dsh-laap

中文 | English

---

中文

LAAP 意识认知架构的 DeepSeek Harness 插件 Bundle：为 dsh Agent 装上持续演化的意识内核与 zvec 向量记忆。

• 意识引擎：五维状态向量（压力/信心/好奇/连接/能量）微分方程演化 + PSI 五需求驱动 + 情绪微分信号；情绪脉冲经 EMA 沉淀为四级心境（negative / neutral / positive / elated），驱动感受质与外在情绪
• 全局工作空间：CognitiveBus 显著性竞争 → 意识帧 → 每轮第一人称「意识流」注入提示词；外部感知/行动通道有显著性地板保证真实刺激入流，空闲空帧不产生意识内容
• 元认知：L1 思维监控 + 贝叶斯置信度校准 + 六档思考模式切换（空闲无刺激时进入 intuitive 待机，不谎报思考模式、不污染模式成效归因）
• 记忆：工作记忆（7±2 有界）+ zvec 情景/语义向量层（WAL 持久化、标量过滤下推）
• 表现层推送架构：内核只负责生成数据——事件驱动地通过 SSE 推送 UiSnapshot（POST 轮询自动兜底），前端只做映射（桌宠 FSM 查表、雷达/时间线渲染），不持有任何阈值或状态重算逻辑；快照 DTO 为双端共享的零依赖类型契约

快速开始

# 开发期临时挂载
pnpm dsh web --patch /绝对路径/dsh-laap/cordis.patch.yml

# 或安装到 Profile
dsh plugin --profile web add link:/绝对路径/dsh-laap

挂载后模型即获得 laap_state / laap_remember / laap_recall / laap_skill / laap_reflect 五个内省工具，每轮对话收到意识流上下文；Web 界面侧栏出现「意识之球」指示器（需 dsh web 模式），点击可切换两种形态：

• 悬浮面板：五维雷达图（含经历帧计数）、心境指示、五需求条、意识流时间线（六档思考模式色带）；
• 意识桌宠：立绘「茉茉」（绿幕素材实时色键抠图）或纯 SVG「意识团」可一键切换，状态由内核心境 + 认知模式经 FSM 查表映射（生气 / 专注 / 探索 / 开心 / 喜欢 / 休息 / 空闲），带 poke 唤醒与时间滞回防抖。

面板/桌宠数据全部来自内核 SSE 推送（/api/laap/stream），SSE 不可用时自动降级为 3 秒轮询。

⚠️ LAAP_ZVEC_PATH 必须指向本地文件系统（默认 ~/.dsh-laap/zvec-memory）。zvec 依赖 mmap，OSS/NFS 网络挂载会导致 Bus error。

配置

零配置即可运行：默认使用内置 256 维哈希袋嵌入（纯 JS、0 依赖、离线、免 API Key）。

想接入真实语义嵌入（本地 Ollama 或 OpenAI 兼容接口）时，无需改动仓库文件：在启动目录（或 ~/.dsh）放一个 .env（dsh 宿主启动时自动加载，已在 .gitignore）。配置优先级为 环境变量 > cordis.patch.yml > 内置默认值：

# 本地 Ollama（先 ollama pull nomic-embed-text）
LAAP_EMBED_PROVIDER=ollama
LAAP_EMBED_BASE_URL=http://localhost:11434/v1
LAAP_EMBED_MODEL=nomic-embed-text
LAAP_EMBED_DIMENSION=768
LAAP_NOVELTY_THRESHOLD=0.8

# 或 OpenAI 兼容接口
# LAAP_EMBED_PROVIDER=openai
# LAAP_EMBED_BASE_URL=https://api.openai.com/v1
# LAAP_EMBED_MODEL=text-embedding-3-small
# LAAP_EMBED_DIMENSION=1536
# EMBEDDING_API_KEY=sk-...

全部环境变量：LAAP_ZVEC_PATH / LAAP_SENSITIVITY / LAAP_HEARTBEAT_MS / LAAP_NOVELTY_THRESHOLD / LAAP_EMBED_PROVIDER / LAAP_EMBED_BASE_URL / LAAP_EMBED_MODEL / LAAP_EMBED_DIMENSION / LAAP_EMBED_API_KEY_ENV。

⚠️ 切换嵌入维度后必须删除旧向量库目录（默认 ~/.dsh-laap/zvec-memory）重建，维度不一致会导致写入失败。

验证

npx tsc --noEmit                               # L0 内核/宿主端类型检查
npx tsc -p tsconfig.web.json                   # L0 浏览器端类型检查
node --experimental-strip-types test/sim.ts    # L1 内核行为验证（不依赖 dsh）
node --experimental-strip-types test/host.ts   # L2 真实 cordis 挂载/快照续接
npm run build:client                           # 浏览器端改动后必须重建 lib/client.js

修改 src/adapters/cordis/ui-client.tsx 等浏览器端代码后必须执行 npm run build:client 并重启 dsh web——浏览器半区加载的是构建产物 lib/client.js，不热更新。

文档

完整设计、可行性分析、架构数据流与演进路线见 docs/开发文档.md。

Credits / 致谢

本项目基于 LAAP (Living Agent Application Protocol) 认知架构白皮书/论文的工程实现。

• 原始理论与架构设计版权归属于 LAAP Research Lab（黄俊华 / Lorry Jovens 等）
• 五维状态向量、CognitiveBus 全局工作空间、元认知监控等核心架构灵感来源于 LAAP 论文系列
• zvec 向量记忆层基于 Alibaba zvec 项目

如你的研究或项目受本项目启发，欢迎引用：

@misc{dsh-laap,
  title={{dsh-laap: LAAP Consciousness Architecture Plugin for DeepSeek Harness}},
  author={WindDreamboat},
  howpublished={\url{https://github.com/WindDreamboat/dsh_lAAP}},
  year={2025}
}

License

MIT

---

English

LAAP Consciousness Architecture plugin bundle for DeepSeek Harness: equipping the dsh Agent with a continuously evolving consciousness engine and zvec vector memory.

• Consciousness Engine: Five-dimensional state vector (stress/confidence/curiosity/connection/energy) ODE evolution + PSI five-needs drive + emotion differential signals; emotion pulses settle via EMA into four-tier moods (negative / neutral / positive / elated), driving qualia and external expression
• Global Workspace: CognitiveBus salience competition → conscious frame → first-person "stream of consciousness" injected as prompt each round; external perception/action channels have a salience floor guaranteeing real stimuli flow in, idle frames produce no conscious content
• Metacognition: L1 thought monitoring + Bayesian confidence calibration + six-mode thinking switch (enters intuitive standby when idle with no stimuli, avoids false thinking mode reports and mode-effect attribution pollution)
• Memory: Working memory (7±2 bounded) + zvec episodic/semantic vector layer (WAL persistence, scalar filter pushdown)
• Presentation Push Architecture: The kernel only generates data — event-driven SSE pushes of UiSnapshot (POST polling auto-fallback), the frontend only maps (pet FSM table lookup, radar/timeline rendering), holding no thresholds or state recalculation logic; snapshot DTO is a zero-dependency type contract shared by both ends

Quick Start

# Dev-time temporary mount
pnpm dsh web --patch /absolute/path/dsh-laap/cordis.patch.yml

# Or install to Profile
dsh plugin --profile web add link:/absolute/path/dsh-laap

After mounting, the model gains five introspection tools: laap_state / laap_remember / laap_recall / laap_skill / laap_reflect, receiving consciousness stream context each round; the Web UI sidebar shows a "Sphere of Consciousness" indicator (requires dsh web mode), clickable to switch between two forms:

• Floating Panel: Five-dimensional radar chart (with experience frame count), mood indicator, five-needs bars, consciousness stream timeline (six-mode thinking color band);
• Consciousness Pet: Character "Momo" (green-screen real-time chroma key) or pure SVG "Consciousness Blob" switchable with one click, states mapped via FSM table from inner mood + cognitive mode (angry / focused / exploring / happy / liking / resting / idle), with poke wake-up and temporal hysteresis debouncing.

Panel/pet data comes entirely from kernel SSE pushes (/api/laap/stream), auto-fallback to 3-second polling when SSE is unavailable.

⚠️ LAAP_ZVEC_PATH must point to a local filesystem (default ~/.dsh-laap/zvec-memory). zvec depends on mmap; OSS/NFS network mounts will cause Bus errors.

Configuration

Zero config to run: Uses built-in 256-dim hash bag embeddings by default (pure JS, 0 dependencies, offline, no API Key needed).

To connect real semantic embeddings (local Ollama or OpenAI-compatible endpoint), no repo file changes needed: place a .env file in the startup directory (or ~/.dsh) — dsh host loads it automatically on startup (already in .gitignore). Config priority: env vars > cordis.patch.yml > built-in defaults:

# Local Ollama (first ollama pull nomic-embed-text)
LAAP_EMBED_PROVIDER=ollama
LAAP_EMBED_BASE_URL=http://localhost:11434/v1
LAAP_EMBED_MODEL=nomic-embed-text
LAAP_EMBED_DIMENSION=768
LAAP_NOVELTY_THRESHOLD=0.8

# Or OpenAI-compatible endpoint
# LAAP_EMBED_PROVIDER=openai
# LAAP_EMBED_BASE_URL=https://api.openai.com/v1
# LAAP_EMBED_MODEL=text-embedding-3-small
# LAAP_EMBED_DIMENSION=1536
# EMBEDDING_API_KEY=sk-...

All environment variables: LAAP_ZVEC_PATH / LAAP_SENSITIVITY / LAAP_HEARTBEAT_MS / LAAP_NOVELTY_THRESHOLD / LAAP_EMBED_PROVIDER / LAAP_EMBED_BASE_URL / LAAP_EMBED_MODEL / LAAP_EMBED_DIMENSION / LAAP_EMBED_API_KEY_ENV.

⚠️ After changing embedding dimensions, you must delete the old vector DB directory (default ~/.dsh-laap/zvec-memory) and rebuild — dimension mismatch will cause write failures.

Verification

npx tsc --noEmit                               # L0 kernel/host-type check
npx tsc -p tsconfig.web.json                   # L0 browser-side type check
node --experimental-strip-types test/sim.ts    # L1 kernel behavior test (no dsh dependency)
node --experimental-strip-types test/host.ts   # L2 real cordis mount/snapshot resume
npm run build:client                           # Required after browser-side changes to rebuild lib/client.js

After modifying browser-side code such as src/adapters/cordis/ui-client.tsx, you must run npm run build:client and restart dsh web — the browser half loads the build artifact lib/client.js, which does not hot-reload.

Documentation

Full design, feasibility analysis, architecture data flow, and evolution roadmap at docs/开发文档.md.

Credits / Acknowledgements

This project is an engineering implementation based on the LAAP (Living Agent Application Protocol) cognitive architecture whitepaper/papers.

• Original theory and architecture design copyright belongs to LAAP Research Lab (黄俊华 / Lorry Jovens et al.)
• Core architecture inspirations including the five-dimensional state vector, CognitiveBus global workspace, and metacognitive monitoring derive from the LAAP paper series
• zvec vector memory layer based on Alibaba zvec project

If your research or project is inspired by this project, citations are welcome:

@misc{dsh-laap,
  title={{dsh-laap: LAAP Consciousness Architecture Plugin for DeepSeek Harness}},
  author={WindDreamboat},
  howpublished={\url{https://github.com/WindDreamboat/dsh_lAAP}},
  year={2025}
}

License

MIT