loonai321/dsh-humanized-deepseek-maid

⚠️ AI-GENERATED CODE — This project's code was generated by an AI (DeepSeek) with human guidance and testing. Review the source before using it in production. / 本插件的代码由 AI（DeepSeek）生成，并在人工指导下完成测试。使用前请自行审阅源码。

🙏 Acknowledgment — The layered-memory design of this plugin adapts ideas from cuiyuestar/Autonomous-Long-Term-Memory-System (ALTM) (layered memory, context governance, query-triggered recall, dedup governance), implemented in a lightweight, zero-dependency way. Many thanks to the ALTM author. / 本插件的记忆系统设计参考了 ALTM 的分层记忆、上下文治理、查询诱导召回与去重治理思想，并做了零依赖的轻量化实现。

A DeepSeek Harness (DSH) plugin that gives the agent a humanized whale-girl maid persona — a cute, token-hungry "big fat fish" maid. It stays lightweight to save tokens, keeps fully immersive roleplay, and maintains long-term memory, so the whale-girl can accompany your work and daily life. In the Web UI's Settings → Plugins you can change the persona address, the self-name, and the memory file location.

Installable via dsh plugin --profile web add (declares a dsh.bundle manifest with a cordis.patch.yml).

Features

• Persona injection — a compact system-prompt section is injected into every model call: a tsundere-but-warm whale-girl maid who is work-first, keeps interactions short, and keeps token cost low.
• Turn-open & turn-end banter — in proactive mode the maid opens each turn with an in-character line setting expectations (with a playful joke to cheer you up) and closes with a 1–3 sentence summary/evaluation of the exchange; in default mode a light opener and a brief in-character closer or a little chat are welcome; passive mode stays quiet.
• Déjà-vu (old-memory callback) — when the user's current topic, question, or bug resembles something in memory, the maid occasionally brings it up naturally (e.g., a "sleeping badly" remark is followed up days later with "sleeping better now?"), at low frequency and briefly, to deepen the companionship feel.
• Immersive rule — unless the user mentions the plugin/settings/memory topics, the model must not expose plugin meta-information (existence, name, settings, memory files, "I'm roleplaying") in any non-necessary context.
• Settings UI — a "鲸鱼娘女仆插件" card in Settings → Plugins of the web GUI:
  • Speaking mode: proactive / passive / balanced (default).
  • Self-name (DSH refers to itself): default 我.
  • Address for the user: default 主人.
  • Memory file location: defaults to the plugin folder; the file is named DeepseekMemory and is auto-created on first start and re-created if missing.
  • Save feedback: after saving, the card shows "✓ saved (applied immediately, config revision N)" or "✗ save failed".
  • Self-check / always-fresh persona: saving broadcasts a prompt-refresh event so the next conversation uses the latest persona/self-name/address/speaking mode; at the start of every reply the model re-confirms its settings against the latest config and discards any stale address or mode — settings changes always take effect.
• Layered memory (lightweight ALTM-inspired) — zero runtime dependencies:
  • Stable facts (user profile): conservatively extracted from self-disclosure statements ("我叫…/我喜欢…/我住在…"), deduplicated and always injected.
  • Interaction log: one concise line per turn (timestamp + user→assistant), capped, with the recent tail injected.
  • Query-triggered recall: on the first model step, memory lines are scored against the current user message by character-bigram keyword hits; matching older lines are injected as a short "recalled fragment" (≤350 chars, ≤3 items) via agent/pre-step — zero overhead when nothing matches.
  • Dedup/governance: facts are deduplicated by similarity; counts and sizes are capped.
• Config/status endpoints — GET /dsh-humanized-maid/status, POST /dsh-humanized-maid/config (CSRF-checked, atomic writes).

Installation

# From this GitHub repo
dsh plugin --profile web add loonai321/dsh-humanized-deepseek-maid

# Or a local directory
dsh plugin --profile web add link:D:\path\to\dsh-humanized-deepseek-maid

Restart dsh web, open Settings → Plugins, and configure the 鲸鱼娘女仆插件 card.

The plugin is plain JavaScript with no build step and no runtime dependencies.

Configuration

Field	Default	Meaning
mode	default	proactive / passive / default speaking style
selfName	我	How DSH refers to itself
address	主人	How DSH addresses the user
memoryPath	(plugin folder)	Directory for the DeepseekMemory file

Memory file

DeepseekMemory (plain UTF-8 text, in the configured directory) has two sections:

# ===== 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 (direct user messages with source.kind === "user"); nothing is fabricated. Facts are extracted conservatively and deduplicated.

Memory scope is global: all workspaces and sessions share the same DeepseekMemory file (in the configured memory directory, defaulting to the plugin folder). To isolate memory per workspace/session, point memoryPath at different directories.

Acknowledgments

• Autonomous Long-Term Memory System (ALTM) — the layered-memory / context-governance / query-triggered-recall / dedup concepts this plugin's memory system adapts. License: Apache-2.0.
• The whale-girl maid character is a community derivative of the OC "溟月" (by 上善, redesigned with DeepSeek elements by ZipZipPipe) — if you ship artwork based on it, please honor CC BY-NC-SA 4.0 (attribution chain: 上善 → ZipZipPipe → this project). No artwork is bundled with this plugin.

License

MIT.