dsh-connect-trae 4.5.1 → 4.6.0

package/CHANGELOG.md

@@ -1,5 +1,103 @@
 # Changelog
 
+## 4.6.0 (2026-10-26)
+
+> **文生图先只在国内版面世，并且必须由用户亲手勾选。** 企业版的租户网关是另一条
+> 计费面，国际版按另一套套餐计费，两者的生图行为都没有测过——所有取证都在
+> `trae-api-cn.mchost.guru` 上做的。
+
+### 新增
+
+- **文生图接进会话：`trae_draw_image`**（`feat(tools)`）。
+  - 用登录账号画一张图，返回 Trae 的**图片对象 key**（形如 `tos-cn-i-…/cn/<uid>/gen_image/…`，
+    图在 Trae 客户端的历史里能看到；字节插件拿不到，见下）。默认关闭、**按面板单独开**：
+    卡片模型栏勾上「允许模型自主生图」（`drawImageEnabled`）该面板才注册这个工具，
+    取消勾选立即撤下。
+  - **协议是实测出来的，不是从二进制里猜的**：`POST /api/ide/v1/tool_text_to_image`
+    （另有 SSE 变体），请求体**只认** `{prompt, image_size?}`——多带 `mode_type`/`session_id`/
+    `image_urls` 之类的已知字段直接 400 空 body，未知字段反被忽略；成功响应是**扁平**的
+    `{"message":"success","uri":"tos-cn-i-…"}`，没有 `{code,msg,data}` 信封。模型 `seedream_5`。
+  - **尺寸边界**：总像素必须落在 **3686400 ~ 16777216**（`1024x1024` 报
+    `image size must be at least 3686400 pixels`、`4608x4608` 报 at most，两者都**不扣费**）；
+    预设 `2k`/`3k`/`4k` 可用，缺省 `1920x1920`。
+  - **计费口径写进工具文案**：实测 `1920x1920` 一张约 **8.8 积分**（国内版价），上游不回传
+    扣费数字。失败后加 **30 秒冷却**（按张计费，不过冷却的重试等于再买一张）。
+  - **不取图，只回 key**：同路径的 GET 不是取图，**是又一次生成**。决定性探针（CN 个人版）：
+    POST 拿到 `…/1792954795741_44f94116_jpeg_1920x1920` 之后，用同一 `(prompt, size)` 带
+    全套网关头 + `x-ide-sign` 打 GET，回来的是**另一个** uri
+    （`…/1792954810505_b107ac11_jpeg_1920x1920`），耗时 27 秒、**一个像素都没有**——等于又画
+    一张、又扣一次（两次共约 17.6）。非法尺寸探针（`100x100`）同样印证：报错直接来自 Ark 的
+    `[callArkImageGeneration]`。所以发版代码**不再发这条 GET**：`generate` 只发一次 POST，
+    返回 `{url, credit?}`，渲染层把 key 显示成**行内代码**并注明「不能直接打开」，绝不拼进
+    `![](...)`（key 不是路径，拼进去只会得到以当前目录为基准的碎图标）。要真正把 uri 变成
+    字节，需要另一条通道（iCube material 的 `get_url` + `x-icube-token`，基址来自运行时
+    boot config），目前**没有打通**，代码里也不留假装能取的入口。
+    两例 `tests/image-tool-render.spec.ts` 与两例 `tests/image-tool-download.spec.ts`
+    （现已改名口径）用 stub fetch 断言**每次生成只发一次请求**。
+  - **只有国内版面板提供**（`DRAW_IMAGE_PANELS = ['cn']`，见 `src/status-paths.ts`）。
+    企业版走租户网关、国际版按另一套套餐计费，两者的生图行为都**没有测过**：企业版原先
+    就因为这一点被拒绝，国际版没有理由凭「面板是消费者版」放行——而且工具文案里的
+    8.8 积分是国内价，放到国际版就是错的。两处的勾选框都**置灰并写明原因**（不隐藏：
+    隐藏会让人以为插件漏了这个功能），理由与「加进白名单必须先补实测证据」写在同一处注释里。
+  - **注册开关是一个可测的纯函数**：`drawImagePanelsOf()` 三道门——面板在白名单内、
+    该面板的 provider 开着（opt-out 语义，旧配置仍算开）、用户显式勾选（**只认字面量
+    `true`**，`'yes'`、`1` 之类的手改值不算同意）。企业版即使被勾上也**结构性**拒绝，
+    不靠白名单列表的书写顺序。首版把这条规则写在 `apply()` 的闭包里，只能靠读代码确认，
+    现在由 `tests/image-tool-optin.spec.ts` 的 12 例钉住（缺省/非法形状/非布尔真值/关掉的
+    provider/企业版/国际版/多面板顺序/整段 section 与裸 `regions` map 同答）。
+  - **开关就在卡片里，不是一个只有 schema 的字段**：真值是 `regions.<panel>.drawImageEnabled`
+    （Host 的工具解析按面板读它，所以写的是 `regions` 而不是顶层设置键）。勾选即写、不经过
+    「保存」，写后**回读文档**核对——一个谎报授权的勾选框会让「工具没出现」看起来像 bug。
+    4 例卡片测试 `tests/card-draw-image-switch.spec.tsx` 断言写入字段是 `regions`、该槽位的
+    模型目录/图片模型/上下文预算/账号池一个都不能丢。
+  - **范围**：改图（需先走 TOS 上传预签名）与文生视频（异步素材轮询）**不做**，也不留半截入口。
+
+### 文档
+
+- `README.md` / `README.en.md` 新增「文生图」一节与功能列表条目：开关语义（含只有国内版
+  提供）、实测协议表、计费，以及为什么那条「取图」GET **是又一次生成**、插件只回 key。
+
+## 4.5.2 (2026-10-25)
+
+> **思考档位不再被吞，也不再被谎报。** 插件原先只认识三档，未知档位一律按 `high`
+> 发出去——用户选 Max、实际收到 high，无报错也无日志。这一版把词表与 Trae 客户端
+> 对齐（`low / medium / high / extra_high / max`），并把「不认识的档位」从谎报改成
+> 丢弃，让上游用它自己的默认值。
+>
+> 词表变宽，**声称没变**：能选哪些档位仍然只取决于上游为该模型实际下发的
+> `options`，所以今天只发三档的模型不会凭空多出 Max。
+
+### 修复
+
+- **思考档位词表配平，未知档位不再静默降档成 high**（#60）。
+  - **三处吞档位**：解析侧 `EFFORT_MAP` / `TRAE_REASONING_EFFORTS` 缺 `medium` 与
+    `max`；`toPiModel` 的 `thinkingLevelMap` 把 `minimal / medium / max` 硬编码成
+    `null`；反向映射写成三元链
+    `effort === 'low' ? 'light' : effort === 'xhigh' ? 'extra_high' : 'high'`。
+  - **最坏的是第三处**：任何未预期档位都被映射成 `high` 发给上游。Trae 客户端自己
+    认识五档（`reasoning_effort.{low,medium,high,extra_high,max}`），并按 agent model
+    灰度下发，所以上游哪天发 `max`，用户选 Max 就会被静默降成 high。**静默降档比
+    不给选择器更坏**——它给出一个便宜了、但用户没要的答案。
+  - **改法**：反向映射抽成 `WIRE_EFFORT_MAP` 查表 + `wireEffortsOf()`，三处调用点
+    统一走它（报告人点到两处，第三处在 wire 行提升的分支里，一起改才不会有同一份
+    目录两种行为）。查不到即**丢弃**，而不是猜一个默认值。
+  - **实测**（2026-10-25，本机 CN 凭据，98 行目录）：当前 `options` 只有
+    `["light","high"]`、`["high","extra_high"]`、`["light","high","extra_high"]`
+    三种，`max` 尚未下发——所以这是「今天就备好、明天才用上」，不是伪造档位。
+  - 测试：新增 10 例（词表、丢弃未知档位、既有三档逐字不变、`discoveredCatalog`
+    同步、不伪造档位、`thinkingLevelMap` 按模型而非硬编码）。三处均做变异验证：
+    恢复旧映射 → 5 例红；恢复硬编码 `null` → 1 例红；去掉解析侧新键 → 1 例红。
+  - 贡献者：[@ahian-lee](https://github.com/ahian-lee)（issue #60，附诊断与补丁）。
+
+### 文档
+
+- **卡片模型栏写明「上下文预算只影响 DSH 侧」**（#60 的顺带问题）。
+  - 每行显示的是 Trae 自己公布的最大窗口，而下面的预算单选只写 DSH 的本地预算；
+    请求体里既没有窗口字段、也没有 `max_mode`，所以把预算提到 1M 并不会让上游按
+    1M 服务。原先的标签读起来像「已经放开到 1M」。
+  - 上游是否接受 `v2_max_mode_enabled` 这类字段**尚未取证**，所以这次不动请求体
+    （乱加字段可能让每次请求 400），只把事实说出来：常驻说明，不用悬停提示。
+
 ## 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，

package/README.en.md

@@ -42,6 +42,7 @@ A [DeepSeek Harness (DSH)](https://github.com/deepseek-ai/deepseek-harness) bund
 - **Multiplier in the model name** — model names show the credit multiplier in Trae's own menu format (e.g. `GLM-5.2 · x0.79`), updated with each directory refresh.
 - **Promotion badge** — Trae publishes its discount rules in the model directory (`activity_discount`: pay ratio, list price, off-peak windows with their time zone). The plugin judges them against **the clock at render time** and paints one capsule beside the model name (`5 折` / `off-peak 5 折` / `limited 5 折`), with the old price, the new price and the window on hover. Judging at render time is what keeps a night-time discount from still hanging there the next morning; when upstream says nothing is discounted right now no badge is painted, the multiplier is never used to guess a promotion, and no campaign without an end time is labelled "limited". It is information only: no request and no billing changes, and the international region (subscription) has no such data and therefore shows no badge. See [docs/PROMO_BADGE.md](docs/PROMO_BADGE.md).
 - **DSH local tool loop** — gets pending structured `tool_calls` from Trae `llm_utils_chat`, lets DSH execute its own local tools, then returns tool results to the model.
+- **Text-to-image (`trae_draw_image`)** — brings the signed-in account's Trae text-to-image service into the conversation: the model draws one image and the plugin hands back the Trae object key (the image itself is visible in the Trae client; the bytes are not reachable from the plugin). **Off by default, enabled per panel** — generation really spends credits (measured: ~8.8 for one `1920x1920` on CN personal), so unlike the image-input declaration it only happens when you say so. Protocol, size bounds and why the "fetch" GET is really a second generation are in "Text-to-image" below.
 - **Dual-region detection (CN / international)** — detects local sign-ins across Trae CN, TRAE SOLO CN, Trae, and TRAE SOLO installs. The region is derived from the credential's own `userRegion` claim (host suffix and edition label as fallbacks), with no manual switch.
 - **Region-isolated directories and selections** — CN and international each keep their own model directory, enabled picks, image opt-ins, and context budgets, and each provider reads only its own slot.
 - **Vision follows Trae's own declaration** — when Trae marks a model `multimodal`, the plugin advertises it as such with no manual tick per model (issue #16). The rule is "**models you tick ∪ models Trae marks multimodal**": models Trae marks `false`, or does not mark at all, stay text-only and can still be turned on by hand. A row enabled automatically is shown ticked and read-only, with the reason. See "Image input" below.
@@ -115,6 +116,27 @@ Notes:
 - **An automatically enabled row is shown ticked and read-only.** Under the union, unticking it would be re-applied immediately, so offering a clickable box would be a lie. There is currently no per-model "force off" escape hatch; if you want one, say so on issue #16.
 - Manual ticks still work for models Trae marks `false` (or does not mark), which is the escape hatch for "upstream does not say so, but it actually works".
 
+## Text-to-image (`trae_draw_image`)
+
+The text-to-image service behind TraeWork / SOLO is now reachable from a DSH conversation: the model **draws one image** and the result is written to a local file.
+
+**Off by default, switched on per panel.** Generation **really spends credits** (measured: ~8.8 for one `1920x1920` on CN personal), so this is not the image-input kind of declaration — it is an explicit spend gate. Tick "Allow the model to generate images" (`drawImageEnabled`) in the card's model column and that panel's `trae_draw_image` gets registered with DSH; unticking removes it immediately. Each panel stores its own switch; the enterprise and international panels do not offer one — the former's tenant gateway is a separate billing surface, the latter bills under another plan, and neither has been measured, so the plugin does not pretend otherwise.
+
+**The protocol (all measured, none of it guessed from the binary):**
+
+| Item | Fact |
+|---|---|
+| Endpoint | `POST /api/ide/v1/tool_text_to_image` (plus an SSE variant `..._stream`), model `seedream_5` |
+| Request body | Only `{prompt, image_size?}`. Adding known fields (`mode_type` / `session_id` / `image_urls` / `watermark`) earns a 400 with an empty body; unknown fields are ignored |
+| Success | Flat `{"message":"success","uri":"tos-cn-i-…/cn/<uid>/gen_image/<ts>_<hex>_jpeg_1920x1920"}` — **no** `{code,msg,data}` envelope |
+| Sizes | Total pixels must land in **3686400 ~ 16777216**: `1024x1024` is rejected as too small (`code 3003`, no charge) and `4608x4608` as too large; presets `2k`/`3k`/`4k` work, default is `1920x1920` |
+| Fetch | **There is no working fetch channel.** The `uri` is an object key, not an openable URL, and the same-path GET is not a fetch but another generation (next row). So the tool returns the key and nothing else; the picture is visible in the Trae client's history |
+| What that GET really is | Decisive probe: after the POST returned `…/1792954795741_44f94116_jpeg_1920x1920`, the GET with the same `(prompt, size)` came back with a **different** uri (`…/1792954810505_b107ac11_jpeg_1920x1920`) after 27 s and zero pixels — a second image and a second charge (~17.6 for the pair). The illegal-size probe agreed: the error is Ark's own `[callArkImageGeneration]` |
+
+**Key only, no fetch.** On success the tool returns `url` (the Trae object key, shaped `tos-cn-i-…/cn/<uid>/gen_image/…`), `model`, `prompt`, `size` and a `hint`; the renderer shows the key as **inline code** labelled "cannot be opened directly" and never splices it into `![](...)` — a key is not a path, and pasting it there only yields a broken icon resolved against the cwd. Missing bytes are a capability boundary here, not an error: turning a uri into pixels needs another channel (iCube material `get_url` plus `x-icube-token`, base URL from the runtime boot config) that is not wired up yet. Failures still start a 30-second cooldown — generation is billed per image, so a retry buys another one.
+
+**Image editing and text-to-video are not opened up.** Editing needs a TOS upload presign first (this plugin has no such path); video polls a material result — a separate async pipeline. Neither ships, and neither leaves half an entry point behind.
+
 ## Model coverage
 
 This plugin serves models from Trae's **SOLO channel** (`DeepSeek-V4-Flash-Official`, `DeepSeek-V4-Pro-Official`, `GLM-5.3`, `GLM-5.2`, `Kimi-K3`, `MiniMax-M3`, `Qwen3.8-Max`, `Doubao-Seed-*`, …), for both the domestic and international sides.

package/README.md

@@ -46,6 +46,7 @@
 - **倍率内嵌模型名** —— 模型名称按 Trae 自身菜单的格式显示积分倍率（如 `GLM-5.2 · x0.79`），倍率随目录刷新更新。
 - **活动优惠徽标** —— Trae 把促销规则发在目录里（`activity_discount`：折扣比例、原价、闲时时段与时区），插件按**当前时钟**判定后，在模型名旁边画一枚胶囊：`5 折` / `闲时 5 折` / `限时 5 折`，悬停写明原价、现价与时段。判定放在**渲染时刻**而不是解析时刻，所以夜间折扣不会在第二天早上继续挂着；上游说「此刻没打折」就不画，**不从倍率反推**，也不给没有结束时间的活动编「限时」。这是纯提示：不改请求、不改计费，国际区（订阅制）没有这项数据因此不显示。详见 [docs/PROMO_BADGE.md](docs/PROMO_BADGE.md)。**同一枚胶囊也画在 DSH 的模型选择器里**（宿主 `conversation.input.model.label` 插槽），挑模型时就能看见——毕竟徽标画在设置卡片里时，你已经选完了。选择器里**倍率只出现一次**：名字里已有的目录倍率会被拆出来，有折扣时换成**折后价**（与卡片同一套时钟判定，同一份数据），打开选择器不会额外触发上游读。
 - **DSH 本地工具循环** —— 通过 Trae `llm_utils_chat` 获取待执行的结构化 `tool_calls`，交由 DSH 自带的本地工具执行，再将工具结果回传模型。
+- **文生图（`trae_draw_image`）** —— 把登录账号的 Trae 文生图能力接进会话：模型画一张图，插件回你 Trae 的图片对象 key（图在 Trae 客户端里看，字节插件拿不到）。**默认关闭、按面板单独开**——生图真实消耗积分（实测 CN 个人版 `1920x1920` 一张约 8.8），所以它和识图那种「让不让看图」的开关不同，只能由你点头。协议、尺寸边界，以及为什么「取图」那条 GET 其实是又一次生成，见下方「[文生图](#文生图trae_draw_image)」。
 - **国内国际双区域自动识别** —— 自动发现 Trae CN / TRAE SOLO CN / Trae / TRAE SOLO 四个本地安装的登录账号；区域由凭证自带的 `userRegion` 声明自动判定（host 后缀与 edition 标签兜底），无需手动指定。
 - **目录与勾选按面板隔离** —— 三个面板各一套模型目录、勾选、图片开关与上下文预算，各自读各自的槽位。名录从**该面板所钉的那条网关**读取——原本钉在企业池的客户端去读消费池名录，会永远发现不了自己池的独占模型。
 - **识图能力跟随 Trae 自身的标注** —— Trae 把某个模型标为 `multimodal` 时，插件直接按多模态对外声明，无需你在卡片上逐一手勾（issue #16）。判据是「**你手动勾选的 ∪ Trae 标注多模态的**」：Trae 明确标 `false` 或未标注的模型保持纯文本，你可以照旧手动为它们打开图片输入。卡片上被自动开启的那一行显示为已勾选且只读，并说明原因。详见下方「图片输入（识图）」。
@@ -120,6 +121,27 @@ DSH PiAiAdapter（每个 provider 一套）
 - **被自动开启的行在卡片上显示为已勾选且只读。** 因为并集语义下取消勾选会被立刻重新加上，给它一个可点的框是在撒谎。目前不为单个模型提供「强制关闭」的出口；如果你需要这个开关，可以在 issue #16 下提出。
 - 手动勾选对 Trae 标注 `false`（或未标注）的模型依然有效，这是为那些「上游没标但实际可用」的情况留的口子。
 
+## 文生图（`trae_draw_image`）
+
+TraeWork / SOLO 侧那套文生图服务带出来了：在会话里让模型**画一张图**。插件拿得到的是 Trae 的图片对象 key（图在 Trae 客户端里看），拿不到图片字节——原因见下。
+
+**默认关闭，按面板单独开。** 生图**真实消耗积分**（实测 CN 个人版 `1920x1920` 一张约 8.8 积分），所以它不是识图那种「让不让看图」的声明，而是一个要你点头的支出口：卡片模型栏勾上「允许模型自主生图」（`drawImageEnabled`），该面板的 `trae_draw_image` 才注册进 DSH；取消勾选立即撤下。面板各存各的开关；企业版与国际版面板不提供——企业版租户网关是另一条计费面，国际版按另一套套餐计费，两者的生图行为都没有测过，不在这里假装支持。
+
+**口径（全部实测，不是从二进制里猜的）：**
+
+| 项目 | 事实 |
+|---|---|
+| 端点 | `POST /api/ide/v1/tool_text_to_image`（另有 SSE 变体 `..._stream`），模型 `seedream_5` |
+| 请求体 | 只有 `{prompt, image_size?}`。多带已知字段（`mode_type` / `session_id` / `image_urls` / `watermark`）网关直接回 400 空 body——未知字段反而被忽略 |
+| 成功响应 | 扁平结构 `{"message":"success","uri":"tos-cn-i-…/cn/<uid>/gen_image/<ts>_<hex>_jpeg_1920x1920"}`，**没有** `{code,msg,data}` 信封 |
+| 尺寸 | 总像素必须落在 **3686400 ~ 16777216**：`1024x1024` 太小被拒（`code 3003`，不扣费）、`4608x4608` 太大被拒；预设 `2k`/`3k`/`4k` 可用，缺省 `1920x1920` |
+| 取图 | **没有可用的取图通道。** `uri` 是对象 key，不是可直接打开的 URL；同路径的 GET 不是取图，是又一次生成（见下一行），所以工具**只回 key、不再请求**，把 key 原样交给用户，图在 Trae 客户端的历史里能看到 |
+| 那条 GET 的真实身份 | 决定性探针：POST 拿到 `…/1792954795741_44f94116_jpeg_1920x1920` 后，用同一 `(prompt, size)` 打 GET，回来的是**另一个** uri（`…/1792954810505_b107ac11_jpeg_1920x1920`），耗时 27 秒、一个像素都没有——等于又画一张、再扣一次（两次共约 17.6）。非法尺寸探针（`100x100`）也印证：报错直接来自 Ark 的 `[callArkImageGeneration]` |
+
+**只回 key，不取图。** 工具成功时返回 `url`（Trae 对象 key，形如 `tos-cn-i-…/cn/<uid>/gen_image/…`）、`model`、`prompt`、`size` 和一句 `hint`；渲染层把 key 显示成**行内代码**并注明「不能直接打开」，绝不拼进 `![](...)`——key 不是路径，拼进去只会得到以当前目录为基准的碎图标。取不到字节这件事在这里不是错误，是能力边界：真的取字节要另一条通道（iCube material 的 `get_url` + `x-icube-token`，基址来自运行时 boot config），还没打通。失败仍有 30 秒冷却——生图按张计费，重试等于再买一张。
+
+**改图与文生视频没有放开。** 改图要先走 TOS 上传预签名（本插件没有这条链路）；视频要轮询素材结果，是另一条异步管线。两者都不做，也不留半截入口。
+
 ## 模型覆盖范围
 
 插件提供的是 Trae **SOLO 通道**的模型（`DeepSeek-V4-Flash-Official`、`DeepSeek-V4-Pro-Official`、`GLM-5.3`、`GLM-5.2`、`Kimi-K3`、`MiniMax-M3`、`Qwen3.8-Max`、`Doubao-Seed-*` 等），三个面板均走这条通道。企业池另有一份名录（例如 `DeepSeek-V4.1-Flash` 只有企业池列），所以两边勾选互不覆盖。

package/lib/client.js

@@ -369,6 +369,48 @@ window.__ModuleLoader__.load({
 			});
 		}
 		/**
+		* Panels whose text-to-image path this plugin has actually MEASURED, in
+		* `TRAE_PANELS` tab order.
+		*
+		* ONE entry today. The other two are refused for the SAME reason — no evidence
+		* — which is exactly the asymmetry the review of #62 held the PR on:
+		*
+		*  - `enterprise` bills through the tenant gateway, a second billing surface
+		*    whose image behaviour has never been probed;
+		*  - `ai` answers on `coresg-normal.trae.ai`, where no `tool_text_to_image`
+		*    request has ever run either — every measurement behind this tool was
+		*    taken against `trae-api-cn.mchost.guru`. Refusing one unverified gateway
+		*    while serving another is not a rule, it is an accident.
+		*
+		* Adding a panel means adding its evidence in the same change (the precedent
+		* for international claims is `docs/INTL_SG_EVIDENCE.md`). Order is
+		* significant: the tool's resolver walks `TRAE_PANELS` order and takes the
+		* FIRST consenting panel, so `cn` — the only measured one — always answers
+		* first.
+		*/
+		const DRAW_IMAGE_PANELS = ["cn"];
+		/**
+		* Build the next `regions` settings value for one panel's 「允许模型自主生图」
+		* switch (`drawImageEnabled`, `trae_draw_image`).
+		*
+		* Same merge rule as every sibling region write: only the addressed panel's
+		* `drawImageEnabled` changes, so flipping the switch carries that slot's model
+		* directory, selection, budgets, pool and credit seat through, and leaves the
+		* other panels alone. The flag is a PANEL's consent, not the account's: cn and
+		* enterprise bill the same Trae CN account but route through different gateways,
+		* so consent must not cross the boundary (the tool resolver reads it per panel).
+		*
+		* `value` may be the whole settings section or the `regions` map; the RETURN
+		* value is always the `regions` map, i.e. exactly what `settingsScope.set(
+		* 'regions', ...)` needs.
+		*/
+		function nextRegionDrawImage(value, panel, enabled) {
+			return nextRegionSlots(regionsMapOf(value), panel, {
+				...regionSlotOf(value, panel),
+				drawImageEnabled: enabled
+			});
+		}
+		/**
 		* Build the next `regions` settings value for one panel's MODEL DIRECTORY.
 		*
 		* The model-directory save REPLACES the whole region slot, so every field of
@@ -1195,6 +1237,14 @@ window.__ModuleLoader__.load({
    capability note. Wraps on narrow cards; keeps the two buttons visually a
    pair and distinct from the save footer below. */
 .dsm-trae-context-bulk{display:flex;flex-wrap:wrap;gap:8px;margin-top:10px}
+/* Per-panel 「自主生图」 consent (trae_draw_image). Sits under the bulk-context
+   row, above the capability note, because it is a per-REGION runtime choice like
+   the model picks above it — not a global preference. The input stays a real
+   checkbox so the card's tests can find it by its accessible name. */
+.dsm-trae-draw-image{display:flex;align-items:flex-start;gap:8px;margin-top:10px;font-size:11px;line-height:16px;cursor:pointer}
+.dsm-trae-draw-image input{flex:none;margin:2px 0 0}
+.dsm-trae-draw-image[data-disabled="true"]{cursor:not-allowed;opacity:.6}
+.dsm-trae-draw-image-hint{color:var(--dsw-alias-label-tertiary,#999)}
 /* Verdict badge. Only a real refusal is coloured as a warning; an unknown
    verdict stays neutral because an inconclusive probe is not evidence against
    the model. */
@@ -4133,6 +4183,8 @@ window.__ModuleLoader__.load({
 			const optimisticSettings = (0, react.useRef)({});
 			/** True while a preference is crossing the wire; serializes both writers. */
 			const [savingSettings, setSavingSettings] = (0, react.useState)(false);
+			/** The in-flight flip of the per-panel 「自主生图」 switch (undefined = read the document). */
+			const [drawImageDraft, setDrawImageDraft] = (0, react.useState)(void 0);
 			/**
 			* Per-account status (check-in / balances / subscription) shown in each row of
 			* the account table, keyed by region then account id. Loaded per account via
@@ -4250,6 +4302,37 @@ window.__ModuleLoader__.load({
 					if (mounted.current) setSavingSettings(false);
 				}
 			};
+			/**
+			* Flip ONE panel's 「允许模型自主生图」 switch (`drawImageEnabled`).
+			*
+			* Not a top-level preference like the two above: the flag lives INSIDE
+			* `regions.<panel>`, because it is per-panel consent and the Host's tool
+			* resolver reads it per panel. `nextRegionDrawImage` carries the rest of that
+			* slot through, so flipping this cannot drop the model directory, budgets,
+			* pool or credit seat — the failure #44 fixed for the pool.
+			*
+			* Serialized by `savingSettings` with the sibling writers: they all write the
+			* same settings document. The read-back re-reads the document, so a write that
+			* silently reverted cannot leave the box claiming a grant the Host never got —
+			* the one failure that matters here, because a box that lies about consent
+			* makes the tool's absence look like a bug.
+			*/
+			const writeDrawImage = async (value) => {
+				if (settingsScope === void 0 || settingsScope.getSnapshot().writable !== true) return;
+				setSavingSettings(true);
+				setWriteError(void 0);
+				setDrawImageDraft(value);
+				try {
+					await writeSettingsField(settingsScope, "regions", nextRegionDrawImage(settingsScope.getSnapshot().value, activeRegion, value), (readBack) => regionSlotsOf(readBack)[activeRegion]?.drawImageEnabled === value);
+				} catch (error) {
+					if (mounted.current) setWriteError(error instanceof Error ? error.message : t("row.requestFailed"));
+				} finally {
+					if (mounted.current) {
+						setDrawImageDraft(void 0);
+						setSavingSettings(false);
+					}
+				}
+			};
 			(0, react.useEffect)(() => settingsScope?.subscribe(() => {
 				setSettingsRevision((value) => value + 1);
 			}), [settingsScope]);
@@ -4854,6 +4937,26 @@ window.__ModuleLoader__.load({
 			const slotImageIds = Array.isArray(regionSlot?.imageModelIds) ? regionSlot.imageModelIds.filter((id) => typeof id === "string") : void 0;
 			const legacyImageIds = legacySlot && typeof settingsValue === "object" && settingsValue !== null && Array.isArray(settingsValue.imageModelIds) ? settingsValue.imageModelIds.filter((id) => typeof id === "string") : [];
 			const savedImageIds = new Set(slotImageIds ?? legacyImageIds);
+			/**
+			* The active panel's 「自主生图」 consent, read from the SAME place the Host
+			* reads it (`regions.<panel>.drawImageEnabled`, see regionStateOf) with any
+			* in-flight flip laid over it. There is deliberately no legacy flat fallback:
+			* this flag was born after the region split, and reading a top-level key here
+			* would show consent the Host's per-panel resolver never grants.
+			*/
+			const committedDrawImage = regionSlot?.drawImageEnabled === true;
+			const drawImageOn = drawImageDraft ?? committedDrawImage;
+			/**
+			* Whether the active panel can serve text-to-image at all. This is the SAME
+			* list the Host registers from (`DRAW_IMAGE_PANELS`, imported rather than
+			* restated), so the box can never offer a grant the Host will then refuse:
+			* a switch that ticks and does nothing is the exact failure this whole
+			* opt-in must not have. Two panels are out — `enterprise` bills through a
+			* tenant gateway, and `ai` has never been probed at all (every measurement
+			* behind the credit figures is CN), so neither may spend credits.
+			*/
+			const drawImageServable = DRAW_IMAGE_PANELS.includes(activeRegion);
+			const drawImageUnavailableHint = activeRegion === "ai" ? t("row.drawImageHintAi") : t("row.drawImageHintEnterprise");
 			const draft = drafts[activeRegion];
 			const visibleModels = draft?.models ?? (status.status === "signed-in" ? status.models : []);
 			/**
@@ -5659,6 +5762,7 @@ window.__ModuleLoader__.load({
 																	}), /* @__PURE__ */ (0, react_jsx_runtime.jsx)("fieldset", {
 																		className: "dsm-trae-context-budget",
 																		"aria-label": t("row.contextBudget"),
+																		title: t("row.contextBudgetHint"),
 																		children: contextBudgetOptions(model).map((option) => {
 																			const selectable = model.maxContextWindow !== void 0;
 																			const checked = (activeContextBudgets[model.id] ?? model.contextWindow) === option;
@@ -5726,6 +5830,26 @@ window.__ModuleLoader__.load({
 													})
 												]
 											}),
+											/* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
+												className: "dsm-trae-model-capability-note",
+												children: t("row.contextBudgetHint")
+											}),
+											/* @__PURE__ */ (0, react_jsx_runtime.jsxs)("label", {
+												className: "dsm-trae-draw-image",
+												"data-disabled": drawImageServable ? void 0 : "true",
+												title: drawImageServable ? t("row.drawImageHint") : drawImageUnavailableHint,
+												children: [/* @__PURE__ */ (0, react_jsx_runtime.jsx)("input", {
+													type: "checkbox",
+													checked: drawImageOn,
+													disabled: !drawImageServable || busy || saving || savingSettings || settingsScope?.getSnapshot().writable !== true,
+													onChange: (event) => {
+														writeDrawImage(event.target.checked);
+													}
+												}), /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("span", { children: [t("row.drawImage"), /* @__PURE__ */ (0, react_jsx_runtime.jsxs)("span", {
+													className: "dsm-trae-draw-image-hint",
+													children: [" ", drawImageServable ? t("row.drawImageHint") : drawImageUnavailableHint]
+												})] })]
+											}),
 											/* @__PURE__ */ (0, react_jsx_runtime.jsx)("p", {
 												className: "dsm-trae-model-capability-note",
 												children: t("row.modelCapabilityPending")
@@ -6833,6 +6957,10 @@ window.__ModuleLoader__.load({
 			"row.modelsTestAllHint": "The「test all」button probes every model in the catalog, enabled or not, so it spends noticeably more quota than the enabled-only run.",
 			"row.showPointsInMainUi": "Show credits in the composer row",
 			"row.showPointsInMainUiHint": "A clickable readout in the composer tool row while a model from this plugin is selected: Trae CN · <balance> on the CN side, Trae Global · <plan> on the international one.",
+			"row.drawImage": "Let the model draw images (trae_draw_image)",
+			"row.drawImageHint": "Adds a trae_draw_image tool to this panel: the model can generate one image when you ask for one, at this account's real credit cost (measured ~8.8 for 1920x1920 on Trae CN). Off by default — while this is unticked the tool does not exist, so the model cannot spend credits on its own.",
+			"row.drawImageHintEnterprise": "Not available on the enterprise panel: that path bills through the tenant gateway and has not been verified.",
+			"row.drawImageHintAi": "Not available on the international panel: image generation has only been measured on Trae CN, and this path bills on a different plan, so the credit figures shown here would be wrong. Switch to the CN panel to use it.",
 			"row.showPointsInMainUiHintAi": "A clickable readout in the composer tool row while a model from this plugin is selected: Trae Global · <plan>.",
 			"row.creditPositionLabel": "position",
 			"row.creditPositionComposer": "composer toolbar",
@@ -6898,6 +7026,7 @@ window.__ModuleLoader__.load({
 			"row.veilSaving": "Saving changes…",
 			"row.modelContext": "Max context {context}",
 			"row.contextBudget": "DSH context budget",
+			"row.contextBudgetHint": "Affects the DSH-side context budget only. The plugin does not send a window or Max-mode field upstream, so this does not change the window Trae serves the request with.",
 			"row.modelOutput": "Output {output}",
 			"row.modelImage": "Image",
 			"row.modelImageAuto": "Enabled automatically: Trae marks this model multimodal.",
@@ -7213,6 +7342,10 @@ window.__ModuleLoader__.load({
 			"row.showPointsInMainUi": "在输入框工具栏显示积分读数",
 			"row.showPointsInMainUiHint": "在输入框工具栏显示可点击读数（选中本插件模型时出现）：国内版为「Trae CN · 余额」，国际版为「Trae 国际版 · 套餐」。",
 			"row.showPointsInMainUiHintAi": "在输入框工具栏显示可点击读数（选中本插件模型时出现）：国际版为「Trae 国际版 · 套餐」。",
+			"row.drawImage": "允许模型自主生图（trae_draw_image）",
+			"row.drawImageHint": "为本面板挂上 trae_draw_image 工具：你要求时模型能生成一张图，按本账号真实扣积分（国内版实测 1920x1920 约 8.8）。默认关闭——不勾选时工具根本不存在，模型无法自己花额度。",
+			"row.drawImageHintEnterprise": "企业面板不提供：那条路径走租户网关计费，未经验证。",
+			"row.drawImageHintAi": "国际版面板不提供：生图只在 Trae CN 实测过，国际版按另一套套餐计费，这里标注的积分价格会对不上。要用请切到国内版面板。",
 			"row.creditPositionLabel": "显示位置",
 			"row.creditPositionComposer": "输入框工具栏",
 			"row.creditPositionStats": "输入框下方",
@@ -7277,6 +7410,7 @@ window.__ModuleLoader__.load({
 			"row.veilSaving": "正在保存更改…",
 			"row.modelContext": "最大上下文 {context}",
 			"row.contextBudget": "DSH 上下文预算",
+			"row.contextBudgetHint": "只影响 DSH 侧的上下文预算。插件不会向上游发送窗口或 Max 模式字段，所以这不会改变 Trae 实际服务这次请求所用的窗口。",
 			"row.modelOutput": "最大输出 {output}",
 			"row.modelImage": "图片",
 			"row.modelImageAuto": "已自动开启：Trae 把该模型标为多模态。",

package/lib/index.d.ts

@@ -44,7 +44,23 @@ interface TraeActivityDiscount {
 }
 //#endregion
 //#region src/reasoning.d.ts
-declare const TRAE_REASONING_EFFORTS: readonly ["minimal", "low", "medium", "high", "xhigh"];
+/**
+ * The DSH thinking levels this plugin can carry, in ascending order.
+ *
+ * `max` is the fifth level Trae's own client advertises — its bundle ships the
+ * i18n keys `trae-chat-core.model.reasoning_effort.{low,medium,high,extra_high,max}`
+ * and `ai.model.reasoning_effort.{low,medium,high,xhigh,max}`, and publishes the
+ * per-model list from a remote config (`AI.agent.model.reasoning_effort_level_by_agent_model_v2`).
+ * Leaving it out of this list meant a level the vendor could start serving was
+ * filtered away at PARSE time, before anything downstream could decide (issue
+ * #60). Listing it does not fabricate the level: `supported` is still built
+ * only from what the upstream response actually contains, so a model that does
+ * not advertise `max` never offers it.
+ *
+ * `minimal` is kept for the same reason in reverse — it is a DSH level, so the
+ * type must be able to represent it even though no CN model has advertised it.
+ */
+declare const TRAE_REASONING_EFFORTS: readonly ["minimal", "low", "medium", "high", "xhigh", "max"];
 type TraeReasoningEffort = typeof TRAE_REASONING_EFFORTS[number];
 interface TraeReasoningCapability {
   supported: readonly TraeReasoningEffort[];
@@ -4411,6 +4427,13 @@ interface TraeRegionState {
   enabledModelIds?: string[];
   /** Models the user explicitly enabled for image input in this region. */
   imageModelIds?: string[];
+  /**
+   * Whether the model may call trae_draw_image for this panel. Opt-IN:
+   * image generation spends real credits, so an absent value is never consent
+   * (the same rule workbuddy's imageToolEnabled follows). Declared in the
+   * schema below — a field missing from the schema is dropped on every save.
+   */
+  drawImageEnabled?: boolean;
   /** Local DSH context budget per model in this region. */
   contextBudgets?: Record<string, number>;
   /**

package/lib/index.js

@@ -282,6 +282,53 @@ function displayKey(name) {
 	return name.trim().toLowerCase();
 }
 /**
+* DSH thinking level -> the `reasoning_effort` name Trae's wire accepts.
+*
+* The INVERSE of `EFFORT_MAP` in `model-metadata.ts`, and it must stay an exact
+* mirror of it: a level this table does not name has no wire spelling, so it is
+* DROPPED rather than guessed (issue #60).
+*
+* The previous form was a ternary chain ending in a bare `'high'`:
+*
+*     effort === 'low' ? 'light' : effort === 'xhigh' ? 'extra_high' : 'high'
+*
+* so EVERY unnamed level was sent as `high`. That is not a harmless default: it
+* is a silent downgrade. Trae's own vocabulary carries five levels
+* (`low / medium / high / extra_high / max`, verified in the client bundle's
+* `reasoning_effort` i18n keys), and its remote config publishes the list per
+* model (`AI.agent.model.reasoning_effort_level_by_agent_model_v2`). The day
+* upstream starts advertising `max`, a user picking Max would have sent `high`
+* — no error, no log, and a cheaper answer than the one they asked for. A
+* dropped field lets the upstream apply its own default, which is at least the
+* model's real behaviour instead of a wrong claim about it.
+*
+* Measured 2026-10-25 on a live CN credential (98 directory rows): options are
+* `["light","high"]`, `["high","extra_high"]`, `["light","high","extra_high"]`
+* or absent — `max` is not published YET, which is exactly why this must not
+* wait for it to be.
+*/
+const WIRE_EFFORT_MAP = {
+	low: "light",
+	medium: "medium",
+	high: "high",
+	xhigh: "extra_high",
+	max: "max"
+};
+/**
+* Map a model's advertised levels to their wire names, dropping unnamed ones.
+*
+* `flatMap` rather than `map` + a placeholder: an unmappable level must not
+* produce an entry at all, because every consumer reads "is there a value here"
+* as "may this level be sent" (see `thinkingLevelMap` in `adapter.ts` and the
+* `delete input['reasoning_effort']` fallback in `solo-bridge.ts`).
+*/
+function wireEffortsOf(supported) {
+	return Object.fromEntries(supported.flatMap((effort) => {
+		const wire = WIRE_EFFORT_MAP[effort];
+		return wire === void 0 ? [] : [[effort, wire]];
+	}));
+}
+/**
 * Merge the two Trae model sources into one authoritative catalog.
 *
 * `remote` (the solo.trae.cn `/models` directory) is the authoritative model
@@ -360,7 +407,7 @@ function mergeTraeModelSources(remote, wire) {
 			reasoningSupported: model.reasoningSupported,
 			...model.reasoning === void 0 ? {} : {
 				reasoning: model.reasoning,
-				reasoningEfforts: Object.fromEntries(model.reasoning.supported.map((effort) => [effort, effort === "low" ? "light" : effort === "xhigh" ? "extra_high" : "high"]))
+				reasoningEfforts: wireEffortsOf(model.reasoning.supported)
 			},
 			...wireModel !== void 0 && wireModel.id !== "" && wireModel.id !== model.id ? { wireConfigName: wireModel.id } : {},
 			...wireFunction === void 0 ? {} : { wireFunction }
@@ -433,7 +480,7 @@ function mergeTraeModelSources(remote, wire) {
 					supported: [...wireModel.reasoning.supported],
 					...wireModel.reasoning.defaultEffort === void 0 ? {} : { defaultEffort: wireModel.reasoning.defaultEffort }
 				},
-				reasoningEfforts: Object.fromEntries(wireModel.reasoning.supported.map((effort) => [effort, effort === "low" ? "light" : effort === "xhigh" ? "extra_high" : "high"]))
+				reasoningEfforts: wireEffortsOf(wireModel.reasoning.supported)
 			},
 			...wireModel.function === void 0 ? {} : { wireFunction: wireModel.function }
 		});
@@ -507,7 +554,7 @@ function discoveredCatalog(models) {
 		reasoningSupported: model.reasoningSupported,
 		...model.reasoning === void 0 ? {} : {
 			reasoning: model.reasoning,
-			reasoningEfforts: Object.fromEntries(model.reasoning.supported.map((effort) => [effort, effort === "low" ? "light" : effort === "xhigh" ? "extra_high" : "high"]))
+			reasoningEfforts: wireEffortsOf(model.reasoning.supported)
 		}
 	});
 	return result;
@@ -836,14 +883,28 @@ function toPiModel(info, baseUrl, providerId, withRate = true) {
 			cacheWrite: 0
 		},
 		reasoning: info.reasoningEfforts !== void 0,
-		...info.reasoningEfforts === void 0 ? {} : { thinkingLevelMap: {
+		...info.reasoningEfforts === void 0 ? {} : { 
+		/**
+		* Which DSH levels this model may be asked for, and under which wire name.
+		*
+		* Every level is read from the model's OWN advertised capability, never
+		* hardcoded. `minimal / medium / max` used to be pinned to `null`, so a
+		* level the upstream published could not be selected even after it was
+		* parsed — the second of the three places issue #60 reported. `null`
+		* still means "not offered", which is what a model that does not
+		* advertise the level reports: nothing here fabricates a level.
+		*
+		* `off` stays `null` deliberately: Trae's wire has no "no reasoning"
+		* spelling, so an explicit off is not something this provider can send.
+		*/
+thinkingLevelMap: {
 			off: null,
-			minimal: null,
+			minimal: info.reasoningEfforts.minimal ?? null,
 			low: info.reasoningEfforts.low ?? null,
-			medium: null,
+			medium: info.reasoningEfforts.medium ?? null,
 			high: info.reasoningEfforts.high ?? null,
 			xhigh: info.reasoningEfforts.xhigh ?? null,
-			max: null
+			max: info.reasoningEfforts.max ?? null
 		} },
 		...info.contextWindow === void 0 ? {} : { contextWindow: info.contextWindow },
 		...info.maxTokens === void 0 ? {} : { maxTokens: info.maxTokens },
@@ -1499,6 +1560,64 @@ function nextRegionCreditPosition(value, panel, position) {
 		creditPosition: position
 	});
 }
+/**
+* Panels whose text-to-image path this plugin has actually MEASURED, in
+* `TRAE_PANELS` tab order.
+*
+* ONE entry today. The other two are refused for the SAME reason — no evidence
+* — which is exactly the asymmetry the review of #62 held the PR on:
+*
+*  - `enterprise` bills through the tenant gateway, a second billing surface
+*    whose image behaviour has never been probed;
+*  - `ai` answers on `coresg-normal.trae.ai`, where no `tool_text_to_image`
+*    request has ever run either — every measurement behind this tool was
+*    taken against `trae-api-cn.mchost.guru`. Refusing one unverified gateway
+*    while serving another is not a rule, it is an accident.
+*
+* Adding a panel means adding its evidence in the same change (the precedent
+* for international claims is `docs/INTL_SG_EVIDENCE.md`). Order is
+* significant: the tool's resolver walks `TRAE_PANELS` order and takes the
+* FIRST consenting panel, so `cn` — the only measured one — always answers
+* first.
+*/
+const DRAW_IMAGE_PANELS = ["cn"];
+/**
+* Whether ONE panel's 「允许模型自主生图」 grant is in force, i.e. its box was
+* ticked. ONLY the literal `true` counts: a hand-edited `'yes'` or a `1` is a
+* typo, not consent to spend credits. (A `{ get(): true }` live reference — the
+* DSH 0.1.7 slot shape — IS decoded to the true it holds first, see
+* {@link unwrapVolatileDeep}: that is the Host's own document, not a hand edit.)
+*
+* `value` may be the whole settings section or the `regions` map (see
+* {@link regionsMapOf}).
+*/
+function drawImageEnabledOf(value, panel) {
+	return regionSlotOf(value, panel)["drawImageEnabled"] === true;
+}
+/**
+* The panels that may register `trae_draw_image` RIGHT NOW — the single
+* judgment behind both the tool's registration and its account resolver, and
+* therefore the only place a user's credits can be exposed by accident.
+*
+* Pure and exported on purpose (review of #62, item 3): the first version kept
+* this rule inside `apply()`'s closure, so "does ticking the box actually
+* grant anything" could only be code-reviewed, never tested.
+*
+* THREE gates, all required: the panel must be measured
+* ({@link DRAW_IMAGE_PANELS}), its provider must be switched ON (opt-out rule,
+* so a config written before the switch existed still counts), and the panel
+* must carry an explicit grant. A tick on a switched-off provider is STORED
+* consent, not a live grant: flipping the provider back on restores the tool
+* without asking again, which is why both gates are read here rather than at
+* the checkbox.
+*
+* `enterprise` is refused structurally, even if someone later adds it to
+* {@link DRAW_IMAGE_PANELS}: consent must never cross into the tenant billing
+* surface by editing a list.
+*/
+function drawImagePanelsOf(value) {
+	return TRAE_PANELS.filter((panel) => panel !== "enterprise" && DRAW_IMAGE_PANELS.includes(panel) && regionEnabledOf(value, panel) && drawImageEnabledOf(value, panel));
+}
 //#endregion
 //#region src/decrypt.ts
 const TRAE_AUTH_STORAGE_KEY = "iCubeAuthInfo://icube.cloudide";
@@ -4752,12 +4871,29 @@ function parseTraeActivityDiscount(value) {
 }
 //#endregion
 //#region src/reasoning.ts
+/**
+* The DSH thinking levels this plugin can carry, in ascending order.
+*
+* `max` is the fifth level Trae's own client advertises — its bundle ships the
+* i18n keys `trae-chat-core.model.reasoning_effort.{low,medium,high,extra_high,max}`
+* and `ai.model.reasoning_effort.{low,medium,high,xhigh,max}`, and publishes the
+* per-model list from a remote config (`AI.agent.model.reasoning_effort_level_by_agent_model_v2`).
+* Leaving it out of this list meant a level the vendor could start serving was
+* filtered away at PARSE time, before anything downstream could decide (issue
+* #60). Listing it does not fabricate the level: `supported` is still built
+* only from what the upstream response actually contains, so a model that does
+* not advertise `max` never offers it.
+*
+* `minimal` is kept for the same reason in reverse — it is a DSH level, so the
+* type must be able to represent it even though no CN model has advertised it.
+*/
 const TRAE_REASONING_EFFORTS = [
 	"minimal",
 	"low",
 	"medium",
 	"high",
-	"xhigh"
+	"xhigh",
+	"max"
 ];
 function parseReasoningCapability(value) {
 	if (typeof value !== "object" || value === null || Array.isArray(value)) return void 0;
@@ -5863,10 +5999,31 @@ function parseFeatures(value) {
 		return;
 	}
 }
+/**
+* Trae's wire effort name -> the DSH thinking level.
+*
+* The FORWARD half of the mapping; `WIRE_EFFORT_MAP` in `catalog.ts` is its
+* exact inverse. Both halves must name the same levels, or a level survives one
+* direction and vanishes in the other.
+*
+* `medium` and `max` are here because Trae's own client advertises them
+* (`reasoning_effort.{low,medium,high,extra_high,max}` in its i18n bundle) and
+* publishes the per-model list remotely — so the day upstream starts sending
+* either, the parser must not filter it away before anything can use it (issue
+* #60). Measured 2026-10-25 on a live CN credential, the published options were
+* only `light / high / extra_high`; these entries are therefore inert today and
+* ready tomorrow, not a claim that the levels are served now.
+*
+* An unknown name is deliberately ABSENT rather than defaulted: `parse` drops
+* it, and the reverse map drops it too, so an unrecognised level can never be
+* reported to the user or sent upstream under a name the vendor never said.
+*/
 const EFFORT_MAP = {
 	light: "low",
+	medium: "medium",
 	high: "high",
-	extra_high: "xhigh"
+	extra_high: "xhigh",
+	max: "max"
 };
 /**
 * Parse only capabilities explicitly advertised by Trae's remote model API.
@@ -9110,7 +9267,7 @@ async function checkinOne(target, deps) {
 		return {
 			...base,
 			status: "failed",
-			message: messageOf(error)
+			message: messageOf$1(error)
 		};
 	}
 	if (credential === void 0) return {
@@ -9125,7 +9282,7 @@ async function checkinOne(target, deps) {
 		return {
 			...base,
 			status: "failed",
-			message: messageOf(error)
+			message: messageOf$1(error)
 		};
 	}
 	if (!status.enabled) return {
@@ -9161,7 +9318,7 @@ async function checkinOne(target, deps) {
 		return {
 			...base,
 			status: "failed",
-			message: messageOf(error)
+			message: messageOf$1(error)
 		};
 	}
 }
@@ -9213,7 +9370,7 @@ async function testOne(target, modelId, probe, deps) {
 		return {
 			...base,
 			outcome: "auth",
-			detail: messageOf(error)
+			detail: messageOf$1(error)
 		};
 	}
 	if (credential === void 0) return {
@@ -9281,7 +9438,7 @@ function defaultWait(ms) {
 	return new Promise((resolve) => setTimeout(resolve, ms));
 }
 /** A safe message from an unknown throwable, bounded in length. */
-function messageOf(error) {
+function messageOf$1(error) {
 	return (error instanceof Error ? error.message : String(error)).slice(0, 300);
 }
 //#endregion
@@ -9748,6 +9905,316 @@ var UnconfiguredTraeUpstreamClient = class {
 	}
 };
 //#endregion
+//#region src/image-gen-trae.ts
+/** 文生图（同步）：一次 POST 拿回结果。Trae 客户端的 ToolText2Image 用的就是它。 */
+const TRAE_TOOL_TEXT_TO_IMAGE_PATH = "/api/ide/v1/tool_text_to_image";
+/** 上游为图像工具固定的模型 id；请求体里没有 model 字段，服务端自选。 */
+const TRAE_IMAGE_MODEL = "seedream_5";
+/** 尺寸预设：活体确认 '2k','3k','4k' 被 Ark 接受，square/square_hd 等
+*  客户端预设名服务端也会翻译（square_hd 实测出 1920x1920）。 */
+const TRAE_IMAGE_SIZE_PRESETS = [
+	"2k",
+	"3k",
+	"4k",
+	"square",
+	"square_hd",
+	"landscape",
+	"portrait"
+];
+var TraeImageError = class extends Error {
+	code;
+	/** 上游业务码（响应里的 code 字段），只在 UPSTREAM 类失败上有值。 */
+	serverCode;
+	constructor(message, code, serverCode) {
+		super(message);
+		this.name = "TraeImageError";
+		this.code = code;
+		this.serverCode = typeof serverCode === "number" ? serverCode : void 0;
+	}
+};
+/** 尺寸合法性：预设名，或 宽x高 且总像素在 Ark 的 [下限,上限] 区间。 */
+function isTraeImageSize(value) {
+	if (TRAE_IMAGE_SIZE_PRESETS.includes(value)) return true;
+	const m = /^([0-9]{3,5})x([0-9]{3,5})$/.exec(value);
+	if (m === null) return false;
+	const pixels = Number(m[1]) * Number(m[2]);
+	return pixels >= 3686400 && pixels <= 16777216;
+}
+/** 组一次文生图请求体；参数毛病全在这里拦下（INVALID_ARGS，不花积分）。 */
+function buildTraeImageBody(input) {
+	const prompt = input.prompt.trim();
+	if (prompt === "") throw new TraeImageError("trae image: prompt 必须是非空字符串。", "INVALID_ARGS");
+	const out = { prompt };
+	if (input.size !== void 0 && input.size !== "") {
+		const size = input.size.trim().toLowerCase();
+		if (!isTraeImageSize(size)) throw new TraeImageError("trae image: 尺寸「" + size + "」不被接受（需 宽x高 且总像素 ∈ [3686400,16777216]，或预设 " + TRAE_IMAGE_SIZE_PRESETS.join("/") + "）。", "INVALID_ARGS");
+		out.image_size = size;
+	}
+	return out;
+}
+/** 应答里图片地址的字段名。活体看到的就是 uri；其余字段是 Rust 端
+*  ToolcallResultImage 的字段池（tos_uri/display_url/…），留给未来形态。 */
+const TRAE_IMAGE_URL_FIELDS = [
+	"uri",
+	"tos_uri",
+	"display_url",
+	"image_url",
+	"original_url",
+	"preview_url",
+	"internal_url"
+];
+/** 深度不超过 4 地扫第一个可用 URL 字段；再深的嵌套不是这个 API 的形态。 */
+function findImageUrl(node, depth = 0) {
+	if (depth > 4 || node === null || typeof node !== "object") return void 0;
+	const record = node;
+	for (const field of TRAE_IMAGE_URL_FIELDS) {
+		const value = record[field];
+		if (typeof value === "string" && value.trim() !== "") return value.trim();
+	}
+	for (const value of Object.values(record)) if (value !== null && typeof value === "object") {
+		const found = findImageUrl(value, depth + 1);
+		if (found !== void 0) return found;
+	}
+}
+function findCredit(node, depth = 0) {
+	if (depth > 4 || node === null || typeof node !== "object") return void 0;
+	const record = node;
+	const credit = record.credit;
+	if (typeof credit === "number" && Number.isFinite(credit)) return credit;
+	for (const value of Object.values(record)) if (value !== null && typeof value === "object") {
+		const found = findCredit(value, depth + 1);
+		if (found !== void 0) return found;
+	}
+}
+/**
+* 解析 tool_text_to_image 的扁平 JSON 应答。
+*
+* 实测形态：成功 {message:"success", uri:"…"}；失败 {code:3003, message:"…"}。
+* code 存在且非 0 → UPSTREAM（原样抛出 code/message——余额不足、风控、
+* Ark 参数错误全靠它浮出）；无 code 但 uri 缺失 → EMPTY_RESULT。
+*/
+function parseTraeImageResponse(json) {
+	const record = json !== null && typeof json === "object" ? json : {};
+	const code = typeof record.code === "number" ? record.code : void 0;
+	const message = typeof record.message === "string" ? record.message : void 0;
+	if (code !== void 0 && code !== 0) throw new TraeImageError("trae image: 上游业务错误 code=" + code + (message === void 0 || message === "" ? "" : ": " + message), "UPSTREAM", code);
+	const url = findImageUrl(record);
+	if (url === void 0) throw new TraeImageError("trae image: 应答里没有图片地址" + (message === void 0 ? "" : "（message=" + message + "）"), "EMPTY_RESULT", code);
+	const credit = findCredit(record);
+	return credit === void 0 ? { url } : {
+		url,
+		credit
+	};
+}
+//#endregion
+//#region src/image-tool.ts
+/** 工具名。与其它 connect 插件的命名习惯一致。 */
+const TRAE_IMAGE_TOOL_NAME = "trae_draw_image";
+/** 单次生图上限。Trae 客户端里一张图 15-45 秒，留足余量。 */
+const IMAGE_TOOL_TIMEOUT_MS = 18e4;
+/** 失败后的冷却：生图按张扣积分，连续重试会白白烧积分。 */
+const FAILURE_COOLDOWN_MS = 3e4;
+function messageOf(error) {
+	return error instanceof Error ? error.message : String(error);
+}
+/** HTTP 层的状态码映射。语义与 workbuddy 侧一致。 */
+function mapHttpToImageError(status, detail) {
+	const tail = detail === "" ? "" : "：" + detail.slice(0, 200);
+	if (status === 401 || status === 403) return new TraeImageError("trae image: 上游拒绝了凭据（HTTP " + status + "）" + tail, "AUTH");
+	if (status === 429) return new TraeImageError("trae image: 上游限流（HTTP 429）" + tail, "RATE_LIMIT");
+	return new TraeImageError("trae image: 上游 HTTP " + status + tail, "UPSTREAM");
+}
+/**
+* 发一次真机文生图请求。所有可预期的失败都归一成 TraeImageError，
+* 让上层的冷却判据只看 code（见 createTraeDrawImageTool 的注释）。
+*/
+var TraeImageClient = class {
+	options;
+	constructor(options = {}) {
+		this.options = options;
+	}
+	async generate(account, input, signal) {
+		const body = buildTraeImageBody(input);
+		const endpoint = traeEndpoint(account.baseUrl, TRAE_TOOL_TEXT_TO_IMAGE_PATH);
+		const headers = buildTraeHeaders(account.credential, account.identity, { profile: "model-detail" });
+		const fetchImpl = this.options.fetchImpl ?? globalThis.fetch;
+		const timeoutMs = this.options.timeoutMs ?? IMAGE_TOOL_TIMEOUT_MS;
+		const requestSignal = signal ?? AbortSignal.timeout(timeoutMs);
+		let response;
+		try {
+			response = await fetchImpl(endpoint, {
+				method: "POST",
+				headers,
+				body: JSON.stringify(body),
+				signal: requestSignal
+			});
+		} catch (error) {
+			throw new TraeImageError("trae image: 生图请求发出失败：" + messageOf(error), "NO_ROUTE");
+		}
+		let payload;
+		try {
+			payload = await response.text();
+		} catch (error) {
+			throw new TraeImageError("trae image: 读取应答失败：" + messageOf(error), "UPSTREAM");
+		}
+		if (!response.ok) throw mapHttpToImageError(response.status, payload);
+		let json;
+		try {
+			json = JSON.parse(payload);
+		} catch {
+			throw new TraeImageError("trae image: 应答不是 JSON（HTTP 200）：" + payload.slice(0, 200), "UPSTREAM");
+		}
+		const outcome = parseTraeImageResponse(json);
+		const result = { url: outcome.url };
+		if (outcome.credit !== void 0) result.credit = outcome.credit;
+		return result;
+	}
+};
+function asRecord(value) {
+	return typeof value === "object" && value !== null ? value : {};
+}
+function asNonEmptyString(value) {
+	return typeof value === "string" && value.trim() !== "" ? value.trim() : void 0;
+}
+/**
+* 这些失败码都发生在「生图请求已经发出去」之后或尝试发出它——上游可能已经
+* 扣了积分或排了队，连续重试只会把它们再花一遍，所以进冷却。
+* INVALID_ARGS / NOT_LOGGED_IN 一分钱没花，不冷却（判据同 workbuddy）。
+*/
+const BILLABLE_FAILURE_CODES = /* @__PURE__ */ new Set([
+	"AUTH",
+	"RATE_LIMIT",
+	"NO_ROUTE",
+	"UPSTREAM",
+	"EMPTY_RESULT"
+]);
+function isBillableFailure(error) {
+	if (error instanceof TraeImageError) return BILLABLE_FAILURE_CODES.has(error.code);
+	return true;
+}
+/** 造出工具定义。每次 apply() 只造一个实例，冷却状态挂在闭包里。 */
+function createTraeDrawImageTool(deps) {
+	let cooldownUntil = 0;
+	return {
+		name: TRAE_IMAGE_TOOL_NAME,
+		description: "用当前登录的 Trae 账号文生图（TraeWork/SOLO 同款服务，模型 seedream_5）。只在用户明确要一张图时调用：每张真实扣该账号积分，实测 CN 个人版 1920x1920 一张约 8.8 积分（尺寸越大越贵），而上游不回传扣费数字，以 Trae 额度面板为准。不是常开能力：只有插件设置里为该区打开「允许模型自主生图」(drawImageEnabled) 时本工具才存在，没打开就没得调——不要用报错去试探它开没开。取图口径：生成只回一个图片对象 key（tos uri，形如 tos-cn-i-xxx/cn/<uid>/gen_image/…，打不开），本工具不会再发取图请求——实测对同一 (prompt, size) 再 GET 一次不是缓存查询，而是又一次生成，只会再扣一张的钱并回一个**新的** uri。图片字节目前插件拿不到，图只在 Trae 客户端的历史里能看到；把 key 原样告诉用户即可。不要为了\"取图\"或\"换张\"重跑同一条 prompt。失败后有 30 秒冷却，冷却内重试不会打到上游。尺寸：默认 1920x1920，也可给 宽x高（总像素需在 3686400~16777216 之间，如 2048x2048）或预设 2k/3k/4k；过小（如 1024x1024）会被上游直接拒绝。（改图与文生视频暂未开放：改图需要先走 TOS 上传预签名，视频要轮询素材结果。）",
+		parameters: {
+			type: "object",
+			properties: {
+				prompt: {
+					type: "string",
+					description: "画面描述（中文或英文都可以）。写清楚主体、风格、构图、光线。"
+				},
+				size: {
+					type: "string",
+					description: "图片尺寸：默认 1920x1920；也可给 宽x高（总像素 3686400~16777216，如 2048x2048）或预设 2k/3k/4k。过小（如 1024x1024）会被上游拒绝。"
+				}
+			},
+			required: ["prompt"],
+			additionalProperties: false
+		},
+		output: {
+			schema: {
+				type: "object",
+				properties: {
+					url: {
+						type: "string",
+						description: "Trae 图片对象 key（tos uri），不是可直接打开的 URL；图片字节插件拿不到。"
+					},
+					model: {
+						type: "string",
+						description: "实际使用的生图模型 id。"
+					},
+					prompt: {
+						type: "string",
+						description: "实际提交的画面描述。"
+					},
+					size: {
+						type: "string",
+						description: "实际提交的尺寸。"
+					},
+					hint: {
+						type: "string",
+						description: "给模型的后续提示（如何展示、扣费口径）。"
+					},
+					credit: {
+						type: "number",
+						description: "上游若回传扣费数字，这里给出积分数。"
+					}
+				},
+				required: [
+					"url",
+					"model",
+					"prompt",
+					"size",
+					"hint"
+				],
+				additionalProperties: false
+			},
+			render(_args, value) {
+				const record = asRecord(value);
+				const url = asNonEmptyString(record["url"]) ?? "";
+				const model = asNonEmptyString(record["model"]) ?? "unknown";
+				const size = asNonEmptyString(record["size"]) ?? "";
+				const credit = typeof record["credit"] === "number" ? record["credit"] : void 0;
+				const hint = asNonEmptyString(record["hint"]) ?? "";
+				const lines = [
+					"图片对象 key（不能直接打开）：`" + url + "`",
+					"",
+					"文生图 · 模型：" + model + (size === "" ? "" : " · 尺寸：" + size) + (credit === void 0 ? "" : " · 消耗 " + credit + " 积分")
+				];
+				if (hint !== "") lines.push("", hint);
+				return [{
+					type: "text",
+					text: lines.join("\n")
+				}];
+			}
+		},
+		timeoutMs: IMAGE_TOOL_TIMEOUT_MS,
+		async execute(args, exec) {
+			const record = asRecord(args);
+			const prompt = asNonEmptyString(record["prompt"]);
+			if (prompt === void 0) throw new TraeImageError("trae_draw_image: 参数 prompt 必填，且必须是非空字符串。", "INVALID_ARGS");
+			const size = asNonEmptyString(record["size"]) ?? "1920x1920";
+			const now = Date.now();
+			if (now < cooldownUntil) throw new TraeImageError("trae image: 上一次生图刚失败过，请 " + Math.ceil((cooldownUntil - now) / 1e3) + " 秒后再试（失败冷却是为了避免连续重试白白烧积分）。", "COOLDOWN");
+			const account = await deps.resolveAccount();
+			if (account === void 0) throw new TraeImageError("trae image: 没有可用的 Trae 账号——要么相关区都没登录，要么已登录的那个区没有打开「允许模型自主生图」（drawImageEnabled）。请先登录 Trae 并在插件设置里打开那个开关。", "NOT_LOGGED_IN");
+			const signal = exec.signal === void 0 ? AbortSignal.timeout(IMAGE_TOOL_TIMEOUT_MS) : AbortSignal.any([exec.signal, AbortSignal.timeout(IMAGE_TOOL_TIMEOUT_MS)]);
+			try {
+				const outcome = await deps.generate(account, {
+					prompt,
+					size
+				}, signal);
+				const value = {
+					url: outcome.url,
+					model: TRAE_IMAGE_MODEL,
+					prompt,
+					size,
+					hint: "这张图来自 Trae 文生图（模型 seedream_5），按张计入该账号积分（实测 1920x1920 约 8.8，上游不回传数字，以 Trae 额度面板为准）。url 是 Trae 的图片对象 key，不能直接打开，图片字节插件也拿不到——这张图能在 Trae 客户端的历史/作品里看到。不要把 key 拼进 markdown 图片语法；也不要为了取图或换张重跑同一条 prompt（取图那条请求本身就是又一次生成，会再扣一张）。"
+				};
+				if (outcome.credit !== void 0) value.credit = outcome.credit;
+				return value;
+			} catch (error) {
+				if (isBillableFailure(error)) cooldownUntil = Date.now() + FAILURE_COOLDOWN_MS;
+				throw error;
+			}
+		}
+	};
+}
+/**
+* 把工具注册进宿主。返回 disposer；调用方负责在插件卸载时调用它。
+* 注册失败只记警告——生图工具缺失不该让整个插件起不来（同 workbuddy）。
+*/
+function registerTraeDrawImageTool(tools, deps) {
+	try {
+		return tools.register(createTraeDrawImageTool(deps));
+	} catch (error) {
+		deps.logger.warn("dsh-connect-trae: 注册 trae_draw_image 失败: " + messageOf(error));
+		return;
+	}
+}
+//#endregion
 //#region src/index.ts
 const name = "dsh-connect-trae";
 const inject = ["llm"];
@@ -9879,6 +10346,7 @@ const regionStateConfig = z.object({
 	lastCatalog: z.array(modelConfig).default([]),
 	enabledModelIds: z.array(z.string()).default([]),
 	imageModelIds: z.array(z.string()).default([]),
+	drawImageEnabled: z.boolean().default(false).description("Allow the model to generate images with trae_draw_image on this panel (consumes credits; default off)"),
 	contextBudgets: z.dict(z.number().step(1).min(1)).default({}),
 	creditPosition: z.union([z.const("composer"), z.const("stats")]).default("composer").description("积分读数显示位置：\"composer\" = 输入框工具栏（默认），\"stats\" = 输入框下方（与「轮次 / 步数 / token」统计同一行）。装了别的插件后读数不显示时改成 stats。"),
 	pool: z.object({
@@ -11021,8 +11489,48 @@ function apply(ctx, config) {
 		const settings = settingsCtx.settings;
 		ctx.effect(() => settings.configure({ auto: true }, ctx.fiber));
 	});
+	const drawImageClient = new TraeImageClient({ logger: ctx.logger });
+	let drawImageTools;
+	let disposeDrawImageTool;
+	const syncTraeDrawImageTool = (value) => {
+		if (drawImagePanelsOf(value).length === 0) {
+			disposeDrawImageTool?.();
+			disposeDrawImageTool = void 0;
+			return;
+		}
+		if (disposeDrawImageTool !== void 0 || drawImageTools === void 0) return;
+		disposeDrawImageTool = registerTraeDrawImageTool(drawImageTools, {
+			resolveAccount: async () => {
+				for (const panel of drawImagePanelsOf(current())) {
+					const stack = stacks[panel];
+					try {
+						return {
+							credential: await stack.store.resolve(),
+							identity: await stack.identity(),
+							baseUrl: REGION_GATEWAYS[stack.region].chat
+						};
+					} catch {
+						continue;
+					}
+				}
+			},
+			generate: (account, request, signal) => drawImageClient.generate(account, request, signal),
+			logger: ctx.logger
+		});
+	};
+	ctx.inject(["tools"], (toolsCtx) => {
+		drawImageTools = toolsCtx.tools;
+		if (drawImageTools === void 0) return;
+		syncTraeDrawImageTool(current());
+		toolsCtx.effect(() => () => {
+			disposeDrawImageTool?.();
+			disposeDrawImageTool = void 0;
+			drawImageTools = void 0;
+		});
+	});
 	ctx.on("loader/volatile-update", () => {
 		applySelection(current());
+		syncTraeDrawImageTool(current());
 	});
 	let stopped = false;
 	ctx.effect(() => () => {

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": "4.5.1",
+  "version": "4.6.0",
   "author": "dingminhua",
   "copyright": "Copyright (c) 2026 LaoDing",
   "license": "MIT",