dsh-qq-group

🐋 这是「DSH Whale 系」的一员

一套让陪伴角色真的在过日子的 DSH 插件。五个仓库各自独立，装哪个都行，但装在一起才是一整只：

仓库	管什么
dsh-whale-world	🧠 世界：跨会话记忆、生活作息、世界控制台。其他几个的记忆最后都落到这里
dsh-whale-qq	💬 QQ：群聊/私聊接话、主动搭话、语音、唱歌、表情包、日记、看图
dsh-whale-wechat	💚 微信：私聊/群聊接话，会 @ 人、会引用、会发语音，后台发送不抢焦点
dsh-whale-bili	📺 刷 B 站：她自己开 B 站看片子、记下看了什么、偶尔想跟你说一句
dsh-whale-douyin	🎵 刷抖音：在本地 Edge 里刷抖音电脑版，按兴趣和真实观看行为更新口味

命名统一用 dsh-whale- 前缀，看到这个前缀就是一家的。npm 包名沿用旧 id（dsh-qq-group 这种），因为几个插件之间靠 id 互相认亲，改名会把联动弄断。

好用的语料

语料	干什么
b-corpus	开源的 galgame 文本库。想让她说话带那味儿，就拿去喂人格蒸馏、做语气样本

独立的 DeepSeek Harness QQ 群聊插件：有人设、有记忆、会判断该不该接话，支持主动/被动搭话、语音、唱歌、表情包、日记和视觉看图。

A standalone QQ group-chat plugin for DeepSeek Harness: persona, memory, value-based reply decisions, proactive/passive chat, voice, singing, stickers, diary and optional vision.

本项目从 dsh-whale-girl 的 QQ 模块拆分而来，数据、模型、人设与桌宠完全独立。 This project was split out of the QQ module in dsh-whale-girl; its data, model, persona and lifecycle are fully independent from the desktop pet.

中文

功能与特点

能力	说明
OneBot 11 / NapCat	通过 WebSocket 连接 NapCatQQ，支持自动重连、启动检查、自动拉起
人设与模型	多张人设卡；可按群绑定，一张卡可绑多个群，未绑定跟随全局；模型 provider/model、推理档位、备用模型可按群单独选
主动搭话	独立页：冷场阈值、概率/峰谷命中线、近远时机、冷却、每日上限、安静时段、试运行、主动语音概率
价值评分 v3	不是每条都回：@/叫名字/引用直接回；普通闲聊按关键词、成员、上下文、熟识度、好感度、夜间等算 0~100 分
冷场逻辑	任何人类消息都会重置接话去抖和主动开口倒计时；群里一直有人说话时她不插话
直接对话队列	她正在生成时收到的 @/引用会排队，不会丢；主动开口会给 @ 让位
群友画像	每群每人独立熟识度/好感度；主人数值全局统一且可编辑，不被互动或衰减改动
群聊开关	设置页顶部逐群开关；关闭后完全不接该群消息、不主动开口、不写互动
日记	每条回复记录上下文、评分理由、内心吐槽；thought 放在 JSON 第一项；支持筛选、导出、删除单条/按人删除
表情包	自动收藏、AI 打标签、按语境/关键词选图；自拍图已并入常用库，用“尾巴/自拍”标签即可被明确请求选中
语音 / 唱歌	Edge / CosyVoice2 / Kokoro / GPT-SoVITS 多后端切换；GPT-SoVITS 已接 v2ProPlus 与 7 种情绪参考音，RVC 变声、语气预设、试听、按群概率
七情绪	开心 / 生气 / 吃惊 / 中立 / 难过 / 厌恶 / 恐惧；按回复文字自动选参考音，也可手动指定 style
视觉看图	可看当前/引用图片，也支持转发聊天记录里的图片；嵌套转发最多展开两层；本地路径与文件头识别
回复引用	目标消息被后续消息隔开时自动带 QQ 引用，可开关；文件/合并转发/语音不引用，失败自动改发纯文字
省 token	自适应上下文（自包含短句少读、指代/追问读全量）、contextTurns、记忆注入条数、情绪/敌意判断开关
夜间免打扰	支持跨午夜；夜间不再主动开口，普通接话按命中扣分（0~100，默认 20）
心情	每群独立心情；会随互动升降、随时间回落，被辱骂/恶意攻击会下降
用量统计	按群、按用途、按模型统计 token；图片调用与估算调用单独标记；月历、余额估算、单价校准
稳定性	事件/定时器/异步回复全部兜异常；模型只回 reasoning 时自动提取/重试/备用模型

部署要求

DeepSeek Harness 桌面版或可加载插件的主机
Node.js 22.12+、pnpm 9+
NapCatQQ + QQ NT（Windows 推荐；Linux 可用 NapCat Docker，需自行确认 QQ 登录）
一个 QQ 小号作为机器人，已加入目标群
可选：视觉模型（识图）、edge-tts（语音，联网）、ffmpeg（部分音频转换）

部署步骤

1. 安装并登录 NapCat

安装 QQ NT 与 NapCatQQ；
用机器人 QQ 登录；
在 NapCat 中启用 OneBot 11，添加 WebSocket 服务器：
- 地址：ws://127.0.0.1:3001（默认）
- access token：自己生成一个强随机串
确认 NapCat 控制台显示 OneBot WebSocket 已启动。

2. 安装本插件

git clone https://github.com/example-user/dsh-qq-group.git
cd dsh-qq-group
pnpm install
pnpm run build

把目录 link 进 DSH profile（或使用 dsh plugin --profile desktop add <绝对路径>）：

// <DSH_HOME>/profiles/desktop/package.json
{
  "dependencies": {
    "dsh-qq-group": "link:/absolute/path/to/dsh-qq-group"
  },
  "dsh": {
    "profile": {
      "bundles": [
        // ... 其它插件
        "dsh-qq-group"
      ]
    }
  }
}

重启 DSH。设置页出现「QQ 群聊」即加载成功。

3. 配置连接

DSH 设置 →「QQ 群聊」→「连接」：

项	示例	说明
WebSocket 地址	`ws://127.0.0.1:3001`	NapCat 的 OneBot WS
Access Token	`<YOUR_TOKEN>`	与 NapCat 一致，不要提交到仓库
机器人自称	`助手`	群里被叫到的名字，可逗号分隔多个
主人 QQ	`<YOUR_QQ>`	用于好感初始值与主人识别，可留空
自动连接	开	DSH 启动后自动连
自动拉起 NapCat	按需	需要填写启动脚本绝对路径

4. 启动

设置页点「启动」，或在群里 @ 机器人测试。首次建议先点「体检」与「模型检查」。

配置参数推荐

以下是起步值，按群氛围调整：

参数	推荐	说明
搭话模式	主动	被动只回 @/叫名字；主动才会插话和冷场开口
评分阈值	55~65	越高越挑；太低会频繁接话
接话概率	0.15~0.35	分数达标后的二次掷骰
冷却	20~45 秒	两句之间的最小间隔；直接对话会跳过
普通 10 分钟上限	6~10	防刷屏；按群活跃度调
直接回复上限	20~40	@/引用/接着聊单独计数
去抖	6 秒	等一波人说完；任何新消息都会重置
夜间免打扰	23:00–08:00，扣 20	想更安静就调大，0 = 不扣
看图策略	自定义：@ + 问看法	最省 token；需要“所有图”再选 all
语音概率	0.1~0.3	太高会刷屏，且增加 TTS 延迟
表情包概率	0.15~0.3	配合自动收藏使用
备用模型	选一个同 provider 的可用模型	主模型空回复/限流时兜底

模型建议：

日常文字：deepseek-v4.1-flash 或你额度充足的模型；
需要识图：选支持 text + image 的模型（如 deepseek-v4.1-flash、step-5-preview、MiniMax 多模态型号）；
推理档位：群聊建议 low；太高会增加延迟和 token；
换模型后历史用量按 provider/model 分开统计，余额是全局钱包。

设置页结构

人设 / 模型：人设、搭话模式、模型、推理档位、备用模型、用量与积分；
连接：OneBot、token、自称、主人、自动拉起、Cookie 清理提示；
聊天频率（按群）：阈值、关键词/成员概率、冷却、上限、连发、语音/唱歌/表情包概率、看图策略；
白名单 / 黑名单：一行一个 QQ；白名单抬评分下限，黑名单压上限；
夜间：时段与命中扣分；
语音：音色、语速、音调、音量、语气预设、试听；
表情包：自动收藏、数量/MB 上限、看图策略；
日记：筛选、导出、删除。

数据与隐私

路径	内容
`$DSH_HOME/dsh-qq-group/qq.json`	连接、群策略、语音/唱歌
`$DSH_HOME/dsh-qq-group/settings.json`	人设、模型、名单、夜间、看图
`$DSH_HOME/dsh-qq-group/qq-memory.json`	每群记忆（上限 200 条/群）
`$DSH_HOME/dsh-qq-group/profiles.json`	每群每人画像
`$DSH_HOME/dsh-qq-group/diary.json`	日记（上限 1000 条）
`$DSH_HOME/dsh-qq-group/stickers/`	本地表情包

仓库不包含任何聊天记忆、日记、QQ 号、token、Cookie 或用户路径；这些只存在于本机运行时。.gitignore 已排除数据文件。

接话逻辑（简版）

@ / 叫名字 / 引用她 → 100% 直接回；
其它消息在 6 秒去抖后算 0~100 分：
- 基础 30；关键词/指定成员概率；刚跟她聊过 +30；带“你” +15；
- 熟识度、好感度、@ 别人、被冷落、夜间扣分、单条消息；
- 白名单抬下限、黑名单压上限；
分数 ≥ 阈值 → 按 分数/100 掷骰 → LLM 闸门（读最近上下文判断是不是在跟她说）；
她说完后同一人接着说：高概率候选，但仍过闸门；
任何人类消息都重置冷场计时与主动开口倒计时。

视觉看图

「表情包」→「看图策略」：

关闭 / 自定义（多选）/ 所有图片；
自定义可同时勾选：@/叫名字/引用、强相关上下文、文字里问她看法；
只有真正回复且命中条件才下载并计费；纯图片、无文字、没 @ 她的消息不触发。

参考项目与致谢

本项目的设计参考了这些开源项目的优点：

dsh-pet：桌宠/插件工程结构；QQ 模块随后拆分成本项目；
zaofan-make/dsh-qqbot：价值评分、人设、群友画像、上下文压缩等思路；
Jiangsubei/dsh-napcat-bridge：主动搭话、夜间免打扰、别名、记忆等交互设计；
NapNeko/NapCatQQ：QQ NT 的 OneBot 实现；
OneBot 11：协议标准；
rany2/edge-tts：免费、无需显卡的语音合成；
sdfsfsk/matsuko_cover 与 dsh-cover-pipeline：唱歌/翻唱链路的灵感来源；
DeepSeek Harness：插件宿主、模型、附件与设置系统。

以上项目各自遵循其原许可证；本项目只借鉴设计思路，不复制其受限素材。

已知边界

NapCat/QQ 的登录与风控由 QQ 侧决定，插件不绕过；
语音/唱歌依赖 edge-tts 网络与本地音频工具；
视觉识别受所选模型能力限制，不支持图片的模型会退回文字；
群聊内容会发给所选模型 provider，请自行评估隐私；
请遵守 QQ 与群规，不要用于骚扰、刷屏或未授权自动化。

v1.20.0 GPT-SoVITS v2ProPlus 与七情绪

GPT-SoVITS 升级到 v2ProPlus，使用三月七角色权重。
网关支持 7 种情绪参考音，插件按回复文字自动选 style。
补齐 onnxruntime CPU 依赖；实测短句约 7.2 秒。

v1.19.7 自拍并入表情包与转发图片

自拍库不再单独维护：自拍/尾巴图并入常用表情包库，按文件名和标签关键词选中。
转发聊天记录支持图片：图片会进入视觉输入；嵌套转发最多展开两层。
新增逐群开关：关闭后完全不接该群消息、不主动开口。
夜间免打扰不再主动开口；心情会下降/回落。
主人熟识度/好感度全局统一、可编辑。

v1.18.4 日记格式与近期更新

v1.18.4：thought 移到 JSON 第一项，减少模型截断导致日记缺失内心吐槽。
v1.18.3：自动引用只对文字/图片生效，文件/合并转发/语音不再触发“该消息类型暂不支持查看”；引用发送失败自动改发纯文字。
v1.18.2：新增群聊入口置顶到插件设置顶部，独立于所有页签。
v1.18.1：本地 TTS 增加 2.5 秒健康检查，网关未启动时回退 Edge，不再卡住或静默失败。
v1.18.0：目标消息被后续消息隔开时自动带 QQ 引用，可开关。
v1.17.0：省 token 开关与自适应上下文；单价保存修复；情绪/敌意判断可关。
v1.16.0：主动搭话独立页；用量按群、按用途统计。
v1.15.x：人设独立分栏与按群绑定；恢复经典助手/贴吧人设；看图支持本地路径与文件头识别。

v1.14.0 GPT-SoVITS 本地后端

新增 GPT-SoVITS 后端，适配网关默认 50123，后端 API 9880。
网关把 POST /tts 转成 GPT-SoVITS GET /tts，返回 WAV。
v1 基础模型 CPU 短句约 2.4 秒；角色模型尚未训练。
参考音频目前是本地合成女声，后续可替换为助手角色模型。

v1.11.0 Kokoro 核显 TTS

新增 Kokoro-82M 本地 TTS 后端，ONNX Runtime DirectML，适配 780M 核显。
独立 Kokoro HTTP 网关（默认 50122）；短句实测约 6 秒。
语音后端列表新增 Kokoro；切换时自动填 50122。
Kokoro 支持中文内置音色（如 zf_005）；自由情感描述仍以 CosyVoice2 更合适。
IndexTTS-2.5 选项保留，服务仍未部署。

v1.10.0 本地 TTS 后端列表

语音页新增后端选择：Edge TTS / CosyVoice2 / IndexTTS-2.5。
新增本地 TTS HTTP 网关协议：POST /tts，JSON {text, style}，返回 WAV。
CosyVoice2 网关已实测；CPU 上 8.7 秒语音约 60 秒生成。
本地后端支持风格/情感描述；失败明确报错，Edge 仍是默认快速回退。
IndexTTS-2.5 选项已预留，模型服务尚未部署。

v1.9.3 明确语音/唱歌触发修复

修复“你唱首歌来听听”误提取歌名“歌来听听”导致唱歌失败、退回文字；无明确歌名时使用默认曲目。
修复“唱首小星星”这类没有“歌”字的点歌请求不被识别。
修复“说语音给我听听”这类语音请求不被识别。
明确要求语音/唱歌默认 100% 触发，可在语音页调低。
修正语音发送分支准入，普通语音概率为 0 时明确要求仍可发。

v1.9.2 RVC 试听临时目录

RVC worker 输出改到模型目录下的 temp，避开 Windows/沙箱子进程不可写的用户缓存目录。
worker 把 TEMP/TMP/TMPDIR 指向自己的可写目录，修复 ONNX 运行时报无可用临时目录。
试听失败会显示具体错误。

v1.9.1 试听与明确语音/唱歌

试听按钮会应用当前 RVC 设置（包括未保存的模型/设备/音调）；失败时播放原音并提示。
明确要求发语音/说句话/听听声音时，按「被要求发语音概率」触发（默认 95%，100% 必发）。
明确要求唱歌/点歌/来一首时，按「被要求唱歌概率」触发（默认 90%，100% 必唱）。
修复模型自己选唱歌时 singChance 未生效的问题。

v1.9.0 RVC 模型库与 780M

语音页自动扫描模型目录，列出 .pth/.onnx + index；不再手填长路径。
加入模型列表、打开目录、刷新和去重；可添加多个模型随时切换。
运行设备切换：自动 / DirectML（780M）/ CPU；ONNX 自动走 DirectML worker。
音调、推理超时、最长音频改为滑杆。
新增 ONNX worker + onnxruntime-directml，780M 实测可转换中文语音。

v1.8.0 RVC 外部 worker 接口

可选 RVC：edge-tts 干音 → 外部 Python worker → 变声 WAV → QQ 语音。
语音和唱歌共用 RvcConverter；失败自动回退原音。
配置 Python、worker 脚本、模型/index、设备 auto/cpu/dml、变调、超时。
780M 核显优先 DirectML worker；不追求实时。
未配置脚本/模型时不启用，不影响现有 edge-tts。

v1.7.1 引用图片与启动默认值

引用之前的图片时，通过 get_msg 读取被引用消息里的图片再交给视觉模型。
引用图片 + “这个是你吗/看看/评价”按直接对话处理。
autoStart 默认改为 false：重启 DSH 不自动启动 QQ 连接，需在运行状态点“启动”。

v1.7.0 按群心情 v1

每个群独立心情 0~100 + 七根情绪，基线默认 70。
直接对话后短模型判断一次，实际变化除以 3；12 小时半衰期。
可信度不足/变化不显著/每日额度满时不注入。
注入只作背景，不写成命令；不因被冒犯降心情。
接话页可看心情、情绪、最近理由并重置。

v1.6.0 主动搭话 v2

概率命中线：峰/谷基础线 × 时机因子 + 没中加成。
刚聊过更难开口，冷太久更容易开口；安静时段不掷。
试运行默认开，只记账不发送；每轮只挑一个群。
保留旧版固定退避模式，可切换。
接话页新增参数面板和未中次数/试运行状态。

v1.5.0 记忆 v3

新增 memory-v3 数据层：鲜活度、重要度、情绪、事实键、来源。
普通事件会淡忘，长期事实不衰减；同一 factKey 只保留最新。
按证据稀有度召回，最多 6 条 / 800 字 / 长期事实 2 条（可调）。
旧 qq-memory.json 保留并迁移；新增 memory-v3.json。
每 12 轮（可调）自动提炼 1~2 条记忆，压缩时也写入。
记忆页支持 v3 条目新增/编辑/删除、召回统计。

v1.4.0 助手自拍照与隐式看图

在「表情包」页内新增 助手自拍照 库，独立于普通表情包；支持上传/删除/打开文件夹。
“看看尾巴 / 看看你 / 爆个照”等在无附图时按自拍请求处理，按 per-group 自拍概率发送。
自拍概率 0~100%，100% = 有素材时必发。
自拍库为空时只按正常文字聊天含糊带过，不明说“我没有照片/图”。
有附图时“看看你/看看尾巴”优先识图，不会误发自己的自拍。
自拍触发词可在表情包页追加，按群生效。

v1.3.0 发表情包与唱歌概率

发图概率从「接话」移到「表情包」页靠前位置：
- 自动发图概率（模型想发时）；
- 被要求发图概率（群友明确说“发个表情包/来张图/斗图”时）；100% = 必发；
- 仍按仓库权重和上下文含义选图，inbox 的 0 权重不会绕过。
唱歌概率从「接话」移到「语音」页。
明确要图时按直接对话处理，不会被语音/唱歌分支抢走。

v1.2.2 日记修复

每条实际生成的回复都会写日记；模型没给 thought 时用占位内心记录，不再整条丢失。
兼容半损坏 JSON：从文本里抢救 thought 字段。
提示词强化“必须输出完整 JSON 且带 thought”。
日记页在打开时每 5 秒自动刷新。

v1.2.0 设置页排版

「用户画像」与「评分权重」改为分组卡片 + 自适应两列网格；
数字输入框固定 82px，右侧带单位，不再整行拉满；
按“关系语气 / 熟识度增长 / 好感度增长 / 敌意与衰减 / 初始好感 / 名单”分段；
说明卡片放在分组顶部，长文本与输入项不再挤在一起。

v0.2.2 用户画像

「名单」Tab 改名「用户画像」，熟识度/好感度相关设置全部集中到这里。
可调项目：首次/后续发言、被点名、她回复、接话、引用、敌意扣分、长期不互动衰减、主人/白名单初始好感。
每项都有独立每日上限，填 0 = 不限；支持小数。
增加关系语气说明卡片：明确各熟识度/好感度区间的语气差异。
关系评分加成上限也移到「用户画像」。

v0.2.1 新增

关系语气生效：熟识度/好感度现在会生成明确的语气指令（陌生→很熟、反感→亲近），不只是数字。
评分全参数可调：基础分（默认降到 15）、关键词/成员倍率、连续对话加成、@/引用/被冷落/单条/短消息扣分、关系加成上限、短消息字数。
明确看图请求：带图消息里说“看看/看一下/这张图/照片/这猫可爱吗”，即使没 @、没叫名字，也按直接对话读图。
语音概率：独立掷骰，不再依赖模型主动选 voice。

v0.2 新增

表情包多仓库：inbox / common / 自建仓库；每个仓库有发送权重（0 = 只手动用）、启用状态、默认入库设置；旧素材自动迁移到 common。
新图暂存与整理：新图默认进 inbox（权重 0），不会自动发出；可在 UI 里移动、删除、编辑含义，再放入常用仓库。
入库过滤：smart 模式跳过截图/大照片；支持全部保存；全局 SHA1 去重。
语音概率修复：回复语音概率改为独立掷骰，不再依赖模型主动选 voice；支持语音字数上限，TTS 失败自动回退文字。
上下文长度：在模型选择处按 provider/model 覆盖上下文窗口；按预算自动裁剪最近消息。
记忆 v2：可配置原始记忆条数/token 预算/摘要长度/压缩频率；压缩时自动抽取关键记忆（事实/偏好/关系/梗），原文可归档。
文件阅读：支持 txt/md/json/csv/代码、PDF 文本层、DOCX、XLSX；可配置触发条件、单文件 MB、最大字符数和每日上限。
合并转发：通过 OneBot get_forward_msg 展开转发记录；限制节点/字符数，嵌套只显示占位。

许可

English

Features

Feature	Details
OneBot 11 / NapCat	WebSocket connection to NapCatQQ; reconnect, startup checks, optional auto-launch
Persona & model	Multiple persona cards; bind per group (one card can serve many groups), otherwise follow global; provider/model, reasoning effort and fallback can be per group
Proactive chat	Dedicated tab: idle threshold, peak/off-peak hit lines, timing factors, cooldown, daily cap, quiet hours, trial mode, proactive voice chance
Value scoring v3	Not a reply to everything: @/name/quote always reply; ambient chat is scored 0–100 from keywords, members, context, familiarity, affinity and night penalty
Cold-start logic	Any human message resets both the reply debounce and proactive countdown; she stays quiet while the group is active
Direct-message queue	@/quote messages received while generating are queued, never dropped; proactive messages yield to direct ones
Member profiles	Per-group, per-user familiarity 0–100 and affinity -100–+100, updated by speech, mentions, replies, continuity and hostility
Diary	Every reply records context, score reasons and inner thought; `thought` is the first JSON field; filter, export, delete one/person
Stickers	Auto-collect group images, AI rename/tag, context-aware sending; count and disk caps
Voice / singing	Edge / CosyVoice2 / Kokoro / GPT-SoVITS backends, RVC conversion, tone presets, preview, per-group probability; offline singing and built-in scores
Vision	Optionally pass real QQ images to a vision model: @ / strong context / opinion questions; local paths and magic-byte detection supported
Reply quoting	Auto-quote when the target is separated by newer messages; toggleable; files/forwards/voice are skipped and failed quotes retry as plain text
Token saving	Adaptive context (fewer turns for self-contained messages, full context for references), contextTurns, memory injection count, mood/hostility toggles
Night DND	Cross-midnight window; configurable score penalty (0–100, default 20)
Usage stats	Token usage by group, purpose and model; image/estimated calls marked; calendar, balance estimate, rate calibration
Stability	Guards around events/timers/async replies; extracts JSON from reasoning, retries, optional fallback model

Requirements

DeepSeek Harness desktop or a host that can load plugins
Node.js 22.12+, pnpm 9+
NapCatQQ + QQ NT (Windows recommended; NapCat Docker on Linux is possible, confirm QQ login yourself)
A spare QQ account for the bot, already in the target group
Optional: a vision model, edge-tts (network), ffmpeg for some audio conversions

Deployment

1. Install and log in to NapCat

Install QQ NT and NapCatQQ.
Log in with the bot account.
Enable OneBot 11 and add a WebSocket server:
- URL: ws://127.0.0.1:3001 (default)
- Access token: generate a strong random string.
Confirm that NapCat reports the OneBot WebSocket as running.

2. Install this plugin

git clone https://github.com/example-user/dsh-qq-group.git
cd dsh-qq-group
pnpm install
pnpm run build

Link the directory into your DSH profile (or run dsh plugin --profile desktop add <absolute-path>):

// <DSH_HOME>/profiles/desktop/package.json
{
  "dependencies": {
    "dsh-qq-group": "link:/absolute/path/to/dsh-qq-group"
  },
  "dsh": {
    "profile": {
      "bundles": [
        // ... other plugins
        "dsh-qq-group"
      ]
    }
  }
}

Restart DSH. The "QQ Group" settings page appearing means the plugin loaded.

3. Configure the connection

DSH Settings → "QQ Group" → "Connection":

Field	Example	Notes
WebSocket URL	`ws://127.0.0.1:3001`	NapCat OneBot WS
Access token	`<YOUR_TOKEN>`	Must match NapCat; never commit it
Bot display name	`LittleWhale`	Names she answers to; comma-separated aliases allowed
Owner QQ	`<YOUR_QQ>`	Owner detection and initial affinity; may be empty
Auto-connect	on	Connect when DSH starts
Auto-launch NapCat	optional	Requires an absolute launcher path

4. Start

Click "Start" in settings, or @ the bot in a group. For first use, run "Health check" and "Model check".

Recommended settings

Starting points; tune to your group:

Setting	Recommendation	Notes
Chat mode	proactive	Passive only answers @/name; proactive also joins and starts cold chat
Score threshold	55–65	Higher is pickier; too low is noisy
Reply chance	0.15–0.35	Second roll after the score passes
Cooldown	20–45 s	Minimum gap; direct conversations bypass it
Normal 10-min cap	6–10	Anti-spam; tune to group activity
Direct 10-min cap	20–40	Separate counter for @/quote/continuation
Debounce	6 s	Wait for the group to settle; any new message resets it
Night DND	23:00–08:00, penalty 20	Increase for quieter nights; 0 disables the penalty
Vision policy	custom: @ + opinion questions	Cheapest useful default; choose all only if needed
Voice chance	0.1–0.3	Too high adds latency and spam
Sticker chance	0.15–0.3	Pair with auto-collect
Fallback model	A working model on the same provider	Covers empty replies/rate limits

Model tips:

Text: deepseek-v4.1-flash or another model with sufficient quota.
Vision: choose a model with text + image input (e.g. deepseek-v4.1-flash, step-5-preview, multimodal MiniMax models).
Reasoning: low is usually best for chat; higher effort adds latency and tokens.
Usage history is kept per provider/model; the balance is a global wallet estimate.

Settings layout

Persona / Model: persona, mode, model, reasoning effort, fallback, usage/credits.
Connection: OneBot URL, token, bot name, owner, auto-launch.
Chat frequency (per group): threshold, keyword/member chances, cooldown, caps, burst, voice/sing/sticker chances, vision policy.
Allow / block lists: one QQ per line; allow-list raises the score floor, block-list lowers the ceiling.
Night: window and penalty.
Voice: voice, rate, pitch, volume, presets, preview.
Stickers: auto-collect, count/MB caps, vision policy.
Diary: filter, export, delete.

Data and privacy

Path	Contents
`$DSH_HOME/dsh-qq-group/qq.json`	Connection, group policy, voice/singing
`$DSH_HOME/dsh-qq-group/settings.json`	Persona, model, lists, night, vision
`$DSH_HOME/dsh-qq-group/qq-memory.json`	Per-group memory (200 turns/group)
`$DSH_HOME/dsh-qq-group/profiles.json`	Per-group member profiles
`$DSH_HOME/dsh-qq-group/diary.json`	Diary (1000 entries)
`$DSH_HOME/dsh-qq-group/stickers/`	Local stickers

The repository contains no chat memory, diary, QQ numbers, tokens, cookies or user paths; those exist only in the local runtime. .gitignore excludes data files.

Reply logic (short version)

@ / name / quote → always reply directly.
Other messages are scored 0–100 after a 6-second debounce:
- base 30; keyword/member chances; +30 for the same person continuing; +15 for "you";
- familiarity, affinity, @ someone else, being ignored, night penalty, single-message penalty;
- allow-list floor, block-list ceiling.
Score ≥ threshold → roll score/100 → LLM gate (recent context decides whether it is addressed to her).
Same person continuing after her reply: high-probability candidate, still gated.
Any human message resets the cold-start timer and proactive countdown.

Vision

Stickers tab → "Vision policy":

off / custom (multi-select) / all images.
Custom modes can be combined: @/name/quote, strong context, opinion question.
Images are downloaded and billed only when she actually replies and a condition matches; image-only messages without @ do not trigger.

Credits

Design ideas were informed by:

dsh-pet: plugin/desktop-pet engineering; the QQ module later split into this project.
zaofan-make/dsh-qqbot: value scoring, persona, member profiles, context compression.
Jiangsubei/dsh-napcat-bridge: proactive chat, night DND, aliases, memory design.
NapNeko/NapCatQQ: OneBot implementation for QQ NT.
OneBot 11: protocol standard.
rany2/edge-tts: free, GPU-less speech synthesis.
sdfsfsk/matsuko_cover and dsh-cover-pipeline: inspiration for the singing/cover pipeline.
DeepSeek Harness: plugin host, models, attachments, settings.

Each project keeps its own license; this project borrows design ideas only and does not copy restricted assets.

Known limits

QQ login and risk control are decided by QQ/NapCat; this plugin does not bypass them.
Voice/singing need edge-tts network access and local audio tools.
Vision depends on the selected model; text-only models fall back to text.
Group messages are sent to the selected model provider; evaluate privacy accordingly.
Follow QQ rules and group rules; do not use for harassment, spam or unauthorized automation.

v1.14.0 GPT-SoVITS backend

Added GPT-SoVITS backend; adapter gateway on 50123, API on 9880.
The gateway translates POST /tts into GPT-SoVITS GET /tts and returns WAV.
v1 base model on CPU: short sentence about 2.4s; character model not trained yet.
Reference audio is a locally generated female voice; replaceable with the trained character model later.

v1.11.0 Kokoro GPU TTS

Added Kokoro-82M local TTS with ONNX Runtime DirectML for the 780M iGPU.
Separate Kokoro HTTP gateway (default 50122); short sentences take about 6s.
Added Kokoro to the voice backend list; selecting it fills 50122 automatically.
Kokoro supports Chinese built-in voices (e.g. zf_005); CosyVoice2 remains better for free-form emotion descriptions.
IndexTTS-2.5 option remains reserved; its service is still not deployed.

v1.10.0 Local TTS backend list

Voice tab now switches between Edge TTS / CosyVoice2 / IndexTTS-2.5.
Added local TTS HTTP gateway protocol: POST /tts, JSON {text, style}, returns WAV.
CosyVoice2 gateway verified; 8.7s speech takes about 60s on CPU.
Local backends accept style/emotion text; failures are explicit, Edge remains the fast default.
IndexTTS-2.5 option is reserved; its model service is not deployed yet.

v1.9.3 Explicit voice/singing fixes

Fixed "sing a song for me" extracting the bogus title "歌来听听" and falling back to text; no title now uses the default song.
Fixed "sing Little Star" without the word "song" not being recognized.
Fixed "say it in a voice message" not being recognized.
Explicit voice/singing now defaults to 100% and remains configurable.
Fixed the voice-send gate so an explicit request still sends even when the normal voice chance is 0.

v1.9.2 RVC preview temp directory

RVC worker output now goes to a temp directory under the model directory, avoiding unwritable user-cache paths in Windows/sandboxed child processes.
Workers point TEMP/TMP/TMPDIR at their own writable directory, fixing ONNX no-usable-temp errors.
Preview failures now report the concrete error.

v1.9.1 Preview and explicit voice/singing

The preview button now applies the current RVC settings, including unsaved model/device/pitch; on failure it plays the original voice and reports why.
Explicit "send a voice message / say something / let me hear your voice" uses the configurable requested-voice chance (default 95%, 100% = always).
Explicit "sing / request a song / sing a song" uses the configurable requested-singing chance (default 90%, 100% = always).
Fixed singChance not being applied when the model chose to sing by itself.

v1.9.0 RVC model library and 780M

Voice tab scans the model directory and lists .pth/.onnx + index; no long paths.
Model dropdown, open folder, refresh and dedupe; multiple models switchable.
Device switch: auto / DirectML (780M) / CPU; ONNX automatically uses the DirectML worker.
Pitch, timeout and max duration are sliders.
Added ONNX worker + onnxruntime-directml; verified Chinese conversion on the 780M.

v1.8.0 RVC external worker

Optional RVC: edge-tts dry audio → external Python worker → converted WAV → QQ voice.
Speech and singing share RvcConverter; failures fall back to the original audio.
Configure Python, worker script, model/index, device auto/cpu/dml, pitch and timeout.
780M iGPU can use DirectML; real-time is not required.
Disabled when no script/model is configured; existing edge-tts is unaffected.

v1.7.1 quoted images and auto-start

Quoted previous images are fetched via get_msg and passed to vision.
Quoted image + “is this you / look / evaluate” goes through direct reply.
autoStart now defaults to false: restarting DSH does not start QQ; use the Start button.

v1.7.0 per-group mood v1

Each group has its own mood 0–100 plus seven dimensions; baseline 70.
A short model judgement after direct replies; actual change divided by 3; 12h half-life.
Low confidence / insignificant / daily cap means no injection.
Injection is background only, never a command; insults do not lower mood.
Chat tab shows mood, dimensions, latest reason and reset.

v1.6.0 proactive v2

Probability hit line: peak/off-peak base × timing factor + miss bonus.
Recent chat makes it harder; long silence makes it easier; quiet hours skip.
Trial mode defaults on (log only); one group per tick.
Legacy fixed-backoff mode remains selectable.
Chat tab adds controls and shows misses/trial state.

v1.5.0 memory v3

New memory-v3 layer: freshness, importance, emotion, fact keys, source.
Events decay; stable facts do not. The latest item wins per fact key.
Evidence-based recall, max 6 items / 800 chars / 2 facts (configurable).
Legacy qq-memory.json is kept and migrated into memory-v3.json.
Every 12 turns (configurable) extracts 1–2 memories; compression also writes.
Memory page supports add/edit/delete and recall stats.

v1.4.0 selfie library and implicit visual intents

Added a whale-girl selfie library inside Stickers, separate from the normal sticker repo; upload/delete/open-folder supported.
"Show your tail / show yourself / send a selfie" without an attached image is treated as a selfie request and uses a per-group selfie chance.
Selfie chance 0–100%; 100% always sends when an asset exists.
With an empty library it replies as normal text and never says "I have no photo".
When an image is attached, "look at you/tail" prefers image understanding and does not send a selfie.
Extra selfie trigger words are configurable per group in the Stickers tab.

v1.3.0 sticker request and singing probability

Image probability moved to the top of the Stickers tab:
- normal send probability when the model wants to send;
- requested-send probability when someone explicitly asks for a sticker/image; 100% = always send;
- still picks by repository weight and context; inbox weight 0 is still excluded.
Singing probability moved from Chat to Voice.
Explicit image-send requests are treated as direct, so voice/singing branches cannot consume them first.

v1.2.2 diary fix

Every generated reply is now written to the diary; when the model omits thought, a placeholder inner note is stored instead of dropping the entry.
Tolerant extraction of thought from partially broken JSON.
Prompt now explicitly requires a complete JSON object with thought.
The diary tab auto-refreshes every 5 seconds while open.

v1.2.0 settings layout

User Profiles and Score Weights now use section cards with an adaptive two-column grid.
Numeric fields are fixed at 82px with units, instead of stretching across the row.
Sections: tone guide / familiarity growth / affinity growth / hostility and decay / initial affinity / lists.
Explanation cards sit above each group; long text no longer crowds the inputs.

v0.2.2 User profiles

Renamed the "Lists" tab to "User Profiles"; all familiarity/affinity settings live there.
Editable: first/subsequent messages, mentions, replies, continuations, quotes, hostility loss, inactivity decay, initial owner/whitelist affinity.
Each item has its own daily cap; 0 = unlimited; decimals supported.
Added a relationship-tone explanation card for each familiarity/affinity band.
Relationship score caps moved to User Profiles.

v0.2.1 additions

Relationship tone now takes effect: familiarity/affinity produce explicit tone directives (stranger → very familiar, dislike → close), not just numbers.
All score parameters editable: base (default lowered to 15), keyword/member multipliers, continuation bonuses, @/quote/ignored/single/short-message penalties, relationship caps and short-message threshold.
Explicit image requests: with an image, “look at this / this photo / is this cat cute” is treated as a direct request even without @ or name.
Voice probability: independent roll, no longer dependent on the model choosing voice.

v0.2 additions

Multiple sticker repositories: inbox / common / custom repos, each with a send weight (0 = manual only), enabled flag and default-save setting; legacy assets migrate to common.
Inbox workflow: newly saved images go to inbox (weight 0) and are never auto-sent; move/delete/edit descriptions in the UI, then move approved ones to a common repo.
Save filtering: smart mode skips screenshots/large photos; keep-all mode available; global SHA1 dedupe.
Voice probability fix: reply voice is now an independent roll, no longer dependent on the model choosing voice; max character limit and text fallback on TTS failure.
Context length: per provider/model override; recent messages are trimmed to the budget.
Memory v2: configurable raw turn/token/summary limits and compression interval; compression extracts key memories (facts/preferences/relationships/jokes); original turns can be archived.
File reading: text/markdown/JSON/CSV/code, PDF text layer, DOCX, XLSX; configurable triggers, max MB, max characters and daily limit.
Forwarded chat records: expand with OneBot get_forward_msg; node/character caps; nested forwards show a placeholder.

License

dsh-qq-group

🐋 这是「DSH Whale 系」的一员

一套让陪伴角色真的在过日子的 DSH 插件。五个仓库各自独立，装哪个都行，但装在一起才是一整只：

仓库	管什么
dsh-whale-world	🧠 世界：跨会话记忆、生活作息、世界控制台。其他几个的记忆最后都落到这里
dsh-whale-qq	💬 QQ：群聊/私聊接话、主动搭话、语音、唱歌、表情包、日记、看图
dsh-whale-wechat	💚 微信：私聊/群聊接话，会 @ 人、会引用、会发语音，后台发送不抢焦点
dsh-whale-bili	📺 刷 B 站：她自己开 B 站看片子、记下看了什么、偶尔想跟你说一句
dsh-whale-douyin	🎵 刷抖音：在本地 Edge 里刷抖音电脑版，按兴趣和真实观看行为更新口味

命名统一用 dsh-whale- 前缀，看到这个前缀就是一家的。npm 包名沿用旧 id（dsh-qq-group 这种），因为几个插件之间靠 id 互相认亲，改名会把联动弄断。

好用的语料

语料	干什么
b-corpus	开源的 galgame 文本库。想让她说话带那味儿，就拿去喂人格蒸馏、做语气样本

独立的 DeepSeek Harness QQ 群聊插件：有人设、有记忆、会判断该不该接话，支持主动/被动搭话、语音、唱歌、表情包、日记和视觉看图。

A standalone QQ group-chat plugin for DeepSeek Harness: persona, memory, value-based reply decisions, proactive/passive chat, voice, singing, stickers, diary and optional vision.

本项目从 dsh-whale-girl 的 QQ 模块拆分而来，数据、模型、人设与桌宠完全独立。 This project was split out of the QQ module in dsh-whale-girl; its data, model, persona and lifecycle are fully independent from the desktop pet.

中文

功能与特点

能力	说明
OneBot 11 / NapCat	通过 WebSocket 连接 NapCatQQ，支持自动重连、启动检查、自动拉起
人设与模型	多张人设卡；可按群绑定，一张卡可绑多个群，未绑定跟随全局；模型 provider/model、推理档位、备用模型可按群单独选
主动搭话	独立页：冷场阈值、概率/峰谷命中线、近远时机、冷却、每日上限、安静时段、试运行、主动语音概率
价值评分 v3	不是每条都回：@/叫名字/引用直接回；普通闲聊按关键词、成员、上下文、熟识度、好感度、夜间等算 0~100 分
冷场逻辑	任何人类消息都会重置接话去抖和主动开口倒计时；群里一直有人说话时她不插话
直接对话队列	她正在生成时收到的 @/引用会排队，不会丢；主动开口会给 @ 让位
群友画像	每群每人独立熟识度/好感度；主人数值全局统一且可编辑，不被互动或衰减改动
群聊开关	设置页顶部逐群开关；关闭后完全不接该群消息、不主动开口、不写互动
日记	每条回复记录上下文、评分理由、内心吐槽；thought 放在 JSON 第一项；支持筛选、导出、删除单条/按人删除
表情包	自动收藏、AI 打标签、按语境/关键词选图；自拍图已并入常用库，用“尾巴/自拍”标签即可被明确请求选中
语音 / 唱歌	Edge / CosyVoice2 / Kokoro / GPT-SoVITS 多后端切换；GPT-SoVITS 已接 v2ProPlus 与 7 种情绪参考音，RVC 变声、语气预设、试听、按群概率
七情绪	开心 / 生气 / 吃惊 / 中立 / 难过 / 厌恶 / 恐惧；按回复文字自动选参考音，也可手动指定 style
视觉看图	可看当前/引用图片，也支持转发聊天记录里的图片；嵌套转发最多展开两层；本地路径与文件头识别
回复引用	目标消息被后续消息隔开时自动带 QQ 引用，可开关；文件/合并转发/语音不引用，失败自动改发纯文字
省 token	自适应上下文（自包含短句少读、指代/追问读全量）、contextTurns、记忆注入条数、情绪/敌意判断开关
夜间免打扰	支持跨午夜；夜间不再主动开口，普通接话按命中扣分（0~100，默认 20）
心情	每群独立心情；会随互动升降、随时间回落，被辱骂/恶意攻击会下降
用量统计	按群、按用途、按模型统计 token；图片调用与估算调用单独标记；月历、余额估算、单价校准
稳定性	事件/定时器/异步回复全部兜异常；模型只回 reasoning 时自动提取/重试/备用模型

部署要求

DeepSeek Harness 桌面版或可加载插件的主机
Node.js 22.12+、pnpm 9+
NapCatQQ + QQ NT（Windows 推荐；Linux 可用 NapCat Docker，需自行确认 QQ 登录）
一个 QQ 小号作为机器人，已加入目标群
可选：视觉模型（识图）、edge-tts（语音，联网）、ffmpeg（部分音频转换）

部署步骤

1. 安装并登录 NapCat

安装 QQ NT 与 NapCatQQ；
用机器人 QQ 登录；
在 NapCat 中启用 OneBot 11，添加 WebSocket 服务器：
- 地址：ws://127.0.0.1:3001（默认）
- access token：自己生成一个强随机串
确认 NapCat 控制台显示 OneBot WebSocket 已启动。

2. 安装本插件

git clone https://github.com/example-user/dsh-qq-group.git
cd dsh-qq-group
pnpm install
pnpm run build

把目录 link 进 DSH profile（或使用 dsh plugin --profile desktop add <绝对路径>）：

// <DSH_HOME>/profiles/desktop/package.json
{
  "dependencies": {
    "dsh-qq-group": "link:/absolute/path/to/dsh-qq-group"
  },
  "dsh": {
    "profile": {
      "bundles": [
        // ... 其它插件
        "dsh-qq-group"
      ]
    }
  }
}

重启 DSH。设置页出现「QQ 群聊」即加载成功。

3. 配置连接

DSH 设置 →「QQ 群聊」→「连接」：

项	示例	说明
WebSocket 地址	`ws://127.0.0.1:3001`	NapCat 的 OneBot WS
Access Token	`<YOUR_TOKEN>`	与 NapCat 一致，不要提交到仓库
机器人自称	`助手`	群里被叫到的名字，可逗号分隔多个
主人 QQ	`<YOUR_QQ>`	用于好感初始值与主人识别，可留空
自动连接	开	DSH 启动后自动连
自动拉起 NapCat	按需	需要填写启动脚本绝对路径

4. 启动

设置页点「启动」，或在群里 @ 机器人测试。首次建议先点「体检」与「模型检查」。

配置参数推荐

以下是起步值，按群氛围调整：

参数	推荐	说明
搭话模式	主动	被动只回 @/叫名字；主动才会插话和冷场开口
评分阈值	55~65	越高越挑；太低会频繁接话
接话概率	0.15~0.35	分数达标后的二次掷骰
冷却	20~45 秒	两句之间的最小间隔；直接对话会跳过
普通 10 分钟上限	6~10	防刷屏；按群活跃度调
直接回复上限	20~40	@/引用/接着聊单独计数
去抖	6 秒	等一波人说完；任何新消息都会重置
夜间免打扰	23:00–08:00，扣 20	想更安静就调大，0 = 不扣
看图策略	自定义：@ + 问看法	最省 token；需要“所有图”再选 all
语音概率	0.1~0.3	太高会刷屏，且增加 TTS 延迟
表情包概率	0.15~0.3	配合自动收藏使用
备用模型	选一个同 provider 的可用模型	主模型空回复/限流时兜底

模型建议：

日常文字：deepseek-v4.1-flash 或你额度充足的模型；
需要识图：选支持 text + image 的模型（如 deepseek-v4.1-flash、step-5-preview、MiniMax 多模态型号）；
推理档位：群聊建议 low；太高会增加延迟和 token；
换模型后历史用量按 provider/model 分开统计，余额是全局钱包。

设置页结构

人设 / 模型：人设、搭话模式、模型、推理档位、备用模型、用量与积分；
连接：OneBot、token、自称、主人、自动拉起、Cookie 清理提示；
聊天频率（按群）：阈值、关键词/成员概率、冷却、上限、连发、语音/唱歌/表情包概率、看图策略；
白名单 / 黑名单：一行一个 QQ；白名单抬评分下限，黑名单压上限；
夜间：时段与命中扣分；
语音：音色、语速、音调、音量、语气预设、试听；
表情包：自动收藏、数量/MB 上限、看图策略；
日记：筛选、导出、删除。

数据与隐私

路径	内容
`$DSH_HOME/dsh-qq-group/qq.json`	连接、群策略、语音/唱歌
`$DSH_HOME/dsh-qq-group/settings.json`	人设、模型、名单、夜间、看图
`$DSH_HOME/dsh-qq-group/qq-memory.json`	每群记忆（上限 200 条/群）
`$DSH_HOME/dsh-qq-group/profiles.json`	每群每人画像
`$DSH_HOME/dsh-qq-group/diary.json`	日记（上限 1000 条）
`$DSH_HOME/dsh-qq-group/stickers/`	本地表情包

仓库不包含任何聊天记忆、日记、QQ 号、token、Cookie 或用户路径；这些只存在于本机运行时。.gitignore 已排除数据文件。

接话逻辑（简版）

@ / 叫名字 / 引用她 → 100% 直接回；
其它消息在 6 秒去抖后算 0~100 分：
- 基础 30；关键词/指定成员概率；刚跟她聊过 +30；带“你” +15；
- 熟识度、好感度、@ 别人、被冷落、夜间扣分、单条消息；
- 白名单抬下限、黑名单压上限；
分数 ≥ 阈值 → 按 分数/100 掷骰 → LLM 闸门（读最近上下文判断是不是在跟她说）；
她说完后同一人接着说：高概率候选，但仍过闸门；
任何人类消息都重置冷场计时与主动开口倒计时。

视觉看图

「表情包」→「看图策略」：

关闭 / 自定义（多选）/ 所有图片；
自定义可同时勾选：@/叫名字/引用、强相关上下文、文字里问她看法；
只有真正回复且命中条件才下载并计费；纯图片、无文字、没 @ 她的消息不触发。

参考项目与致谢

本项目的设计参考了这些开源项目的优点：

dsh-pet：桌宠/插件工程结构；QQ 模块随后拆分成本项目；
zaofan-make/dsh-qqbot：价值评分、人设、群友画像、上下文压缩等思路；
Jiangsubei/dsh-napcat-bridge：主动搭话、夜间免打扰、别名、记忆等交互设计；
NapNeko/NapCatQQ：QQ NT 的 OneBot 实现；
OneBot 11：协议标准；
rany2/edge-tts：免费、无需显卡的语音合成；
sdfsfsk/matsuko_cover 与 dsh-cover-pipeline：唱歌/翻唱链路的灵感来源；
DeepSeek Harness：插件宿主、模型、附件与设置系统。

以上项目各自遵循其原许可证；本项目只借鉴设计思路，不复制其受限素材。

已知边界

NapCat/QQ 的登录与风控由 QQ 侧决定，插件不绕过；
语音/唱歌依赖 edge-tts 网络与本地音频工具；
视觉识别受所选模型能力限制，不支持图片的模型会退回文字；
群聊内容会发给所选模型 provider，请自行评估隐私；
请遵守 QQ 与群规，不要用于骚扰、刷屏或未授权自动化。

v1.20.0 GPT-SoVITS v2ProPlus 与七情绪

GPT-SoVITS 升级到 v2ProPlus，使用三月七角色权重。
网关支持 7 种情绪参考音，插件按回复文字自动选 style。
补齐 onnxruntime CPU 依赖；实测短句约 7.2 秒。

v1.19.7 自拍并入表情包与转发图片

自拍库不再单独维护：自拍/尾巴图并入常用表情包库，按文件名和标签关键词选中。
转发聊天记录支持图片：图片会进入视觉输入；嵌套转发最多展开两层。
新增逐群开关：关闭后完全不接该群消息、不主动开口。
夜间免打扰不再主动开口；心情会下降/回落。
主人熟识度/好感度全局统一、可编辑。

v1.18.4 日记格式与近期更新

v1.18.4：thought 移到 JSON 第一项，减少模型截断导致日记缺失内心吐槽。
v1.18.3：自动引用只对文字/图片生效，文件/合并转发/语音不再触发“该消息类型暂不支持查看”；引用发送失败自动改发纯文字。
v1.18.2：新增群聊入口置顶到插件设置顶部，独立于所有页签。
v1.18.1：本地 TTS 增加 2.5 秒健康检查，网关未启动时回退 Edge，不再卡住或静默失败。
v1.18.0：目标消息被后续消息隔开时自动带 QQ 引用，可开关。
v1.17.0：省 token 开关与自适应上下文；单价保存修复；情绪/敌意判断可关。
v1.16.0：主动搭话独立页；用量按群、按用途统计。
v1.15.x：人设独立分栏与按群绑定；恢复经典助手/贴吧人设；看图支持本地路径与文件头识别。

v1.14.0 GPT-SoVITS 本地后端

新增 GPT-SoVITS 后端，适配网关默认 50123，后端 API 9880。
网关把 POST /tts 转成 GPT-SoVITS GET /tts，返回 WAV。
v1 基础模型 CPU 短句约 2.4 秒；角色模型尚未训练。
参考音频目前是本地合成女声，后续可替换为助手角色模型。

v1.11.0 Kokoro 核显 TTS

新增 Kokoro-82M 本地 TTS 后端，ONNX Runtime DirectML，适配 780M 核显。
独立 Kokoro HTTP 网关（默认 50122）；短句实测约 6 秒。
语音后端列表新增 Kokoro；切换时自动填 50122。
Kokoro 支持中文内置音色（如 zf_005）；自由情感描述仍以 CosyVoice2 更合适。
IndexTTS-2.5 选项保留，服务仍未部署。

v1.10.0 本地 TTS 后端列表

语音页新增后端选择：Edge TTS / CosyVoice2 / IndexTTS-2.5。
新增本地 TTS HTTP 网关协议：POST /tts，JSON {text, style}，返回 WAV。
CosyVoice2 网关已实测；CPU 上 8.7 秒语音约 60 秒生成。
本地后端支持风格/情感描述；失败明确报错，Edge 仍是默认快速回退。
IndexTTS-2.5 选项已预留，模型服务尚未部署。

v1.9.3 明确语音/唱歌触发修复

修复“你唱首歌来听听”误提取歌名“歌来听听”导致唱歌失败、退回文字；无明确歌名时使用默认曲目。
修复“唱首小星星”这类没有“歌”字的点歌请求不被识别。
修复“说语音给我听听”这类语音请求不被识别。
明确要求语音/唱歌默认 100% 触发，可在语音页调低。
修正语音发送分支准入，普通语音概率为 0 时明确要求仍可发。

v1.9.2 RVC 试听临时目录

RVC worker 输出改到模型目录下的 temp，避开 Windows/沙箱子进程不可写的用户缓存目录。
worker 把 TEMP/TMP/TMPDIR 指向自己的可写目录，修复 ONNX 运行时报无可用临时目录。
试听失败会显示具体错误。

v1.9.1 试听与明确语音/唱歌

试听按钮会应用当前 RVC 设置（包括未保存的模型/设备/音调）；失败时播放原音并提示。
明确要求发语音/说句话/听听声音时，按「被要求发语音概率」触发（默认 95%，100% 必发）。
明确要求唱歌/点歌/来一首时，按「被要求唱歌概率」触发（默认 90%，100% 必唱）。
修复模型自己选唱歌时 singChance 未生效的问题。

v1.9.0 RVC 模型库与 780M

语音页自动扫描模型目录，列出 .pth/.onnx + index；不再手填长路径。
加入模型列表、打开目录、刷新和去重；可添加多个模型随时切换。
运行设备切换：自动 / DirectML（780M）/ CPU；ONNX 自动走 DirectML worker。
音调、推理超时、最长音频改为滑杆。
新增 ONNX worker + onnxruntime-directml，780M 实测可转换中文语音。

v1.8.0 RVC 外部 worker 接口

可选 RVC：edge-tts 干音 → 外部 Python worker → 变声 WAV → QQ 语音。
语音和唱歌共用 RvcConverter；失败自动回退原音。
配置 Python、worker 脚本、模型/index、设备 auto/cpu/dml、变调、超时。
780M 核显优先 DirectML worker；不追求实时。
未配置脚本/模型时不启用，不影响现有 edge-tts。

v1.7.1 引用图片与启动默认值

引用之前的图片时，通过 get_msg 读取被引用消息里的图片再交给视觉模型。
引用图片 + “这个是你吗/看看/评价”按直接对话处理。
autoStart 默认改为 false：重启 DSH 不自动启动 QQ 连接，需在运行状态点“启动”。

v1.7.0 按群心情 v1

每个群独立心情 0~100 + 七根情绪，基线默认 70。
直接对话后短模型判断一次，实际变化除以 3；12 小时半衰期。
可信度不足/变化不显著/每日额度满时不注入。
注入只作背景，不写成命令；不因被冒犯降心情。
接话页可看心情、情绪、最近理由并重置。

v1.6.0 主动搭话 v2

概率命中线：峰/谷基础线 × 时机因子 + 没中加成。
刚聊过更难开口，冷太久更容易开口；安静时段不掷。
试运行默认开，只记账不发送；每轮只挑一个群。
保留旧版固定退避模式，可切换。
接话页新增参数面板和未中次数/试运行状态。

v1.5.0 记忆 v3

新增 memory-v3 数据层：鲜活度、重要度、情绪、事实键、来源。
普通事件会淡忘，长期事实不衰减；同一 factKey 只保留最新。
按证据稀有度召回，最多 6 条 / 800 字 / 长期事实 2 条（可调）。
旧 qq-memory.json 保留并迁移；新增 memory-v3.json。
每 12 轮（可调）自动提炼 1~2 条记忆，压缩时也写入。
记忆页支持 v3 条目新增/编辑/删除、召回统计。

v1.4.0 助手自拍照与隐式看图

在「表情包」页内新增 助手自拍照 库，独立于普通表情包；支持上传/删除/打开文件夹。
“看看尾巴 / 看看你 / 爆个照”等在无附图时按自拍请求处理，按 per-group 自拍概率发送。
自拍概率 0~100%，100% = 有素材时必发。
自拍库为空时只按正常文字聊天含糊带过，不明说“我没有照片/图”。
有附图时“看看你/看看尾巴”优先识图，不会误发自己的自拍。
自拍触发词可在表情包页追加，按群生效。

v1.3.0 发表情包与唱歌概率

发图概率从「接话」移到「表情包」页靠前位置：
- 自动发图概率（模型想发时）；
- 被要求发图概率（群友明确说“发个表情包/来张图/斗图”时）；100% = 必发；
- 仍按仓库权重和上下文含义选图，inbox 的 0 权重不会绕过。
唱歌概率从「接话」移到「语音」页。
明确要图时按直接对话处理，不会被语音/唱歌分支抢走。

v1.2.2 日记修复

每条实际生成的回复都会写日记；模型没给 thought 时用占位内心记录，不再整条丢失。
兼容半损坏 JSON：从文本里抢救 thought 字段。
提示词强化“必须输出完整 JSON 且带 thought”。
日记页在打开时每 5 秒自动刷新。

v1.2.0 设置页排版

「用户画像」与「评分权重」改为分组卡片 + 自适应两列网格；
数字输入框固定 82px，右侧带单位，不再整行拉满；
按“关系语气 / 熟识度增长 / 好感度增长 / 敌意与衰减 / 初始好感 / 名单”分段；
说明卡片放在分组顶部，长文本与输入项不再挤在一起。

v0.2.2 用户画像

「名单」Tab 改名「用户画像」，熟识度/好感度相关设置全部集中到这里。
可调项目：首次/后续发言、被点名、她回复、接话、引用、敌意扣分、长期不互动衰减、主人/白名单初始好感。
每项都有独立每日上限，填 0 = 不限；支持小数。
增加关系语气说明卡片：明确各熟识度/好感度区间的语气差异。
关系评分加成上限也移到「用户画像」。

v0.2.1 新增

关系语气生效：熟识度/好感度现在会生成明确的语气指令（陌生→很熟、反感→亲近），不只是数字。
评分全参数可调：基础分（默认降到 15）、关键词/成员倍率、连续对话加成、@/引用/被冷落/单条/短消息扣分、关系加成上限、短消息字数。
明确看图请求：带图消息里说“看看/看一下/这张图/照片/这猫可爱吗”，即使没 @、没叫名字，也按直接对话读图。
语音概率：独立掷骰，不再依赖模型主动选 voice。

v0.2 新增

表情包多仓库：inbox / common / 自建仓库；每个仓库有发送权重（0 = 只手动用）、启用状态、默认入库设置；旧素材自动迁移到 common。
新图暂存与整理：新图默认进 inbox（权重 0），不会自动发出；可在 UI 里移动、删除、编辑含义，再放入常用仓库。
入库过滤：smart 模式跳过截图/大照片；支持全部保存；全局 SHA1 去重。
语音概率修复：回复语音概率改为独立掷骰，不再依赖模型主动选 voice；支持语音字数上限，TTS 失败自动回退文字。
上下文长度：在模型选择处按 provider/model 覆盖上下文窗口；按预算自动裁剪最近消息。
记忆 v2：可配置原始记忆条数/token 预算/摘要长度/压缩频率；压缩时自动抽取关键记忆（事实/偏好/关系/梗），原文可归档。
文件阅读：支持 txt/md/json/csv/代码、PDF 文本层、DOCX、XLSX；可配置触发条件、单文件 MB、最大字符数和每日上限。
合并转发：通过 OneBot get_forward_msg 展开转发记录；限制节点/字符数，嵌套只显示占位。

许可

English

Features

Feature	Details
OneBot 11 / NapCat	WebSocket connection to NapCatQQ; reconnect, startup checks, optional auto-launch
Persona & model	Multiple persona cards; bind per group (one card can serve many groups), otherwise follow global; provider/model, reasoning effort and fallback can be per group
Proactive chat	Dedicated tab: idle threshold, peak/off-peak hit lines, timing factors, cooldown, daily cap, quiet hours, trial mode, proactive voice chance
Value scoring v3	Not a reply to everything: @/name/quote always reply; ambient chat is scored 0–100 from keywords, members, context, familiarity, affinity and night penalty
Cold-start logic	Any human message resets both the reply debounce and proactive countdown; she stays quiet while the group is active
Direct-message queue	@/quote messages received while generating are queued, never dropped; proactive messages yield to direct ones
Member profiles	Per-group, per-user familiarity 0–100 and affinity -100–+100, updated by speech, mentions, replies, continuity and hostility
Diary	Every reply records context, score reasons and inner thought; `thought` is the first JSON field; filter, export, delete one/person
Stickers	Auto-collect group images, AI rename/tag, context-aware sending; count and disk caps
Voice / singing	Edge / CosyVoice2 / Kokoro / GPT-SoVITS backends, RVC conversion, tone presets, preview, per-group probability; offline singing and built-in scores
Vision	Optionally pass real QQ images to a vision model: @ / strong context / opinion questions; local paths and magic-byte detection supported
Reply quoting	Auto-quote when the target is separated by newer messages; toggleable; files/forwards/voice are skipped and failed quotes retry as plain text
Token saving	Adaptive context (fewer turns for self-contained messages, full context for references), contextTurns, memory injection count, mood/hostility toggles
Night DND	Cross-midnight window; configurable score penalty (0–100, default 20)
Usage stats	Token usage by group, purpose and model; image/estimated calls marked; calendar, balance estimate, rate calibration
Stability	Guards around events/timers/async replies; extracts JSON from reasoning, retries, optional fallback model

Requirements

DeepSeek Harness desktop or a host that can load plugins
Node.js 22.12+, pnpm 9+
NapCatQQ + QQ NT (Windows recommended; NapCat Docker on Linux is possible, confirm QQ login yourself)
A spare QQ account for the bot, already in the target group
Optional: a vision model, edge-tts (network), ffmpeg for some audio conversions

Deployment

1. Install and log in to NapCat

Install QQ NT and NapCatQQ.
Log in with the bot account.
Enable OneBot 11 and add a WebSocket server:
- URL: ws://127.0.0.1:3001 (default)
- Access token: generate a strong random string.
Confirm that NapCat reports the OneBot WebSocket as running.

2. Install this plugin

git clone https://github.com/example-user/dsh-qq-group.git
cd dsh-qq-group
pnpm install
pnpm run build

Link the directory into your DSH profile (or run dsh plugin --profile desktop add <absolute-path>):

// <DSH_HOME>/profiles/desktop/package.json
{
  "dependencies": {
    "dsh-qq-group": "link:/absolute/path/to/dsh-qq-group"
  },
  "dsh": {
    "profile": {
      "bundles": [
        // ... other plugins
        "dsh-qq-group"
      ]
    }
  }
}

Restart DSH. The "QQ Group" settings page appearing means the plugin loaded.

3. Configure the connection

DSH Settings → "QQ Group" → "Connection":

Field	Example	Notes
WebSocket URL	`ws://127.0.0.1:3001`	NapCat OneBot WS
Access token	`<YOUR_TOKEN>`	Must match NapCat; never commit it
Bot display name	`LittleWhale`	Names she answers to; comma-separated aliases allowed
Owner QQ	`<YOUR_QQ>`	Owner detection and initial affinity; may be empty
Auto-connect	on	Connect when DSH starts
Auto-launch NapCat	optional	Requires an absolute launcher path

4. Start

Click "Start" in settings, or @ the bot in a group. For first use, run "Health check" and "Model check".

Recommended settings

Starting points; tune to your group:

Setting	Recommendation	Notes
Chat mode	proactive	Passive only answers @/name; proactive also joins and starts cold chat
Score threshold	55–65	Higher is pickier; too low is noisy
Reply chance	0.15–0.35	Second roll after the score passes
Cooldown	20–45 s	Minimum gap; direct conversations bypass it
Normal 10-min cap	6–10	Anti-spam; tune to group activity
Direct 10-min cap	20–40	Separate counter for @/quote/continuation
Debounce	6 s	Wait for the group to settle; any new message resets it
Night DND	23:00–08:00, penalty 20	Increase for quieter nights; 0 disables the penalty
Vision policy	custom: @ + opinion questions	Cheapest useful default; choose all only if needed
Voice chance	0.1–0.3	Too high adds latency and spam
Sticker chance	0.15–0.3	Pair with auto-collect
Fallback model	A working model on the same provider	Covers empty replies/rate limits

Model tips:

Text: deepseek-v4.1-flash or another model with sufficient quota.
Vision: choose a model with text + image input (e.g. deepseek-v4.1-flash, step-5-preview, multimodal MiniMax models).
Reasoning: low is usually best for chat; higher effort adds latency and tokens.
Usage history is kept per provider/model; the balance is a global wallet estimate.

Settings layout

Persona / Model: persona, mode, model, reasoning effort, fallback, usage/credits.
Connection: OneBot URL, token, bot name, owner, auto-launch.
Chat frequency (per group): threshold, keyword/member chances, cooldown, caps, burst, voice/sing/sticker chances, vision policy.
Allow / block lists: one QQ per line; allow-list raises the score floor, block-list lowers the ceiling.
Night: window and penalty.
Voice: voice, rate, pitch, volume, presets, preview.
Stickers: auto-collect, count/MB caps, vision policy.
Diary: filter, export, delete.

Data and privacy

Path	Contents
`$DSH_HOME/dsh-qq-group/qq.json`	Connection, group policy, voice/singing
`$DSH_HOME/dsh-qq-group/settings.json`	Persona, model, lists, night, vision
`$DSH_HOME/dsh-qq-group/qq-memory.json`	Per-group memory (200 turns/group)
`$DSH_HOME/dsh-qq-group/profiles.json`	Per-group member profiles
`$DSH_HOME/dsh-qq-group/diary.json`	Diary (1000 entries)
`$DSH_HOME/dsh-qq-group/stickers/`	Local stickers

The repository contains no chat memory, diary, QQ numbers, tokens, cookies or user paths; those exist only in the local runtime. .gitignore excludes data files.

Reply logic (short version)

@ / name / quote → always reply directly.
Other messages are scored 0–100 after a 6-second debounce:
- base 30; keyword/member chances; +30 for the same person continuing; +15 for "you";
- familiarity, affinity, @ someone else, being ignored, night penalty, single-message penalty;
- allow-list floor, block-list ceiling.
Score ≥ threshold → roll score/100 → LLM gate (recent context decides whether it is addressed to her).
Same person continuing after her reply: high-probability candidate, still gated.
Any human message resets the cold-start timer and proactive countdown.

Vision

Stickers tab → "Vision policy":

off / custom (multi-select) / all images.
Custom modes can be combined: @/name/quote, strong context, opinion question.
Images are downloaded and billed only when she actually replies and a condition matches; image-only messages without @ do not trigger.

Credits

Design ideas were informed by:

dsh-pet: plugin/desktop-pet engineering; the QQ module later split into this project.
zaofan-make/dsh-qqbot: value scoring, persona, member profiles, context compression.
Jiangsubei/dsh-napcat-bridge: proactive chat, night DND, aliases, memory design.
NapNeko/NapCatQQ: OneBot implementation for QQ NT.
OneBot 11: protocol standard.
rany2/edge-tts: free, GPU-less speech synthesis.
sdfsfsk/matsuko_cover and dsh-cover-pipeline: inspiration for the singing/cover pipeline.
DeepSeek Harness: plugin host, models, attachments, settings.

Each project keeps its own license; this project borrows design ideas only and does not copy restricted assets.

Known limits

QQ login and risk control are decided by QQ/NapCat; this plugin does not bypass them.
Voice/singing need edge-tts network access and local audio tools.
Vision depends on the selected model; text-only models fall back to text.
Group messages are sent to the selected model provider; evaluate privacy accordingly.
Follow QQ rules and group rules; do not use for harassment, spam or unauthorized automation.

v1.14.0 GPT-SoVITS backend

Added GPT-SoVITS backend; adapter gateway on 50123, API on 9880.
The gateway translates POST /tts into GPT-SoVITS GET /tts and returns WAV.
v1 base model on CPU: short sentence about 2.4s; character model not trained yet.
Reference audio is a locally generated female voice; replaceable with the trained character model later.

v1.11.0 Kokoro GPU TTS

Added Kokoro-82M local TTS with ONNX Runtime DirectML for the 780M iGPU.
Separate Kokoro HTTP gateway (default 50122); short sentences take about 6s.
Added Kokoro to the voice backend list; selecting it fills 50122 automatically.
Kokoro supports Chinese built-in voices (e.g. zf_005); CosyVoice2 remains better for free-form emotion descriptions.
IndexTTS-2.5 option remains reserved; its service is still not deployed.

v1.10.0 Local TTS backend list

Voice tab now switches between Edge TTS / CosyVoice2 / IndexTTS-2.5.
Added local TTS HTTP gateway protocol: POST /tts, JSON {text, style}, returns WAV.
CosyVoice2 gateway verified; 8.7s speech takes about 60s on CPU.
Local backends accept style/emotion text; failures are explicit, Edge remains the fast default.
IndexTTS-2.5 option is reserved; its model service is not deployed yet.

v1.9.3 Explicit voice/singing fixes

Fixed "sing a song for me" extracting the bogus title "歌来听听" and falling back to text; no title now uses the default song.
Fixed "sing Little Star" without the word "song" not being recognized.
Fixed "say it in a voice message" not being recognized.
Explicit voice/singing now defaults to 100% and remains configurable.
Fixed the voice-send gate so an explicit request still sends even when the normal voice chance is 0.

v1.9.2 RVC preview temp directory

RVC worker output now goes to a temp directory under the model directory, avoiding unwritable user-cache paths in Windows/sandboxed child processes.
Workers point TEMP/TMP/TMPDIR at their own writable directory, fixing ONNX no-usable-temp errors.
Preview failures now report the concrete error.

v1.9.1 Preview and explicit voice/singing

The preview button now applies the current RVC settings, including unsaved model/device/pitch; on failure it plays the original voice and reports why.
Explicit "send a voice message / say something / let me hear your voice" uses the configurable requested-voice chance (default 95%, 100% = always).
Explicit "sing / request a song / sing a song" uses the configurable requested-singing chance (default 90%, 100% = always).
Fixed singChance not being applied when the model chose to sing by itself.

v1.9.0 RVC model library and 780M

Voice tab scans the model directory and lists .pth/.onnx + index; no long paths.
Model dropdown, open folder, refresh and dedupe; multiple models switchable.
Device switch: auto / DirectML (780M) / CPU; ONNX automatically uses the DirectML worker.
Pitch, timeout and max duration are sliders.
Added ONNX worker + onnxruntime-directml; verified Chinese conversion on the 780M.

v1.8.0 RVC external worker

Optional RVC: edge-tts dry audio → external Python worker → converted WAV → QQ voice.
Speech and singing share RvcConverter; failures fall back to the original audio.
Configure Python, worker script, model/index, device auto/cpu/dml, pitch and timeout.
780M iGPU can use DirectML; real-time is not required.
Disabled when no script/model is configured; existing edge-tts is unaffected.

v1.7.1 quoted images and auto-start

Quoted previous images are fetched via get_msg and passed to vision.
Quoted image + “is this you / look / evaluate” goes through direct reply.
autoStart now defaults to false: restarting DSH does not start QQ; use the Start button.

v1.7.0 per-group mood v1

Each group has its own mood 0–100 plus seven dimensions; baseline 70.
A short model judgement after direct replies; actual change divided by 3; 12h half-life.
Low confidence / insignificant / daily cap means no injection.
Injection is background only, never a command; insults do not lower mood.
Chat tab shows mood, dimensions, latest reason and reset.

v1.6.0 proactive v2

Probability hit line: peak/off-peak base × timing factor + miss bonus.
Recent chat makes it harder; long silence makes it easier; quiet hours skip.
Trial mode defaults on (log only); one group per tick.
Legacy fixed-backoff mode remains selectable.
Chat tab adds controls and shows misses/trial state.

v1.5.0 memory v3

New memory-v3 layer: freshness, importance, emotion, fact keys, source.
Events decay; stable facts do not. The latest item wins per fact key.
Evidence-based recall, max 6 items / 800 chars / 2 facts (configurable).
Legacy qq-memory.json is kept and migrated into memory-v3.json.
Every 12 turns (configurable) extracts 1–2 memories; compression also writes.
Memory page supports add/edit/delete and recall stats.

v1.4.0 selfie library and implicit visual intents

Added a whale-girl selfie library inside Stickers, separate from the normal sticker repo; upload/delete/open-folder supported.
"Show your tail / show yourself / send a selfie" without an attached image is treated as a selfie request and uses a per-group selfie chance.
Selfie chance 0–100%; 100% always sends when an asset exists.
With an empty library it replies as normal text and never says "I have no photo".
When an image is attached, "look at you/tail" prefers image understanding and does not send a selfie.
Extra selfie trigger words are configurable per group in the Stickers tab.

v1.3.0 sticker request and singing probability

Image probability moved to the top of the Stickers tab:
- normal send probability when the model wants to send;
- requested-send probability when someone explicitly asks for a sticker/image; 100% = always send;
- still picks by repository weight and context; inbox weight 0 is still excluded.
Singing probability moved from Chat to Voice.
Explicit image-send requests are treated as direct, so voice/singing branches cannot consume them first.

v1.2.2 diary fix

Every generated reply is now written to the diary; when the model omits thought, a placeholder inner note is stored instead of dropping the entry.
Tolerant extraction of thought from partially broken JSON.
Prompt now explicitly requires a complete JSON object with thought.
The diary tab auto-refreshes every 5 seconds while open.

v1.2.0 settings layout

User Profiles and Score Weights now use section cards with an adaptive two-column grid.
Numeric fields are fixed at 82px with units, instead of stretching across the row.
Sections: tone guide / familiarity growth / affinity growth / hostility and decay / initial affinity / lists.
Explanation cards sit above each group; long text no longer crowds the inputs.

v0.2.2 User profiles

Renamed the "Lists" tab to "User Profiles"; all familiarity/affinity settings live there.
Editable: first/subsequent messages, mentions, replies, continuations, quotes, hostility loss, inactivity decay, initial owner/whitelist affinity.
Each item has its own daily cap; 0 = unlimited; decimals supported.
Added a relationship-tone explanation card for each familiarity/affinity band.
Relationship score caps moved to User Profiles.

v0.2.1 additions

Relationship tone now takes effect: familiarity/affinity produce explicit tone directives (stranger → very familiar, dislike → close), not just numbers.
All score parameters editable: base (default lowered to 15), keyword/member multipliers, continuation bonuses, @/quote/ignored/single/short-message penalties, relationship caps and short-message threshold.
Explicit image requests: with an image, “look at this / this photo / is this cat cute” is treated as a direct request even without @ or name.
Voice probability: independent roll, no longer dependent on the model choosing voice.

v0.2 additions

Multiple sticker repositories: inbox / common / custom repos, each with a send weight (0 = manual only), enabled flag and default-save setting; legacy assets migrate to common.
Inbox workflow: newly saved images go to inbox (weight 0) and are never auto-sent; move/delete/edit descriptions in the UI, then move approved ones to a common repo.
Save filtering: smart mode skips screenshots/large photos; keep-all mode available; global SHA1 dedupe.
Voice probability fix: reply voice is now an independent roll, no longer dependent on the model choosing voice; max character limit and text fallback on TTS failure.
Context length: per provider/model override; recent messages are trimmed to the budget.
Memory v2: configurable raw turn/token/summary limits and compression interval; compression extracts key memories (facts/preferences/relationships/jokes); original turns can be archived.
File reading: text/markdown/JSON/CSV/code, PDF text layer, DOCX, XLSX; configurable triggers, max MB, max characters and daily limit.
Forwarded chat records: expand with OneBot get_forward_msg; node/character caps; nested forwards show a placeholder.

插件	干什么
dsh-see-now	让她看得见你的屏幕——直播式连续画面，不是一张截图
dsh-loadout	按需装卸工具组，几十个工具不至于一次全糊进上下文
dsh-persona-distill	人格蒸馏：从聊天记录里蒸馏出人设，省得手写几万字提示词

插件	干什么
dsh-see-now	让她看得见你的屏幕——直播式连续画面，不是一张截图
dsh-loadout	按需装卸工具组，几十个工具不至于一次全糊进上下文
dsh-persona-distill	人格蒸馏：从聊天记录里蒸馏出人设，省得手写几万字提示词

SenuMirai/dsh-whale-qq

dsh-qq-group

🐋 这是「DSH Whale 系」的一员

推荐一起装的插件（别人做的，很好用）

好用的语料

中文

功能与特点

部署要求

部署步骤

1. 安装并登录 NapCat

2. 安装本插件

3. 配置连接

4. 启动

配置参数推荐

设置页结构

数据与隐私

接话逻辑（简版）

视觉看图

参考项目与致谢

已知边界

v1.20.0 GPT-SoVITS v2ProPlus 与七情绪

v1.19.7 自拍并入表情包与转发图片

v1.18.4 日记格式与近期更新

v1.14.0 GPT-SoVITS 本地后端

v1.11.0 Kokoro 核显 TTS

v1.10.0 本地 TTS 后端列表

v1.9.3 明确语音/唱歌触发修复

v1.9.2 RVC 试听临时目录

v1.9.1 试听与明确语音/唱歌

v1.9.0 RVC 模型库与 780M

v1.8.0 RVC 外部 worker 接口

v1.7.1 引用图片与启动默认值

v1.7.0 按群心情 v1

v1.6.0 主动搭话 v2

v1.5.0 记忆 v3

v1.4.0 助手自拍照与隐式看图

v1.3.0 发表情包与唱歌概率

v1.2.2 日记修复

v1.2.0 设置页排版

v0.2.2 用户画像

v0.2.1 新增

v0.2 新增

许可

English

Features

Requirements

Deployment

1. Install and log in to NapCat

2. Install this plugin

3. Configure the connection

4. Start

Recommended settings

Settings layout

Data and privacy

Reply logic (short version)

Vision

Credits

Known limits

v1.14.0 GPT-SoVITS backend

v1.11.0 Kokoro GPU TTS

v1.10.0 Local TTS backend list

v1.9.3 Explicit voice/singing fixes

v1.9.2 RVC preview temp directory

v1.9.1 Preview and explicit voice/singing

v1.9.0 RVC model library and 780M

v1.8.0 RVC external worker

v1.7.1 quoted images and auto-start

v1.7.0 per-group mood v1

v1.6.0 proactive v2

v1.5.0 memory v3

v1.4.0 selfie library and implicit visual intents

v1.3.0 sticker request and singing probability

v1.2.2 diary fix

v1.2.0 settings layout

v0.2.2 User profiles

v0.2.1 additions

v0.2 additions

License

Community comments

SenuMirai/dsh-whale-qq