deepseek-harness-pets

English | 简体中文

A desktop pet for DeepSeek Harness (dsh) — a whale that swims in the corner of your Web UI and reacts to your agent's turns.

A persistent little whale floats in the bottom-right corner of the DSH Web UI (shell.overlay slot — never touches your chat stream). It bobs while idle, braces while loading, swims hard while inferring, reviews while scoring, waves when a turn completes, leaps on SOTA, and floats belly-up on failure. Click it, hover it, drag it — it's a real pet.

Pet skins follow the Petdex standard (pet.json + 8×9 spritesheet), so assets interop with the Petdex desktop overlay and the petdex.dev community gallery.

See it in action

Live captures from a real dsh web session (dsh 0.1.0-rc.5, zero console errors):

Idle — bobbing doze, zzZ	Hover — "huh?"	Click — happy hop + cheer	Dragged — position persists

Drag it anywhere inside the viewport — the position is clamped and survives reloads (localStorage). Full set: docs/screenshots/.

Install

Two commands (v0.2.2+ on npm):

dsh plugin --profile web add deepseek-harness-pets
dsh web

Open http://localhost:3080 — the whale is already there. Full guide (prerequisites, source checkout, hand-written patch row, troubleshooting): docs/installation.md.

State mapping

Three-layer contract: dsh event signal → plugin semantic state → overlay CSS animation + bubble. Machine-readable: mappings/harness-states.json.

dsh signal	State	Overlay visual / bubble
session idle	idle	bobbing doze / zzZ
turn/start first prep	loading	buoyant anticipation / Loading data…
step/start, tool activity	inferring	swimming sway / Inferring…
report/scoring tools	scoring	leaning review / Scoring…
turn/end completed	done	fin waving / Done!
completion with SOTA text	sota	leap + splash / SOTA!
turn/end failed	error	belly-up grayscale / [ERROR] check logs

Animations are driven by the OVERLAY_ANIMATIONS table in src/pet-art.js (CSS transform keyframes over the official DeepSeek whale SVG). The Petdex 9-row set remains the asset contract for the future spritesheet switch point.

Configuration

~/.dsh/settings.yaml (user layer) or the profile patch layer:

deepseek-harness-pets:
  pet: random   # deepseek-whale (default) | deepseek-octo | random (deterministic per turn id)

Skin discovery scans three sources, highest priority first: ~/.dsh/pets (custom) > ~/.petdex/pets (community) > bundled assets/pets. Details: docs/installation.md.

Creating a skin

Each pet is a standard Petdex package under assets/pets/<slug>/ — contribute one via PR (pet/<slug> branch), or keep it local in ~/.dsh/pets/ where it always wins. Frame spec: 8×9 grid, 192×208 per frame, magenta #FF00FF background, row order idle→review. See CONTRIBUTING.md and each pet's spritesheet.README.md.

Release

Published to npm by GitHub Actions via trusted publishing (OIDC, zero tokens): push a v* tag → CI verifies (tests, build, artifact-drift gate) → publishes with provenance. Maintainer runbook: CONTRIBUTING.md. Changelog: CHANGELOG.md.

Roadmap

• Full-pipeline verification against a real dsh checkout (install → composition → boot graph → client bundle → settings, 2026-08-15)
• Persistent floating overlay form (shell.overlay, v0.2)
• Overlay live verification (bottom-right residency, interactions, drag persistence, zero console errors, 2026-08-15)
• Published to npm via CI trusted publishing (v0.2.2, zero token; topics: dsh-plugin/pet/deepseek/petdex)
• Settings-panel pet picker row (settings.general.item slot)
• Pixel spritesheets (spritesheet.webp × 2, community-owned)
• Host static serving for custom/petdex skins (removes the single-SVG limitation)
• Host-side event family for precise eval-phase & SOTA detection
• npx petdex submit to the community gallery

Contributing & license

• Guide: CONTRIBUTING.md · Architecture evidence: docs/architecture/
• Code MIT (LICENSE); pixel assets © their contributors, per each pet.json.

Disclaimer: community-driven, unofficial, not affiliated with DeepSeek. "DeepSeek" is a registered trademark of DeepSeek AI, used here as thematic homage.

Package disambiguation: the official harness is the npm package @deepseek-ai/dsh (Node.js). The PyPI deepseek-harness is an unrelated third-party Python library.