@mojieai/dsh-mojie-model-provider 0.0.0-stage → 0.2.0

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 MoJIeAIGC
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -1,3 +1,328 @@
-# Temporary Holding Version
+<p align="center">
+  <a href="https://mojieaigc.com"><img src="docs/assets/mojie-logo.png" alt="Mojie" width="112" /></a>
+</p>
+
+<h1 align="center">Mojie dsh-mojie-model-provider</h1>
+<p align="center">MojieAgentFlow全站模型deepseek-harness插件</p>
+
+<p align="center">
+  <a href="https://github.com/MoJIeAIGC">GitHub</a> ·
+  <a href="https://space.bilibili.com/483532108">B站</a> ·
+  <a href="https://www.douyin.com/user/MS4wLjABAAAA3JbjlCqKd8WXklSYmO64wxWeMBg4efViyTg8Gz6VZdI">抖音</a>
+</p>
+
+由MojieAI开发的deepseek-harness插件，用于将Mojieaigc.com的所有平台能力无缝接入到DSH中，无需多余配置，直接让你的DSH拥有原生多模型，生图生视频能力。
+
+| 用一个 Key，聊全球主流模型 | 一句话，让 AI 出图 | 像原生模型一样用 |
+| :-- | :-- | :-- |
+| GPT、Claude、Gemini、Qwen、DeepSeek、豆包、Kimi、Grok、GLM 等主流模型，全部出现在 DSH 的模型下拉里。 | `/image 一只戴帽子的猫` 直接把图画进对话；也可以对 AI 说「帮我画……」，它自己调用画图工具。 | 全部沿用 DSH 原生能力，习惯不用改 |
+
+
+---
+
+## 快速开始
+
+> 🚀 三步：装插件 → 填 Key → 开始使用。
+
+## 安装前的准备
+
+### 1. DSH 已经装好
+
+桌面版或命令行版都行，能打开、能收发消息就可以。
+
+- **桌面版**：在「设置 → 通用」里能看到「当前版本：x.x.x」；
+- **命令行版**：在终端里执行：
+
+```powershell
+dsh --version
+```
+
+打印出 `0.2.1-rc.2` 或更高版本就可以继续。
+
+### 1. 创建APIKEY
+1. 登录 [mojieaigc.com](https://www.mojieaigc.com)；
+2. 点击右上角**个人头像**；
+3. 进入 **我的APIKEY**；
+4. 点击**新建密钥**，给 Key 起个名字（例如 `codex`），复制生成的那串。
+
+![头像菜单里的「我的APIKEY」入口](https://mojieagentflow.tos-cn-guangzhou.volces.com/uploads/apidoc/apikey1.png)
+
+进入后是 APIKEY 管理页，点右上角 **新建密钥**：
+
+![APIKEY 管理页，右上角新建密钥](https://mojieagentflow.tos-cn-guangzhou.volces.com/uploads/apidoc/apikey2.png)
+
+![新建 API Key 弹窗](https://mojieagentflow.tos-cn-guangzhou.volces.com/uploads/apidoc/create_apikey.png)
+
+创建后回到列表，点 Key 后面的复制图标把它复制走：
+
+![点复制图标复制 Key](https://mojieagentflow.tos-cn-guangzhou.volces.com/uploads/apidoc/copy_apikey.png)
+
+
+> [!WARNING]
+> APIKEY 是你与 MojieAgentFlow 通信的重要凭证，请妥善保管；如怀疑泄露，请及时删除并重新创建，
+
+### 2. 安装插件
+
+#### 方式 A · 插件市场（推荐）
+
+1. 打开 DSH，进入 **设置**。
+2. 找到 **插件 / 插件市场** 页面。
+3. 搜索 **Mojie**（或包名 `@mojieai/dsh-mojie-model-provider`）。
+4. 点**安装**，等它提示「已安装，请重启 DSH 加载它」，再点**完成**。
+
+
+![image](https://mojieagentflow.tos-cn-guangzhou.volces.com/uploads/apidoc/dsh-1.png)
+
+![image](https://mojieagentflow.tos-cn-guangzhou.volces.com/uploads/apidoc/dsh-2.png)
+
+#### 方式 B · 一行命令
+
+打开终端（PowerShell / Terminal）：
+
+```sh
+dsh plugin --profile desktop add https://github.com/MoJIeAIGC/dsh-mojie-model-provider/releases/latest/download/mojieai-dsh-mojie-model-provider-0.2.0.tgz
+```
+
+> [!NOTE]
+> URL 里的版本号要和发布页上的文件名一致。如果提示 404，请到 [Releases](https://github.com/MoJIeAIGC/dsh-mojie-model-provider/releases) 页面复制最新的 `.tgz` 链接替换。
+
+#### 装完必须重启一次
+
+新插件是在 DSH 启动时加载的，**不重启看不到效果**。
+
+- **Windows 桌面版**：右键托盘图标 → **退出**（只关窗口不算退出），再双击 `DeepSeek Harness` 重新打开。
+- **命令行版**：结束当前的 `dsh` 进程，重新启动。
+
+### 3. 填入API Key（一次就好）
+
+1. 重启后进入 **设置 → 模型**。
+2. 找到 **Mojie** provider 行；卡片会显示「API 密钥」。
+3. 在 **API 密钥** 输入框里粘贴第 1 步复制的 Key。
+4. 点卡片底部的 **保存**，看到「已保存。」即可，以后不用再填。
+
+
+### 4. 正确使用
+
+**对话界面** —— 在 **设置 → 模型** 或聊天界面的模型下拉里，选一个 Mojie 提供的模型（例如 `claude-opus-5.5`、`gpt-6-astra`、`Kimi-k3.1-pro`），然后像平时一样发消息。它和原生模型一样支持：
+
+- **工具调用**（读写文件、跑命令、搜索等）
+- **看图**（把图片拖进输入框，选一个支持图片的模型）
+
+**画图** —— 在输入框里敲：
+
+```text
+/image 一只戴帽子的猫，扁平插画风，白底
+```
+
+几秒到几十秒后，图片会直接出现在对话里，并附上模型名与尺寸。
+
+---
+
+## 画图手册
+
+> 🎨 `/image` 是给人用的画图入口：提示词原样发给画图模型，不经过聊天模型改写，也不靠模型猜测要调哪个工具。
+
+### 常用写法
+
+| 想做什么 | 输入 |
+| :-- | :-- |
+| 画一张（用当前默认模型） | `/image 一只戴帽子的猫` |
+| 编辑 / 融合（附带图片） | 把图片拖进输入框，输入 `/image 把这两张图融合成一张` |
+| 本次换模型 | `/image 一只猫 --model mojie-nanobanana-2` |
+| 指定分辨率与比例 | `/image 赛博朋克风的城市夜景 --resolution 2K --ratio 16:9` |
+| 列出可用的画图模型 | `/image models` |
+| 切换默认模型 | `/image model mojie-nanobanana-2` |
+| 恢复内置默认 | `/image model auto` |
+| 续查没跑完的任务 | `/image task <任务ID>` |
+| 查看用法 | `/image help` |
+
+> 🖼 只要这次 `/image` 带着图片附件，命令就自动走**编辑**：附件按上传顺序作为 `input_image_urls` 提交给上游，不需要 `/image edit` 语法。`--model` / `--resolution` / `--ratio` 依然可叠加；`models`、`model`、`task`、`help` 这些子命令照常可用，附带的图片会被静默忽略（不会报错，也不影响你暂存的原图）。
+
+参数可以简写：`--model`/`-m`、`--resolution`/`-r`、`--ratio`/`--aspect-ratio`/`-a`，也支持 `--model=mojie-nanobanana-2` 这种等号写法。
+
+### 任务没跑完怎么办
+
+图片生成是异步任务。`/image` 提交后**立即返回**（输入框的附件和发送记录会马上清空），图片在后台完成后自动放进对话；如果提交后进程退出、或想手动取回，用 `/image task <任务ID>` 续查即可，**图不会丢**。
+
+- 原图会写进当前会话的工作目录（文件名带任务 ID），重启 DSH 后对话里仍能看到图。
+- 想要多张，就多敲几次 `/image`，每张都会单独落一个文件、单独发进对话。
+
+### 设置默认画图模型
+
+1. **设置 → 模型 → Mojie** 行，展开 **自定义设置**。
+2. 找到 **图片生成默认模型** 下拉框；列表为空就点右侧 **刷新图片模型**。
+3. 选一个（例如 `seedream-5-image`、`mojie-nanobanana-2`），点卡片底部的 **保存**。
+
+之后 `/image …` 和 AI 自己画图都会用这个模型。也可以随时用 `/image model <模型名>` 切换——两处改的是同一个设置。
+
+> [!IMPORTANT]
+> 对话里 AI 调用画图工具时，**只有你明确点名某个模型**它才会传 `model` 参数；没点名就省略，使用这里设置的默认模型。
+>
+>
+> 模型名可以模糊说（例如“用 image 2.5”）：AI 会先用 `mojie_list_image_models` 把模糊名解析成唯一的确切 id；如果有多个模型都匹配，它会先问你要哪一个。
+
+
+### 让 AI 自己画 / 改图
+
+插件注册了 4 个原生工具，AI 会在需要时自己调用：
+
+| 工具 | 用途 |
+| :-- | :-- |
+| `mojie_generate_image` | 文字生图，可指定模型、分辨率、比例 |
+| `mojie_edit_image` | 改图 / 重绘 / 多图融合；所有参考图都放 `input_images` 数组（一张也传数组），按顺序提交 |
+| `mojie_list_image_models` | 列出当前 Key 可用的画图模型 |
+| `mojie_image_task` | 查询任务状态，完成后把图取回对话 |
+
+- **画**：直接说「帮我画一张……」即可，AI 会把图画进对话。
+- **改**：把图片拖进输入框，说「把这张图的背景换成沙滩」「改成蓝色电影色调」，AI 会把结果图发回来。
+- **多图融合**：一句话里给多张参考图（对话里贴多张，或让 AI 用 `input_images` 传多个路径），按你给的顺序提交。
+- **图在哪**：生成/编辑的原图按厂商字节直接写进**当前会话的工作目录**（`mojie-<任务ID>-<序号>.<扩展名>`，也会保存url在历史记录中，不覆盖、不压缩、Provider 返回几张就存几张）；对话里显示的压缩副本（由 DSH 附件存储按默认规则规范化，仅用于看），原图路径在「查看详情」里。
+
+
+---
+
+## 模型清单
+
+<details>
+<summary><b>OpenAI · GPT</b>（11）</summary>
+
+`gpt-6.1-sol`、`gpt-6-sol`、`gpt-6-luna`、`gpt-6-astra`、`gpt-5.6-terra-pro`、`gpt-5.6-terra`、`gpt-5.6-sol`、`gpt-5.6-luna-pro`、`gpt-5.6-luna`、`gpt-5.5`、`gpt-5.4`
+
+</details>
+
+<details>
+<summary><b>Anthropic · Claude</b>（14）</summary>
+
+`claude-opus-5.5`、`claude-sonnet-5.5`、`claude-haiku-5.5`、`claude-opus-5`、`claude-sonnet-5`、`claude-fable-5.1`、`claude-fable-5`、`claude-opus-4.8`、`claude-opus-4.7`、`claude-sonnet-4.6`、`claude-opus-4.6`、`claude-sonnet-4.5`、`claude-opus-4.5`、`claude-haiku-4.5`
+
+</details>
+
+<details>
+<summary><b>Google · Gemini</b>（9）</summary>
+
+`gemini-4-argon`、`gemini-3.8-flash`、`gemini-3.7-flash`、`gemini-3.6-flash`、`gemini-3.5-flash`、`gemini-3.5-flash-lite`、`gemini-3.1-pro-preview`、`gemini-3.1-flash-lite`、`gemini-3-flash-preview`
+
+</details>
+
+<details>
+<summary><b>阿里 · Qwen</b>（8）</summary>
+
+`qwen3.8-max`、`qwen3.8-omni-flash`、`qwen3.8-flash`、`qwen3.7-max`、`qwen3.7-plus`、`qwen3.7-flash`、`qwen3.6-plus`、`qwen3.6-flash`
+
+</details>
+
+<details>
+<summary><b>字节 · 豆包</b>（7）</summary>
+
+`doubao-seed-2-1-pro`、`doubao-seed-2-1-turbo`、`doubao-seed-2-1-lite`、`doubao-seed-2-0-lite`、`doubao-seed-2-0-mini`、`doubao-seed-2-0-code-preview`、`doubao-seed-character`
+
+</details>
+
+<details>
+<summary><b>智谱 · GLM</b>（6）</summary>
+
+`GLM-5.3`、`GLM-5.3-FlashX`、`GLM-5.3-Flash`、`GLM-5.2`、`GLM-5.1`、`GLM-5`
+
+</details>
+
+<details>
+<summary><b>Moonshot · Kimi</b>（4）</summary>
+
+`kimi-k3.1-pro`、`kimi-k3.1-pro-highspeed`、`kimi-k3`、`kimi-k2.7-code`
+
+</details>
+
+<details>
+<summary><b>xAI · Grok</b>（4）</summary>
+
+`grok-4.7`、`grok-4.6`、`grok-4.5`、`grok-4.3`
+
+</details>
+
+<details>
+<summary><b>DeepSeek</b>（2）</summary>
+
+`deepseek-v4-pro`、`deepseek-v4-flash`
+
+</details>
+
+> [!NOTE]
+> 清单会随主站更新。每个模型的 **上下文窗口、最大输出 token 数、输入类型、推理等级** 会一并自动填写。点 **获取可用模型** 可以搜索、勾选并 **更新列表**；
+---
+
+
+## 常见问题
+
+<details>
+<summary><b>设置里模型目录是空的？</b></summary>
+
+模型目录默认由内置快照自动填写，即使还没联网也能看到全部模型。要拿到主站最新清单，先确认 API Key 已保存（卡片上会显示「已配置凭据」），再展开 **自定义设置** 点 **获取可用模型**，勾选后点 **更新列表**；画图模型则点 **刷新图片模型**。如果仍然为空，检查主站余额是否用完，或 API 地址是否被改错。
+
+</details>
+
+<details>
+<summary><b>画图报 AUTH / 401？</b></summary>
+
+Key 复制错了或已失效。回主站「个人中心 → API-KEY」重新复制一次，粘进卡片保存即可。注意别把首尾空格一起粘进去。
+
+</details>
+
+<details>
+<summary><b>提示 RATE_LIMIT / 上游繁忙？</b></summary>
+
+主站临时限流，等几十秒再试。画图失败通常不会产生可用的创作结果，可以直接重发。
+
+</details>
+
+<details>
+<summary><b>/image 命令打不出来？</b></summary>
+
+按顺序自查：
+
+1. 插件已安装，并且**重启过** DSH；
+2. 设置 → 模型里能看到 **Mojie** 这一行（行内是 API 密钥 + 自定义设置）；
+3. 在输入框敲 `/`，菜单里能看到 **image**。
+
+</details>
+
+<details>
+<summary><b>怎么升级插件？</b></summary>
+
+在插件市场点更新；或先卸载再装新版 tarball，然后重启。。
+
+</details>
+
+<details>
+<summary><b>怎么卸载？</b></summary>
+
+在插件市场点卸载，或执行：
+
+```sh
+dsh plugin --profile desktop remove "@mojieai/dsh-mojie-model-provider"
+```
+
+然后重启 DSH。已生成的原图仍在各会话的工作目录里，不会被删掉。
+
+</details>
+
+<details>
+<summary><b>费用怎么算？</b></summary>
+
+由 Mojie 主站按点数计费（聊天按 token、画图按张 / 分辨率），DSH 这边不额外收费。具体价格见主站。
+余额不足请前往[Mojieaigc.com](https://www.mojieaigc.com) 充值
+
+</details>
+
+---
+
+## 隐私与安全
+
+- **Key 只在本机**：保存在 DSH 的凭据存储里（或 `MOJIE_API_KEY` 环境变量），不写进插件配置，也不进入对话内容。
+- **请求直连主站**：聊天与画图请求直接发往 `mojieaigc.com`，不经过第三方中转。
+- **图片留在本地**：生成的原图（与厂商字节一致）写进当前会话的工作目录，方便之后继续查看或二次编辑；对话显示副本存放在本机附件目录。
+- **别把 Key 贴出去**：不要贴进对话、截图或代码仓库；怀疑泄露时，去主站重置 Key 再改一遍设置。
+
+---
+
+遇到问题或想提需求，欢迎到 [MojieAgentflow](https://www.mojieaigc.com) 查看联系方式。
 
-This version is a temporary placeholder for this package. An operational version to replace this has been submitted for review and is awaiting a staged release.

package/cordis.patch.yml

@@ -0,0 +1,22 @@
+# Bundle layer installed by `dsh plugin add`.
+#
+# This inserts the Mojie provider row under the settings namespace `mojie`.
+# The API key is a credential reference: keep MOJIE_API_KEY in the launching
+# environment, or store it under that reference through the Models page card.
+# A profile's own cordis.patch.yml can override this row by id.
+- insert:
+    - id: mojie
+      name: '@mojieai/dsh-mojie-model-provider'
+      config:
+        apiKeyEnv: MOJIE_API_KEY
+        baseURL: 'https://mojieaigc.com'
+        # Empty means: inherit /v1/models discovery plus the built-in catalog.
+        # The Models page card fills this list for you (model rows, context
+        # windows, input types, and thinking levels); hand-written entries here
+        # still win and pin the values. Thinking levels come from the per-vendor
+        # recommendation unless a row pins `reasoningEfforts`, and
+        # `reasoning: 'off'` declares a model that cannot think at all.
+        models: []
+        # Send `reasoning.effort` for every model that does not opt out, even
+        # though the site publishes no effort vocabulary. Set false to stop.
+        reasoningDefaults: true

package/docs/assets/mojie-logo.svg

@@ -0,0 +1,28 @@
+<svg id="图层_1" data-name="图层 1" xmlns="http://www.w3.org/2000/svg" xmlns:xlink="http://www.w3.org/1999/xlink"
+     viewBox="0 0 97.95 62.39">
+    <defs>
+        <style>
+            .cls-1{fill:url(#未命名的渐变_25);}.cls-2{fill:url(#未命名的渐变_25-2);}.cls-3{fill:url(#未命名的渐变_25-3);}.cls-4{fill:url(#未命名的渐变_25-4);}
+        </style>
+        <linearGradient id="未命名的渐变_25" x1="1.05" y1="33.11" x2="20.59" y2="33.11" gradientUnits="userSpaceOnUse">
+            <stop offset="0" stop-color="#4d6bfe"/>
+            <stop offset="1" stop-color="#a035ff"/>
+        </linearGradient>
+        <linearGradient id="未命名的渐变_25-2" x1="3.41" y1="24.77" x2="46.04" y2="24.77"
+                        xlink:href="#未命名的渐变_25"/>
+        <linearGradient id="未命名的渐变_25-3" x1="69.44" y1="31.51" x2="99" y2="31.51" xlink:href="#未命名的渐变_25"/>
+        <linearGradient id="未命名的渐变_25-4" x1="36.61" y1="38.4" x2="72" y2="38.4" xlink:href="#未命名的渐变_25"/>
+    </defs>
+    <path class="cls-1"
+          d="M10.82,62.68h0a9.77,9.77,0,0,1-9.77-9.77V13.32a9.77,9.77,0,0,1,9.77-9.78h0a9.77,9.77,0,0,1,9.77,9.78V52.91A9.77,9.77,0,0,1,10.82,62.68Z"
+          transform="translate(-1.05 -0.29)"/>
+    <path class="cls-2"
+          d="M3.41,7C7,.89,16.39-1.7,23.22,2a23.17,23.17,0,0,1,8.1,7.32c5.17,8,9.75,16.29,14.48,24.52a2.78,2.78,0,0,1,0,2.31c-1.94,3.46-3.89,6.92-6.09,10.22-2.54,3.83-5.1,3.84-7.58-.07C28.74,41,25.66,35.49,22.47,30.05c-3-5.21-6-10.46-9.14-15.64C11.06,10.62,8.21,7.49,3.41,7Z"
+          transform="translate(-1.05 -0.29)"/>
+    <path class="cls-3"
+          d="M69.44,6.87A16.35,16.35,0,0,1,87.21,1c6.71,1.9,10.28,6.62,11.41,13.33A24.84,24.84,0,0,1,99,18.48c0,11.22.08,22.45,0,33.68,0,7.65-6.55,12.29-13.62,9.86-4.16-1.43-6.48-4.86-6.49-9.84,0-10.53,0-21.07,0-31.6,0-.78,0-1.56,0-2.34C78.69,11.89,75.79,8.32,69.44,6.87Z"
+          transform="translate(-1.05 -0.29)"/>
+    <path class="cls-4"
+          d="M36.61,56s4.53,2.11,12.11-10.51c4.18-7,11.19-19.33,16.63-29,2.1-3.74,6.65-2.25,6.65,2.05V41.44s-9.44,21-20.75,21S36.61,56,36.61,56Z"
+          transform="translate(-1.05 -0.29)"/>
+</svg>

package/lib/adapter.d.ts

@@ -0,0 +1,57 @@
+import { LlmAdapter, type GenerateOptions, type ImageAttachmentAccessResolver, type LlmModelInfo, type LlmResolvedModelInfo, type StreamChunk } from '@deepseek-ai/dsh-llm';
+import type { AttachmentStore, ImageAttachmentRef } from '@deepseek-ai/dsh-attachment';
+import { type ResolvedConfig } from './config.js';
+/**
+ * Host-owned request-time services the adapter can read without depending on the harness repo.
+ */
+export interface MojieAdapterOptions {
+    /** Read the current live configuration for one request operation. */
+    readonly config: () => ResolvedConfig;
+    /** Resolve one credential reference through the host seam. */
+    readonly resolveApiKey: (ref: string) => Promise<string>;
+    readonly resolveAttachments?: () => AttachmentStore | undefined;
+    readonly resolveImageAccess?: ImageAttachmentAccessResolver;
+}
+/**
+ * Mojie Responses adapter.
+ *
+ * Timing model (one clock per failure mode, never a wall-clock cap on a live
+ * stream):
+ * - `firstByteTimeoutMs`: request sent → response headers. Catches a gateway
+ *   that accepts the connection and never answers.
+ * - `streamIdleTimeoutMs`: silence between received bytes. SSE comment
+ *   heartbeats reset it, so a generation that keeps streaming may run for
+ *   hours.
+ * - `requestTimeoutMs`: optional operational backstop (`0` = off), never the
+ *   liveness judge.
+ */
+export declare class MojieAdapter extends LlmAdapter {
+    private readonly dependencies;
+    private catalog;
+    private catalogRefresh;
+    constructor(dependencies: MojieAdapterOptions);
+    providerInfo(provider: string): {
+        id: string;
+        name: string;
+    };
+    listModels(provider: string): Promise<readonly LlmModelInfo[]>;
+    resolveModel(provider: string, model: string): Promise<LlmResolvedModelInfo>;
+    private resolveModelWith;
+    /** Discovery row for one model from the last successful listing, when it still matches the endpoint. */
+    private discoveredModel;
+    private loadCatalog;
+    /**
+     * Refresh the advisory catalog off the request path. Metadata only affects
+     * compaction and image handling, so a failed refresh must not fail a request.
+     */
+    private ensureCatalog;
+    imageRequestTarget(_provider: string, _model: string, image: ImageAttachmentRef): {
+        maxBytes: number;
+        width: number;
+        height: number;
+    };
+    private imageTarget;
+    stream(options: GenerateOptions): AsyncIterable<StreamChunk>;
+    private prepareRequestImages;
+    private projectOffloaded;
+}

package/lib/adapter.js

@@ -0,0 +1,172 @@
+import { contentHasImage, LlmAdapter, LlmError, offloadedImageText, projectOffloadedImages, requiredImageOffload, } from '@deepseek-ai/dsh-llm';
+import { requestImageDimensions } from '@deepseek-ai/dsh-attachment';
+import { fetchModelListing } from './discovery.js';
+import { mergeCatalog, mergeResolvedModelInfo } from './catalog.js';
+import { responsesURL } from './config.js';
+import { serializeRequest } from './serialize.js';
+import { streamResponses } from './core/stream.js';
+/** How long one `/v1/models` listing answers advisory model metadata. */
+const CATALOG_TTL_MS = 10 * 60_000;
+/**
+ * Mojie Responses adapter.
+ *
+ * Timing model (one clock per failure mode, never a wall-clock cap on a live
+ * stream):
+ * - `firstByteTimeoutMs`: request sent → response headers. Catches a gateway
+ *   that accepts the connection and never answers.
+ * - `streamIdleTimeoutMs`: silence between received bytes. SSE comment
+ *   heartbeats reset it, so a generation that keeps streaming may run for
+ *   hours.
+ * - `requestTimeoutMs`: optional operational backstop (`0` = off), never the
+ *   liveness judge.
+ */
+export class MojieAdapter extends LlmAdapter {
+    dependencies;
+    catalog;
+    catalogRefresh;
+    constructor(dependencies) {
+        super();
+        this.dependencies = dependencies;
+    }
+    providerInfo(provider) {
+        return { id: provider, name: 'Mojie' };
+    }
+    async listModels(provider) {
+        const config = this.dependencies.config();
+        const apiKey = await this.dependencies.resolveApiKey(config.apiKeyEnv);
+        const discovered = await this.loadCatalog(config, apiKey, false);
+        return mergeCatalog(provider, config.models, discovered);
+    }
+    async resolveModel(provider, model) {
+        const config = this.dependencies.config();
+        return this.resolveModelWith(provider, model, config, this.discoveredModel(config, model));
+    }
+    resolveModelWith(provider, model, config, discovered) {
+        const configured = config.models.find((entry) => entry.id === model);
+        return mergeResolvedModelInfo(provider, model, configured, discovered, {
+            reasoningDefaults: config.reasoningDefaults,
+        });
+    }
+    /** Discovery row for one model from the last successful listing, when it still matches the endpoint. */
+    discoveredModel(config, model) {
+        const cached = this.catalog;
+        if (cached === undefined || cached.key !== config.baseURL)
+            return undefined;
+        return cached.models.find((entry) => entry.id === model);
+    }
+    async loadCatalog(config, apiKey, force) {
+        const cached = this.catalog;
+        if (!force
+            && cached !== undefined
+            && cached.key === config.baseURL
+            && Date.now() - cached.at < CATALOG_TTL_MS) {
+            return cached.models;
+        }
+        const models = await fetchModelListing(config.baseURL, apiKey);
+        this.catalog = { key: config.baseURL, at: Date.now(), models };
+        return models;
+    }
+    /**
+     * Refresh the advisory catalog off the request path. Metadata only affects
+     * compaction and image handling, so a failed refresh must not fail a request.
+     */
+    ensureCatalog(config) {
+        const cached = this.catalog;
+        if (cached !== undefined && cached.key === config.baseURL && Date.now() - cached.at < CATALOG_TTL_MS)
+            return;
+        if (this.catalogRefresh !== undefined)
+            return;
+        this.catalogRefresh = (async () => {
+            try {
+                const apiKey = await this.dependencies.resolveApiKey(config.apiKeyEnv);
+                await this.loadCatalog(config, apiKey, true);
+            }
+            catch {
+                // Advisory metadata: keep serving built-in/configured values.
+            }
+            finally {
+                this.catalogRefresh = undefined;
+            }
+        })();
+    }
+    imageRequestTarget(_provider, _model, image) {
+        return this.imageTarget(this.dependencies.config(), image);
+    }
+    imageTarget(config, image) {
+        return {
+            ...requestImageDimensions(image.width, image.height, config.requestImagePixelBudget),
+            maxBytes: config.requestImageMaxBytes,
+        };
+    }
+    async *stream(options) {
+        if (options.stop !== undefined) {
+            throw new LlmError('Mojie Responses does not support GenerateOptions.stop', 'UNSUPPORTED_OPTION');
+        }
+        // One snapshot per request: a settings edit that lands mid-stream must not
+        // mix two generations of endpoint, limits, and model metadata.
+        const config = this.dependencies.config();
+        const provider = this.resolveModelWith(options.provider, options.model, config, this.discoveredModel(config, options.model));
+        if (options.reasoningEffort !== undefined && provider.reasoning === undefined) {
+            throw new LlmError(`provider "${options.provider}" model "${options.model}" does not support reasoning effort "${options.reasoningEffort}"`, 'UNSUPPORTED_REASONING_EFFORT');
+        }
+        this.ensureCatalog(config);
+        // The transport, clocks, SSE framing, and chunk translation are shared with
+        // standalone consumers (src/core/stream.ts); the adapter only contributes the
+        // DSH-specific image projection and request serialization.
+        const apiKey = await this.dependencies.resolveApiKey(config.apiKeyEnv);
+        yield* streamResponses({
+            endpoint: responsesURL(config.baseURL),
+            apiKey,
+            firstByteTimeoutMs: config.firstByteTimeoutMs,
+            streamIdleTimeoutMs: config.streamIdleTimeoutMs,
+            requestTimeoutMs: config.requestTimeoutMs,
+            ...options.signal === undefined ? {} : { signal: options.signal },
+            prepareBody: async (signal) => {
+                const requestImages = await this.prepareRequestImages(options, provider, config, signal);
+                return serializeRequest({ ...options, messages: this.projectOffloaded(options, requestImages) }, requestImages === undefined ? undefined : {
+                    requestImages,
+                    ...this.dependencies.resolveImageAccess === undefined ? {} : { resolveImageAccess: this.dependencies.resolveImageAccess },
+                });
+            },
+        });
+    }
+    async prepareRequestImages(options, model, config, signal) {
+        if (!options.messages.some((message) => contentHasImage(message.content)))
+            return undefined;
+        if (!model.inputModalities?.includes('image'))
+            return undefined;
+        const attachments = this.dependencies.resolveAttachments?.();
+        if (attachments === undefined) {
+            throw new LlmError('Mojie image input requires the durable attachment service', 'UNSUPPORTED_CONTENT');
+        }
+        const refs = new Map();
+        for (const message of options.messages) {
+            for (const block of message.content) {
+                if (block.type !== 'image' || block.offloaded === true)
+                    continue;
+                refs.set(block.attachment.attachmentId, block.attachment);
+            }
+        }
+        const versions = new Map();
+        for (const ref of refs.values()) {
+            versions.set(ref.attachmentId, await attachments.readImageRequest(ref, this.imageTarget(config, ref), signal));
+        }
+        const offloadImages = requiredImageOffload(options.messages, { representation: 'base64', maxBytes: config.maxRequestImageBytes }, block => {
+            const version = versions.get(block.attachment.attachmentId);
+            if (version === undefined)
+                return 0;
+            return version.bytes;
+        });
+        if (offloadImages > 0) {
+            throw new LlmError(`Mojie request images exceed the ${config.maxRequestImageBytes}-byte base64 budget; ${offloadImages} more oldest occurrence(s) must be offloaded.`, 'IMAGE_OFFLOAD_REQUIRED', { offloadImages });
+        }
+        return versions;
+    }
+    projectOffloaded(options, requestImages) {
+        if (requestImages === undefined)
+            return options.messages;
+        return projectOffloadedImages(options.messages, (ref) => {
+            return offloadedImageText(ref, this.dependencies.resolveImageAccess?.(ref));
+        });
+    }
+}