agent-pet

DSH PLUGIN通知与集成

仓库 rpegorov/agent-pet

通知与集成 L4 · 沙盒安装通过 安全审查:低风险 可信度:银 可安装插件
安装这个插件

agent-pet 是 rpegorov 维护的 通知与集成 类 DSH 可安装插件, 仓库为 rpegorov/agent-pet,许可证 MIT。 下面按「装什么、怎么装、验没验过、跑起来要注意什么」逐项列出。

一句话选型结论

按资料字段整理

agent-pet 是通知与集成类的 DSH 可安装插件,验证等级 L4 · 沙盒安装通过,安全审查为低风险,有可直接复制的安装命令。验证通过不等于安全:插件会读写文件、执行命令或联网,装前先到仓库确认它做了什么;涉及密钥或系统权限的,建议先在隔离环境试装。

分类通知与集成 收录类型可安装插件 验证等级L4 · 沙盒安装通过 安全审查低风险
验证等级 L4 L4 · 沙盒安装通过
GitHub 星标 0 复刻 未标注 · 未关闭 issue 未标注
最近更新 2026-10-17 创建于 未标注
README 7.7 KB 正文体量,越大说明说明书写得越细

安装

可安装插件
在 DSH 里执行,装完用 dsh plugin list 确认已挂上;标了 L4 / L5 的命令已在沙盒里真实跑通
dsh plugin --profile web add github:rpegorov/agent-pet
  1. 1
    确认版本与运行环境

    仓库没有标注兼容版本与 Node.js 要求,先确认你的 DSH 版本能装,再继续。

  2. 2
    执行安装命令

    在 DSH 里执行上面的命令;多包仓库要注意区分具体子包名,别装错包。

  3. 3
    确认已挂载

    用 dsh plugin list 查看列表里是否出现这个插件,出现即挂载成功。

  4. 4
    试跑并检查权限

    加载后跑一个最小用例,确认它能正常工作;同时看它申请了哪些文件、命令或网络权限。

兼容 DSH 版本未标注版本不匹配时安装会报错
Node.js 要求未标注来自仓库 package.json 的 engines 字段
前置依赖无需要同时安装的同级依赖
npm 包名未发布到 npm可用 npm view 查看包信息

能做什么

通知与集成
字段来自项目仓库与公开索引,中文说明由本站标注

适用场景

通知

验证等级与安全审查

L4 · 沙盒安装通过
自动化流程结论,不是人工审核
验证等级L4 · 沙盒安装通过安装命令在隔离沙盒内真实执行成功,无依赖错误
安全审查低风险静态分析结论,未标注不代表已通过审查
仓库健康度活跃按仓库活跃度与最近更新时间给出
可信度银综合仓库活跃度与验证结果

选择建议

装之前先看这几条
按条目字段给出的判断,不替你决定用不用

值得选的地方

  • 验证等级 L4:在沙盒里真实装过,依赖能装上,不是纸面声明。
  • 静态安全审查为低风险,没有发现明显可疑行为。
  • 有可直接复制的安装命令,不用自己翻文档找入口。

先想清楚的地方

  • 没有声明 Node.js 版本要求,版本不匹配可能装不上。
  • 验证通过不等于安全:插件会读写文件、执行命令或联网,装前先到仓库确认它做了什么。

更适合

  • 需要在通知与集成场景里给 DSH 补能力的人。
  • 想装完就能用、不想自己排依赖的人。
  • 能接受在本地装依赖、愿意自己确认权限的人。

不太适合

  • 只使用 DSH 官方内置能力、不接受第三方插件的人。

同类对比

6 个同类
同分类里挑最接近的几条,字段以各自详情页为准
插件验证等级安全审查星标兼容版本安装命令
agent-pet当前 L4 · 沙盒安装通过 低风险 0 未标注 有
treg L3 · 安装规范通过 低风险 4,976 未标注 无
dsh-purge 未做沙盒验证 未标注 3,412 未标注 有
dsh-im L5 · 运行验证通过 低风险 1,837 0.1.7-rc.2 等 10 个 有
dsh-dafeiyu L5 · 运行验证通过 低风险 425 0.1.7-rc.2 等 10 个 有
dsh-trading L3 · 安装规范通过 低风险 235 未标注 无

仓库档案

通知与集成
字段来自项目仓库与公开索引,中文说明由本站标注
作者 / 维护方rpegorov仓库 agent-pet · 默认分支 未标注
分类通知与集成notify
收录类型可安装插件声明了 dsh.bundle、可用 dsh plugin add 安装的条目
开发语言未标注仓库主语言
许可证MIT商用或二次分发前请确认条款原文
星标 / 复刻 / issue0 / 未标注 / 未标注收录时的数值,非实时
创建 / 更新未标注 · 2026-10-17最后推送 未标注
本站收录2026-10-30首次进入本站目录的时间

README 全文

7.7 KB
作者原文,中文版优先;只做安全转义,不改内容

agent-pet

A floating desktop pet for coding agents. One character per active agent (Claude Code, Codex, DeepSeek Harness, or anything that can run a shell command); her mood follows the agent, and when an agent asks for permission the pet shows the prompt in a bubble you can answer without switching windows.

Pets appear only while a session is working, waiting for you, or just finished. Everything is local: agents append JSON lines to a file and the overlay tails it. Nothing is sent over the network and there are no accounts; the optional DSH plugin opens a loopback-only port (127.0.0.1) for the overlay's open-session commands.

Requirements

  • macOS 14 or later
  • Swift 6 toolchain (Xcode or the Command Line Tools)
  • Python 3 (ships with macOS; used by the installer)

ChatGPT (the chat app) has no local plugin API, so there is no integration with it. Codex running inside ChatGPT.app is covered by the Codex plugin below.

Install

git clone https://github.com/rpegorov/agent-pet.git
cd agent-pet
scripts/install.sh                  # build, install to ~/.local/agent-pet, hook up Claude Code and Codex
scripts/install.sh --launch-agent   # also start the overlay at login

install.sh also accepts --no-claude-hooks and --no-codex-hooks (use these with the plugins below), and --uninstall.

Connect your agents

Use either the plugin or the install.sh hooks for a given agent, not both. Both together would deliver every event twice and ask every permission twice; the emitter does not de-duplicate because it cannot tell which hook fired first. Plugins carry only the hook wiring, so run scripts/install.sh --no-claude-hooks --no-codex-hooks once for the binary and the overlay.

Claude Code. install.sh adds hooks to ~/.claude/settings.json. Or, as a plugin:

/plugin marketplace add rpegorov/agent-pet
/plugin install agent-pet@agent-pet

Codex. install.sh adds hooks to ~/.codex/hooks.json and trusts them in ~/.codex/config.toml. Or, as a plugin:

codex plugin marketplace add rpegorov/agent-pet
codex plugin add codex-agent-pet@agent-pet

Codex does not trust the plugin's hooks on a command-line install, so they stay silent until you approve them (see Troubleshooting). The install.sh route has no such step.

DeepSeek Harness (DSH). plugins/dsh-agent-pet is a standalone DSH plugin. In your profile's package.json add a dependency "dsh-agent-pet": "link:/path/to/agent-pet/plugins/dsh-agent-pet" and add "dsh-agent-pet" to dsh.profile.bundles, install, and restart DSH. It reports session state and lets the overlay open a session or start a chat in the Harness UI.

Anything else (scripts, task runners, other agents):

agent-pet-emit send --agent glm --session task-1 --label "refactor" --mood working
agent-pet-emit send --agent glm --session task-1 --mood complete

Event protocol

agent-pet-emit appends one line per event to ~/.agent-pet/events.jsonl (override with AGENT_PET_LOG); the overlay tails the file. The file is rotated to events.jsonl.1 above 512 KiB.

{"v":1,"agent":"claude","session":"5e9d","mood":"working","label":"Bash","ts":1792276251.42}

mood is idle, working, attention, complete or ended (removes the session). label is at most 40 characters. Optional fields, omitted when unknown: cwd, app (bundle id of the hosting app), tty, detail (permission summary, at most 200 characters), permission (id of a pending request).

Decay: working with no event for 120 s is hidden, complete is hidden after 10 s, attention after 10 min, a session silent for 30 min is dropped. An agent's pet shows the highest-priority mood of its live sessions: attention, working, complete.

Permission bubble

With the overlay running, agent-pet-emit permit (the PermissionRequest hook) shows the pending tool call in a bubble with Deny, In agent and Allow, and a countdown.

  • The bubble shows the whole tool input (pretty-printed JSON, capped at 8 KiB). If the input is capped or unreadable, Allow is withheld.
  • In agent hands the prompt back to the agent's own dialog. If nobody answers before the countdown ends (at most 40 s, AGENT_PET_PERMIT_WAIT seconds, default 30), or the overlay stops running, the native dialog appears as usual. The pet never auto-allows.
  • Tools that ask you something (AskUserQuestion, ExitPlanMode) are never turned into a bubble.
  • Requests live in ~/.agent-pet/permissions/ (owner-only) and are removed when answered, expired or when the hook is killed.
  • The overlay's own texts (Deny, In agent, Allow, status labels) are English and not localised; labels that come from agents (tool names, session titles) appear as sent.
  • To turn the bubble off, quit the overlay (the hook then stays silent), or change the PermissionRequest command from permit to hook in the agent's hook config.

Privacy

agent-pet keeps everything on your Mac and sends nothing over the network.

  • What is written. ~/.agent-pet/events.jsonl holds one line per hook event: agent name, session id, mood, a short label (tool name, session title or notification text), the working directory, the hosting app's bundle id, the terminal tty and, for permission requests, a summary of the tool call (at most 200 characters). Prompts, tool output and file contents are not recorded.
  • Permission requests. While a request is pending, ~/.agent-pet/permissions/ holds its full tool input (at most 8 KiB). The file is deleted when the request is answered, expires or the hook is killed; leftovers older than 2 minutes are cleaned up.
  • Access. The directory is owner-only (0700) and the log is 0600. The log is rotated at 512 KiB, keeping one previous file.
  • What is read. The overlay reads the event log, and, to open a Claude Desktop session, the session index in ~/Library/Application Support/Claude/claude-code-sessions (read only). Selecting a Terminal or iTerm2 tab uses AppleScript and asks macOS for Automation permission once.
  • System log. Clicks and jumps are logged under the subsystem agent-pet with agent, session id, app and the opened URL; no tool input or paths.
  • Network. None outbound. The DSH plugin listens on a loopback-only port and accepts commands only from local processes and the Harness web UI; the overlay talks to it over 127.0.0.1.

Delete ~/.agent-pet to remove all recorded data.

Uninstall

scripts/install.sh --uninstall

Removes the binaries, hooks, trust entries and LaunchAgent. A no-op agent-pet-emit stub stays behind so sessions that are still open do not report a missing command; restart them and delete the stub. Plugins are removed with /plugin uninstall, codex plugin remove or by editing the DSH profile.

Troubleshooting

  • Nothing appears. Is the overlay running (pgrep -x AgentPet)? Does ~/.agent-pet/events.jsonl grow when an agent works? If not, the hooks are not installed or the agent was started before they were; restart the agent.
  • Codex plugin hooks do nothing. Codex 0.163 does not trust hooks from a plugin installed with codex plugin add; they ran only with --dangerously-bypass-hook-trust. Approve them in Codex (/hooks) or use scripts/install.sh, which writes the trust entries itself. Re-run install.sh after changing a hook command: the trust is tied to its text.
  • DSH changes are not picked up. DSH caches plugin ES modules; restart it after installing or editing a plugin.
  • DSH commands are refused. The plugin's bridge answers only the Harness web UI origin (http://127.0.0.1:50306) and local processes; a UI served from another origin cannot read its queue.

License

MIT. The sprite art in Sources/AgentPet/Resources is included under the same license.

常见问题

5 问
答案由本站依据条目字段整理,不引用上游原文
agent-pet 怎么安装?

在 DSH 里执行 `dsh plugin --profile web add github:rpegorov/agent-pet` 即可。

agent-pet 的验证等级是什么意思?

L4 · 沙盒安装通过。安装命令在隔离沙盒内真实执行成功,无依赖错误

agent-pet 安全吗?

低风险。

agent-pet 装之前要先确认什么?

先看它会读写哪些文件、执行哪些命令、连接哪些网络地址;验证等级只说明装得上、跑得起来,不等于安全。涉及密钥或系统权限的,建议先在隔离环境试装。

agent-pet 用的是什么许可证?

上游标注为 MIT。商用或二次分发前请到仓库确认许可证原文。

通知与集成
按星标排序,可作替代方案

treg

treg

Treg 是面向 Agent 工具的 OpenRouter,聚合 60 个提供商的 2896 个端点,按调用计费,统一令牌访问,无需注册各供

L3 · 安装规范通过 4,976 星

dsh-purge

dsh-purge

dsh-purge Community · DeepSeek Harness 社区维护版:宿主兼容、状态诊断、提示词管理、更新与本机备份回滚

未标注验证等级 3,412 星

dsh-im

dsh-im

将IM机器人接入DeepSeek Harness,支持飞书、微信、钉钉等九种渠道,通过扫码或凭据配置,统一管理多机器人连接。

L5 · 运行验证通过 1,837 星

dsh-dafeiyu

dsh-dafeiyu

DSH 大肥鱼是一款由 DeepSeek Harness 真实工作状态驱动的桌面伴侣插件,以始终置顶的透明窗口实时展示 Agent 的思考、

L5 · 运行验证通过 425 星

dsh-trading

dsh-trading

L3 · 安装规范通过 235 星

rpegorov 的其它插件

5 个

插件的元信息来自作者仓库;分类、验证等级与安全结论的中文说明为中文整理,最近核验 —。