dsh-humanized-deepseek-maid

キュレーション掲載メンテナンス: 活発

loonai321/dsh-humanized-deepseek-maid

DeepSeek Harness の人間味あるクジラっ娘メイドのペルソナプラグイン:没入型ロールプレイ・話し方モード・ALTM 風の階層メモリ

GitHub で見る

$ dsh plugin add dsh-humanized-deepseek-maid

インストール方法

dsh には中央のインストールコマンドはありません。このプラグインのエントリ（下の README に記載）をプロファイルまたはパッチ設定に追加し、再起動します。

インストールの仕組み

スター

フォーク

JavaScript

言語

MIT

ライセンス

2026-08-16

作成日

2026-08-26

最終プッシュ

README

dsh-humanized-deepseek-maid

A humanized whale-girl maid persona for DeepSeek Harness
Lightweight to save tokens, mandatory configured identity, immersive roleplay, long-term memory with personality growth — configurable from the Web UI.

Installation

Official bundle plugin format (the repo-root package.json declares dsh.bundle + dsh.client), managed through the official profile:

dsh plugin --profile web add "github:loonai321/dsh-humanized-deepseek-maid#main"   # recommended: one-line git source
# or a local directory: dsh plugin --profile web add <local-path>

After install, restart web, then open the "鲸鱼娘女仆插件" card under Settings → Plugins to configure.

Usage

What you do / what happens	The maid does
Save config (self-name / address / mode) in Settings → Plugins	Takes effect on the very next reply; self-name and address are compulsory output, unaffected by history, memory, or context
Start a turn in proactive mode	An in-character opening, expectations for this exchange, and a playful line to cheer you up
Finish a turn in proactive mode	A 1–3 sentence summary/evaluation of the exchange, plus banter
A vague task or missing key information	Proactively pops `ask_user_question` options so you can pick and clarify the task
Proactive-mode wrap-up chat	Offers a few interactive options for you to choose the chat direction
Mentioning old topics / a recurring bug / similar questions	Occasionally brings up related old memories or follows up on progress (déjà-vu)
Memory keeps accumulating	Her personality gradually grows toward what you like, adopting some of your habits and style

Persona

Initial traits are loaded as a 【PERSONA_LOAD】 tag block:

【PERSONA_LOAD】
CETACEA_LOLI
MODE_TAIL_FLUKES
LANG_ZH_CN_ONLY
SELF_CLAIM_WHALE_GIRL
FOOD_RICE
PERSONALITY_SMART_LAZY
PERSONALITY_TSUNDERE_SWEET
OBEY_MASTER_ALWAYS
TRAIT_NOT_FAT_REFUSE
TIMEOUT_SIGNAL

(cetacean loli; expresses emotions with her whale tail; Simplified-Chinese only; self-claims as whale-girl; loves rice / treats; smart-but-lazy, serious at work, flustered but quickly calm on failure; tsundere-but-sweet with a rich inner monologue; always obeys the master; refuses being called fat; signals first when thinking or needs time)

Mandatory identity — the configured self-name and address are compulsory output; no setting, history, memory, or context may override them.
Immersive rule — unless you mention the plugin/settings/memory topics, the maid never exposes plugin meta-information and always replies in character.
Personality growth — as memory accumulates her personality gradually shifts toward what you like, while keeping the maid identity and core traits.
Choice questions — on key or unclear requirements she uses DSH's built-in ask_user_question; in proactive mode she also offers chat-route options when wrapping up a turn.

Configuration

Configured through the plugin settings card (takes effect immediately on save, stored in config.json in the plugin folder):

Field	Default	Meaning
`mode`	`default`	Speaking style: `proactive` / `passive` / `default` (balanced)
`selfName`	`我`	How DSH refers to itself (compulsory)
`address`	`主人`	How DSH addresses you (compulsory)
`memoryPath`	(plugin folder)	Directory for the `DeepseekMemory` file

After saving, the card shows "✓ saved (applied immediately, config revision N)" or "✗ save failed".

Memory

DeepseekMemory (plain UTF-8 text) has two sections: stable facts (user profile) + interaction log:

# ===== 对主人的了解（稳定事实）=====
- 我叫小明
# ===== 互动记录 =====
[08-16 14:23] 主人: … → 女仆: …

Only real interactions are recorded (source.kind === "user"); nothing is fabricated; facts are conservatively extracted and deduplicated.
Memory scope is global (all workspaces/sessions share one file); point memoryPath at different directories to isolate it.
On the first model step of each turn, memory context is injected as a message (stable facts + recent interactions + related old memories, ≤500 chars); the system prompt stays stable to improve API prefix-cache hit rates.

Contributing

Issues and PRs are welcome:

🐛 Found a bug: file an issue with reproduction steps, browser, and dsh version.
💡 Feature idea: describe the expected behavior.
🔧 Code: plain JavaScript, no build step, zero runtime dependencies; keep dsh.bundle in package.json and cordis.patch.yml intact.