dsh-connect-trae 4.4.7 → 4.5.1

package/CHANGELOG.md

@@ -1,5 +1,113 @@
 # Changelog
 
+## 4.5.1 (2026-10-25)
+
+> **只在租户池上架的模型不再隐形。** 插件的模型目录只问了 `solo.trae.cn` 一处，
+> 而它并不发布租户池的完整名单——`DeepSeek-V4.1-Flash-1017` 就在其中。这一版把
+> `console.enterprise.trae.cn` 的目录**并入**主目录，两个池的模型都保留。
+>
+> 成因不是合并层丢行，而是**读错了 host**：目录基址循环在第一个答非空的 host 就
+> `return`，而主目录总会答——所以只在租户池上架的模型，即便拿着租户凭据也永远看不到
+> （#46）。
+
+### 修复
+
+- **模型目录并入租户池，只在租户侧上架的模型不再隐形**（#46）。
+  - **实测**（2026-10-24，真实 CN 租户凭据）：`solo.trae.cn` 答 36 个 id、**没有**
+    `DeepSeek-V4.1-Flash-1017`；`console.enterprise.trae.cn` 答 24 个、**有**。
+    4.4.7 与 4.5.0 表现一致，所以升级修不了。
+  - **并集而非替换**：两侧没有一方包含另一方——主目录的 36 个里有 **19** 个不在租户
+    应答中，其中就包括 `deepseek-v4.1-flash`。换 host 等于用 1 个模型换掉 19 个。
+  - **也不能做成兜底候选**：基址循环在第一个答非空的 host 就返回，而主目录总会答，
+    候选恰好在唯一会出错的情形下被跳过。因此额外 host 单独读取后再并入
+    （`TraeSoloRemoteCatalogOptions.unionBaseUrls`）。
+  - **冲突保留主目录的拼写**，按不区分大小写比较。这里不能用 `preferStrongerRow`
+    打擂台：`deriveCatalog` 用 `enabled.has(model.id)` **精确匹配**，改名会让用户
+    已存进 `enabledModelIds` 的那一项静默失配、模型从选择里消失。
+  - **逐 host 尽力而为**：传输失败、非 2xx、响应体读不出、应答为空，都记日志后跳过，
+    只有调用方 abort 才向上抛。某个额外 host 坏掉，损失的是它的模型，而不是整份目录。
+  - **门控按凭据、不按面板**：只登录租户账号的机器，在 Trae（消费端）面板下
+    `store.resolve()` 依然会返回租户凭据（`offered()` 是「藏光了就回退」），所以按
+    面板判会漏掉这类用户。现在判「解析出的凭据 `tenantId` 非空」，个人凭据在发出请求
+    前就跳过。
+  - **`unionRegion` 挡住跨区域请求**：`DSH_TRAE_TENANT_IN_CONSUMER` 是环境变量、不分
+    区域，原先国际版面板也会挂上 CN 租户 host（用 SG 凭据实测确实发出、拿 401 后被
+    跳过）。现在区域不符即跳过。
+  - **额外 host 有负缓存**：`discoverModels()` 不传 signal、超时回落到 30 秒，原先企业
+    host 不可达时每次刷新都要多等。失败后与主 host 同款冷却 5 分钟。
+  - **它不是别名**：用插件自己的 bridge 驱动真实网关，两个 id 派往不同后端
+    （`DeepSeek-V4.1-Flash-1017` → `deepseek-v4.1-flash-economic-ark`，
+    `DeepSeek-V4.1-Flash` → `deepseek-v4.1-flash-ark`）。
+  - 测试：`tests/solo-remote.spec.ts` 新增 8 例，覆盖并集、仅大小写不同的冲突、
+    额外 host 不可达/被拒/应答为空、个人凭据跳过、跨区域跳过、冷却、已经是主 host 时
+    不重复请求，以及未配置 `unionBaseUrls` 时行为与改动前完全一致。
+  - 贡献者：[@yuantao313](https://github.com/yuantao313)（PR #59）。
+
+## 4.5.0 (2026-10-24)
+
+> **宽窗模型多一个 500K 上下文中间档。** Trae 每个模型公布两个窗口——默认（dev）档和
+> Max 档——从 200K 直接跳到 1M 是 5 倍，而 1M 的计费并不是每个会话都用得上。这一版在两者
+> 之间加了一个本地的半步。
+>
+> 500K **不是上游公布的窗口**，所以它有门槛、也有校验：只有当模型的 Max 真的够宽、500K
+> 又确实落在它的两个窗口之间时才出现；落盘校验用同一条规则，旧值撑不破模型的真实窗口。
+
+### 新增
+
+- **宽窗模型多一个 500K 上下文中间档（国内版 / 国际版都有）**。
+  - Trae 每个模型公布两个窗口：默认（dev，实测 200K / 256K / 272K 三种）与 Max
+    （实测 1M，MiMo 为 981504）。从默认直接跳到 Max 是 5 倍，而 1M 的计费并不是每个
+    会话都用得上——所以加一个 500K 的半步。
+  - **500K 不是上游公布的窗口**，是本地的选择，因此有门槛：只有当该模型的
+    `maxContextWindow >= 500000`、且 500K 严格宽于它的 dev 窗、又严格窄于它的 Max
+    时才出现（`contextBudgetOptions`）。dev 本身已是 500K、或 Max 正好 500K 时不会
+    出现重复选项。
+  - **落盘校验用的是同一条规则**：`applyContextBudgets` 只接受该模型可选窗口里的值，
+    所以旧配置里留下的 500K 在某模型 Max 被上游调小之后会安静失效、退回 dev 档，
+    而不是把模型撑到它没有的窗口。
+  - 卡片每行因此可能是 1 / 2 / 3 个单选（无 Max 的行仍是一个只读选项）；新增整区按钮
+    「整区切到 500K」，只作用于 Max ≥ 500K 的行，写入值与逐行单选完全一致。
+  - 测试：`tests/context-budget.spec.ts`（9 例，含「Max 不足 500K 不出现」「Max 正好
+    500K 不重复」「Max 调小后旧 500K 失效」）与 `tests/card-context-500k.spec.tsx`
+    （3 例，含「只给合格行第三个单选」「整区按钮只写合格行」「逐行与整区落盘一致」）。
+    两组均做过变异验证：去掉 500K 档 → 5 例红；放宽接受规则 → 3 例红。
+
+### 修复
+
+- **发布核验不再重建工作树，宿主不再被核验动作掀翻**（`fix(scripts)`）。
+  - **事故**（2026-10-24 14:40:08 / 14:41:01 / 14:42:16）：桌面版渲染进程连崩三次
+    （`EXC_BREAKPOINT`/`SIGTRAP`），应用白屏，最后要人工重启。取证在
+    `~/Library/Logs/DiagnosticReports/DeepSeek Harness Helper (Renderer)-*.ips`。
+  - **成因**：开发者常态是把 bundle 以**软链**装进 profile
+    （`node_modules/dsh-connect-trae -> <本仓库>`），于是运行中的 DSH 宿主服务的正是本仓库
+    `lib/` 的字节；而 `scripts/verify-release.mjs` 的第 4 步（确定性重建比对）原本
+    **在工作树里**跑 `npm run build`——等于当着活着的宿主改写它正在服务的客户端 bundle，
+    宿主重新 import，渲染进程崩。
+  - **修法**：重建改到**临时 worktree**（`git worktree add --detach <tag>` + 软链
+    `node_modules`）里做，工作树的 `lib/` 一个字节都不动；`git worktree remove` 在收尾时
+    清理。用 worktree 而不是 `git archive | tar -x`，是因为测试机上的 PATH 上有**假的
+    `tar`**（发布测试用它冒充 registry），走 tar 会解出错的树。
+  - **守卫**：`tests/verify-release.spec.ts` 新增一例「工作树不被重建」。变异验证：把重建
+    改回 `cwd: repo` → 该例立刻变红（fixture 的 build 会真的写 `lib/`）。
+- **构建前会提醒「这次会改写正在被宿主服务的字节」**（`chore(scripts)`）。
+  新增 `scripts/warn-live-host.mjs`，接到 `prebuild` 钩子：检查「有没有 profile 软链到本仓库」
+  且「有没有活着的宿主进程」，两者同时成立就打印警告。**它从不失败构建**——热更新的正当
+  流程必须还能走；崩了重启应用即可。
+
+### 文档
+
+- `DEVELOPMENT.md` 新增首节「危险动作：不要在宿主运行时重建 `lib/`」：讲清软链装法、
+  事故时间线与三种场景下该怎么做（不开了随便构建 / 只验证发布用 `verify:release` /
+  要热更新就接受代价 / 发版前先关宿主）。
+- `RELEASING.md` 第 8 步补一句：第 4 步的重建在临时 worktree 里做，理由与对策指向上面那节。
+- **在核心位置写明「不做插件之间的联动」这条边界**。
+  - `README.md` 开头加一条醒目的边界声明，并新增 `## 与其它插件的关系` 一节（英文版同步）：
+    不共享状态、不做跨插件的账号/额度调度、不协商槽位、不共享开关，以及立这条边的理由。
+  - `src/index.ts` 的**模块头**加 SCOPE BOUNDARY 注释：本模块只与 DSH Host 和 Trae 上游
+    两方对话，不读写任何其它插件的文件或状态。
+  - 补一条区分：同一族插件之间**互相学习**（workbuddy 的教训被写进本仓库的注释与测试），
+    但**不互相调用**——学习是编译期的事，联动是运行时的事。
+
 ## 4.4.7 (2026-10-24)
 
 > **4.4.6 把「读不出来的账号副本」改名留住了，但这件事一直是静默的——现在卡片会把它点名，

package/README.en.md

@@ -30,6 +30,8 @@ A [DeepSeek Harness (DSH)](https://github.com/deepseek-ai/deepseek-harness) bund
 
 > Apart from the daily check-in claim, every query this plugin makes is read-only and consumes nothing; the claim only happens when you click the button.
 
+> **Boundary: this plugin does not interoperate with any other plugin.** Accounts, credentials, the model directory, the account pool and the settings namespace belong to this plugin alone; a `trae` failure never borrows another plugin's account or credit, and vice versa. Why, and what that rules out, is in [Relationship to other plugins](#relationship-to-other-plugins).
+
 ## Features
 
 - **Three parallel providers** — the CN side registers as the `trae` provider (e.g. `DeepSeek-V4-Flash`, `DeepSeek-V4-Pro`), the enterprise side as `trae-enterprise`, and the international side as `trae-global` (the Gemini/GPT/MiniMax roster); **all three rosters appear in the DSH model picker simultaneously**: different sessions can each pick a side without interfering.
@@ -82,6 +84,19 @@ Usage overview hits the read-only `https://api.trae.cn/trae/api/v2/pay/*` and `/
 
 > See `docs/IMPLEMENTATION_PLAN.md`, `docs/SOLO_ROUTE_DECISION.md`, `docs/USAGE_API_RESEARCH.md`.
 
+## Relationship to other plugins
+
+**This plugin does not interoperate with any other plugin.** That is a boundary, not a slogan:
+
+- **No shared state.** Accounts, credentials, the model directory, the account pool, access tokens and the device id belong to this plugin alone; it reads and writes only its own files (`$DSH_HOME/.trae-auth.*.json`, `.trae-auth.deleted.*.json`, `.trae-pool.*.json`) and its own settings namespace — it does **not** read, write, or supply data to any other plugin. It does edit the Trae desktop app's own sign-in file, but that is another **application**, not another plugin.
+- **No cross-plugin account or credit scheduling.** The pool ranks and switches only among this plugin's own accounts; a `trae` failure is a failure — it never borrows another connector plugin's account, credit or channel, and never hands a request to another plugin to retry. The reverse holds too: no plugin can send requests through this one.
+- **No slot negotiation.** Whether the card and the composer readout appear is decided by **whether this plugin's provider is selected** (read from DSH's own `modelSelection` projection), not by plugin-to-plugin messages allocating space. Two connector plugins installed together coexist because each appears only for its own provider — that is **non-interference**, not collaboration.
+- **No shared switches.** Turning one panel off withdraws only this plugin's provider and requests; it never changes another plugin's provider, models, binding or settings.
+
+**Why the boundary exists.** Once plugins depend on each other, any upgrade, uninstall or version-line change silently alters the behaviour of its neighbours — and DSH's load order, parallel versions and user-authored profiles make that coupling impossible to express reliably. Holding the boundary keeps "who is broken" answerable with **one** plugin instead of a chain.
+
+> Sibling plugins **learn from** each other (lessons from `dsh-connect-workbuddy` are quoted in this repository's comments and tests) but never **call** each other — learning is a compile-time affair, interoperation is a runtime one, and the two must stay apart.
+
 ## Image input
 
 The plugin declares whether a model accepts images, and DSH uses that to decide whether an attachment is **sent as-is** or **degraded into a text description**. That makes the declaration a hard switch, not a cosmetic label.
@@ -134,7 +149,9 @@ To use these two models today, use the **Trae IDE itself**.
 
 The "最大输出" (max output) on the model card comes from `model_detail_list[].max_tokens` in Trae's `get_detail_param` — upstream **does** publish it (live-verified: 80/80 CN configs carry it). As of 4.0.1 the value travels through the catalog merge all the way to the card, `/models/refresh` and DSH model registration; before that the merge step dropped the field, so the whole region rendered blank (issue #42 — full diagnosis in `docs/ISSUE42_DIAGNOSIS.md`). **Rows that are still blank are being honest**: models the directory lists but the SOLO roster has no config for (the three TraeCode rows above) have no candidate value to show — we prefer a blank over an invented cap. The value is metadata only; the plugin never puts `max_tokens` into request bodies (the server decides the real cap).
 
-Context windows are a per-model choice between Trae's default (dev) window and its Max window, stored per panel as `contextBudgets`. For a whole-region switch, the model section now has two buttons: "**Switch all to Max**" (writes each Max-capable row's `maxContextWindow`, leaves no-Max rows and existing manual budgets alone) and "**Restore default context**" (clears the region's budget dict). Both act on the draft and take effect on Save, producing exactly the same stored shape as the per-row radios (issue #42 side request — 18 Max rows in CN used to mean 18 clicks).
+Context windows are a per-model choice between Trae's default (dev) window and its Max window — **and, as of 4.4.8, a 500K middle step for wide-window models** — stored per panel as `contextBudgets`. 500K is **not** an upstream-advertised window; it is a local half-way point. Dev windows are typically 200K/256K/272K and Max is typically 1M, so going straight from default to Max is a 5× jump, and Max-tier billing is not what every session needs. It is therefore **conditional**: the option only appears when that model's Max is at least 500K, and 500K is genuinely wider than its dev window and narrower than its Max. `applyContextBudgets` validates against the **same rule** (`contextBudgetOptions`), so a stored value that is not one of that model's selectable windows always falls back to the dev window — it can never push a model past a window it does not have (if upstream later shrinks a model's Max to 256K, a stale 500K quietly stops applying instead of overflowing).
+
+For a whole-region switch, the model section has three buttons: "**Switch all to 500K**" (writes 500K for every row whose Max is at least 500K, leaves the rest alone), "**Switch all to Max**" (writes each Max-capable row's `maxContextWindow`, leaves no-Max rows and existing manual budgets alone) and "**Restore default context**" (clears the region's budget dict). All act on the draft and take effect on Save, producing exactly the same stored shape as the per-row radios (issue #42 side request — 18 Max rows in CN used to mean 18 clicks).
 
 ## Install
 

package/README.md

@@ -30,6 +30,8 @@
 
 > 除每日签到领取外，插件的所有查询都是只读的、不消耗额度；签到领取也只在你点击按钮时才发生。
 
+> **边界：本插件不与任何其它插件联动。** 账号、凭据、模型目录、账号池与设置命名空间都只属于本插件；`trae` 失败不会借用别的插件的账号或额度，反之亦然。理由与具体含义见[与其它插件的关系](#与其它插件的关系)。
+
 ## 功能特性
 
 - **三供应商并行接入** —— 国内版注册为 DSH 的 `trae` provider（`DeepSeek-V4-Flash`、`DeepSeek-V4-Pro` 等），企业版注册为 `trae-enterprise`，国际版注册为 `trae-global`（Gemini / GPT / MiniMax 阵容），**三边模型同时出现在 DSH 模型选择器里**：不同会话可以各选一边，互不干扰。
@@ -87,6 +89,19 @@ DSH PiAiAdapter（每个 provider 一套）
 
 > 详见 `docs/IMPLEMENTATION_PLAN.md`、`docs/SOLO_ROUTE_DECISION.md`、`docs/USAGE_API_RESEARCH.md`。
 
+## 与其它插件的关系
+
+**本插件不与任何其它插件联动。** 这是一条边界，不是一句口号：
+
+- **不共享状态。** 账号、凭据、模型目录、账号池、访问令牌、设备号都只属于本插件；插件只读写自己的文件（`$DSH_HOME/.trae-auth.*.json`、`.trae-auth.deleted.*.json`、`.trae-pool.*.json`）与自己的设置命名空间，**不读、不写、也不为任何别的插件供数**。它确实会编辑 Trae 桌面应用自己的登录文件，但那是另一个**应用**，不是另一个插件。
+- **不做跨插件的账号或额度调度。** 账号池只在本插件自己的账号之间排序与换号；`trae` 失败就是失败，不会去「借」另一个连接器插件的账号、额度或通道，也不会把请求转交给别的插件重试。反方向同样成立：没有任何插件能通过本插件代发请求。
+- **不协商槽位。** 卡片与 composer 读数是否出现，由**本插件的 provider 是否被选中**决定（读 DSH 自己的 `modelSelection` 投影），不靠插件间通信分配位置。两个连接器插件同时安装时，它们靠「各自只在自己的 provider 被选中时出现」共存——这是**互不干扰**，不是协作。
+- **不共享开关。** 关掉本插件的某个面板，只撤下本插件的 provider 与请求，不会改动另一个插件的供应商、模型、绑定或设置。
+
+**为什么立这条边界。** 插件之间一旦互相依赖，任何一个插件的升级、卸载或版本线变化，都会悄悄改变另一个插件的行为；而 DSH 的加载顺序、多版本并存与用户自定义 profile 让这种耦合无法可靠表达。立住这条边界，「出问题的是谁」永远可以定位到**一个**插件，而不是一串。
+
+> 同一族插件之间**互相学习**（`dsh-connect-workbuddy` 的教训被写进本仓库的注释与测试），但**不互相调用**——学习是编译期的事，联动是运行时的事，这两件事必须分开。
+
 ## 图片输入（识图）
 
 插件对外声明「这个模型是否接受图片」，DSH 据此决定附图是**原样发出**还是**降级成文字描述**。所以这条声明是硬开关，不是一个无关紧要的标记。
@@ -151,7 +166,9 @@ Trae does not serve this model under the SOLO function the request used
 
 模型卡上的「最大输出」来自 Trae `get_detail_param` 的 `model_detail_list[].max_tokens` —— 上游**是**发布这个值的（CN 实测 80/80 config 都带）。4.0.1 起它随目录合并一路带到卡片、`/models/refresh` 与 DSH 模型注册；此前合并层漏传了这个字段，整区模型全部显示空白（issue #42，诊断见 `docs/ISSUE42_DIAGNOSIS.md`）。**仍然空白的行是诚实的**：目录里有、但 SOLO 名单里没有对应 config 的模型（上面那 3 个 TraeCode 通道行）拿不到任何候选值，宁可空白也不编一个上限。这个值是元数据，插件不把它写进请求体（输出上限由 Trae 服务端决定）。
 
-上下文窗口在 Trae 公布的默认（dev）档与 Max 档之间二选一，按面板逐模型存成 `contextBudgets`。想整区一次切换：模型管理区有「**整区切到 Max**」（给每个带 Max 档的行写入其 `maxContextWindow`，没有 Max 档的行不动）和「**全部恢复默认上下文**」（清空本区预算）两个按钮，作用在草稿上、按「保存」生效，与逐行单选落盘的结果完全一致（issue #42 附带需求——CN 有 18 个 Max 模型，以前要点 18 次）。
+上下文窗口在 Trae 公布的默认（dev）档与 Max 档之间选，**4.4.8 起对宽窗模型多一个 500K 中间档**，按面板逐模型存成 `contextBudgets`。500K **不是**上游公布的窗口，是本地的半步：dev 常见 200K/256K/272K，Max 常见 1M，从默认直接跳到 Max 是 5 倍，而 1M 的计费并不是每个会话都用得上。所以它是**有门槛的**：只有当该模型的 Max ≥ 500K、且 500K 确实比 dev 宽、又确实比 Max 窄时才出现——`applyContextBudgets` 用**同一条规则**（`contextBudgetOptions`）校验，落盘值不是该模型可选窗口就一律退回 dev 档，绝不会把模型撑到它没有的窗口（比如某模型后来被上游把 Max 调到 256K，旧的 500K 会安静地失效而不是溢出）。
+
+想整区一次切换：模型管理区有「**整区切到 500K**」（给本区所有 Max ≥ 500K 的行写入 500K，其余不动）、「**整区切到 Max**」（给每个带 Max 档的行写入其 `maxContextWindow`，没有 Max 档的行不动）和「**全部恢复默认上下文**」（清空本区预算）三个按钮，作用在草稿上、按「保存」生效，与逐行单选落盘的结果完全一致（issue #42 附带需求——CN 有 18 个 Max 模型，以前要点 18 次）。
 
 ## 安装
 

package/lib/client.js

@@ -411,6 +411,55 @@ window.__ModuleLoader__.load({
 			return tier !== void 0 && tier !== "" && !/^free$/i.test(tier);
 		}
 		//#endregion
+		//#region src/context-budget.ts
+		/**
+		* Context-budget arithmetic shared by the HOST (which applies a saved budget to
+		* the catalog DSH is told about) and the CLIENT card (which renders the radios).
+		*
+		* It lives in its own module because the client bundle cannot import
+		* `catalog.ts` (host-only) and the two sides must agree EXACTLY on which
+		* windows are selectable: a card that offers a window the host then refuses
+		* would be a control that silently does nothing.
+		*/
+		/**
+		* The one intermediate budget the card offers, in tokens: 500K.
+		*
+		* WHY IT EXISTS: Trae advertises a dev window (typically 200K/256K/272K) and a
+		* Max window (typically 1M). Switching dev→Max is a 5× jump for a user who
+		* just wants more room without paying Max-tier pricing for the whole session.
+		* 500K is the half-way point, and it is only ever offered for a model whose
+		* Max window is at least that wide.
+		*
+		* WHY A CONSTANT AND NOT A PER-MODEL VALUE: this is a LOCAL preference that
+		* DSH uses to decide when to compress; it is not a Trae-advertised window. The
+		* value must be sane for every model that can select it, so eligibility is
+		* gated on `maxContextWindow >= 500000` rather than inventing a window per
+		* model. `applyContextBudgets` additionally refuses any budget wider than what
+		* the model advertises, so a stale saved value cannot overflow a shrunken Max.
+		*/
+		const TRAE_CONTEXT_BUDGET_MID = 5e5;
+		/**
+		* The selectable windows for one model, smallest first: dev, then 500K (only
+		* when it is a genuine middle step), then Max. Duplicates are dropped so a
+		* model whose dev window is already 500K offers it once.
+		*
+		* A model with no Max window yields exactly one option — its dev window — which
+		* is what makes the card's radios read-only for such models.
+		*/
+		function contextBudgetOptions(model) {
+			const dev = model.contextWindow;
+			const max = model.maxContextWindow;
+			const options = [];
+			const push = (value) => {
+				if (value === void 0 || value < 1) return;
+				if (!options.includes(value)) options.push(value);
+			};
+			push(dev);
+			if (max !== void 0 && max >= 5e5 && (dev === void 0 || 5e5 > dev)) push(TRAE_CONTEXT_BUDGET_MID);
+			push(max);
+			return options.sort((left, right) => left - right);
+		}
+		//#endregion
 		//#region src/model-order.ts
 		/**
 		* Wire name of the default SOLO chat function.
@@ -4969,6 +5018,27 @@ window.__ModuleLoader__.load({
 				});
 			};
 			/**
+			* Bulk 500K switch: every model that offers the middle step goes to it, and
+			* every other model is left exactly as configured.
+			*
+			* The reporter-shaped case this serves: a region can hold 19 models with a
+			* 1M Max, and picking the half-way window one radio at a time is 19 clicks.
+			* It writes the literal {@link TRAE_CONTEXT_BUDGET_MID} — the same value the
+			* per-row radio writes — so bulk and per-row land on identical saved state,
+			* and `applyContextBudgets` re-validates it against each model's windows on
+			* every read.
+			*/
+			const setAllContextMid = () => {
+				editDraft((current) => {
+					const next = { ...current.contextBudgets };
+					for (const model of current.models) if (contextBudgetOptions(model).includes(5e5)) next[model.id] = TRAE_CONTEXT_BUDGET_MID;
+					return {
+						...current,
+						contextBudgets: next
+					};
+				});
+			};
+			/**
 			* 「恢复默认上下文」 clears every override in this region's draft. Empty
 			* dict = every row falls back to its default window (the same state as a
 			* region that was never configured), and clearing is what makes the bulk
@@ -5586,26 +5656,22 @@ window.__ModuleLoader__.load({
 																				toggleImage(model.id);
 																			}
 																		}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", { children: t("row.modelImage") })]
-																	}), /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("fieldset", {
+																	}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("fieldset", {
 																		className: "dsm-trae-context-budget",
 																		"aria-label": t("row.contextBudget"),
-																		children: [model.maxContextWindow !== void 0 ? /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("label", { children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("input", {
-																			type: "radio",
-																			name: `context-${model.id}`,
-																			checked: activeContextBudgets[model.id] !== model.maxContextWindow,
-																			disabled: settingsScope?.getSnapshot().writable !== true || saving,
-																			onChange: () => {
-																				setContextBudget(model.id, model.contextWindow);
-																			}
-																		}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", { children: formatCapacity(model.contextWindow, t("row.modelUnknown")) })] }) : null, /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("label", { children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("input", {
-																			type: "radio",
-																			name: `context-${model.id}`,
-																			checked: model.maxContextWindow === void 0 || activeContextBudgets[model.id] === model.maxContextWindow,
-																			disabled: model.maxContextWindow === void 0 || settingsScope?.getSnapshot().writable !== true || saving,
-																			onChange: () => {
-																				setContextBudget(model.id, model.maxContextWindow);
-																			}
-																		}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", { children: formatCapacity(model.maxContextWindow ?? model.contextWindow, t("row.modelUnknown")) })] })]
+																		children: contextBudgetOptions(model).map((option) => {
+																			const selectable = model.maxContextWindow !== void 0;
+																			const checked = (activeContextBudgets[model.id] ?? model.contextWindow) === option;
+																			return /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("label", { children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("input", {
+																				type: "radio",
+																				name: `context-${model.id}`,
+																				checked,
+																				disabled: !selectable || settingsScope?.getSnapshot().writable !== true || saving,
+																				onChange: () => {
+																					setContextBudget(model.id, option);
+																				}
+																			}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("span", { children: formatCapacity(option, t("row.modelUnknown")) })] }, option);
+																		})
 																	})]
 																})]
 															}), /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
@@ -5633,21 +5699,32 @@ window.__ModuleLoader__.load({
 											}),
 											/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("div", {
 												className: "dsm-trae-context-bulk",
-												children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
-													type: "button",
-													className: "dsm-btn dsm-btn-outline",
-													disabled: busy || saving || settingsScope?.getSnapshot().writable !== true || !visibleModels.some((model) => model.maxContextWindow !== void 0),
-													title: t("row.contextAllMaxHint"),
-													onClick: setAllContextMax,
-													children: t("row.contextAllMax")
-												}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
-													type: "button",
-													className: "dsm-btn dsm-btn-outline",
-													disabled: busy || saving || settingsScope?.getSnapshot().writable !== true || Object.keys(activeContextBudgets).length === 0,
-													title: t("row.contextAllDefaultHint"),
-													onClick: resetAllContextBudgets,
-													children: t("row.contextAllDefault")
-												})]
+												children: [
+													/* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
+														type: "button",
+														className: "dsm-btn dsm-btn-outline",
+														disabled: busy || saving || settingsScope?.getSnapshot().writable !== true || !visibleModels.some((model) => contextBudgetOptions(model).includes(5e5)),
+														title: t("row.contextAllMidHint"),
+														onClick: setAllContextMid,
+														children: t("row.contextAllMid")
+													}),
+													/* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
+														type: "button",
+														className: "dsm-btn dsm-btn-outline",
+														disabled: busy || saving || settingsScope?.getSnapshot().writable !== true || !visibleModels.some((model) => model.maxContextWindow !== void 0),
+														title: t("row.contextAllMaxHint"),
+														onClick: setAllContextMax,
+														children: t("row.contextAllMax")
+													}),
+													/* @__PURE__ */ (0, react_jsx_runtime.jsx)("button", {
+														type: "button",
+														className: "dsm-btn dsm-btn-outline",
+														disabled: busy || saving || settingsScope?.getSnapshot().writable !== true || Object.keys(activeContextBudgets).length === 0,
+														title: t("row.contextAllDefaultHint"),
+														onClick: resetAllContextBudgets,
+														children: t("row.contextAllDefault")
+													})
+												]
 											}),
 											/* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
 												className: "dsm-trae-model-capability-note",
@@ -6794,6 +6871,8 @@ window.__ModuleLoader__.load({
 			"row.expiryEarliest": "Earliest: {time}",
 			"row.settingsDirty": "Unsaved changes",
 			"row.modelsTestEmpty": "Enable at least one model first.",
+			"row.contextAllMid": "Switch all to 500K",
+			"row.contextAllMidHint": "Sets the DSH context budget of every model in this region whose Max window is at least 500K to the 500K middle step, leaving the rest alone. Half-way between the default window and Max — for when 1M is more than you need. Still a draft — press Save to apply.",
 			"row.contextAllMax": "Switch all to Max",
 			"row.contextAllMaxHint": "Sets the DSH context budget of every model in this region that has a Max window to its Max window at once. Still a draft — press Save to apply. Models without a Max window are left alone.",
 			"row.contextAllDefault": "Restore default context",
@@ -7171,6 +7250,8 @@ window.__ModuleLoader__.load({
 			"row.expiryEarliest": "最早到期：{time}",
 			"row.settingsDirty": "有未保存的修改",
 			"row.modelsTestEmpty": "请先启用至少一个模型。",
+			"row.contextAllMid": "整区切到 500K",
+			"row.contextAllMidHint": "把本区所有 Max 档不小于 500K 的模型的 DSH 上下文预算一次性设为 500K 中间档，其余模型不动。这是默认窗口与 Max 之间的半步——觉得 1M 用不上、又嫌默认太小时用。这只是草稿——要按「保存」才生效。",
 			"row.contextAllMax": "整区切到 Max",
 			"row.contextAllMaxHint": "把本区所有带 Max 档的模型的 DSH 上下文预算一次性设为各自的 Max 档。这只是草稿——要按「保存」才生效。没有 Max 档的模型不会被改动。",
 			"row.contextAllDefault": "全部恢复默认上下文",

package/lib/index.d.ts

@@ -481,9 +481,13 @@ declare function withChatCallableFunctions(rows: readonly TraeModelInfo[], chatF
 /** Local DSH context budget per model; a value may only select an advertised window. */
 type TraeContextBudget = number;
 /**
- * Apply the saved local budget. Trae advertises two windows per model (dev and
- * Max), so the budget may only switch a model to its own advertised Max value —
- * never to a fabricated number. Everything else keeps the dev window.
+ * Apply the saved local budget. Trae advertises a dev window and a Max window
+ * per model, and the card may also offer the 500K middle step (see
+ * {@link contextBudgetOptions}), so a budget may only switch a model to a
+ * window that model actually offers — never to a fabricated number, and never
+ * to something wider than its Max (a saved 500K for a model later refreshed
+ * down to a 256K Max must fall back, not overflow). Everything else keeps the
+ * dev window.
  */
 declare function applyContextBudgets(catalog: readonly TraeModelInfo[], budgets?: Readonly<Record<string, TraeContextBudget | undefined>>): TraeModelInfo[];
 /** Convert Trae metadata into catalog rows; image input is decided by {@link applyImageSelection}. */
@@ -3482,6 +3486,51 @@ interface TraeSoloRemoteCatalogOptions {
    * of official-host-first.
    */
   log?: (message: string, detail?: unknown) => void;
+  /**
+   * Extra directory hosts whose rows are UNIONED into the primary answer.
+   *
+   * The base loop below stops at the FIRST host that answers non-empty. That is
+   * the right rule for a FALLBACK chain (one host, mirrored), but it cannot
+   * express "these hosts serve DIFFERENT pools" — and the primary host
+   * answering successfully is exactly the case where another host's models
+   * stay invisible. So these hosts are read SEPARATELY and merged, never added
+   * as another candidate base: a candidate is skipped precisely when the
+   * primary succeeds, which is always in the scenario that fails today.
+   *
+   * Measured 2026-10-24 on a live CN tenant credential:
+   * `solo.trae.cn` (the primary) answered 36 ids and does NOT list
+   * `DeepSeek-V4.1-Flash-1017`, while `console.enterprise.trae.cn` answered 24
+   * and DOES list it (issue #46). Neither host is a superset of the other: 19
+   * of the primary's ids are absent from the tenant answer, so REPLACING the
+   * primary would trade one missing model for nineteen — a union is the only
+   * shape that loses nothing.
+   *
+   * The primary answer wins every id collision, compared case-insensitively, and
+   * a colliding extra row is dropped rather than capacity-competed: a
+   * `preferStrongerRow` contest could rename a model the user has already saved
+   * in `enabledModelIds` (`deriveCatalog` matches those ids EXACTLY), silently
+   * dropping the model out of the selection.
+   *
+   * An extra host is asked ONLY when the credential being used can call that
+   * pool at all (`region` must match the host, and a tenant host additionally
+   * requires a tenant credential). The gate lives here rather than at the
+   * construction site because the deciding fact — "is this credential a tenant
+   * one?" — only exists once `credential()` has resolved, which is per refresh.
+   * Gating on the PANEL instead missed the case that matters: a consumer store
+   * does not hard-exclude tenant accounts (`auth.ts`'s `offered()` backs off
+   * when its filter would hide everything), so a machine whose only login is a
+   * tenant account resolves a tenant credential under the Trae panel too.
+   */
+  unionBaseUrls?: readonly string[];
+  /**
+   * Region an extra host serves. An `ai` credential never asks a `cn` host.
+   *
+   * `unionBaseUrls` used to be ungated by region, so a `DSH_TRAE_TENANT_IN_CONSUMER`
+   * machine on the international panel issued the CN tenant request on every
+   * refresh — a 401 the union skipped, but one wasted request and one warn log
+   * per refresh (measured with an SG credential 2026-10-24).
+   */
+  unionRegion?: TraeRegion;
 }
 /**
  * Read-only model catalog client for the SOLO Web API.
@@ -3522,9 +3571,42 @@ declare class TraeSoloRemoteCatalogClient {
    * (30 s) in the worst case, so 5 minutes bounds that to ~1 attempt per 5 min.
    */
   private static readonly OFFICIAL_RETRY_MS;
+  /**
+   * When each extra union host may be asked again after it failed, as epoch
+   * milliseconds. Absent = ask it on the next refresh.
+   *
+   * The official host has {@link OFFICIAL_RETRY_MS} because a blocked network
+   * would otherwise pay its failing connect on every refresh; an extra host has
+   * the same problem and, since discovery (`discoverModels()`) passes no signal,
+   * the same 30 s ceiling. A tenant host that is unreachable from this network
+   * would therefore stall every refresh for up to 30 s. Same cooldown, same
+   * reasoning — an unhelpful host is skipped for a while instead of retried
+   * immediately.
+   */
+  private readonly extraRetryAtMs;
+  private readonly unionBaseUrls;
+  private readonly unionRegion;
   constructor(options: TraeSoloRemoteCatalogOptions);
   private headers;
   fetchModels(signal?: AbortSignal): Promise<TraeDiscoveredModel[]>;
+  /**
+   * Union the extra directory hosts' rows into the primary answer (issue #46).
+   *
+   * The base loop above returns on the FIRST host that answers non-empty, which
+   * is correct for a FALLBACK chain but cannot express "these hosts serve
+   * different pools" — the primary host answering is exactly the case where the
+   * extra host's models would stay invisible forever. Hence a separate read and
+   * a merge, not another candidate base (see `unionBaseUrls`).
+   *
+   * Every extra host is BEST-EFFORT: transport failure, non-2xx, an unreadable
+   * body and an empty answer are each logged and skipped. Only an aborted
+   * `signal` propagates — that is the caller cancelling, not a host being
+   * unhelpful. Failing here must never fail a refresh the primary host already
+   * satisfied, which is why a broken extra host costs models, not the catalog.
+   * A failure also puts that host on {@link extraRetryAtMs} so a network that
+   * cannot reach it stops paying its connect timeout on every refresh.
+   */
+  private unionWithExtraHosts;
   /**
    * Bases to try for one refresh, in order.
    *

package/lib/index.js

@@ -15,6 +15,55 @@ import { execFile } from "node:child_process";
 import { promisify } from "node:util";
 import { createServer } from "node:http";
 import { Readable } from "node:stream";
+//#region src/context-budget.ts
+/**
+* Context-budget arithmetic shared by the HOST (which applies a saved budget to
+* the catalog DSH is told about) and the CLIENT card (which renders the radios).
+*
+* It lives in its own module because the client bundle cannot import
+* `catalog.ts` (host-only) and the two sides must agree EXACTLY on which
+* windows are selectable: a card that offers a window the host then refuses
+* would be a control that silently does nothing.
+*/
+/**
+* The one intermediate budget the card offers, in tokens: 500K.
+*
+* WHY IT EXISTS: Trae advertises a dev window (typically 200K/256K/272K) and a
+* Max window (typically 1M). Switching dev→Max is a 5× jump for a user who
+* just wants more room without paying Max-tier pricing for the whole session.
+* 500K is the half-way point, and it is only ever offered for a model whose
+* Max window is at least that wide.
+*
+* WHY A CONSTANT AND NOT A PER-MODEL VALUE: this is a LOCAL preference that
+* DSH uses to decide when to compress; it is not a Trae-advertised window. The
+* value must be sane for every model that can select it, so eligibility is
+* gated on `maxContextWindow >= 500000` rather than inventing a window per
+* model. `applyContextBudgets` additionally refuses any budget wider than what
+* the model advertises, so a stale saved value cannot overflow a shrunken Max.
+*/
+const TRAE_CONTEXT_BUDGET_MID = 5e5;
+/**
+* The selectable windows for one model, smallest first: dev, then 500K (only
+* when it is a genuine middle step), then Max. Duplicates are dropped so a
+* model whose dev window is already 500K offers it once.
+*
+* A model with no Max window yields exactly one option — its dev window — which
+* is what makes the card's radios read-only for such models.
+*/
+function contextBudgetOptions(model) {
+	const dev = model.contextWindow;
+	const max = model.maxContextWindow;
+	const options = [];
+	const push = (value) => {
+		if (value === void 0 || value < 1) return;
+		if (!options.includes(value)) options.push(value);
+	};
+	push(dev);
+	if (max !== void 0 && max >= 5e5 && (dev === void 0 || 5e5 > dev)) push(TRAE_CONTEXT_BUDGET_MID);
+	push(max);
+	return options.sort((left, right) => left - right);
+}
+//#endregion
 //#region src/catalog.ts
 /**
 * Bootstrap catalog: identity only where current Trae metadata has not been
@@ -425,15 +474,22 @@ function withChatCallableFunctions(rows, chatFunctions) {
 	});
 }
 /**
-* Apply the saved local budget. Trae advertises two windows per model (dev and
-* Max), so the budget may only switch a model to its own advertised Max value —
-* never to a fabricated number. Everything else keeps the dev window.
+* Apply the saved local budget. Trae advertises a dev window and a Max window
+* per model, and the card may also offer the 500K middle step (see
+* {@link contextBudgetOptions}), so a budget may only switch a model to a
+* window that model actually offers — never to a fabricated number, and never
+* to something wider than its Max (a saved 500K for a model later refreshed
+* down to a 256K Max must fall back, not overflow). Everything else keeps the
+* dev window.
 */
 function applyContextBudgets(catalog, budgets = {}) {
-	return catalog.map((model) => ({
-		...model,
-		...model.maxContextWindow !== void 0 && budgets[model.id] === model.maxContextWindow ? { contextWindow: model.maxContextWindow } : {}
-	}));
+	return catalog.map((model) => {
+		const budget = budgets[model.id];
+		return {
+			...model,
+			...budget !== void 0 && contextBudgetOptions(model).includes(budget) ? { contextWindow: budget } : {}
+		};
+	});
 }
 /** Convert Trae metadata into catalog rows; image input is decided by {@link applyImageSelection}. */
 function discoveredCatalog(models) {
@@ -5969,6 +6025,17 @@ function traeSoloRemoteFallbackBases(region, officialBase) {
 	return backupBase === officialBase ? [] : [backupBase];
 }
 /**
+* Case-folded identity for a directory id.
+*
+* A local copy of `catalog.ts`'s private `displayKey`: the two hosts of the
+* union spell the same model differently (they disagree on at least 3 ids), and
+* ids that differ only by case MUST be treated as the same model — otherwise a
+* union would list one model twice under two spellings.
+*/
+function directoryKey(id) {
+	return id.trim().toLowerCase();
+}
+/**
 * Of two rows for the SAME model id, the one the catalog should keep.
 *
 * The gateway groups its directory by function and reports a model's capacity
@@ -6107,10 +6174,27 @@ var TraeSoloRemoteCatalogClient = class TraeSoloRemoteCatalogClient {
 	* (30 s) in the worst case, so 5 minutes bounds that to ~1 attempt per 5 min.
 	*/
 	static OFFICIAL_RETRY_MS = 3e5;
+	/**
+	* When each extra union host may be asked again after it failed, as epoch
+	* milliseconds. Absent = ask it on the next refresh.
+	*
+	* The official host has {@link OFFICIAL_RETRY_MS} because a blocked network
+	* would otherwise pay its failing connect on every refresh; an extra host has
+	* the same problem and, since discovery (`discoverModels()`) passes no signal,
+	* the same 30 s ceiling. A tenant host that is unreachable from this network
+	* would therefore stall every refresh for up to 30 s. Same cooldown, same
+	* reasoning — an unhelpful host is skipped for a while instead of retried
+	* immediately.
+	*/
+	extraRetryAtMs = /* @__PURE__ */ new Map();
+	unionBaseUrls;
+	unionRegion;
 	constructor(options) {
 		this.options = options;
 		this.fetchImpl = options.fetchImpl ?? fetch;
 		this.baseUrl = options.baseUrl;
+		this.unionBaseUrls = options.unionBaseUrls ?? [];
+		this.unionRegion = options.unionRegion;
 	}
 	async headers(region) {
 		const credential = await this.options.credential();
@@ -6126,7 +6210,8 @@ var TraeSoloRemoteCatalogClient = class TraeSoloRemoteCatalogClient {
 		};
 	}
 	async fetchModels(signal) {
-		const region = regionOfCredential(await this.options.credential());
+		const credential = await this.options.credential();
+		const region = regionOfCredential(credential);
 		const officialBase = this.baseUrl ?? REGION_GATEWAYS[region].remote;
 		const headers = await this.headers(region);
 		const functions = TRAE_REMOTE_DIRECTORY_FUNCTIONS[region].join(",");
@@ -6161,12 +6246,88 @@ var TraeSoloRemoteCatalogClient = class TraeSoloRemoteCatalogClient {
 			if (base !== officialBase) this.options.log?.(this.lastSuccessfulBase === base ? `model directory served by the remembered fallback host ${base}; the official host (${officialBase}) is re-probed every ${TraeSoloRemoteCatalogClient.OFFICIAL_RETRY_MS / 6e4} minutes` : `model directory served by fallback host ${base}; the official host (${officialBase}) failed this refresh`, { officialBase });
 			else this.officialRetryAtMs = 0;
 			this.lastSuccessfulBase = base;
-			return models;
+			return await this.unionWithExtraHosts(region, credential, headers, models, signal);
 		}
 		if (emptyError !== void 0) throw emptyError;
 		throw lastError instanceof Error ? lastError : new Error(String(lastError));
 	}
 	/**
+	* Union the extra directory hosts' rows into the primary answer (issue #46).
+	*
+	* The base loop above returns on the FIRST host that answers non-empty, which
+	* is correct for a FALLBACK chain but cannot express "these hosts serve
+	* different pools" — the primary host answering is exactly the case where the
+	* extra host's models would stay invisible forever. Hence a separate read and
+	* a merge, not another candidate base (see `unionBaseUrls`).
+	*
+	* Every extra host is BEST-EFFORT: transport failure, non-2xx, an unreadable
+	* body and an empty answer are each logged and skipped. Only an aborted
+	* `signal` propagates — that is the caller cancelling, not a host being
+	* unhelpful. Failing here must never fail a refresh the primary host already
+	* satisfied, which is why a broken extra host costs models, not the catalog.
+	* A failure also puts that host on {@link extraRetryAtMs} so a network that
+	* cannot reach it stops paying its connect timeout on every refresh.
+	*/
+	async unionWithExtraHosts(region, credential, headers, primary, signal) {
+		if (this.unionBaseUrls.length === 0) return primary;
+		if (this.unionRegion !== void 0 && this.unionRegion !== region) return primary;
+		if (!isEnterpriseAccount(credential)) return primary;
+		const functions = TRAE_REMOTE_DIRECTORY_FUNCTIONS[region].join(",");
+		const byKey = /* @__PURE__ */ new Map();
+		for (const model of primary) byKey.set(directoryKey(model.id), model);
+		for (const base of this.unionBaseUrls) {
+			if (base === this.lastSuccessfulBase) continue;
+			if ((this.extraRetryAtMs.get(base) ?? 0) > Date.now()) continue;
+			let response;
+			try {
+				response = await this.fetchImpl(`${base}/models?functions=${functions}`, {
+					headers,
+					signal: signal ?? AbortSignal.timeout(3e4)
+				});
+			} catch (error) {
+				if (signal?.aborted) throw error;
+				this.extraRetryAtMs.set(base, Date.now() + TraeSoloRemoteCatalogClient.OFFICIAL_RETRY_MS);
+				this.options.log?.(`model directory union host ${base} was unreachable; its models are missing from this refresh`, {
+					base,
+					error: String(error)
+				});
+				continue;
+			}
+			if (!response.ok) {
+				this.extraRetryAtMs.set(base, Date.now() + TraeSoloRemoteCatalogClient.OFFICIAL_RETRY_MS);
+				this.options.log?.(`model directory union host ${base} answered HTTP ${response.status}; its models are missing from this refresh`, { base });
+				continue;
+			}
+			let extra;
+			try {
+				const json = await response.json();
+				extra = this.collectModels(region, json.data?.list ?? []);
+			} catch (error) {
+				this.extraRetryAtMs.set(base, Date.now() + TraeSoloRemoteCatalogClient.OFFICIAL_RETRY_MS);
+				this.options.log?.(`model directory union host ${base} answered a body that could not be read; its models are missing from this refresh`, {
+					base,
+					error: String(error)
+				});
+				continue;
+			}
+			this.extraRetryAtMs.delete(base);
+			let added = 0;
+			for (const model of extra) {
+				const key = directoryKey(model.id);
+				if (byKey.has(key)) continue;
+				byKey.set(key, model);
+				added += 1;
+			}
+			const collided = extra.length - added;
+			this.options.log?.(`model directory union host ${base} contributed ${added} model(s) the primary directory did not list${collided > 0 ? ` (${collided} collided and kept the primary spelling)` : ""}`, {
+				base,
+				served: extra.length,
+				added
+			});
+		}
+		return [...byKey.values()];
+	}
+	/**
 	* Bases to try for one refresh, in order.
 	*
 	* While the official host is in its post-failure cooldown the backup hosts go
@@ -9974,7 +10135,57 @@ function apply(ctx, config) {
 		});
 		const remoteCatalog = new TraeSoloRemoteCatalogClient({
 			credential: () => store.resolve(),
-			log: (message, detail) => ctx.logger.warn(message, detail)
+			log: (message, detail) => ctx.logger.warn(message, detail),
+			/**
+			* The TENANT directory, UNIONED into the consumer one (issue #46).
+			*
+			* This is the "reads both pools" half of the routing note below, which
+			* was intent without an implementation: only the consumer host was ever
+			* asked. The `Solo` client above is routed, but THIS one is not — and
+			* `solo.trae.cn`, the primary it lands on, does not publish every model a
+			* tenant credential can call.
+			*
+			* Measured 2026-10-24 on the reporter's live CN tenant credential:
+			* `solo.trae.cn` answered 36 ids with no `DeepSeek-V4.1-Flash-1017`;
+			* `console.enterprise.trae.cn` answered 24 that include it, and the
+			* gateway really dispatching a distinct backend for it
+			* (`deepseek-v4.1-flash-economic-ark`, vs `deepseek-v4.1-flash-ark` for
+			* the plain id) is how issue #46's report was confirmed.
+			*
+			* UNION, never replace, and never a fallback candidate. The primary is
+			* the richer host — 19 of its ids are absent from the tenant answer,
+			* including the `deepseek-v4.1-flash` the reporter has saved — so
+			* replacing it would trade one missing model for nineteen. And a
+			* candidate base cannot work at all: the base loop returns on the first
+			* host that answers, and the primary always answers.
+			*
+			* Gated on a tenant pool being bindable here (an enterprise panel, or a
+			* consumer panel under `DSH_TRAE_TENANT_IN_CONSUMER`). A personal
+			* consumer account has nothing to gain from the tenant host, so it is not
+			* made to pay the extra request.
+			*
+			* The fix is not about 1017 specifically: the tenant host is a second
+			* pool, and every model it alone serves was invisible for the same
+			* reason. Verified against the live hosts: 0 rows lost, 0 duplicated.
+			*
+			* The client decides WHEN to ask it, not this construction site: the
+			* deciding fact is whether the credential being resolved is a tenant one,
+			* which does not exist until `credential()` runs. Gating on the panel here
+			* missed the reporters most likely to hit #46 in the first place — a
+			* consumer store does not hard-exclude tenant accounts (`auth.ts`'s
+			* `offered()` backs off when the filter would hide everything), so a
+			* machine whose only login is a tenant account resolves one under the
+			* Trae panel too, and an enterprise-panel-only gate never opened the
+			* union for it. `unionRegion` additionally keeps an `ai` credential from
+			* asking a `cn` host, which the old `DSH_TRAE_TENANT_IN_CONSUMER` gate on
+			* its own did not.
+			*
+			* Providing the host unconditionally is therefore safe: a personal
+			* credential and an `ai` credential both skip it before any request is
+			* made, so neither pays for a pool it cannot use.
+			*/
+			unionBaseUrls: [`${TRAE_ENTERPRISE_HOST}/api/remote/v1`],
+			unionRegion: "cn"
 		});
 		/**
 		* The IDE's own roster: which models the vendor's picker shows, per
@@ -10003,7 +10214,10 @@ function apply(ctx, config) {
 		* gateway and stamps the `session_id` it requires. The discovery client
 		* below stays UNROUTED on purpose — it owns the roster/probe path and, for
 		* a tenant credential, reads both pools so the shared catalogue keeps every
-		* model either one serves.
+		* model either one serves. That second pool is supplied by `unionBaseUrls`
+		* at the catalog client above, which is what makes this paragraph true;
+		* routing THIS client instead would have read one pool and dropped the
+		* other (see the note there, and issue #46).
 		*/
 		/**
 		* Latest successful discovery for this panel, mutated by discoverModels

package/package.json

@@ -2,7 +2,7 @@
   "name": "dsh-connect-trae",
   "displayName": "DSH Connect Trae",
   "description": "把本机登录的 Trae 模型接入 DeepSeek Harness：国内版与国际版双供应商并行，提供用量/积分概览与每日签到领取。",
-  "version": "4.4.7",
+  "version": "4.5.1",
   "author": "dingminhua",
   "copyright": "Copyright (c) 2026 LaoDing",
   "license": "MIT",
@@ -52,7 +52,8 @@
     "verify:release": "node scripts/verify-release.mjs",
     "card:shot": "vitest run --config vitest.card.config.ts",
     "promo:shot": "vitest run --config vitest.card.config.ts -t \"promo shot\"",
-    "panels:shot": "vitest run --config vitest.card.config.ts -t \"panels shot\""
+    "panels:shot": "vitest run --config vitest.card.config.ts -t \"panels shot\"",
+    "prebuild": "node scripts/warn-live-host.mjs"
   },
   "dsh": {
     "bundle": {