dsh-plugin-tool-management 0.1.2 → 0.1.3

package/README.md

@@ -3,24 +3,36 @@
 [![npm version](https://img.shields.io/npm/v/dsh-plugin-tool-management?logo=npm&color=cb3837)](https://www.npmjs.com/package/dsh-plugin-tool-management)
 [![License](https://img.shields.io/badge/license-MIT-blue)](LICENSE)
 [![Node](https://img.shields.io/badge/node-%3E%3D18-339933?logo=node.js)](package.json)
-[![GitHub](https://img.shields.io/badge/GitHub-ouli--1242%2Fdsh--plugin--tool--management-181717?logo=github)](https://github.com/ouli-1242/dsh-plugin-tool-management)
+[![GitHub](https://img.shields.io/badge/GitHub-ouli--1242%2Fdsh--plugin--tool--management-181717?logo=github)](https://github.com/ouli-1242/dsh-plugin-tool-management)
+
+**简体中文** · [English](README_EN.md)
 
-**DeepSeek Harness 的 MCP 服务与技能管理插件。** 一个设置面板同时管好两件事：
+**DeepSeek Harness 的 MCP 服务与技能管理插件。** 一个设置面板管好四件事：
 
 - **MCP**：连接了哪些服务、每个服务有哪些工具、哪些工具该让模型用——增删改查、启停、重启，全部即改即生效；
-- **Skills**：本机各处的技能（DSH / Agents / Codex / Claude / 项目级 / 你自己指定的任意目录）一目了然，逐个或整组启停、创建、导入、回收。
+- **Skills**：本机各处的技能（DSH / Agents / Codex / Claude / 项目级 / 你自己指定的任意目录）一目了然，逐个或整组启停、创建、导入、回收；
+- **AGENTS.md**：管理多套全局指令基线预设，一键「应用」写入 `~/.dsh/AGENTS.md`，新会话生效、当前会话不变；
+- **History**：已归档会话统一管理，按项目分组、批量恢复/删除，对话导入/导出，保留期自动清理。
 
 不手改 `cordis.patch.yml`，不碰任何技能源文件，重启与升级后配置依旧。
 
 ---
 
-<!-- 图片占位 1：MCP 管理页截图 → docs/images/mcp-page.png -->
+<!-- 图片占位 1：MCP 管理页截图 → docs/images/mcp.png -->
 
-![MCP 管理](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/mcp-page.png)
+![MCP 管理](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/mcp.png)
 
-<!-- 图片占位 2：Skills 管理页截图 → docs/images/skills-page.png -->
+<!-- 图片占位 2：Skills 管理页截图 → docs/images/skills.png -->
 
-![Skills 管理](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/skills-page.png)
+![Skills 管理](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/skills.png)
+
+<!-- 图片占位 3：AGENTS.md 预设页截图 → docs/images/agents-md.png -->
+
+![AGENTS.md 预设](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/agents-md.png)
+
+<!-- 图片占位 4：History 归档会话页截图 → docs/images/history.png -->
+
+![History 归档会话](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/history.png)
 
 ## 核心亮点
 
@@ -34,9 +46,12 @@
 | 技能来源 | 接入 `~/.agents` / `~/.codex` / `~/.claude` 三个官方不加载的技能目录，并支持**自定义任意技能目录**（只读接入、重叠拒绝） |
 | 技能操作 | 创建技能、ZIP/文件夹导入、插件回收站（恢复 / 永久删除 / 系统回收站兜底）、系统编辑器打开源文件 |
 | 即时刷新 | 技能目录由后台线程监听，编辑器里改完技能页面自动刷新 |
-| 斜杠命令 | 聊天框直接输入 `/mcp`、`/skills` 查看状态 |
+| AGENTS.md 预设 | 多套全局指令基线预设库：新建 / 导入 / 编辑 / 应用 / 删除；「应用」写入 `~/.dsh/AGENTS.md`（新会话生效，当前会话不变） |
+| 会话归档管理 | History 页按项目分组展示已归档会话：搜索、全选、批量恢复 / 永久删除、保留期自动清理（改保留期后倒计时以修改时间为基准重置） |
+| 对话导入 / 导出 | 从 Claude Code / Cursor（JSONL）、Codex（Markdown）、任意文本无痛接管对话；导出可选会话范围，目录默认桌面，支持 Markdown / JSONL |
+| 斜杠命令 | 聊天框直接输入 `/mcp`、`/skills`、`/agents-md` 查看状态 |
 | 模型工具 | **7 个**：`skill_mcp_manager_*` 管 MCP，`skill_manager_*` 管技能（创建前需用户确认） |
-| 界面 | 独立的 `dsm-*` 设计系统，两页风格统一 |
+| 界面 | 独立的 `dsm-*` 设计系统，四页风格统一 |
 
 ## 快速开始
 
@@ -51,7 +66,7 @@ dsh plugin --profile web add dsh-plugin-tool-management@latest
 dsh plugin --profile web remove dsh-plugin-tool-management
 ```
 
-装完硬刷新浏览器（Cmd/Ctrl+Shift-R），设置里出现 **MCP** 与 **Skills** 两页即安装成功（客户端改动由 DSH 热加载，无需重启）。
+装完硬刷新浏览器（Cmd/Ctrl+Shift-R），设置里出现 **MCP**、**Skills**、**AGENTS.md** 与 **History** 四页即安装成功（客户端改动由 DSH 热加载，无需重启）。
 
 也可以直接对任意 DSH 会话说：
 
@@ -76,13 +91,26 @@ dsh plugin --profile web add dsh-plugin-tool-management@latest
 - **看全貌**：按来源分组列出所有技能——项目级、运行时、内置、插件自带，以及四个用户目录（`~/.dsh` / `~/.agents` / `~/.codex` / `~/.claude`，后三个由本插件接入）和你自己添加的自定义目录。
 - **启停**：单个技能、整个来源、整个项目，随时切换；实现是 override provider 的遮蔽策略，源文件一个字节都不动，换机或重装只要复制状态文件。
 - **自定义目录**：点「添加目录」输入绝对路径，该目录即成为只读技能来源——适合管理散落在仓库、网盘同步目录里的技能合集；与已有来源重叠的路径会被拒绝，避免遮蔽失效。
-- **创建与导入**：表单直接创建；ZIP、`.md`、技能文件夹拖进来就能装；删除先进回收站，可恢复，永久删除前还会尝试移入系统回收站兜底。
+- **创建与导入**：表单直接创建；ZIP、`.md`、技能文件夹拖进来就能装；删除先进回收站，可恢复，永久删除前还会尝试移入系统回收站兜底。
+
+### 管 AGENTS.md 预设
+
+- **预设库**：新建、导入、编辑多套全局指令基线（如不同团队的 coding standard 或不同角色的行为规范）。
+- **应用即写入**：「应用」把选中预设写入 `~/.dsh/AGENTS.md`，**新会话生效、当前会话不变**；编辑后点「重新应用」即可同步最新内容，删除前请先切换到其他预设。
+
+### 管历史会话
+
+- **按项目分组**：归档会话按工作区自动分组展示，搜索标题 / 会话 ID / 项目路径快速定位；工作区目录已不存在的会话会打上 ⚠ 标记。
+- **批量操作**：「全选」后批量恢复或永久删除；恢复的会话回到工作区列表，删除会连同其子代理子会话一并清理。
+- **保留期**：顶部下拉选择自动清理周期（0 = 永久保留），到期自动清除；修改保留期后，所有会话的倒计时以修改时间为基准重新计算。
+- **导入对话**：无痛接管其他工具的会话——Claude Code / Cursor 的 JSONL、Codex 的 Markdown、以及任意文本格式，导入后即可继续对话。
+- **导出对话**：按会话范围（全部 / 仅归档 / 按工作区）导出，每个会话一个 Markdown 或 JSONL 文件；导出目录默认桌面，旁边带「选择文件夹」按钮弹出目录树，逐级浏览选中后自动回填绝对路径。
 
 ### 让模型和脚本参与管理
 
 | 入口 | 能做什么 |
 |---|---|
-| `/mcp`、`/skills` | 聊天框查看当前状态 |
+| `/mcp`、`/skills`、`/agents-md` | 聊天框查看当前状态 |
 | `skill_mcp_manager_list / set_enabled / restart / add` | 模型查询与操作 MCP 服务 |
 | `skill_manager_list / set_enabled / create` | 模型查询与操作技能（创建前会征求你同意） |
 | `POST /dsh-plugin-tool-management/api` | 脚本调用的 HTTP API（`{op, args}` 协议） |
@@ -106,6 +134,8 @@ dsh plugin --profile web add dsh-plugin-tool-management@latest
 | 服务器备注 / 页面设置 / 工具停用列表 / 导出 | DSH 主目录下的旁路 JSON（`skill-mcp-manager-*.json`） |
 | 技能启停策略 / 自定义目录 | `~/.dsh/tool-management/state.json` |
 | 技能回收站 / 导入暂存 | `~/.dsh/tool-management/trash`、`uploads` |
+| AGENTS.md 预设库 / 应用结果 | 插件目录 `data/agents-md-presets/`；「应用」写入 `~/.dsh/AGENTS.md` |
+| 归档会话账本 / 保留期 | 插件目录 `data/history-archived-at.json`、`data/history-retention.json` |
 | 运行日志 | `~/.dsh/dsh-plugin-tool-management.log`（滚动） |
 
 ## 常见问题
@@ -127,7 +157,7 @@ npm run test:fast    # 跳过构建直接跑测试
 npm run build        # 仅构建（tsc + 同步客户端 bundle）
 ```
 
-结构：宿主端 `src/index.ts`（Cordis 对象插件，`lib/index.js` 为发布产物）；技能核心 `src/skills/core.js`（纯 Node，可独立单测）；浏览器端 `src/client.js`（ModuleLoader CJS bundle，`dsm-*` 设计系统，经同源 API 与宿主通信）。运行时依赖仅 `fflate`（ZIP 解压）。
+结构：宿主端 `src/index.ts`（Cordis 对象插件，`lib/index.js` 为发布产物）；技能核心 `src/skills/core.js`（纯 Node，可独立单测）；AGENTS.md 预设库 `src/agents-md/service.ts`；归档会话管理 `lib/history/`（`workspace.js` / `projcache.js` / `tombstone.js`）；对话导入解析 `src/imports/parsers.js`；浏览器端 `src/client.js`（ModuleLoader CJS bundle，`dsm-*` 设计系统，经同源 API 与宿主通信）。运行时依赖仅 `fflate`（ZIP 解压）。
 
 发布：`npm version patch && npm publish`（`prepublishOnly` 自动构建）。
 

package/README_EN.md

@@ -3,24 +3,36 @@
 [![npm version](https://img.shields.io/npm/v/dsh-plugin-tool-management?logo=npm&color=cb3837)](https://www.npmjs.com/package/dsh-plugin-tool-management)
 [![License](https://img.shields.io/badge/license-MIT-blue)](LICENSE)
 [![Node](https://img.shields.io/badge/node-%3E%3D18-339933?logo=node.js)](package.json)
-[![GitHub](https://img.shields.io/badge/GitHub-ouli--1242%2Fdsh--plugin--tool--management-181717?logo=github)](https://github.com/ouli-1242/dsh-plugin-tool-management)
+[![GitHub](https://img.shields.io/badge/GitHub-ouli--1242%2Fdsh--plugin--tool--management-181717?logo=github)](https://github.com/ouli-1242/dsh-plugin-tool-management)
+
+[简体中文](README.md) · **English**
 
-**An MCP server & skills manager for DeepSeek Harness.** One settings panel keeps two things under control:
+**An MCP server & skills manager for DeepSeek Harness.** One settings panel keeps four things under control:
 
 - **MCP**: which servers are configured, what tools each one exposes, and which tools the model may call — add, edit, remove, toggle, restart; every change takes effect immediately;
-- **Skills**: every skill on the machine (DSH / Agents / Codex / Claude / project-level / any directory you add) at a glance — toggle individually or per source, create, import, recycle.
+- **Skills**: every skill on the machine (DSH / Agents / Codex / Claude / project-level / any directory you add) at a glance — toggle individually or per source, create, import, recycle;
+- **AGENTS.md**: keep multiple global instruction baselines as presets, apply one with a click to write `~/.dsh/AGENTS.md` — new sessions pick it up, current sessions stay unchanged;
+- **History**: archived sessions in one place — grouped by project, batch restore / delete, import & export transcripts, retention-based auto-cleanup.
 
 No hand-editing of `cordis.patch.yml`, and skill source files are never touched. Configuration survives restarts and upgrades.
 
 ---
 
-<!-- Image slot 1: MCP management page screenshot → docs/images/mcp-page.png -->
+<!-- Image slot 1: MCP management page screenshot → docs/images/mcp.png -->
 
-![MCP management](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/mcp-page.png)
+![MCP management](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/mcp.png)
 
-<!-- Image slot 2: Skills management page screenshot → docs/images/skills-page.png -->
+<!-- Image slot 2: Skills management page screenshot → docs/images/skills.png -->
 
-![Skills management](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/skills-page.png)
+![Skills management](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/skills.png)
+
+<!-- Image slot 3: AGENTS.md presets page screenshot → docs/images/agents-md.png -->
+
+![AGENTS.md presets](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/agents-md.png)
+
+<!-- Image slot 4: History archived sessions page screenshot → docs/images/history.png -->
+
+![History archived sessions](https://raw.githubusercontent.com/ouli-1242/dsh-plugin-tool-management/main/docs/images/history.png)
 
 ## Highlights
 
@@ -34,9 +46,12 @@ No hand-editing of `cordis.patch.yml`, and skill source files are never touched.
 | Skill sources | Hooks up `~/.agents` / `~/.codex` / `~/.claude` (three directories official DSH does not load) plus **any custom skill directory** you add (read-only, overlapping paths rejected) |
 | Skill operations | Create skills, import ZIP / folders, plugin recycle bin (restore / permanent delete with OS-trash fallback), open the source file in the system editor |
 | Live refresh | Skill directories are watched from a background thread — edits made in an editor show up automatically |
-| Slash commands | `/mcp` and `/skills` right from the chat box |
+| AGENTS.md presets | Multiple global instruction baselines as presets — create / import / edit / apply / delete; "Apply" writes `~/.dsh/AGENTS.md` (new sessions pick it up, current sessions stay unchanged) |
+| Archived session management | History page groups archived sessions by project: search, select-all, batch restore / permanent delete, retention-based auto-cleanup (changing the retention resets the countdown from the change time) |
+| Transcript import / export | Seamlessly take over conversations from Claude Code / Cursor (JSONL), Codex (Markdown), or any text; export picks the session scope, defaults to the desktop, in Markdown / JSONL |
+| Slash commands | `/mcp`, `/skills`, `/agents-md` right from the chat box |
 | Model tools | **7 tools**: `skill_mcp_manager_*` for MCP servers, `skill_manager_*` for skills (creating asks for user confirmation first) |
-| UI | Its own `dsm-*` design system, consistent across both pages |
+| UI | Its own `dsm-*` design system, consistent across all four pages |
 
 ## Getting started
 
@@ -51,7 +66,7 @@ dsh plugin --profile web add dsh-plugin-tool-management@latest
 dsh plugin --profile web remove dsh-plugin-tool-management
 ```
 
-Hard-refresh the browser (Cmd/Ctrl+Shift-R) after installing — the **MCP** and **Skills** pages appear in Settings (client changes are hot-loaded by DSH, no restart needed).
+Hard-refresh the browser (Cmd/Ctrl+Shift-R) after installing — the **MCP**, **Skills**, **AGENTS.md** and **History** pages appear in Settings (client changes are hot-loaded by DSH, no restart needed).
 
 You can also tell any DSH session:
 
@@ -76,13 +91,26 @@ Then remind me to hard-refresh the browser.
 - **See everything**: skills are grouped by source — project, runtime, built-in, plugin-shipped, the four user directories (`~/.dsh` / `~/.agents` / `~/.codex` / `~/.claude`; the last three are hooked up by this plugin) and any custom directories you added.
 - **Toggle**: individual skills, whole sources or whole projects — implemented as an override-provider shadow policy, so not a single byte of the source file changes; moving machines is just copying the state file.
 - **Custom directories**: click "Add directory", enter an absolute path, and that directory becomes a read-only skill source — ideal for skill collections living in repos or synced folders; overlapping paths are rejected to keep the shadow policy sound.
-- **Create / import / recycle**: create from a form; drag in a ZIP, a `.md` file or a skill folder; deleted skills go to the plugin recycle bin first, and permanent delete still tries the OS trash as a last safety net.
+- **Create / import / recycle**: create from a form; drag in a ZIP, a `.md` file or a skill folder; deleted skills go to the plugin recycle bin first, and permanent delete still tries the OS trash as a last safety net.
+
+### Managing AGENTS.md presets
+
+- **Preset library**: create, import and edit multiple global instruction baselines (e.g. different teams' coding standards or role behaviors).
+- **Apply = write**: "Apply" writes the selected preset to `~/.dsh/AGENTS.md` — **new sessions pick it up, current sessions stay unchanged**; "Re-apply" syncs the latest content after editing; switch to another preset before deleting.
+
+### Managing archived sessions
+
+- **Grouped by project**: archived sessions are grouped by workspace automatically; search by title / session ID / project path; sessions whose workspace folder no longer exists are flagged with ⚠.
+- **Batch operations**: "Select all" then batch-restore or permanently delete; restored sessions return to the workspace list, and deletion cascades to their subagent sessions.
+- **Retention**: pick the cleanup period from the dropdown (0 = keep forever); the expiry baseline is the later of the archive time and the last retention change, so changing the retention resets the countdown.
+- **Import conversations**: take over sessions from other tools — Claude Code / Cursor JSONL, Codex Markdown, and arbitrary text — and keep chatting right after import.
+- **Export conversations**: pick a session scope (all / archived only / by workspace); each session becomes a Markdown or JSONL file; the export directory defaults to the desktop, and the adjacent "Select" button opens a directory tree to browse and fill in the absolute path.
 
 ### Let the model and scripts help
 
 | Entry point | What it does |
 |---|---|
-| `/mcp`, `/skills` | Check the current state from the chat box |
+| `/mcp`, `/skills`, `/agents-md` | Check the current state from the chat box |
 | `skill_mcp_manager_list / set_enabled / restart / add` | Let the model query and operate MCP servers |
 | `skill_manager_list / set_enabled / create` | Let the model query and operate skills (creating asks for your consent) |
 | `POST /dsh-plugin-tool-management/api` | HTTP API for scripts (`{op, args}` protocol) |
@@ -106,6 +134,8 @@ Why a token: the cross-site protection (POST-only + custom header + same-origin
 | Server notes / page settings / disabled tools / export | Sidecar JSON files under the DSH home (`dsh-plugin-tool-management-*.json`) |
 | Skill toggle policy / custom directories | `~/.dsh/tool-management/state.json` |
 | Skill recycle bin / import staging | `~/.dsh/tool-management/trash`, `uploads` |
+| AGENTS.md presets / applied file | Plugin dir `data/agents-md-presets/`; "Apply" writes `~/.dsh/AGENTS.md` |
+| Archived session ledger / retention | Plugin dir `data/history-archived-at.json`, `data/history-retention.json` |
 | Runtime log | `~/.dsh/dsh-plugin-tool-management.log` (rolling) |
 
 ## FAQ
@@ -127,7 +157,7 @@ npm run test:fast    # run tests without building
 npm run build        # build only (tsc + sync client bundle)
 ```
 
-Layout: host half `src/index.ts` (object-form Cordis plugin, `lib/index.js` is the shipped artifact); skill core `src/skills/core.js` (pure Node, unit-testable); browser half `src/client.js` (ModuleLoader CJS bundle, `dsm-*` design system, talks to the host through the same-origin API). The only runtime dependency is `fflate` (ZIP extraction).
+Layout: host half `src/index.ts` (object-form Cordis plugin, `lib/index.js` is the shipped artifact); skill core `src/skills/core.js` (pure Node, unit-testable); AGENTS.md presets `src/agents-md/service.ts`; archived session management `lib/history/` (`workspace.js` / `projcache.js` / `tombstone.js`); transcript import parsing `src/imports/parsers.js`; browser half `src/client.js` (ModuleLoader CJS bundle, `dsm-*` design system, talks to the host through the same-origin API). The only runtime dependency is `fflate` (ZIP extraction).
 
 Publish: `npm version patch && npm publish` (`prepublishOnly` builds automatically).
 

package/cordis.patch.yml

@@ -25,7 +25,25 @@
 # short-circuits on every unrelated entry (e.g. dsh-better-sidebar's own
 # guard), so `!e.disabled` is only ever read on a genuinely matching
 # legacy/aggregate row (a plain boolean, no re-entry).
+# 折叠 dsh-archive-manager：替换官方 workspace 与 session-projection-cache
+# 服务为归档感知子类（永久删除 + unarchive + 批量 + archivedAt 保留期账本）。
+# 官方 ui-workspace 保持启用；归档菜单仍由官方 UI 提供，删除/恢复/保留期
+# 由本插件的 TOOLS → History 页面经 HTTP API 驱动。
+#
+# 冲突告警：不要同时安装独立的 @michengai/dsh-archive-manager——两个补丁都
+# 禁用官方 workspace 行并插入替换，会导致重复服务注册、启动失败。
+- id: workspace
+  disabled: true
+- id: session-projection-cache
+  disabled: true
 - insert:
     - id: dsh-plugin-tool-management
       name: 'dsh-plugin-tool-management'
       disabled: !!js "[...ctx.loader.entries()].some((e) => e.options.id !== 'dsh-plugin-tool-management' && e.options.name === 'dsh-plugin-tool-management' && !e.disabled)"
+    - id: workspace-tool-management-history
+      name: 'dsh-plugin-tool-management/workspace'
+    - id: session-projection-cache-tool-management-history
+      name: 'dsh-plugin-tool-management/projcache'
+      config:
+        writeEveryEvents: 200
+        writeIntervalMs: 5000

package/lib/agents-md/service.js

@@ -0,0 +1,240 @@
+// dsh-plugin-tool-management —— AGENTS.md 预设库 + 切换 服务层。
+//
+// DSH 全局指令基线只有一个文件 ~/.dsh/AGENTS.md（USER_GLOBAL_FILE 固定），
+// 没有内置的「多份全局 AGENTS.md 切换」机制。本服务在插件目录内维护一个
+// 预设库（每套一个子目录 + AGENTS.md），「应用」= 把选中预设内容写入
+// ~/.dsh/AGENTS.md，新会话生效（当前会话不变，DSH 本身如此）。
+//
+// 约定镜像 @deepseek-ai/dsh-agent-presets：id 即目录名，正则
+// /^[a-z0-9][a-z0-9-]*$/；__last-applied__ 是备份槽，不算用户预设。
+// 「当前生效」靠比对 ~/.dsh/AGENTS.md 的 sha256 与各预设 sha256 推断，
+// 无状态文件——用户手改全局文件也能如实反映。
+import { mkdir, readdir, readFile, rm, stat, writeFile } from 'node:fs/promises';
+import { join } from 'node:path';
+import { createHash } from 'node:crypto';
+const PRESET_ID_RE = /^[a-z0-9][a-z0-9-]*$/;
+const FILENAME = 'AGENTS.md';
+const LAST_APPLIED_ID = '__last-applied__';
+export function createAgentsMdService(_ctx, deps) {
+    const message = (e) => String((e && e.message) || e);
+    async function sha256OfFile(abs) {
+        try {
+            const buf = await readFile(abs, 'utf8');
+            return createHash('sha256').update(buf).digest('hex');
+        }
+        catch {
+            return null;
+        }
+    }
+    // 幂等初始化：仅当库完全无预设且全局 AGENTS.md 存在时，拷贝全局为 default
+    // 预设（首次运行引导）。用户已 create/apply 过任何预设则不插手，避免把
+    // 当前生效内容误存成 default 与真正生效的预设冲突。
+    async function ensureInit() {
+        const [presetsDir, globalPath] = await Promise.all([deps.presetsDir, deps.getGlobalAgentsMdPath()]);
+        let existing = [];
+        try {
+            existing = await readdir(presetsDir, { withFileTypes: true });
+        }
+        catch { /* 库不存在 */ }
+        const hasPreset = existing.some((e) => e.isDirectory() && PRESET_ID_RE.test(e.name) && e.name !== LAST_APPLIED_ID);
+        if (hasPreset)
+            return;
+        let globalContent;
+        try {
+            globalContent = await readFile(globalPath, 'utf8');
+        }
+        catch {
+            return;
+        }
+        try {
+            await mkdir(presetsDir, { recursive: true });
+            const defaultDir = join(presetsDir, 'default');
+            await mkdir(defaultDir, { recursive: true });
+            await writeFile(join(defaultDir, FILENAME), globalContent, 'utf8');
+        }
+        catch { /* best effort */ }
+    }
+    async function list() {
+        try {
+            await ensureInit();
+            const [presetsDir, globalPath] = await Promise.all([deps.presetsDir, deps.getGlobalAgentsMdPath()]);
+            let entries = [];
+            try {
+                entries = await readdir(presetsDir, { withFileTypes: true });
+            }
+            catch {
+                // 库目录不存在视作空（首次运行尚未创建）
+            }
+            const globalHash = await sha256OfFile(globalPath);
+            const ids = entries
+                .filter((e) => e.isDirectory() && PRESET_ID_RE.test(e.name) && e.name !== LAST_APPLIED_ID)
+                .map((e) => e.name);
+            const presets = await Promise.all(ids.map(async (id) => {
+                const contentHash = await sha256OfFile(join(presetsDir, id, FILENAME));
+                return {
+                    id,
+                    active: globalHash !== null && contentHash !== null && globalHash === contentHash,
+                };
+            }));
+            presets.sort((a, b) => a.id.localeCompare(b.id));
+            return { ok: true, presets };
+        }
+        catch (e) {
+            return { ok: false, error: message(e) };
+        }
+    }
+    async function read(id) {
+        const safeId = String(id || '');
+        // id 即目录名，严格校验防目录穿越：仅允许小写字母/数字/连字符。
+        // __last-applied__ 含下划线会被拒（它是备份槽，不走 read，由 apply 内部直写）。
+        if (!PRESET_ID_RE.test(safeId))
+            return { ok: false, error: '非法 id（仅允许小写字母、数字、连字符）：' + safeId };
+        try {
+            const content = await readFile(join(deps.presetsDir, safeId, FILENAME), 'utf8');
+            return { ok: true, content };
+        }
+        catch {
+            return { ok: false, error: '预设不存在：' + safeId };
+        }
+    }
+    async function create(id, from) {
+        const safeId = String(id || '');
+        if (!PRESET_ID_RE.test(safeId))
+            return { ok: false, error: '非法 id（仅允许小写字母、数字、连字符）：' + safeId };
+        const dir = join(deps.presetsDir, safeId);
+        try {
+            await stat(dir);
+            return { ok: false, error: 'id 已存在：' + safeId };
+        }
+        catch { /* 不存在，继续 */ }
+        let content;
+        const srcId = String(from || '');
+        if (srcId) {
+            if (!PRESET_ID_RE.test(srcId))
+                return { ok: false, error: '非法 from id：' + srcId };
+            try {
+                content = await readFile(join(deps.presetsDir, srcId, FILENAME), 'utf8');
+            }
+            catch {
+                return { ok: false, error: '来源预设不存在：' + srcId };
+            }
+        }
+        else {
+            content = '# AGENTS.md\n\n（DSH 全局指令基线预设，待编辑）\n';
+        }
+        await mkdir(dir, { recursive: true });
+        await writeFile(join(dir, FILENAME), content, 'utf8');
+        return { ok: true, id: safeId };
+    }
+    async function update(id, content) {
+        const safeId = String(id || '');
+        if (!PRESET_ID_RE.test(safeId))
+            return { ok: false, error: '非法 id（仅允许小写字母、数字、连字符）：' + safeId };
+        const dir = join(deps.presetsDir, safeId);
+        try {
+            await stat(dir);
+        }
+        catch {
+            return { ok: false, error: '预设不存在：' + safeId };
+        }
+        await writeFile(join(dir, FILENAME), String(content ?? ''), 'utf8');
+        return { ok: true, id: safeId };
+    }
+    async function apply(id) {
+        const safeId = String(id || '');
+        if (!PRESET_ID_RE.test(safeId))
+            return { ok: false, error: '非法 id（仅允许小写字母、数字、连字符）：' + safeId };
+        let content;
+        try {
+            content = await readFile(join(deps.presetsDir, safeId, FILENAME), 'utf8');
+        }
+        catch {
+            return { ok: false, error: '预设不存在：' + safeId };
+        }
+        const globalPath = await deps.getGlobalAgentsMdPath();
+        // 备份当前全局内容（若存在）到 __last-applied__，再覆盖写入。
+        let prev = null;
+        try {
+            prev = await readFile(globalPath, 'utf8');
+        }
+        catch { /* 当前不存在，不备份 */ }
+        if (prev !== null) {
+            const backupDir = join(deps.presetsDir, LAST_APPLIED_ID);
+            await mkdir(backupDir, { recursive: true });
+            await writeFile(join(backupDir, FILENAME), prev, 'utf8');
+        }
+        await writeFile(globalPath, content, 'utf8');
+        return { ok: true, id: safeId, backedUp: prev !== null };
+    }
+    async function getCurrent() {
+        try {
+            await ensureInit();
+            const [globalPath, presetsDir] = await Promise.all([deps.getGlobalAgentsMdPath(), deps.presetsDir]);
+            let content;
+            try {
+                content = await readFile(globalPath, 'utf8');
+            }
+            catch {
+                return { ok: true, content: '', presetId: null, exists: false };
+            }
+            const globalHash = createHash('sha256').update(content).digest('hex');
+            let presetId = null;
+            let entries = [];
+            try {
+                entries = await readdir(presetsDir, { withFileTypes: true });
+            }
+            catch { /* 库不存在 */ }
+            for (const e of entries) {
+                if (!e.isDirectory() || !PRESET_ID_RE.test(e.name) || e.name === LAST_APPLIED_ID)
+                    continue;
+                let c;
+                try {
+                    c = await readFile(join(presetsDir, e.name, FILENAME), 'utf8');
+                }
+                catch {
+                    continue;
+                }
+                if (createHash('sha256').update(c).digest('hex') === globalHash) {
+                    presetId = e.name;
+                    break;
+                }
+            }
+            return { ok: true, content, presetId, exists: true };
+        }
+        catch (e) {
+            return { ok: false, error: message(e) };
+        }
+    }
+    async function remove(id) {
+        const safeId = String(id || '');
+        if (safeId === LAST_APPLIED_ID)
+            return { ok: false, error: '备份槽不可删除' };
+        if (!PRESET_ID_RE.test(safeId))
+            return { ok: false, error: '非法 id（仅允许小写字母、数字、连字符）：' + safeId };
+        const dir = join(deps.presetsDir, safeId);
+        try {
+            await stat(dir);
+        }
+        catch {
+            return { ok: false, error: '预设不存在：' + safeId };
+        }
+        await rm(dir, { recursive: true, force: true });
+        return { ok: true, id: safeId };
+    }
+    // 从外部文本内容（如导入的 .md 文件）建预设：id 校验 + 重复检查 + 写 AGENTS.md。
+    async function importPreset(id, content) {
+        const safeId = String(id || '');
+        if (!PRESET_ID_RE.test(safeId))
+            return { ok: false, error: '非法 id（仅允许小写字母、数字、连字符）：' + safeId };
+        const dir = join(deps.presetsDir, safeId);
+        try {
+            await stat(dir);
+            return { ok: false, error: 'id 已存在：' + safeId };
+        }
+        catch { /* 不存在，继续 */ }
+        await mkdir(dir, { recursive: true });
+        await writeFile(join(dir, FILENAME), String(content ?? ''), 'utf8');
+        return { ok: true, id: safeId };
+    }
+    return { list, read, create, update, apply, getCurrent, remove, importPreset };
+}