@caesarloo/dsh-deepseek-account-vision
English | 中文
English
Give any main model in DeepSeek Harness the vision of the DeepSeek account: recognize images with DeepSeek-V4.1-Flash (deepseek-flash, release 1017) — through a deepseek_vision tool, or automatically for images attached to a text-only chat model. No API key, no extra VLM account, no image library.
What it does
- Adds a model-facing tool
deepseek_vision: pass a local PNG/JPEG/WebP/GIF path or an http(s) image URL, get back a complete textual rendition (verbatim OCR, layout, visual elements, key details).
- Answers with the DeepSeek account route (
deepseek-account / deepseek-flash by default), so it reuses the DeepSeek login you already have in DSH — the plugin never stores credentials.
- Registers automatic image-transcription routes: for every registered provider it adds a
<provider>-vision route that declares image input, turns each attached image into text with DeepSeek-V4.1-Flash, then delegates the conversation to the original model. A pure-text main model can finally "see" a pasted screenshot.
- Caches results by image content hash (in-process only), merges concurrent identical requests, cools down a failed image for 60 s, and fails fast instead of hanging.
- Works alongside everything: text-only conversations pass through untouched; the original model, tools and history shape stay the same.
Install
dsh plugin --profile <profile> add @caesarloo/dsh-deepseek-account-vision
From a local checkout, install by directory instead: dsh plugin --profile <profile> add <project-directory>.
Restart DSH once (bundle layers have no hot reload), then verify that deepseek_vision appears in the tool list of a new session. A patch-layer development install (- insert: with the plugin entry file) reloads hot without a restart.
Requirements: DSH >= 0.2.1-rc.2, a signed-in DeepSeek account, and the standard llm / attachments / fs services (all present in the shipped bundle).
Usage
1. Tool — any main model can call it
{
"image_path": "screenshot.png",
"prompt": "Transcribe every visible character and describe the layout"
}
or, for a remote image:
{ "image_url": "https://example.com/chart.png" }
The tool returns { text, model, source, truncated }. The image never enters the main model's context — only the transcription does. A custom prompt is passed to the vision model verbatim; leave it out to use the built-in full-transcription prompt.
2. Automatic transcription route — paste images into a text-only chat
Pick the route named <provider> + 自动识图 (id <provider>-vision, configurable) in the model selector, then attach an image as usual. The image block is transcribed before the original adapter sees the request:
GUI image ──▶ <provider>-vision route (declares image input → attachment admitted)
│ each image block is transcribed by DeepSeek-V4.1-Flash
▼
original model (receives plain text, answers normally)
Transcribed text is prefixed with [图片转译]. When transcription fails and failureMode is placeholder (default), an explicit [图片转译失败: …] marker is inserted and the conversation continues.
| Situation | What to use |
|---|
| Any model, image exists as a file or URL | deepseek_vision tool |
| Text-only model, image attached in the GUI | the <provider>-vision route |
| Main model already accepts images | no plugin needed for that turn |
Configuration
Defaults ship in cordis.patch.yml; override them in your profile's cordis.patch.yml with an - id: dsh-deepseek-account-vision entry. Unlisted keys fall back to the defaults below.
| Key | Default | Meaning |
|---|
visionProvider | deepseek-account | Registered provider route used for image recognition |
visionModel | deepseek-flash | Vision model id (DeepSeek-V4.1-Flash / 1017) |
visionMaxTokens | 4096 | Output cap for the vision model |
timeoutMs | 120000 | Per-image recognition timeout |
tool | true | Register the deepseek_vision tool |
toolName | deepseek_vision | Tool name exposed to models |
autoTranscribe | true | Register automatic image-transcription routes |
wrapAllProviders | true | Wrap every registered provider (including ones registered later) |
innerProviders | [] | Providers to wrap when wrapAllProviders=false |
routeSuffix | -vision | Route id suffix: <provider><routeSuffix> |
routeNameTemplate | {name} + 自动识图 | Route display name; {name} / {provider} are substituted |
marker | [图片转译] | Prefix added to each transcription |
failureMode | placeholder | placeholder keeps the turn alive; error fails the whole turn |
transcribePrompt | '' | Override the automatic-transcription prompt (empty = built-in) |
recognitionPrompt | '' | Override the deepseek_vision prompt (empty = built-in) |
cacheSize |
Dependencies
- No API key, no separate VLM account. Recognition goes through
ctx.llm.stream() on the DeepSeek account route; DSH's account service owns the credentials.
- No image library. Attachment normalization (size, pixels, format validation) is done by DSH's attachment service; the plugin only ships
schemastery for config.
- Node built-ins only beyond that (
crypto, path, global fetch / AbortSignal).
Boundaries
- It does not turn the main model into a multimodal model: the original model still receives only text. It does make image input usable for it via transcription.
- It does not replace
read_image: read_image hands the native image block to an image-capable model; deepseek_vision returns text so any model can consume it.
- It only accepts PNG / JPEG / WebP / GIF — no PDF, audio or video.
- Transcription is a lossy summary for dense screenshots: very small text or fine UI details may be missed. Use a stronger
visionModel or a more explicit prompt when exactness matters.
- Images are sent to DeepSeek with the recognition request (that is the point) and are not otherwise persisted by the plugin; the result cache lives in process memory only.
Development
pnpm install
npm test # unit tests + BOM + fingerprint/credential + bundle-manifest checks
npm run check-leaks # fingerprint scan only
prepublishOnly runs npm test, so a package with a failing test, a BOM-prefixed source file, or a machine-specific fingerprint cannot be published. Publish explicitly against the official registry:
npm publish --registry https://registry.npmjs.org/
License
MIT
中文
让 DeepSeek Harness 里的任意主模型都用上 DeepSeek 账号的视觉能力：用 DeepSeek-V4.1-Flash（deepseek-flash，1017 版）识别图片——既可以由模型调用 deepseek_vision 工具，也可以给纯文本主模型自动转译 GUI 里贴的图片。不需要 API Key，不需要另找视觉模型账号，不装图像库。
它做什么
- 注册模型可见的工具
deepseek_vision：给一个本地 PNG/JPEG/WebP/GIF 路径或 http(s) 图片地址，返回完整文字转译（逐字 OCR、版面结构、视觉元素、关键细节）。
- 识图后端走 DeepSeek 账号路由（默认
deepseek-account / deepseek-flash），复用 DSH 里已登录的 DeepSeek 账号，插件自身不保存任何凭据。
- 注册自动转译路由：为每个已注册 provider 加一条
<provider>-vision 路由，对外声明支持图片输入，收到图片后先用 DeepSeek-V4.1-Flash 转成文字，再交给原模型作答。纯文本主模型也能“看到”粘贴的截图。
- 按图片内容哈希缓存（仅进程内）、合并并发的同图请求、失败图片冷却 60 秒、超时快速失败而不是卡死。
- 与现有能力并存：纯文本对话零改动直达原模型，原模型、工具与会话历史结构都不变。
安装
dsh plugin --profile <profile> add @caesarloo/dsh-deepseek-account-vision
本地检出目录安装：dsh plugin --profile <profile> add <项目目录>。
安装后重启一次 DSH（bundle 层不支持热重载），再在新会话的工具列表里确认出现 deepseek_vision。开发期也可以用 patch 层 - insert: 指向插件入口文件装载，改完热生效、无需重启。
依赖：DSH >= 0.2.1-rc.2、已登录的 DeepSeek 账号，以及标准的 llm / attachments / fs 服务（随官方 bundle 提供）。
使用
1. 工具——任意主模型都能调
{
"image_path": "screenshot.png",
"prompt": "逐字识别所有可见文字，并说明版式"
}
远程图片：
{ "image_url": "https://example.com/chart.png" }
返回 { text, model, source, truncated }。图片本身不会进入主模型上下文，只有转译文字会。自定义 prompt 会原样交给识图模型；留空才使用内置的完整转译提示词。
2. 自动转译路由——给纯文本模型直接贴图
在模型选择器里选 <provider> + 自动识图（id 为 <provider>-vision，可配置），然后照常附加图片。图片块会在原适配器收到请求之前被转译：
GUI 图片 ──▶ <provider>-vision 路由（声明支持图片 → 附件放行）
│ 每个 image 块交给 DeepSeek-V4.1-Flash 转译
▼
原模型（收到纯文本，正常作答）
转译文本带 [图片转译] 前缀。转译失败且 failureMode 为默认的 placeholder 时，会插入 [图片转译失败: …] 占位文本并继续对话。
| 场景 | 用什么 |
|---|
| 任意模型，图片是文件或 URL | deepseek_vision 工具 |
| 纯文本模型，在 GUI 里贴图 | <provider>-vision 路由 |
| 主模型本来就支持图片 | 该轮不需要本插件 |
配置
默认值随 cordis.patch.yml 提供；在 profile 的 cordis.patch.yml 里用 - id: dsh-deepseek-account-vision 条目覆盖。未列出的键回落到下表默认值。
| 键 | 缺省 | 含义 |
|---|
visionProvider | deepseek-account | 识图使用的已注册 provider 路由 |
visionModel | deepseek-flash | 识图模型 id（DeepSeek-V4.1-Flash / 1017） |
visionMaxTokens | 4096 | 识图模型输出上限 |
timeoutMs | 120000 | 单张图片识图超时 |
tool | true | 是否注册 deepseek_vision 工具 |
toolName | deepseek_vision | 暴露给模型的工具名 |
autoTranscribe | true | 是否注册自动转译路由 |
wrapAllProviders | true | 包裹全部已注册 provider（含之后动态注册的） |
innerProviders | [] | wrapAllProviders=false 时要包裹的 provider |
routeSuffix | -vision | 路由 id 后缀：<provider><routeSuffix> |
routeNameTemplate | {name} + 自动识图 | 路由显示名；{name} / {provider} 会被替换 |
marker | [图片转译] | 每条转译文本的前缀 |
failureMode | placeholder | placeholder 保住本轮对话；error 让整轮失败 |
transcribePrompt | '' | 覆盖自动转译提示词（留空用内置） |
recognitionPrompt | '' | 覆盖 deepseek_vision 提示词（留空用内置） |
cacheSize | 200 | 进程内结果缓存条数（0 关闭） |
urlTimeoutMs | 30000 | 下载 http(s) 图片的超时 |
依赖约定
- 无 API Key、无独立视觉账号：识图经
ctx.llm.stream() 打在 DeepSeek 账号路由上，凭据由 DSH 账号服务持有。
- 无图像库：尺寸/像素/格式归一化由 DSH 附件服务完成；插件自身只依赖
schemastery 做配置校验。
- 其余只用 Node 内建能力（
crypto、path、全局 fetch / AbortSignal）。
边界（明确不做）
- 不把主模型变成多模态模型：原模型收到的仍然只有文字；本插件做的是“把图片转成文字再交给它”。
- 不替代
read_image：read_image 把原生图片块交给支持图片的模型；deepseek_vision 返回文字，所以任何模型都能消费。
- 只接受 PNG / JPEG / WebP / GIF，不支持 PDF、音频、视频。
- 转译对密集截图是有损摘要：极小的字或精细 UI 细节可能丢失；要求精确时换更强的
visionModel 或把 prompt 写具体。
- 图片只随识图请求发送到 DeepSeek（这正是功能本身），插件不另行持久化；结果缓存只在进程内存中。
开发与发布
pnpm install
npm test # 单测 + BOM + 本机指纹/凭据 + bundle 清单一致性检查
npm run check-leaks # 只跑指纹扫描
prepublishOnly 跑 npm test：单测失败、源码带 BOM、或命中本机指纹/凭据时都发不出去。发布时显式指定官方 registry：
npm publish --registry https://registry.npmjs.org/
License
MIT
