feat(thinking): 思考等级扩展五档（Max/XHigh）+ 按默认模型 support_efforts 能力感知；删除过时的 max→high 迁移；Segmented 支持逐选项禁用/tooltip — bump to v0.7.23

fix(release): publish-gitea.sh 兼容 docs/release-notes/release-notes-<tag>.md 命名（此前 Release 正文回退为 tag 名）
chore(data): 同步 models-dev 快照至 2026-10-02（8,359 模型 / 225 供应商）
2026-10-02 13:00:35 +08:00 · 2026-10-02 13:00:33 +08:00 · 2026-10-02 13:00:32 +08:00 · 2026-09-29 10:42:13 +08:00 · 2026-09-24 00:19:10 +08:00 · 2026-09-20 10:29:40 +08:00
52 changed files with 291043 additions and 263642 deletions
@@ -73,7 +73,8 @@ jobs:
        id: notes
        shell: bash
        run: |
-          FILE="release-notes-${{ github.ref_name }}.md"
+          FILE="docs/release-notes/${{ github.ref_name }}.md"
+          [ -f "$FILE" ] || FILE="release-notes-${{ github.ref_name }}.md"
          {
            echo 'body<<RELEASE_BODY_EOF'
            if [ -f "$FILE" ]; then
@@ -0,0 +1,156 @@
+# Changelog
+
+**Kimi Switch** — Windows 桌面端的 Kimi Code CLI 配置管理器，统一管理多家 LLM 供应商、模型、图标、连通性、用量统计与版本更新。
+
+本文件汇总各版本要点，按版本倒序排列；每个版本的**完整 release notes**（含安装说明、平台产物、已知问题）见 [`docs/release-notes/`](./docs/release-notes/)。
+
+---
+
+## v0.7.23 (2026-09-30)
+
+- **思考等级五档 + 能力感知**：全局配置的思考等级扩展为低 / 中 / 高 / Max / XHigh，按当前默认模型声明的 `support_efforts` 自动置灰不支持档位并提示原因；models.dev 标记不支持思考的模型整体禁用思考区；未声明档位的模型保留可选并提示"上游可能拒绝"
+- **修正过时适配**：删除 max→high 读取迁移（kimi-code 2.1.1 中 max 为合法档位），存储的 max 与任意手写档位值原样保留显示；子代理页继承档位显示同步修正
+
+## v0.7.22 (2026-09-24)
+
+- **修复 Sub2API / NewAPI 模板查询不显示**：为无预设识别类型的自定义供应商配置 Sub2API（或 NewAPI）模板后，弹窗测试可查但供应商列表始终不显示——前端列表的查询/显示门槛只看预设 `usageKinds`，显式模板被漏掉；现在与后端口径对齐（模板查询绕过 usage_kinds），列表正常显示余额 / 配额
+
+## v0.7.21 (2026-09-24)
+
+- **models.dev 在线同步**：高级设置新增「模型参考数据」卡片——一键在线拉取 models.dev 最新快照（上下文长度 / 能力标记 / 单价），前端与仪表盘计价即时生效，无需等待新版本；支持一键恢复随版本内置数据；显示当前数据来源与版本（同步 / 内置 + 日期 + 模型/供应商数）
+- **修复构建链路数据漂移**：`fetch-models-dev.mjs` 现在同时写入 `src/lib/models-dev.json`（Rust 计价内嵌）与 `public/models-dev.json`（前端静态资源）——此前 public 副本需手工拷贝，已两次滞后于 src/lib
+- **kimi-code 2.1.0 已适配确认**：实验 flag 注册表（5 个）、config 模式、用量/quota 格式均无变化；上游 `[watch]` 默认关闭与现有开关行为一致；新增 `tui_mode`（tui.toml）不涉及配置管理面
+- models-dev 快照同步至 2026-09-23（8,126 模型 / 223 供应商）
+
+## v0.7.20 (2026-09-20)
+
+- **套餐用量阈值预警**：用量配置新增「预警阈值 %」（0-100，0/空为关闭），任一用量窗口达到阈值即红色高亮 + ⚠ 标记，低于阈值自动消除
+- **Kimi For Coding 加力钱包**：0.43.1 起解析 `boosterWallet` 并展示为独立余额行（💰 月度用量 / 总额 / 余额，含币种）
+- **供应商凭证环境变量化（`api_key_env`）**：适配 kimi-code 2.0.0 #3762，凭证来源可切换「直接填写 / 环境变量」，密钥不写入 `config.toml`，与 `api_key` 互斥；SQLite 新增列含自动迁移
+- **Kimi Code 版本跟踪器**：设置页新增版本卡片——本机已装版本、上游最新版本、适配状态徽标（绿已适配 / 黄建议升级 / 橙部分兼容 / 灰未安装），可一键打开上游 Releases
+- **供应商一键体检**：并发探活所有启用供应商的推理端点（`GET /v1/models`）并顺带重查账单，三态判定避免误报（端点不支持模型列表计中性）
+- 新增 `loop_control.compaction_max_attempts`（上游 0.43.0 #3750，默认 5）与 `[watch] enabled` 开关（上游 2.0.1 #3892）；已确认适配 kimi-code 2.0.2
+
+## v0.7.19 (2026-09-14)
+
+- **适配 kimi-code 0.43.0 实验 flag 镜像（7→5）**：移除已转正的 `auto_session_title`（上游 #3749），上游注册表现剩 `wait_for` / `tool-select` / `notify_user` / `tower` / `subagent_fork`
+- 兼容性核查：0.43.0 的 wire 持久化重建（#3737）不影响用量统计，相关记录全部保留；`loop_control.compaction_max_attempts`（#3750）保存不丢失
+- 残留的 `auto_session_title` 旧键会被上游静默忽略，无需手动清理
+
+## v0.7.18 (2026-09-14)
+
+- **会话管理「一键归档」**：支持按 1 个月前 / 半个月前 / 1 星期前 / 指定日期批量归档，替代逐个选择，并反馈已归档 / 跳过数量
+- **归档用量快照落库**：归档瞬间按「天 × 模型」粒度把 token 用量与费用写入 SQLite `archived_sessions` 表
+- 删除已归档会话释放磁盘后不再损失历史统计——趋势图、热力图、模型统计、区间合计自动合并快照
+- 快照与现存文件实时统计严格去重（只合并文件已删除的归档会话，绝不双计）
+- models-dev 快照同步至 2026-09-14（7784 模型 / 213 供应商），pricing 断言跟随 DeepSeek 官方调价（deepseek-v4-flash input 0.14→0.15 / output 0.28→0.60）
+
+## v0.7.17 (2026-09-13)
+
+- **账单查询适配 Sub2API 中转站（`balance:sub2api`）**：支持 Wei-Shaw/sub2api 面板余额查询（`GET {base}/v1/usage`），自动复用推理 `sk-` API Key，无需网页后台 Access Token
+- 解析三种模式：单 Key 总额度（quota）、5 小时 / 每日 / 7 天速率窗口（含重置倒计时）、订阅组日/周/月额度与钱包余额（USD）
+- `detect_provider` 家族规则拆分：`codingplan.site` 主域识别为 Sub2API，`ai.codingplan.site` 识别为 NewAPI，修正主域被误判导致查询 404 的问题
+- 用量配置面板新增「Sub2API 中转站」模板，可选覆盖查询地址（Base URL）
+- **模型映射「获取模型列表」新增批量选择**：全选 / 取消全选 / 反选，配合已添加标记批量接入中转站模型
+
+## v0.7.16 (2026-09-10)
+
+- **适配 kimi-code 0.42.0 实验 flag 镜像（9→6）**：移除 4 个已转正/删除的开关（子代理次主力模型、MiniDB 读模型、远程控制、搜索 Worker），新增 `notify_user`（实验性 Updates 面板，默认关闭）
+- 用量统计「模型用量统计」图表固定最大高度 360px 并支持滚动，与右侧模型用量表一致
+- models-dev 快照同步至 2026-09-10（7615 模型 / 213 供应商），`public/` 静态目录同步
+
+## v0.7.15 (2026-09-09)
+
+- **用量统计新增「昨天」选项卡**：时间范围为 今天 / 昨天 / 7 天 / 30 天 / 全部，「昨天」按本地时区严格统计昨日 0 点至今日 0 点，概览卡扩展为 5 张
+- **新增「模型用量统计」趋势选项卡**：跨供应商按模型名汇总（同一模型合并为一行），横向条形图按 Token 降序，悬停显示请求数 / Token / 缓存命中 / 费用
+- **修复 opencode Go 套餐子代理请求 400**：Go 网关强制要求 `x-opencode-session` 头，导出时凡指向 Go 端点的供应商自动补写（已有自定义值保留），按量计费的 `zen/v1` 不受影响
+- models-dev 快照同步至 2026-09-09（7583 模型 / 213 供应商）；`kimi/k2.5` 官方条目下架后暂解析到 302ai 转售价（0.66 / 3.3）
+
+## v0.7.14 (2026-09-06)
+
+- **修复关闭「启用用量查询」后仍残留查询行为**：此前关闭后前端仍发起查询并持续显示「查询失败 · 用量查询已关闭」红色错误行，现在关闭即停止查询并隐藏整行用量信息，重开后自动恢复
+- 在途查询结果在开关切换瞬间作废且不再写入共享缓存，避免过期失败状态在 5 分钟内复活
+
+## v0.7.13 (2026-09-05)
+
+- **适配 kimi-code 0.41.0**：上游移除 `file_history` flag（轮级文件历史转为恒开），镜像全链路同步删除；Auto（永不询问）模式不再拦截危险命令，设置页文案（中英）同步修正
+- **账单查询改进**：`codingplan.site` 系供应商自动识别为 NewAPI 余额查询；凭据缺失不再误报「网络异常」而给出本地化配置提示；金额显示统一为两位小数并去掉 `¤` 占位符
+- 新增 usage 错误本地化单元测试
+
+## v0.7.12 (2026-09-03)
+
+- **适配 kimi-code 0.40.1（flag 9→10）**：新增 `file_history` / `search_worker`；`secondary-model` 默认开启并显示徽章；新增危险命令守卫开关 `[permission] dangerous_command_guard`（默认开启）
+- **flag 优先级语义修正**：单 flag 环境变量 > `[experimental]` 显式配置 > 总开关环境变量 > 默认值，总开关不再锁定全部开关
+- **修复 WebUI 打开竞态**：并发调用串行化 + 后到点击自动聚焦已有窗口，失败路径不再误杀正在使用的 server，前端增加「打开中...」防重入态
+- models-dev 快照更新至 7495 模型 / 212 供应商
+
+## v0.7.11 (2026-08-29)
+
+- **适配 kimi-code 0.39.1**：`[thinking]` 保留开关改写 `keep = "off"`（修复布尔值导致整个节被 v2 校验丢弃）；`loop_control` 默认只写新键 `max_attempts_per_step` 并清理旧键；思考强度移除 `max` 档（上游迁移为 `high`）
+- **会话归档与上游 v2 对齐**：不再搬动文件，改为在 `state.json` 写入 `archived` / `archivedAt` 元数据，CLI 侧与 Kimi Switch 侧归档状态互通；旧 `.kcd-archive/` 物理归档仍可识别恢复
+- **修复子代理模型池无法添加第二个条目**：默认模型自动物化进池，不再被校验拦截
+- models.dev 快照刷新至 7482 个模型 / 207 个供应商
+
+## v0.7.10 (2026-08-29)
+
+- **修复保存配置时误删 CLI 新增的顶层配置节**：以导入时顶层键为基线，CLI 后加的 `[task]` / `[swarm]` / `[cron]` / `[tools]` / `[identity]` / `[token_counting]` 等节原样保留，界面中显式删除的节仍会移除
+- **修复未知供应商类型被改写**：`config.toml` 中新增的供应商 `type` 往返读写后原样保留（此前会被改成 `kimi`），下拉框标注「未知类型」，SQLite 快照恢复路径同步修复
+- 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
+
+## v0.7.9 (2026-08-27)
+
+- 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（每百万 token）
+- 新增 Qwen3.8 系列、豆包 Seed 2.x 全系、DeepSeek-V4 GA、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
+
+## v0.7.8 (2026-08-25)
+
+- **修复模型用量上下显示不同步**：上方用量摘要刷新后下方进度条仍停留旧数据（如 24% vs 8%），用量查询状态提升至卡片层统一持有，两处同帧更新
+- **实验功能开关同步 kimi-code 上游 7 项旗标**：新增 Tower 模式、子代理 Fork 上下文、WaitFor 工具、自动会话标题；移除已废弃的 ACP v2 开关（手写配置仍完整保留）
+- 自动刷新间隔改为每卡片恰好一份定时器，行为不变
+
+## v0.7.7 (2026-08-22)
+
+- **双区域 OAuth 支持（适配 kimi-code 0.38.0）**：适配 mainland-cn `auth.kimi.com` / global `auth.kimi.ai`（#2862），按 provider 的 oauth ref 推导凭据文件与刷新端点，global 账号（`credentials/kimi-code-env-<sha256>.json`）可正常查询用量
+- **内置登录区域选择**：应用内 Kimi 登录对话框新增「中国大陆 / 国际版 (kimi.ai)」，登录成功后按官方 CLI 行为 provision `[providers."managed:kimi-code"]`（global 写 `oauthHost`，cn 不写）
+- **修复**：手动添加模型后别名自动跟随为 `<provider>/<model-id>`；managed（OAuth）供应商自填 `api_key` 保存时不再被清空
+- models.dev 快照刷新至 **7246 模型 / 193 供应商**，新增 DeepSeek V4 Flash Vision Exp、Ox Alpha Free（opencode-go，免费）
+
+## v0.7.6 (2026-08-18)
+
+- **新增 OpenCode Go 套餐用量查询**：卡片用量页脚显示 **5 小时滚动 / 7 天 / 30 天** 三窗口的已用百分比与重置时间（数据源 `GET https://opencode.ai/zen/go/v1/usage`）
+- 已配置 OpenCode Go API Key 的老用户无需手动设置，程序按 `base_url` 自动识别并启用；已适配 opencode.ai 的 Cloudflare 拦截（携带浏览器 User-Agent）
+- 插件目录被运行中进程占用时的错误提示给出明确指引
+
+## v0.7.5 (2026-08-15)
+
+- **子代理模型池适配 kimi-code 0.36.0 新引擎**：适配 `[secondary_model]` 的 `default_model` + `[secondary_model.models]` 表 + `force`；写入时 `model` 与 `default_model` 双写同值，v1 legacy 与 v2 引擎均可识别
+- **高级设置页新增模型池管理**：升级旧配置、增删池条目、编辑路由描述（渲染进主代理 Agent/AgentSwarm 工具描述）；支持 force 强制默认模型（带二次确认）；保存前 6 类错误前置拦截
+- 新增 32 个单元测试覆盖池读写与校验逻辑（vitest）
+- **热力图详情改为按需加载**：双击方块调用后端 `get_day_detail` 聚合，不再受 7d / 30d 范围限制
+- models.dev 快照刷新至 2026-08-15（6583 模型 / 185 供应商），新增 GLM-5.3 等
+
+## v0.7.4 (2026-08-09)
+
+- **热力图方块双击弹窗**：查看当日模型用量分布，每种模型独立展示 Token 用量 / 请求次数 / 费用 / 缓存命中率（此前仅有 token 数）
+- 后端 `[dashboard] by_model` 从纯 token 数升级为结构化对象（含 requests / cost / cacheHitRate），弹窗与柱状图双击均展示全量指标
+- 范围外日期（不在当前 7d / 30d / all 内）不可双击；Escape / 遮罩 / 关闭按钮均可关闭弹窗
+
+## v0.7.3 (2026-08-07)
+
+- **Kimi Code WebUI 应用内嵌窗口**：新版 Web 界面以独立顶层窗口打开（1100×750、可缩放、居中），不再跳转外部浏览器
+- **单例 + 服务器自动管理**：窗口最多一个，重复点击自动聚焦；复用已运行的 `kimi web` 服务，自启动的服务器在窗口关闭或应用退出时自动清理
+- 「在应用内打开」与「在浏览器打开」两个入口并存（后者复用本地服务器，直接打开 `http://127.0.0.1:58627`）
+
+## v0.7.2 (2026-08-07)
+
+- 「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
+- **新增 Kimi Code WebUI 快捷打开**：高级设置页可将新版 Web 界面以独立窗口在应用内打开，也可在系统浏览器打开（需 kimi-code 0.33+，kimi 命令在 PATH 中）
+- **子代理模型指定**：为子代理选择次主力模型后默认绑定该模型，不再继承主模型；实验开关改为滑动开关并修复开启态不可见问题
+- 兼容 kimi-code 0.33+（v2 引擎）：`loop_control` 改用新键名 `max_attempts_per_step`（旧配置自动迁移，保留旧键兼容 legacy 引擎）；Release 附带 `install-macos.sh` 一键安装脚本
+
+## v0.7.1 (2026-08-06)
+
+- **新增「子代理模型（次主力模型）」设置**：选择后子代理默认绑定该模型，不再继承主模型，仅保存模型引用即可继承上下文长度与思考模式
+- **仪表盘用量统计**：识别子代理请求（`__secondary__`）归为「子代理模型（估算）」独立展示，并缓存次主力模型定价，删除配置后历史记录仍稳定估算
+- 自定义网关模型（未收录 models.dev）按官方同名模型跨 provider 匹配价格，不再落入兜底估算
@@ -0,0 +1,23 @@
+# KimiSwitch v0.7.21
+
+## 新增
+
+- **models.dev 在线同步**：高级设置新增「模型参考数据（models.dev）」卡片——一键在线拉取 models.dev 最新快照，新模型发布后无需等待 KimiSwitch 发版：
+  - 覆盖三类参考数据：模型上下文长度、能力标记（思考 / 工具 / 图像 / 视频）、单价（$/M tokens）
+  - 同步后立即生效：供应商编辑页的参数自动填充、仪表盘用量计价（Rust 侧价格索引按快照文件 mtime 自动重建，无需重启应用）
+  - 数据落在 `~/.kimi-switch/models-dev.json`，打包内置快照始终作为兜底；无效副本自动回退
+  - 支持一键「恢复内置数据」，随时可重新同步
+  - 状态行显示当前数据来源（在线同步 / 随版本内置）、快照日期与模型 / 供应商数量
+  - 代理环境自动适配：依次读取 `HTTPS_PROXY` / `HTTP_PROXY` 环境变量与 git `http.proxy` 配置
+
+## 修复
+
+- **构建链路数据漂移**：`fetch-models-dev.mjs` 此前只写 `src/lib/models-dev.json`（Rust 计价内嵌），前端实际加载的 `public/models-dev.json` 需手工拷贝，历史上已两次滞后。现在两个文件由脚本同步写入，彻底消除漂移
+
+## 适配
+
+- **kimi-code 2.1.0 已适配确认**：实验 flag 注册表（5 个）、config 模式、用量 / quota 格式均无变化；上游 `[watch]` 默认关闭与现有开关行为一致；新增 `tui_mode`（`~/.kimi-code/tui.toml`）不涉及配置管理面，暂无适配需求
+
+## 数据
+
+- models-dev 快照同步至 2026-09-23（8,126 模型 / 223 供应商）
@@ -0,0 +1,9 @@
+# KimiSwitch v0.7.22
+
+## 修复
+
+- **Sub2API / NewAPI 模板查询在供应商列表不显示**：为没有预设识别类型的自定义供应商（如自家中转站）配置「Sub2API 中转站」或「NewAPI 中转站」模板后，配置弹窗里「测试查询」正常返回余额，但回到供应商列表却始终空白。根因：列表的查询与显示门槛只认预设注入的 `usageKinds`，而后端模板查询本就绕过 `usageKinds` 直接执行——前端口径与后端不一致。现在二者对齐，显式模板的供应商在列表正常显示余额 / 配额窗口、支持自动查询间隔与预警阈值。
+
+## 备注
+
+- 仅前端修复（`useUsageQuery` 支持判定 + 卡片渲染门槛），后端查询逻辑无变化；已补 3 个单测覆盖支持判定。
@@ -0,0 +1,18 @@
+# KimiSwitch v0.7.23
+
+## 新功能
+
+- **思考等级五档 + 能力感知**：全局配置的思考等级由「低 / 中 / 高」三档扩展为「低 / 中 / 高 / Max / XHigh」五档，并按当前默认模型的实际能力自适应——
+  - 模型在 config.toml 声明了 `support_efforts` 时，不支持的档位自动置灰并给出原因提示（如 deepseek-v4.1-flash 仅支持「高」）
+  - models.dev 标记不支持思考（reasoning=false）的模型，整个思考等级区禁用并给出警告
+  - 未声明档位的模型保留全部可选，Max / XHigh 附带「上游可能拒绝此档位」提示
+  - 等级行下方新增说明：实际生效档位取决于当前默认模型
+
+## 修复
+
+- **max 档位被误显示为「高」**：此前代码基于过时信息（"上游已移除 max 档位"）把存储的 `effort = "max"` 读取时强制改为 `high`；实测 kimi-code 2.1.1 中 max 为合法档位（如 kimi-for-coding 默认即为 max），现已原样保留与显示。手写的任意档位值（上游本就是自由字符串）也会作为额外可选项正常显示，不再被丢弃
+- **子代理设置页**：继承档位显示同步修正，max / xhigh 正常显示，未知值按原样展示
+
+## 备注
+
+- 仅前端改动（新增 `thinking-efforts` 能力判定模块 + `Segmented` 控件逐选项禁用/tooltip），无 Rust 变更；新增 14 个单测覆盖三层能力判定规则，全量 117 个测试通过
@@ -1,7 +1,7 @@
 {
  "name": "kimiswitch",
  "private": true,
-  "version": "0.7.20",
+  "version": "0.7.23",
  "type": "module",
  "scripts": {
    "dev": "vite",
@@ -1,9 +1,17 @@
 #!/usr/bin/env node
 /**
 * Fetch the latest model reference data from models.dev and write:
- *   1. src/lib/models-dev.json      — compact per-model snapshot (frontend)
- *   2. src/lib/models-dev-full.json — provider-grouped full list incl. pricing
- *   3. src/lib/models-dev.last-good.json — local backup of the last success
+ *   1. src/lib/models-dev.json      — compact per-model snapshot; embedded
+ *                                     into the Rust binary by dashboard.rs
+ *                                     (include_str!) and loaded by the
+ *                                     frontend as fallback
+ *   2. public/models-dev.json       — the same snapshot as a static asset;
+ *                                     this is what the frontend actually
+ *                                     serves at runtime (must stay in sync
+ *                                     with #1 — they serve different halves
+ *                                     of the app)
+ *   3. src/lib/models-dev-full.json — provider-grouped full list incl. pricing
+ *   4. src/lib/models-dev.last-good.json — local backup of the last success
 *
 * The data source is https://models.dev/api.json (176 providers; each model
 * carries `cost` = { input, output, cache_read?, cache_write? } in $/M tokens).
@@ -63,6 +71,8 @@ if (proxy && process.env.NODE_USE_ENV_PROXY !== "1") {
 const SOURCE_URL = "https://models.dev/api.json";
 const ROOT = join(dirname(fileURLToPath(import.meta.url)), "..");
 const SNAPSHOT = join(ROOT, "src", "lib", "models-dev.json");
+// Static-asset copy served to the frontend (see header — must equal SNAPSHOT).
+const SNAPSHOT_PUBLIC = join(ROOT, "public", "models-dev.json");
 const FULL = join(ROOT, "src", "lib", "models-dev-full.json");
 // Local backup of the last successful snapshot. Not committed to git (see
 // .gitignore); the committed models-dev.json itself is the versioned fallback.
@@ -168,12 +178,15 @@ async function main() {
  }

  await mkdir(dirname(SNAPSHOT), { recursive: true });
+  await mkdir(dirname(SNAPSHOT_PUBLIC), { recursive: true });
  const json = JSON.stringify(snapshot, null, 2) + "\n";
  await writeFile(SNAPSHOT, json, "utf8");
+  await writeFile(SNAPSHOT_PUBLIC, json, "utf8");
  await writeFile(FULL, JSON.stringify(full, null, 2) + "\n", "utf8");
  await writeFile(BACKUP, json, "utf8");

  console.log(`models.dev snapshot: ${modelCount} models -> ${SNAPSHOT}`);
+  console.log(`models.dev static asset: -> ${SNAPSHOT_PUBLIC}`);
  console.log(
    `models.dev full list: ${Object.keys(full.providers).length} providers -> ${FULL}`,
  );
@@ -3,7 +3,8 @@
 # (git.codingplan.site/admin/KimiCodeSwitch).
 #
 # Usage: publish-gitea.sh <tag> <asset> [asset...]
-#   tag    : e.g. v0.7.15; notes read from release-notes-<tag>.md
+#   tag    : e.g. v0.7.15; notes read from docs/release-notes/<tag>.md
+#            (legacy root release-notes-<tag>.md as fallback)
 #   assets : files to attach (MSI, install-macos.sh, ...)
 #
 # Auth: reuses the stored git credential for git.codingplan.site
@@ -15,7 +16,9 @@ REPO="admin/KimiCodeSwitch"
 TAG="${1:?usage: publish-gitea.sh <tag> <asset> [asset...]}"
 shift

-NOTES_FILE="release-notes-${TAG}.md"
+NOTES_FILE="docs/release-notes/${TAG}.md"
+[ -f "$NOTES_FILE" ] || NOTES_FILE="docs/release-notes/release-notes-${TAG}.md"
+[ -f "$NOTES_FILE" ] || NOTES_FILE="release-notes-${TAG}.md"
 json_escape() {
  python - "$1" <<'PYEOF'
 import json, sys
@@ -1978,7 +1978,7 @@ dependencies = [

 [[package]]
 name = "kimiswitch"
-version = "0.7.20"
+version = "0.7.23"
 dependencies = [
 "anyhow",
 "chrono",
@@ -1,6 +1,6 @@
 [package]
 name = "kimiswitch"
-version = "0.7.20"
+version = "0.7.23"
 description = "Kimi Switch - model config manager"
 authors = ["codingplan.site"]
 edition = "2021"
@@ -6,7 +6,8 @@ use std::collections::HashMap;
 use std::fs;
 use std::io::{BufRead, BufReader};
 use std::path::{Path, PathBuf};
-use std::sync::{Mutex, OnceLock};
+use std::sync::{Arc, Mutex, OnceLock, RwLock};
+use std::time::SystemTime;
 use walkdir::WalkDir;

 // ---------------------------------------------------------------------------
@@ -534,50 +535,62 @@ struct ModelsDevCost {
    cache_write: Option<f64>,
 }

-/// Compiled-in models.dev snapshot (`src/lib/models-dev.json`, generated by
-/// scripts/fetch-models-dev.mjs before every build via the `pretauri` hook).
-/// Keyed by "<provider>/<model>", lowercased.
-const MODELS_DEV_SNAPSHOT: &str = include_str!("../../src/lib/models-dev.json");
-
-fn models_dev_cost_index() -> &'static HashMap<String, ModelsDevCost> {
-    static INDEX: OnceLock<HashMap<String, ModelsDevCost>> = OnceLock::new();
-    INDEX.get_or_init(|| {
-        let mut map = HashMap::new();
-        let Ok(v) = serde_json::from_str::<serde_json::Value>(MODELS_DEV_SNAPSHOT) else {
-            return map;
-        };
-        let Some(obj) = v.as_object() else {
-            return map;
-        };
-        for (key, entry) in obj {
-            // Skip the "last_updated" metadata key and entries without cost.
-            if !entry.is_object() || entry.get("cost").is_none() {
-                continue;
+/// models.dev price index over the effective snapshot (runtime-synced copy
+/// when present, else the compiled-in `src/lib/models-dev.json` generated by
+/// scripts/fetch-models-dev.mjs). Keyed by "<provider>/<model>", lowercased.
+/// Cached and rebuilt only when the synced file's mtime changes, so an online
+/// sync (models_dev::sync_models_dev) refreshes prices without a restart.
+fn models_dev_cost_index() -> Arc<HashMap<String, ModelsDevCost>> {
+    static CACHE: OnceLock<RwLock<Option<(Option<SystemTime>, Arc<HashMap<String, ModelsDevCost>>)>>>=
+        OnceLock::new();
+    let cache = CACHE.get_or_init(|| RwLock::new(None));
+    let stamp = fs::metadata(crate::models_dev::synced_path())
+        .and_then(|m| m.modified())
+        .ok();
+    {
+        let cached = cache.read().expect("price index lock");
+        if let Some((cached_stamp, idx)) = cached.as_ref() {
+            if *cached_stamp == stamp {
+                return idx.clone();
            }
-            let Some(cost) = entry.get("cost").and_then(|c| c.as_object()) else {
-                continue;
-            };
-            let num = |k: &str| cost.get(k).and_then(|x| x.as_f64());
-            let (Some(input), Some(output)) = (num("input"), num("output")) else {
-                continue;
-            };
-            map.insert(
-                key.to_ascii_lowercase(),
-                ModelsDevCost {
-                    input,
-                    output,
-                    cache_read: num("cache_read").unwrap_or(0.0),
-                    cache_write: num("cache_write"),
-                },
-            );
        }
-        map
-    })
+    }
+    let mut map = HashMap::new();
+    let v = serde_json::from_str::<serde_json::Value>(&crate::models_dev::effective_snapshot());
+    if let Ok(v) = v {
+        if let Some(obj) = v.as_object() {
+            for (key, entry) in obj {
+                // Skip the "last_updated" metadata key and entries without cost.
+                if !entry.is_object() || entry.get("cost").is_none() {
+                    continue;
+                }
+                let Some(cost) = entry.get("cost").and_then(|c| c.as_object()) else {
+                    continue;
+                };
+                let num = |k: &str| cost.get(k).and_then(|x| x.as_f64());
+                let (Some(input), Some(output)) = (num("input"), num("output")) else {
+                    continue;
+                };
+                map.insert(
+                    key.to_ascii_lowercase(),
+                    ModelsDevCost {
+                        input,
+                        output,
+                        cache_read: num("cache_read").unwrap_or(0.0),
+                        cache_write: num("cache_write"),
+                    },
+                );
+            }
+        }
+    }
+    let idx = Arc::new(map);
+    *cache.write().expect("price index lock") = Some((stamp, idx.clone()));
+    idx
 }

-/// Warm the compiled-in models.dev price index off the first dashboard open:
-/// parsing the ~1.3 MB snapshot costs tens of milliseconds, so build it on a
-/// background thread at startup instead of lazily on the first get_summary.
+/// Warm the models.dev price index off the first dashboard open: parsing the
+/// ~1.3 MB snapshot costs tens of milliseconds, so build it on a background
+/// thread at startup instead of lazily on the first get_summary.
 pub fn warm_price_index() {
    std::thread::spawn(|| {
        let _ = models_dev_cost_index();
@@ -608,14 +621,15 @@ fn provider_rank(key: &str) -> usize {
 ///    (deterministic tie-break). Returns None when nothing matches.
 fn models_dev_lookup(model_name: &str) -> Option<(String, ModelsDevCost)> {
    let lower = model_name.to_ascii_lowercase();
-    if let Some(cost) = models_dev_cost_index().get(&lower) {
+    let index = models_dev_cost_index();
+    if let Some(cost) = index.get(&lower) {
        return Some((lower, *cost));
    }
    let bare = model_name.rsplit_once('/').map(|(_, b)| b).unwrap_or(model_name);
    let bare_l = bare.to_ascii_lowercase();
    // Match on the model part after the last '/', not the full key: aliases
    // strip the family prefix (record "k2.5" vs models.dev "kimi-k2.5").
-    let mut matches: Vec<(&String, &ModelsDevCost)> = models_dev_cost_index()
+    let mut matches: Vec<(&String, &ModelsDevCost)> = index
        .iter()
        .filter(|(key, _)| {
            key.rsplit_once('/')
@@ -695,7 +709,8 @@ fn match_price(model_name: &str) -> (String, f64, f64, f64, bool) {
 fn priced_suffix_lookup(model_name: &str) -> Option<(String, ModelsDevCost)> {
    let bare = model_name.rsplit_once('/').map(|(_, b)| b).unwrap_or(model_name);
    let bare_l = bare.to_ascii_lowercase();
-    let mut matches: Vec<(&String, &ModelsDevCost)> = models_dev_cost_index()
+    let index = models_dev_cost_index();
+    let mut matches: Vec<(&String, &ModelsDevCost)> = index
        .iter()
        .filter(|(key, cost)| {
            (cost.input > 0.0 || cost.output > 0.0)
@@ -4,6 +4,7 @@ pub mod dashboard;
 pub mod db;
 pub mod kimi_code_io;
 pub mod models;
+pub mod models_dev;
 pub mod oauth;
 pub mod pi_io;
 pub mod plugins;
@@ -119,6 +120,10 @@ pub fn run() {
            commands::kimi_oauth_start,
            commands::kimi_oauth_poll,
            commands::get_experimental_env_status,
+            models_dev::sync_models_dev,
+            models_dev::get_models_dev_status,
+            models_dev::get_models_dev_snapshot,
+            models_dev::reset_models_dev,
            dashboard::get_paths,
            dashboard::get_prices,
            dashboard::get_summary,
@@ -0,0 +1,352 @@
+//! models.dev reference-data sync.
+//!
+//! The app carries a compiled-in models.dev snapshot (generated by
+//! scripts/fetch-models-dev.mjs before every build). This module lets the
+//! user refresh it at runtime from https://models.dev/api.json without
+//! rebuilding: the synced copy lives at `~/.kimi-switch/models-dev.json` and
+//! takes precedence whenever it exists and parses. The dashboard price index
+//! (see dashboard::models_dev_cost_index) re-checks that file's mtime so new
+//! prices take effect without an app restart.
+
+use std::collections::HashSet;
+use std::fs;
+use std::path::PathBuf;
+use std::time::Duration;
+
+use chrono::Utc;
+use serde::Serialize;
+use serde_json::Value;
+
+use crate::db::kimi_switch_data_dir;
+
+/// Compiled-in snapshot — the same file the dashboard price index used to
+/// embed directly. Always available; the runtime-synced copy (if any) wins.
+const BUILTIN_SNAPSHOT: &str = include_str!("../../src/lib/models-dev.json");
+
+const SOURCE_URL: &str = "https://models.dev/api.json";
+const SYNCED_FILE: &str = "models-dev.json";
+
+/// Where the runtime-synced snapshot lives. The dashboard stats this path's
+/// mtime to decide when to rebuild its price index.
+pub fn synced_path() -> PathBuf {
+    kimi_switch_data_dir().join(SYNCED_FILE)
+}
+
+/// Read + validate the synced copy. Anything unparsable or missing
+/// `last_updated` is treated as absent so the built-in snapshot takes over.
+fn read_synced() -> Option<String> {
+    let raw = fs::read_to_string(synced_path()).ok()?;
+    let v: Value = serde_json::from_str(&raw).ok()?;
+    if v.get("last_updated").and_then(|x| x.as_str()).is_some() {
+        Some(raw)
+    } else {
+        None
+    }
+}
+
+/// The snapshot the app should use right now: the synced copy when present
+/// and valid, otherwise the compiled-in one.
+pub fn effective_snapshot() -> String {
+    read_synced().unwrap_or_else(|| BUILTIN_SNAPSHOT.to_string())
+}
+
+#[derive(Debug, Serialize, Clone)]
+#[serde(rename_all = "camelCase")]
+pub struct ModelsDevStatus {
+    /// "synced" = runtime online copy; "builtin" = compiled at build time.
+    pub source: &'static str,
+    pub last_updated: String,
+    pub model_count: u64,
+    pub provider_count: u64,
+}
+
+/// Count models / providers and read the `last_updated` stamp from a
+/// snapshot string ("<provider>/<model>" keys, plus the `last_updated` meta).
+fn status_of(source: &'static str, raw: &str) -> ModelsDevStatus {
+    let v: Value = serde_json::from_str(raw).unwrap_or(Value::Null);
+    let mut model_count = 0u64;
+    let mut providers = HashSet::new();
+    if let Some(obj) = v.as_object() {
+        for (k, entry) in obj {
+            if k == "last_updated" || !entry.is_object() {
+                continue;
+            }
+            model_count += 1;
+            if let Some((p, _)) = k.split_once('/') {
+                providers.insert(p.to_string());
+            }
+        }
+    }
+    ModelsDevStatus {
+        source,
+        last_updated: v
+            .get("last_updated")
+            .and_then(|x| x.as_str())
+            .unwrap_or("unknown")
+            .to_string(),
+        model_count,
+        provider_count: providers.len() as u64,
+    }
+}
+
+fn status(source: &'static str) -> ModelsDevStatus {
+    match source {
+        "synced" => {
+            let raw = read_synced().unwrap_or_default();
+            status_of("synced", &raw)
+        }
+        _ => status_of("builtin", BUILTIN_SNAPSHOT),
+    }
+}
+
+/// Numbers serialize as integers when they are whole (matches the mjs
+/// script's JSON output, e.g. context 1000000 not 1000000.0).
+fn num(v: f64) -> Value {
+    if v.fract() == 0.0 && v.abs() < 9e15 {
+        Value::from(v as i64)
+    } else {
+        Value::from(v)
+    }
+}
+
+/// Transform models.dev api.json into the compact snapshot shape produced by
+/// scripts/fetch-models-dev.mjs (the canonical field-picking logic lives
+/// there; this is the runtime port). Pure — unit-tested below.
+fn build_snapshot(raw: &Value) -> Result<Value, String> {
+    let providers = raw.as_object().ok_or("api.json is not a JSON object")?;
+    let mut snapshot = serde_json::Map::new();
+    snapshot.insert(
+        "last_updated".into(),
+        Value::String(Utc::now().format("%Y-%m-%d").to_string()),
+    );
+
+    for (provider_id, provider) in providers {
+        let Some(models) = provider.get("models").and_then(|m| m.as_object()) else {
+            continue;
+        };
+        for (model_id, m) in models {
+            let mut entry = serde_json::Map::new();
+            if let Some(name) = m.get("name").and_then(|x| x.as_str()) {
+                entry.insert("name".into(), Value::String(name.to_string()));
+            }
+            // Context 0 means "not applicable" (image/audio models) — treat as
+            // missing so callers fall back to defaults.
+            if let Some(ctx) = m
+                .pointer("/limit/context")
+                .and_then(|x| x.as_f64())
+                .filter(|c| *c > 0.0)
+            {
+                entry.insert("context".into(), num(ctx));
+            }
+            for flag in ["reasoning", "tool_call", "structured_output"] {
+                if m.get(flag).and_then(|x| x.as_bool()) == Some(true) {
+                    entry.insert(flag.into(), Value::Bool(true));
+                }
+            }
+            if let Some(input) = m.pointer("/modalities/input").and_then(|x| x.as_array()) {
+                let has = |s: &str| input.iter().any(|v| v.as_str() == Some(s));
+                if has("image") {
+                    entry.insert("image".into(), Value::Bool(true));
+                }
+                if has("video") {
+                    entry.insert("video".into(), Value::Bool(true));
+                }
+            }
+            if let Some(cost) = m.get("cost").and_then(|c| c.as_object()) {
+                let mut c = serde_json::Map::new();
+                for k in ["input", "output", "cache_read", "cache_write"] {
+                    if let Some(n) = cost.get(k).and_then(|x| x.as_f64()) {
+                        c.insert(k.into(), num(n));
+                    }
+                }
+                if !c.is_empty() {
+                    entry.insert("cost".into(), Value::Object(c));
+                }
+            }
+            snapshot.insert(format!("{provider_id}/{model_id}"), Value::Object(entry));
+        }
+    }
+    Ok(Value::Object(snapshot))
+}
+
+/// reqwest is built without the default system-proxy feature, so proxies must
+/// be attached explicitly. Mirror fetch-models-dev.mjs: env vars first, then
+/// git's http.proxy.
+fn proxy_url() -> Option<String> {
+    let env = ["HTTPS_PROXY", "https_proxy", "HTTP_PROXY", "http_proxy"]
+        .iter()
+        .find_map(|k| std::env::var(k).ok().filter(|p| !p.trim().is_empty()));
+    env.or_else(git_http_proxy)
+}
+
+fn git_http_proxy() -> Option<String> {
+    let out = std::process::Command::new("git")
+        .args(["config", "--get", "http.proxy"])
+        .output()
+        .ok()?;
+    let s = String::from_utf8_lossy(&out.stdout).trim().to_string();
+    (!s.is_empty()).then_some(s)
+}
+
+fn http_client() -> Result<reqwest::Client, String> {
+    let mut b = reqwest::Client::builder()
+        .user_agent(concat!("KimiSwitch/", env!("CARGO_PKG_VERSION")))
+        .timeout(Duration::from_secs(60));
+    if let Some(p) = proxy_url() {
+        b = b.proxy(reqwest::Proxy::all(&p).map_err(|e| format!("invalid proxy '{p}': {e}"))?);
+    }
+    b.build().map_err(|e| format!("failed to build HTTP client: {e}"))
+}
+
+/// Download api.json, transform it and atomically replace the synced copy.
+/// On any failure the previous snapshot (synced or built-in) stays untouched.
+async fn sync_from_remote() -> Result<ModelsDevStatus, String> {
+    let client = http_client()?;
+    let raw: Value = client
+        .get(SOURCE_URL)
+        .header("Accept", "application/json")
+        .send()
+        .await
+        .map_err(|e| format!("request failed: {e}"))?
+        .error_for_status()
+        .map_err(|e| format!("models.dev returned {e}"))?
+        .json()
+        .await
+        .map_err(|e| format!("failed to parse api.json: {e}"))?;
+
+    let snapshot = build_snapshot(&raw)?;
+    let json = serde_json::to_string_pretty(&snapshot).map_err(|e| e.to_string())? + "\n";
+
+    let path = synced_path();
+    if let Some(dir) = path.parent() {
+        fs::create_dir_all(dir).map_err(|e| format!("failed to create {}: {e}", dir.display()))?;
+    }
+    let tmp = path.with_extension("json.tmp");
+    fs::write(&tmp, &json).map_err(|e| format!("failed to write {}: {e}", tmp.display()))?;
+    // Windows rename replaces an existing destination (MOVEFILE_REPLACE_EXISTING).
+    fs::rename(&tmp, &path).map_err(|e| {
+        let _ = fs::remove_file(&tmp);
+        format!("failed to replace {}: {e}", path.display())
+    })?;
+    Ok(status_of("synced", &json))
+}
+
+/// Refresh the models.dev snapshot from the network.
+#[tauri::command]
+pub async fn sync_models_dev() -> Result<ModelsDevStatus, String> {
+    sync_from_remote().await
+}
+
+/// Where the current reference data comes from (for the settings card).
+#[tauri::command]
+pub fn get_models_dev_status() -> Result<ModelsDevStatus, String> {
+    Ok(if read_synced().is_some() {
+        status("synced")
+    } else {
+        status("builtin")
+    })
+}
+
+/// The synced snapshot's raw JSON, or None when the frontend should keep
+/// using the bundled static asset.
+#[tauri::command]
+pub fn get_models_dev_snapshot() -> Result<Option<String>, String> {
+    Ok(read_synced())
+}
+
+/// Drop the synced copy and fall back to the compiled-in snapshot.
+#[tauri::command]
+pub fn reset_models_dev() -> Result<ModelsDevStatus, String> {
+    match fs::remove_file(synced_path()) {
+        Ok(()) => {}
+        Err(e) if e.kind() == std::io::ErrorKind::NotFound => {}
+        Err(e) => return Err(format!("failed to remove synced snapshot: {e}")),
+    }
+    Ok(status("builtin"))
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+    use serde_json::json;
+
+    #[test]
+    fn build_snapshot_picks_reference_fields() {
+        let raw = json!({
+            "moonshotai": {
+                "models": {
+                    "kimi-k3": {
+                        "name": "Kimi K3",
+                        "limit": { "context": 262144, "output": 16384 },
+                        "reasoning": true,
+                        "tool_call": true,
+                        "structured_output": true,
+                        "modalities": { "input": ["text", "image"], "output": ["text"] },
+                        "cost": { "input": 3.0, "output": 15.0, "cache_read": 0.3, "cache_write": 0.0 }
+                    },
+                    "kimi-image": {
+                        "name": "Kimi Image",
+                        "limit": { "context": 0 },
+                        "modalities": { "input": ["text", "image", "video"] }
+                    }
+                }
+            },
+            "zhipuai": {
+                "models": {
+                    "glm-5.3": {
+                        "name": "GLM-5.3",
+                        "reasoning": false,
+                        "cost": { "input": 0.6, "output": 2.2 }
+                    }
+                }
+            }
+        });
+
+        let snap = build_snapshot(&raw).unwrap();
+        let obj = snap.as_object().unwrap();
+        assert!(obj.contains_key("last_updated"));
+
+        let k3 = &obj["moonshotai/kimi-k3"];
+        assert_eq!(k3["name"], "Kimi K3");
+        assert_eq!(k3["context"], 262144);
+        assert_eq!(k3["reasoning"], true);
+        assert_eq!(k3["tool_call"], true);
+        assert_eq!(k3["structured_output"], true);
+        assert_eq!(k3["image"], true);
+        assert!(k3.get("video").is_none());
+        assert_eq!(k3["cost"]["input"], 3.0);
+        assert_eq!(k3["cost"]["cache_read"], 0.3);
+        assert_eq!(k3["cost"]["cache_write"], 0.0);
+
+        // Context 0 → dropped; video modality → flagged.
+        let img = &obj["moonshotai/kimi-image"];
+        assert!(img.get("context").is_none());
+        assert_eq!(img["video"], true);
+        assert!(img.get("cost").is_none());
+
+        // Explicit false is omitted (same as the mjs script).
+        let glm = &obj["zhipuai/glm-5.3"];
+        assert!(glm.get("reasoning").is_none());
+        assert_eq!(glm["cost"]["input"], 0.6);
+    }
+
+    #[test]
+    fn status_counts_models_and_providers() {
+        let raw = r#"{
+            "last_updated": "2026-09-21",
+            "moonshotai/kimi-k3": { "cost": { "input": 3.0 } },
+            "moonshotai/kimi-k2": { "name": "K2" },
+            "zhipuai/glm-5.3": {}
+        }"#;
+        let st = status_of("builtin", raw);
+        assert_eq!(st.last_updated, "2026-09-21");
+        assert_eq!(st.model_count, 3);
+        assert_eq!(st.provider_count, 2);
+    }
+
+    #[test]
+    fn num_serializes_whole_numbers_as_integers() {
+        assert_eq!(num(262144.0), json!(262144));
+        assert_eq!(num(0.3), json!(0.3));
+    }
+}
@@ -1,6 +1,6 @@
 {
  "productName": "Kimi Switch",
-  "version": "0.7.20",
+  "version": "0.7.23",
  "identifier": "com.kimiswitch.app",
  "build": {
    "beforeDevCommand": "npm run dev",
@@ -1,12 +1,20 @@
-import { useEffect, useState } from "react";
+import { useEffect, useReducer, 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 { getModelRef, modelsDevReady } from "../lib/models-dev";
+import {
+  THINKING_EFFORTS,
+  readSupportEfforts,
+  resolveThinkingEffortSupport,
+  type EffortAvailability,
+} from "../lib/thinking-efforts";
 import type {
  AgentSettings,
  ExperimentalEnvStatus,
  Hook,
+  Model,
  PermissionRule,
 } from "../types";
 import { Card, Checkbox, NumberField, Segmented } from "./ui/controls";
@@ -14,14 +22,18 @@ import { Card, Checkbox, NumberField, Segmented } from "./ui/controls";
 interface AgentSettingsPanelProps {
  rawOther: unknown;
  onChange: (nextRawOther: unknown) => void;
+  /** Configured models (alias → entry) — used to resolve effort support. */
+  models?: Record<string, Model>;
+  /** Alias of the model the effort tier actually applies to. */
+  defaultModel?: string | null;
 }

-// 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> = {
+const THINKING_LABELS: Record<(typeof THINKING_EFFORTS)[number], TranslationKey> = {
  low: "thinkingLow",
  medium: "thinkingMedium",
  high: "thinkingHigh",
+  max: "thinkingMax",
+  xhigh: "thinkingXHigh",
 };

 const PERMISSION_DECISIONS = ["allow", "deny", "ask"] as const;
@@ -44,7 +56,7 @@ const COMMON_EVENTS = [
  "SessionEnd",
 ] as const;

-export function AgentSettingsPanel({ rawOther, onChange }: AgentSettingsPanelProps) {
+export function AgentSettingsPanel({ rawOther, onChange, models, defaultModel }: AgentSettingsPanelProps) {
  const { t } = useTranslation();
  const settings = getAgentSettings(rawOther);
  /**
@@ -52,6 +64,9 @@ export function AgentSettingsPanel({ rawOther, onChange }: AgentSettingsPanelPro
   * v1-engine users; the default (v2) writes only max_attempts_per_step.
   */
  const [legacyV1, setLegacyV1] = useState(false);
+  // The models.dev index loads in the background; re-render once it lands so
+  // the reasoning flag can disable the thinking area.
+  const [, forceModelsDevReady] = useReducer((x: number) => x + 1, 0);

  useEffect(() => {
    invoke<ExperimentalEnvStatus>("get_experimental_env_status")
@@ -62,6 +77,18 @@ export function AgentSettingsPanel({ rawOther, onChange }: AgentSettingsPanelPro
      .catch(() => setLegacyV1(false));
  }, []);

+  useEffect(() => {
+    // The panel only needs the flag once; App owns the permanent
+    // onModelsDevReady listener for later hot swaps.
+    let alive = true;
+    modelsDevReady().then(() => {
+      if (alive) forceModelsDevReady();
+    });
+    return () => {
+      alive = false;
+    };
+  }, []);
+
  const update = (patch: Partial<AgentSettings>) => {
    onChange(setAgentSettings(rawOther, patch, { legacyV1 }));
  };
@@ -92,6 +119,42 @@ export function AgentSettingsPanel({ rawOther, onChange }: AgentSettingsPanelPro

  const thinkingEnabled = settings.thinking?.enabled ?? true;

+  // Which tiers the current default model accepts: `[models."<alias>"]
+  // support_efforts` first, then the models.dev `reasoning` flag.
+  const defaultEntry = defaultModel ? models?.[defaultModel] : undefined;
+  const effortSupport = resolveThinkingEffortSupport({
+    supportEfforts: readSupportEfforts(defaultEntry),
+    reasoning: defaultEntry ? getModelRef(defaultEntry.model)?.reasoning : undefined,
+  });
+
+  const effortNote = (availability: EffortAvailability): string | undefined => {
+    const note = availability.note;
+    if (!note) return undefined;
+    if (note.kind === "thinkingUnsupported") return t("thinkingEffortUnsupported");
+    if (note.kind === "tierUnsupported") {
+      return t("thinkingEffortTierUnsupported", { levels: note.supported.join(", ") });
+    }
+    return t("thinkingEffortTierUndeclared");
+  };
+
+  // Upstream takes a free-form string: keep a hand-written tier visible and
+  // selectable instead of dropping it from the control.
+  const effort = settings.thinking?.effort ?? "medium";
+  const effortOptions: {
+    key: string;
+    label: string;
+    disabled?: boolean;
+    title?: string;
+  }[] = THINKING_EFFORTS.map((tier) => ({
+    key: tier,
+    label: t(THINKING_LABELS[tier]),
+    disabled: !effortSupport.levels[tier].enabled,
+    title: effortNote(effortSupport.levels[tier]),
+  }));
+  if (!THINKING_EFFORTS.includes(effort as (typeof THINKING_EFFORTS)[number])) {
+    effortOptions.push({ key: effort, label: effort });
+  }
+
  return (
    <div className="mt-6 space-y-4">
      <h3 className="text-content-muted text-sm font-medium">{t("agentSettings")}</h3>
@@ -105,17 +168,20 @@ export function AgentSettingsPanel({ rawOther, onChange }: AgentSettingsPanelPro
        <div className="flex items-center gap-3 flex-wrap">
          <span className="text-sm text-content-muted">{t("thinkingLevel")}</span>
          <Segmented
-            options={THINKING_LEVELS.map((lvl) => ({
-              key: lvl,
-              label: t(THINKING_LABELS[lvl]),
-            }))}
-            value={settings.thinking?.effort ?? "medium"}
-            onChange={(effort) =>
-              updateThinking({ effort: effort as NonNullable<AgentSettings["thinking"]>["effort"] })
-            }
-            disabled={!thinkingEnabled}
+            options={effortOptions}
+            value={effort}
+            onChange={(next) => updateThinking({ effort: next })}
+            disabled={!thinkingEnabled || !effortSupport.thinkingSupported}
          />
        </div>
+        {!effortSupport.thinkingSupported && (
+          <p className="text-xs text-amber-500 dark:text-amber-400">
+            {t("thinkingEffortUnsupported")}
+          </p>
+        )}
+        <p className="text-xs text-content-muted">
+          {t("thinkingEffortDependsOnDefaultModel")}
+        </p>
        <Checkbox
          label={t("thinkingKeep")}
          checked={settings.thinking?.keep === "all"}
@@ -1,4 +1,4 @@
-import { useEffect, useId, useState } from "react";
+import { useEffect, useId, useMemo, useState } from "react";
 import { createPortal } from "react-dom";
 import { invoke } from "@tauri-apps/api/core";
 import { useTranslation } from "../i18n";
@@ -142,6 +142,13 @@ export function ProviderEdit({
  const apiKeyEnvSupported = agent === "kimi_code";
  const apiKeyEnvMode = apiKeyEnvSupported && provider.api_key_env != null;

+  // Alias → entry lookup for panels that resolve a model by alias
+  // (AgentSettingsPanel reads the default model's support_efforts).
+  const modelsByAlias = useMemo(
+    () => Object.fromEntries(models.map((m) => [m.alias, m])),
+    [models]
+  );
+
  useEffect(() => {
    const def = defaultBaseUrl(agent, provider.provider_type);
    if (!def) return;
@@ -462,6 +469,8 @@ export function ProviderEdit({
              <AgentSettingsPanel
                rawOther={rawOther}
                onChange={onRawOtherChange}
+                models={modelsByAlias}
+                defaultModel={defaultModel}
              />
            )}
          </>
@@ -96,7 +96,9 @@ function ProviderCard({
    provider.usageKinds,
    provider.usageConfig?.autoQueryIntervalMinutes,
    provider.usageConfig?.enabled === false,
-    provider.usageConfig?.threshold
+    provider.usageConfig?.threshold,
+    provider.usageConfig?.templateType === "newapi" ||
+      provider.usageConfig?.templateType === "sub2api"
  );
  const providerModels = Object.values(models).filter(
    (m) => m.provider === provider.name
@@ -221,7 +223,7 @@ function ProviderCard({
        )}
        {/* compact usage summary — sits left of the switch button,
            mirroring cc-switch's card layout (usage → action buttons) */}
-        {(provider.usageKinds?.length ?? 0) > 0 && (
+        {usage.supported && (
          <UsageFooter usage={usage} variant="compact" />
        )}
        <button
@@ -24,6 +24,15 @@ import {
  type ValidationErrorKey,
 } from "../lib/subagent-settings";
 import { Card, Toggle } from "./ui/controls";
+import { applyModelsDevSnapshot, reloadModelsDev } from "../lib/models-dev";
+
+/** Mirror of models_dev::ModelsDevStatus (camelCase over IPC). */
+interface ModelsDevStatus {
+  source: "synced" | "builtin";
+  lastUpdated: string;
+  modelCount: number;
+  providerCount: number;
+}

 interface SubagentSettingsPageProps {
  /** config.raw_other — hosts the `[experimental]` and `[secondary_model]` sections. */
@@ -34,12 +43,15 @@ interface SubagentSettingsPageProps {
  onBack: () => void;
 }

-// Upstream removed the "max" effort tier (auto-migrates to "high").
-const EFFORTS = ["low", "medium", "high"] as const;
+// Effort tiers the CLI accepts; the value is forwarded verbatim upstream, and
+// per-model support comes from `[models."<alias>"] support_efforts`.
+const EFFORTS = ["low", "medium", "high", "max", "xhigh"] as const;
 const EFFORT_LABELS: Record<(typeof EFFORTS)[number], TranslationKey> = {
  low: "thinkingLow",
  medium: "thinkingMedium",
  high: "thinkingHigh",
+  max: "thinkingMax",
+  xhigh: "thinkingXHigh",
 };

 const FLAG_LABELS: Record<string, { name: TranslationKey; desc: TranslationKey }> = {
@@ -155,13 +167,65 @@ export function SubagentSettingsPage({
  const [addSelection, setAddSelection] = useState("");
  /** WebUI-open button in flight; disables both buttons while non-null. */
  const [webuiBusy, setWebuiBusy] = useState<"embedded" | "browser" | null>(null);
+  /** models.dev snapshot provenance (synced copy vs bundled asset). */
+  const [modelsDevStatus, setModelsDevStatus] = useState<ModelsDevStatus | null>(null);
+  /** Sync/restore button in flight; blocks the card's buttons. */
+  const [modelsDevBusy, setModelsDevBusy] = useState(false);
+  /** Last sync error (inline) and success flag (auto-cleared on next action). */
+  const [syncError, setSyncError] = useState<string | null>(null);
+  const [syncOk, setSyncOk] = useState(false);

  useEffect(() => {
    invoke<ExperimentalEnvStatus>("get_experimental_env_status")
      .then(setEnv)
      .catch(() => setEnv({}));
+    invoke<ModelsDevStatus>("get_models_dev_status")
+      .then(setModelsDevStatus)
+      .catch(() => setModelsDevStatus(null));
  }, []);

+  const refreshModelsDevStatus = () => {
+    invoke<ModelsDevStatus>("get_models_dev_status")
+      .then(setModelsDevStatus)
+      .catch(() => {});
+  };
+
+  const handleModelsDevSync = async () => {
+    setModelsDevBusy(true);
+    setSyncError(null);
+    setSyncOk(false);
+    try {
+      const st = await invoke<ModelsDevStatus>("sync_models_dev");
+      setModelsDevStatus(st);
+      // Hot-swap the frontend index; the dashboard price index rebuilds on
+      // its next query (mtime-triggered on the Rust side).
+      const raw = await invoke<string | null>("get_models_dev_snapshot");
+      if (raw) applyModelsDevSnapshot(JSON.parse(raw) as Record<string, unknown>);
+      setSyncOk(true);
+    } catch (err) {
+      setSyncError(err instanceof Error ? err.message : String(err));
+    } finally {
+      setModelsDevBusy(false);
+    }
+  };
+
+  const handleModelsDevRestore = async () => {
+    if (!confirm(t("modelsDataRestoreConfirm"))) return;
+    setModelsDevBusy(true);
+    setSyncError(null);
+    setSyncOk(false);
+    try {
+      const st = await invoke<ModelsDevStatus>("reset_models_dev");
+      setModelsDevStatus(st);
+      await reloadModelsDev();
+    } catch (err) {
+      setSyncError(err instanceof Error ? err.message : String(err));
+      refreshModelsDevStatus();
+    } finally {
+      setModelsDevBusy(false);
+    }
+  };
+
  const flags = getExperimentalFlags(rawOther);
  const pool = getSubagentModelPool(rawOther);
  const poolKeys = pool ? Object.keys(pool.models) : [];
@@ -220,9 +284,8 @@ export function SubagentSettingsPage({
        ? `${Math.round(n / 1000)}K`
        : String(n);
  const effortLabel = (e: string): string => {
-    // 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])];
+    // Unknown values (hand-written config) are shown verbatim.
+    const key = EFFORT_LABELS[e as (typeof EFFORTS)[number]];
    return key ? t(key) : e;
  };

@@ -398,6 +461,66 @@ export function SubagentSettingsPage({
          </div>
        </Card>

+        {/* models.dev reference data: online sync / restore bundled */}
+        <Card title={t("modelsDataSection")}>
+          <p className="text-xs text-content-muted">{t("modelsDataDesc")}</p>
+          {modelsDevStatus && (
+            <div className="flex flex-wrap items-center gap-2 text-sm">
+              <span
+                className={`shrink-0 rounded border px-1.5 py-0.5 text-xs ${
+                  modelsDevStatus.source === "synced"
+                    ? "border-blue-500/30 text-blue-600 dark:text-blue-400"
+                    : "border-border text-content-muted"
+                }`}
+              >
+                {modelsDevStatus.source === "synced"
+                  ? t("modelsDataSourceSynced")
+                  : t("modelsDataSourceBuiltin")}
+              </span>
+              <span className="text-xs text-content-muted">
+                {t("modelsDataStats", {
+                  date: modelsDevStatus.lastUpdated,
+                  models: modelsDevStatus.modelCount.toLocaleString(),
+                  providers: modelsDevStatus.providerCount.toLocaleString(),
+                })}
+              </span>
+            </div>
+          )}
+          {syncOk && (
+            <div className="bg-green-100 dark:bg-green-900/20 border border-green-300 dark:border-green-500/30 rounded-lg px-3 py-2 text-xs text-green-700 dark:text-green-400">
+              {t("modelsDataSyncOk")}
+            </div>
+          )}
+          {syncError && (
+            <div
+              role="alert"
+              className="bg-red-100 dark:bg-red-900/20 border border-red-300 dark:border-red-500/30 rounded-lg px-3 py-2 text-xs text-red-700 dark:text-red-400"
+            >
+              {t("modelsDataSyncFailed")}: {syncError}
+            </div>
+          )}
+          <div className="flex flex-wrap gap-2">
+            <button
+              type="button"
+              disabled={modelsDevBusy}
+              onClick={handleModelsDevSync}
+              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"
+            >
+              {modelsDevBusy ? t("modelsDataSyncing") : t("modelsDataSyncNow")}
+            </button>
+            {modelsDevStatus?.source === "synced" && (
+              <button
+                type="button"
+                disabled={modelsDevBusy}
+                onClick={handleModelsDevRestore}
+                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("modelsDataRestore")}
+              </button>
+            )}
+          </div>
+        </Card>
+
        {/* Subagent: secondary model picker / pool + inherited settings */}
        <Card title={t("secondaryModelSection")}>
          {isPoolView && pool ? (
@@ -106,7 +106,7 @@ export function Segmented({
  onChange,
  disabled,
 }: {
-  options: { key: string; label: string }[];
+  options: { key: string; label: string; disabled?: boolean; title?: string }[];
  value: string;
  onChange: (value: string) => void;
  disabled?: boolean;
@@ -117,21 +117,27 @@ export function Segmented({
        disabled ? "opacity-50" : ""
      }`}
    >
-      {options.map((opt) => (
-        <button
-          key={opt.key}
-          type="button"
-          disabled={disabled}
-          onClick={() => onChange(opt.key)}
-          className={`px-3 py-1 text-sm ${
-            value === opt.key
-              ? "bg-blue-600 text-white"
-              : "bg-input text-content-muted hover:bg-hover-2"
-          }`}
-        >
-          {opt.label}
-        </button>
-      ))}
+      {options.map((opt) => {
+        const optDisabled = disabled || opt.disabled === true;
+        return (
+          // The title lives on the wrapper: a disabled button swallows pointer
+          // events in Chromium, so its own tooltip would never show.
+          <span key={opt.key} title={opt.title} className="inline-flex">
+            <button
+              type="button"
+              disabled={optDisabled}
+              onClick={() => onChange(opt.key)}
+              className={`px-3 py-1 text-sm ${
+                value === opt.key
+                  ? "bg-blue-600 text-white"
+                  : "bg-input text-content-muted hover:bg-hover-2"
+              } ${optDisabled ? "cursor-not-allowed opacity-40 hover:bg-input" : ""}`}
+            >
+              {opt.label}
+            </button>
+          </span>
+        );
+      })}
    </div>
  );
 }
@@ -1,5 +1,5 @@
 import { describe, expect, it } from "vitest";
-import { isAlertTier, type UsageData } from "./useUsageQuery";
+import { isAlertTier, usageQuerySupported, type UsageData } from "./useUsageQuery";

 /** 百分比窗口行（five_hour 等）；套餐类 tier 的 unit 恒为 "%"。 */
 function percentTier(used: number): UsageData {
@@ -42,3 +42,22 @@ describe("isAlertTier", () => {
    expect(isAlertTier({ planName: "five_hour", unit: "%", remaining: 20 }, 50)).toBe(false);
  });
 });
+
+describe("usageQuerySupported", () => {
+  it("有识别类型即可查询", () => {
+    expect(usageQuerySupported(["balance:deepseek"])).toBe(true);
+    expect(usageQuerySupported([], true)).toBe(true);
+  });
+
+  it("显式模板（newapi/sub2api）绕过 usage_kinds——自定义供应商也该查", () => {
+    expect(usageQuerySupported(undefined, true)).toBe(true);
+  });
+
+  it("无识别类型且非显式模板时不查询", () => {
+    expect(usageQuerySupported(undefined)).toBe(false);
+    expect(usageQuerySupported([])).toBe(false);
+    // auto 模板没有识别到类型 = 不支持（与弹窗短路口径一致）
+    expect(usageQuerySupported([], false)).toBe(false);
+    expect(usageQuerySupported(undefined, undefined)).toBe(false);
+  });
+});
@@ -75,15 +75,30 @@ export function isAlertTier(
  return d.unit === "%" && d.used != null && d.used >= threshold;
 }

+/**
+ * 是否值得为该供应商发起用量查询：有预设识别类型，或后端按显式模板
+ * （newapi / sub2api，二者绕过 usage_kinds）查询。
+ */
+export function usageQuerySupported(
+  usageKinds?: string[],
+  explicitTemplate?: boolean
+): boolean {
+  return (usageKinds?.length ?? 0) > 0 || explicitTemplate === true;
+}
+
 export function useUsageQuery(
  agent: Agent,
  providerName: string,
  usageKinds?: string[],
  autoIntervalMinutes?: number,
  disabled?: boolean,
-  threshold?: number
+  threshold?: number,
+  /** Backend resolves the query from an explicit template (newapi / sub2api)
+   * regardless of usageKinds — those bypass kinds entirely, so the row is
+   * supported even when the provider has no preset-detected kinds. */
+  explicitTemplate?: boolean
 ): UsageQueryState {
-  const supported = (usageKinds?.length ?? 0) > 0;
+  const supported = usageQuerySupported(usageKinds, explicitTemplate);
  // 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}`;
@@ -164,6 +164,15 @@ export const enTranslations: Record<TranslationKey, string> = {
  thinkingLow: "Low",
  thinkingMedium: "Medium",
  thinkingHigh: "High",
+  thinkingMax: "Max",
+  thinkingXHigh: "XHigh",
+  thinkingEffortUnsupported: "The current default model does not support thinking.",
+  thinkingEffortTierUnsupported:
+    "The current default model does not support this tier (supported: {levels}).",
+  thinkingEffortTierUndeclared:
+    "This model declares no thinking effort tiers; the upstream may reject this one.",
+  thinkingEffortDependsOnDefaultModel:
+    "The tier that actually applies depends on the current default model.",
  thinkingContextHint: "Thinking uses more context. Ensure the model context length and reserved size are sufficient.",
  loopControlSettings: "Loop Control",
  maxAttemptsPerStep: "Max attempts per step",
@@ -594,6 +603,21 @@ export const enTranslations: Record<TranslationKey, string> = {
  openWebUIBrowser: "Open in Browser",
  webuiOpening: "Opening...",

+  // models.dev reference data (context / capabilities / pricing)
+  modelsDataSection: "Model Reference Data (models.dev)",
+  modelsDataDesc:
+    "Model context limits, capability flags, and pricing come from an online models.dev snapshot. Syncing applies new models immediately — no app update needed.",
+  modelsDataSourceSynced: "Online sync",
+  modelsDataSourceBuiltin: "Bundled with release",
+  modelsDataStats: "{date} · {models} models / {providers} providers",
+  modelsDataSyncNow: "Sync Now",
+  modelsDataSyncing: "Syncing...",
+  modelsDataSyncOk: "Synced — model parameters and pricing refreshed",
+  modelsDataSyncFailed: "Sync failed",
+  modelsDataRestore: "Restore Built-in Data",
+  modelsDataRestoreConfirm:
+    "Delete the local synced copy and fall back to the snapshot bundled with this release? You can sync again at any time.",
+
  // Plugin marketplace
  pluginMarketplace: "Plugin Marketplace",
  pluginMarketplaceSubtitle:
@@ -161,6 +161,12 @@ export const zhTranslations = {
  thinkingLow: "低",
  thinkingMedium: "中",
  thinkingHigh: "高",
+  thinkingMax: "Max",
+  thinkingXHigh: "XHigh",
+  thinkingEffortUnsupported: "当前默认模型不支持思考。",
+  thinkingEffortTierUnsupported: "当前默认模型不支持该档位（支持：{levels}）。",
+  thinkingEffortTierUndeclared: "该模型未声明思考等级，上游可能拒绝此档位。",
+  thinkingEffortDependsOnDefaultModel: "实际生效的档位取决于当前默认模型。",
  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",
@@ -583,6 +589,21 @@ export const zhTranslations = {
  openWebUIBrowser: "在浏览器打开",
  webuiOpening: "打开中...",

+  // models.dev reference data (context / capabilities / pricing)
+  modelsDataSection: "模型参考数据（models.dev）",
+  modelsDataDesc:
+    "模型上下文长度、能力标记与单价参考来自 models.dev 在线快照。同步后新模型的参数与价格立即生效，无需等待新版本。",
+  modelsDataSourceSynced: "在线同步",
+  modelsDataSourceBuiltin: "随版本内置",
+  modelsDataStats: "{date} · {models} 模型 / {providers} 供应商",
+  modelsDataSyncNow: "立即同步",
+  modelsDataSyncing: "同步中...",
+  modelsDataSyncOk: "已同步，模型参数与单价已刷新",
+  modelsDataSyncFailed: "同步失败",
+  modelsDataRestore: "恢复内置数据",
+  modelsDataRestoreConfirm:
+    "删除本地同步副本并回退到随版本内置的快照数据？之后可随时重新同步。",
+
  // Plugin marketplace
  pluginMarketplace: "插件市场",
  pluginMarketplaceSubtitle:
@@ -20,7 +20,7 @@ function loopOf(raw: unknown): Record<string, unknown> {
 }

 // ---------------------------------------------------------------------------
-// Reading — legacy off values normalized to "off", "max" → "high"
+// Reading — legacy off values normalized to "off"; effort tiers pass through
 // ---------------------------------------------------------------------------

 describe("getAgentSettings — thinking.keep normalization", () => {
@@ -52,12 +52,29 @@ describe("getAgentSettings — thinking.keep normalization", () => {
  });
 });

-describe("getAgentSettings — effort \"max\" read mapping", () => {
-  it("normalizes a stored \"max\" to \"high\" on read", () => {
+describe("getAgentSettings — thinking.effort read mapping", () => {
+  it("keeps a stored \"max\" as-is (the tier is valid upstream)", () => {
    expect(getAgentSettings({ thinking: { effort: "max" } }).thinking?.effort).toBe(
-      "high"
+      "max"
    );
  });
+
+  it("passes an \"xhigh\" tier through untouched", () => {
+    expect(getAgentSettings({ thinking: { effort: "xhigh" } }).thinking?.effort).toBe(
+      "xhigh"
+    );
+  });
+
+  it("keeps the default \"medium\" when the key is absent", () => {
+    expect(getAgentSettings({}).thinking?.effort).toBe("medium");
+  });
+
+  it("round-trips an arbitrary hand-written tier on save", () => {
+    const next = setAgentSettings({ thinking: { effort: "ultra" } }, {
+      thinking: { effort: "ultra" },
+    });
+    expect(thinkingOf(next).effort).toBe("ultra");
+  });
 });

 // ---------------------------------------------------------------------------
@@ -65,11 +65,6 @@ export function getAgentSettings(rawOther: unknown): AgentSettings {
  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"
@@ -3,13 +3,18 @@
 * (see scripts/fetch-models-dev.mjs).
 *
 * The snapshot (~1.3 MB JSON) is NOT bundled into the main chunk anymore:
- * parsing a 1.3 MB JSON literal at startup blocks first paint. It is served
- * as a static asset (`/models-dev.json`) and loaded once in the background.
- * `getModelRef` stays synchronous and returns undefined until the index is
- * ready — callers already fall back to defaults — and `modelsDevReady()`
- * lets the app re-render once the index arrives.
+ * parsing a 1.3 MB JSON literal at startup blocks first paint. Loading order
+ * (see modelsDevReady): the runtime-synced copy under ~/.kimi-switch/
+ * (via models_dev.rs, written by the 高级设置 sync button) when present,
+ * otherwise the bundled static asset (`/models-dev.json`). `getModelRef`
+ * stays synchronous and returns undefined until the index is ready — callers
+ * already fall back to defaults — and `modelsDevReady()` lets the app
+ * re-render once the index arrives. `applyModelsDevSnapshot` hot-swaps the
+ * index after an online sync without a restart.
 */

+import { invoke } from "@tauri-apps/api/core";
+
 export interface ModelCost {
  input?: number;
  output?: number;
@@ -49,32 +54,64 @@ function buildIndex(raw: Record<string, unknown>) {
  return lower;
 }

+/**
+ * Swap in a new snapshot (initial load, online sync, or restore-to-builtin)
+ * and re-notify listeners so context/capability/price columns re-render.
+ * Listeners stay registered across reloads by design.
+ */
+function applySnapshot(raw: Record<string, unknown>) {
+  byLowerKey = buildIndex(raw);
+  for (const cb of readyListeners) cb();
+}
+
+async function fetchBundled(): Promise<Record<string, unknown>> {
+  const r = await fetch(`${import.meta.env.BASE_URL}models-dev.json`);
+  if (!r.ok) throw new Error(`models-dev.json HTTP ${r.status}`);
+  return (await r.json()) as Record<string, unknown>;
+}
+
 let loadPromise: Promise<void> | null = null;

 export function modelsDevReady(): Promise<void> {
  if (!loadPromise) {
-    loadPromise = fetch(`${import.meta.env.BASE_URL}models-dev.json`)
-      .then((r) => {
-        if (!r.ok) throw new Error(`models-dev.json HTTP ${r.status}`);
-        return r.json() as Promise<Record<string, unknown>>;
-      })
-      .then((raw) => {
-        byLowerKey = buildIndex(raw);
-        const listeners = readyListeners;
-        readyListeners = [];
-        for (const cb of listeners) cb();
-      })
-      .catch((err) => {
-        // A failed load is permanent for this session: fall back to defaults.
-        byLowerKey = {};
-        loadPromise = null; // allow one retry next time
-        console.warn("models-dev.json load failed:", err);
-      });
+    // Prefer the runtime-synced copy (~/.kimi-switch/models-dev.json, via
+    // models_dev.rs); outside Tauri (or without one) fall back to the bundled
+    // static asset. A failed load is permanent for this session.
+    loadPromise = (async () => {
+      let raw: Record<string, unknown> | null = null;
+      try {
+        const synced = await invoke<string | null>("get_models_dev_snapshot");
+        if (synced) raw = JSON.parse(synced) as Record<string, unknown>;
+      } catch {
+        /* not running inside Tauri — use the bundled asset */
+      }
+      applySnapshot(raw ?? (await fetchBundled()));
+    })().catch((err) => {
+      byLowerKey = {};
+      loadPromise = null; // allow one retry next time
+      console.warn("models-dev.json load failed:", err);
+    });
  }
  return loadPromise;
 }

-/** Register a callback invoked once the models.dev index is ready. */
+/**
+ * Apply a freshly synced snapshot (already fetched by the caller) and notify
+ * listeners so the UI refreshes without a restart.
+ */
+export function applyModelsDevSnapshot(raw: Record<string, unknown>): void {
+  applySnapshot(raw);
+}
+
+/**
+ * Reload from the bundled static asset (after "restore built-in data" —
+ * the synced copy was already deleted on the Rust side).
+ */
+export function reloadModelsDev(): Promise<void> {
+  return fetchBundled().then(applySnapshot);
+}
+
+/** Register a callback invoked whenever the models.dev index changes. */
 export function onModelsDevReady(cb: () => void): void {
  if (byLowerKey) {
    cb();
@@ -0,0 +1,139 @@
+import { describe, expect, it } from "vitest";
+import {
+  THINKING_EFFORTS,
+  normalizeSupportEfforts,
+  readSupportEfforts,
+  resolveThinkingEffortSupport,
+} from "./thinking-efforts";
+
+// ---------------------------------------------------------------------------
+// support_efforts normalization — an absent / malformed key means "unknown",
+// never "no tiers supported"
+// ---------------------------------------------------------------------------
+
+describe("normalizeSupportEfforts", () => {
+  it("keeps a non-empty list of strings", () => {
+    expect(normalizeSupportEfforts(["low", "high"])).toEqual(["low", "high"]);
+  });
+
+  it("drops blanks and non-strings but keeps the order", () => {
+    expect(normalizeSupportEfforts(["low", "", 3, null, "max"])).toEqual([
+      "low",
+      "max",
+    ]);
+  });
+
+  it("returns null for absent / empty / non-array values", () => {
+    expect(normalizeSupportEfforts(undefined)).toBeNull();
+    expect(normalizeSupportEfforts([])).toBeNull();
+    expect(normalizeSupportEfforts(["  "])).toBeNull();
+    expect(normalizeSupportEfforts("low")).toBeNull();
+  });
+});
+
+describe("readSupportEfforts", () => {
+  const model = (raw_other: unknown) =>
+    ({ alias: "a", provider: "p", model: "m", max_context_size: 0, display_name: null, raw_other }) as const;
+
+  it("reads the key from a model entry's preserved config fields", () => {
+    expect(readSupportEfforts(model({ support_efforts: ["low", "max"] }))).toEqual([
+      "low",
+      "max",
+    ]);
+  });
+
+  it("returns null when the entry or the key is absent", () => {
+    expect(readSupportEfforts(undefined)).toBeNull();
+    expect(readSupportEfforts(model(undefined))).toBeNull();
+    expect(readSupportEfforts(model({ other: 1 }))).toBeNull();
+  });
+});
+
+// ---------------------------------------------------------------------------
+// Tier resolution — three layers, in priority order
+// ---------------------------------------------------------------------------
+
+describe("resolveThinkingEffortSupport — declared support_efforts wins", () => {
+  const support = resolveThinkingEffortSupport({
+    supportEfforts: ["low", "high", "max"],
+    reasoning: true,
+  });
+
+  it("exposes the declared list and keeps thinking enabled", () => {
+    expect(support.declared).toEqual(["low", "high", "max"]);
+    expect(support.thinkingSupported).toBe(true);
+  });
+
+  it("enables exactly the declared tiers", () => {
+    for (const tier of THINKING_EFFORTS) {
+      expect(support.levels[tier].enabled).toBe(tier === "low" || tier === "high" || tier === "max");
+    }
+  });
+
+  it("explains a disabled tier with the supported list", () => {
+    expect(support.levels.medium).toEqual({
+      enabled: false,
+      note: { kind: "tierUnsupported", supported: ["low", "high", "max"] },
+    });
+    expect(support.levels.xhigh.enabled).toBe(false);
+  });
+
+  it("does not disable thinking even when models.dev says reasoning: false", () => {
+    const conflicting = resolveThinkingEffortSupport({
+      supportEfforts: ["high"],
+      reasoning: false,
+    });
+    expect(conflicting.thinkingSupported).toBe(true);
+    expect(conflicting.levels.low.enabled).toBe(false);
+    expect(conflicting.levels.high.enabled).toBe(true);
+  });
+
+  it("matches tiers case-insensitively", () => {
+    const upper = resolveThinkingEffortSupport({ supportEfforts: ["HIGH"] });
+    expect(upper.levels.high.enabled).toBe(true);
+  });
+});
+
+describe("resolveThinkingEffortSupport — model without thinking", () => {
+  const support = resolveThinkingEffortSupport({ reasoning: false });
+
+  it("disables the whole thinking area", () => {
+    expect(support.thinkingSupported).toBe(false);
+    expect(support.declared).toBeNull();
+    for (const tier of THINKING_EFFORTS) {
+      expect(support.levels[tier]).toEqual({
+        enabled: false,
+        note: { kind: "thinkingUnsupported" },
+      });
+    }
+  });
+});
+
+describe("resolveThinkingEffortSupport — nothing known about the model", () => {
+  it("offers low/medium/high plain and flags max/xhigh as undeclared", () => {
+    const support = resolveThinkingEffortSupport({});
+    expect(support.thinkingSupported).toBe(true);
+    expect(support.declared).toBeNull();
+    for (const tier of ["low", "medium", "high"] as const) {
+      expect(support.levels[tier]).toEqual({ enabled: true });
+    }
+    for (const tier of ["max", "xhigh"] as const) {
+      expect(support.levels[tier]).toEqual({
+        enabled: true,
+        note: { kind: "tierUndeclared" },
+      });
+    }
+  });
+
+  it("treats an undefined reasoning flag as unknown, not as unsupported", () => {
+    expect(
+      resolveThinkingEffortSupport({ reasoning: undefined }).thinkingSupported
+    ).toBe(true);
+  });
+
+  it("ignores a malformed support_efforts value and falls back to unknown", () => {
+    const support = resolveThinkingEffortSupport({ supportEfforts: "low" });
+    expect(support.declared).toBeNull();
+    expect(support.levels.low).toEqual({ enabled: true });
+  });
+});
@@ -0,0 +1,127 @@
+/**
+ * Thinking-effort capability resolution.
+ *
+ * kimi-code's `[thinking] effort` is a free-form string (low/medium/high/
+ * max/xhigh in practice) that the CLI forwards verbatim to OpenAI-compatible
+ * upstreams as `reasoning_effort`. Which tiers a given model accepts is
+ * declared per model in config.toml (`[models."<alias>"] support_efforts`);
+ * models.dev only carries a boolean `reasoning` flag with no tier list. The
+ * two sources are merged here so the UI can grey out tiers the current default
+ * model would reject, without ever rejecting a stored value on read.
+ */
+
+import type { Model } from "../types";
+
+export const THINKING_EFFORTS = [
+  "low",
+  "medium",
+  "high",
+  "max",
+  "xhigh",
+] as const;
+
+export type ThinkingEffort = (typeof THINKING_EFFORTS)[number];
+
+/** Why a tier is blocked, or (for undeclared models) merely unverified. */
+export type EffortNote =
+  | { kind: "thinkingUnsupported" }
+  | { kind: "tierUnsupported"; supported: string[] }
+  | { kind: "tierUndeclared" };
+
+export interface EffortAvailability {
+  /** Whether the tier can be picked. */
+  enabled: boolean;
+  /** Advisory message — set for blocked tiers and for unverified ones. */
+  note?: EffortNote;
+}
+
+export interface ThinkingEffortSupport {
+  /** False only when the model is known not to support thinking at all. */
+  thinkingSupported: boolean;
+  /** Declared tiers, or null when nothing is known about the model. */
+  declared: string[] | null;
+  levels: Record<ThinkingEffort, EffortAvailability>;
+}
+
+/** Tiers offered unverified when a model declares nothing at all. */
+const ALWAYS_OFFERED = new Set<string>(["low", "medium", "high"]);
+
+/**
+ * Normalize a `support_efforts` value into a comparable list, or null when the
+ * key is absent / malformed (treating it as "unknown", not "none supported").
+ */
+export function normalizeSupportEfforts(value: unknown): string[] | null {
+  if (!Array.isArray(value)) return null;
+  const tiers = value
+    .filter((v): v is string => typeof v === "string")
+    .map((v) => v.trim())
+    .filter((v) => v.length > 0);
+  return tiers.length > 0 ? tiers : null;
+}
+
+/** Read `support_efforts` from a model entry's preserved config.toml fields. */
+export function readSupportEfforts(model: Model | undefined): string[] | null {
+  if (!model) return null;
+  const raw = model.raw_other;
+  if (!raw || typeof raw !== "object" || Array.isArray(raw)) return null;
+  return normalizeSupportEfforts(
+    (raw as Record<string, unknown>).support_efforts
+  );
+}
+
+function buildLevels(
+  resolve: (effort: ThinkingEffort) => EffortAvailability
+): Record<ThinkingEffort, EffortAvailability> {
+  const levels = {} as Record<ThinkingEffort, EffortAvailability>;
+  for (const effort of THINKING_EFFORTS) levels[effort] = resolve(effort);
+  return levels;
+}
+
+/**
+ * Resolve per-tier availability from the two capability sources, in priority
+ * order:
+ *  1. `supportEfforts` declared → tiers follow the list exactly.
+ *  2. models.dev `reasoning === false` → the whole thinking area is off.
+ *  3. Nothing known → low/medium/high are offered plain, max/xhigh stay
+ *     selectable but carry a "not declared" note (upstream may reject them).
+ */
+export function resolveThinkingEffortSupport(input: {
+  supportEfforts?: unknown;
+  reasoning?: boolean;
+}): ThinkingEffortSupport {
+  const declared = normalizeSupportEfforts(input.supportEfforts);
+  if (declared) {
+    const supported = new Set(declared.map((t) => t.toLowerCase()));
+    return {
+      thinkingSupported: true,
+      declared,
+      levels: buildLevels((effort) =>
+        supported.has(effort)
+          ? { enabled: true }
+          : {
+              enabled: false,
+              note: { kind: "tierUnsupported", supported: declared },
+            }
+      ),
+    };
+  }
+  if (input.reasoning === false) {
+    return {
+      thinkingSupported: false,
+      declared: null,
+      levels: buildLevels(() => ({
+        enabled: false,
+        note: { kind: "thinkingUnsupported" },
+      })),
+    };
+  }
+  return {
+    thinkingSupported: true,
+    declared: null,
+    levels: buildLevels((effort) =>
+      ALWAYS_OFFERED.has(effort)
+        ? { enabled: true }
+        : { enabled: true, note: { kind: "tierUndeclared" } }
+    ),
+  };
+}
@@ -111,11 +111,12 @@ 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 tier, forwarded verbatim to OpenAI-compatible upstreams as
+   * `reasoning_effort`. Upstream accepts a free-form string; the UI offers
+   * low/medium/high/max/xhigh and derives per-model support from
+   * `[models."<alias>"] support_efforts`.
   */
-  effort?: "low" | "medium" | "high" | "max";
+  effort?: string;
  /**
   * Keep thinking content. The legacy off values (`false`, `0`, "no", "none",
   * `null`) are read-compatible — old configs may carry them; the UI
Author	SHA1	Message	Date
KimiSwitch Dev	26cb88ad0e	feat(thinking): 思考等级扩展五档（Max/XHigh）+ 按默认模型 support_efforts 能力感知；删除过时的 max→high 迁移；Segmented 支持逐选项禁用/tooltip — bump to v0.7.23 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	2026-10-02 13:00:35 +08:00
KimiSwitch Dev	50f59cac79	fix(release): publish-gitea.sh 兼容 docs/release-notes/release-notes-<tag>.md 命名（此前 Release 正文回退为 tag 名）	2026-10-02 13:00:33 +08:00
KimiSwitch Dev	ecc98fd8cd	chore(data): 同步 models-dev 快照至 2026-10-02（8,359 模型 / 225 供应商）	2026-10-02 13:00:32 +08:00
KimiSwitch Dev	13e15657af	fix(usage): 列表支持显式模板查询（Sub2API/NewAPI）——对齐后端绕过 usage_kinds 的口径 — bump to v0.7.22 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	2026-09-29 10:42:13 +08:00
KimiSwitch Dev	11286419b1	feat(models-dev): 高级设置新增 models.dev 在线同步 — 上下文/能力/单价即时生效 + 一键恢复内置；修复 public 快照构建漂移；确认适配 kimi-code 2.1.0 — bump to v0.7.21 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	2026-09-24 00:19:10 +08:00
KimiSwitch Dev	26e783c2ba	chore(data): 同步 models-dev 快照至 2026-09-20（7870 模型 / 222 供应商）	2026-09-20 10:29:40 +08:00
KimiSwitch Dev	11f87fd9af	chore(docs): release notes 归档 docs/release-notes/ + 根目录汇总 CHANGELOG.md - 20 份 release-notes-v*.md 移入 docs/release-notes/（git mv 保留历史） - 新增 CHANGELOG.md：v0.7.1~v0.7.20 倒序浓缩汇总（日期取各文件原始添加提交） - publish-gitea.sh 与 release.yml 改为 docs/release-notes/<tag>.md 优先、根目录回退	2026-09-20 09:58:01 +08:00