---

DSH version support (important for GitHub users)

• Works with DSH rc.5 (legacy, no auth) and 0.1.2 / 0.1.3+ (per-launch token auth). The launcher auto-detects: GET / returns 200 -> legacy plain URL; returns 401 -> opens the per-launch token URL printed by dsh web.
• Upgrading to the latest DSH? Read skills/dsh-upgrade-runbook first (BOM/schema pitfalls, official git-clone steps, preflight script).
• NOT upgrading? Nothing to do - old DSH keeps working in legacy mode.
• Backup: this repo's git history is the old-version fallback; always back up your DSH_HOME (sessions/settings/storages) before upgrading.

✨ Highlights

🚀 One-Click Entry	Launch Harness + Quark + Qianwen floating sidebar with one icon
🧠 AI Brain	DeepSeek (cloud) + local LLMs — reason, plan, generate
👁️ Eyes	Vision skills — Ollama local, Quark-browser Qwen, or any OpenAI-compatible API
🎬 Studio	Video production — Doubao Seedance 2.0, RunningHub MiniMax H3, all web-automation driven
🧩 Everything is a Plugin	Skills (how-to) + Plugins (can-do), freely composed & replaced
🔥 Zero API Key Matrix	DeepSeek brain + Qwen vision/image-gen + Doubao video + H3 cloud video — all via CDP web automation, no Ollama, no API keys, no GPU needed

---

📦 What's Inside

SuperMate Harness System
 ├── skills/                        → Skill layer (task-level: how-to)
 │    ├── Supermate-harness-launcher/ 🚀 one-click launch: Harness + Quark + Qianwen sidebar
 │    ├── quark-qwen-vision/         🖼️ Quark-browser vision + image gen (zero API key)
 │    ├── Deepseek-eyes/             👁️ vision skill (Ollama / vision API)
 │    ├── doubao-creator/            🎬 Doubao text-to-video (Seedance 2.0, zero API key)
 │    ├── rh-workflow/               🎥 RunningHub cloud H3 workflow API
 │    ├── rh-workflow-9b/            📋 9B RH executor charter
 │    ├── MiniMax h3-video-producer/ 🎞️ H3 video production pipeline
 │    ├── Supermate/                 🤖 the SuperMate agent identity
 │    └── DSH Official/              official deepseek-ai skills (13 · fork copy)
 └── plugins/                        → Plugin layer (system-level: can-do)
      ├── Supermate/                 in-house plugins (core enhancements = closed-source commercial)
      └── DSH Official/              official deepseek-ai plugins (49 families · fork copy)

---

⚡ Quick Start

One-click environment — Harness + Quark + Qianwen floating sidebar:

# 1. Copy the launcher skill into DSH's skill directory
Copy-Item skills\Supermate-harness-launcher\ ~\.dsh\skills\ -Recurse

# 2. Edit config.ps1 — set your Harness start command / Quark path
#    $HarnessStartCmd = 'E:\deepseek-harness-v013\start-web.cmd'   (empty = start Harness manually)

# 3. Run it — Harness starts, Quark opens, the Qianwen floating sidebar pops up
powershell -ExecutionPolicy Bypass -File scripts\launch.ps1

Any other skill works the same way: copy its folder into ~/.dsh/skills/ and restart the DSH session.

---

🧩 Original Skills

Built around the Quark browser — Quark ships with the built-in Qianwen sidebar + Quark Netdisk, making it the ideal foundation for the SuperMate Harness System's extensions (vision, image generation, netdisk, video).

Skill	What it does
Supermate-harness-launcher	🚀 One-click entry (foundation) — start Harness → open Quark (debug port 9222) → open the Harness GUI + Qianwen sidebar → auto-click Quark's native "问AI" button so the floating Qianwen side panel pops up next to your conversation. Cross-machine / resolution / DPI aware
quark-qwen-vision	🖼️ Zero API key vision + image gen — drive Quark browser's built-in Qwen (qwen-vl + Qwen-Image 2.0) via CDP: image analysis, prompt reverse-engineering, text-to-image
Deepseek-eyes	👁️ Give text models eyes — image → local vision (Ollama) / OpenAI-compatible vision API → structured text
dsh-deepseek-vision-bridge	🧠 ARCHITECTURAL BREAKTHROUGH — text DeepSeek natively reads images — a patch on DSH's own llm-deepseek adapter: paste an image into the Harness GUI and deepseek-v4-flash (pure text, inputModalities:['text']) accepts it. The adapter auto-routes each image through the Quark Qianwen sidebar (CDP, zero VRAM / zero API key) and injects a clean Chinese description before serialization. The model never "sees" the image — yet fully understands it. Works for GUI, IM, any DeepSeek-bound image. Multi-image serial + cooldown, success-only caching, auto-retry, zero regression without the attachment seam
doubao-creator	🎬 Zero API key text-to-video — drive Doubao web (Seedance 2.0) via CDP: image analysis, 10s vertical video ads, portrait-protection workaround (describe-then-anchor)
dsh-im-wecom	💬 WeCom bot × DSH — connect WeCom smart bots to Harness via the @xmanrui/dsh-im plugin (official long connection, zero hand-written bridge); Qianwen sidebar image proxy: WeCom images → Quark Qwen CDP analysis → text-only into the model (text-only models can "see" images)
rh-workflow	🎥 RunningHub cloud H3 workflow API — I2V / T8 / Ref2VA video generation
rh-workflow-9b	📋 9B RH executor charter — no-drift execution rules for local 9B models
MiniMax h3-video-producer	🎞️ Local H3 full video production: storyboard → H3 prompts → segment generation → assembly + BGM
video-production-studio	🎥 General-purpose video production pipeline — distilled from zenstory-ai/drama-skills (MIT): five creator documents + preview→confirm→produce gate + native MiniMax H3 prompts + local channel wiring (Z-Image / RH Ref2VA·I2V / Doubao Seedance / ffmpeg)
lyric-mv-storyboard	🎵 Lyric-driven MV storyboard — lyrics (or .lrc timeline) + character image → three-layer acting system (goal/beat/facial timing) → timed storyboard + Ref2VA singing prompts (one line per clip to avoid recitation); portable skill.json sources included
SuperMate	🤖 The SuperMate agent identity — how the assembled agent behaves end-to-end

---

🚀 Featured · One-Click Launch — Harness + Quark + Qianwen Sidebar

Double-click one icon → Harness starts → Quark opens → the Qianwen floating sidebar slides out next to your conversation.

Supermate-harness-launcher is a single entry point for your whole working environment:

1. Start Harness — if http://127.0.0.1:3080 is not up, launch your configured start script and wait until ready (≤120s)
2. Open Quark — ensure the browser runs with --remote-debugging-port=9222, open the Harness GUI tab + the Qianwen sidebar page
3. Pop the floating sidebar — screenshot-locate Quark's native "问AI" toolbar button (blue ✨ star, top-right) and auto-click it

Cross-machine / resolution / DPI aware: the button's pixel position varies, but its ratio inside the window is stable — DPI-aware capture + color scan of the blue star (topmost cluster) + ratio-scaled fallback.

⚠️ The "问AI" button is a toggle (running again closes the panel). The Qianwen sidebar page must use the parameterized URL (entry=frame&tab_id=...) — the plain URL only loads a broken shell.

---

👁️ Featured · Give DeepSeek Eyes

Image → vision model converts it to text → the text model "sees"

deepseek-eyes converts images — screenshots, photos, charts, design drafts, illustrations, character sheets, AI-generated images — into structured text that DeepSeek (or any text model) can reason about.

Image / graphic file → Skill → local vision model or vision API → structured text → DeepSeek reads & reasons

Measured on RTX 5080 16GB — English manual · 中文手册

---

🧠 The Architectural Breakthrough · dsh-deepseek-vision-bridge

"Seeing" is a capability, not a model property. DeepSeek's weights contain no vision encoder — that is a physical limit no prompt trick can cross. But this bridge proves the harness can cross it at the architecture level: the vision channel (Quark Qianwen sidebar) is promoted from "call it by hand" to "an automatic stage of the model-request pipeline".

Paste image into GUI → host accepts (adapter now declares image input)
  → image kept durably (visible in history, never lost)
  → llm-deepseek adapter, before every request, scans image blocks
  → Quark Qianwen sidebar (CDP) describes each one (~6-12s, cached after)
  → clean text replaces the image block → serialize (assertTextOnly never fires)
  → DeepSeek answers as if it saw the picture

Why this is a breakthrough — not a trick:

• Native layer, universal effect — the patch lives in DSH's own llm-deepseek adapter, so every DeepSeek-bound image — GUI paste, IM channels, anything — goes through it. No per-channel bridge to maintain (unlike bridge-only approaches).
• The image is never lost — original bytes stay in the durable attachment store; only the copy sent to the model becomes text.
• Zero VRAM, zero tokens — vision rides the already-open Quark browser over CDP; nothing loads into the 16 GB GPU that ComfyUI H3 sampling needs.
• Built to run, not demo — strict serial + 1.5s cooldown (the sidebar is a web session and stalls under bursts), success-only caching (same image = ms on repeat), failure degrades to a placeholder and auto-retries next turn (never poisons later requests), and the composition without the attachment seam keeps native text-only behavior (zero regression, 158 tests green).
• Upstream-general, not SuperMate-private — the patch touches only the official @deepseek-ai/dsh-llm-deepseek package; every dependency is @deepseek-ai/* + Node built-ins + stock Quark CDP. Clone official deepseek-ai/deepseek-harness, drop in patch/, rebuild, restart — same capability. This repo is just its distribution point (upstream-contributable).
• Two agents beat a projector — local multimodal models bolt on an mmproj for weight-level static alignment (a fixed local eye you cannot interrogate). This bridge instead has the text agent (DeepSeek) auto-invoke the cloud full-strength Qianwen (qwen-vl-max class) at runtime: a borrowed eye far sharper than any local 9B/35B vision model, and one you can conversationally interrogate, swap, and iterate — impossible with a static mmproj.
• Honest boundary — first image costs ~10 s inside the model request; the cache is per-process; the sidebar being a session is the ceiling for extreme multi-image bursts.

Source patch + docs: skills/dsh-deepseek-vision-bridge/patch/ · rollback included.

---

🏛 Official Content Ships With This Repo

	Count	Source
skills/DSH Official	13 official skills	deepseek-ai · Apache-2.0 · NOTICE-attributed fork copy
plugins/DSH Official	49 official plugin families	deepseek-ai · Apache-2.0 · NOTICE-attributed fork copy

---

🔗 Ecosystem & License

Part of the DeepSeek Harness plugin ecosystem — topic dsh-plugin · DeepSeek Harness · Agent Skills

License · Apache-2.0 with patent terms · dual-license boundary (closed-source commercial core: scheduling / memory governance / sandbox / evaluation) — see 双许可说明

---