ysyyhhh/dsh-pet

English | 中文

An optional desktop companion for DeepSeek Harness. It is compatible with Codex desktop-pet packages (pet.json plus an 8×9 or 8×11 sprite sheet) and can import pets directly from Petdex without installing the Petdex CLI.

The pet acts as an ambient agent-status indicator: it relaxes when DSH is idle, thinks while the model reasons, works while tools run, waits when input is needed, and celebrates or frowns when a turn finishes.

It is not a second chat UI, a task manager, or a full desktop app. It is a status indicator.

• Plugin-first: a normal deepseek-harness plugin — no separately launched daemon, browser, or desktop app.
• Built-in Petdex import: /pet import <slug> downloads an approved pet from the official Petdex API; normal rendering is local and has no telemetry.
• Zero added LLM cost: event → state resolution is fully deterministic.
• Runtime-discovered pets: pets under assets/pets/ are discovered at startup, so adding a pet is dropping a folder in — no rebuild.

---

Installation

The plugin is a Cordis bundle that ships both a host half (the pet window) and a client half (the settings card). dsh plugin add installs it and, because the manifest declares dsh.bundle, adds it to the profile's bundle list automatically.

From a local directory

Install the plugin directly from its source directory (the profile keeps it as a link: dependency):

dsh plugin --profile <name> add /path/to/dsh-pet

On Windows, for example:

dsh plugin --profile web add D:/deepseek-pet

You can also run dsh plugin --profile <name> add . from inside the plugin directory.

From a tarball

Pack the plugin, then install the tarball:

npm pack
dsh plugin --profile <name> add /path/to/ysyyhhh-dsh-pet-0.2.0.tgz

From a Git repository

dsh plugin --profile <name> add github:ysyyhhh/dsh-pet

If the repository does not commit build artifacts, configure a prepare script so dsh plugin add builds the plugin during install.

Run

dsh --profile <name>

If you are running Harness from a source checkout, prefix the commands above with pnpm — i.e. run pnpm dsh plugin ... and pnpm dsh ... from the Harness repository.

The settings card requires Harness to expose the dsh-pet settings namespace (via its WEB_SETTINGS_NAMESPACES allowlist). The pet window itself does not depend on that allowlist.

Enable / disable

Set enabled: false in the plugin's config, or remove the bundle from the profile. The plugin then loads but shows nothing; removing it entirely leaves Harness fully functional.

---

Codex pet compatibility and Petdex import

The plugin reads the same Codex pet package shape used by Petdex: a pet.json manifest plus a WebP or PNG sprite sheet. Use the built-in commands from DSH:

/pet import boba     # download, validate, install, and select from petdex.dev
/pet list            # list bundled and imported pets
/pet use boba        # select an installed pet

Imported pets persist under ~/.dsh/dsh-pet/pets/, so plugin upgrades do not remove them. You can also copy any compatible Codex pet folder there manually.

See Adding a Pet for the full walkthrough, and the asset format reference for the exact pet.json and sprite-sheet layout.

---

Supported platforms

Platform	Status
Windows 11	✅ primary target (Win32 layered window via koffi)
Linux (X11 / XWayland)	✅ (XCB ARGB overlay; requires a compositor)
macOS	❌ not implemented (backend interface is reserved)

Linux per-pixel transparency needs a running compositor (GNOME/KDE ship one by default; lightweight WMs need picom or similar). On Wayland the overlay runs through XWayland.

---

Configuration

All fields are optional and validated with a Schemastery schema (invalid values fail loudly at load). The user-editable fields (enabled, petScale, petId, hideWhenIdle) are exposed on the Web settings card.

Field	Default	Description
enabled	true	Master switch.
alwaysOnTop	true	Keep the pet above other windows.
petScale	1	Pet size, 0.5–4× in 0.25 steps.
petId	text	Which pet to display (a directory name under assets/pets/).
hideWhenIdle	false	Automatically hide the pet when it sleeps (no task), and show it again on activity.
animationEnabled	true	Run the frame animation (static frame when false).
idleFrequencySec	20	Seconds (≥8) between randomized idle variations.
clickThrough	false	Pass pointer events through (Windows only).
startSleeping	false	Start in the sleeping state.
animationSpeed	1	Global speed multiplier (0.25–4).

Example:

- insert:
    - id: dsh-pet
      name: "@ysyyhhh/dsh-pet"
      config:
        petScale: 1
        petId: text
        idleFrequencySec: 30

Window position is persisted privately under ~/.dsh/dsh-pet/position.json (best-effort; failures are ignored). It does not depend on any Harness storage service.

Developer mode

When the core ctx.commands service is present, the plugin also accepts renderer-state commands without invoking an LLM:

/pet thinking
/pet working
/pet waiting_for_user
/pet success
/pet error
/pet reset

Valid states: STARTING IDLE THINKING WORKING CODING RUNNING_COMMAND WAITING_FOR_USER SUCCESS ERROR SLEEPING.

---

Architecture

harness events / lifecycle
        ↓  (the only harness-specific layer)
integration/  HarnessBridge · capability-detection · event-mapping
        ↓  NormalizedEvent
core/         PetStateResolver · PetStateMachine · TaskStateRegistry
        ↓  SemanticState
renderer/     AnimationController · PetWindow
        ↓  finished RGBA frames
renderer/backend/  Win32Backend · X11Backend   (native overlays via koffi)
        ↑
renderer/codex-pet/  PetContract · PetLoader   (pet.json + sprite sheet)

• HarnessBridge is the only module that knows raw harness event names. Everything above it is harness-independent.
• Pet core (core/) is a standalone library: testable with no harness, no window, no network.
• Backends are platform-isolated behind WindowBackend; the renderer never sees Win32 or X11 details.
• Client half (src/client/) is a separate browser bundle registered through the harness module loader; the host and client halves communicate through the settings namespace.

Harness dependencies

Only the Cordis plugin lifecycle and these core services/events are used:

• Plugin entry: apply(ctx, config) + name / inject / Config.
• Lifecycle: ctx.effect(), ctx.on(), ctx.logger(name).
• Activity observation: session/event, agent/status.
• Settings: the dsh-pet settings namespace (host-side), bound by the client card.
• Optional (detected, not required): ctx.agents, ctx.sessions, ctx.approval, ctx.commands.

No non-core plugin is required. If an optional service is absent, the pet degrades gracefully (coarser states, no /pet command).

External dependencies

Package	Purpose	Runtime
koffi	Win32 + X11 FFI for the overlay window	Node ≥22
sharp	Decode WebP/PNG sprite sheets to RGBA	Node ≥22
@deepseek-ai/schemastery	Config schema validation	Node ≥22
clsx	Class-name helper for the client card (inlined into the browser bundle)	build

Peer (type-only, not bundled): @deepseek-ai/cordis.

The client bundle's react and @deepseek-ai/dsh-client-* imports are externalized: they are provided at runtime by the harness module loader, so the plugin does not ship them as runtime dependencies (they appear only as dev dependencies for type checking and bundling).

Explicitly avoided: Electron, Tauri, WebView2/webview, GLFW/SDL/raylib, game engines, GPU/OpenGL, Docker, databases, Redis, any external server, browser automation.

Event → state mapping

Normalized event (from harness)	Pet state (semantic → animation)
startup	STARTING → waving
idle (agent/status: idle)	IDLE → idle
assistant/chunk (text/reasoning/tool-call delta)	THINKING → running
tool/call (editing tools)	CODING → running
tool/call (shell/command tools)	RUNNING_COMMAND → running
tool/call (other)	WORKING → running
approval/asked / waiting	WAITING_FOR_USER → waiting
turn/end reason completed	SUCCESS → review
turn/end reason error/aborted	ERROR → failed
long quiet period	SLEEPING → idle

SUCCESS / ERROR / STARTING are transient (default 2s) then return to IDLE. Concurrent agents are tracked per session/task and folded by priority WAITING_FOR_USER > ERROR > WORKING > THINKING > SUCCESS > IDLE.

Extending

• Adding a pet — see Adding a Pet; no code change is required.
• Adding an animation state — extend SemanticState in src/core/types.ts, its resolver mapping in src/core/PetStateResolver.ts, and its renderer pose in SEMANTIC_TO_CODEX.
• Adding a window backend — implement WindowBackend (src/renderer/backend/WindowBackend.ts) and register it in src/renderer/backend/selectBackend.ts.

---

Testing

npm test               # vitest unit tests (core + loader + integration)
npm run typecheck      # tsc --noEmit (host half)
npm run typecheck:client # tsc -p tsconfig.client.json --noEmit (client half)
npm run build          # tsdown bundle (host + client)
npm run gen:assets     # regenerate the bundled text pet

The pet core is tested without a harness or a display. The native overlay backends require a real desktop session and are not exercised by the headless test suite — they need manual verification on Windows/Linux.

---

Known limitations

• Linux transparency requires a compositor; on Wayland the pet runs as an XWayland client (no native wlr-layer-shell).
• macOS is not implemented.
• The bundled placeholder is the text test pet only — original SVG-drawn text, with no OpenAI/Codex/DeepSeek character artwork or trademarks.
• Native window rendering (frameless/transparent/topmost/drag) has not been exercised by automated CI and needs a manual check on a real desktop.

---

License

MIT.