---

English

Turn a Harness conversation into a reusable AI companion: choose a persona, retain useful memories, and let it start a brief conversation after a period of inactivity.

Preview release. Targets Harness Desktop 0.2.1-rc.3. Offline tests and Windows Desktop installation/loading checks pass; real-model proactive chat remains unverified. This is an independent community project, not an official DeepSeek product.

What is new in v0.3.0?

Enable once in a chat; resume after restart. Saved opt-ins restore when you reopen the original session, with the same settings and unanswered-message count. New chats stay off; assist_idle_stop forgets the saved opt-in. The backend must be running.

Download the plugin · Install and enable · Send trial feedback

We are looking for early testers on Windows + Harness Desktop 0.2.1-rc.3. Tell us where installation, activation, restart recovery, or stopping worked or failed. The one-minute trial helps collect reproducible results; model calls use your normal quota. No need to share private conversations or memory files.

What it does

Capability	Behavior
Portable personas	Versioned JSON packs with identity, voice, greeting, and memory guidance
Persistent memory	Long-term notes shared within a persona; short-term checkpoints scoped to a session
Proactive chat	Opt-in idle triggers with cooldowns, quiet hours, and an unanswered-message limit
Shareable packs	Public exports omit stored memories; private backups require an explicit option
Inspectable data	Local JSON storage, CLI tools, and documented formats

Meet the default companion

Rice-Loving Chubby Fish Daughter — 爱吃白饭的大肥鱼女儿 — is a playful fictional fish who loves a bowl of plain steamed rice. She calls you “爸爸” (“Dad”) by default; you can change that address at any time. The bundled persona speaks Chinese by default; ask for another language if preferred.

“Dad, your chubby fish daughter is here! I'd love a warm bowl of rice today. Shall we chat, or is there something I can help you with?”

This is an example of the persona's voice, not a fixed response. Fresh storage selects this persona automatically. Existing installations retain their saved roles and memories.

Role ID: little-fish · Definition: packs/fish.json

Download

Get the files from the v0.3.0 release.

File	Use
dsh-assist-0.3.0.tgz	Desktop plugin installer — do not extract
dsh-assist-0.3.0-source.zip	Complete source for development
SHA-256 manifest	Verify downloaded artifacts

Installation

The commands below use Windows PowerShell. Desktop supplies its own runtime; a separate system Node.js/npm installation is not required for plugin installation.

1. Launch Harness Desktop once to initialize its profile. Use Manage dsh Command… in the application menu to install or repair its dsh command.
2. Fully quit Desktop. Closing the window may only hide it.
3. Open a new PowerShell terminal. Verify that dsh comes from Desktop and reports 0.2.1-rc.3, then install the downloaded package:

Get-Command dsh
dsh --version

$package = Join-Path $env:USERPROFILE 'Downloads\dsh-assist-0.3.0.tgz'
dsh plugin --profile desktop add $package
dsh plugin --profile desktop list

Adjust the path if your download is elsewhere. Stop if a command fails. Reopen Desktop after installation.

An npm-installed standalone dsh cannot modify Desktop's profile. See the installation and troubleshooting guide for source-folder installation and existing-user migration.

Enable proactive chat

In the target conversation, ask:

Use assist_list, select little-fish with assist_use, and read assist_memory_read. Then call assist_idle_start: chat after 15 idle minutes, wait at least 30 minutes between triggers, pause after two unanswered triggers, and stay quiet from 23:00 to 08:00.

Enabled means the tool returned enabled: true, not merely that the model said it would remind you.

Action	Tool
Enable or change settings	assist_idle_start
Inspect this conversation	assist_idle_status
Stop future triggers	assist_idle_stop

Quiet hours use the backend's local timezone. The backend must remain running; saved opt-ins resume when the original session is loaded after restart, after a fresh idle/cooldown wait. New sessions stay off. Stop also forgets the saved opt-in. Each trigger uses the selected model and incurs normal usage. Stop does not retract a turn already submitted.

For a one-minute check, use idleMinutes: 1, cooldownMinutes: 1, maxUnanswered: 1, quietStart: 0, and quietEnd: 0. After the model finishes replying, stop typing. A trigger should occur after roughly 60–75 seconds; the response itself also depends on model latency. No second trigger should occur until you reply. Then call assist_idle_stop.

See idle-chat settings for full behavior. Reading SKILL.md in TRAE or another tool does not, by itself, supply a running Harness timer. This package does not connect to QQ, WeChat, or X.

Data and development

Memory lives in DSH_ASSIST_HOME, otherwise DSH_HOME/dsh-assist, or ~/.dsh/dsh-assist when neither variable is set. Public export filters stored memory fields; it does not redact personal information written into the persona itself.

Standalone CLI and development require Node.js 20.18+. Core and CLI need no third-party runtime dependencies.

node src/cli.js init --home "$PWD\.work\demo"
npm test
npm run check
npm run release

The release script builds local artifacts only. It does not publish to GitHub or npm.

Documentation

This README is bilingual. The detailed guides are currently in Simplified Chinese.

Installation & troubleshooting · Proactive chat · Persona format · CLI reference · Verification scope · Changelog

---

简体中文

为 Harness 聊天加入可复用的人设、持续保存的记忆，以及无人回复时主动聊一句的能力。

当前为预发布版。 面向 Harness Desktop 0.2.1-rc.3，已通过离线验证和 Windows 桌面端安装、加载检查；真实模型主动聊天尚未实测。本项目是独立社区插件，并非 DeepSeek 官方产品。

v0.3.0 新增了什么？

在会话里开启一次，重启后继续陪聊。 重新打开原会话后恢复已保存的设置和连续无人回应次数；新会话默认关闭，assist_idle_stop 会取消保存的启用状态。应用后端需要保持运行。

下载安装包 · 安装与开启步骤 · 反馈试用结果

正在寻找使用 Windows + Harness Desktop 0.2.1-rc.3 的早期测试者：安装、首次开启、重启恢复、停止，在哪一步成功或遇到问题都可以反馈。一分钟试用步骤方便复现，模型调用会消耗正常用量；无需提供私人聊天和记忆文件。

功能概览

能力	说明
可移植人设	用版本化 JSON 封装身份、说话风格、开场白和记忆规则
持久记忆	长期记忆按角色共享，短期摘要按会话保存
主动聊天	显式开启后按空闲时间触发，支持冷却、安静时段和无人回应上限
角色分享	公开导出不含已存记忆，私人备份需要明确选择
可检查数据	本地 JSON 存储，配套 CLI 和格式文档

默认角色：爱吃白饭的大肥鱼女儿

她圆滚滚、俏皮、有自己的小主意，喜欢捧着小饭碗吃热乎乎的白饭。示例默认称呼你“爸爸”，可以随时要求更改称呼。

“爸爸，大肥鱼女儿来啦！今天也想吃一碗热乎乎的白饭。你想先聊聊天，还是让我陪你做点事情？”

这是虚构角色的示例语气，实际措辞由模型生成。全新存储首次初始化会默认选中她；已有角色与记忆不会被覆盖。

角色 ID： little-fish · 人设文件： packs/fish.json

下载与安装

在 v0.3.0 发布页 下载：

文件	用途
dsh-assist-0.3.0.tgz	普通用户安装插件，无需解压
dsh-assist-0.3.0-source.zip	完整开发源码
dsh-assist-0.3.0-manifest.json	SHA-256 校验清单

以下为 Windows PowerShell 步骤。桌面端自带运行环境，安装插件无需另行安装系统 Node.js/npm。

1. 先启动一次 Harness Desktop，完成 profile 初始化；在应用菜单的 Manage dsh Command… 中安装或修复桌面端命令。
2. 完全退出桌面端，只关闭窗口可能仍在后台运行。
3. 打开新 PowerShell，确认 dsh 来自桌面端且版本为 0.2.1-rc.3，再安装：

Get-Command dsh
dsh --version

$package = Join-Path $env:USERPROFILE 'Downloads\dsh-assist-0.3.0.tgz'
dsh plugin --profile desktop add $package
dsh plugin --profile desktop list

下载到其他位置时修改路径；命令报错时先处理错误。确认安装后重新打开应用。 不要使用全局 npm 安装的另一个 dsh 修改 desktop profile。

源码目录安装、已有用户切换人设和故障排查见 安装指南。

开启、查询与关闭主动聊天

在目标会话中发送：

请先调用 assist_list，用 assist_use 切换到 little-fish，再调用 assist_memory_read。 然后调用 assist_idle_start：空闲 15 分钟后主动聊天，两次至少间隔 30 分钟，连续两次没人回复就暂停，23:00–08:00 不打扰。

只有工具返回 enabled: true 才算开启，模型口头说“已开启”不代表计时器已经运行。

操作	工具
开启或修改设置	assist_idle_start
查询当前会话	assist_idle_status
停止未来触发	assist_idle_stop

安静时段按后端本地时区计算。应用后端必须保持运行；已启用的原会话重新加载后会恢复，重新等待空闲与冷却；新会话仍需首次启用。停止会同时删除保存的启用状态。每次触发产生正常模型用量，停止不会撤回已提交的回合。

一分钟体验可设为 idleMinutes: 1、cooldownMinutes: 1、maxUnanswered: 1、quietStart: 0、quietEnd: 0。等回复完成后停止输入，约 60–75 秒应触发一个模型回合，实际回复还取决于模型速度。你再次回复前不应有第二次触发；体验后调用 assist_idle_stop。

完整参数见 主动聊天说明。只让 TRAE 等工具读取 SKILL.md，不会自动获得 Harness 计时与唤醒能力。本包不连接 QQ、微信或 X。

数据与开发

记忆优先存放在 DSH_ASSIST_HOME，其次为 DSH_HOME/dsh-assist，均未设置时使用 ~/.dsh/dsh-assist。公开导出会过滤已存记忆，但不会自动脱敏人设正文中的个人信息。

独立 CLI 和开发需要 Node.js 20.18+，核心与 CLI 无第三方运行依赖：

node src/cli.js init --home "$PWD\.work\demo"
npm test
npm run check
npm run release

发布脚本仅生成本地产物，不自动上传 GitHub 或 npm。

文档导航

安装与排错 · 主动聊天 · 角色包格式 · CLI 命令 · 验证范围 · 更新日志

---

Released under the MIT License. · 本项目采用 MIT 许可。