dsh-auto-memory

一句话定位

为 DeepSeek Harness (DSH) Web GUI 提供集中式三层自动记忆：用户级规则、项目笔记与每日日志都由插件自动写入、自动注入、自动检索，并能继承 CodeBuddy / Claude Code / Codex 等其他 AI 工具留下的记忆。

核心能力

• 自动注入记忆上下文：每次组装系统提示词时，把用户规则、项目笔记、最近一天日志、反思精华与未完成日历事项按需拼到末尾，让模型在回复前最后读到记忆纪律
• 每轮对话自动沉淀：每轮对话结束后由子代理判断本轮价值，写入今日日志（标记 [自动沉淀]），长期决策升格到项目笔记、跨项目规则升格到用户级记忆，寒暄轮自动跳过
• 主动日历与提醒：AI 从对话中自动提取 deadline、约定、任务节点写入日历，区分重要/紧急四象限颜色，并在之后会话中赶在时间到达前提醒
• 跨工具记忆继承：自动检测并接入 WorkBuddy / CodeBuddy / Claude Code / Codex 等其他 AI 工具的记忆（只记路径不复制内容，按需读取）
• 可视化与设置：侧边栏「记忆」入口提供 Overview / Logs / Notes / Reflections / Connect / Calendar / Search 七个面板，AI 写时段问候、三级抽屉总结、液态玻璃视觉、记忆地图（v0.1.24）
• 卫生闸门与一键升级：写入端拦截乱码/复读/外部 AI profile 原始 JSON 污染；设置页提供版本检查、npm registry 对比、一键升级命令

技术实现

• 语言: JavaScript（ESM，type: module，源码是 lib/index.js 与 lib/client.js 编译产物，未提供 .ts 源）
• 关键依赖: 零运行时第三方依赖（仅 Node 内置 node:fs/promises、node:fs、node:child_process、node:zlib、node:os、node:path、node:url）；宿主 peer 依赖 @deepseek-ai/cordis ^4.0.1；前端通过 require('react') 与 DSH 客户端运行时共用 React
• 架构模式: 双面（dual-face）cordis 插件——lib/index.js（host 半）注册路由、工具、系统提示词注入、生命周期钩子；lib/client.js（browser 半）通过 cordis.patch.yml 与 package.json#dsh.client.inject 注入到 DSH Web GUI；平台固定为 web
• 入口文件: lib/index.js（导出 apply(ctx, config) 在宿主进程挂载；导出 name = 'auto-memory' 作为稳定 cordis 插件标识），对应 lib/client.js 注入浏览器侧边栏面板

适用场景

日常在 DSH 里持续做项目、跨工作区切换、跨会话回访的开发者。痛点是 DSH 默认不持久化项目记忆、模型每次都从零开始、之前做的决策要靠用户主动复述——这个插件让模型自己记得之前做过什么，并把 deadline、约定、用户偏好沉淀为可检索的本地笔记。还适合同时用 CodeBuddy / Claude Code / Codex 的用户，跨工具复用历史记忆而无需手动搬运。

前置依赖与兼容性

安装方式

dsh plugin --profile web add @a9i5k4/dsh-auto-memory

实际等价步骤：在 DSH 的 web profile 目录 ~/.dsh/profiles/web 下安装包，并把 @a9i5k4/dsh-auto-memory 追加到该目录 package.json 的 dsh.profile.bundles 数组，最后重启 dsh web 才会出现侧边栏「记忆」入口。

配置项

配置文件位于 ~/.dsh/dsh-auto-memory.json（首次启动自动创建），完整默认值由源码 lib/index.js 的 DEFAULT_CONFIG 决定：

修改后无需重启 dsh：设置 → 自动记忆 页面保存即生效；面板字体大小（小/正常/大/特大）实时切换不需保存。

常见问题

Q: 安装后没看到「记忆」侧边栏入口怎么办?

A: 必须三步齐全：在 ~/.dsh/profiles/web 下安装包、把 @a9i5k4/dsh-auto-memory 追加到该目录 package.json 的 dsh.profile.bundles 数组、然后重启 dsh web。只装包不重启或没追加 bundle 行都不显示。

Q: 记忆文件存在哪里？重装 DSH 会丢吗?

A: 记忆以纯文本 Markdown 集中在用户主目录，与 DSH 安装目录隔离：用户级在 ~/.dsh/memory/MEMORY.md，项目记忆统一在 ~/.dsh/memory/workspaces/{workspace}/。重装 DSH 不影响；旧版分散在工作区里的 .dsh-memory 在首次运行时会自动迁移到集中根目录，旧文件保留不删。

Q: 自动沉淀会不会消耗很多 token?

A: 默认开启且每天最多调用 8 次、冷却默认 30 分钟（非工作时间自动翻倍），可在设置页调整。静态规则挂在系统提示词里保持字节级稳定，动态记忆走运行时快照，让 DeepSeek 前缀缓存全程命中以减少重复编码开销。

Q: 能继承 CodeBuddy / Claude Code / Codex 的记忆吗?

A: 能。插件会扫描 ~/.workbuddy、~/.codebuddy、~/.claude、~/.codex 等目录。memory_external 的 import 动作以「纯链接模式」接入，只在本地记忆里记录源文件路径指针，不复制内容，需要时按需读取——这样既能用上外部历史，又不会被其他 AI 工具 profile 的原始 JSON / 乱码污染本地记忆。

Q: 记忆里会泄露密钥吗?

A: 注入到 prompt 的内容会过滤掉疑似令牌 / 密钥 / 凭据的段落，文件本身完整保留；外部会话检索为关键词级（非语义）。但密钥本身仍属敏感信息，建议不要主动写入或粘贴到对话里。

Q: 怎么升级这个插件?

A: 在 ~/.dsh/profiles/web 目录执行 pnpm up @a9i5k4/dsh-auto-memory（或 npm install @a9i5k4/dsh-auto-memory@latest），然后重启 dsh web；设置 → 自动记忆 页面提供「检查更新」按钮，会对比你当前版本与 npm registry 最新版并显示更新命令。注意 pnpm v11 默认限制安装发布不足 1 天的版本，当天发版需要在 pnpm-workspace.yaml 加 minimumReleaseAge: 0 或指定显式版本号。

Q: 卸载后记忆还在吗?

A: 记忆文件保存在用户主目录（~/.dsh/memory/），不在 DSH 安装目录；从 web profile 移除 bundle 并卸载包不会删除这些 Markdown 文件，下次再装可继续使用。

上手难度

入门 — 安装后无需任何配置即可获得默认三层记忆 + 自动沉淀 + 主动日历能力；所有偏好都能在 GUI 的设置 → 自动记忆 页面调整并实时生效，不需要碰 JSON。

已知问题与限制

• 记忆检索是关键词级（非语义）：memory_recall 按空白/中文标点分词后做 OR 匹配，按命中词数排序，源码未提供向量检索或语义匹配
• memory_recall 跨历史 DSH 会话检索依赖宿主部署的 session-query 索引；未启用该索引时只能搜本地记忆文件（README 原文限制）
• 插件增删需要重启 dsh web 才会生效（README 原文限制）
• Node.js < 22.19.0 缺少 zstdDecompressSync，DSH 新版会话持久化的 session.jsonl.zstd 压缩帧无法解压，工作区总览会扫不到压缩会话（v0.1.16 已修复读取逻辑，但旧 Node 仍只能回退明文读）
• pnpm v11 默认不允许安装发布不足 1 天的版本：同一天发布想立刻装新版，需要在 pnpm-workspace.yaml 加 minimumReleaseAge: 0 或使用 pnpm add @a9i5k4/dsh-auto-memory@<version> 显式版本号
• 系统文件夹选择器依赖 DSH 主机的 directory-picker-auto 原生后端；不可用时设置页自动回退到内嵌文件夹浏览
• 0.1.14 及以下版本会击穿 DeepSeek 前缀缓存（旧版把秒级时间戳注入系统提示词），大量消耗 token；如仍在运行老版本，建议升级到 0.1.16+ 解决
• 外部记忆接入（memory_external import）只记录路径指针，不复制内容：好处是抗污染，代价是其他 AI 工具里的内容需要时按需读取，导入后不立即可全文搜索

A cache-friendly three-layer memory engine for the DeepSeek Harness Web GUI — lean auto injection, per-turn AI consolidation, on-demand reads and cross-tool memory inheritance — wrapped in human touches: proactive calendar reminders, warm AI greetings, and a daily journal that writes itself.

DSH Web GUI 的记忆与人性化交互插件：三层记忆自动注入与检索、每日反思、AI 时段问候与三级抽屉、每轮自动沉淀、智能检索、日历视图与设置页，支持继承其他 AI 工具的历史记忆。

Quick install: cd ~/.dsh/profiles/web → pnpm add @a9i5k4/dsh-auto-memory → append "@a9i5k4/dsh-auto-memory" to dsh.profile.bundles in that directory's package.json → restart dsh web (the 「Memory」entry appears in the sidebar). Full instructions in Installation; no-pnpm alternative: npm install @a9i5k4/dsh-auto-memory.

English | 中文版

---

Installation (one command)

Prerequisite: install DeepSeek Harness (dsh) and start dsh web at least once.

Run in the profile directory (~/.dsh/profiles/web):

cd ~/.dsh/profiles/web
pnpm add @a9i5k4/dsh-auto-memory

pnpm v11 默认限制安装发布不足 1 天的版本。当天发布想立即拿到新版,在 profile 目录的 pnpm-workspace.yaml 加 minimumReleaseAge: 0,或用显式版本 pnpm add @a9i5k4/[email protected]。(pnpm v11 blocks packages published <1 day ago; set minimumReleaseAge: 0 in pnpm-workspace.yaml or use an explicit version for same-day releases.)

Then edit package.json in that directory and append to the dsh.profile.bundles array:

"@a9i5k4/dsh-auto-memory"

Restart dsh web to activate (the 「Memory」entry appears in the sidebar).

No pnpm? Use npm: npm install @a9i5k4/dsh-auto-memory

Updating (check & upgrade)

The plugin is a normal npm package, so updating is one command in the same profile directory:

cd ~/.dsh/profiles/web
pnpm up @a9i5k4/dsh-auto-memory   # or: npm install @a9i5k4/dsh-auto-memory@latest

Then restart dsh web.

The Settings → Auto Memory page has a Check for updates button that compares your installed version with the latest one on the npm registry (it shows the update command when a newer version exists).

AI-Era Installation (copy-paste for your AI)

It's the AI era — just copy the sentence below to your AI assistant (DeepSeek / Claude / Codex etc.) and it will do the installation for you.

Install the npm package @a9i5k4/dsh-auto-memory in the DeepSeek Harness web profile
directory ~/.dsh/profiles/web (run "pnpm add @a9i5k4/dsh-auto-memory" or "npm install @a9i5k4/dsh-auto-memory"),
append "@a9i5k4/dsh-auto-memory" to the dsh.profile.bundles array in package.json,
then restart dsh web to activate the plugin.

---

A companion that takes initiative

Auto Memory is designed to feel less like a database and more like an assistant who knows you:

• Proactive reminders — when the AI notices a deadline, appointment or promise in your conversation, it files it into the calendar automatically and reminds you in later sessions before the time arrives (calendar_add / calendar_done / calendar_remove).
• Warm greetings — an AI-written greeting for the current period (morning / afternoon / evening) that mentions your most important work of the day; come back after an hour away and it says "welcome back" and catches you up.
• It writes its own journal — after every conversation turn, a small subagent quietly decides what's worth keeping and appends topic-grouped entries to today's log; long-term decisions are promoted to project notes and cross-project rules to user-level memory. You never have to remember to log anything.
• Daily reflections — structured results / lessons / next steps are kept in the background, so you can review how a day really went.
• Calendar with human touches — semantic colors for urgency, a day timeline (07:00–22:00), location and reminder fields; the AI maintains it from context instead of you filling forms.
• Inherited memory — memories accumulated in WorkBuddy / CodeBuddy / Claude Code / Codex are discoverable and importable, so the plugin picks up where your other AI tools left off.

Under the hood — engineering that stays out of your way

• Prefix-cache friendly — static rules live in the system prompt, dynamic memory in runtime snapshots; the prompt stays byte-stable so DeepSeek's prefix cache keeps hitting (no repeated re-encoding of your whole history).
• Lean injection, low token cost — only the last day of logs + a reflection digest are injected; everything else is fetched on demand via memory_read / memory_recall.
• Rate-limited AI — auto-consolidation runs at most 8×/day with a configurable cooldown (doubled outside work hours), so memory stays useful without burning your budget.
• Credentials never enter the prompt — sensitive sections (tokens / secrets / credentials) are filtered out of injection while staying safe in the files.
• Cross-workspace & cross-tool — centralized storage readable from any session; memories from WorkBuddy / CodeBuddy / Claude Code / Codex are discoverable and importable.

Features

Three-layer Memory

Centralized storage (WorkBuddy-style): all workspace memories live under one root — ~/.dsh/memory/workspaces/, one subdirectory per workspace (readable by any model in any session via injection + cross-workspace memory_recall). Legacy per-workspace .dsh-memory/ folders are auto-migrated on first run after upgrade (the old copies are kept, not deleted).

• Auto injection (at the end of the system prompt): every prompt gets a <memory_system> block (user rules + project notes + reflection digest + recent 1 day of log tails + external memory paths + pending calendar items + writing discipline); it is placed at the very end of the system prompt so the model reads the memory discipline right before replying
• Lean by default, details on demand (v0.1.24): injection stays small — the last 1 day of logs, a reflection digest (achievements section only), and paths for external memory instead of full text. memory_read fetches full log/reflection/user/notes/calendar files when needed; sensitive sections (credentials/tokens/secrets) are filtered out of the prompt but kept in the files
• Visible memory ops: when the AI updates or searches memory, it says so in plain text in the chat reply (e.g. "Logged X to today's journal", "I checked memory and found..."), not hidden inside tool calls

Auto-Consolidation — memory writes itself after every turn (v0.1.9)

Every finished conversation turn is automatically evaluated (via a small subagent) and anything worth keeping is written for you — no reliance on the model remembering to log:

• Today's log gets entries like - 21:03 [自动沉淀] … — no manual memory_log needed for routine work
• Long-term value is promoted: project decisions/architecture → project notes (with a ## YYYY-MM-DD heading); cross-project rules → user-level memory
• Small talk is skipped (content threshold autoConsolidateMinChars), each turn is deduplicated by turn number, subagent turns are ignored
• Agent traces in the GUI: the overview shows "Auto-consolidated N points today (latest HH:MM)"; the panel refreshes on open, every 30s while open, and via the ⟳ button
• memory_consolidate tool: read recent logs and distill long-term decisions / architecture / user preferences into MEMORY.md on demand ("dream-like" consolidation)
• Configurable in ~/.dsh/dsh-auto-memory.json: autoConsolidate (default true), autoConsolidateMinChars (default 240), autoConsolidateCooldownMinutes (default 30, doubled 22:00–08:00), autoConsolidateDailyMax (default 8) — all adjustable in the Settings → Automation section

AI Greetings & Three-level Drawers (Overview page, v0.1.9)

The first thing you see when opening the memory panel is an AI-generated period greeting, not a template and not technical info:

• AI-written greeting: a subagent writes a warm, casual greeting for the current period (morning / forenoon / noon / afternoon / evening), mentioning your most important work of the day; generated once per period per day and cached in .dsh-memory/greetings/ — no repeated API cost
• Drawer titles are the AI summaries: the "Today afternoon / Today evening" window titles are replaced with the AI's casual summary text itself
• Three-level drawer structure:
  • Level 1: period drawer, titled with the AI summary
  • Level 2: inside it, small drawers — one per work item the AI distilled (with a point count)
  • Level 3: expand a work item to read its detail points
• Summaries are cached: structured results live in .dsh-memory/summaries/; opening the panel reads the cache (offline-friendly, no regeneration); the ⟳ refresh button or returning after >1h away forces a fresh generation; every summary shows its generation time
• Smart timing: if you were away for more than 1 hour and come back, the greeting says "Welcome back" with what was finished meanwhile
• Daily reflection stays in background: structured reflections (results / lessons / next steps) are kept, while the front page only shows a light greeting

Smart Search (Search tab, v0.1.9)

The Search tab adds a Smart search button next to the keyword search:

• The AI expands your natural-language query into 3-6 keywords (e.g. "last time publishing npm hit a snag" → 发布 / 踩坑 / GitHub / npm / 推送)
• Scans all three memory layers plus reflections with those keywords
• The AI then composes a conversational answer in natural language, citing where each fact came from (log date / project note / user-level memory) — it never fabricates facts not present in memory
• Raw keyword hits with their sources are listed under the answer

Calendar View (Four Quadrants)

「Calendar」tab (liquid-glass monthly view):

• Monthly grid, today highlighted, click any date to add an item
• Four-quadrant colors: Urgent & Important (red) / Important (blue) / Urgent (orange) / Neither (gray)
• Click an item to toggle done, click again to delete; legend + weekday header
• Cross-conversation persistence: data lives at user level ~/.dsh/memory/CALENDAR.md, shared across workspaces, survives DSH reinstall
• AI-maintained: the AI extracts deadlines and appointments from conversations and writes them to the calendar automatically (calendar_add / calendar_list / calendar_done / calendar_remove), restating it in plain text; pending items are injected into every session's system prompt

Agent Tools

memory_log / memory_note / memory_user / memory_recall / memory_external / memory_maintain / memory_status / memory_reflect / memory_consolidate / calendar_add / calendar_list / calendar_done / calendar_remove

UI

• Sidebar 「Memory」entry → floating panel (Overview / Logs / Notes / Reflections / Connect / Calendar / Search)
• Settings page (Settings → Auto Memory): storage paths, injection budget, reflection style, UI language (中文 / English), panel font size (Small / Normal / Large / Extra large, default Large) — applies immediately, no save needed
• External memory inheritance: import memories accumulated by other AI tools (CodeBuddy / Claude Code / Codex / project convention files)

---

Screenshots

All screenshots below are real captures of the plugin running inside the DSH Web GUI (Chinese UI for now; English UI captures will be added later).

Memory panel overview — away greeting & AI period summaries

Connect — inherit global memory & history sessions from other AI agents

Calendar — AI adds items, toggles status and marks completion from context

Workspace mind map — auto-generated, workspace-centered cross-workspace overview

Settings — highly customizable, covers most technical details

Beyond the screenshots

• Auto-consolidation: every finished turn is evaluated by a small subagent and topic-grouped entries are written to today's log automatically (## 主题（HH:MM） + bullet points) — no memory_log needed for routine work. Long-term value is promoted to project notes / user-level memory, small talk is skipped, failures are queued and retried every 5 minutes (a 15-second heartbeat file proves the loop is alive).
• Smart search: ask in natural language — the AI expands your query into keywords, scans every memory layer, then answers conversationally with sources cited.
• Calendar with day timeline (v0.1.24): click a date to open a 07:00–22:00 timeline with events placed on their time slots; add events with location/reminder/details; the AI proactively files deadlines via calendar_add and marks them done via calendar_done/calendar_remove.
• Workspace mind map (v0.1.24): an AI-generated graph of workspaces, topics and cross-workspace links; pan by dragging, resize with the slider, click a card for details.
• Memory panel UI (v0.1.24): liquid-glass panel with smooth animations everywhere — tab strip with arrow scrolling, expanding sections (greeting drawers, workspace summaries), floating save bar in Settings that highlights when there are unsaved changes.
• External memory management (v0.1.24): per-source view/import/remove — imported sources show "✓" and become "Delete" buttons per target (notes vs user-level); full content is read on demand and never injected in bulk.
• Calendar reminders: pending items are injected into future sessions' system prompts until completed — the AI reminds you without being asked.
• One-click update: the settings page shows your installed version vs. the npm registry latest; registry installs get a one-click update button (pnpm/npm runs under the hood), then restart to apply.

---

Configuration

Defaults (JSON file ~/.dsh/dsh-auto-memory.json):

{
  "userMemoryDir": "~/.dsh/memory",
  "projectMemoryDir": ".dsh-memory",
  "injectEnabled": true,
  "injectBudgetChars": 2400,
  "recentDaysInjected": 1,
  "reflectEnabled": true,
  "reflectStyle": "auto",
  "locale": "zh",
  "autoConsolidate": true,
  "autoConsolidateMinChars": 240,
  "autoConsolidateCooldownMinutes": 30,
  "autoConsolidateDailyMax": 8,
  "externalInjectionChars": 1400,
  "memoryRoot": "~/.dsh/memory/workspaces",
  "dayBoundaryMinutes": 450
}

Adjustable in the GUI (Settings → Auto Memory), including the UI language (zh / en), the panel font size and the day boundary.

v0.1.27 hardening (memory hygiene gate — keeps external dirt out)

• Write gate on the three write tools: memory_log_dev / memory_note_dev / memory_user_dev now run content through sanitizeForWrite before writing — suspected mojibake (GBK round-trip artifacts), stutter degeneration (word/char loops, punctuation-separated included) and consecutive duplicate lines (≥3) are rejected with a reason; append entries cap at 8,000 chars, full rewrites (replace) at 200,000; appends are also deduped against the file's last ~60 lines (the - HH:MM log prefix does not affect matching).
• Hygiene family fixes: removed the 进行中 false positive from the mojibake pattern (added the real artifact 杩涜涓), switched mojibake counting to global matching (it previously counted at most 1 hit, letting dirty long texts slip past the threshold), stutter detection now works across punctuation, and the duplicate-line check is reset by blank lines (no more false positives on short lines repeated across paragraphs).
• External memory import is link-only: memory_external_dev import now records only source file path pointers instead of copying content — keeping dirty content from other AI tools (WorkBuddy/CodeBuddy/Claude Code/Codex) out of local memory.
• Injection-side scrub + voice discipline: mojibake lines, code blocks and stutter lines are scrubbed before injection; the injected block adds a "how to read memory" note and voice discipline — write all memory entries as third-person objective statements, no first-person thinking narration.

v0.1.28 dirty-token checker (prion-scan integrated — stops the "model silently degrades" class)

• Write gate now rejects raw JSON envelopes & base64 residue: besides mojibake / stutter / duplicate lines, the write gate also rejects lines matching external-AI-profile JSON signatures (memoryBlock / "uid": / updatedAt / "role":"... ") and base64 residue lines — so an external tool profile can never be pasted into memory wholesale again.
• Full mojibake table (34 features, verbatim from prion-scan.mjs): the GBK round-trip residue list is completed, catching the exact artifacts this class of "garbage token" shows up as.
• New "Scan dirty tokens" in Settings → Debug Center: one-click prion-style scan of user-level memory / project notes / today's log / reflections — returns a per-file report by line range (mojibake / raw JSON envelope / long lines >500 / base64 / duplicate lines, duplicate ## headings) with locations only, no content, so you can spot residual dirt without reading it.

v0.1.9 hardening (budget / boundary / picker)

• Daily write budget with auto-compaction: user memory ≤ 4000 chars/day, project notes ≤ 3000 chars/day (shared across sessions, reset at the day boundary). Going over the budget never rejects the write — the framework compacts the pre-today sections with an AI pass (merge duplicates, drop stale entries, keep hard facts) and then writes; if AI is unavailable, the oldest sections are archived to archived-user.md / archive/notes-archived.md (nothing is lost). Compaction is throttled to once per 10 minutes.
• Day boundary (late-night belongs to yesterday): dayBoundaryMinutes (default 450 = 07:30). Work logged before the boundary is appended to the previous day's log, and the daily reflection for the previous day starts only after the boundary — no more "it's 00:30, tell me what you did yesterday" right after midnight.
• Native OS folder picker: the "Browse…" button next to the memory root opens the real system folder picker (via the DSH directory-picker native backend); falls back to the in-app browser when no native picker is available. Changing the root auto-migrates existing workspace memory folders to the new location (old files are kept) and all path variables follow the new config on the next refresh.
• 30-day distillation: memory_maintain distills logs older than 30 days with an AI pass into the project notes, archives the originals under archive/, and removes them from the active log list.
• First-turn injection guarantee: a pre-step hook awaits the memory state refresh before the first step, so the model sees memory from the very first token (previously the async load could leave the first turn empty).
• Per-step reminder with timestamp: the injected discipline block carries a live HH:MM:SS timestamp that refreshes on every prompt assembly, and a 15-second heartbeat file proves the background loop is alive.

---

Structure

• lib/index.js — Host half: engine, injection, tools, routes (zero runtime deps, Node built-ins only)
• lib/client.js — Browser half: memory panel (with calendar view) + settings page (built-in zh/en i18n)
• cordis.patch.yml — Plugin row (auto-memory)

---

Limitations

• Memory files are plain-text Markdown; no secrets stored unless explicitly requested.
• memory_recall session search depends on the deployed session-query index; without it, only local search works.
• Plugin-set changes require a dsh restart.

---

Release Info

• GitHub: https://github.com/Aik358/dsh-auto-memory
• npm: @a9i5k4/dsh-auto-memory
• License: BSD-3-Clause