+            Some("custom-xyz"),
+            "unknown provider type must round-trip verbatim"
+        );
+    }
+
+    #[test]
+    fn kimi_code_export_preserves_sections_added_after_import() {
+        // Top-level sections the CLI added AFTER we imported (not in the
+        // import-time baseline) must survive export instead of being wiped
+        // by the stale-key cleanup.
+        let imported_str = r#"
+default_model = "glm-5.2"
+
+[providers."glmzhongzhuan"]
+type = "anthropic"
+api_key = "sk-test"
+
+[models."glm-5.2"]
+provider = "glmzhongzhuan"
+model = "glm-5.2"
+max_context_size = 900000
+
+[thinking]
+enabled = true
+"#;
+        let imported: TomlValue = imported_str.parse().unwrap();
+        let config = kimi_code_to_config(&imported);
+
+        // The on-disk file now contains an extra [swarm] section the CLI
+        // wrote after our import.
+        let existing: TomlValue = format!("{}\n[swarm]\ntimeout_ms = 123\n", imported_str.trim_end())
+            .parse()
+            .unwrap();
+
+        let exported = config_to_kimi_code(&config, Some(&existing));
+        let root = exported.as_table().unwrap();
+        let swarm = root.get("swarm").unwrap().as_table().unwrap();
+        assert_eq!(
+            swarm.get("timeout_ms").and_then(|v| v.as_integer()),
+            Some(123),
+            "CLI-added [swarm] section must be preserved"
+        );
+        // Baseline sections are still synced as before.
+        let thinking = root.get("thinking").unwrap().as_table().unwrap();
+        assert_eq!(thinking.get("enabled").and_then(|v| v.as_bool()), Some(true));
+    }
+
+    #[test]
+    fn kimi_code_export_drops_user_removed_sections() {
+        // A section that WAS in the import baseline but was removed from
+        // raw_other (the user deleted it in the UI) must still be dropped
+        // from the exported file.
+        let toml_str = r#"
+default_model = "glm-5.2"
+
+[providers."glmzhongzhuan"]
+type = "anthropic"
+api_key = "sk-test"
+
+[models."glm-5.2"]
+provider = "glmzhongzhuan"
+model = "glm-5.2"
+max_context_size = 900000
+
+[thinking]
+enabled = true
+"#;
+        let value: TomlValue = toml_str.parse().unwrap();
+        let mut config = kimi_code_to_config(&value);
+        assert!(
+            config.imported_section_keys.contains(&"thinking".to_string()),
+            "baseline must record the imported [thinking] section"
+        );
+
+        // UI removed the thinking section.
+        let mut raw = config.raw_other.as_object().unwrap().clone();
+        raw.remove("thinking");
+        config.raw_other = Value::Object(raw);
+
+        let exported = config_to_kimi_code(&config, Some(&value));
+        let root = exported.as_table().unwrap();
+        assert!(
+            !root.contains_key("thinking"),
+            "user-removed section must be dropped on export"
+        );
+    }
 }
@@ -1,5 +1,7 @@
+use std::borrow::Cow;
+
 use indexmap::IndexMap;
-use serde::{Deserialize, Serialize};
+use serde::{Deserialize, Deserializer, Serialize, Serializer};
 use serde_json::Value;

 /// Target agent whose provider/model config is being edited.
@@ -19,18 +21,35 @@ impl Agent {
    }
 }

-#[derive(Debug, Clone, PartialEq, Eq, Serialize, Deserialize)]
-#[serde(rename_all = "snake_case")]
+#[derive(Debug, Clone, PartialEq, Eq)]
 pub enum ProviderType {
    Anthropic,
    Openai,
-    #[serde(rename = "openai_responses")]
    OpenaiResponses,
-    #[serde(rename = "google-genai")]
    GoogleGenai,
    Vertexai,
    /// Kept for compatibility; mapped to OpenAI-compatible in Pi.
    Kimi,
+    /// A provider type string the CLI knows but Kimi Switch does not.
+    /// Kept verbatim so new upstream `type` values survive the round-trip
+    /// instead of being rewritten to "kimi".
+    Unknown(String),
+}
+
+// Serde is hand-written to keep a pure string wire format: known variants
+// map to their CLI string, `Unknown(s)` maps to `s` itself. A derived
+// representation would expose the internal variant names.
+impl Serialize for ProviderType {
+    fn serialize<S: Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
+        serializer.serialize_str(&self.as_str())
+    }
+}
+
+impl<'de> Deserialize<'de> for ProviderType {
+    fn deserialize<D: Deserializer<'de>>(deserializer: D) -> Result<Self, D::Error> {
+        let s = String::deserialize(deserializer)?;
+        Ok(ProviderType::from_kimi_type(&s))
+    }
 }

 const fn default_true() -> bool {
@@ -38,14 +57,29 @@ const fn default_true() -> bool {
 }

 impl ProviderType {
-    pub fn as_str(&self) -> &'static str {
+    pub fn as_str(&self) -> Cow<'static, str> {
        match self {
-            ProviderType::Anthropic => "anthropic",
-            ProviderType::Openai => "openai",
-            ProviderType::OpenaiResponses => "openai_responses",
-            ProviderType::GoogleGenai => "google-genai",
-            ProviderType::Vertexai => "vertexai",
-            ProviderType::Kimi => "kimi",
+            ProviderType::Anthropic => Cow::Borrowed("anthropic"),
+            ProviderType::Openai => Cow::Borrowed("openai"),
+            ProviderType::OpenaiResponses => Cow::Borrowed("openai_responses"),
+            ProviderType::GoogleGenai => Cow::Borrowed("google-genai"),
+            ProviderType::Vertexai => Cow::Borrowed("vertexai"),
+            ProviderType::Kimi => Cow::Borrowed("kimi"),
+            ProviderType::Unknown(s) => Cow::Owned(s.clone()),
+        }
+    }
+
+    /// Map a CLI `type` string to a `ProviderType`; unrecognized values
+    /// become `Unknown(s)` so the original string round-trips on export.
+    pub fn from_kimi_type(s: &str) -> ProviderType {
+        match s {
+            "anthropic" => ProviderType::Anthropic,
+            "openai" => ProviderType::Openai,
+            "openai_responses" => ProviderType::OpenaiResponses,
+            "google-genai" => ProviderType::GoogleGenai,
+            "vertexai" => ProviderType::Vertexai,
+            "kimi" => ProviderType::Kimi,
+            other => ProviderType::Unknown(other.to_string()),
        }
    }

@@ -55,7 +89,7 @@ impl ProviderType {
                Some("https://api.openai.com/v1")
            }
            ProviderType::GoogleGenai => Some("https://generativelanguage.googleapis.com"),
-            ProviderType::Anthropic | ProviderType::Vertexai => None,
+            ProviderType::Anthropic | ProviderType::Vertexai | ProviderType::Unknown(_) => None,
        }
    }

@@ -203,6 +237,12 @@ pub struct Config {
    pub models: IndexMap<String, Model>,
    #[serde(default, skip_serializing_if = "Value::is_null")]
    pub raw_other: Value,
+    /// Top-level section keys captured at import time. Export uses this
+    /// baseline to distinguish "user removed this section in the UI" from
+    /// "the CLI added this section after we imported" — only baseline keys
+    /// absent from the current raw_other are dropped on export.
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub imported_section_keys: Vec<String>,
 }

 impl PartialEq for Config {
@@ -211,6 +251,7 @@ impl PartialEq for Config {
            && self.providers == other.providers
            && self.models == other.models
            && self.raw_other == other.raw_other
+            && self.imported_section_keys == other.imported_section_keys
    }
 }

@@ -205,8 +205,10 @@ pub fn pi_api_for_provider(provider_type: &ProviderType) -> &'static str {
        ProviderType::Anthropic => "anthropic-messages",
        ProviderType::GoogleGenai => "google-generative-ai",
        ProviderType::Vertexai => "google-vertex",
-        // Treat Kimi as OpenAI-compatible since it is not a native Pi API.
-        ProviderType::Kimi => "openai-completions",
+        // Treat Kimi as OpenAI-compatible since it is not a native Pi API;
+        // unknown upstream types get the same lenient handling (Pi is a
+        // legacy path).
+        ProviderType::Kimi | ProviderType::Unknown(_) => "openai-completions",
    }
 }

@@ -383,6 +385,7 @@ pub fn pi_file_to_config(file: &PiModelsFile) -> Config {
        providers,
        models,
        raw_other: file.extra.clone(),
+        imported_section_keys: Vec::new(),
    }
 }

@@ -1,6 +1,6 @@
 {
  "productName": "Kimi Switch",
-  "version": "0.7.7",
+  "version": "0.7.12",
  "identifier": "com.kimiswitch.app",
  "build": {
    "beforeDevCommand": "npm run dev",
@@ -1,7 +1,14 @@
+import { useEffect, useState } from "react";
+import { invoke } from "@tauri-apps/api/core";
 import { useTranslation } from "../i18n";
 import type { TranslationKey } from "../i18n/zh";
 import { getAgentSettings, setAgentSettings } from "../lib/agent-settings";
-import type { AgentSettings, Hook, PermissionRule } from "../types";
+import type {
+  AgentSettings,
+  ExperimentalEnvStatus,
+  Hook,
+  PermissionRule,
+} from "../types";
 import { Card, Checkbox, NumberField, Segmented } from "./ui/controls";

 interface AgentSettingsPanelProps {
@@ -9,12 +16,12 @@ interface AgentSettingsPanelProps {
  onChange: (nextRawOther: unknown) => void;
 }

-const THINKING_LEVELS = ["low", "medium", "high", "max"] as const;
+// Upstream removed the "max" effort tier (auto-migrates to "high").
+const THINKING_LEVELS = ["low", "medium", "high"] as const;
 const THINKING_LABELS: Record<(typeof THINKING_LEVELS)[number], TranslationKey> = {
  low: "thinkingLow",
  medium: "thinkingMedium",
  high: "thinkingHigh",
-  max: "thinkingMax",
 };

 const PERMISSION_DECISIONS = ["allow", "deny", "ask"] as const;
@@ -40,9 +47,23 @@ const COMMON_EVENTS = [
 export function AgentSettingsPanel({ rawOther, onChange }: AgentSettingsPanelProps) {
  const { t } = useTranslation();
  const settings = getAgentSettings(rawOther);
+  /**
+   * KIMI_CODE_LEGACY_FLAG=1 keeps the v1 loop_control keys dual-written for
+   * v1-engine users; the default (v2) writes only max_attempts_per_step.
+   */
+  const [legacyV1, setLegacyV1] = useState(false);
+
+  useEffect(() => {
+    invoke<ExperimentalEnvStatus>("get_experimental_env_status")
+      .then((env) => {
+        const value = env?.KIMI_CODE_LEGACY_FLAG ?? "";
+        setLegacyV1(["1", "true", "yes", "on"].includes(value.trim().toLowerCase()));
+      })
+      .catch(() => setLegacyV1(false));
+  }, []);

  const update = (patch: Partial<AgentSettings>) => {
-    onChange(setAgentSettings(rawOther, patch));
+    onChange(setAgentSettings(rawOther, patch, { legacyV1 }));
  };

  const updateThinking = (patch: Partial<AgentSettings["thinking"]>) => {
@@ -95,7 +116,7 @@ export function AgentSettingsPanel({ rawOther, onChange }: AgentSettingsPanelPro
          label={t("thinkingKeep")}
          checked={settings.thinking?.keep === "all"}
          disabled={!thinkingEnabled}
-          onChange={(checked) => updateThinking({ keep: checked ? "all" : false })}
+          onChange={(checked) => updateThinking({ keep: checked ? "all" : "off" })}
        />
        <p className="text-xs text-content-muted">{t("thinkingContextHint")}</p>
      </Card>
@@ -136,6 +157,25 @@ export function AgentSettingsPanel({ rawOther, onChange }: AgentSettingsPanelPro
      </Card>

      <Card title={t("permissionRules")}>
+        {/* kimi-code 0.40.1 `[permission] dangerous_command_guard`. The env
+            var KIMI_CODE_DANGEROUS_COMMAND_GUARD (literal "true"/"false")
+            outranks this config at runtime; no env-lock UI here (deliberate
+            minimal scope). Absent key = upstream default on. */}
+        <Checkbox
+          label={t("permissionDangerousGuard")}
+          checked={settings.permission?.dangerous_command_guard ?? true}
+          onChange={(checked) =>
+            update({
+              permission: {
+                ...settings.permission,
+                dangerous_command_guard: checked,
+              },
+            })
+          }
+        />
+        <p className="text-xs text-content-muted">
+          {t("permissionDangerousGuardDesc")}
+        </p>
        <div className="space-y-2">
          {(settings.permission?.rules ?? []).map((rule, idx) => (
            <div key={idx} className="flex items-center gap-2">
@@ -54,6 +54,8 @@ function defaultBaseUrl(agent: Agent, type: ProviderType): string {
      case "anthropic":
      case "vertexai":
        return "";
+      default:
+        return "";
    }
  }
  // Pi defaults
@@ -68,6 +70,8 @@ function defaultBaseUrl(agent: Agent, type: ProviderType): string {
    case "anthropic":
    case "vertexai":
      return "";
+    default:
+      return "";
  }
 }

@@ -307,6 +311,11 @@ export function ProviderEdit({
                          {t}
                        </option>
                      ))}
+                      {!PROVIDER_TYPES.includes(provider.provider_type) && (
+                        <option value={provider.provider_type}>
+                          {t("apiFormatUnknown", { type: provider.provider_type })}
+                        </option>
+                      )}
                    </select>
                  </div>
                </div>
@@ -4,6 +4,7 @@ import { Pencil, Copy, Activity, Loader2, Trash2, BarChart3 } from "lucide-react
 import { useTranslation } from "../i18n";
 import { ProviderIcon } from "./ProviderIcon";
 import { UsageFooter } from "./UsageFooter";
+import { useUsageQuery } from "../hooks/useUsageQuery";
 import type { Agent, Model, Provider } from "../types";

 interface ConnectivityResult {
@@ -32,6 +33,228 @@ interface ProviderListProps {
  agent: Agent;
 }

+interface ProviderCardProps {
+  provider: Provider;
+  agent: Agent;
+  defaultModel: string | null;
+  models: Record<string, Model>;
+  /** Inline connectivity-test state for this provider, or undefined when idle. */
+  ts: TestState | undefined;
+  onEdit: (name: string) => void;
+  onDelete: (name: string) => void;
+  onDuplicate: (name: string) => void;
+  onSwitchProvider: (name: string) => void;
+  onConfigureUsage: (name: string) => void;
+  onTest: (provider: Provider) => void;
+}
+
+function ProviderCard({
+  provider,
+  agent,
+  defaultModel,
+  models,
+  ts,
+  onEdit,
+  onDelete,
+  onDuplicate,
+  onSwitchProvider,
+  onConfigureUsage,
+  onTest,
+}: ProviderCardProps) {
+  const { t } = useTranslation();
+  // Usage query state lives once per card; both the compact (card header) and
+  // detail (footer) UsageFooter variants read this same state, so a refresh
+  // from either side updates both.
+  const usage = useUsageQuery(
+    agent,
+    provider.name,
+    provider.usageKinds,
+    provider.usageConfig?.autoQueryIntervalMinutes
+  );
+  const providerModels = Object.values(models).filter(
+    (m) => m.provider === provider.name
+  );
+  const defaultModelName = defaultModel
+    ? models[defaultModel]?.model || models[defaultModel]?.display_name || defaultModel
+    : null;
+  const isActive = provider.active === true;
+
+  return (
+    <div
+      className={`group relative flex flex-col gap-3 p-4 rounded-xl border transition-colors cursor-pointer w-full ${
+        isActive
+          ? "bg-green-50 dark:bg-green-900/10 border-green-300 dark:border-green-500/30 hover:border-green-400 dark:border-green-500/50 hover:bg-green-100 dark:bg-green-900/20"
+          : "bg-panel border-border hover:border-strong hover:bg-hover"
+      }`}
+      onClick={() => onEdit(provider.name)}
+    >
+      <div className="flex items-center gap-4 w-full">
+      <ProviderIcon
+        name={provider.name}
+        icon={provider.icon}
+        color={provider.icon_color}
+        size={44}
+      />
+
+      <div className="flex-1 min-w-0">
+        <div className="flex items-center gap-2">
+          <h3 className="font-semibold text-content-primary truncate">
+            {provider.name}
+          </h3>
+          {isActive && (
+            <span className="text-xs px-2 py-0.5 rounded-full bg-green-100 dark:bg-green-900/30 text-green-600 dark:text-green-400 border border-green-300 dark:border-green-500/30">
+              {t("inUse")}
+            </span>
+          )}
+          {provider.note && (
+            <span className="text-xs text-content-muted truncate max-w-[200px]">
+              {provider.note}
+            </span>
+          )}
+        </div>
+        <div className="mt-1 flex items-center gap-3 text-sm text-content-muted flex-wrap">
+          <span className="font-mono text-xs">
+            {provider.official_url || provider.base_url || t("noUrl")}
+          </span>
+          <span className="text-xs px-2 py-0.5 rounded-full bg-hover-2 text-content-muted border border-border">
+            {provider.provider_type}
+          </span>
+          <span className="text-xs">
+            {t("modelCount", { count: providerModels.length })}
+          </span>
+        </div>
+      </div>
+
+      <div className="flex items-center flex-wrap justify-end gap-2">
+        {defaultModel &&
+          models[defaultModel]?.provider === provider.name && (
+            <span className="text-xs px-2 py-1 rounded-full bg-green-100 dark:bg-green-900/30 text-green-600 dark:text-green-400 border border-green-300 dark:border-green-500/20">
+              {t("defaultModel", { name: defaultModelName ?? "" })}
+            </span>
+          )}
+        {/* inline connectivity test result */}
+        {ts && ts.status !== "testing" && (
+          <span
+            className={`text-xs px-2 py-0.5 rounded-full border tabular-nums ${
+              ts.status === "ok"
+                ? ts.latency && ts.latency > 6000
+                  ? "bg-orange-100 dark:bg-orange-900/30 text-orange-600 dark:text-orange-400 border-orange-300 dark:border-orange-500/30"
+                  : "bg-green-100 dark:bg-green-900/30 text-green-600 dark:text-green-400 border-green-300 dark:border-green-500/30"
+                : "bg-red-100 dark:bg-red-900/30 text-red-500 dark:text-red-400 border-red-300 dark:border-red-500/30"
+            }`}
+            title={
+              ts.status === "fail"
+                ? ts.error || t("connectivityFail", { name: provider.name, error: "" })
+                : ts.latency && ts.latency > 6000
+                  ? t("connectivitySlow", { name: provider.name, latency: ts.latency ?? 0 })
+                  : t("connectivityOk", { name: provider.name, latency: ts.latency ?? 0 })
+            }
+          >
+            {ts.status === "ok" ? `${ts.latency}ms` : "✕"}
+          </span>
+        )}
+        {/* compact usage summary — sits left of the switch button,
+            mirroring cc-switch's card layout (usage → action buttons) */}
+        {(provider.usageKinds?.length ?? 0) > 0 && (
+          <UsageFooter usage={usage} variant="compact" />
+        )}
+        <button
+          type="button"
+          onClick={(e) => {
+            e.stopPropagation();
+            onSwitchProvider(provider.name);
+          }}
+          className={`px-3 py-1.5 text-sm rounded focus:ring-2 focus:outline-none ${
+            isActive
+              ? "bg-green-600 hover:bg-green-700 text-white focus:ring-green-500"
+              : "bg-blue-600 hover:bg-blue-700 text-white focus:ring-blue-500"
+          }`}
+        >
+          {isActive ? t("inUse") : t("switchTo")}
+        </button>
+        {agent === "kimi_code" && (
+          <span
+            className="text-sm text-content-muted hover:text-content-primary cursor-help select-none"
+            title={t("switchReloadHint")}
+            aria-label={t("switchReloadHint")}
+          >
+            ⓘ
+          </span>
+        )}
+
+        {/* icon button group */}
+        <button
+          type="button"
+          onClick={(e) => {
+            e.stopPropagation();
+            onEdit(provider.name);
+          }}
+          title={t("edit")}
+          className={iconBtn}
+          aria-label={t("edit")}
+        >
+          <Pencil className="w-4 h-4" />
+        </button>
+        <button
+          type="button"
+          onClick={(e) => {
+            e.stopPropagation();
+            onDuplicate(provider.name);
+          }}
+          title={t("copyProvider")}
+          className={iconBtn}
+          aria-label={t("copyProvider")}
+        >
+          <Copy className="w-4 h-4" />
+        </button>
+        <button
+          type="button"
+          onClick={(e) => {
+            e.stopPropagation();
+            onTest(provider);
+          }}
+          disabled={ts?.status === "testing"}
+          title={t("testConnectivity")}
+          className={`${iconBtn} disabled:opacity-50`}
+          aria-label={t("testConnectivity")}
+        >
+          {ts?.status === "testing" ? (
+            <Loader2 className="w-4 h-4 animate-spin" />
+          ) : (
+            <Activity className="w-4 h-4" />
+          )}
+        </button>
+        <button
+          type="button"
+          onClick={(e) => {
+            e.stopPropagation();
+            onConfigureUsage(provider.name);
+          }}
+          title={t("usageConfigBtn")}
+          className={iconBtn}
+          aria-label={t("usageConfigBtn")}
+        >
+          <BarChart3 className="w-4 h-4" />
+        </button>
+        <button
+          type="button"
+          onClick={(e) => {
+            e.stopPropagation();
+            onDelete(provider.name);
+          }}
+          title={t("delete")}
+          className={`${iconBtn} hover:text-red-400 hover:border-red-500/30 hover:bg-red-900/20`}
+          aria-label={t("delete")}
+        >
+          <Trash2 className="w-4 h-4" />
+        </button>
+      </div>
+      </div>
+      <UsageFooter usage={usage} variant="detail" />
+    </div>
+  );
+}
+
 const iconBtn =
  "w-8 h-8 flex items-center justify-center rounded border border-border text-content-muted hover:text-content-primary hover:bg-hover-2 focus:ring-2 focus:ring-blue-500 focus:outline-none transition-colors";

@@ -139,205 +362,22 @@ export function ProviderList({
          </div>
        ) : (
          <div className="grid grid-cols-1 gap-3 w-full">
-            {sortedProviders.map((provider) => {
-              const providerModels = Object.values(models).filter(
-                (m) => m.provider === provider.name
-              );
-              const defaultModelName = defaultModel
-                ? models[defaultModel]?.model || models[defaultModel]?.display_name || defaultModel
-                : null;
-              const isActive = provider.active === true;
-              const ts = testState[provider.name];
-
-              return (
-                <div
-                  key={provider.name}
-                  className={`group relative flex flex-col gap-3 p-4 rounded-xl border transition-colors cursor-pointer w-full ${
-                    isActive
-                      ? "bg-green-50 dark:bg-green-900/10 border-green-300 dark:border-green-500/30 hover:border-green-400 dark:border-green-500/50 hover:bg-green-100 dark:bg-green-900/20"
-                      : "bg-panel border-border hover:border-strong hover:bg-hover"
-                  }`}
-                  onClick={() => onEdit(provider.name)}
-                >
-                  <div className="flex items-center gap-4 w-full">
-                  <ProviderIcon
-                    name={provider.name}
-                    icon={provider.icon}
-                    color={provider.icon_color}
-                    size={44}
-                  />
-
-                  <div className="flex-1 min-w-0">
-                    <div className="flex items-center gap-2">
-                      <h3 className="font-semibold text-content-primary truncate">
-                        {provider.name}
-                      </h3>
-                      {isActive && (
-                        <span className="text-xs px-2 py-0.5 rounded-full bg-green-100 dark:bg-green-900/30 text-green-600 dark:text-green-400 border border-green-300 dark:border-green-500/30">
-                          {t("inUse")}
-                        </span>
-                      )}
-                      {provider.note && (
-                        <span className="text-xs text-content-muted truncate max-w-[200px]">
-                          {provider.note}
-                        </span>
-                      )}
-                    </div>
-                    <div className="mt-1 flex items-center gap-3 text-sm text-content-muted flex-wrap">
-                      <span className="font-mono text-xs">
-                        {provider.official_url || provider.base_url || t("noUrl")}
-                      </span>
-                      <span className="text-xs px-2 py-0.5 rounded-full bg-hover-2 text-content-muted border border-border">
-                        {provider.provider_type}
-                      </span>
-                      <span className="text-xs">
-                        {t("modelCount", { count: providerModels.length })}
-                      </span>
-                    </div>
-                  </div>
-
-                  <div className="flex items-center flex-wrap justify-end gap-2">
-                    {defaultModel &&
-                      models[defaultModel]?.provider === provider.name && (
-                        <span className="text-xs px-2 py-1 rounded-full bg-green-100 dark:bg-green-900/30 text-green-600 dark:text-green-400 border border-green-300 dark:border-green-500/20">
-                          {t("defaultModel", { name: defaultModelName ?? "" })}
-                        </span>
-                      )}
-                    {/* inline connectivity test result */}
-                    {ts && ts.status !== "testing" && (
-                      <span
-                        className={`text-xs px-2 py-0.5 rounded-full border tabular-nums ${
-                          ts.status === "ok"
-                            ? ts.latency && ts.latency > 6000
-                              ? "bg-orange-100 dark:bg-orange-900/30 text-orange-600 dark:text-orange-400 border-orange-300 dark:border-orange-500/30"
-                              : "bg-green-100 dark:bg-green-900/30 text-green-600 dark:text-green-400 border-green-300 dark:border-green-500/30"
-                            : "bg-red-100 dark:bg-red-900/30 text-red-500 dark:text-red-400 border-red-300 dark:border-red-500/30"
-                        }`}
-                        title={
-                          ts.status === "fail"
-                            ? ts.error || t("connectivityFail", { name: provider.name, error: "" })
-                            : ts.latency && ts.latency > 6000
-                              ? t("connectivitySlow", { name: provider.name, latency: ts.latency ?? 0 })
-                              : t("connectivityOk", { name: provider.name, latency: ts.latency ?? 0 })
-                        }
-                      >
-                        {ts.status === "ok" ? `${ts.latency}ms` : "✕"}
-                      </span>
-                    )}
-                    {/* compact usage summary — sits left of the switch button,
-                        mirroring cc-switch's card layout (usage → action buttons) */}
-                    {(provider.usageKinds?.length ?? 0) > 0 && (
-                      <UsageFooter
-                        agent={agent}
-                        providerName={provider.name}
-                        usageKinds={provider.usageKinds}
-                        variant="compact"
-                        autoIntervalMinutes={
-                          provider.usageConfig?.autoQueryIntervalMinutes
-                        }
-                      />
-                    )}
-                    <button
-                      type="button"
-                      onClick={(e) => {
-                        e.stopPropagation();
-                        onSwitchProvider(provider.name);
-                      }}
-                      className={`px-3 py-1.5 text-sm rounded focus:ring-2 focus:outline-none ${
-                        isActive
-                          ? "bg-green-600 hover:bg-green-700 text-white focus:ring-green-500"
-                          : "bg-blue-600 hover:bg-blue-700 text-white focus:ring-blue-500"
-                      }`}
-                    >
-                      {isActive ? t("inUse") : t("switchTo")}
-                    </button>
-                    {agent === "kimi_code" && (
-                      <span
-                        className="text-sm text-content-muted hover:text-content-primary cursor-help select-none"
-                        title={t("switchReloadHint")}
-                        aria-label={t("switchReloadHint")}
-                      >
-                        ⓘ
-                      </span>
-                    )}
-
-                    {/* icon button group */}
-                    <button
-                      type="button"
-                      onClick={(e) => {
-                        e.stopPropagation();
-                        onEdit(provider.name);
-                      }}
-                      title={t("edit")}
-                      className={iconBtn}
-                      aria-label={t("edit")}
-                    >
-                      <Pencil className="w-4 h-4" />
-                    </button>
-                    <button
-                      type="button"
-                      onClick={(e) => {
-                        e.stopPropagation();
-                        onDuplicate(provider.name);
-                      }}
-                      title={t("copyProvider")}
-                      className={iconBtn}
-                      aria-label={t("copyProvider")}
-                    >
-                      <Copy className="w-4 h-4" />
-                    </button>
-                    <button
-                      type="button"
-                      onClick={(e) => {
-                        e.stopPropagation();
-                        handleTest(provider);
-                      }}
-                      disabled={ts?.status === "testing"}
-                      title={t("testConnectivity")}
-                      className={`${iconBtn} disabled:opacity-50`}
-                      aria-label={t("testConnectivity")}
-                    >
-                      {ts?.status === "testing" ? (
-                        <Loader2 className="w-4 h-4 animate-spin" />
-                      ) : (
-                        <Activity className="w-4 h-4" />
-                      )}
-                    </button>
-                    <button
-                      type="button"
-                      onClick={(e) => {
-                        e.stopPropagation();
-                        onConfigureUsage(provider.name);
-                      }}
-                      title={t("usageConfigBtn")}
-                      className={iconBtn}
-                      aria-label={t("usageConfigBtn")}
-                    >
-                      <BarChart3 className="w-4 h-4" />
-                    </button>
-                    <button
-                      type="button"
-                      onClick={(e) => {
-                        e.stopPropagation();
-                        onDelete(provider.name);
-                      }}
-                      title={t("delete")}
-                      className={`${iconBtn} hover:text-red-400 hover:border-red-500/30 hover:bg-red-900/20`}
-                      aria-label={t("delete")}
-                    >
-                      <Trash2 className="w-4 h-4" />
-                    </button>
-                  </div>
-                  </div>
-                  <UsageFooter
-                    agent={agent}
-                    providerName={provider.name}
-                    usageKinds={provider.usageKinds}
-                    variant="detail"
-                  />
-                </div>
-              );
-            })}
+            {sortedProviders.map((provider) => (
+              <ProviderCard
+                key={provider.name}
+                provider={provider}
+                agent={agent}
+                defaultModel={defaultModel}
+                models={models}
+                ts={testState[provider.name]}
+                onEdit={onEdit}
+                onDelete={onDelete}
+                onDuplicate={onDuplicate}
+                onSwitchProvider={onSwitchProvider}
+                onConfigureUsage={onConfigureUsage}
+                onTest={handleTest}
+              />
+            ))}
          </div>
        )}
      </div>
@@ -9,6 +9,7 @@ import {
  forcedEnvValue,
  getExperimentalFlags,
  getSubagentModelPool,
+  isExperimentalFlagSet,
  isFlagLockedByEnv,
  isMasterEnvOn,
  removeSubagentPoolEntry,
@@ -33,22 +34,31 @@ interface SubagentSettingsPageProps {
  onBack: () => void;
 }

-const EFFORTS = ["low", "medium", "high", "max"] as const;
+// Upstream removed the "max" effort tier (auto-migrates to "high").
+const EFFORTS = ["low", "medium", "high"] as const;
 const EFFORT_LABELS: Record<(typeof EFFORTS)[number], TranslationKey> = {
  low: "thinkingLow",
  medium: "thinkingMedium",
  high: "thinkingHigh",
-  max: "thinkingMax",
 };

 const FLAG_LABELS: Record<string, { name: TranslationKey; desc: TranslationKey }> = {
  "secondary-model": { name: "flagSecondaryModel", desc: "flagSecondaryModelDesc" },
  "tool-select": { name: "flagToolSelect", desc: "flagToolSelectDesc" },
-  "acp-v2": { name: "flagAcpV2", desc: "flagAcpV2Desc" },
  persistence_minidb_readmodel: {
    name: "flagMinidbReadmodel",
    desc: "flagMinidbReadmodelDesc",
  },
+  tower: { name: "flagTower", desc: "flagTowerDesc" },
+  subagent_fork: { name: "flagSubagentFork", desc: "flagSubagentForkDesc" },
+  wait_for: { name: "flagWaitFor", desc: "flagWaitForDesc" },
+  auto_session_title: {
+    name: "flagAutoSessionTitle",
+    desc: "flagAutoSessionTitleDesc",
+  },
+  "remote-control": { name: "flagRemoteControl", desc: "flagRemoteControlDesc" },
+  search_worker: { name: "flagSearchWorker", desc: "flagSearchWorkerDesc" },
+  file_history: { name: "flagFileHistory", desc: "flagFileHistoryDesc" },
 };

 /** Validation-error i18n key per engine rule, for the pre-write self-check. */
@@ -154,6 +164,8 @@ export function SubagentSettingsPage({
  const [poolErrors, setPoolErrors] = useState<PoolValidationError[]>([]);
  /** Local draft for the "add pool entry" select (reset after each pick). */
  const [addSelection, setAddSelection] = useState("");
+  /** WebUI-open button in flight; disables both buttons while non-null. */
+  const [webuiBusy, setWebuiBusy] = useState<"embedded" | "browser" | null>(null);

  useEffect(() => {
    invoke<ExperimentalEnvStatus>("get_experimental_env_status")
@@ -177,13 +189,18 @@ export function SubagentSettingsPage({
  const secondaryFlag = EXPERIMENTAL_FLAGS[0]; // "secondary-model"
  const otherFlags = EXPERIMENTAL_FLAGS.slice(1);

-  /** Effective state of the secondary-model flag (env overrides config). */
-  const secondaryEnabled = masterOn
-    ? true
-    : isFlagLockedByEnv(env, secondaryFlag)
-      ? forcedEnvValue(env, secondaryFlag)
-      : flags["secondary-model"] === true;
-  const secondaryLocked = masterOn || isFlagLockedByEnv(env, secondaryFlag);
+  /** Effective state of the secondary-model flag. Mirrors the kimi-code
+   *  0.40.1 priority: single-flag env > explicit [experimental] value >
+   *  master env (force-on only) > upstream default. The master env no
+   *  longer locks the toggle — only the flag's own env var does. */
+  const secondaryEnabled = isFlagLockedByEnv(env, secondaryFlag)
+    ? forcedEnvValue(env, secondaryFlag)
+    : isExperimentalFlagSet(rawOther, secondaryFlag.id)
+      ? flags["secondary-model"] === true
+      : masterOn
+        ? true
+        : (secondaryFlag.defaultEnabled ?? false);
+  const secondaryLocked = isFlagLockedByEnv(env, secondaryFlag);

  const aliasList = Object.keys(models).sort();

@@ -229,7 +246,9 @@ export function SubagentSettingsPage({
        ? `${Math.round(n / 1000)}K`
        : String(n);
  const effortLabel = (e: string): string => {
-    const key = EFFORT_LABELS[e as (typeof EFFORTS)[number]];
+    // Stored "max" tiers from old configs are shown as "high" (upstream
+    // removed the tier; it auto-migrates to "high").
+    const key = EFFORT_LABELS[e === "max" ? "high" : (e as (typeof EFFORTS)[number])];
    return key ? t(key) : e;
  };

@@ -290,16 +309,35 @@ export function SubagentSettingsPage({

  const renderFlagRow = (flag: ExperimentalFlagDef) => {
    const labels = FLAG_LABELS[flag.id];
-    const locked = masterOn || isFlagLockedByEnv(env, flag);
-    const on = masterOn
-      ? true
-      : isFlagLockedByEnv(env, flag)
-        ? forcedEnvValue(env, flag)
-        : flags[flag.id] === true;
+    // Only the flag's own env var locks the toggle. The master env force-ons
+    // undefined flags but an explicit config entry outranks it, so the user
+    // can still flip values here while it is set.
+    const locked = isFlagLockedByEnv(env, flag);
+    // Resolution order mirrors kimi-code 0.40.1 flagService:
+    // single-flag env > explicit [experimental] value (true *or* false) >
+    // master env (on-only) > upstream default (`defaultEnabled`).
+    const explicitlySet = isExperimentalFlagSet(rawOther, flag.id);
+    const on = locked
+      ? forcedEnvValue(env, flag)
+      : explicitlySet
+        ? flags[flag.id] === true
+        : masterOn
+          ? true
+          : (flag.defaultEnabled ?? false);
+    // Turning the flag off only sticks via an explicit `false` when the
+    // fallback (master env or the upstream default) would otherwise keep it on.
+    const explicitFalse = masterOn || flag.defaultEnabled === true;
    return (
      <div key={flag.id} className="flex items-center gap-3">
        <div className="flex-1 min-w-0">
-          <div className="text-sm text-content-primary">{t(labels.name)}</div>
+          <div className="flex items-center gap-2 text-sm text-content-primary">
+            {t(labels.name)}
+            {!locked && !explicitlySet && flag.defaultEnabled && (
+              <span className="shrink-0 text-xs text-blue-600 dark:text-blue-400 border border-blue-500/30 rounded px-1">
+                {t("flagDefaultOn")}
+              </span>
+            )}
+          </div>
          <div className="text-xs text-content-muted">
            {t(labels.desc)}
            <code className="ml-2">{flag.envVar}</code>
@@ -315,7 +353,9 @@ export function SubagentSettingsPage({
          disabled={locked}
          ariaLabel={t(labels.name)}
          onChange={(checked) =>
-            onChange(setExperimentalFlag(rawOther, flag.id, checked))
+            onChange(
+              setExperimentalFlag(rawOther, flag.id, checked, { explicitFalse })
+            )
          }
        />
      </div>
@@ -349,29 +389,37 @@ export function SubagentSettingsPage({
          <div className="mt-3 flex flex-wrap gap-2">
            <button
              type="button"
+              disabled={webuiBusy !== null}
              onClick={async () => {
+                setWebuiBusy("embedded");
                try {
                  await invoke("open_kimi_web_embedded");
                } catch (err) {
                  alert(err instanceof Error ? err.message : String(err));
+                } finally {
+                  setWebuiBusy(null);
                }
              }}
-              className="px-3 py-1.5 text-sm rounded bg-blue-600 text-white hover:bg-blue-500 focus:ring-2 focus:ring-blue-500 focus:outline-none"
+              className="px-3 py-1.5 text-sm rounded bg-blue-600 text-white hover:bg-blue-500 focus:ring-2 focus:ring-blue-500 focus:outline-none disabled:opacity-50 disabled:cursor-not-allowed"
            >
-              {t("openWebUIEmbedded")}
+              {webuiBusy === "embedded" ? t("webuiOpening") : t("openWebUIEmbedded")}
            </button>
            <button
              type="button"
+              disabled={webuiBusy !== null}
              onClick={async () => {
+                setWebuiBusy("browser");
                try {
                  await invoke("open_kimi_web");
                } catch (err) {
                  alert(err instanceof Error ? err.message : String(err));
+                } finally {
+                  setWebuiBusy(null);
                }
              }}
-              className="px-3 py-1.5 text-sm border border-border rounded hover:bg-hover-2 focus:ring-2 focus:ring-blue-500 focus:outline-none"
+              className="px-3 py-1.5 text-sm border border-border rounded hover:bg-hover-2 focus:ring-2 focus:ring-blue-500 focus:outline-none disabled:opacity-50 disabled:cursor-not-allowed"
            >
-              {t("openWebUIBrowser")}
+              {webuiBusy === "browser" ? t("webuiOpening") : t("openWebUIBrowser")}
            </button>
          </div>
        </Card>
@@ -528,6 +576,16 @@ export function SubagentSettingsPage({
          <div className="border-t border-border" />
          <div className="flex items-center gap-3">
            <div className="flex-1 min-w-0">
+              <div className="flex items-center gap-2 text-sm text-content-primary">
+                {t(FLAG_LABELS[secondaryFlag.id].name)}
+                {!secondaryLocked &&
+                  !isExperimentalFlagSet(rawOther, secondaryFlag.id) &&
+                  secondaryFlag.defaultEnabled && (
+                    <span className="shrink-0 text-xs text-blue-600 dark:text-blue-400 border border-blue-500/30 rounded px-1">
+                      {t("flagDefaultOn")}
+                    </span>
+                  )}
+              </div>
              <div className="text-xs text-content-muted">
                {t(FLAG_LABELS[secondaryFlag.id].desc)}
                <code className="ml-2">{secondaryFlag.envVar}</code>
@@ -543,7 +601,11 @@ export function SubagentSettingsPage({
              disabled={secondaryLocked}
              ariaLabel={t(FLAG_LABELS[secondaryFlag.id].name)}
              onChange={(checked) =>
-                onChange(setExperimentalFlag(rawOther, secondaryFlag.id, checked))
+                onChange(
+                  setExperimentalFlag(rawOther, secondaryFlag.id, checked, {
+                    explicitFalse: masterOn || secondaryFlag.defaultEnabled === true,
+                  })
+                )
              }
            />
          </div>
@@ -1,175 +1,23 @@
-import { useCallback, useEffect, useRef, useState } from "react";
-import { invoke } from "@tauri-apps/api/core";
+import { useEffect, useState } from "react";
 import { Loader2, RefreshCw } from "lucide-react";
 import { useTranslation } from "../i18n";
 import { localizeUsageError, planLabel } from "../lib/usage-display";
-import type { Agent } from "../types";
-
-interface UsageData {
-  planName?: string | null;
-  remaining?: number | null;
-  total?: number | null;
-  used?: number | null;
-  unit?: string | null;
-  isValid?: boolean | null;
-  resetsAt?: string | null;
-}
-
-interface UsageResult {
-  success: boolean;
-  data?: UsageData[] | null;
-  error?: string | null;
-}
-
-type UsageStatus = "idle" | "loading" | "success" | "error";
-
-interface CacheEntry {
-  status: "success" | "error";
-  data: UsageData[];
-  /** Raw error text from Rust; null = transient failure (invoke rejected). */
-  error: string | null;
-  updatedAt: number;
-}
-
-// Module-level cache shared across mounts: re-entering the list within the
-// stale TTL shows the last result without firing new requests. Both the
-// compact (card header) and detail (footer) variants read this same cache.
-const STALE_TTL_MS = 5 * 60 * 1000;
-const cache = new Map<string, CacheEntry>();
-
-// Simple semaphore: at most MAX_CONCURRENT queries in flight at once.
-const MAX_CONCURRENT = 3;
-let running = 0;
-const waiters: Array<() => void> = [];
-async function acquireSlot(): Promise<void> {
-  if (running >= MAX_CONCURRENT) {
-    await new Promise<void>((resolve) => waiters.push(resolve));
-  }
-  running += 1;
-}
-function releaseSlot(): void {
-  running -= 1;
-  waiters.shift()?.();
-}
+import type { UsageData, UsageQueryState } from "../hooks/useUsageQuery";

 interface UsageFooterProps {
-  agent: Agent;
-  providerName: string;
-  usageKinds?: string[];
+  /** Query state lifted to the provider card via useUsageQuery. Both the
+   * compact (card header) and detail (footer) variants read the same state,
+   * so a refresh from either updates both. */
+  usage: UsageQueryState;
  /** "detail" (default) renders the multi-line footer; "compact" renders a
   * one-line summary + last-updated + refresh button for the card header. */
  variant?: "detail" | "compact";
-  /** Auto query interval in minutes; only wired on the compact variant so
-   * both variants do not double-fire. 0/undefined = manual only. */
-  autoIntervalMinutes?: number;
 }

-export function UsageFooter({
-  agent,
-  providerName,
-  usageKinds,
-  variant = "detail",
-  autoIntervalMinutes,
-}: UsageFooterProps) {
+export function UsageFooter({ usage, variant = "detail" }: UsageFooterProps) {
  const { t } = useTranslation();
-  const supported = (usageKinds?.length ?? 0) > 0;
-  // Cache key includes the agent so a Kimi Code provider and a Pi provider
-  // with the same name do not clobber each other's cached result.
-  const cacheKey = `${agent}:${providerName}`;
-
-  const [status, setStatus] = useState<UsageStatus>("idle");
-  const [data, setData] = useState<UsageData[]>([]);
-  /** undefined = no error; null = network error; string = Rust error text. */
-  const [error, setError] = useState<string | null | undefined>(undefined);
-  const [updatedAt, setUpdatedAt] = useState<number | null>(null);
+  const { status, data, error, updatedAt, supported, refresh } = usage;
  const [now, setNow] = useState(() => Date.now());
-  // Generation counter: stale responses (unmounted / superseded query) are ignored.
-  const genRef = useRef(0);
-
-  const runQuery = useCallback(
-    async (forceRefresh: boolean) => {
-      const gen = ++genRef.current;
-      setStatus("loading");
-      setError(undefined);
-      const finish = (entry: CacheEntry) => {
-        cache.set(cacheKey, entry);
-        if (genRef.current !== gen) return;
-        setStatus(entry.status);
-        setData(entry.data);
-        setError(entry.error);
-        setUpdatedAt(entry.updatedAt);
-      };
-      try {
-        await acquireSlot();
-        let result: UsageResult;
-        try {
-          result = await invoke<UsageResult>("query_provider_usage", {
-            agent,
-            providerName,
-            forceRefresh,
-          });
-        } finally {
-          releaseSlot();
-        }
-        if (genRef.current !== gen) return;
-        if (result.success) {
-          finish({
-            status: "success",
-            data: result.data ?? [],
-            error: null,
-            updatedAt: Date.now(),
-          });
-        } else {
-          // Deterministic failure: keep last good data for ghost display.
-          finish({
-            status: "error",
-            data: cache.get(cacheKey)?.data ?? [],
-            error: result.error ?? "",
-            updatedAt: Date.now(),
-          });
-        }
-      } catch {
-        // Transient failure (network / timeout): invoke rejected.
-        finish({
-          status: "error",
-          data: cache.get(cacheKey)?.data ?? [],
-          error: null,
-          updatedAt: Date.now(),
-        });
-      }
-    },
-    [agent, providerName, cacheKey]
-  );
-
-  // On mount / provider change: serve fresh cache, otherwise query once.
-  useEffect(() => {
-    if (!supported) return;
-    const cached = cache.get(cacheKey);
-    if (cached) {
-      setStatus(cached.status);
-      setData(cached.data);
-      setError(cached.error);
-      setUpdatedAt(cached.updatedAt);
-      if (Date.now() - cached.updatedAt < STALE_TTL_MS) return;
-    }
-    void runQuery(false);
-  }, [supported, cacheKey, runQuery]);
-
-  // Ignore late responses after unmount.
-  useEffect(() => {
-    return () => {
-      genRef.current += 1;
-    };
-  }, []);
-
-  // Auto query interval (compact variant only, to avoid double-firing).
-  useEffect(() => {
-    if (variant !== "compact" || !supported) return;
-    const mins = autoIntervalMinutes ?? 0;
-    if (mins <= 0) return;
-    const id = setInterval(() => void runQuery(false), mins * 60_000);
-    return () => clearInterval(id);
-  }, [variant, supported, autoIntervalMinutes, runQuery]);

  // Tick relative times once a minute while showing data.
  useEffect(() => {
@@ -226,7 +74,7 @@ export function UsageFooter({
      type="button"
      onClick={(e) => {
        e.stopPropagation();
-        void runQuery(true);
+        void refresh(true);
      }}
      disabled={status === "loading"}
      title={t("usageRefresh")}
@@ -344,7 +192,7 @@ export function UsageFooter({
          type="button"
          onClick={(e) => {
            e.stopPropagation();
-            void runQuery(true);
+            void refresh(true);
          }}
          className="ml-auto shrink-0 px-2 py-0.5 rounded border border-red-300 dark:border-red-500/30 text-red-500 dark:text-red-400 hover:bg-red-50 dark:hover:bg-red-900/20 transition-colors"
        >
@@ -0,0 +1,174 @@
+import { useCallback, useEffect, useRef, useState } from "react";
+import { invoke } from "@tauri-apps/api/core";
+import type { Agent } from "../types";
+
+export interface UsageData {
+  planName?: string | null;
+  remaining?: number | null;
+  total?: number | null;
+  used?: number | null;
+  unit?: string | null;
+  isValid?: boolean | null;
+  resetsAt?: string | null;
+}
+
+export interface UsageResult {
+  success: boolean;
+  data?: UsageData[] | null;
+  error?: string | null;
+}
+
+export type UsageStatus = "idle" | "loading" | "success" | "error";
+
+export interface CacheEntry {
+  status: "success" | "error";
+  data: UsageData[];
+  /** Raw error text from Rust; null = transient failure (invoke rejected). */
+  error: string | null;
+  updatedAt: number;
+}
+
+// Module-level cache shared across mounts: re-entering the list within the
+// stale TTL shows the last result without firing new requests. Both the
+// compact (card header) and detail (footer) variants read this same cache.
+const STALE_TTL_MS = 5 * 60 * 1000;
+const cache = new Map<string, CacheEntry>();
+
+// Simple semaphore: at most MAX_CONCURRENT queries in flight at once.
+const MAX_CONCURRENT = 3;
+let running = 0;
+const waiters: Array<() => void> = [];
+async function acquireSlot(): Promise<void> {
+  if (running >= MAX_CONCURRENT) {
+    await new Promise<void>((resolve) => waiters.push(resolve));
+  }
+  running += 1;
+}
+function releaseSlot(): void {
+  running -= 1;
+  waiters.shift()?.();
+}
+
+export interface UsageQueryState {
+  status: UsageStatus;
+  data: UsageData[];
+  error: string | null | undefined;
+  updatedAt: number | null;
+  refresh: (force: boolean) => void;
+  supported: boolean;
+}
+
+export function useUsageQuery(
+  agent: Agent,
+  providerName: string,
+  usageKinds?: string[],
+  autoIntervalMinutes?: number
+): UsageQueryState {
+  const supported = (usageKinds?.length ?? 0) > 0;
+  // Cache key includes the agent so a Kimi Code provider and a Pi provider
+  // with the same name do not clobber each other's cached result.
+  const cacheKey = `${agent}:${providerName}`;
+
+  const [status, setStatus] = useState<UsageStatus>("idle");
+  const [data, setData] = useState<UsageData[]>([]);
+  /** undefined = no error; null = network error; string = Rust error text. */
+  const [error, setError] = useState<string | null | undefined>(undefined);
+  const [updatedAt, setUpdatedAt] = useState<number | null>(null);
+  // Generation counter: stale responses (unmounted / superseded query) are ignored.
+  const genRef = useRef(0);
+
+  const runQuery = useCallback(
+    async (forceRefresh: boolean) => {
+      const gen = ++genRef.current;
+      setStatus("loading");
+      setError(undefined);
+      const finish = (entry: CacheEntry) => {
+        cache.set(cacheKey, entry);
+        if (genRef.current !== gen) return;
+        setStatus(entry.status);
+        setData(entry.data);
+        setError(entry.error);
+        setUpdatedAt(entry.updatedAt);
+      };
+      try {
+        await acquireSlot();
+        let result: UsageResult;
+        try {
+          result = await invoke<UsageResult>("query_provider_usage", {
+            agent,
+            providerName,
+            forceRefresh,
+          });
+        } finally {
+          releaseSlot();
+        }
+        if (genRef.current !== gen) return;
+        if (result.success) {
+          finish({
+            status: "success",
+            data: result.data ?? [],
+            error: null,
+            updatedAt: Date.now(),
+          });
+        } else {
+          // Deterministic failure: keep last good data for ghost display.
+          finish({
+            status: "error",
+            data: cache.get(cacheKey)?.data ?? [],
+            error: result.error ?? "",
+            updatedAt: Date.now(),
+          });
+        }
+      } catch {
+        // Transient failure (network / timeout): invoke rejected.
+        finish({
+          status: "error",
+          data: cache.get(cacheKey)?.data ?? [],
+          error: null,
+          updatedAt: Date.now(),
+        });
+      }
+    },
+    [agent, providerName, cacheKey]
+  );
+
+  // On mount / provider change: serve fresh cache, otherwise query once.
+  useEffect(() => {
+    if (!supported) return;
+    const cached = cache.get(cacheKey);
+    if (cached) {
+      setStatus(cached.status);
+      setData(cached.data);
+      setError(cached.error);
+      setUpdatedAt(cached.updatedAt);
+      if (Date.now() - cached.updatedAt < STALE_TTL_MS) return;
+    }
+    void runQuery(false);
+  }, [supported, cacheKey, runQuery]);
+
+  // Ignore late responses after unmount.
+  useEffect(() => {
+    return () => {
+      genRef.current += 1;
+    };
+  }, []);
+
+  // Auto query interval — the hook is called once per provider card and both
+  // variants read the same state, so there is no second mount to double-fire.
+  useEffect(() => {
+    if (!supported) return;
+    const mins = autoIntervalMinutes ?? 0;
+    if (mins <= 0) return;
+    const id = setInterval(() => void runQuery(false), mins * 60_000);
+    return () => clearInterval(id);
+  }, [supported, autoIntervalMinutes, runQuery]);
+
+  return {
+    status,
+    data,
+    error,
+    updatedAt,
+    refresh: runQuery,
+    supported,
+  };
+}
@@ -74,6 +74,7 @@ export const enTranslations: Record<TranslationKey, string> = {
  kimiOAuthStart: "Start sign-in",
  apiSettings: "API Settings",
  apiFormat: "API Format",
+  apiFormatUnknown: "Unknown format ({type}, kept verbatim)",
  authField: "Auth Field",
  apiKey: "API Key",
  getApiKeyLink: "Get API Key",
@@ -137,7 +138,6 @@ export const enTranslations: Record<TranslationKey, string> = {
  thinkingLow: "Low",
  thinkingMedium: "Medium",
  thinkingHigh: "High",
-  thinkingMax: "Max",
  thinkingContextHint: "Thinking uses more context. Ensure the model context length and reserved size are sufficient.",
  loopControlSettings: "Loop Control",
  maxAttemptsPerStep: "Max attempts per step",
@@ -151,6 +151,9 @@ export const enTranslations: Record<TranslationKey, string> = {
  permissionAllow: "Allow",
  permissionDeny: "Deny",
  permissionAsk: "Ask",
+  permissionDangerousGuard: "Dangerous Command Guard",
+  permissionDangerousGuardDesc:
+    "When on (default): Auto mode refuses dangerous commands (rm -rf, shutdown, dd of=, ...) and other modes always prompt; turning it off restores the previous behavior. Env KIMI_CODE_DANGEROUS_COMMAND_GUARD overrides this config",
  addRule: "+ Add rule",
  addCommonRules: "Add common rules",
  hooks: "Lifecycle Hooks",
@@ -465,12 +468,27 @@ export const enTranslations: Record<TranslationKey, string> = {
  flagSecondaryModelDesc: "Newly spawned subagents bind a second (usually cheaper) model config",
  flagToolSelect: "Tool Select",
  flagToolSelectDesc: "Enable the experimental tool selection feature",
-  flagAcpV2: "ACP v2 Protocol",
-  flagAcpV2Desc: "Enable Agent Communication Protocol v2",
  flagMinidbReadmodel: "MiniDB Read Model",
  flagMinidbReadmodelDesc: "Enable the v2 engine MiniDB read-model backend",
+  flagTower: "Tower Mode",
+  flagTowerDesc: "Enable tower mode: coordinate multiple agents on a shared objective (/tower command)",
+  flagSubagentFork: "Subagent Fork Context",
+  flagSubagentForkDesc: "Agent/AgentSwarm tools can start a subagent with a snapshot of the calling agent's history (fork parameter)",
+  flagWaitFor: "WaitFor Tool",
+  flagWaitForDesc: "The agent can wait for background tasks to finish within the current turn",
+  flagAutoSessionTitle: "Auto Session Title",
+  flagAutoSessionTitleDesc: "Generate session titles automatically via the managed chat_title tool",
+  flagRemoteControl: "Remote Control",
+  flagRemoteControlDesc: "Enable the experimental remote control feature",
+  flagSearchWorker: "Search Worker",
+  flagSearchWorkerDesc: "Run the kap-server global search backend in a background worker host",
+  flagFileHistory: "File History",
+  flagFileHistoryDesc:
+    "Snapshot edited files before/after each turn (last 5 turns across 30 sessions, stored in the session data dir) for real whole-file diffs",
+  flagDefaultOn: "On by default",
  lockedByEnv: "Locked by env var",
-  masterEnvOnHint: "KIMI_CODE_EXPERIMENTAL_FLAG is set: all experimental features are forced on by the env var",
+  masterEnvOnHint:
+    "KIMI_CODE_EXPERIMENTAL_FLAG is set: flags without an explicit [experimental] entry are forced on; explicit entries in config.toml still win",
  secondaryModelSection: "Subagent Model (Secondary Model)",
  secondaryModelDisabledHint:
    "The experimental flag \"Subagent Secondary Model\" is off and this config will not take effect — enable the toggle above first",
@@ -519,6 +537,7 @@ export const enTranslations: Record<TranslationKey, string> = {
    "Open the new Web UI in an independent in-app window, or in your system browser (built into kimi-code 0.33+). Requires the kimi CLI on PATH.",
  openWebUIEmbedded: "Open in App",
  openWebUIBrowser: "Open in Browser",
+  webuiOpening: "Opening...",

  // Plugin marketplace
  pluginMarketplace: "Plugin Marketplace",
@@ -72,6 +72,7 @@ export const zhTranslations = {
  kimiOAuthStart: "开始登录",
  apiSettings: "API 设置",
  apiFormat: "API 格式",
+  apiFormatUnknown: "未知格式（{type}，原样保留）",
  authField: "认证字段",
  apiKey: "API Key",
  getApiKeyLink: "获取 API Key",
@@ -135,7 +136,6 @@ export const zhTranslations = {
  thinkingLow: "低",
  thinkingMedium: "中",
  thinkingHigh: "高",
-  thinkingMax: "最大",
  thinkingContextHint: "启用思考会占用更多上下文，请确保模型上下文长度和预留空间足够。",
  loopControlSettings: "循环控制",
  maxAttemptsPerStep: "单步最大尝试次数",
@@ -149,6 +149,9 @@ export const zhTranslations = {
  permissionAllow: "允许",
  permissionDeny: "拒绝",
  permissionAsk: "询问",
+  permissionDangerousGuard: "危险命令防护",
+  permissionDangerousGuardDesc:
+    "开启（默认）时：Auto 模式直接拒绝危险命令（rm -rf、shutdown、dd of= 等），其它模式强制询问；关闭后回归旧行为。环境变量 KIMI_CODE_DANGEROUS_COMMAND_GUARD 会覆盖本配置",
  addRule: "+ 添加规则",
  addCommonRules: "添加常用规则",
  hooks: "生命周期钩子",
@@ -460,12 +463,27 @@ export const zhTranslations = {
  flagSecondaryModelDesc: "子代理派发时绑定第二份（通常更便宜的）模型配置",
  flagToolSelect: "工具选择",
  flagToolSelectDesc: "启用实验性工具选择功能",
-  flagAcpV2: "ACP v2 协议",
-  flagAcpV2Desc: "启用 Agent Communication Protocol v2",
  flagMinidbReadmodel: "MiniDB 读模型",
  flagMinidbReadmodelDesc: "启用 v2 引擎 MiniDB 读模型后端",
+  flagTower: "Tower 模式",
+  flagTowerDesc: "启用 Tower 多智能体协同模式（/tower 命令切换）",
+  flagSubagentFork: "子代理 Fork 上下文",
+  flagSubagentForkDesc: "Agent/AgentSwarm 工具可携带主代理会话快照启动子代理（fork 参数）",
+  flagWaitFor: "WaitFor 工具",
+  flagWaitForDesc: "代理可在当前回合内等待后台任务完成，而无需结束回合再被唤醒",
+  flagAutoSessionTitle: "自动会话标题",
+  flagAutoSessionTitleDesc: "通过托管 chat_title 工具自动生成会话标题",
+  flagRemoteControl: "远程控制",
+  flagRemoteControlDesc: "启用实验性远程控制功能",
+  flagSearchWorker: "搜索 Worker",
+  flagSearchWorkerDesc: "kap-server 全局搜索的后台 worker 宿主服务",
+  flagFileHistory: "文件历史",
+  flagFileHistoryDesc:
+    "记录每轮编辑文件的前后快照（最近 30 个会话×最近 5 轮，存于会话数据目录），用于真实的整文件 diff",
+  flagDefaultOn: "默认开启",
  lockedByEnv: "被环境变量锁定",
-  masterEnvOnHint: "KIMI_CODE_EXPERIMENTAL_FLAG 已设置：全部实验功能被环境变量强制开启",
+  masterEnvOnHint:
+    "KIMI_CODE_EXPERIMENTAL_FLAG 已设置：未显式配置的实验项将被强制开启；config.toml 中的显式配置优先生效",
  secondaryModelSection: "子代理模型（次主力模型）",
  secondaryModelDisabledHint:
    "实验开关「子代理次主力模型」未启用，本配置不会生效——请先在上方打开对应开关",
@@ -511,6 +529,7 @@ export const zhTranslations = {
    "以独立窗口在应用内打开新版 Web 界面，也可以选择在系统浏览器打开（kimi-code 0.33+ 内置 code-app；需要 kimi 命令在 PATH 中）。",
  openWebUIEmbedded: "在应用内打开",
  openWebUIBrowser: "在浏览器打开",
+  webuiOpening: "打开中...",

  // Plugin marketplace
  pluginMarketplace: "插件市场",
@@ -0,0 +1,206 @@
+import { describe, expect, it } from "vitest";
+import { getAgentSettings, setAgentSettings } from "./agent-settings";
+
+// ---------------------------------------------------------------------------
+// Helpers — drill into the written `[thinking]` / `[loop_control]` sections
+// ---------------------------------------------------------------------------
+
+function thinkingOf(raw: unknown): Record<string, unknown> {
+  const t = (raw as { thinking?: unknown })?.thinking;
+  return t && typeof t === "object" && !Array.isArray(t)
+    ? (t as Record<string, unknown>)
+    : {};
+}
+
+function loopOf(raw: unknown): Record<string, unknown> {
+  const l = (raw as { loop_control?: unknown })?.loop_control;
+  return l && typeof l === "object" && !Array.isArray(l)
+    ? (l as Record<string, unknown>)
+    : {};
+}
+
+// ---------------------------------------------------------------------------
+// Reading — legacy off values normalized to "off", "max" → "high"
+// ---------------------------------------------------------------------------
+
+describe("getAgentSettings — thinking.keep normalization", () => {
+  it("normalizes a legacy boolean `false` to \"off\"", () => {
+    expect(getAgentSettings({ thinking: { keep: false } }).thinking?.keep).toBe(
+      "off"
+    );
+  });
+
+  it("normalizes the other legacy off values (0, null, \"no\", \"none\") to \"off\"", () => {
+    for (const v of [0, null, "no", "none"]) {
+      expect(
+        getAgentSettings({ thinking: { keep: v } }).thinking?.keep
+      ).toBe("off");
+    }
+  });
+
+  it("keeps \"all\" and \"off\" as-is on read", () => {
+    expect(getAgentSettings({ thinking: { keep: "all" } }).thinking?.keep).toBe(
+      "all"
+    );
+    expect(getAgentSettings({ thinking: { keep: "off" } }).thinking?.keep).toBe(
+      "off"
+    );
+  });
+
+  it("keeps the default \"all\" when the key is absent", () => {
+    expect(getAgentSettings({}).thinking?.keep).toBe("all");
+  });
+});
+
+describe("getAgentSettings — effort \"max\" read mapping", () => {
+  it("normalizes a stored \"max\" to \"high\" on read", () => {
+    expect(getAgentSettings({ thinking: { effort: "max" } }).thinking?.effort).toBe(
+      "high"
+    );
+  });
+});
+
+// ---------------------------------------------------------------------------
+// Writing — keep: "all" round-trips, "off" is written for legacy off values,
+// absent keep is not materialized
+// ---------------------------------------------------------------------------
+
+describe("setAgentSettings — keep serialization", () => {
+  it("round-trips keep = \"all\"", () => {
+    const raw = { thinking: { keep: "all" } };
+    const next = setAgentSettings(raw, { thinking: { keep: "all" } });
+    expect(thinkingOf(next).keep).toBe("all");
+  });
+
+  it("writes \"off\" (string) for a legacy keep = false, never a boolean", () => {
+    const next = setAgentSettings({ thinking: { keep: false } }, {});
+    expect(thinkingOf(next).keep).toBe("off");
+  });
+
+  it("does not write the keep key when it was never set (absent in raw and patch)", () => {
+    expect(thinkingOf(setAgentSettings({}, {}))).not.toHaveProperty("keep");
+    const next = setAgentSettings(
+      { thinking: { enabled: true } },
+      { thinking: { effort: "high" } }
+    );
+    expect(thinkingOf(next)).not.toHaveProperty("keep");
+    expect(thinkingOf(next).effort).toBe("high");
+  });
+
+  it("writes the key when an absent raw config's keep is explicitly toggled off", () => {
+    const next = setAgentSettings({}, { thinking: { keep: "off" } });
+    expect(thinkingOf(next).keep).toBe("off");
+  });
+});
+
+// ---------------------------------------------------------------------------
+// loop_control — v2 default writes only max_attempts_per_step; legacyV1 opts
+// dual-writes; read fallback from the v1 key is preserved
+// ---------------------------------------------------------------------------
+
+describe("loop_control — v1/v2 key handling", () => {
+  const raw = { loop_control: { max_attempts_per_step: 5, max_retries_per_step: 5 } };
+
+  it("default (v2): a save strips the legacy max_retries_per_step key", () => {
+    const next = setAgentSettings(raw, { loop_control: { reserved_context_size: 60000 } });
+    expect(loopOf(next).max_attempts_per_step).toBe(5);
+    expect(loopOf(next)).not.toHaveProperty("max_retries_per_step");
+  });
+
+  it("default (v2): a dual-write patch ends up with only the v2 key", () => {
+    const next = setAgentSettings(
+      raw,
+      { loop_control: { max_attempts_per_step: 7, max_retries_per_step: 7 } }
+    );
+    expect(loopOf(next).max_attempts_per_step).toBe(7);
+    expect(loopOf(next)).not.toHaveProperty("max_retries_per_step");
+  });
+
+  it("legacyV1: both keys are written", () => {
+    const next = setAgentSettings(
+      raw,
+      { loop_control: { max_attempts_per_step: 7, max_retries_per_step: 7 } },
+      { legacyV1: true }
+    );
+    expect(loopOf(next).max_attempts_per_step).toBe(7);
+    expect(loopOf(next).max_retries_per_step).toBe(7);
+  });
+
+  it("read fallback keeps a max_retries_per_step-only config editable", () => {
+    const s = getAgentSettings({ loop_control: { max_retries_per_step: 4 } });
+    expect(s.loop_control?.max_attempts_per_step).toBe(4);
+    expect(s.loop_control?.max_retries_per_step).toBe(4);
+  });
+});
+
+// ---------------------------------------------------------------------------
+// [permission] dangerous_command_guard (kimi-code 0.40.1) — absent key means
+// upstream default ON; an explicit false must survive even with no rules
+// ---------------------------------------------------------------------------
+
+describe("permission.dangerous_command_guard", () => {
+  it("reads as undefined when the key is absent (default on)", () => {
+    expect(
+      getAgentSettings({}).permission?.dangerous_command_guard
+    ).toBeUndefined();
+    expect(
+      getAgentSettings({ permission: { rules: [] } }).permission
+        ?.dangerous_command_guard
+    ).toBeUndefined();
+  });
+
+  it("reads an explicit true / false", () => {
+    expect(
+      getAgentSettings({ permission: { dangerous_command_guard: false } })
+        .permission?.dangerous_command_guard
+    ).toBe(false);
+    expect(
+      getAgentSettings({ permission: { dangerous_command_guard: true } })
+        .permission?.dangerous_command_guard
+    ).toBe(true);
+  });
+
+  it("writes guard=false even with no rules (section kept)", () => {
+    const next = setAgentSettings(
+      {},
+      { permission: { rules: [], dangerous_command_guard: false } }
+    ) as { permission?: Record<string, unknown> };
+    expect(next.permission).toEqual({ dangerous_command_guard: false });
+  });
+
+  it("writes guard together with rules", () => {
+    const rules = [{ decision: "allow" as const, pattern: "Read" }];
+    const next = setAgentSettings(
+      {},
+      { permission: { rules, dangerous_command_guard: true } }
+    ) as { permission?: Record<string, unknown> };
+    expect(next.permission).toEqual({
+      rules,
+      dangerous_command_guard: true,
+    });
+  });
+
+  it("a rules-only update carries the existing guard through", () => {
+    const raw = { permission: { dangerous_command_guard: false } };
+    const next = setAgentSettings(raw, {
+      permission: { rules: [{ decision: "deny" as const, pattern: "Bash" }] },
+    }) as { permission?: Record<string, unknown> };
+    expect(next.permission?.dangerous_command_guard).toBe(false);
+    expect(next.permission?.rules).toHaveLength(1);
+  });
+
+  it("omits an empty rules array; keeps an explicit guard on plain saves", () => {
+    const raw = { permission: { dangerous_command_guard: true } };
+    const off = setAgentSettings(raw, {
+      permission: { rules: [], dangerous_command_guard: false },
+    }) as { permission?: Record<string, unknown> };
+    expect(off.permission).toEqual({ dangerous_command_guard: false });
+    expect(off.permission).not.toHaveProperty("rules");
+    // An explicit true is materialized on plain saves (harmless, and it
+    // documents the state the UI toggle shows).
+    const on = setAgentSettings(raw, {}) as {
+      permission?: Record<string, unknown>;
+    };
+    expect(on.permission).toEqual({ dangerous_command_guard: true });
+  });
+});
@@ -49,20 +49,43 @@ export function getAgentSettings(rawOther: unknown): AgentSettings {
  ) {
    loop.max_attempts_per_step = sectionLoop.max_retries_per_step;
  }
+  const thinking = {
+    ...DEFAULT_SETTINGS.thinking,
+    ...getSection<AgentSettings["thinking"]>(rawOther, "thinking"),
+  };
+  // Upstream engines only accept string off values ("off"/"none"/"no"; see
+  // KEEP_OFF_VALUES) — a `false` from an old config fails the v2 validator
+  // and the v1 strict parse. Normalize every legacy off value to "off" for
+  // display; the serialization path (setAgentSettings) writes the string.
+  if (thinking.keep !== undefined && thinking.keep !== "all") {
+    thinking.keep = "off";
+  }
+  // The "max" effort tier was removed upstream (auto-migrates to "high");
+  // old configs still carrying it are shown as "high" (not rewritten on read).
+  if (thinking.effort === "max") {
+    thinking.effort = "high";
+  }
+  const sectionPermission = getSection<AgentSettings["permission"]>(
+    rawOther,
+    "permission"
+  );
  return {
-    thinking: {
-      ...DEFAULT_SETTINGS.thinking,
-      ...getSection<AgentSettings["thinking"]>(rawOther, "thinking"),
-    },
+    thinking,
    loop_control: loop,
    background: {
      ...DEFAULT_SETTINGS.background,
      ...getSection<AgentSettings["background"]>(rawOther, "background"),
    },
    permission: {
-      rules:
-        getSection<AgentSettings["permission"]>(rawOther, "permission")?.rules ??
-        [],
+      rules: sectionPermission?.rules ?? [],
+      // `dangerous_command_guard` stays undefined unless the config carries
+      // an explicit boolean — upstream defaults it to ON (kimi-code 0.40.1),
+      // so an absent key means "on". The env var
+      // KIMI_CODE_DANGEROUS_COMMAND_GUARD (literal "true"/"false" only)
+      // outranks this config value at kimi-code runtime.
+      ...(typeof sectionPermission?.dangerous_command_guard === "boolean"
+        ? { dangerous_command_guard: sectionPermission.dangerous_command_guard }
+        : {}),
    },
    hooks: getSection<AgentSettings["hooks"]>(rawOther, "hooks") ?? [],
  };
@@ -70,25 +93,58 @@ export function getAgentSettings(rawOther: unknown): AgentSettings {

 export function setAgentSettings(
  rawOther: unknown,
-  patch: Partial<AgentSettings>
+  patch: Partial<AgentSettings>,
+  opts?: { legacyV1?: boolean }
 ): unknown {
  const root = { ...asRecord(rawOther) };
  const current = getAgentSettings(rawOther);
+  // Explicit guard value wins over the carried-over one; undefined keeps the
+  // key absent (upstream default = on).
+  const guard =
+    patch.permission?.dangerous_command_guard ??
+    current.permission?.dangerous_command_guard;
+  const permission: NonNullable<AgentSettings["permission"]> = {
+    rules: patch.permission?.rules ?? current.permission?.rules ?? [],
+    ...(typeof guard === "boolean" ? { dangerous_command_guard: guard } : {}),
+  };
  const next: AgentSettings = {
    thinking: { ...current.thinking, ...patch.thinking },
    loop_control: { ...current.loop_control, ...patch.loop_control },
    background: { ...current.background, ...patch.background },
-    permission: {
-      rules: patch.permission?.rules ?? current.permission?.rules ?? [],
-    },
+    permission,
    hooks: patch.hooks ?? current.hooks ?? [],
  };

+  // `keep` is only written when the source config explicitly carries it or the
+  // patch sets it — an absent key keeps the engine default ("all") instead of
+  // being materialized, so a plain save does not add the key.
+  const rawThinking = asRecord(getSection(rawOther, "thinking"));
+  const patchHasKeep = patch.thinking !== undefined && "keep" in patch.thinking;
+  if (!("keep" in rawThinking) && !patchHasKeep) {
+    delete next.thinking?.keep;
+  }
+
+  // The v2 engine (default) only reads max_attempts_per_step; the legacy v1
+  // key is stripped on save unless KIMI_CODE_LEGACY_FLAG=1 (v1 engine compat).
+  if (!opts?.legacyV1) {
+    delete next.loop_control?.max_retries_per_step;
+  }
  if (next.thinking) root.thinking = next.thinking;
  if (next.loop_control) root.loop_control = next.loop_control;
  if (next.background) root.background = next.background;
-  if (next.permission?.rules && next.permission.rules.length > 0) {
-    root.permission = next.permission;
+  // Keep `[permission]` when it carries rules *or* an explicit
+  // dangerous_command_guard — dropping the section would silently lose a
+  // guard=false write (absent key = upstream default on). An empty rules
+  // array is omitted rather than written as `rules = []`.
+  const hasRules = !!permission.rules && permission.rules.length > 0;
+  const hasGuard = typeof permission.dangerous_command_guard === "boolean";
+  if (hasRules || hasGuard) {
+    root.permission = {
+      ...(hasRules ? { rules: permission.rules } : {}),
+      ...(hasGuard
+        ? { dangerous_command_guard: permission.dangerous_command_guard }
+        : {}),
+    };
  } else {
    delete root.permission;
  }
@@ -1,8 +1,12 @@
 import { describe, expect, it } from "vitest";
 import {
+  EXPERIMENTAL_FLAGS,
+  getExperimentalFlags,
  getSecondaryModel,
  getSubagentModelPool,
+  isExperimentalFlagSet,
  removeSubagentPoolEntry,
+  setExperimentalFlag,
  setSecondaryModelOnly,
  setSubagentDefault,
  setSubagentForce,
@@ -231,6 +235,31 @@ describe("pool mutations", () => {
    expect(section(updated).default_model).toBe("kimi-k1");
  });

+  it("upsertSubagentPoolEntry materializes the implicit default into a new table", () => {
+    // Implicit single-entry form (default_model set, no models table): adding a
+    // second alias must pull the default into the table, otherwise the result
+    // violates "default_model must be a pool key" (rule 3) and gets rejected.
+    const raw = rawWith({ default_model: "kimi-k1", force: false });
+    const added = upsertSubagentPoolEntry(raw, "kimi-k2", "");
+    expect(modelsOf(added)).toEqual({ "kimi-k1": "", "kimi-k2": "" });
+    expect(section(added).default_model).toBe("kimi-k1");
+    const pool = getSubagentModelPool(added);
+    expect(pool).toBeDefined();
+    expect(
+      validateSubagentPool(pool!, ["kimi-k1", "kimi-k2"]),
+    ).toEqual([]);
+  });
+
+  it("upsertSubagentPoolEntry upgrades a legacy `model`-only section to a valid pool", () => {
+    const raw = rawWith({ model: "kimi-k1" });
+    const added = upsertSubagentPoolEntry(raw, "kimi-k2", "K2");
+    expect(modelsOf(added)).toEqual({ "kimi-k1": "", "kimi-k2": "K2" });
+    expect(section(added).default_model).toBe("kimi-k1");
+    expect(section(added).model).toBe("kimi-k1");
+    const pool = getSubagentModelPool(added);
+    expect(validateSubagentPool(pool!, ["kimi-k1", "kimi-k2"])).toEqual([]);
+  });
+
  it("removeSubagentPoolEntry keeps the default when removing a non-default alias", () => {
    const raw = rawWith({
      default_model: "kimi-k1",
@@ -424,3 +453,76 @@ describe("validateSubagentPool", () => {
    expect(errors).toContainEqual({ key: "forceExcludesModels" });
  });
 });
+
+// ---------------------------------------------------------------------------
+// [experimental] — flag registry (kimi-code 0.40.1) + write semantics
+// ---------------------------------------------------------------------------
+
+describe("experimental flag registry", () => {
+  it("mirrors the 0.40.1 v2 registry (10 flags incl. file_history)", () => {
+    expect(EXPERIMENTAL_FLAGS.map((f) => f.id)).toEqual([
+      "secondary-model",
+      "tool-select",
+      "persistence_minidb_readmodel",
+      "tower",
+      "subagent_fork",
+      "wait_for",
+      "auto_session_title",
+      "remote-control",
+      "search_worker",
+      "file_history",
+    ]);
+    const fh = EXPERIMENTAL_FLAGS.find((f) => f.id === "file_history");
+    expect(fh?.envVar).toBe("KIMI_CODE_EXPERIMENTAL_FILE_HISTORY");
+    expect(fh?.defaultEnabled).toBeUndefined();
+  });
+
+  it("secondary-model is on by default since 0.40.1", () => {
+    expect(
+      EXPERIMENTAL_FLAGS.find((f) => f.id === "secondary-model")?.defaultEnabled
+    ).toBe(true);
+  });
+});
+
+describe("setExperimentalFlag", () => {
+  const rawExp = (section: Record<string, unknown>) => ({
+    experimental: section,
+  });
+
+  it("enabling writes true; disabling a default-off flag deletes the key", () => {
+    const on = setExperimentalFlag({}, "tool-select", true) as {
+      experimental: Record<string, unknown>;
+    };
+    expect(on.experimental["tool-select"]).toBe(true);
+    const off = setExperimentalFlag(on, "tool-select", false) as Record<
+      string,
+      unknown
+    >;
+    expect(off).not.toHaveProperty("experimental"); // empty section dropped
+  });
+
+  it("disabling a default-on flag with explicitFalse writes a literal false", () => {
+    const next = setExperimentalFlag(
+      rawExp({ wait_for: true }),
+      "wait_for",
+      false,
+      { explicitFalse: true }
+    ) as { experimental: Record<string, unknown> };
+    expect(next.experimental.wait_for).toBe(false);
+    expect(isExperimentalFlagSet(next, "wait_for")).toBe(true);
+    expect(getExperimentalFlags(next).wait_for).toBe(false);
+  });
+
+  it("an explicit false coexists with other flags and keeps the section", () => {
+    const next = setExperimentalFlag(
+      rawExp({ "secondary-model": false }),
+      "tower",
+      true,
+      { explicitFalse: true }
+    ) as { experimental: Record<string, unknown> };
+    expect(next.experimental).toEqual({
+      "secondary-model": false,
+      tower: true,
+    });
+  });
+});
@@ -28,22 +28,57 @@ export interface ExperimentalFlagDef {
  id: string;
  /** Single-feature env var that can force the flag (locks the UI toggle). */
  envVar: string;
+  /** Feature turned on by default upstream; an absent config key falls back
+   *  to this, while an explicit `false` in `[experimental]` still wins. */
+  defaultEnabled?: boolean;
 }

-/** Known experimental flags (Kimi Code v1 + v2 registries, merged). */
+/** Known experimental flags — mirrors the kimi-code v2 flag registry
+ * (the per-feature flag.ts files under packages/agent-core-v2/src:
+ * secondary-model, tool-select, persistence_minidb_readmodel, tower,
+ * subagent_fork, wait_for, auto_session_title, remote-control,
+ * search_worker, file_history). `acp-v2` was removed upstream and is
+ * dropped here. Keep the env-var list in sync with
+ * `get_experimental_env_status` in src-tauri/src/commands.rs. */
 export const EXPERIMENTAL_FLAGS: ExperimentalFlagDef[] = [
-  { id: "secondary-model", envVar: "KIMI_CODE_EXPERIMENTAL_SECONDARY_MODEL" },
+  {
+    id: "secondary-model",
+    envVar: "KIMI_CODE_EXPERIMENTAL_SECONDARY_MODEL",
+    defaultEnabled: true, // on by default since kimi-code 0.40.1
+  },
  { id: "tool-select", envVar: "KIMI_CODE_EXPERIMENTAL_TOOL_SELECT" },
-  { id: "acp-v2", envVar: "KIMI_CODE_EXPERIMENTAL_ACP_V2" },
  {
    id: "persistence_minidb_readmodel",
    envVar: "KIMI_CODE_EXPERIMENTAL_PERSISTENCE_MINIDB_READMODEL",
+    defaultEnabled: true,
  },
+  { id: "tower", envVar: "KIMI_CODE_EXPERIMENTAL_TOWER" },
+  { id: "subagent_fork", envVar: "KIMI_CODE_EXPERIMENTAL_SUBAGENT_FORK" },
+  {
+    id: "wait_for",
+    envVar: "KIMI_CODE_EXPERIMENTAL_WAIT_FOR",
+    defaultEnabled: true,
+  },
+  {
+    id: "auto_session_title",
+    envVar: "KIMI_CODE_EXPERIMENTAL_AUTO_SESSION_TITLE",
+  },
+  { id: "remote-control", envVar: "KIMI_CODE_EXPERIMENTAL_REMOTE_CONTROL" },
+  {
+    id: "search_worker",
+    envVar: "KIMI_CODE_EXPERIMENTAL_SEARCH_WORKER",
+    defaultEnabled: true,
+  },
+  { id: "file_history", envVar: "KIMI_CODE_EXPERIMENTAL_FILE_HISTORY" },
 ];

 export const EXPERIMENTAL_MASTER_ENV = "KIMI_CODE_EXPERIMENTAL_FLAG";

-/** Whether the master env var is set to a truthy value (locks every flag on). */
+/** Whether the master env var is set to a truthy value. Since kimi-code
+ *  0.40.1 it only force-ONs flags that have *no* explicit `[experimental]`
+ *  entry (an explicit config value — true or false — outranks it, and the
+ *  master env can never force a flag off). It locks no UI toggles; only a
+ *  flag's own single-feature env var does. */
 export function isMasterEnvOn(env: ExperimentalEnvStatus | null): boolean {
  return isTruthyEnv(env?.[EXPERIMENTAL_MASTER_ENV]);
 }
@@ -77,16 +112,33 @@ export function getExperimentalFlags(rawOther: unknown): Record<string, boolean>
  return out;
 }

-/** Set one flag in `[experimental]`. Removing all flags drops the section. */
+/** Whether a flag is explicitly written in `[experimental]` (true *or* false).
+ *  An absent flag falls back to the upstream default (`defaultEnabled`). */
+export function isExperimentalFlagSet(rawOther: unknown, id: string): boolean {
+  return Object.prototype.hasOwnProperty.call(
+    asRecord(asRecord(rawOther)["experimental"]),
+    id
+  );
+}
+
+/** Set one flag in `[experimental]`. Removing all flags drops the section.
+ *  Pass `explicitFalse` when deleting the key would NOT express "off" —
+ *  i.e. the flag is on by upstream default (`defaultEnabled`) or the master
+ *  env force-ons undefined flags — and write a literal `false` instead,
+ *  which outranks both under the kimi-code 0.40.1 priority
+ *  (single env > explicit config > master env > default). */
 export function setExperimentalFlag(
  rawOther: unknown,
  id: string,
-  enabled: boolean
+  enabled: boolean,
+  opts?: { explicitFalse?: boolean }
 ): unknown {
  const root = { ...asRecord(rawOther) };
  const flags = { ...getExperimentalFlags(rawOther) };
  if (enabled) {
    flags[id] = true;
+  } else if (opts?.explicitFalse) {
+    flags[id] = false;
  } else {
    delete flags[id];
  }
@@ -246,8 +298,11 @@ export function setSubagentModelPool(
  return root;
 }

-/** Add or update one alias in the pool `models` table. The rest of the
- *  section (default, force, unknown fields) is preserved untouched. */
+/** Add or update one alias in the pool `models` table. When the section is
+ *  the implicit single-entry form (a `default_model`/`model` but no table),
+ *  the effective default is materialized into the table first — otherwise the
+ *  new table would fail the engine rule "default_model must be a pool key".
+ *  The rest of the section (force, unknown fields) is preserved untouched. */
 export function upsertSubagentPoolEntry(
  rawOther: unknown,
  alias: string,
@@ -259,8 +314,21 @@ export function upsertSubagentPoolEntry(
  for (const [k, v] of Object.entries(section)) next[k] = v;
  const models: Record<string, unknown> = {};
  for (const [k, v] of Object.entries(asRecord(next.models))) models[k] = v;
+  const effectiveDefault =
+    typeof next.default_model === "string" && next.default_model !== ""
+      ? next.default_model
+      : typeof next.model === "string" && next.model !== ""
+        ? next.model
+        : undefined;
+  if (effectiveDefault !== undefined && !(effectiveDefault in models)) {
+    models[effectiveDefault] = "";
+  }
  models[alias] = description;
  next.models = models;
+  if (effectiveDefault !== undefined) {
+    next.default_model = effectiveDefault;
+    next.model = effectiveDefault;
+  }
  root.secondary_model = next;
  return root;
 }
@@ -6,7 +6,9 @@ export type ProviderType =
  | "openai"
  | "openai_responses"
  | "google-genai"
-  | "vertexai";
+  | "vertexai"
+  /** Any other string the CLI writes — preserved verbatim on round-trip. */
+  | (string & {});

 export interface Provider {
  name: string;
@@ -62,6 +64,12 @@ export interface Config {
  providers: Record<string, Provider>;
  models: Record<string, Model>;
  raw_other?: unknown;
+  /**
+   * Top-level section keys captured at import time by the Rust side.
+   * Export drops only baseline keys absent from raw_other (sections the
+   * user removed in the UI); CLI-added sections are preserved.
+   */
+  imported_section_keys?: string[];
 }

 /**
@@ -92,7 +100,18 @@ export interface DiscoveredModel {

 export interface ThinkingConfig {
  enabled?: boolean;
+  /**
+   * Effort tier. `max` is read-compatible only — upstream removed the tier
+   * (old configs auto-migrate to `high`); the UI normalizes it and the
+   * serialization path never writes it.
+   */
  effort?: "low" | "medium" | "high" | "max";
+  /**
+   * Keep thinking content. The legacy off values (`false`, `0`, "no", "none",
+   * `null`) are read-compatible — old configs may carry them; the UI
+   * normalizes them to `"off"` and the serialization path never emits a
+   * boolean (upstream engines only accept string off values).
+   */
  keep?: "all" | false | 0 | "no" | "off" | "none" | null;
 }

@@ -101,7 +120,8 @@ export interface LoopControlConfig {
  max_attempts_per_step?: number;
  /** v2 engine key (kimi-code 0.33+): `max_steps_per_run` was renamed. */
  max_steps_per_turn?: number;
-  /** Legacy v1 key — written in sync so KIMI_CODE_LEGACY_FLAG=1 still works. */
+  /** Legacy v1 key — read fallback only; written/kept only when
+   *  KIMI_CODE_LEGACY_FLAG=1 (v1 engine compat). v2 saves strip it. */
  max_retries_per_step?: number;
  reserved_context_size?: number;
 }
@@ -129,7 +149,15 @@ export interface AgentSettings {
  thinking?: ThinkingConfig;
  loop_control?: LoopControlConfig;
  background?: BackgroundConfig;
-  permission?: { rules?: PermissionRule[] };
+  permission?: {
+    rules?: PermissionRule[];
+    /** kimi-code 0.40.1 `[permission]` key. Default true when the key is
+     *  absent: Auto mode refuses dangerous commands (rm -rf, shutdown,
+     *  dd of=, …) and other modes always prompt; false restores the
+     *  previous behavior. Env KIMI_CODE_DANGEROUS_COMMAND_GUARD (literal
+     *  "true"/"false" only) outranks this config. */
+    dangerous_command_guard?: boolean;
+  };
  hooks?: Hook[];
 }
Author	SHA1	Message	Date
KimiSwitch Dev	5138a295e0	chore(release): bump to v0.7.12 — 适配 kimi-code 0.40.1（flag 清单/优先级/危险命令守卫）+ WebUI 竞态修复 + models-dev 7495/212 Release / Version consistency (push) Canceled after 0s Details Release / Build (macos-latest) (push) Canceled after 0s Details Release / Build (ubuntu-latest) (push) Canceled after 0s Details Release / Build (windows-latest) (push) Canceled after 0s Details Release / Attach macOS install script (push) Canceled after 0s Details	2026-09-03 10:12:17 +08:00
KimiSwitch Dev	a02179c9da	chore(data): 同步 models-dev 最新快照（7495 模型 / 212 供应商）	2026-09-03 10:11:10 +08:00
KimiSwitch Dev	86a128efea	feat: 适配 kimi-code 0.40.1 —— flag 清单/优先级语义/危险命令守卫 - flag 清单 9→10：新增 file_history、search_worker；secondary-model 0.40.1 起默认开启（defaultEnabled） - env 探测补齐为 12 项元组（10 flag + master + legacy），修正此前 tower/subagent_fork/wait_for/auto_session_title 四项漏探 - 优先级语义跟随上游 flagService：单 flag env > [experimental] 显式值 > master env（仅强制开）> 默认。master env 不再锁定 UI；setExperimentalFlag 增加 explicitFalse 路径，关闭默认开/master 强开的 flag 时写字面 false - 权限设置新增 dangerous_command_guard 开关（默认开，写 [permission] 蛇形键；env KIMI_CODE_DANGEROUS_COMMAND_GUARD 运行时覆盖 config） - 新增 11 个测试（注册表/写语义/守卫读写）	2026-09-03 10:10:52 +08:00
KimiSwitch Dev	6a7c8f5393	fix(webui): 修复首次打开慢时多次点击导致的竞态报错 open_kimi_web_embedded 加 launch_lock 串行化并发调用，拿锁后重查单例； oneshot 在锁内等待主线程建窗完成，封死命令返回后的残留空档。 spawned_by_me 门控失败路径的 server 清理，复用的 server 不再被误杀； spawn 前用 try_wait 检查已有子进程避免覆盖泄漏。前端两个打开按钮加 busy 态防重入（打开中...）。	2026-09-03 10:10:40 +08:00
KimiSwitch Dev	8b597496b3	chore(release): bump to v0.7.11 — 适配 kimi-code 0.39.1（thinking keep/loop_control/flags）+ 归档元数据化 + 子代理池修复 + models-dev 7482/207 Release / Version consistency (push) Canceled after 0s Details Release / Build (macos-latest) (push) Canceled after 0s Details Release / Build (ubuntu-latest) (push) Canceled after 0s Details Release / Build (windows-latest) (push) Canceled after 0s Details Release / Attach macOS install script (push) Canceled after 0s Details	2026-08-29 11:15:40 +08:00
KimiSwitch Dev	8ce3f290ed	chore(data): 同步 models-dev 最新快照（7482 模型 / 207 供应商）	2026-08-29 11:14:11 +08:00
KimiSwitch Dev	8639188968	fix(sessions): 归档迁移为 state.json 元数据（archived/archivedAt，与上游 v2 setArchived 逐字段一致），不再物理搬目录；遗留 .kcd-archive 目录兼容识别与恢复；归档不再动 session_index	2026-08-29 11:14:01 +08:00
KimiSwitch Dev	45ec073e6f	fix(config): 适配 kimi-code 0.39.1——thinking keep 写字符串 off（布尔会整节作废）、存量布尔归一化；loop_control 默认只写 v2 键（legacy flag 才双写）；effort 移除 max 档；实验 flags 补 remote-control、wait_for/minidb 标默认开启	2026-08-29 11:13:58 +08:00
KimiSwitch Dev	7435198fcc	chore(release): bump to v0.7.10 — 配置写回保护（导入基线 + 未知供应商类型原样保留） Release / Version consistency (push) Canceled after 0s Details Release / Attach macOS install script (push) Canceled after 0s Details Release / Build (macos-latest) (push) Canceled after 0s Details Release / Build (ubuntu-latest) (push) Canceled after 0s Details Release / Build (windows-latest) (push) Canceled after 0s Details	2026-08-29 00:10:52 +08:00
KimiSwitch Dev	67db19eda4	fix(config): 保存不再误删 CLI 后加的顶层配置节 + 未知供应商类型原样保留 - 导出 stale-key 清理增加导入基线（imported_section_keys）：CLI 在导入后新增的顶层节（kimi-code 0.38+ 的 [task]/[swarm]/[cron] 等）在保存时原样保留；用户在 UI 显式删除的节仍被移除；空基线（SQLite 快照恢复）不删任何键（安全退化） - ProviderType 增加 Unknown(String) 变体，手写 serde 保持纯字符串线格式，上游新增的 type 值往返不再被改写为 kimi；SQLite 快照恢复路径同步修复；模型发现对未知类型返回明确错误，不再误用 KIMI_API_KEY - 新增 3 个往返测试（106 个 Rust 测试全绿），前端 tsc/vitest 通过	2026-08-29 00:09:55 +08:00
KimiSwitch Dev	9764929919	chore(release): bump to v0.7.9 — models.dev 数据同步（GLM-5.3/5.3-Flash、Qwen3.8、豆包 Seed 2.x） Release / Version consistency (push) Canceled after 0s Details Release / Build (macos-latest) (push) Canceled after 0s Details Release / Build (ubuntu-latest) (push) Canceled after 0s Details Release / Build (windows-latest) (push) Canceled after 0s Details Release / Attach macOS install script (push) Canceled after 0s Details	2026-08-27 09:42:14 +08:00
KimiSwitch Dev	f40e4be577	chore(data): 同步 models-dev 2026-08-27 快照（7343 模型 / 203 供应商）— 新增 GLM-5.3/5.3-Flash、Qwen3.8、豆包 Seed 2.x 等 74 模型，58 处价格与 56 处能力更新；补齐 public/ 静态资源（原停留在 08-01）	2026-08-27 09:37:26 +08:00
KimiSwitch Dev	138a2b8547	chore(data): 同步 models-dev 最新模型数据（7284 模型 / 199 供应商）	2026-08-25 13:12:11 +08:00
KimiSwitch Dev	769852c4b5	feat(flags): v0.7.8 — 实验开关同步 kimi-code 上游 7 旗标（tower/subagent_fork/wait_for/auto_session_title 新增，acp-v2 移除）+ 修复模型用量上下显示不同步（状态提升至卡片层） Release / Version consistency (push) Canceled after 0s Details Release / Build (macos-latest) (push) Canceled after 0s Details Release / Build (ubuntu-latest) (push) Canceled after 0s Details Release / Build (windows-latest) (push) Canceled after 0s Details Release / Attach macOS install script (push) Canceled after 0s Details	2026-08-25 11:20:34 +08:00