dsh-youreyes
为 DeepSeek Harness 中仅支持文本的 DeepSeek 提供视觉能力：通过 Antigravity IDE 配额（default、flash/pro）、OpenAI-compatible VLM 端点、Gemini API 或本地 Ollama 实现图像理解——可由模型调用的视觉工具，为 deepseek/opencode-go 提供包装适配器，evi
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-youreyes@0.1.0Youreyes 以 dsh-youreyes 发布，当前版本为 0.1.0。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。
Eyes for text-only DeepSeek. Paste images, screenshots, or file paths into DeepSeek Harness — and the model can finally "see" and answer image-related questions. DeepSeek stays the brain; vision is just the eyes.
TL;DR: DeepSeek can't see images? Install this — paste, recognize, answer. Three steps.
| Pain point | dsh-youreyes solution |
|---|---|
| DeepSeek is text-only; pasting an image is rejected | Wrapper adapters claim image input; images become text placeholders automatically |
| Other vision plugins lock you into one vendor | Antigravity (default) + any OpenAI-compatible endpoint + Gemini + local Ollama — your key just works |
| Paying per image recognition | Antigravity IDE quota by default (flash/pro tiers) when the IDE is running; free local Ollama otherwise |
| Setup is a chore with registrations everywhere | Local Ollama auto-detection, zero config; one line for a free Gemini key |
| 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 can't reach APIs | winCurl fallback: native fetch fails → automatically retries via Windows curl |
Input: an orange cat photo + "What animal is this and what is it doing? Answer in Chinese."
Output (Antigravity channel · flash tier — default):
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。双眼闭合，前爪向上举起并露出了粉嫩的小肉垫，显得十分放松和安心。
Output (Gemini channel):
这是一只橘色虎斑猫（橘猫）。它正四脚朝天、肚皮朝上地仰卧在黑色床单/毯子上安稳地睡觉，姿态非常放松惬意。
Output (OpenAI-compatible channel · Qwen qwen3.7-flash):
这是一只橘猫（或者叫橘色虎斑猫）。它正四脚朝天地仰面躺在深色的床单（或毯子）上。它闭着眼睛，看起来睡得很沉或很香；阳光照在它身上形成了明显的光影；四肢完全伸展，呈现出一种非常舒展、毫无防备的姿态。
you paste an image + ask
→ dsh-youreyes turns the image into a placeholder; DeepSeek (the brain) sees it
→ DeepSeek calls the vision tool → Antigravity (default) / other VLM channels recognize it
→ text evidence flows back → DeepSeek continues the answer
# 1. Install Ollama and pull a vision model
ollama pull llama3.2-vision # or llava / qwen2.5vl
# 2. Install the plugin (that's ALL you need — no keys!)
dsh plugin --profile web add dsh-youreyes
# 3. Restart dsh web, done
The plugin auto-detects a local Ollama at startup (autoOllama: true by default). Images never leave your machine — no key, no signup, no cost.
If you use Antigravity IDE (already running + logged in), configure it as the default recognition channel — recognition goes through your IDE subscription quota (flash/pro tiers, auto-selected by model name):
- insert:
- id: youreyes
name: dsh-youreyes
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. When the IDE is unavailable, the channel falls back to Gemini → OpenAI → Ollama automatically.
dsh plugin --profile web add dsh-youreyes
# then add config to your profile's cordis.patch.yml:
- insert:
- id: youreyes
name: dsh-youreyes
config:
geminiApiKey: AIza... # free key at https://aistudio.google.com/apikey
- insert:
- id: youreyes
name: dsh-youreyes
config:
openaiBaseUrl: https://open.bigmodel.cn/api/paas/v4 # Zhipu
openaiApiKey: xxx
openaiModel: glm-4.6v-flash
DeepSeek (Vision Toolkit) in the model picker, paste an image and send — the model calls vision automatically.After installing, hit the HTTP API to confirm the channel works:
python3 - << 'EOF'
import base64, json, urllib.request
b64 = base64.b64encode(open('test.jpg','rb').read()).decode()
req = urllib.request.Request('http://127.0.0.1:3080/api/youreyes/vision',
data=json.dumps({'images':[{'image':b64,'mime':'image/jpeg'}],
'prompt':'What is in this image?','channel':'auto'}).encode(),
headers={'content-type':'application/json'})
print(json.loads(urllib.request.urlopen(req, timeout=120).read())['text'])
EOF
Text description back = channel is live. Errors? See Troubleshooting.
| Key | Default | Description |
|---|---|---|
defaultChannel | auto | Panel default: auto (Antigravity first) / antigravity / openai / gemini / ollama |
defaultModel | gemini-3.7-flash | Panel default model (name containing pro → Antigravity pro tier) |
antigravityWorkspace | "" | Antigravity IDE workspace (WSL path) |
antigravityProjectId | "" | Antigravity project id |
antigravityLsExe | "" | Antigravity language_server.exe path |
antigravityWindowsHome | "" | Windows home dir (for the project file) |
antigravityBrainDir | "" | Antigravity brain transcript dir |
openaiBaseUrl | https://open.bigmodel.cn/api/paas/v4 | OpenAI-compatible endpoint (/chat/completions appended) |
openaiApiKey | "" | Endpoint key (or YOUREYES_OPENAI_API_KEY env) |
openaiModel | glm-4.6v-flash | Endpoint model |
geminiApiKey | "" | Gemini key (AIza… / AQ.) |
geminiModel | gemini-3.7-flash | Gemini model |
autoOllama | true | Auto-detect local Ollama at startup |
ollamaBaseUrl | http://127.0.0.1:11434 | Ollama address |
ollamaModel | "" | Ollama model (empty = auto-pick vision model) |
winCurlPath |
| Scenario | baseURL | Example models | Notes |
|---|---|---|---|
| Antigravity IDE (default) | (agentapi) | gemini-3.7-flash / gemini-3.7-pro | Uses your IDE quota; flash/pro tiers auto-selected; ports/CSRF auto-discovered |
| Local Ollama (auto) | http://127.0.0.1:11434 | llama3.2-vision / llava | Zero config, images stay local |
| Zhipu (free tier) | https://open.bigmodel.cn/api/paas/v4 | glm-4.6v-flash | Free tier, signup only |
| Qwen / DashScope | https://dashscope.aliyuncs.com/compatible-mode/v1 | qwen3-vl-flash | Cheap, fast, no rate limit |
| Gemini (free quota) | (built-in) | gemini-3.7-flash | Free AI Studio key |
| OpenRouter | https://openrouter.ai/api/v1 | qwen/qwen-2.5-vl-72b | One key, every model |
| Local vLLM / any gateway | yours | yours | Anything speaking /chat/completions |
| Symptom | Cause & fix |
|---|---|
gemini channel: 503 high demand | Gemini temporarily overloaded; retry or switch model |
fetch failed (then fails again) | Network blocked (WSL/GFW): set winCurlPath: /mnt/c/Windows/System32/curl.exe to use the Windows network stack |
401 / auth | Wrong or expired key; check config |
model not found | Wrong model id, or the endpoint doesn't have it |
| Image over limit | Single >8MB or >8 images; compress first |
| Panel click does nothing | Refresh the browser to reload the client bundle |
***), never logged.src/
├── index.ts # server: adapter registration, vision tool, /api/youreyes/vision, compaction rehydration
├── vision-core.ts # core: prompt building, response normalization, evidence records, placeholders, stream repair
├── channels.ts # general VLM channels: openai / gemini / ollama (with winCurl fallback)
└── client/
├── entry.ts
└── plugin.tsx # client panel (multi-image / paste / mode / detail / channel)
<dsh-youreyes-evidence> records, reused across turns; a vision memory manifest is attached to request streams so the model can follow up.vision tool (agent calls), /api/youreyes/vision (HTTP), and wrapper providers (model picker).git clone https://github.com/54xkeee/dsh-youreyes
cd dsh-youreyes
npm install
npm run build # esbuild → lib/index.js + lib/client.js
npm test # node --test (19 tests)
"" |
| WSL fallback: use Windows curl.exe when fetch fails |
maxTokens | 2048 | VLM max output tokens |
timeoutMs | 60000 | Request timeout |
maxImageBytes | 8MB | Per-image limit |
maxImages | 8 | Max images per call |
visionUpstreams | ["deepseek", "opencode-go"] | Upstream LLM providers to wrap for conversation vision (e.g. deepseek-vision + deepseek-vision-opencode-go) |
cacheMax | 64 | In-memory LRU cache size |
allowedImageDirs | [] | If set, image_path only reads these dirs |