One-Sentence Positioning

Provides cross-session long-term memory capability for DeepSeek Harness (DSH), with four-layer storage (L0 meta-rules / L1 index / L2 environmental facts / L3 task experience), namespace isolation by default, and mandatory "action verification" requirement before allowing writes.

Core Capabilities

• Each turn injects L1 existence index into model context, allowing the model to see what reusable experiences it has at any time
• Provides 12 memory tools, covering read, list, write, modify, archive, rollback, statistics, maintenance, candidate confirmation, and trace expansion
• Forces "action verification evidence" during writes, filtering out model guesses and common sense contamination
• Automatically distills each turn's successful tool calls into pending candidates, requiring human or model confirmation before entering formal memory
• Preserves historical snapshots (supersede) on every memory update; archiving doesn't physically delete, allowing one-click rollback
• Automatically isolates namespaces by workspace directory name + git branch, preventing memory mixing across projects
• Automatically runs maintenance every 20 turns (deduplication, L1 compression, statistics, candidate merging)

Technical Implementation

• Language: JavaScript (ESM single file)
• Key Dependencies: @deepseek-ai/cordis (host framework), @deepseek-ai/dsh-tools (defineTool), @deepseek-ai/dsh-system-prompt (L1 injection), @deepseek-ai/schemastery (config validation)
• Architecture Pattern: Injects bundle via cordis.patch.yml; apply hooks into ctx.systemPrompt.context (L1 visible each turn), ctx.skills.register (memory runtime skill), ctx.tools.register (tools), ctx.on('session/event') (turn/end distillation & maintenance), ctx.on('tools/result') (successful tool buffering), ctx.on('agent/disposed') (cleanup)
• Entry File: lib/index.js (1637 lines, single-file implementation)

Use Cases

When you want DSH Agent to retain machine environment, tool configurations, pitfalls encountered, and other reusable experiences across multiple sessions, avoiding starting from scratch each time. After installation, the Agent can proactively consult historical experiences when starting new tasks, and沉淀 "verified successful" knowledge after task completion for reuse. Suitable for developers maintaining the same machine or group of projects long-term.

Prerequisites & Compatibility

Installation

dsh plugin --profile web add dsh-layered-memory

Configuration Options

FAQ

Q: Where is memory stored after installation?

A: Defaults to user home directory at ~/.dsh/memory/ (macOS/Linux uses ~/.dsh/memory, Windows uses %USERPROFILE%\.dsh\memory). When autoNamespace is enabled, subdirectories are automatically created for isolation using workspace name + git branch; customize by setting memoryDir in profile config.

Q: Must evidence be provided when writing memory?

A: Yes. Both memory_write and memory_accept require evidence; lacking evidence causes the tool to reject outright. Evidence should describe the tool call or measured result that verified this information—unverified information is not allowed to enter formal memory.

Q: Tools like memory_read don't appear in the tool list after installation?

A: Uses progressive exposure mechanism by default (progressive=true), where all 12 tools are only mounted to the current Agent after calling memory_activate or after the memory skill loads. Skill loading usually activates automatically; if tools still don't appear, call memory_activate once.

Q: How to rollback mistaken changes or archived memories?

A: Use the memory_rollback tool, passing topic and entry_type (fact/sop), which automatically restores from the latest snapshot in .history/; memory_archive only hides entries from L1 index and memory_read, files remain in archive/, and can also be restored via memory_rollback.

Q: Does uninstalling delete written memories?

A: No. Memory files are stored as regular Markdown/JSON in ~/.dsh/memory/ (or its namespace subdirectories); uninstalling only removes the plugin itself, memory files can still be browsed or reused in other DSH instances; to completely clear, manually delete that directory.

Q: What happens when L1 index exceeds 30 lines?

A: After writing, it prompts over_limit=true. Running memory_index only rebuilds the index; running memory_maintain compresses L1 auto-segments by access heat, keeping only active entries, compressed fact/sop still exist in files but no longer appear in L1 index.

Q: What is automatic distillation of pending candidates?

A: When autoPending=true (default), each turn's end writes successful tool calls to the pending/ candidate area, not automatically entering formal memory. Use memory_pending to view candidates, memory_accept with topic/entry_type/evidence is required before being written to L2/L3.

Difficulty

Beginner — Works with default configuration, install and use; customize namespace, maintenance frequency, or storage location as needed.

Known Issues & Limitations

• autoNamespace depends on git command to read current branch (git branch --show-current); non-git directories automatically degrade to only using cwd directory name, cannot differentiate namespace by branch
• memory_expand requires host to provide sessionQuery service; returns "sessionQuery service unavailable" when not injected, trace information cannot be expanded
• With progressive: true (default), if memory skill loading doesn't automatically activate current Agent's tools, need to manually call memory_activate; setting progressive to false registers all tools at once but loses the Agent isolation advantage
• When L1 index exceeds maxIndexLines, only prompts over_limit=true, no automatic compression, requires manually running memory_index or memory_maintain
• Evidence during write/accept candidate is a string, self-certified by caller; plugin itself doesn't verify evidence actually corresponds to a successful tool call
• Pending candidates are archived to archive/ and original files deleted after successful memory_accept, but buffer doesn't rollback when accept fails

简体中文 | English

Agent_Extensions

📦 DSH 插件已拆分为独立仓库

本仓库 dsh-plugins/ 下的 DSH 插件已迁移为独立 GitHub 仓库，推荐直接安装独立仓库：

插件	独立仓库
dsh-vision-skill	https://github.com/DDDFXYqiming/dsh-vision-skill
dsh-layered-memory	https://github.com/DDDFXYqiming/dsh-layered-memory
dsh-annotation-patched	https://github.com/DDDFXYqiming/dsh-annotation-patched
dsh-side-panel-patched	https://github.com/DDDFXYqiming/dsh-side-panel-patched
dsh-ocr1-memory	https://github.com/DDDFXYqiming/dsh-ocr1-memory

安装示例：

dsh plugin --profile web add github:DDDFXYqiming/dsh-ocr1-memory

本仓库不再维护 dsh 插件，仅保留历史快照与说明；后续更新请以独立仓库为准。

DeepSeek Harness（DSH）插件与 AI Agent 技能集合 —— 面向 DeepSeek Harness 的原生插件（走官方扩展接缝，零框架补丁）+ 跨框架通用 Skill + Hermes 插件，开箱即用。

本仓库收集、翻译并自包含封装 AI Agent 相关的技能资源，并面向 DeepSeek Harness（DSH） 提供标准插件形式的扩展。所有内容均为自包含（技能/插件内自带脚本、模板与文档，不依赖仓库外文件），克隆即可用。

✨ 内容总览

1️⃣ DSH 原生插件（dsh-plugins）—— 零框架补丁

面向 DeepSeek Harness 官方扩展接缝（ctx.skills / ctx.tools / ctx.credentials / ctx.slots / ctx.layout），可随 DSH 版本升级：

2️⃣ 通用技能（General_skills）—— 跨框架可用

任何智能体框架（Claude Code / Codex / opencode / DSH / Hermes 等）均可把目录作为 Skill 挂载：

所有技能均内置 SKILL.md（agent 运行时加载的指令），部分附 scripts/、templates/、references/。

3️⃣ Hermes 插件（hermes_plugins）

📁 目录结构

Agent_Extensions/
├── dsh-plugins/               # DeepSeek Harness（DSH）原生插件
│   ├── dsh-vision-skill/      # 识图插件 v0.4.4（8 工具 + 渐进式暴露 + Credential 化）
│   ├── dsh-layered-memory/    # 分层长期记忆（v0.4：命名空间/自动蒸馏/自动维护）
│   ├── dsh-annotation-patched/ # 选中批注/引用插件（fork 增强，Codex 式选中即引用）
│   ├── dsh-side-panel-patched/ # 右侧工作区面板（fork 增强，多文件 tab + 会话跟踪）
│   └── dsh-ocr1-memory/        # 光学压缩记忆（SoM 图像 + active recall）

├── General_skills/            # 通用智能体技能（跨框架，挂载即用）
│   ├── vision-skill/          # 识图
│   ├── video-notes-generator/ # 视频转结构化笔记
│   ├── ppt-master/            # 文档 → SVG → PPTX
│   ├── markitdown-skill/      # 任意文档 → Markdown
│   └── generic-agent-code-run/ # Windows 桌面/浏览器自动化
├── hermes_plugins/            # Hermes 框架插件
│   └── language-router/       # 语言路由（planner-first）
└── README.md

🚀 快速开始

方式一：安装 DSH 插件（以 dsh-vision-skill 为例）

DSH 插件已拆分为独立仓库，推荐直接安装独立仓库：

# 1. 从独立仓库安装（自带 cordis.patch.yml，自动贡献 id: vision-skill）
dsh plugin --profile web add github:DDDFXYqiming/dsh-vision-skill

# 2. 配置 Credential（$DSH_HOME/.credentials.yaml）
VISION_API_KEY: sk-xxxx

本仓库 dsh-plugins/ 下保留的是历史快照；各插件最新文档见对应独立仓库。

⚠️ bundle 安装后不要在 profile 的 cordis.patch.yml 里再 insert 同名条目，否则会触发 duplicate loader entry id 启动崩溃；需要自定义配置时用裸条目按 id 覆盖（见插件 README）。

方式二：挂载通用技能（任何框架）

以 vision-skill 为例：

# 1. 复制技能目录到你的 agent 的 skills 目录
#    （Claude Code: ~/.claude/skills/ ；Codex: ~/.codex/skills/ ；其他框架见其文档）
cp -r General_skills/vision-skill <你的 skills 目录>/

# 2. 配置视觉模型（OpenAI 兼容接口）
cd General_skills/vision-skill
cp templates/.env.example .env   # 填入 VISION_API_URL / VISION_MODEL / VISION_API_KEY

# 3. 自检
python scripts/vision.py --check

其他技能用法详见各目录内 SKILL.md。

方式三：安装 Hermes 插件

将 hermes_plugins/language-router 目录放入 Hermes 插件目录即可（plugin.yaml 声明了全部 hooks 与版本信息）。

🧩 DSH 插件能力一览

dsh-vision-skill v0.4.4（8 工具 + 1 运行时 skill）

工程化特性：渐进式工具暴露（全局只挂 1 个轻量激活工具，省上下文）、密钥 Credential 化（credential: VISION_API_KEY 引用，每操作解析）、路径围栏（realpath 校验防穿越）、超时与并发门控、严格 JSON Schema 结构化输出。

dsh-layered-memory v0.4（分层长期记忆）

核心特性：命名空间隔离（<memoryDir>/<namespace>/...，默认 workspace 目录 + git 分支）、自动蒸馏（turn/end 成功工具调用 → pending/，确认后才入正式记忆）、自动维护（maintainEveryTurns 默认 20）、渐进式工具暴露（progressive: false 可回退全局注册）。 核心公理：行动验证（No Execution, No Memory）、神圣不可删改（已验证事实可压缩迁移、严禁丢弃）、禁易变状态（时间戳/PID/临时路径不存）、最小充分指针（L1 只写存在性）。注入走 user-role 快照，不破坏 DSH 的 KV 缓存命中率（设计细节见插件 README）。

dsh-annotation-patched（选中批注/引用，fork 增强）

来源：omdsh-dev/dsh-annotation v1.3.13（MIT），全部改动带 PATCH(2026-08-14) 标记，详见目录内 README.md。

dsh-side-panel-patched（右侧工作区面板，fork 增强）

来源：ccq1/dsh-side-panel v0.2.0（BSD-3-Clause），全部改动带 PATCH(2026-08-14) 标记，详见目录内 README.md。

⚙️ 环境要求

❓ FAQ

Q：通用技能和 DSH 插件怎么选？ A：通用技能跨框架，任何 agent 都能挂载；DSH 插件是 DSH 原生扩展，走官方接缝（工具注册 / Credential / 上下文注入 / 布局插槽），能力更强但只适用于 DSH。两者能力互通——dsh-vision-skill 就是把 General_skills/vision-skill 包装成 DSH 原生插件。

Q：我的模型不支持图片，能识图吗？ A：能。直接贴图会被框架拒绝（MODEL_DOES_NOT_SUPPORT_IMAGES），改用两种方式：① 发图片的本地路径文本；② 截图后说"看图"，vision_clipboard 自动保存到工作区再识别。这是纯文本模型的能力门禁，不是插件问题。

Q：视觉 API 用哪家？ A：任意 OpenAI 兼容的多模态模型接口，通过 VISION_API_URL / VISION_MODEL / VISION_API_KEY 注入，不写死任何厂商。

Q：子目录之间有关联吗？ A：没有。每个子目录是独立的自包含单元，可单独使用、单独发布、单独删除。

🤝 贡献

• 每个子目录是独立的自包含单元，欢迎以 PR 提交新技能/插件，或提 Issue 反馈问题。
• 贡献要求：自包含（自带脚本/模板/文档）、许可证清晰（建议 MIT）、不写死密钥与本机绝对路径。
• 新增技能的入口文档统一为 SKILL.md，DSH 插件另附 cordis.patch.yml 与 package.json。

📄 许可

本仓库全部内容以 MIT License 分发。社区来源内容保留原作者署名（见各子目录 SKILL.md / plugin.yaml 头部）。