@eddyskywalker/dsh-chatgpt-subscription 0.37.0 → 0.38.0

package/CHANGELOG.md

@@ -2,6 +2,18 @@
 
 ## Unreleased
 
+## 0.38.0 - 2026-10-24
+- **[kimi-code] 新增 K3.1 Pro 系列（与其它模型同一 OpenAI 兼容 chat completions）、只支持思考的档位、以及缓存写入这个计费项**
+  - **现象**：`k3.1-pro` / `k3.1-pro-highspeed` 不在模型目录、设置卡与能力表里；「默认思考档位」永远列着 `low` / `high` / `max` 与「关闭思考」，而 K3.1 Pro 根本关不掉（服务端 400）；缓存区块只有命中率、命中 / 新处理 / 输出 tokens 与请求数，**写入**这个独立计费项一个数字都没有，看不出 5m / 1h 两档各写了多少。
+  - **模型目录**：两条新条目按官方型号表写入（`k3.1-pro` 默认 256K、可覆盖到 1M；`k3.1-pro-highspeed` 1M），含 `thinkingType: 'only'`、`defaultReasoningEffort: 'max'` 与 `quotaMultiplier: null`（官方未公布相对倍率，卡片显示「未公布」而不是猜成 1×）。**协议不新增分支**：本线路对全部模型只走 OpenAI 兼容 chat completions（与官方 agent 的 `type = "kimi"` 一致），因此 K3.1 Pro 的请求照常发往 `/v1/chat/completions`，视频输入按型号表声明可用。
+  - **thinking-only**：`types.ts` 增 `thinkingTypeFor` / `defaultReasoningEffortFor` / `coerceThinkingEffort`，`reasoningEffortsFor` 对 `only` 行剔除 `none`；`adapter.ts` 在构建 body 之前加一道闸门（THE THINKING GATE）：被禁的 `none` 抬到模型默认档并通过 `recordThinkingCoercion` 记录（`KIMI_CODE_THINKING_COERCED` 警告），而不是让整轮失败。
+  - **缓存写入**：总量读 `usage.prompt_tokens_details.cache_write_tokens`（chat）与 `cache_creation_input_tokens`（Anthropic），5m / 1h 分档读 `Msh-Usage-Cache-Write-Tokens-5m/-1h` 响应头；`KimiCodeCacheStatsDto` 的 `cacheWriteTokens` / `cacheWrite5mTokens` / `cacheWrite1hTokens` 由 `routes.ts` 填真值（原 `cacheWriteTokensNote: 'always-zero-on-this-route'` 占位删除）。chat wire 的新处理量改为 `total − cached − written`（三者互斥），Anthropic 的 `input_tokens` 仍不减——它本身不含缓存。
+  - **改法（客户端）**：新增 `src/client/kimi-code/labels.ts`——`wireLabel`、`effortLadder`（读模型声明的档位，`thinkingType === 'only'` 时剔除 `none`）、`canDisableThinking`、`effortOptions`（取已勾选模型的并集；「关闭思考」是开关不是档位，任一勾选模型允许它就保留），两张卡片都从这里取标签。档位下拉改由 `effortOptions` 驱动；存着的档位若所有勾选模型都不接受，仍保留该选项并标注「当前勾选的模型不接受」。只支持思考的模型在档位旁给一句说明（含模型名与 400），能力表新增「只支持思考」标记与脚注。缓存区块新增一行 `缓存写入 = 总量 · 5 分钟档 · 1 小时档`，**0 也照实显示**（空白会被读成「没有测量」）。
+  - **改法（三语）**：`kimi-code/locales.ts`（zh + en）与 `locales-ru/kimi-code.ts` 同步新增 `capThinkingOnly` / `capEfforts` / `capThinkingOnlyFootnote` / `effortThinkingOnlyNote` / `effortNotOffered` / `quotaUnknown` / `cacheWrite` / `cacheWrite5m` / `cacheWrite1h` / `cacheWriteHint` / `unknownMetric`，并按官方口径重写 `cacheTtlHint`、`cacheTtlDefault`、`defaultReasoningEffortHint`；俄语包键集由 `Record<KimiCodeLocaleKey, string>` 在编译期强制，缺一个键 `tsc` 就红。
+  - **测试**：新增 `test/kimi-code-k31.test.ts`（17 例：两条条目的全部字段、thinking-only 语义、双值 wire 与视频按声明保留、卡片契约字段）、`test/kimi-code-k31-card-ui.test.tsx`（8 例）、`test/kimi-code-effort-ladder.test.ts`（6 例）与 `test/kimi-code-cache-write.test.ts`（9 例：chat wire 的 usage 三分解、Anthropic 不减、响应头分档、以及 thinking 闸门在 chat / Anthropic 两条 wire 上的行为）；`test/kimi-code-capability-ui.test.tsx` 补「未公布倍率」与「只支持思考 + 400 脚注」两例。
+  - **验证**：`npm run typecheck` 三工程 0 错误；全量 `npm test` 281 文件通过 / 2 skipped，4633 passed / 8 skipped / 0 failed（删掉的用例全部只属于已撤除的 Responses 分支）。
+  - **文档**：README 的线路总览把 Kimi 的协议列改为「OpenAI 兼容 Chat Completions（可选 Anthropic Messages）」，模型目录由四款改为六款，补缓存写入那一行、倍率表里的两个「未公布」，并明确「本线路对全部模型使用 OpenAI 兼容 chat completions（与官方 agent 的 `type = "kimi"` 一致）」、视频清单加入 `k3.1-pro`，以及设置步骤里「默认思考档位跟随模型声明」。
+
 ## 0.37.0 - 2026-10-23
 - **[progress-updates] 进展更新补上「开局」与「收尾」两条硬要求**
   - **现象**：任务开始和结束都没有进展条目，整场只在被催到的时候冒出两三条；面板看着像「只有出结论才有动静」。

package/README.md

@@ -36,7 +36,7 @@
 | Provider ID | 订阅 | 协议 | 登录方式 | 模型目录来源 |
 | --- | --- | --- | --- | --- |
 | `codex-chatgpt` | ChatGPT（Plus / Pro / Business…） | Responses | 浏览器 OAuth（localhost:1455 回调），或导入本机 Codex CLI 登录 | **实时** `/backend-api/codex/models` |
-| `kimi-code` | Kimi 会员 | Anthropic Messages | **设备码**（RFC 8628） | 实时 `/v1/models` + 本地兜底 |
+| `kimi-code` | Kimi 会员 | OpenAI 兼容 Chat Completions（可选 Anthropic Messages） | **设备码**（RFC 8628） | 实时 `/v1/models` + 本地兜底 |
 | `command-code` | Command Code | Anthropic / OpenAI 双轨 | 浏览器 OAuth 或粘贴 API Key | 实时 `/provider/v1/models` |
 | `claude-subscription` | Claude Pro / Max | Anthropic Messages | 手动粘贴 / loopback 回调 | 实时 `/v1/models` + 本地兜底 |
 | `minimax-code` | MiniMax Code 编程订阅 | Anthropic Messages | **复用桌面端登录态** 或设备码 | **硬编码**（见下） |
@@ -119,13 +119,15 @@
   - 真正的 429 背压（`too many requests` / `engine is currently overloaded`）、连接失败与流停滞；
 - **不重试**，直接给出可操作提示（换模型 / 降上下文 / 等窗口重置 / 重新登录）：配额耗尽型的 429、403 的账号额度上限、401 里的套餐权限不足、400 请求格式错误；
   - 服务把这些含义压进**同一个状态码**，因此分类读响应正文而不只看状态码；
-- 模型目录为订阅侧的四款模型：`k3`（1M 上下文，需 Allegretto+；Moderato 上限 256K，故默认按 256K 计算，可用上下文覆盖升到 1M）、`k3-256k`、`kimi-for-coding`（K2.8 Preview）、`kimi-for-coding-highspeed`（约 6× 速度、3× 额度），运行时以 `GET /v1/models` 为准；
+- 模型目录为订阅侧的六款模型：`k3`（1M 上下文，需 Allegretto+；Moderato 上限 256K，故默认按 256K 计算，可用上下文覆盖升到 1M）、`k3-256k`、`kimi-for-coding`（K2.8 Preview）、`kimi-for-coding-highspeed`（约 6× 速度、3× 额度）、`k3.1-pro`（K3.1 Pro，默认 256K、可覆盖到 1M，需 Moderato 及以上，1M 需 Allegretto 及以上）与 `k3.1-pro-highspeed`（K3.1 Pro Highspeed，1M 上下文、约 2 倍配额，需 Pro / Allegretto 及以上），运行时以 `GET /v1/models` 为准；
 - **K3 行为按其官方文档实现**：
   - 思考档位只发 `low` / `high` / `max`（其余写法收敛映射，未知档位不发送）；关闭思考发 `thinking:{type:"disabled"}`，开启时发 `thinking:{type,effort,keep:"all"}`；
   - **开启思考时每条 assistant 消息都回传 `reasoning_content`**（无推理则回传空串——服务要求的是空值而非省略，否则 400）；
   - 不发送 `temperature`（采样参数按模型固定，显式值会报错）；
   - 工具调用 id 截断到 64 字符；
   - `stop` 按上限裁剪为最多 5 条、每条 ≤32 字节，超长整条丢弃（截断的停止串会在错误位置终止生成）；
+  - **K3.1 Pro 只支持思考模式**：关闭思考（`none`）会被服务端以 `400` 拒绝，所以它的档位只有 `low` / `high` / `max`、默认 `max`；「默认思考档位」下拉只列出勾选模型真正声明的档位，任一勾选模型允许关闭思考时它才出现，被拒时卡片上直接写明原因；
+- **本线路对全部模型使用 OpenAI 兼容 chat completions**（与官方 agent 的 `type = "kimi"` 一致）：K3.1 Pro 系列与其它模型同一协议、同一端点（`/v1/chat/completions`），Anthropic Messages 只在实时目录为某个 id 声明时启用；视频输入按官方型号表声明走，K3.1 Pro 系列可用；
 - **K3 的长思考不会被截断**：输出上限跟随上下文窗口（保留 4096 余量），因为 `reasoning_content` 计入输出，固定 32K 会把 `max` 档的长推理中途截断并返回 `length`；调用方已知 prompt 规模时上限会被下调到放得下，未知时不做猜测；
 - **请求体超 2 MB 本地即拒绝**：该端点最常见的 400 是 `total message size N exceeds limit 2097152`，官方文案不给建议，这里直接按真实序列化体积拦截，并按「循环实际做了什么」给出处方（有图可丢但丢了也补不上缺口 / 图已丢完仍超限 / 本来无图，三种情况文案不同，不再声称发生了没发生的省略）；确有实测上限的部署（前置代理/网关）可设 `DSH_KIMI_CODE_MAX_BODY_BYTES` 指定该档上限，取值非正整数时**回落到默认值而不是关闭守卫**；带视频的请求走 64 MiB 档，不受该变量影响；**超限时报的是可恢复的溢出码**（`CONTEXT_WINDOW_EXCEEDED`）：DSH 会对该码强制压缩本轮历史并重试同一步，所以长会话的表现是「压缩一次后继续」，而不是整轮失败——官方 `kimi-cli` 从不特殊处理 `2097152`，它靠 85% 提前压缩避开，详见下文「上下文窗口」那条；
 - **缓存是自动的，且无法手动干预**：
@@ -134,8 +136,9 @@
   - 真正决定命中率的是**内容稳定性**：同一会话内 `system` 或工具列表一旦变化会使整个前缀缓存失效（实测归零）。因此应保持工具集合稳定、把新增内容追加在末尾；
   - 卡片会显示滚动命中率，方便验证效果；
   - **写入时长可手动选择**（设置页）：`prompt_cache_options`（OpenAI 兼容线路）或**顶层** `cache_control`（Anthropic 兼容线路）控制缓存**写入时长**，与上面的「无法干预」不矛盾——那一条说的是**缓存身份**（由内容前缀哈希决定），TTL 管的是另一回事；
-  - 档位在**首次写入时锁定**，之后无法改写、命中时按原档免费续期；1 小时写入约为 5 分钟的两倍价，只有同一前缀会在一小时内被反复读取才划算。不选择时不发任何缓存字段；
-- **视频输入可用**（`k3`、`kimi-for-coding`）：
+  - 档位在**首次写入时锁定**、之后无法改写，命中时按原 TTL 续期（官方口径：不传时按 **5 分钟档**写入，只有 5m / 1h 两档）；1 小时写入约为 5 分钟的两倍价，只有同一前缀会在一小时内被反复读取才划算。不选择时不发任何缓存字段；
+  - 设置卡的缓存区块把**写入**单独列成一行（总量 + 5 分钟档 + 1 小时档）：写入是独立计费项，折进「新处理 tokens」会让命中率说错话；某个分档是 0 时照实显示 0——响应头缺失只表示服务端没回报该档，不等于那里确实没写；
+- **视频输入可用**（`k3`、`kimi-for-coding`、`k3.1-pro`）：
   - DSH 的模态词表只有 `text`/`image`，但它是可合并扩展的接口，本插件用 TypeScript 模块增强把它扩到 `video`（**未改动 DSH 任何代码**），因此视频走 DSH 真实的能力通道，而不是只能显示在提示里；
   - 适配器把视频映射为服务文档的 `{type:'video_url',video_url:{url:'data:…'}}`。视频与图片**各有独立预算**（图片 1.5 MB、视频 48 MiB base64，最旧优先省略），请求体校验只在确实带视频时才放宽到 64 MiB；
   - 图片预算可用 `DSH_KIMI_CODE_MAX_IMAGE_BYTES` 覆盖（取值非正整数时回落到默认值）。视频预算**没有**环境变量出口，与 MiniMax 线路一致；
@@ -545,9 +548,9 @@ DSH 设置页的「Subagent」卡片会把勾选的模型写成会话级的允
 
 `kimi-code` 会像其他 Provider 一样出现在 DSH 模型选择器中。**上下文窗口**默认取模型目录值，可逐模型覆盖（用于 DSH 的压缩与溢出判断，支持 `1M` / `256K` / `200000` 等写法）——注意 `k3` 的 1M 上下文需要 Allegretto 及以上套餐，因此默认按 256K 计算，升级后再在此覆盖为 1M。
 
-**申报的窗口默认就是目录值或你的覆盖值，不做折算**——`k3` 覆盖成 1M 就按 1M 申报。撞上 2 MB 请求体上限时，本地守卫抛的是 `CONTEXT_WINDOW_EXCEEDED`：DSH 会对该码**自动压缩本轮历史并重试同一步**，所以表现是「压缩一次后继续」，不是整轮失败。想要「撞墙之前就自动压缩」，可设 `DSH_KIMI_CODE_BYTES_PER_TOKEN` 启用窗口折算：申报值 = `min(声明值, floor(2MB / d))`。取值按内容密度选——DSH 的 token 计量是固定密度（4 字符/token，见 `@deepseek-ai/dsh-token-meter`），看不见中文占 3 字节，所以「一个计量 token 占多少字节」= 4 × 每字符字节数：纯 ASCII 约 4、中英混合约 8、密集中文约 12；对应的字节墙分别是 524288 / 262144 / 174763 计量 token。实测：`d=4` → 524288（代码为主的会话在 1.68 MB 处压缩）、`d=8` → 262144（混合）、`d=12` → 174762（中文，三种密度都在墙前）。压力阈值是 `min(floor(0.8 × 窗口), 窗口 − 65536)`（`compaction-basic`：`thresholdRatio` 默认 0.8、`headroomTokens` 默认 65536、本线路不申报 `defaultMaxTokens` 故无输出预留）。该变量是**可选的提前压缩开关**而不是安全守卫：不设或拼错等于不启用，真正兜底的是上面那个恢复码。**默认思考档位**只提供 `low` / `high` / `max` 三档与「关闭思考」（服务对这三档以外的取值直接报 400）；切换模型或切换档位都会使上下文缓存失效，建议在同一会话内保持一致。
+**申报的窗口默认就是目录值或你的覆盖值，不做折算**——`k3` 覆盖成 1M 就按 1M 申报。撞上 2 MB 请求体上限时，本地守卫抛的是 `CONTEXT_WINDOW_EXCEEDED`：DSH 会对该码**自动压缩本轮历史并重试同一步**，所以表现是「压缩一次后继续」，不是整轮失败。想要「撞墙之前就自动压缩」，可设 `DSH_KIMI_CODE_BYTES_PER_TOKEN` 启用窗口折算：申报值 = `min(声明值, floor(2MB / d))`。取值按内容密度选——DSH 的 token 计量是固定密度（4 字符/token，见 `@deepseek-ai/dsh-token-meter`），看不见中文占 3 字节，所以「一个计量 token 占多少字节」= 4 × 每字符字节数：纯 ASCII 约 4、中英混合约 8、密集中文约 12；对应的字节墙分别是 524288 / 262144 / 174763 计量 token。实测：`d=4` → 524288（代码为主的会话在 1.68 MB 处压缩）、`d=8` → 262144（混合）、`d=12` → 174762（中文，三种密度都在墙前）。压力阈值是 `min(floor(0.8 × 窗口), 窗口 − 65536)`（`compaction-basic`：`thresholdRatio` 默认 0.8、`headroomTokens` 默认 65536、本线路不申报 `defaultMaxTokens` 故无输出预留）。该变量是**可选的提前压缩开关**而不是安全守卫：不设或拼错等于不启用，真正兜底的是上面那个恢复码。**默认思考档位**的选项跟随勾选模型自己声明的档位——`low` / `high` / `max`，以及在允许关闭思考的模型上才出现的「关闭思考」（服务对这些以外的取值直接报 400）；K3.1 Pro 系列只支持思考模式，硬关会被拒（400），所以勾选它时这一档不出现，卡片上会写明原因；切换模型或切换档位都会使上下文缓存失效，建议在同一会话内保持一致。
 
-**额度倍率**取自官方模型表并显示在设置页模型能力表的「说明」列（悬停可一并看到）：`k3` = **2×**、`k3-256k` = **1×**、`kimi-for-coding` = **1×**、`kimi-for-coding-highspeed` = **3×**。倍率绑定**模型 id**，与上下文长度、缓存命中率都无关——选 `k3` 就是 2×，哪怕只用几千 token。注册表不认识的新模型 id 显示为**未知**而不是猜成 1×。
+**额度倍率**取自官方模型表并显示在设置页模型能力表的「说明」列（悬停可一并看到）：`k3` = **2×**、`k3-256k` = **1×**、`kimi-for-coding` = **1×**、`kimi-for-coding-highspeed` = **3×**；`k3.1-pro` 与 `k3.1-pro-highspeed` 官方未公布相对倍率（后者只说了「约为常规版 2 倍」），因此显示为**未知**而不是猜成 2×。倍率绑定**模型 id**，与上下文长度、缓存命中率都无关——选 `k3` 就是 2×，哪怕只用几千 token。注册表不认识的新模型 id 显示为**未知**而不是猜成 1×。
 
 **区域**：默认使用中国大陆主机（`auth.kimi.com` / `api.kimi.com`）。国际账号可设置 `DSH_KIMI_CODE_OAUTH_HOST=https://auth.kimi.ai` 与 `DSH_KIMI_CODE_BASE_URL=https://api.kimi.ai/coding` 后重新登录；环境变量同时会钉住区域，设置页会显示当前解析到的主机。
 

package/lib/client.js

@@ -3691,7 +3691,7 @@ window.__ModuleLoader__.load({
 		* exactly one and rejects the others, so the label states which one this model
 		* actually uses.
 		*/
-		function wireLabel(wire, t) {
+		function wireLabel$1(wire, t) {
 			if (wire === "anthropic") return t.wireAnthropic;
 			if (wire === "responses") return t.wireResponses;
 			return t.wireOpenai;
@@ -4120,7 +4120,7 @@ window.__ModuleLoader__.load({
 								items: (status?.models ?? []).map((model) => ({
 									id: model.id,
 									name: model.name,
-									hint: wireLabel(model.wire, t),
+									hint: wireLabel$1(model.wire, t),
 									enabled: model.enabled
 								})),
 								busy: busy !== null,
@@ -4277,15 +4277,6 @@ window.__ModuleLoader__.load({
 			});
 		}
 		//#endregion
-		//#region src/shared/kimi-code-contracts.ts
-		/** Every level, in escalating order; the settings card renders exactly these. */
-		const KIMI_CODE_REASONING_EFFORTS = [
-			"low",
-			"high",
-			"max",
-			"none"
-		];
-		//#endregion
 		//#region src/client/kimi-code/locales.ts
 		const NS_KIMI_CODE = "dsh-kimi-code";
 		const zh$10 = {
@@ -4336,10 +4327,13 @@ window.__ModuleLoader__.load({
 			wireAnthropic: "Anthropic",
 			wireOpenai: "OpenAI",
 			capVideo: "视频*",
+			capThinkingOnly: "只支持思考",
+			capEfforts: "档位 {list}",
 			capDynamicTools: "动态工具",
 			capImageOnly: "仅图片",
 			capNone: "—",
 			capVideoFootnote: "* 视频输入需先由 kimi_attach_video 工具挂载（本地绝对路径或 http(s) 链接）；直接粘贴的上传路径仍只支持图片。",
+			capThinkingOnlyFootnote: "* 「只支持思考」的模型不能关闭思考：请求里带 none 会被服务端拒绝（400），所以它的档位里没有「关闭思考」。",
 			capabilitiesTitle: "模型能力",
 			capColMedia: "多模态",
 			capColTools: "动态工具",
@@ -4350,13 +4344,16 @@ window.__ModuleLoader__.load({
 			speedRegular: "标准",
 			speedHighspeed: "高速 (约 6×)",
 			quotaCost: "额度倍率",
+			quotaUnknown: "未公布",
 			minimumPlan: "所需套餐",
 			enhanced: "思考设置",
 			defaultReasoningEffort: "默认思考档位",
-			defaultReasoningEffortHint: "Kimi Code 只接受 low / high / max 三档；none 表示关闭思考。切换档位会使上下文缓存失效，建议在同一会话中保持一致。",
+			defaultReasoningEffortHint: "可用档位跟随勾选的模型自己声明的能力（一个都没勾时按 low / high / max / none 全列）；none 表示关闭思考，只对允许关闭的模型有效。切换档位会使上下文缓存失效，建议在同一会话中保持一致。",
+			effortThinkingOnlyNote: "勾选的模型里有只支持思考模式的：{models}。它们没有「关闭思考」这一档，请求关闭思考会被服务端拒绝（400），所以这里不提供。",
+			effortNotOffered: "当前勾选的模型不接受",
 			cacheTtl: "上下文缓存时长",
-			cacheTtlHint: "Kimi 按前缀自动缓存，这里选择写入时长。1h 写入约为 5m 的两倍价，只有当同一前缀会在一小时内被反复读取才划算；档位在首次写入时锁定，之后无法改写。留空表示不指定，由 Kimi 自己决定（实测有效期约 5–30 分钟）。依据回合间隔选：间隔通常短于约 30 分钟就留空或选 5 分钟；调度、等子代理这类会出现长空闲的会话选 1 小时。",
-			cacheTtlDefault: "不指定（Kimi 默认）",
+			cacheTtlHint: "Kimi 按前缀自动缓存，这里选择写入档位（只有 5m / 1h 两档）。不传时服务端按 5 分钟档写入。1h 写入单价约为 5m 的两倍，只有当同一前缀会在一小时内被反复读取才划算。档位在首次写入时锁定，之后命中按原 TTL 续期，无法改写。依据回合间隔选：间隔通常短于约 30 分钟选 5 分钟；调度、等子代理这类会出现长空闲的会话选 1 小时。",
+			cacheTtlDefault: "不传（服务端按 5 分钟档写入）",
 			cacheTtl5m: "5 分钟",
 			cacheTtl1h: "1 小时",
 			defaultEffortAuto: "使用模型默认值",
@@ -4382,6 +4379,11 @@ window.__ModuleLoader__.load({
 			cacheFresh: "新处理 tokens",
 			cacheOutput: "输出 tokens",
 			cacheRequests: "请求数",
+			cacheWrite: "缓存写入",
+			cacheWrite5m: "5 分钟档",
+			cacheWrite1h: "1 小时档",
+			cacheWriteHint: "写入是独立计费项：1h 写入单价约为 5m 的两倍，命中后会按原 TTL 续期。分档显示 0 表示服务端没有回报该档的写入数（响应头缺失就是静默），不等于该档确实没写入。",
+			unknownMetric: "暂无",
 			cacheHint: "同一会话内：系统提示与工具列表一旦变化会使缓存整体失效，因此保持工具集合稳定、把新增内容追加在末尾，命中率最高。",
 			cacheExpired: "本会话已空闲超过缓存有效期，下一轮将重新处理约 {tokens} 个 token。",
 			cacheDrift: "最近前缀状态",
@@ -4459,10 +4461,13 @@ window.__ModuleLoader__.load({
 			wireAnthropic: "Anthropic",
 			wireOpenai: "OpenAI",
 			capVideo: "Video*",
+			capThinkingOnly: "Thinking only",
+			capEfforts: "Levels {list}",
 			capDynamicTools: "Dynamic tools",
 			capImageOnly: "Images",
 			capNone: "—",
 			capVideoFootnote: "* Video input is attached with the kimi_attach_video tool (an absolute local path or an http(s) URL); the paste/upload path still accepts images only.",
+			capThinkingOnlyFootnote: "* A \"thinking only\" model cannot switch thinking off: a request carrying none is rejected by the service with HTTP 400, so \"thinking off\" is not among its levels.",
 			capabilitiesTitle: "Capabilities",
 			capColMedia: "Multimodal",
 			capColTools: "Dynamic tools",
@@ -4473,13 +4478,16 @@ window.__ModuleLoader__.load({
 			speedRegular: "Regular",
 			speedHighspeed: "HighSpeed (~6x)",
 			quotaCost: "Quota cost",
+			quotaUnknown: "Not published",
 			minimumPlan: "Required plan",
 			enhanced: "Thinking",
 			defaultReasoningEffort: "Default thinking level",
-			defaultReasoningEffortHint: "Kimi Code accepts only low / high / max; none disables thinking. Switching the level invalidates the context cache, so keep it steady within a session.",
+			defaultReasoningEffortHint: "The levels offered follow what the checked models declare (all of low / high / max / none when none is checked); none disables thinking and only applies to models that allow it. Switching the level invalidates the context cache, so keep it steady within a session.",
+			effortThinkingOnlyNote: "Some checked models think at every level: {models}. They have no \"thinking off\" — the service rejects that request with HTTP 400 — so it is not offered here.",
+			effortNotOffered: "not accepted by the checked models",
 			cacheTtl: "Context cache duration",
-			cacheTtlHint: "Kimi caches prefixes automatically; this chooses the write duration. A 1h write costs about twice a 5m one and only pays off when the same prefix is re-read within the hour. The tier is locked when first written and cannot be changed afterwards. Leave unset to send no cache field and let Kimi decide (measured lifetime is roughly 5-30 minutes). Choose by the gap between turns: leave unset or use 5 minutes when turns are normally less than about 30 minutes apart, and pick 1 hour for a session that idles for long stretches, such as dispatch work waiting on subagents.",
-			cacheTtlDefault: "Unset (Kimi default)",
+			cacheTtlHint: "Kimi caches prefixes automatically; this chooses the write tier, and there are only two (5m and 1h). Unset sends no cache field, and the service then writes at the 5-minute tier. A 1h write costs about twice a 5m one and only pays off when the same prefix is re-read within the hour. The tier is locked at the first write, and later hits renew it under that original TTL. Choose by the gap between turns: 5 minutes when turns are normally less than about 30 minutes apart, 1 hour for a session that idles for long stretches, such as dispatch work waiting on subagents.",
+			cacheTtlDefault: "Unset (service writes at the 5-minute tier)",
 			cacheTtl5m: "5 minutes",
 			cacheTtl1h: "1 hour",
 			defaultEffortAuto: "Use the model default",
@@ -4506,6 +4514,11 @@ window.__ModuleLoader__.load({
 			cacheFresh: "Fresh tokens",
 			cacheOutput: "Output tokens",
 			cacheRequests: "Requests",
+			cacheWrite: "Cache writes",
+			cacheWrite5m: "5-minute tier",
+			cacheWrite1h: "1-hour tier",
+			cacheWriteHint: "A write is billed separately: a 1h write costs about twice a 5m one, and a hit renews the entry under its original TTL. A tier showing 0 means the service did not report a write for it (a missing response header is silence), not that nothing was written there.",
+			unknownMetric: "n/a",
 			cacheHint: "Within one session: any change to the system prompt or the tool list invalidates the whole prefix cache, so keep the tool set stable and append new material at the tail for the best hit rate.",
 			cacheDrift: "Last prefix state",
 			cacheDriftFirst: "First request (cache cold)",
@@ -4541,6 +4554,95 @@ window.__ModuleLoader__.load({
 			"en-US": en$10
 		};
 		//#endregion
+		//#region src/shared/kimi-code-contracts.ts
+		/** Every level, in escalating order; the settings card renders exactly these. */
+		const KIMI_CODE_REASONING_EFFORTS = [
+			"low",
+			"high",
+			"max",
+			"none"
+		];
+		//#endregion
+		//#region src/client/kimi-code/labels.ts
+		/**
+		* Name of the request/response dialect a model is served over.
+		*
+		* @param wire - dialect the adapter reports for this model.
+		* @param t - dictionary of the card rendering it.
+		*/
+		function wireLabel(wire, t) {
+			return wire === "anthropic" ? t.wireAnthropic : t.wireOpenai;
+		}
+		/** Human label for one thinking level. */
+		function effortLabel(effort, t) {
+			switch (effort) {
+				case "low": return t.effortLow;
+				case "high": return t.effortHigh;
+				case "max": return t.effortMax;
+				case "none": return t.effortNone;
+				default: return effort;
+			}
+		}
+		/**
+		* Thinking levels one model publishes, in the order the card renders them.
+		*
+		* The ladder is read, never assumed: a model that declares no levels gets an
+		* empty list, and the picker then offers none of them rather than the generic
+		* four. `none` is dropped for a thinking-only model — the host already strips
+		* it, and filtering again is what keeps an older host, or a live catalog that
+		* says nothing about thinking, from putting "thinking off" back on a menu whose
+		* request the service answers with HTTP 400.
+		*
+		* @param model - one entry from the model directory.
+		*/
+		function effortLadder(model) {
+			const declared = model.reasoningEfforts ?? [];
+			const known = KIMI_CODE_REASONING_EFFORTS.filter((effort) => declared.includes(effort));
+			return model.thinkingType === "only" ? known.filter((effort) => effort !== "none") : known;
+		}
+		/**
+		* Whether one model may be asked to stop reasoning.
+		*
+		* Read from the model's own declaration rather than assumed: `only` is the one
+		* value that makes the question a hard no — the service answers a
+		* disabled-thinking request with HTTP 400 and names the model — while `no`
+		* describes a model that never reasons, so there is no switch to offer either.
+		* `both`, and a registry that says nothing about the id, keep the switch,
+		* because that is what this line offered before the flag existed and the
+		* service, not the card, is the authority on refusing a level.
+		*
+		* @param model - one entry from the model directory.
+		*/
+		function canDisableThinking(model) {
+			return model.thinkingType !== "only" && model.thinkingType !== "no";
+		}
+		/**
+		* Thinking levels a conversation can actually be set to.
+		*
+		* The setting is one level for the whole line, so the three thinking levels are
+		* the union of what the CHECKED models declare: a level no checked model
+		* declares would be stored and then rejected on the first turn. "Thinking off"
+		* is not one of those declared levels on most rows — it is a switch, not a
+		* ladder rung — so it is decided by {@link canDisableThinking} instead, and it
+		* disappears exactly when every checked model refuses to be switched off.
+		*
+		* With nothing checked there is no model to narrow it down and the full ladder
+		* is offered.
+		*
+		* @param models - every model in the directory, checked or not.
+		*/
+		function effortOptions(models) {
+			const enabled = models.filter((model) => model.enabled);
+			if (enabled.length === 0) return [...KIMI_CODE_REASONING_EFFORTS];
+			const declared = new Set(enabled.flatMap((model) => effortLadder(model)));
+			const switchable = enabled.some((model) => canDisableThinking(model));
+			return KIMI_CODE_REASONING_EFFORTS.filter((effort) => effort === "none" ? switchable : declared.has(effort));
+		}
+		/** Models that reason at every level, and so have no "thinking off". */
+		function thinkingOnlyModels(models) {
+			return models.filter((model) => model.thinkingType === "only");
+		}
+		//#endregion
 		//#region src/client/kimi-code/KimiModelCapabilities.tsx
 		/**
 		* Capability table for the Kimi model list.
@@ -4564,54 +4666,71 @@ window.__ModuleLoader__.load({
 			const t = useLineDict(NS_KIMI_CODE, zh$10);
 			return /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
 				className: "dsha-cap-table",
-				children: [/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("table", { children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("thead", { children: /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("tr", { children: [
-					/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t.capabilitiesTitle }),
-					/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t.capColMedia }),
-					/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t.capColTools }),
-					/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t.capColNotes })
-				] }) }), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("tbody", { children: models.map((model) => {
-					return /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("tr", { children: [
-						/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", {
-							title: [
-								model.id,
-								model.wire === "anthropic" ? t.wireAnthropic : t.wireOpenai,
-								...model.defaultReasoningEffort ? [t.defaultReasoningEffort + ": " + model.defaultReasoningEffort] : [],
-								...model.maxInputTokens === void 0 ? [] : [t.capInputCap.replace("{tokens}", String(model.maxInputTokens))],
-								...model.minimumPlan ? [t.capPlan.replace("{plan}", model.minimumPlan)] : [],
-								...model.quotaMultiplier === null ? [] : [`${t.quotaCost}: ${formatMultiplier(model.quotaMultiplier)}`]
-							].join(" · "),
-							children: model.name
-						}),
-						/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", { children: /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
-							className: model.supportsVideo ? "dsha-cap-on" : "dsha-cap-off",
-							children: model.supportsVideo ? t.capVideo : t.capImageOnly
-						}) }),
-						/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", { children: model.supportsDynamicTools ? /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
-							className: "dsha-cap-on",
-							children: t.capDynamicTools
-						}) : model.supportsToolUse ? /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
-							className: "dsha-cap-off",
-							children: t.capNone
-						}) : /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
-							className: "dsha-cap-off",
-							children: t.capNoTools
-						}) }),
-						/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("td", {
-							className: "dsha-cap-notes",
-							children: [model.quotaMultiplier === null ? null : /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("span", {
-								className: "dsha-cap-cost",
-								children: [
-									t.quotaCost,
-									" ",
-									formatMultiplier(model.quotaMultiplier)
-								]
-							}), model.description ?? ""]
-						})
-					] }, model.id);
-				}) })] }), models.some((model) => model.supportsVideo) && /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
-					className: "dsha-muted dsha-cap-footnote",
-					children: t.capVideoFootnote
-				})]
+				children: [
+					/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("table", { children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("thead", { children: /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("tr", { children: [
+						/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t.capabilitiesTitle }),
+						/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t.capColMedia }),
+						/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t.capColTools }),
+						/* @__PURE__ */ (0, react_jsx_runtime.jsx)("th", { children: t.capColNotes })
+					] }) }), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("tbody", { children: models.map((model) => {
+						const ladder = effortLadder(model);
+						return /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("tr", { children: [
+							/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", {
+								title: [
+									model.id,
+									wireLabel(model.wire, t),
+									...model.thinkingType === "only" ? [t.capThinkingOnly] : [],
+									...ladder.length === 0 ? [] : [t.capEfforts.replace("{list}", ladder.join(" / "))],
+									...model.defaultReasoningEffort ? [t.defaultReasoningEffort + ": " + model.defaultReasoningEffort] : [],
+									...model.maxInputTokens === void 0 ? [] : [t.capInputCap.replace("{tokens}", String(model.maxInputTokens))],
+									...model.minimumPlan ? [t.capPlan.replace("{plan}", model.minimumPlan)] : [],
+									...model.quotaMultiplier === null ? [] : [`${t.quotaCost}: ${formatMultiplier(model.quotaMultiplier)}`]
+								].join(" · "),
+								children: model.name
+							}),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", { children: /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+								className: model.supportsVideo ? "dsha-cap-on" : "dsha-cap-off",
+								children: model.supportsVideo ? t.capVideo : t.capImageOnly
+							}) }),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsx)("td", { children: model.supportsDynamicTools ? /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+								className: "dsha-cap-on",
+								children: t.capDynamicTools
+							}) : model.supportsToolUse ? /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+								className: "dsha-cap-off",
+								children: t.capNone
+							}) : /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+								className: "dsha-cap-off",
+								children: t.capNoTools
+							}) }),
+							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("td", {
+								className: "dsha-cap-notes",
+								children: [model.quotaMultiplier === null ? /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+									className: "dsha-cap-off",
+									children: [
+										t.quotaCost,
+										" ",
+										t.quotaUnknown
+									]
+								}) : /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("span", {
+									className: "dsha-cap-cost",
+									children: [
+										t.quotaCost,
+										" ",
+										formatMultiplier(model.quotaMultiplier)
+									]
+								}), model.description ?? ""]
+							})
+						] }, model.id);
+					}) })] }),
+					models.some((model) => model.supportsVideo) && /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
+						className: "dsha-muted dsha-cap-footnote",
+						children: t.capVideoFootnote
+					}),
+					models.some((model) => model.thinkingType === "only") && /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
+						className: "dsha-muted dsha-cap-footnote",
+						children: t.capThinkingOnlyFootnote
+					})
+				]
 			});
 		}
 		//#endregion
@@ -4630,14 +4749,15 @@ window.__ModuleLoader__.load({
 				return `${(cents / 100).toFixed(2)} ${code}`;
 			}
 		}
-		function effortLabel(effort, t) {
-			switch (effort) {
-				case "low": return t.effortLow;
-				case "high": return t.effortHigh;
-				case "max": return t.effortMax;
-				case "none": return t.effortNone;
-				default: return effort;
-			}
+		/**
+		* One token count as the card prints it.
+		*
+		* A reported zero prints as "0": zero writes is a reading, while a blank cell
+		* reads as "nothing was measured", and the two are different claims. Only a
+		* number this build cannot read at all falls back to "暂无".
+		*/
+		function tokenCount(value, t) {
+			return typeof value === "number" && Number.isFinite(value) ? value.toLocaleString() : t.unknownMetric;
 		}
 		/** Human label for the last request's prefix-drift cause. */
 		function driftLabel(cause, t) {
@@ -4984,8 +5104,13 @@ window.__ModuleLoader__.load({
 			});
 			const quota = status?.quota;
 			status?.account;
-			status?.models.filter((model) => model.enabled).length;
+			const models = status?.models ?? [];
+			models.filter((model) => model.enabled).length;
 			const loginPending = flow?.status === "pending";
+			const offeredEfforts = effortOptions(models);
+			const savedEffort = status?.defaultReasoningEffort ?? null;
+			const staleEffort = savedEffort !== null && !offeredEfforts.includes(savedEffort) ? savedEffort : null;
+			const thinkingOnly = thinkingOnlyModels(models.filter((model) => model.enabled));
 			return /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
 				className: "dsha-page",
 				children: [
@@ -5181,6 +5306,30 @@ window.__ModuleLoader__.load({
 											children: driftLabel(status.cache.lastDrift, t)
 										})]
 									}),
+									/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
+										className: "dsha-row",
+										children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", {
+											className: "dsha-label",
+											children: t.cacheWrite
+										}), /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("span", {
+											className: "dsha-value",
+											children: [
+												tokenCount(status.cache.cacheWriteTokens, t),
+												" · ",
+												t.cacheWrite5m,
+												" ",
+												tokenCount(status.cache.cacheWrite5mTokens, t),
+												" · ",
+												t.cacheWrite1h,
+												" ",
+												tokenCount(status.cache.cacheWrite1hTokens, t)
+											]
+										})]
+									}),
+									/* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
+										className: "dsha-muted",
+										children: t.cacheWriteHint
+									}),
 									/* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
 										className: "dsha-muted",
 										children: t.cacheHint
@@ -5206,11 +5355,12 @@ window.__ModuleLoader__.load({
 								children: t.modelsHint
 							}),
 							/* @__PURE__ */ (0, react_jsx_runtime.jsx)(ModelChecklist, {
-								items: (status?.models ?? []).map((model) => ({
+								items: models.map((model) => ({
 									id: model.id,
 									name: model.name,
 									hint: [
-										model.wire === "anthropic" ? t.wireAnthropic : t.wireOpenai,
+										wireLabel(model.wire, t),
+										...model.thinkingType === "only" ? [t.capThinkingOnly] : [],
 										...model.supportsVideo ? [t.capVideo] : [],
 										...model.supportsDynamicTools ? [t.capDynamicTools] : [],
 										...model.minimumPlan ? [t.capPlan.replace("{plan}", model.minimumPlan)] : []
@@ -5226,7 +5376,7 @@ window.__ModuleLoader__.load({
 									list: t.modelsSection
 								}
 							}),
-							status !== null && status.models.length > 0 && /* @__PURE__ */ (0, react_jsx_runtime.jsx)(KimiModelCapabilities, { models: status.models })
+							status !== null && models.length > 0 && /* @__PURE__ */ (0, react_jsx_runtime.jsx)(KimiModelCapabilities, { models })
 						]
 					}),
 					/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("section", {
@@ -5250,15 +5400,30 @@ window.__ModuleLoader__.load({
 										const value = event.currentTarget.value;
 										handleUpdateEffort(value === "" ? null : value);
 									},
-									children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("option", {
-										value: "",
-										children: t.defaultEffortAuto
-									}), KIMI_CODE_REASONING_EFFORTS.map((effort) => /* @__PURE__ */ (0, react_jsx_runtime.jsx)("option", {
-										value: effort,
-										children: effortLabel(effort, t)
-									}, effort))]
+									children: [
+										/* @__PURE__ */ (0, react_jsx_runtime.jsx)("option", {
+											value: "",
+											children: t.defaultEffortAuto
+										}),
+										offeredEfforts.map((effort) => /* @__PURE__ */ (0, react_jsx_runtime.jsx)("option", {
+											value: effort,
+											children: effortLabel(effort, t)
+										}, effort)),
+										staleEffort !== null && /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("option", {
+											value: staleEffort,
+											children: [
+												effortLabel(staleEffort, t),
+												" · ",
+												t.effortNotOffered
+											]
+										})
+									]
 								})]
 							}),
+							thinkingOnly.length > 0 && /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
+								className: "dsha-notice",
+								children: t.effortThinkingOnlyNote.replace("{models}", thinkingOnly.map((model) => model.name).join(" · "))
+							}),
 							/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
 								className: "dsha-pref-row",
 								children: [/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", { children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("strong", { children: t.cacheTtl }), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
@@ -19071,10 +19236,13 @@ window.__ModuleLoader__.load({
 			wireAnthropic: "Anthropic",
 			wireOpenai: "OpenAI",
 			capVideo: "Видео*",
+			capThinkingOnly: "Только рассуждения",
+			capEfforts: "Уровни {list}",
 			capDynamicTools: "Динамические инструменты",
 			capImageOnly: "Изображения",
 			capNone: "—",
 			capVideoFootnote: "* Видео подключается инструментом kimi_attach_video (абсолютный локальный путь или ссылка http(s)); вставка и загрузка по-прежнему принимают только изображения.",
+			capThinkingOnlyFootnote: "* У модели с пометкой «только рассуждения» нельзя отключить рассуждения: запрос с none сервис отклоняет (HTTP 400), поэтому уровня «рассуждения выкл.» у неё нет.",
 			capabilitiesTitle: "Возможности",
 			capColMedia: "Мультимодальность",
 			capColTools: "Динамические инструменты",
@@ -19085,13 +19253,16 @@ window.__ModuleLoader__.load({
 			speedRegular: "Обычная",
 			speedHighspeed: "Ускоренная (~6×)",
 			quotaCost: "Множитель квоты",
+			quotaUnknown: "Не опубликовано",
 			minimumPlan: "Нужный тариф",
 			enhanced: "Рассуждения",
 			defaultReasoningEffort: "Уровень рассуждений по умолчанию",
-			defaultReasoningEffortHint: "Kimi Code принимает только low / high / max; none отключает рассуждения. Смена уровня сбрасывает кеш контекста, поэтому в пределах сессии держите его неизменным.",
+			defaultReasoningEffortHint: "Набор уровней следует тому, что заявляют отмеченные модели (если не отмечено ни одной — все четыре: low / high / max / none); none отключает рассуждения и действует только для моделей, которые это позволяют. Смена уровня сбрасывает кеш контекста, поэтому в пределах сессии держите его неизменным.",
+			effortThinkingOnlyNote: "Среди отмеченных моделей есть те, что рассуждают всегда: {models}. Уровня «рассуждения выкл.» у них нет — такой запрос сервис отклоняет (HTTP 400), поэтому здесь он не предлагается.",
+			effortNotOffered: "отмеченные модели это не принимают",
 			cacheTtl: "Время жизни кеша контекста",
-			cacheTtlHint: "Kimi кеширует префиксы автоматически, здесь выбирается время записи. Запись на 1 ч стоит примерно вдвое дороже записи на 5 мин и окупается, только если тот же префикс перечитывается в течение часа. Уровень фиксируется при первой записи и позже не меняется. Пустое значение — поле не отправляется, решает Kimi (измеренное время жизни — примерно 5–30 минут). Выбирайте по паузе между ходами: если она обычно короче ~30 минут, оставьте пусто или 5 минут; для сессий с долгим простоем, например диспетчера, ждущего субагентов, — 1 час.",
-			cacheTtlDefault: "Не задавать (по умолчанию Kimi)",
+			cacheTtlHint: "Kimi кеширует префиксы автоматически, здесь выбирается уровень записи, и их всего два: 5m и 1h. Пустое значение — поле не отправляется, и тогда сервис пишет на уровне 5 минут. Запись на 1 ч стоит примерно вдвое дороже записи на 5 мин и окупается, только если тот же префикс перечитывается в течение часа. Уровень фиксируется при первой записи, а попадания продлевают её по исходному TTL. Выбирайте по паузе между ходами: 5 минут, если она обычно короче ~30 минут; 1 час для сессий с долгим простоем, например диспетчера, ждущего субагентов.",
+			cacheTtlDefault: "Не передавать (сервис пишет на уровне 5 минут)",
 			cacheTtl5m: "5 минут",
 			cacheTtl1h: "1 час",
 			defaultEffortAuto: "Как в модели по умолчанию",
@@ -19117,6 +19288,11 @@ window.__ModuleLoader__.load({
 			cacheFresh: "Новые токены",
 			cacheOutput: "Выходные токены",
 			cacheRequests: "Запросы",
+			cacheWrite: "Запись в кеш",
+			cacheWrite5m: "Уровень 5 минут",
+			cacheWrite1h: "Уровень 1 час",
+			cacheWriteHint: "Запись — отдельная статья биллинга: запись на 1 ч стоит примерно вдвое дороже записи на 5 мин, а попадание продлевает её по исходному TTL. 0 на уровне означает, что сервис не сообщил запись для этого уровня (отсутствующий заголовок — молчание), а не то, что там ничего не записали.",
+			unknownMetric: "нет данных",
 			cacheHint: "В пределах одной сессии: любое изменение системного промпта или списка инструментов сбрасывает весь префиксный кеш, поэтому держите набор инструментов стабильным, а новое добавляйте в конец — так доля попаданий выше.",
 			cacheExpired: "Сессия простаивала дольше времени жизни кеша, поэтому следующий ход заново обработает около {tokens} токенов.",
 			cacheDrift: "Состояние последнего префикса",