deepseek-pet

One-line Positioning

Injects a Live2D-style desktop pet character into the DeepSeek Harness Web: it automatically switches expressions and speech bubbles based on current tasks, tool calls, context usage, concurrent sessions, and idle time; supports drag-to-move, scaling, and "action image" customization.

Core Features

• Renders a complete character image in the bottom-right corner of the webpage (not split into parts to avoid expression misalignment), switching to corresponding expressions for scenarios like thinking, answering, coding, web search, sub-agent, completion, and failure
• A speech bubble above the character cycles through status short phrases, with the latest response/thinking output displayed in a single-line typewriter effect; the bubble auto-hides after 10 seconds of inactivity
• Displays the focused session (with highlighted border) and up to 7 concurrent executing sessions below the character; automatically stacks and collapses beyond 3 sessions, shows "+N sessions" when exceeding 7
• When context reaches 62%, switches to "还可以再吃一点" (eating series); at 82%, switches to "上下文吃饱了"; switches to blindfolded "看不见" state when image input is detected
• Shows morning/afternoon/evening/night greetings based on local time; idle 10 minutes shows "肚子饿了", 30 minutes "抱着枕头犯困", 1 hour "已经睡着了"
• Drag to reposition, hover + scroll to scale (65%-140%), single-click triggers status phrase, double-click or "−" button minimizes to static icon in bottom-right corner; position and size are automatically remembered
• Settings panel adds "Pet Settings" page: can switch between "Default" and "Page Always-on-Top" display modes, can enable/disable specific images by action group via checkboxes, at least one image per group must be retained
• Respects prefers-reduced-motion and narrow screen breakpoints (≤760px), all animations can be significantly reduced under system preferences

Technical Implementation

• Language: JavaScript / JSX (React 18)
• Key Dependencies: React 18 (peerDependency, injected by host), esbuild 0.25.8 (build-time only); runtime only depends on host-provided cordis / dsh-client-runtime / dsh-client-ui-layout / dsh-client-ui-slots
• Architecture Pattern: Dual-half cordis bundle: src/host/index.js is an empty host half (apply() has no side effects), all visible behavior is in the browser half of src/client/index.jsx: registers the pet component via ctx.slots.inject('shell.overlay', …) (order=90) + registers settings page via ctx.slots.inject('settings.section', …); subscribes to slots and sessions two runtime capabilities
• Entry Files: Client entry src/client/index.jsx, core component src/client/DeepSeekPet.jsx, settings page src/client/DeepSeekPetSettings.jsx; build output lib/index.js (host half) + lib/client.js (client half, ~1.5 MB, with all base64 WebP images embedded)
• Asset Pipeline: scripts/build_assets.py (Python 3 + Pillow processes source images → transparent WebP) + scripts/embed-assets.mjs (converts WebP to base64 and writes back to src/client/assets.generated.js, then esbuild bundles into lib/client.js); no network resources required

Use Cases

• Users running long model tasks (multi-turn reasoning, long code generation, batch tool calls) who want companion feedback instead of staring at progress bars;
• Users managing multiple concurrent sessions (>3) who need to visually see which is running, which is waiting for interaction, which is focused;
• Desktop users focused on accessibility experience who need prefers-reduced-motion automatic animation degradation.

Prerequisites & Compatibility

Installation

dsh plugin --profile web add github:keleus/deepseek-pet

Configuration Options

The pet's position, scaling, last activity time and other runtime states are automatically written to browser localStorage, no manual configuration needed; all "configuration options" above are modified through DSH settings panel's "Pet Settings" page.

FAQ

Q: Does the pet affect page performance?

A: Very little impact. All 28 expression images + 3 frame images are embedded as base64 WebP directly in lib/client.js (~1.5 MB), no external network requests after loading; runtime only renders the currently active image (others have opacity:0), DOM always exists but overhead is negligible. You can see a long string of <img> in browser DevTools but actual rendering is very light.

Q: Can I replace it with a completely different character?

A: Not directly. The plugin's 28 complete character images (not split) and state mappings are hardcoded in src/client/assets/, src/client/pet-state.js, src/client/pet-presentation.js, there's no "skin change" or "load external pet.json" entry in the source code; to customize the character you need to fork the repo, replace source images, and build/install locally.

Q: Why does the pet say "can't see" when I send images?

A: Not a bug. DeepSeek model doesn't support visual input, so when the plugin detects image attachment in recent human input, it actively switches the pet to blindfolded expression and shows "图片暂时看不见" as a visual cue for the user. See src/client/pet-state.js:81-91 for the source.

Q: Does "page-always-on-top" in settings open a new window?

A: No. Both display modes are pure in-page rendering: default renders in DSH's shell interface layer, page-always-on-top uses React createPortal to mount the pet to document.body and positions it in the viewport bottom-right with position:fixed. Neither mode calls window.open, all browsers behave consistently. See src/client/DeepSeekPet.jsx:482-486 + src/client/styles.js:29.

Q: Will the pet get confused with multiple concurrent sessions?

A: No. Focused sessions are marked with highlighted bars, executing sessions are listed below; when exceeding 3 concurrent executing sessions, it enters "busy crazy" state and alternates between "working" and "rice-eating" image groups; when exceeding 7, the session list shows "+N sessions" summary at the bottom.

Q: How to disable specific expression images (like avoiding "crying face")?

A: Open DSH settings panel → "Pet Settings" → scroll to "Action Images" section, uncheck in the corresponding action group; each group must retain at least one image; click "Restore defaults" to reset all action groups' image pools.

Q: Will my preferences remain after uninstalling the plugin?

A: localStorage preferences (position, scaling, last activity, display mode, enabled action images) are not cleared when the plugin is uninstalled and will reload immediately after reinstall; however, since these keys are prefixed with deepseek-pet:, if no deepseek-pet related code runs afterward, they become orphaned keys that don't affect other functionality.

Difficulty Level

Beginner — Install and use immediately, settings panel only has display mode + action image checkboxes as visible options; all runtime states (position/size/scrolling text) are enabled out of the box, no documentation reading required.

Known Issues & Limitations

• When detecting image input, it forcibly shows "图片暂时看不见" blindfolded state (DeepSeek model doesn't support visual input, not a plugin bug)
• runningSessions bottom panel only renders the first 7 concurrent executing sessions, excess shown as "+N sessions" summary; auto-stacks when exceeding 3 but doesn't affect total count
• Custom character images require forking the repo and replacing source images in src/client/assets/ then building locally, no runtime "load external pet.json" or skin switching entry
• Rebuilding requires Python 3 + Pillow (npm run assets), not a pure npm workflow; published versions have embedded images, regular users don't need this step
• All image resources are embedded as base64 WebP in lib/client.js, single file ~1.5 MB, initial load slightly slower than average plugins but no external image requests after loading

DeepSeek Pet 是一个嵌入 DeepSeek Harness 网页的交互式桌宠插件。它会跟随当前任务、 工具调用、上下文占用和活跃会话自动切换 DeepSeek 表情，并通过呼吸、弹跳、倾斜、 视差和淡入动画呈现 Live2D 风格效果。

角色使用完整表情图切换，不拆分头部、手脚或五官图层，避免部件错位和表情突变。

功能

• 角色上方用与 Pet 等宽的气泡轮播状态短句，并以单行横向打字机效果跟随最新输出；无任务活动时 10 秒后自动隐藏；
• 角色下方用横条显示聚焦会话并以亮边标记，正在执行的会话向下排列，超过三个时自动层叠收起；
• 多个会话并行执行时进入“忙疯了”状态；
• 根据分析、回答、编码、网络搜索、子 Agent、完成和失败等场景切换表情；回答阶段轮换敲电脑、思考和认真核对等工作形象，思考阶段轮换不同状态标签；
• 工具等待批准时，可直接在气泡中允许或拒绝；
• 用户问题支持选项、自定义输入、提交和拒绝；
• reasoning 中疑问较多时显示困惑或压力表情；
• 用户指出回答有误时显示道歉表情；
• 上下文达到 62% 时提示“还可以再吃一点”，达到 82% 时显示“吃饱了”；
• 图片输入时显示蒙眼状态；
• 无任务 10 分钟显示饿了，30 分钟抱枕犯困，1 小时后睡觉；
• 待机、思考、回复、成功、等待、错误和干饭等动作按累计时间稳定轮换，不会因重渲染随机跳图；
• 设置页可按动作组启用或停用具体图片，每组至少保留一张；思考和执行工具时会在工作与白米饭动作间稳定交替；
• 根据本地时间显示早上、中午、下午和晚上的问候，23 点后犯困，凌晨进入睡觉状态；
• 鼠标移入角色时，在角色脚下显示“−”最小化按钮；
• 支持拖动角色、单击互动、双击折叠；鼠标悬停角色时可用滚轮缩放，并记住位置和尺寸；
• 支持窄屏布局和 prefers-reduced-motion；
• 设置面板新增「桌宠设置」页（界面样式与页面设计系统一致），可配置动作图片，并提供两种展示模式：
  • 默认：保持当前展示方式不变；
  • 页面置顶：桌宠固定在视口右下角，悬浮在当前网页所有内容（包括弹窗）之上。

安装

安装需要 Node.js 和 pnpm。如果已经全局安装 dsh，运行：

dsh plugin --profile web add github:keleus/deepseek-pet

没有全局 dsh 命令时，无需额外安装 CLI，直接通过 npx 执行：

npx @deepseek-ai/dsh plugin --profile web add github:keleus/deepseek-pet

两种命令效果相同，都会从 GitHub 拉取插件并安装到 Web profile。

安装完成后启动或重新启动网页。全局 CLI：

dsh web

npx：

npx @deepseek-ai/dsh web

如果网页已经打开，请刷新页面。

拉取源码后安装

git clone https://github.com/keleus/deepseek-pet.git
cd deepseek-pet
npm install
npm run build
dsh plugin --profile web add .

如果没有全局 CLI，将最后一条命令替换为：

npx @deepseek-ai/dsh plugin --profile web add .

本地安装会链接当前目录，修改代码并重新构建后即可继续调试。

更新

重新执行安装命令即可拉取并安装远端最新版本：

dsh plugin --profile web add github:keleus/deepseek-pet

或者：

npx @deepseek-ai/dsh plugin --profile web add github:keleus/deepseek-pet

随后重新启动网页进程并刷新页面。

卸载

dsh plugin --profile web remove deepseek-pet

没有全局 CLI 时：

npx @deepseek-ai/dsh plugin --profile web remove deepseek-pet

使用

• 拖动角色：调整显示位置；
• 鼠标停在角色上滚动滚轮：在 65%～140% 范围内调整 Pet 尺寸；
• 单击角色：触发与当前状态相符的短句反馈；
• 双击角色或点击“−”按钮：最小化为右下角静止图标；
• 点击活跃会话：将该会话切换为聚焦会话；
• 出现批准或提问卡片时：可直接选择、输入、允许或拒绝。

开发

环境要求：Node.js 22.19 或更高版本，以及 Python 3 和 Pillow。

npm install
npm run build

• npm run assets：从源图生成透明 WebP 表情并嵌入客户端代码；
• npm run build：生成 lib/index.js 和 lib/client.js。

主要目录：

src/host/         插件 Host 入口
src/client/       网页入口、组件、状态逻辑和动画样式
public/assets/    表情源素材
scripts/          素材处理和构建脚本
lib/              可安装的构建产物

许可证

MIT © 2026 keleus