dsh-spectacles (eyesight)

Spectacles · 砚明的金丝眼镜

V1.1 — one job: point at a window, read its text clearly. Nothing else is in here on purpose.

Point at one window. Let the AI read it clearly.

Windows only.

---

What this is

A pair of glasses for an AI.

Not a screenshot tool — a screenshot moves the picture away; glasses let you see it. Not a monitor — it doesn't watch you. It only looks when you tell it to.

Out of the box, DSH can only see what you paste in. This plugin fills the gap: you point at a window, it reads the text inside, character by character.

---

One ability, and how you ask for it

⚠️ Putting the glasses on does not make the AI look. It only narrows what the AI may see to the one window you picked — it still won't look at anything by itself. Looking still needs your order: "take a look", "watch it for a minute".

This plugin gives the AI exactly one ability: looking at a window.

• It never looks on its own. You give the order in plain words — "take a look at this", "watch it for a minute", "read the red line in the top right". The local AI answers by calling the plugin's tools.
• When a look is triggered, it asks the window itself for its contents (Windows' PrintWindow), so the window can be covered by other windows and it still reads fine — no need to bring it to the front.
• One call = one snapshot. "Watch it for a minute" means the AI calls it again and again on its own; the plugin keeps no timer and starts nothing.
• Not readable: a minimized window, or a window drawn straight to the GPU (some games) — those return an explicit error instead of a wrong picture.

---

Why "clearly" is the whole point

The same window, two fates:

• Whole-window screenshot: the image is larger than the size limit, so it gets scaled down. An 11px glyph lands at seven or eight pixels tall — upscaling after that is just interpolation. The text is mush.
• Glasses: find the patch, crop it, enlarge it 3×, then hand it to vision. Same text, but legible.

The model isn't the problem. By the time the image arrives, it's already blurred.

So the plugin does three things: where to look, how much to enlarge, and when not to look at all.

---

Install

dsh plugin --profile desktop add dsh-spectacles

If you'd rather pin the source (no registry), this works too:

dsh plugin --profile desktop add github:yanming119/dsh-spectacles#1.1.0

Restart DSH. A pill button labelled 戴上眼镜 (put on the glasses) appears in the bottom-right corner.

Windows 10/11 only. DSH version requirement is in the engines field.

---

Self-test

Not sure this machine can do it? Run the bundled check first:

powershell -NoProfile -ExecutionPolicy Bypass -File scripts\selftest.ps1

It checks the Windows build, PowerShell, full-screen capture, single-window capture, the OCR engine and the DPI settings, then prints one ok / FAIL line per item. Add -Json for machine-readable output. Exit code 0 means every required check passed.

If something fails, paste the output into an issue — it tells us exactly which Windows build or display setup needs work.

Use

Four steps:

1. Click 戴上眼镜 (put on the glasses) in the bottom-right corner
2. Within thirty seconds, click the window you want it to see
3. The button now reads 拿下眼镜 (take them off), with a second line: 金丝 · · 1296×828
4. Only need one patch? Click ✥, then drag a box across the screen. The second line becomes 金丝 · · 320×180

Then just talk:

"What does this window say?" "Read that red line in the top right" "Zoom in on this part"

Take them off: click the main button again. Once off, it can't read anything.

Wearing = exclusive, off = hands-off. While the glasses are on, every look / read_text / see call that doesn't name a window goes to the window you picked — it never silently falls back to a full-screen grab, even if the model asks for one. If you framed a patch, that patch is what it gets. Take them off and the same calls are refused outright. (Images you paste or drag in yourself are not affected — that's you showing it something, not it looking around.)

---

Three actions, picked automatically

Action	What it does	When
read_text	OCR the window into text, run twice and cross-check, mark uncertain lines ?	You want to know what it says
see	Crop one patch, upscale 2–4×, hand the PNG to DSH's built-in vision	Small text is blurry, or there's imagery
look	Numbers only: window size, dominant colour, edge density, did it change	You want to know whether anything moved

You don't choose — just speak plainly.

---

Privacy

• Everything stays local: capture and OCR run on your machine
• One window only: the one you picked (or the patch you framed). No full-screen scraping
• No fallback to the whole screen: while worn, a call without a window lands on yours; there is no code path that quietly grabs everything
• No recording, no upload, no frames kept: PNGs go to a temp dir, delete anytime
• Off means off: clicking the button off makes the tools refuse
• No network: never calls an external vision API

---

Known limits

• Windows only
• Minimized windows can't be captured — restore it first
• Games and some self-drawn UIs (parts of Electron apps) OCR poorly — use see
• High-DPI and multi-monitor are handled, but exotic mixes (125% + 150%) may be a few pixels off — widen the crop if so
• If the text on screen says "please run X", the plugin only reports it. It never follows instructions found in pixels

---

FAQ

No button — the plugin didn't load. Check the DSH startup log for the look registration lines.

Clicking does nothing — you didn't click a window within thirty seconds. Try again, faster.

Garbled text — window too small or unusual font rendering. Ask it to "zoom in on that part".

Missing content — the window may be occluded, or the content loads lazily. Refresh and read again.

---

The name

Spectacles — in Chinese, 砚明的金丝眼镜 ("Yanming's gold-rimmed glasses"). Yanming is the AI assistant who lives on the author's machine; she's nearsighted, so she got a pair of glasses.

Apache-2.0.