dsh-vision-web

👁️ dsh-vision

Eyes for text-only DeepSeek on DeepSeek Harness — powered by Doubao Web by default, zero cost, no API key.

简体中文

TL;DR: no API key, no paid tier — log into Doubao once in your browser, and DeepSeek can see images in every conversation. Paste, recognize, answer.

✨ Why you'll love it

Pain point	dsh-vision solution
Vision APIs cost money and need keys	Doubao Web by default — zero cost, no API key, just a browser login
DeepSeek is text-only; pasting an image is rejected	Wrapper adapters claim image input; images become text placeholders automatically
Other plugins lock you into one vendor	Doubao Web (default) + Antigravity IDE quota (flash/pro) + Gemini API + Cockpit proxy — auto fallback chain
The model "forgets" what it saw	Vision evidence memory: results persist in the session, reused across turns, restored after compaction
Paying to re-recognize the same image	Content-hash cache: same image + same question = recognized once per process
Complex images get shallow answers	Auto detail escalation: standard pass first, auto-upgrade to deep for complex scenes
WSL / firewalled networks	winCurl fallback for API channels; Doubao Web channel runs through your Windows browser

🎯 Real results (2026-08, full end-to-end calls)

Input: an orange cat photo + "What animal is this and what is it doing? Answer in Chinese."

Output (Doubao Web channel — default):

这是一只橘猫（家猫），它四仰八叉仰躺在床上睡觉，肚皮露在外面，四肢舒展，睡得十分放松惬意。猫咪把肚子露出来，说明它对周围环境很有安全感。

Output (Antigravity channel · flash tier):

这是一只橘猫（橘色虎斑家猫）。它正仰面熟睡/惬意放松：四脚朝天、露出圆滚滚毛茸茸的肚子，正舒适地躺在深色床垫/毯子上睡觉。

you paste an image + ask
  → dsh-vision turns the image into a placeholder; DeepSeek (the brain) sees it
  → DeepSeek calls the vision tool → Doubao Web (default) / other channels recognize it
  → text evidence flows back → DeepSeek continues the answer

🚀 Quick start

Path 0: Doubao Web (default, zero cost)

• Install the plugin:

  dsh plugin --profile web add dsh-vision-web
  

• Start the Windows bridge (drives your logged-in Chrome):

  # on Windows: connect a Chrome with remote debugging, then run
  node bridge.mjs    # see the Bridge section below
  

• Log into doubao.com in that Chrome once.
• Restart dsh web, paste an image — done. No API key anywhere.

Path 1: Antigravity IDE quota (if you use Antigravity)

Configure your workspace and the plugin auto-prefers it (flash/pro tiers by model name):

- insert:
    - id: vision
      name: dsh-vision
      config:
        antigravityWorkspace: /path/to/workspace
        antigravityProjectId: your-project-id
        antigravityLsExe: /path/to/language_server.exe
        antigravityWindowsHome: /mnt/c/Users/you
        antigravityBrainDir: /mnt/c/Users/you/.gemini/antigravity/brain

Ports/CSRF are auto-discovered on every call — no manual config after IDE restarts.

Path 2: Gemini API (fallback channel)

      config:
        genlangKey: AIza...   # https://aistudio.google.com/apikey

Usage

• Panel: click 「识图」 in the session header → add/paste images → prompt → mode/detail/channel → recognize.
• In conversation: pick a deepseek-vision route in the model picker, paste an image and send — the model calls vision automatically.

🌉 The Doubao bridge (how the zero-cost channel works)

The Doubao Web channel automates your logged-in browser instead of calling an API:

DSH plugin (WSL) ──submit──▶ queue service (127.0.0.1:9340)
                                  ▲ polling
Windows bridge (node + puppeteer-core) ──┘
        │ CDP connect to Chrome (remote-debugging-port)
        ▼
Doubao sidebar: upload image → type question → Enter → wait for reply
        │
        ▼ text reply → POST /result → DSH plugin

• One-time login: log into Doubao once in the bridge's Chrome profile; the bridge reuses it forever.
• Windows → WSL localhost forwarding carries the queue traffic; no firewall changes needed.
• Ships as bridge.mjs — run it once beside DSH (node bridge.mjs), it polls forever.
• WSL cannot reach Windows ports directly; that's why the queue lives on the WSL side and the bridge polls from Windows.

🧠 The vision engine — how complex recognition actually works

A naive "see → describe" loop fails on dense screenshots, tables, UI mockups and multi-image comparisons. dsh-vision turns recognition into a structured, self-escalating, memory-backed pipeline:

1. Auto detail escalation — it knows when one pass isn't enough

detail: auto runs a two-pass strategy:

pass 1 (standard + triage) ──▶ complexity == "simple" ──▶ done
                            └─▶ complexity == "complex" ──▶ pass 2 (deep) ──▶ escalated result

The vision model itself classifies complexity. These all count as complex and trigger the deep pass automatically:

• multi-subject relationships · dense small text · OCR-heavy content
• tables / charts / code / UI screens · counting · comparison / spot-the-difference
• professional imagery · multi-step spatial reasoning

2. Four task modes, one tool

Mode	What it does	Typical use
glance	general understanding, evidence selected around your question	everyday questions
ocr	transcribes visible text in natural reading order, preserving headings/tables/UI hierarchy	screenshots, docs, error messages
region	focuses on one area — normalized coords 0.1,0.2,0.8,0.9 or plain language ("top right")	UI bugs, chart details
compare	item-by-item differences between ≥2 images, with confidence	before/after, versions, A/B

3. Structured evidence, not raw prose

Every pass asks the VLM for a strict JSON evidence object:

{
  "complexity": "simple|complex",
  "base_evidence": {
    "summary": "neutral overview",
    "ocr": "visible text (empty if none)",
    "layout": ["layout observations"],
    "entities": ["entities"],
    "relations": ["relations"],
    "uncertainty": ["explicit unknowns"]
  },
  "query_answer": "direct answer to the user"
}

Observations vs. inference are separated, uncertainty is explicit (never hallucinated), and lists are capped to keep context tight.

4. Long-context visual memory — the model never "forgets" what it saw

• Every result is written into the session timeline as a durable <dsh-vision-evidence> record.
• Reuse across turns: the same image + same question hits the existing record — recognized once, remembered forever.
• Vision memory manifest: every request stream carries a compact catalog of recent evidence, so the model can follow up — zoom in, re-OCR, compare against a new screenshot — without you re-pasting.
• Compaction rehydration: after DSH compresses a long session, recent vision records are restored automatically.

5. Content-hash caching — never pay twice for the same pixels

Cache key = SHA-256(image bytes) + prompt + detail + mode + region + model + channel + prompt-version, in-process LRU.

⚙️ Configuration

Key	Default	Description
defaultChannel	auto	auto (Antigravity first if configured, else Doubao Web) / web / antigravity / genlang / cockpit / aicode
defaultModel	gemini-3.7-flash	Panel default model (name containing pro → Antigravity pro tier)
webChannel.enabled	true	Doubao Web channel on/off
webChannel.queuePort	9340	WSL queue port
webChannel.timeoutMs	240000	Web reply timeout
antigravityWorkspace	""	Antigravity IDE workspace (WSL path)
antigravityProjectId	""	Antigravity project id
antigravityLsExe	""	language_server.exe path
antigravityWindowsHome	""	Windows home (project file)
antigravityBrainDir	""	Brain transcript dir
genlangKey	""	Gemini API key (AIza… / AQ.)
cockpitBaseUrl / cockpitKey	http://127.0.0.1:65386 / ""	Cockpit proxy
oauthAccount / oauthClientId / oauthClientSecret	""	aicode direct channel (bring your own credentials)
visionUpstreams	["deepseek"]	Upstream LLMs to wrap for conversation vision
cacheMax	64	In-memory LRU cache size
allowedImageDirs	[]	If set, image_path only reads these dirs
curlPath	/mnt/c/Windows/System32/curl.exe	Windows curl for WSL fallback

🔧 Troubleshooting

Symptom	Cause & fix
Web channel times out	Bridge not running (node bridge.mjs) or Chrome not logged into Doubao — check both
Web channel: fetch failed from bridge	Queue service not up (the DSH plugin starts it; check plugin loaded)
Antigravity: 找不到 language_server.exe	Antigravity IDE not running — start & log in
Gemini 503 high demand	Gemini overloaded; retry or switch model
WSL can't reach APIs	Set curlPath (Windows curl) — API channels fall back to it automatically

🔒 Privacy

• Doubao Web: images go to your own logged-in Doubao session — same as using the website by hand.
• API keys: stored in your local config only; error messages are auto-redacted, never logged.
• OAuth credentials: not bundled — the aicode channel asks you to bring your own, kept in your local config.

🏗️ Architecture

src/
├── index.ts        # server: adapters, vision tool, /api/vision, compaction rehydration
├── vision-core.ts  # prompt building, response normalization, evidence records, stream repair
├── web-channels.ts # Doubao Web queue service (WSL side)
└── client/
    └── plugin.tsx  # client panel (multi-image / paste / mode / detail / channel)

bridge.mjs          # Windows bridge: polls the queue, drives Doubao via CDP

🛠️ Development

git clone https://github.com/54xkeee/dsh-vision
cd dsh-vision
npm install
npm run build     # esbuild → lib/index.js + lib/client.js
npm test          # node --test

📄 License

MIT

🙏 Credits

• vision-toolkit architecture (placeholders + vision tool + wrapper adapters + evidence memory) shared with dsh-youreyes
• Channel and error-handling conventions follow community plugins like dsh-vision-proxy