dsh-labrador
A Labrador that lives in the DeepSeek Harness window.
image
She is made of seventeen cut-out photographs of my dog DaiDai, driven by a small state machine. She decides for herself what to do, she can be poked, dragged and thrown, she speaks in dog, and she can be tuned from a settings page.
The frames are photographs of one specific dog. Installing this plugin puts that dog in your window. If you want your own, replace the frames — see Using your own photographs.
Install
dsh plugin --profile web add dsh-labrador
Then restart the harness. She appears in the bottom-right corner of the window.
What she does
- Decides for herself. A state machine moves between seven phases — idle, acting, moving, sleeping, reacting, dragging and thrown — and draws from a weighted pool of poses. There are no transitions to a pose that is not reachable from the one she is in: a sleeping dog wakes to idle, never straight into a pose.
- Poses. Seventeen actions, twelve of which can stand on the floor: sitting, standing, looking up, lying on the sofa, asleep, sniffing, in the car, in a cone.
- Being handled. Click to poke her. Drag to move her, kept inside the window. Throw her and she flies — gravity, a bounce off the walls and floor, friction on landing — and changes her face on each impact.
- Speaks dog.
woof. hff. woof woof! yip! grrr. and so on, in a bubble above her head. Vocalisations only; there are no words and so nothing to translate.
- Right-click for a menu: pet him, play any action at once, or send her back to his corner.
- Reacts to the pointer only on his own body, not on the transparent margin around her.
Settings
Settings → 拉布拉多 / Labrador
Temperament (calm, attentive, lively), size, opacity, and every action with its own weight and a shown/hidden switch. Saving writes to $DSH_HOME/dsh-labrador/main-config.json atomically, and the dog follows without a reload.
{
"temperament": "attentive",
"size": 1,
"opacity": 1,
"weights": { "sleep": 2 },
"disabledActions": []
}
An illegal value is reported in the log and falls back to the documented default; nothing is silently ignored.
How it is built
lib/index.js — the host half: builds the action library and serves the frames and the settings over one HTTP route.lib/client.js — the client half: the state machine, the floor geometry, the physics, and the UI. The machine is pure — no React, no DOM, no timers — and the tests drive it directly.lib/repertoire.js — the seventeen actions, as data. Tuning his behaviour means editing this, not the engine.lib/library.js — resolves the frames, warns loudly about anything broken, and never throws away a dog quietly.assets/frames/ — the normalised frames. Every one is the same 445×328 canvas with the dog at the same size, on the same floor line and the same centre, so changing pose cannot make him jump or resize.
Using your own photographs
Drop a folder at $DSH_HOME/dsh-labrador/actions/. Each folder inside is an action; the images in it are its frames; an optional action.json overrides the defaults:
// $DSH_HOME/dsh-labrador/actions/sit/action.json
{
"category": "ambient", // idle | ambient | move | click | event
"weight": 20,
"loop": "hold",
"bust": false // true if head-and-shoulders: it cannot stand on the floor
}
Frames are served straight from that folder, so nothing needs rebuilding.
Tests
npm test
Forty-four tests of the state machine and the physics, six of the configuration, and thirty-six smoke checks over both halves. They run by being executed directly rather than through node --test, because that runner spawns a child process per file and is blocked in some sandboxed environments.
Known limitations
- Movement is disabled. It needs a walk action with several frames to show a stride, and there is only one walking photograph, so she stays put rather than sliding sideways. A load-time warning says so.
- One frame is a fragment. The
sniff pose lost its head when it was cut out, so it looks like a piece of dog. It is included anyway, at low weight.
- The pet lives in the harness window only. There is no detached always-on-top desktop window.
License
MIT
