Vision tools · DeepSeek Harness plugins — Page 16

PaddleOCR text-recognition and document-parsing skills with native DeepSeek Harness tools and GUI configuration.

PDF as a surface of the pack (row `dsh-pdf`, tab kind `pdf` plus the `pdfs` index page): the agent can actually READ and understand PDFs, SCAN the ones that are pictures of text, and you can read them properly in the right bar. FIVE tools - pdf_info, pdf_read, pdf_find, pdf_render, pdf_scan - are backed by a VENDORED pdf.js 6.3.289 (the same engine and version the shipped preview already uses) with its cMaps, standard fonts AND wasm image decoders, run in a CHILD process with `isEvalSupported:false`, a heap cap and a kill, so a corrupt or hostile document is a reported failure and never a wedged host. Extraction is cached per page under `$DSH_HOME/dsh-pdf/artifacts/<sha256>`, so reading, searching and re-reading a document costs one parse. `pdf_read` offers plain reading order AND a geometric layout reconstruction (lines grouped by baseline, columns and label/value gaps kept), which is what makes an invoice, a statement or a two-column paper legible. `pdf_scan` finds the pages that carry no text layer at all - where recognition is the only way to read them - draws them with the host's own rasterizer (poppler pdftoppm, mutool or Ghostscript) and recognizes them with tesseract, caching each result under every input that can change it (page, language, resolution, segmentation mode). It refuses to OCR a page that already has real text unless asked, and with no engine installed it says which one to install in as many words. The tab type replaces the shipped bare PDF renderer for `*.pdf`: fit-width continuous pages, page navigation, a zoom ladder, a selectable text layer and in-document search; a page that has not been drawn yet reserves the document's OWN page box at the current scale - never a viewport unit, since this reader is one pane of a dock - so the page column can never grow wider than the pane and the fit stays centred while the panel is resized; a THUMBNAIL RAIL and the document's own BOOKMARK OUTLINE in a side panel; every page that has no text layer announces it and offers a one-click scan whose recognized text appears under the page, labelled as a transcription. A second page type (`pdfs`, its own Start-page entry) lists every PDF in the workspace. Read-only: nothing this plugin writes can reach a PDF. No core patches, no fork, no npm dependencies, no network. Alpha.

DeepSeek Harness plugin: content-aware PDF reading tools for vision models (pdf_scan / pdf_read_page / pdf_render_region). Detects figures (vector+raster), tables, math, and two-column layout per page, then routes text pages to extraction and figure/table/math pages to high-DPI region crops rendered for read_image — avoiding DeepSeek's ~800x800 whole-page token ceiling. Backed by PyMuPDF; returns a clear warning (with the exact install command) when Python or a dependency is missing.

AI vision PDF translation and LaTeX reconstruction for DeepSeek Harness

基于 dsh-pet 0.3.8 的互动碎碎念版：每 30 秒识屏，用视觉模型在宠物头顶随口吐槽；新增人设导入。

DeepSeek Harness native plugin: vision capabilities for text-only LLMs (describe, OCR, VQA, layout analysis, clipboard)

Image recognition and image generation for DeepSeek Harness: model-facing tools that call any OpenAI-compatible external API (vision chat/completions and images/generations), with no built-in model defaults — a tool without a configured model fails with a clear error.

Vision sidecar for DeepSeek Harness: accept image attachments on text-only models, describe them, then send text to the main model.

[DEPRECATED] Vision bridge for the DeepSeek Harness web GUI — no longer maintained: newer DeepSeek Harness releases support image recognition natively, making this bridge unnecessary. Kept for reference only.

跑团 / DM 工具插件（DeepSeek Harness）：中立随机裁决 rp_random + 按会话隔离的战役配置（世界 / 角色卡 / 随机表 / 宏表）+ 世界书与状态追踪 + 本会话图片资源库 rp_assets（收录宿主生图结果与玩家自带图，按标签/角色检索复用）+ 角色立绘位（玩家导入自己的图）+ 设置页与会话面板 + PNG 故事书（SillyTavern 角色卡）导入（世界书/卡全文/卡面落进工作区，角色与世界写进会话，自动切 dm 预设开场）。**不做图片生成** —— 需要插图时由宿主自带的 generate_image / edit_image 负责，本插件负责把结果收进资源库复用；RP 工具只在 dm 预设会话注册。

按需视觉桥：图片描述、空间定位与 UI 还原，支持 OpenAI Chat Completions 兼容多模态接口

Folder-based skins for DeepSeek Harness Web: image/video backgrounds, token colours, logo, ambient audio.

Full 2-way Telegram Remote Control, Vision, Voice, Shell, Diff & Sound Notifications for DeepSeek Harness (DSH)

The vncode Web GUI's conversation-header package. It owns four controls on that header. The Page-zoom control (order -40, the leftmost of the four) drops a menu holding the level in force plus the two steps, Zoom in and Zoom out: it writes `zoom` on the document element along Chrome's own ladder cut at 50% and 200%, which is the browser's own Ctrl+ / Ctrl- PAGE zoom - the only way to get it in the native window the desktop launcher opens (a Tauri shell over the same dsh web, which has no such gesture at all) - and remembers the level per origin in localStorage, re-applying it before the first render. The Screenshot control (order -30, left of the Themes control) captures the whole window - the app is a fixed-viewport shell, so its 100% width and 100% height are exactly the tab's box - and saves the PNG to the Desktop of the machine running the app: the browser takes the pixels (getDisplayMedia on the current tab, one frame into a canvas) and this package's own host route writes the file, falling back to an ordinary browser download when that route is absent. The Themes control switches the app between the light, dark and system appearance, the same preference Settings > General > Appearance owns, and its menu also lists every theme REGISTERED into that same shipped registry - through ctx.theme.register, the documented third-party surface - which is where this package's own Nord, Monokai and Hacker come from (Nord's Polar Night surfaces, Snow Storm text, Frost accents and Aurora states; Monokai's near-black warm page, off-white body and pink/cyan/yellow/green syntax palette; Hacker's near-black green-cast page, phosphor body and signal-green/amber/cyan syntax palette - all three as alias-token overrides on the dark base; in-process by the shipped design, since the durable preference schema accepts the built-in three only). The button wears one static appearance mark rather than the active preference's sun/moon, and the menu is built from the registry's own list, so a theme is added by being registered. And the Session-log download seat replaces the shipped three-dot 'more actions' button (its only menu item was Download session log) with one plain download icon button that starts the export on the first click while the shipped sessionLogDownload controller keeps doing the work, its preparing/success/error dialog included. It also carries the pack's appearance overrides: the always-light paper the rendered Markdown view is drawn on, the chrome rule that hides the preview header's viewer menu on Markdown tabs so the page has exactly one renderer, the left column's top bar (the branding row becomes the same 76px band, ending in the same hairline, that the middle and right columns open with), the pack's VN branding on that row - the shipped mark and wordmark are replaced by a 24px black disc and the text vncode - and the header ring, so the right bar's own collapse/expand toggle in the header corner wears the same round hairline outline as this package's own header buttons. Alpha.

DeepSeek Harness (dsh) web plugin: API balance + today's spend sidebar widget, /api/dsh-usage route, external vision-call accounting (JSONL usage log), plus a full per-session usage dashboard (call log, per-model stats, cache rate, CSV export) via the usageDashboard session projection — merged from dsh-stats-dashboard.

DeepSeek Harness (DSH) 识图插件：为不具备原生识图能力的模型提供识图能力，把图片发送给阿里云百炼/智谱多模态模型并返回文字描述。由 claude-vision-skill 移植适配。

Composer-attached images are auto-described by an OpenAI-compatible vision model and handed to text-only models (DeepSeek) as text.

DeepSeek Harness plugin: a model-facing `vision` tool that describes and OCRs image files by calling the free Zhipu GLM vision API directly (no external CLI required).

A DeepSeek Harness Mix plugin for vision routing plus GPT Image generation and editing.

DeepSeek Harness plugin: a configurable provider route that transcribes pasted images via free Zhipu GLM vision models before delegating to a text-only adapter.

vision tool for DeepSeek Harness agents: describe local images via any OpenAI-compatible vision endpoint you configure (no built-in keys, optional multi-model cross-check)

Kimi WebBridge host plugin for DeepSeek Harness: model tools that drive the user's REAL browser (navigate, snapshot, click, fill, evaluate, screenshot, network inspection, tab management) through the local kimi-webbridge daemon

Bridge WeChat (iLink bot) to DeepSeek Harness (DSH)

WeChat Official Account content studio for DeepSeek Harness: anti-homogenization writing methodology, blessing-image visual baseline, gpt-image cover pipeline with OCR acceptance, interaction rules, and the measured xiaolvshu (newspic) draft web API.