dsh-codex-subscription 2.2.11 → 2.4.0

package/README.en.md

@@ -16,12 +16,27 @@ No OpenAI API key or Codex CLI. Models, search, quota, and image generation stay
 [![MIT](https://img.shields.io/badge/license-MIT-111111.svg)](LICENSE)
 [![Star](https://img.shields.io/github/stars/WSL043/dsh-codex-subscription?style=flat&logo=github&label=Star)](https://github.com/WSL043/dsh-codex-subscription/stargazers)
 
-[Features](#feature-overview) · [Install](#install) · [Daily use](#inside-the-plugin) · [Examples](#from-sketch-to-image) · [User guide](docs/GUIDE.en.md) · [Update and uninstall](#update-and-uninstall)
+[What's new](#whats-new-in-240) · [Features](#feature-overview) · [Install](#install) · [Daily use](#inside-the-plugin) · [Examples](#from-sketch-to-image) · [User guide](docs/GUIDE.en.md) · [Update and uninstall](#update-and-uninstall)
 
 </div>
 
 <p align="center"><img src="docs/assets/product-account-demo-en.png" width="1100" alt="Codex subscription account and quota interface with model choices, quota windows and alerts; demonstration data"></p>
 
+## What's new in 2.4.0
+
+- **Stronger agent drawing (Beta).** A new SVG drawing command lets the model draw a whole layer in the notation it knows best; the result is still editable, undoable and exportable native strokes. Approximations such as gradients or clamped points are reported, and unsupported elements are refused by name.
+- **New showcase.** GPT-6.1-Sol at Xhigh and GPT-6-Astra at Low compared side by side on a mechanical cutaway and a blue-green landscape, run on the official desktop app; see [From sketch to image](#from-sketch-to-image).
+- **Desktop app fixes.** On the official desktop app the Codex subtask component now detects the host version and installs; support diagnostics report the real DSH version; a failed image request says whether it timed out or the connection dropped.
+
+### 2.3.0
+
+- **The official desktop app works.** The official DeepSeek Harness desktop app (0.2.0-rc.2) passed a hands-on acceptance run: plugin install, sign-in, model list, hiding models, a chat with a tool call, and image generation. See [Install](#install).
+- **Models come only from your account.** The model list is read entirely from the account catalog; before sign-in or when the read fails no model is shown, and there is no built-in fallback. Under Models & runtime you can show or hide each model in the composer.
+- **Settings reorganised.** Models and search are separate cards, switches use DSH's own control, and dead buttons no longer appear while signed out.
+- **Ultrafast speed tier (Beta).** The new tier from OpenAI's developer event. It appears in the model menu only when your account catalog lists it for that model (today Pro 500 with GPT-6 Astra); other accounts never see it and it is never silently swapped for another tier. It is Beta because it has not been run on a Pro 500 account yet.
+- **Shorter support diagnostics.** "Copy report" now copies the conclusions, recent failures and per-feature status (about 1 KB); the full report can still be copied separately from the details.
+- **Subtasks and stability.** Codex subtasks now really take effect on 0.2.0 and no longer expose the apps connected to your account; image generation being rejected by the host on the desktop app is fixed (2.2.11).
+
 ## Feature overview
 
 Subscription access, model controls, quota management, and image creation in one DSH workflow. Each capability is listed below; model access and quota depend on your account.
@@ -34,7 +49,7 @@ Subscription access, model controls, quota management, and image creation in one
 | **Model catalog sync** | Read available models and reasoning levels from your account, with manual refresh; models marked for retirement show their retirement date next to the model name, without automatically switching your selection | New models appear when your account offers them; the account catalog is the only source, so signed out or when it cannot be read no models are shown and no bundled list stands in |
 | **Model list** | Show or hide individual models in the composer picker under Models & runtime | Display only; selected models and requests still work |
 | **Reasoning levels** | Choose supported reasoning effort in the composer | Levels vary by model |
-| **Fast mode** | Switch supported models between Standard and Fast, with a lightning indicator | Standard by default; Fast increases usage |
+| **Speed tiers** | Switch supported models between Standard and Fast; Ultrafast (Beta) too when the account catalog offers it, with a lightning indicator | Standard by default; Fast and Ultrafast increase usage, and Ultrafast is Pro 500 only today |
 | **Response verbosity** | Request short, medium, or detailed responses | Follows the model default unless changed |
 | **Input image detail** | Choose Low, High, Original, or Auto detail for images sent to the chat model | Affects input tokens; Auto by default |
 | **Subscription web search** | Use Codex search or choose DSH search | Auto, DSH, and Codex routing options |
@@ -70,7 +85,7 @@ Subscription access, model controls, quota management, and image creation in one
 | **Smoothing and navigation** | Smooth complete strokes after release; pan, zoom, undo, and redo | Shortcuts can be customized or disabled |
 | **Multiple drafts** | Save and reopen up to 20 local drafts | Stored in the current browser |
 | **Import and export** | Export PNG, layered PSD, or native drafts that preserve editable strokes | PSD import supports a limited set of ordinary pixel layers |
-| **Agent drawing** | Ask the agent to draw with `@sketch`, follow progress, and stop it | Beta; off by default, requires the canvas |
+| **Agent drawing** | Ask the agent to draw with `@sketch`, follow progress, and stop it; the agent can draw whole stages in SVG, imported as editable native strokes | Beta; off by default, requires the canvas |
 | **Result preview for the model** | Return a canvas image after drawing for subsequent review | Beta; off by default, adds image input usage |
 
 ### Advanced capabilities and maintenance
@@ -81,7 +96,7 @@ Subscription access, model controls, quota management, and image creation in one
 | **Subtask component management** | Install, disable, or uninstall the official Codex subtask component from settings | Managed by supported DSH hosts; no extra login |
 | **Cloud context compaction** | Continue long conversations using Codex compaction while retaining DSH history and native compaction | Beta / experimental; off by default, enabled requests use SSE |
 | **WebSocket transport** | Experimentally reuse connections and context transfers, with SSE fallback on connection failure | Beta / experimental; SSE by default, no speed guarantee |
-| **Support diagnostics** | Inspect prerequisites, recent operation results and host registrations without credentials or conversation text | Settings → Maintenance; unverified does not mean broken |
+| **Support diagnostics** | Inspect prerequisites and recent operation results, and copy a short report (about 1 KB; the full one can be copied separately) without credentials or conversation text | Settings → Maintenance; unverified does not mean broken |
 | **Forecast cache cleanup** | Inspect and clear local quota forecast history | Keeps sign-in, drafts, originals, and shared dependencies |
 
 [Usage details, limitations, and optional components →](docs/GUIDE.en.md)
@@ -104,7 +119,7 @@ This plugin supports the latest DeepSeek Harness release recorded in its package
 2. Paste this package name, with its version, into the **Package name or address** field:
 
    ```text
-   dsh-codex-subscription@2.2.11
+   dsh-codex-subscription@2.4.0
    ```
 
 3. Click **Install** and wait for completion. Follow the page instructions; save your work before restarting if requested.
@@ -139,7 +154,7 @@ dsh --profile headless "Reply with only the word: ok"
 
 ### Everyday controls, right in the composer
 
-Choose a model, adjust reasoning, enable Fast mode, and check remaining quota without opening settings for every request.
+Choose a model, adjust reasoning, switch speed tiers, and check remaining quota without opening settings for every request.
 
 <p align="center"><img src="docs/assets/composer-quota-en.png" width="800" alt="DSH composer with GPT-6-Luna, reasoning controls and remaining quota"></p>
 
@@ -174,59 +189,42 @@ Image generation, editing, and the sketch canvas are **Beta**. The canvas and ag
 Draw by hand with the canvas button, or ask the agent with `@sketch` and your drawing request.
 
 <table>
-<tr><th width="50%">Original sketch</th><th width="50%">Actual plugin output</th></tr>
-<tr><td><img src="docs/assets/sketch-demo-source.png" alt="Mountains and cabin sketch" width="480"></td><td><img src="docs/assets/sketch-demo-result.png" alt="Watercolor mountain cabin generated from the sketch" width="480"></td></tr>
+<tr><th width="50%">Native sketch drawn by the agent</th><th width="50%">Image generated with the sketch as the only reference</th></tr>
+<tr><td><img src="docs/assets/sketch-lighthouse-sketch.webp" alt="Lighthouse sketch drawn by GPT-6.1-Sol" width="480"></td><td><img src="docs/assets/sketch-lighthouse-result.webp" alt="Watercolor lighthouse poster generated from the sketch" width="480"></td></tr>
 </table>
 
-The request keeps the mountain and cabin composition, turning it into a warm watercolor illustration with green peaks, an orange roof, a meadow stream, and soft morning light, without the blue outlines.
+**GPT-6.1-Sol · Xhigh** on the official desktop app 0.2.0-rc.2: one `@sketch` message produced 6 layers and 229 editable native strokes, then called GPT Image 2 (quality: high) once with the sketch as the only reference. About 13 minutes end to end.
+
+The prompt (sent in Chinese): draw "Lighthouse and Homecoming" on a 4:3 canvas — a red-and-white lighthouse on a rocky cape casting a fan of light, an orange-sailed fishing boat heading home with a V-shaped wake, a half-risen sun, three cloud bands, three gulls and surf around the rocks, in six named layers; then turn the sketch into a warm watercolor travel poster with no text, keeping the composition.
 
 <details>
-<summary>Advanced examples · Layered illustration and Mona Lisa</summary>
+<summary>Advanced examples · Mechanical cutaway and blue-green landscape (Sol Xhigh vs Astra Low)</summary>
 
-**Someone Behind the Canvas**
+Both subjects were chosen to expose quality differences: a mechanical cutaway tests structure and detail (meshing gears, pendulum, chains, tiny figures on a staircase), and a Song-dynasty blue-green landscape tests composition and taste (the "three distances", empty space, texture strokes and mineral color). The same prompt went to **GPT-6.1-Sol · Xhigh** and **GPT-6-Astra · Low**, one run each, on the official desktop app, with GPT Image 2 (quality: high).
 
-**Astra draws the sketch; GPT Image 2 generates the illustration.** The editable original has 427 strokes across six layers. The image request retains its composition and adds paper-art and hand-painted detail. The requested model is `gpt-image-2`; the server does not report the executing model.
+**Clock Tower Cutaway**
 
 <table>
-<tr><th width="50%">Native sketch</th><th width="50%">Generated result</th></tr>
-<tr><td><img src="docs/assets/sketch-advanced-source.png" alt="Sketch" width="480"></td><td><img src="docs/assets/sketch-advanced-result.png" alt="Result" width="480"></td></tr>
+<tr><th></th><th width="45%">Native sketch</th><th width="45%">Generated image</th></tr>
+<tr><td><b>Sol<br>Xhigh</b></td><td><img src="docs/assets/sketch-clocktower-sol-sketch.webp" alt="Clock tower sketch by Sol" width="400"></td><td><img src="docs/assets/sketch-clocktower-sol-result.webp" alt="Clock tower illustration from Sol's sketch" width="400"></td></tr>
+<tr><td><b>Astra<br>Low</b></td><td><img src="docs/assets/sketch-clocktower-astra-sketch.webp" alt="Clock tower sketch by Astra" width="400"></td><td><img src="docs/assets/sketch-clocktower-astra-result.webp" alt="Clock tower illustration from Astra's sketch" width="400"></td></tr>
 </table>
 
-<details>
-<summary>Show the reproduction prompt and original image request</summary>
-
-**Sketch reproduction prompt (reconstructed from the artwork, not the original conversation)**
-
-Astra drew the original in stages. This Chinese prompt provides a starting point for the same concept, not a guarantee of an identical result.
-
-```text
-@sketch 用 4:3 横版画板绘制《画布背面有人》：中央偏上是一处撕开的纸洞，洞内是深蓝星空和一位拿颜料桶的小画师；蓝色颜料从桶中流出，形成 S 形河流，流向下方城市。左侧城市保持未上色线稿，右侧城市被暖色点亮，加入纸船与飞鸟。按纸面、洞内世界、颜料河流、城市、画师和细节分层绘制，保留原生可编辑笔画。
-```
+Sol first computed tooth counts (48, 38, 27, 46, 36, 24, 34, 30, 22) and pitch circles for the nine gears with a script, then placed them in mesh; drawing took about an hour. Its first image request failed; it asked, was allowed one retry, and succeeded. Astra Low finished sketch and image in about 5 minutes, with a simpler structure and gears that are not individually calibrated.
 
-**Actual image-generation prompt (original Chinese)**
-
-```text
-请基于本条附加草图实际调用订阅图片工具一次，生成成品插画。quality=low，模型使用当前默认，不切换型号，不额外生成。主题《画布背面有人》：保留4:3横构图、中央偏上的撕纸洞口、洞内拿颜料桶的小画师、流出成为S形河流的蓝色颜料、下方左侧未上色城市与右侧被点亮城市、纸船飞鸟。精修为惊艳的立体纸艺与精细手绘结合的编辑插画，纸张纤维、真实撕边及柔和投影，深靛蓝洞内星月，丰富青蓝颜料层次和流动质感，赭橙画师与暖色建筑，微小清晰的叙事细节。不重构为风景，不添加文字水印。必须使用本条参考图片编辑，不能仅凭文字生成。生成后简短说明完成即可。
-```
-
-</details>
-
-Original example released in: [Beta v2.1.0-beta.2](https://github.com/WSL043/dsh-codex-subscription/releases/tag/v2.1.0-beta.2)
-
-**Mona Lisa: Astra sketch → GPT image generation**
-
-Example version: [2.1.0-beta.5](https://github.com/WSL043/dsh-codex-subscription/releases/tag/v2.1.0-beta.5)
-
-Actual results supplied by the user from another computer: Astra draws on a portrait canvas, then GPT image generation turns the sketch into an oil painting.
+**A Thousand Green Peaks**
 
 <table>
-<tr><th width="50%">Native Astra sketch</th><th width="50%">GPT-generated oil painting</th></tr>
-<tr><td><img src="docs/assets/sketch-mona-lisa-source.png" alt="Mona Lisa sketch drawn by Astra" width="480"></td><td><img src="docs/assets/sketch-mona-lisa-result.png" alt="Mona Lisa oil painting generated from the sketch" width="480"></td></tr>
+<tr><th></th><th width="45%">Native sketch</th><th width="45%">Generated image</th></tr>
+<tr><td><b>Sol<br>Xhigh</b></td><td><img src="docs/assets/sketch-shanshui-sol-sketch.webp" alt="Landscape sketch by Sol" width="400"></td><td><img src="docs/assets/sketch-shanshui-sol-result.webp" alt="Blue-green landscape from Sol's sketch" width="400"></td></tr>
+<tr><td><b>Astra<br>Low</b></td><td><img src="docs/assets/sketch-shanshui-astra-sketch.webp" alt="Landscape sketch by Astra" width="400"></td><td><img src="docs/assets/sketch-shanshui-astra-result.webp" alt="Blue-green landscape from Astra's sketch" width="400"></td></tr>
 </table>
 
-Original sketch prompt: `@sketch 用竖版画板画一幅《蒙娜丽莎》` (Draw the Mona Lisa on a portrait canvas.)
+Sol Xhigh took about 28 minutes and Astra Low about 5. Both kept the main peak, waterfall, plank path and river in place; Sol's receding ranges and mist bands are more complete.
+
+The full Chinese prompts for both subjects are in the [Chinese README](README.md#从草图到作品).
 
-Original image prompt: `帮我变成油画` (Turn it into an oil painting.)
+Each example is a single run, not a measure of consistent quality. Sol Xhigh is much slower; Astra or a lower reasoning level is enough for everyday sketches.
 
 </details>
 

package/README.md

@@ -16,12 +16,27 @@
 [![MIT](https://img.shields.io/badge/license-MIT-111111.svg)](LICENSE)
 [![Star](https://img.shields.io/github/stars/WSL043/dsh-codex-subscription?style=flat&logo=github&label=Star)](https://github.com/WSL043/dsh-codex-subscription/stargazers)
 
-[功能](#功能一览) · [安装](#安装) · [日常使用](#界面一览) · [作品案例](#从草图到作品) · [使用指南](docs/GUIDE.zh-CN.md) · [更新与卸载](#更新与卸载)
+[更新亮点](#240-更新亮点) · [功能](#功能一览) · [安装](#安装) · [日常使用](#界面一览) · [作品案例](#从草图到作品) · [使用指南](docs/GUIDE.zh-CN.md) · [更新与卸载](#更新与卸载)
 
 </div>
 
 <p align="center"><img src="docs/assets/product-account-demo.png" width="1100" alt="Codex 订阅账号与额度界面：模型选择、双周期额度与提醒；使用演示数据"></p>
 
+## 2.4.0 更新亮点
+
+- **Agent 绘图更强（Beta）**：新增 SVG 绘图命令，模型可用最擅长的 SVG 一次画完一整层，导入后仍是可编辑、可撤销、可导出的原生笔画；渐变、越界等近似处理会明确回报，不支持的元素会点名拒绝。
+- **全新作品案例**：在官方桌面版上用 GPT-6.1-Sol · Xhigh 与 GPT-6-Astra · Low 实测，机械剖面与青绿山水两道题并排对照，见[从草图到作品](#从草图到作品)。
+- **桌面版修复**：官方桌面版上 Codex 子任务组件现在能正确识别宿主版本并安装；支持诊断会报告真实的 DSH 版本；图片请求网络失败时会说明是超时还是连接中断。
+
+### 2.3.0
+
+- **官方桌面端可用**：DeepSeek Harness 官方桌面版（0.2.0-rc.2）已完成实机验收——安装插件、登录、模型列表、隐藏模型、带工具调用的对话、图片生成均通过；安装方式见[安装](#安装)。
+- **模型只来自你的账号**：模型列表完全读取账号目录，登录前或读取失败时不显示任何模型，也没有内置兜底；“模型与运行”里可逐个显示 / 隐藏输入框中的模型。
+- **设置页重新整理**：模型与搜索分成两个卡片，开关统一使用 DSH 自带控件，未登录时不再出现无效按钮。
+- **极速档 Ultrafast（Beta）**：OpenAI 发布会新增的速度档。只有账号模型目录明确给出该档位（目前为 Pro 500 的 GPT-6 Astra）才会在模型菜单里出现，其他账号不显示，也不会被悄悄降级到别的档位；由于暂无 Pro 500 账号做实机验证，标为 Beta。
+- **更短的支持诊断**：“复制诊断”现在只复制结论、近期失败和各能力状态（约 1 KB），完整报告仍可在详情里单独复制。
+- **子任务与稳定性**：Codex 子任务在 0.2.0 上真正生效，并且不再暴露账号里连接的应用；修复了图片生成在桌面版上被宿主拒绝的问题（2.2.11）。
+
 ## 功能一览
 
 把订阅接入、模型控制、额度管理和图片创作放在同一个 DSH 工作流里。下面逐项列出插件提供的能力；模型权限和额度以账号实际返回为准。
@@ -34,7 +49,7 @@
 | **模型目录同步** | 自动读取账号可用模型与推理档位，支持手动刷新；目录标注退役的模型会在模型名称旁显示退役日期，不会自动切换你的选择 | 新型号以账号目录开放情况为准；模型只来自账号目录，未登录或读取失败时不显示任何模型，也不使用内置列表 |
 | **模型列表** | 在“模型与运行”中逐个显示或隐藏输入框中的模型 | 仅影响列表显示；已选模型和请求不受影响 |
 | **推理等级** | 在输入框选择模型支持的推理档位 | 档位随模型变化 |
-| **高速模式** | 在支持的模型上切换标准 / Fast，输入框显示闪电标识 | 默认标准；Fast 会增加用量消耗 |
+| **速度档位** | 在支持的模型上切换标准 / Fast；账号目录提供时还有 Ultrafast（Beta），输入框显示闪电标识 | 默认标准；Fast 和 Ultrafast 会增加用量消耗，Ultrafast 目前仅 Pro 500 |
 | **输出详略** | 选择简短、适中或详细的回答偏好 | 默认跟随模型 |
 | **输入图片精度** | 为发送给聊天模型的图片选择低、高、原图或自动精度 | 影响输入 token 用量；默认自动 |
 | **订阅联网搜索** | 使用 Codex 搜索，或选择 DSH 搜索 | 可选自动、DSH、Codex 路由 |
@@ -70,7 +85,7 @@
 | **笔画平滑与导航** | 抬笔后平滑整条路径；拖动、缩放、撤销重做 | 快捷键可自定义和关闭 |
 | **多份草稿** | 保存并重新编辑最多 20 份本地草稿 | 保存在当前浏览器 |
 | **导入与导出** | 导出 PNG、分层 PSD，或交换保留原生笔画的草稿文件 | PSD 导入仅支持有限的普通像素图层 |
-| **Agent 绘图** | 用 `@sketch` 让 Agent 绘制，查看进度并随时停止 | Beta；默认关闭，需同时开启画板 |
+| **Agent 绘图** | 用 `@sketch` 让 Agent 绘制，查看进度并随时停止；Agent 可用 SVG 整段作画，导入为可编辑原生笔画 | Beta；默认关闭，需同时开启画板 |
 | **绘图结果回看** | 完成后向模型返回画布预览，帮助后续检查 | Beta；默认关闭，会增加图片输入用量 |
 
 ### 进阶能力与维护
@@ -81,7 +96,7 @@
 | **子任务组件管理** | 在设置里安装、停用或卸载官方 Codex 子任务组件 | 由支持的 DSH 宿主管理，无需额外登录 |
 | **云端上下文压缩** | 使用 Codex 云端压缩继续长对话，同时保留 DSH 历史与原生压缩 | Beta / 实验性；默认关闭，开启的请求使用 SSE |
 | **WebSocket 连接** | 实验性复用连接和上下文传输，连接失败可回退 SSE | Beta / 实验性；默认 SSE，不保证更快 |
-| **支持诊断** | 查看功能前置条件、近期操作结果与宿主注册信息；报告不含凭据和会话正文 | 设置 → 维护；未验证不代表故障 |
+| **支持诊断** | 查看功能前置条件和近期操作结果，一键复制约 1 KB 的精简报告（完整报告可另行复制）；不含凭据和会话正文 | 设置 → 维护；未验证不代表故障 |
 | **预测缓存清理** | 查看并清理本地额度预测历史 | 保留登录、草稿、原图和共享依赖 |
 
 [详细用法、限制与可选组件 →](docs/GUIDE.zh-CN.md)
@@ -104,7 +119,7 @@
 2. 在 **包名或地址** 输入框中粘贴下面的包名（已带版本号）：
 
    ```text
-   dsh-codex-subscription@2.2.11
+   dsh-codex-subscription@2.4.0
    ```
 
 3. 点击 **安装**，等待安装完成；按页面提示操作，需要重启时先保存工作。
@@ -139,7 +154,7 @@ dsh --profile headless "只回复：ok"
 
 ### 日常操作，留在输入框
 
-选模型、调推理档位、开启高速模式，同时查看剩余额度。无需为每次请求打开设置。
+选模型、调推理档位、切换速度档位，同时查看剩余额度。无需为每次请求打开设置。
 
 <p align="center"><img src="docs/assets/composer-astra-fast.png" width="800" alt="DSH 输入框：GPT-6-Astra Max、闪电标识的 Fast 高速模式与剩余额度"></p>
 
@@ -174,59 +189,55 @@ dsh --profile headless "只回复：ok"
 手动画图时点击画板按钮；让 Agent 绘图时发送 `@sketch` 加绘图要求。
 
 <table>
-<tr><th width="50%">画板原草图</th><th width="50%">插件实际生成结果</th></tr>
-<tr><td><img src="docs/assets/sketch-demo-source.png" alt="山峰与小屋草图" width="480"></td><td><img src="docs/assets/sketch-demo-result.png" alt="根据草图生成的水彩山间小屋" width="480"></td></tr>
+<tr><th width="50%">Agent 绘制的原生草图</th><th width="50%">以草图为唯一参考的生图结果</th></tr>
+<tr><td><img src="docs/assets/sketch-lighthouse-sketch.webp" alt="GPT-6.1-Sol 绘制的灯塔与归航草图" width="480"></td><td><img src="docs/assets/sketch-lighthouse-result.webp" alt="根据草图生成的灯塔水彩海报" width="480"></td></tr>
 </table>
 
-示例要求：保留山峰与小屋的构图，生成温暖的水彩旅行插画，青绿山峰、橙色屋顶、草地小溪与柔和晨光，不保留蓝色线条。
+**GPT-6.1-Sol · Xhigh**，官方桌面版 0.2.0-rc.2 实测：一条 `@sketch` 消息完成 6 个图层、229 笔可编辑原生笔画，再以这张草图为唯一参考调用一次 GPT Image 2（质量：高），全程约 13 分钟。
+
+```text
+@sketch 用 4:3 横版画板绘制《灯塔与归航》：左侧礁石岬角上立一座红白条纹灯塔，塔顶灯室射出扇形光束；右侧海面上一艘挂橙色小帆的渔船正驶回港湾，船后拖出 V 形水痕；地平线上是半个初升的太阳，天空有三条横向云带和三只海鸥，前景礁石间有一圈白色浪花。按天空、太阳与云、海面、礁石与灯塔、渔船、海鸥与浪花分层绘制，保留原生可编辑笔画。画完后，把这张草图作为唯一参考图，调用订阅图片工具一次生成成品：保持构图不变，做成温暖的水彩旅行海报，晨光金橙配靛青海面，有纸纹和柔和晕染，不添加文字。
+```
 
 <details>
-<summary>进阶展示 · 分层插画与蒙娜丽莎</summary>
+<summary>进阶展示 · 机械剖面与青绿山水（Sol Xhigh 与 Astra Low 对照）</summary>
 
-**《画布背面有人》**
+两道题都选了容易暴露差距的题材：机械剖面考结构与细节（齿轮咬合、摆锤、链条、楼梯上的人物），青绿山水考构图与审美（三远法、留白、皴法与设色）。同一条提示词分别交给 **GPT-6.1-Sol · Xhigh** 和 **GPT-6-Astra · Low**，各跑一次，均在官方桌面版上完成，生图统一为 GPT Image 2（质量：高）。
 
-**Astra 绘制草图，GPT Image 2 生成成图。** 原生草图共 6 层、427 笔，可继续编辑；生图请求保留构图，精修纸艺与手绘质感。请求型号为 `gpt-image-2`，服务端未报告实际执行型号。
+**《钟塔剖面图》**
 
 <table>
-<tr><th width="50%">原生草图</th><th width="50%">实际生成结果</th></tr>
-<tr><td><img src="docs/assets/sketch-advanced-source.png" alt="Sketch" width="480"></td><td><img src="docs/assets/sketch-advanced-result.png" alt="Result" width="480"></td></tr>
+<tr><th></th><th width="45%">原生草图</th><th width="45%">生图结果</th></tr>
+<tr><td><b>Sol<br>Xhigh</b></td><td><img src="docs/assets/sketch-clocktower-sol-sketch.webp" alt="Sol 绘制的钟塔剖面草图" width="400"></td><td><img src="docs/assets/sketch-clocktower-sol-result.webp" alt="Sol 草图生成的钟塔剖面插画" width="400"></td></tr>
+<tr><td><b>Astra<br>Low</b></td><td><img src="docs/assets/sketch-clocktower-astra-sketch.webp" alt="Astra 绘制的钟塔剖面草图" width="400"></td><td><img src="docs/assets/sketch-clocktower-astra-result.webp" alt="Astra 草图生成的钟塔剖面插画" width="400"></td></tr>
 </table>
 
-<details>
-<summary>展开复现提示词与原始生图请求</summary>
+Sol 先用脚本算出 9 个齿轮的齿数（48、38、27、46、36、24、34、30、22）与节圆，再按咬合位置排布，绘图约 1 小时；第一次生图请求失败，按它的询问同意重试一次后成功。Astra Low 约 5 分钟完成全部草图与生图，结构更简化，齿轮没有逐一校准。
+
+**《千峰翠色》**
+
+<table>
+<tr><th></th><th width="45%">原生草图</th><th width="45%">生图结果</th></tr>
+<tr><td><b>Sol<br>Xhigh</b></td><td><img src="docs/assets/sketch-shanshui-sol-sketch.webp" alt="Sol 绘制的青绿山水草图" width="400"></td><td><img src="docs/assets/sketch-shanshui-sol-result.webp" alt="Sol 草图生成的青绿山水" width="400"></td></tr>
+<tr><td><b>Astra<br>Low</b></td><td><img src="docs/assets/sketch-shanshui-astra-sketch.webp" alt="Astra 绘制的青绿山水草图" width="400"></td><td><img src="docs/assets/sketch-shanshui-astra-result.webp" alt="Astra 草图生成的青绿山水" width="400"></td></tr>
+</table>
 
-**草图复现提示词（按原画面整理，非完整原始对话）**
+Sol Xhigh 约 28 分钟，Astra Low 约 5 分钟。两者都保留了主峰、瀑布、栈道与江面的位置；Sol 的远山层次和雾带留白更完整。
 
-原案例由 Astra 分阶段绘制，下面提供同主题的复现起点，不保证得到完全相同的画面。
+<details>
+<summary>展开两道题的完整提示词</summary>
 
 ```text
-@sketch 用 4:3 横版画板绘制《画布背面有人》：中央偏上是一处撕开的纸洞，洞内是深蓝星空和一位拿颜料桶的小画师；蓝色颜料从桶中流出，形成 S 形河流，流向下方城市。左侧城市保持未上色线稿，右侧城市被暖色点亮，加入纸船与飞鸟。按纸面、洞内世界、颜料河流、城市、画师和细节分层绘制，保留原生可编辑笔画。
+@sketch 用 3:4 竖版画板绘制《钟塔剖面图》：蒸汽朋克风格的钟塔纵向剖面技术插图。最上方是圆形大钟面（十二个刻度、两根指针）和一口铜钟；钟面下一层是齿轮机房：至少 9 个大小不同、齿数正确、彼此咬合的齿轮（相邻齿轮转向相反，齿廓要画出锯齿），另有一个锚形擒纵机构；中段是一根贯穿的长摆杆，末端为铜制圆盘摆锤，旁边两条挂着配重铁块的链条；塔壁一侧是沿墙盘旋而上的螺旋楼梯，梯上有三个微小的工匠人物；底部是石砌地基与拱门。剖面用厚实的暗色砖石墙体框出，内部露出黄铜与铁质结构，加入铆钉、管道，以及从小窗透进来的光柱。按背景砖墙、结构框架、齿轮组、摆锤与链条、楼梯与人物、光影与细节分层绘制。画完后，把这张草图作为唯一参考图，调用订阅图片工具一次生成成品：保持每个部件的位置、比例与咬合关系不变，做成精密的蒸汽朋克剖面图插画——黄铜与旧钢的金属质感、细致的机械倒影和铆钉、羊皮纸底色配墨线渲染、暖色灯光，不添加文字。
 ```
 
-**实际生图提示词**
-
 ```text
-请基于本条附加草图实际调用订阅图片工具一次，生成成品插画。quality=low，模型使用当前默认，不切换型号，不额外生成。主题《画布背面有人》：保留4:3横构图、中央偏上的撕纸洞口、洞内拿颜料桶的小画师、流出成为S形河流的蓝色颜料、下方左侧未上色城市与右侧被点亮城市、纸船飞鸟。精修为惊艳的立体纸艺与精细手绘结合的编辑插画，纸张纤维、真实撕边及柔和投影，深靛蓝洞内星月，丰富青蓝颜料层次和流动质感，赭橙画师与暖色建筑，微小清晰的叙事细节。不重构为风景，不添加文字水印。必须使用本条参考图片编辑，不能仅凭文字生成。生成后简短说明完成即可。
+@sketch 用 3:4 竖版画板绘制《千峰翠色》：宋代青绿山水的构图，严格运用「三远法」。高远：右侧一座主峰拔地而起，占画面上半部，山体用层叠的块面表现，山腰有横向的云雾留白；深远：主峰后方还有两重淡青色远山，越远越淡，之间隔着大片留白的雾带；平远：下方是舒展的江面，一叶渔舟和两只飞鸟。左侧一挂瀑布从中段峭壁倾泻而下，落入山涧，涧边有三株松树，松针成簇；山腰有一座小亭和一条曲折栈道，栈道上有两个极小的行者；前景是坡石与芦苇，用勾勒加披麻皴式的笔触表现。色彩层次为石青、石绿、赭石，远山用淡青。按远山与天空、雾带留白、主峰、瀑布与松树、亭台栈道与行者、江面舟鸟与前景分层绘制。画完后，把这张草图作为唯一参考图，调用订阅图片工具一次生成成品：保持山体、瀑布、栈道、江面的位置和三远法的层次不变，做成宋代青绿山水绢本设色——石青石绿矿物颜料的沉着厚重、细腻的皴擦与勾勒、绢面纹理与微微泛黄的古意、云雾留白，不添加文字或印章。
 ```
 
 </details>
 
-原案例首发：[Beta v2.1.0-beta.2](https://github.com/WSL043/dsh-codex-subscription/releases/tag/v2.1.0-beta.2)
-
-**《蒙娜丽莎》：Astra 草图 → GPT 生图**
-
-案例版本：[2.1.0-beta.5](https://github.com/WSL043/dsh-codex-subscription/releases/tag/v2.1.0-beta.5)
-
-用户在另一台电脑上的实际效果：先让 Astra 在竖版画板上绘制，再通过 GPT 生图转成油画。
-
-<table>
-<tr><th width="50%">Astra 原生草图</th><th width="50%">GPT 生图：油画效果</th></tr>
-<tr><td><img src="docs/assets/sketch-mona-lisa-source.png" alt="Astra 绘制的蒙娜丽莎草图" width="480"></td><td><img src="docs/assets/sketch-mona-lisa-result.png" alt="草图转换后的蒙娜丽莎油画" width="480"></td></tr>
-</table>
-
-草图提示词：`@sketch 用竖版画板画一幅《蒙娜丽莎》`
-
-生图提示词：`帮我变成油画`
+每个例子只跑一次，不代表稳定水平；Sol Xhigh 耗时明显更长，日常草图用 Astra 或较低推理档即可。
 
 </details>