LingHXiao/dsh-effort-mascot

English | 中文

DSH ships reasoning effort as a dropdown where all four options look identical. This replaces it with a card, and the card has a fish on it.

The fish is the effort level: dozing because it hasn't eaten (Off), shovelling rice (Low), fuelled up and beaming (High), and — having overdone it — carb-comatose (Ultra). The card's palette follows along: grey → blue → purple → a gold ring.

To be clear, the artwork is only how it is drawn. The four levels still map one-to-one onto reasoningEffort — off / low / high / max. Pick a level and the model thinks at that level; nothing about that changed. Only the way it looks did.

One mapping worth stating: the Ultra label is DeepSeek's max level. It only gets a friendlier name and a visual identity; the request still carries reasoningEffort: "max". The other three keep the adapter's own names (Off / Low / High).

Nothing underneath changed: it reuses the official model directory and session.selectModel — the same control, drawn differently.

The card is always dark and deliberately does not follow the light/dark theme. It reads as a dark material floating above the composer, which keeps the illustrations and the gold highlights working in both themes.

Dragging through the levels, landing on Ultra (which opens with a one-off notice):

The GIF is capped at 15fps so it can autoplay inline. Full speed and audio: download the video.

Three things worth knowing

• Four levels, four faces — it is not just a colour swap: the mascot itself changes (dozing → snacking → eureka → overload)
• Colour-only motion — rotating ring, brightness breathing, slider sheen, bolt glow; nothing moves position, nothing is dizzying
• Quiet by default — honours prefers-reduced-motion; with it on, everything freezes and the Ultra notice is skipped

Install

dsh plugin add dsh-effort-mascot

Restart dsh web afterwards (the Host half registers the illustration route at boot), then reload the page.

git clone https://github.com/LingHXiao/dsh-effort-mascot.git
dsh plugin --profile web add ./dsh-effort-mascot

This links the checkout into your profile, so local edits take effect after a restart — no republish needed.

Checking it worked: the reasoning-effort control at the bottom right of the composer should now be the card. If it looks unchanged the plugin did not register — check ~/.dsh/model-effort.log for inject errors and confirm the bundle is enabled in plugin_manager.

Compatibility

Tested with	DSH 0.2.1-rc.2 and 0.2.2-rc.1. CI checks weekly that the seat still exists on the next line, so a rename shows up as a red check instead of a silent no-register
Mechanism	shadows the shipped conversation.input.model slot at priority: -1
Failure mode	if that seat's contract changes, this plugin silently does not register — the official control stays, nothing breaks
Blast radius	none on the Host, the session log, or what the model sees: selection still goes through session.selectModel

Four levels: the rice storyline

Level	Colour	Art	In the story
Off	neutral grey, no fill	half-shut eyes, a drool, drifting Zzz	Hasn't eaten — drowsy, nothing on its mind
Low	blue, filled 1/3	hugging a rice bowl, cheeks full, eyes shut in bliss	Meal time — working through the rice, one bite at a time
High	purple, filled 2/3	starry eyes, finger raised, a light bulb, sparkles	Fed — topped up and energised, ready to think
Ultra (= DeepSeek's max)	gold + badge, blue→purple→gold	spiral eyes, drool, dizzy halo	Overfed — carb coma (yes, really)

Read top to bottom, the four levels are one continuous arc for the same fish: idle from hunger → eating → powered up → comatose. They are not four unrelated pictures.

At the top level the card goes quiet for three seconds and says one thing: "maximum reasoning effort may use more tokens." Then it steps aside and the Ultra card fades back in — a heads-up, without a dialog to dismiss.

• Motion stays colour-only: rotating ring, brightness breathing, slider sheen, bolt glow
• Everything honours prefers-reduced-motion; with it on, the motion stops and the notice is skipped

Custom illustrations

Files live at assets/<effort-id>.webp — 360×360 transparent canvas, subject height normalised so the four characters read at the same size. The Host half serves them from GET /model-effort/asset/<name> (a Client bundle cannot import an image).

• Replace a file with the same name → reload the page (the asset cache invalidates on mtime + size).
• Add art for another level → drop the file and add one line to TIER_IMAGES in client.js.

The bundled art is AI-generated sample material, replace it with whatever you like.

Remove

plugin_manager remove_bundle target=dsh-effort-mascot

The shipped control comes back immediately; nothing is left behind.

How is this different from the other effort-slider plugins?

DSH ships its reasoning effort as a dropdown, and a few community plugins have replaced it with a slider. Each picked a direction:

Project	The direction it picked
HanaAyane/dsh-reasoning-effort	Keeps the native model picker, drags continuously inside the effort panel, and shows intensity with a running-fish animation
Microqian2th/dsh-codex-effort-slider	Codex-style continuous track with a stardust gradient at high levels
yuhub233/dsh-effort-slider	Sci-fi energy bar, an extra Ultra tier, plus a "lightning mode" that hands work to sub-agents
2768651338/dsh-effort-slider	A restrained Claude-Code-style slider

This plugin does not chase continuous drag feel, and has no particles or sub-agent wiring. It does one different thing: the four levels tell a single story — the same fish, from dozing off hungry, through a bowl of rice, to carb coma.

So: if you care about continuous feel, particles or agent orchestration, those are a better fit. If you want to see at a glance what state the current level represents, this one might be it.

The names "fat fish" and "Ultra" also appear in other projects; this one merely reuses them and claims no exclusivity.

Known limitations

• The trigger does not use the shipped conversation.input.model.label slot, so the 1017 badge contributed by ui-deepseek-model-label-1017 is not shown on this button.
• The in-card model list has no search (fine for a couple of models; use /model for the full picker).
• Diagnostics: only render failures are written to ~/.dsh/model-effort.log (capped at 1 MB, then rotated to .log.1). Which model or level you picked is never written by default; for troubleshooting, set globalThis.__MEC_TRACE = true in the console to also record state changes.

License

Code: MIT. Illustrations: AI-generated samples, not claimed as original works.