@freespace8/dsh-free-vision
DeepSeek Harness 的本地图像理解：通过 macOS Vision 进行 OCR、表格布局检测，以及对无文字图像的语义描述，完全在设备本地运行（图像绝不会离开你的 Mac）。本地化识图，图片不出本机。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add @freespace8/dsh-free-vision@0.1.2Free Vision 以 @freespace8/dsh-free-vision 发布，当前版本为 0.1.2。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Local image understanding for DeepSeek Harness. Uses the macOS Vision framework on your own machine — images never leave your Mac — and exposes the capability as agent-callable tools, so models without vision (e.g. DeepSeek v4 Flash) can still see.
English · 简体中文
view_image — semantic understanding of textless images: scene
classification, person/animal/face detection (nine-grid position),
QR/barcode decoding, composition focus, and an aesthetic score; when text
is detected it suggests using OCR instead;ocr_image — extracts all text from an image (reading order, Chinese
and English); with layout=true it also outputs table structure and
normalized coordinates — ideal for long screenshots, document photos, and
table pages;A real ocr_image run on a document containing Chinese, English, and a
table — fully on-device:
Inputs accepted: http(s) URLs / base64 / local absolute paths. Security boundaries: downloads refuse local/private network addresses (SSRF protection); uploads only accept loopback connections, are size-limited, and are validated by magic-number sniffing (client-supplied filenames are never trusted).
web profile;xcode-select --install).
The first call compiles the Swift script (~5–10 s); afterwards it is
cached.Published on npm — install with a single command (target profile web):
dsh plugin --profile web add @freespace8/dsh-free-vision
After installing, restart the web profile (dsh plugin add only updates
the manifest and dependencies; a running instance does not hot-load a new
bundle). To uninstall:
dsh plugin --profile web remove @freespace8/dsh-free-vision
Override any field in the profile's cordis.patch.yml by id (the config
block is replaced wholesale):
- id: free-vision
config:
scriptPath: "" # absolute path to ocr.swift; empty = use the bundled script
timeout: 120000 # swift execution timeout (ms); keep it generous for first-run compilation
saveDir: "" # directory for pasted images; default ~/Pictures/dsh-free-vision
maxImageSize: 20971520 # max pasted-image size (bytes); default 20 MB
src/index.js + src/swift.js + src/upload.js): registers
the view_image / ocr_image tools (ctx.tools.register); normalizes image
input (local path / http / base64), then spawns swift scripts/ocr.swift
via ctx.subprocess (process-group management, bounded output, timeout /
cancel termination); also registers the /plugins/dsh-free-vision/images
upload route (lazily on the webServer; skipped automatically in headless
mode);lib/client.js): registers a
conversation.input.right slot entry (renders null), captures image pastes
in the composer, POSTs each to the upload route, and appends the returned
absolute path to the draft (inputActions.setDraft).Tool output contract: output.schema = { type: 'string' }; render returns
[{ type: 'text', text }] blocks (a hard requirement of dsh-session for
tool-result blocks).
npm run check # artifact gate: syntax, export shapes, exports/files, patch line, client id, script existence
Real-behavior verification (after installing into a profile): restart the web
profile → open a new session → paste a screenshot with text into the composer
→ the draft contains a local absolute path → ask the model to call
ocr_image with that path → the image's text is returned.
MIT — see LICENSE. scripts/ocr.swift and parts of the host/client
logic are derived from
niyongsheng/free-vision-skill
(MIT, Copyright (c) 2026 Nico); the upstream copyright and license text are
retained in LICENSE.