dsh-paoding 0.4.4 → 0.4.6

package/README.md

@@ -19,12 +19,13 @@ dsh-paoding（中文品牌「庖丁」，典出《庄子》——庖丁顺纹理
 - **按工作区各有各的班底**：不同项目可以配不同的主/子 agent——「庖丁配置」里选定工作区单独配置，每个工作区生成自己的预设（`orchestrator-<目录名>`），新建会话时选对应预设即可，全局默认不受影响。
 - **图形界面配置**：不习惯改配置文件的话，点开左侧栏底部动作条（设置行上方）的「庖丁配置」入口，整页点选即可，「保存并应用」即时重生成预设——与首装自动化同一条生成管线。
 - **工具条目支持通配符**：`*` 与 `?` 在生成期展开成具体工具名——`mcp__tavily__tavily_*` 一行顶五行，MCP 工具族不必逐个点名；零匹配有提醒，产物与运行时行为不变。
-- **保存前先体检**：预览/应用时自动过一遍检查——配置了但未挂载的工具、零工具角色、查无此名的模型、预设落后于配置，分级提示一目了然；CLI 侧 `dsh-paoding doctor` 六项只读自检（配置可解析、安装轨道、声明行托管块、孤儿目录、版本标记、预设漂移），一条命令查全身。
+- **保存前先体检**：预览/应用时自动过一遍检查——配置了但未挂载的工具、零工具角色、查无此名的模型、预设落后于配置，分级提示一目了然；预览还顺带列出「本次变更」逐项明细与主 agent 工具面瘦身统计，改了什么、省了多少都在明面上。CLI 侧 `dsh-paoding doctor` 六项只读自检（配置可解析、安装轨道、声明行托管块、孤儿目录、版本标记、预设漂移），一条命令查全身；`--fix` 自动修孤儿目录与过期声明行两类安全项。
 - **场景模板一键换班底**：轻量（只留实现）、调研（只留检索）、全开（全员到齐、主 agent 带满检测库存）三档点一下就切，只动「谁在场 + 主 agent 额外带什么」，精调过的角色配置分毫不动。
-- **配置可导出可导入**：面板「更多」菜单一键下载 `dsh-paoding.config.yml`，换机迁移、备份整份拷走；导入先验格式，覆盖整体配置（含工作区段）需二次确认。
+- **配置可导出可导入，还有后悔药**：面板「更多」菜单一键下载 `dsh-paoding.config.yml`，换机迁移、备份整份拷走；导入先验格式，覆盖整体配置（含工作区段）需二次确认。每次写盘前自动轮换备份三份，「恢复备份」二次确认即回退，误操作不用翻文件。
 - **HTTP 接口直连配置**：预设的列举、新增、更新、删除走 `/api/paoding/presets` 四条接口，随 DSH Web 服务同进程挂载、本机回环直调；与面板「保存并应用」同一条生成管线、同一把写盘互斥闸——脚本和自动化工具改班底，不必开浏览器。
 - **面板中英双语**：界面文案 zh/en 全覆盖，语言自动识别（显式选择 → 中文信号优先：浏览器语言、宿主页面语言、中文时区），「更多」菜单一键切换并记住。
 - **版本更新有提醒，面板一键升级**：面板会对比 npm registry（npmmirror 镜像、GitHub release 依次兜底）上的最新版本，发现新版就地提示并附上发布页链接；点「升级」就地换上新版（profile 自动探测：普通 profile 走 `dsh plugin update dsh-paoding`；desktop profile 被 dsh CLI 的 Electron 专属守卫挡住，改在 profile 目录直跑 `pnpm add dsh-paoding@<版本>`，pnpm 找不到时自动回落 `corepack pnpm`），重启 DSH 后生效，编排预设会在重启时按新版自动重生成。开发检出的 link 形态无法就地升级，面板会提示切回 registry 版的办法；检测失败时页头留一行「版本检测失败 · 重试」（完整原因与代理排查提示在悬浮提示里），不弹窗打扰。
+- **委派效果可量化**：面板「委派效果」卡按角色统计调用数、平均耗时与失败数（近 30 天），数据读宿主会话存档本地聚合——哪个角色用得多、哪个常掉链子，一眼可见。
 - **来去自由**：不改 DSH 一行源码；装完新会话即用，卸载即删目录。
 - **两代宿主都兼容**：DSH 0.1.5 – 0.2.1 全支持（0.1.5 / 0.1.6 / 0.1.7-rc.1 / 0.1.7-rc.2 / 0.2.0-rc.1 / 0.2.0-rc.2 / 0.2.1-rc.2+）——两代宿主发现预设的机制不同（目录扫描 vs patch 声明行），安装器自动检测宿主版本选对落点；编排组合里的 workflow worker 行也按宿主版本自动适配（0.1.5 用 workflow-worker-thread，0.1.6+ 换 workflow-ptc——旧包 0.1.6 起已删除，残留行正是 0.2.0 面板上「加载失败」的元凶）；DSH 0.2.1 起角色委派行同步换代（0.2.1 的 dsh-tool-subagent 拒绝内联 persona/toolFilter——人设与工具面改挂各角色的 AgentPreset，spawn 子代理按 agentPreset 选择子预设；0.2.0 及以下保持内联形状），宿主升级、降级后的首次保存/自愈自动迁移（工作区专属预设 orchestrator-<slug> 随同轮自愈一并换代），全程无感；0.2.1 的 prerelease 再细分——tool-subagent 守卫自 0.2.1-rc.2 落地，rc.0/rc.1 仍按内联形状生成。
 
@@ -95,9 +96,9 @@ curl http://127.0.0.1:19387/api/paoding/presets
 | 文档 | 内容 |
 |---|---|
 | [架构](docs/architecture.md) | 为什么这样设计——概述、成本账、角色与工具、对应 DSH 原生机制、Token 治理、已知边界。 |
-| [安装](docs/installation.md) | 官方插件通道一条命令安装、版本支持与预设落点双轨（DSH 0.1.5 – 0.2.1）、首装自动化与版本标记、升级、开发检出自装、host patch 检测机制、庖丁配置面板、卸载、排查。 |
+| [安装](docs/installation.md) | 官方插件通道一条命令安装、版本支持与预设落点双轨（DSH 0.1.5 – 0.2.1）、首装自动化与版本标记、升级、开发检出自装、host patch 检测机制、庖丁配置面板、卸载（`dsh-paoding uninstall` 一条命令）、排查（doctor 可 `--fix`）。 |
 | [编排](docs/orchestration.md) | 编排总览、失败如何呈现与三层处理、one-shot vs continuable、上下文隔离。 |
-| [配置](docs/configuration.md) | 配置键全参考、工具条目通配符、内置角色微调、主 agent 工具与技能、自定义角色、按角色分模型、角色会话与启动方式、场景模板、配置导出导入、检查提示、HTTP 接口。 |
+| [配置](docs/configuration.md) | 配置键全参考、工具条目通配符、内置角色微调、主 agent 工具与技能、自定义角色、按角色分模型、角色会话与启动方式、场景模板、配置导出导入与备份回滚、检查提示、HTTP 接口。 |
 
 ## 贡献
 

package/README_en.md

@@ -19,9 +19,10 @@ dsh-paoding (Chinese brand 「庖丁」/ Pao Ding, from the Zhuangzi parable of
 - **A different lineup per workspace**: each project can carry its own main/sub-agent setup — pick the workspace in 庖丁配置, configure it separately, and it gets its own preset (`orchestrator-<basename>`); choose that preset when starting a session. The global default stays untouched.
 - **A visual configurator**: prefer clicking to editing YAML? the 庖丁配置 (Paoding Config) entry in the sidebar bottom action bar (above the Settings row) covers everything; Save & Apply regenerates the preset on the spot — the same generation pipeline as first-install automation.
 - **Tool entries support wildcards**: `*` and `?` expand into concrete tool names at generation time — one `mcp__tavily__tavily_*` line covers the whole family; zero-match patterns get a warning, and artifacts and runtime behavior stay unchanged.
-- **A health check before every save**: preview/apply runs a built-in check — tools configured but not mounted, zero-tool roles, unknown model names, a preset lagging behind the config — each flagged by severity; on the CLI side, `dsh-paoding doctor` runs six read-only checks (config parses, install track, managed declaration block, orphan directories, version marker, preset drift) in one command.
+- **A health check before every save**: preview/apply runs a built-in check — tools configured but not mounted, zero-tool roles, unknown model names, a preset lagging behind the config — each flagged by severity; the preview also lists an itemized "changes since the last save" diff and the main-agent tool-surface stat, so what changed and what it saves are both on the table. On the CLI side, `dsh-paoding doctor` runs six read-only checks (config parses, install track, managed declaration block, orphan directories, version marker, preset drift) in one command; `--fix` automatically repairs the two safe categories (orphan directories, stale declaration rows).
 - **Scenario templates swap the roster in one click**: Lite (implement only), Research (search roles only), and Full (everyone on deck, main agent carrying the whole detected inventory) — templates only touch who is present and what the main agent carries; hand-tuned role settings are never touched.
-- **Config export & import**: download `dsh-paoding.config.yml` from the panel's More menu for migration or backup; imports are validated first, and overwriting the whole config (workspaces included) asks for confirmation.
+- **Delegation stats you can act on**: the panel's "Delegation" card counts delegations, average duration, and failures per role (last 30 days), aggregated locally from the host session archive — which roles get used, and which keep dropping the ball, at a glance.
+- **Config export & import, with an undo**: download `dsh-paoding.config.yml` from the panel's More menu for migration or backup; imports are validated first, and overwriting the whole config (workspaces included) asks for confirmation. Every write first rotates three automatic backups, and "Restore backup" puts one back after a confirming dialog — a mistake never means digging through files.
 - **Config over HTTP**: list, create, update, and delete presets through the four `/api/paoding/presets` endpoints, served in-process by the DSH web service and callable over loopback; they ride the panel's Save & Apply generation pipeline and share its write lock, so scripts and automation tools can reshape the roster without opening a browser.
 - **Panel in English and Chinese**: every UI string ships in both languages, auto-detected (explicit choice first, then Chinese signals — browser language, host page language, Chinese time zones — outrank English ones, since the Electron shell pins navigator to en-US regardless of the OS), switchable from the More menu and remembered.
 - **Notified of new releases, one-click upgrade in the panel**: the panel compares your local version against the latest npm registry publish (npmmirror and GitHub release as successive fallbacks) and points to the release page when a newer one is out. The Upgrade button swaps in the new version in place (profile auto-detected): regular profiles go through `dsh plugin update dsh-paoding`, while the Electron-owned desktop profile — which the dsh CLI guards against — gets a direct `pnpm add dsh-paoding@<version>` run in the profile directory, falling back to `corepack pnpm` when no bare pnpm is on PATH; a DSH restart applies it, and the orchestrator preset regenerates itself against the new version on startup. A dev checkout in link: form cannot be upgraded in place — the panel tells you how to switch back to the registry version. If the check fails (offline, rate-limited), a subdued "version check failed · retry" row stays in the header (full causes and proxy hints in its tooltip) without popping anything up.
@@ -95,9 +96,9 @@ Writes go through the same generation pipeline and write lock as the panel's Sav
 | Document | Contents |
 |---|---|
 | [Architecture](docs/architecture_en.md) | Why it is designed this way — overview, the cost ledger, roles and tools, mapping to DSH's native mechanisms, token governance, known limits. |
-| [Installation](docs/installation_en.md) | The single official channel (`dsh plugin add`), version support and the dual preset landing (DSH 0.1.5 – 0.2.1), first-install automation and the version marker, upgrading, dev-checkout install, host patch detection, the Paoding Config panel, uninstalling, troubleshooting. |
+| [Installation](docs/installation_en.md) | The single official channel (`dsh plugin add`), version support and the dual preset landing (DSH 0.1.5 – 0.2.1), first-install automation and the version marker, upgrading, dev-checkout install, host patch detection, the Paoding Config panel, uninstalling (one `dsh-paoding uninstall` command), troubleshooting (doctor with `--fix`). |
 | [Orchestration](docs/orchestration_en.md) | Orchestration overview, how failures surface and the three recovery layers, one-shot vs continuable, context isolation. |
-| [Configuration](docs/configuration_en.md) | The full key reference, tool-entry wildcards, tuning built-in roles, main-agent tools and skills, custom roles, per-role models, per-role session and start modes, scenario templates, config export/import, health-check hints, the HTTP API. |
+| [Configuration](docs/configuration_en.md) | The full key reference, tool-entry wildcards, tuning built-in roles, main-agent tools and skills, custom roles, per-role models, per-role session and start modes, scenario templates, config export/import and backup-rollback, health-check hints, the HTTP API. |
 
 ## Contributing
 

package/bin/dsh-paoding.mjs

@@ -15,16 +15,27 @@ const REFUSE =
 const usage = `dsh-paoding 安装快捷方式（等价于 dsh plugin --profile <name> add dsh-paoding@<version>）
 
 用法：npx dsh-paoding@latest [--profile <name>]
-      npx dsh-paoding@latest doctor [--profile <name>] [--config <file>]
+      npx dsh-paoding@latest doctor [--profile <name>] [--config <file>] [--fix] [--yes]
+      npx dsh-paoding@latest uninstall [--purge] [--yes] [--config <file>]
 
 子命令：
-  doctor            本机安装体检（只读）：配置文件 / 预设轨道 / 托管声明行 /
+  doctor            本机安装体检（默认只读）：配置文件 / 预设轨道 / 托管声明行 /
                     孤儿工作区预设 / 预设漂移 / 版本标记，结尾给汇总；
-                    有错误级发现时退出码 1。
+                    有错误级发现时退出码 1。--fix 对其中可安全自动修的两类动手
+                    （孤儿预设目录、孤儿/过期声明行；写盘前备份 .bak-1），其余
+                    发现只给处理建议，输出逐项标注「已修 / 需手动」。
+  uninstall         清理庖丁写进宿主目录的东西：cordis.patch.yml 托管声明块、
+                    .agent-presets 下的 orchestrator* 预设目录、.generator-version
+                    标记。默认保留配置文件及其备份（用户数据），--purge 才连配置
+                    一起删；不碰插件包自身（包本体用 dsh plugin remove 卸载）。
+                    非交互终端（非 TTY）下没带 --yes 只列计划不执行，退出码 2。
 
 选项：
   --profile <name>  插件装入的 profile，默认 web（dsh web 的默认 profile）
-  --config <file>   配置文件路径（仅 doctor 子命令使用）
+  --config <file>   配置文件路径（doctor / uninstall 子命令使用）
+  --purge           uninstall 连配置文件及其备份一起删（默认保留）
+  --yes             跳过逐类确认直接执行（uninstall / doctor --fix）
+  --fix             doctor 体检后自动修可安全处理的问题
   --help            显示本帮助
   --version         显示包版本
 
@@ -36,17 +47,18 @@ const { version } = JSON.parse(readFileSync(new URL('../package.json', import.me
 
 const argv = process.argv.slice(2);
 
-// doctor 子命令：包不再携带独立 CLI，但 tools/ 随包发布（package.json files），
-// 体检逻辑本体在 tools/lib/cli.mjs——这里 dynamic import 转发（--profile /
-// --config 等参数原样透传，由它自己的 parseArgs 解析），退出码语义：
-// 无错误 0，有 error 级发现 1。放在参数循环之前：doctor 的参数不用本入口的
+// doctor / uninstall 子命令：包不再携带独立 CLI，但 tools/ 随包发布
+//（package.json files），逻辑本体在 tools/lib/cli.mjs——这里 dynamic import
+// 转发（--profile / --config / --fix / --purge / --yes 等参数原样透传，由它
+// 自己的 parseArgs 解析），退出码语义：成功 0，有 error 级发现或清理失败 1，
+// 非 TTY 无 --yes 的防呆拒做 2。放在参数循环之前：子命令的参数不用本入口的
 // profile 解析兜底。
-if (argv.includes('doctor')) {
+if (argv.includes('doctor') || argv.includes('uninstall')) {
   try {
     const { main } = await import('../tools/lib/cli.mjs');
     process.exit((await main({ argv })) ?? 0);
   } catch (err) {
-    console.error(`doctor 运行失败：${err?.message ?? err}`);
+    console.error(`子命令运行失败：${err?.message ?? err}`);
     process.exit(1);
   }
 }

package/docs/architecture.md

@@ -58,6 +58,7 @@ dsh-paoding（中文品牌「庖丁」，取庖丁解牛之意）是 DSH（DeepS
         │               │               │               │
         ▼               ▼               ▼               ▼
    web_search       read/read_image  read/read_image  glob/grep/read/
+   web_fetch
    mcp__tavily__*   glob/grep/bash/  write/edit/      read_image/bash/
    (5 个)           skill/           glob/grep/bash/  ask_user_question/
    glob/grep/read   web_search       skill/web_search todo_write/
@@ -78,7 +79,7 @@ dsh-paoding（中文品牌「庖丁」，取庖丁解牛之意）是 DSH（DeepS
 |---|---|---|---|
 | 单体（现状，未裁剪） | 60 工具全量 | **~16.2k** | 每请求固定注入 |
 | 编排主 agent | 21 项白名单（∩ 注册面） | **~5.5k** | ↓66%；内部搜索零委派延迟 |
-| search_external 子 agent | 10 项 allow | ~3.0k | 仅实际联网调研时才产生 |
+| search_external 子 agent | 11 项 allow | ~3.0k | 仅实际联网调研时才产生 |
 | design 子 agent | 12 项 allow | ~3.3k | 仅实际委派设计时才产生 |
 | implement 子 agent | 17 项 allow | ~3.8k | 仅实际委派实现时才产生 |
 | search_internal_deep 子 agent | 10 项 allow | 未单独估算 | 仅实际委派全仓深搜时才产生 |
@@ -123,7 +124,7 @@ dsh-paoding（中文品牌「庖丁」，取庖丁解牛之意）是 DSH（DeepS
 **host 工具不在白名单常量内硬编码**：`mcp__*` 与插件工具（codegraph / memory_search / mnemon_* 等）
 是"检测期事实"——主 agent 需要时经配置文件 `$DSH_HOME/dsh-paoding.config.yml` 的 `main_agent_extra`
 勾选，生成层按运行时检测把仍启用的注入 restrict 的 `config.allow`，停用即自动清除（详见
-[docs/configuration.md](configuration.md)）。因此主 agent 的默认面**不包含** `web_search`、全部
+[docs/configuration.md](configuration.md)）。因此主 agent 的默认面**不包含** `web_search` / `web_fetch`、全部
 `mcp__tavily__*`、`write` / `edit` / `str_replace_editor`、`skill`、`workflow`、`ralph` 及未过滤的
 `subagent` / `subagent_fork`——这些要么属于角色子 agent，要么成本被刻意清零。
 
@@ -142,7 +143,7 @@ dsh-paoding（中文品牌「庖丁」，取庖丁解牛之意）是 DSH（DeepS
 
 | 委派工具 | 实例 ID | persona 一句话职责 | 加载成本 |
 |---|---|---|---|
-| `search_external` | `delegation-search-external` | 只做联网调研（web_search + tavily），绝不改文件；返回结构化研究摘要（发现、来源、待答问题） | ~3.0k，仅委派时 |
+| `search_external` | `delegation-search-external` | 只做联网调研（web_search + web_fetch + tavily），绝不改文件；返回结构化研究摘要（发现、来源、待答问题） | ~3.0k，仅委派时 |
 | `design` | `delegation-design` | 产出 UI/UX 设计、线框与前端 spec，可用 skill 工具加载设计技能（frontend-design、html-ppt）；先读参考再交付 spec，不写最终代码 | ~3.3k，仅委派时 |
 | `implement` | `delegation-implement` | 写与改代码：先读上下文 → 精准修改 → 验证（build/test/grep）；不改架构 | ~3.8k，仅委派时 |
 | `search_internal_deep` | `delegation-search-internal-deep` | 只做仓库探索：不改文件、不联网；处理主 agent 上下文装不下的任务（全仓 grep dump、超页大文件通读、跨目录符号/定义追踪），只回浓缩摘要 | 未单独估算，仅委派时 |
@@ -157,12 +158,12 @@ dsh-paoding（中文品牌「庖丁」，取庖丁解牛之意）是 DSH（DeepS
 **不做**库存对账：手写进去、两头都不占的名字不会被剔除，会静默落盘、运行时以 unknown tools 拒创建
 （与下方该角色的 codegraph 注同一口径），见 [docs/configuration.md](configuration.md)）。
 
-**`search_external`（外部调研，10 项）**
+**`search_external`（外部调研，11 项）**
 
 | 字段 | 值 |
 |---|---|
 | persona 职责 | 只做联网调研并返回结构化研究摘要（发现、来源、待答问题）；绝不修改文件 |
-| toolFilter.allow | `web_search`、`mcp__tavily__tavily_search`、`mcp__tavily__tavily_crawl`、`mcp__tavily__tavily_extract`、`mcp__tavily__tavily_map`、`mcp__tavily__tavily_research`、`glob`、`grep`、`read`、`ask_user_question` |
+| toolFilter.allow | `web_search`、`web_fetch`、`mcp__tavily__tavily_search`、`mcp__tavily__tavily_crawl`、`mcp__tavily__tavily_extract`、`mcp__tavily__tavily_map`、`mcp__tavily__tavily_research`、`glob`、`grep`、`read`、`ask_user_question` |
 | 加载成本 | ~3.0k tokens，仅实际联网调研时发生 |
 
 **`design`（设计，12 项）**

package/docs/architecture_en.md

@@ -70,6 +70,7 @@ allow lists are in the "Roles & tool surfaces" section below):
             │                   │                   │                   │               
             ▼                   ▼                   ▼                   ▼               
     web_search           read/read_image      read/read_image      glob/grep/read
+    web_fetch
     mcp__tavily__*       glob/grep/bash       write/edit           read_image/bash
     (5 tools)            skill/web_search     glob/grep/bash       ask_user_question
     glob/grep/read       ask_user_question    skill/web_search     todo_write
@@ -90,8 +91,8 @@ In the diagram:
   ask_user_question / get_goal / create_goal / update_goal / exit_plan_mode / job_output / job_list /
   job_kill);
 - Four arrows fan out to the **role sub-agents**, each created on demand (cold path). search_external
-  carries web_search, the five `mcp__tavily__*` tools, plus glob/grep/read and ask_user_question, with no
-  write/edit; design carries read/read_image/glob/grep/bash/skill/web_search plus coordination tools;
+  carries web_search, web_fetch, the five `mcp__tavily__*` tools, plus glob/grep/read and ask_user_question,
+  with no write/edit; design carries read/read_image/glob/grep/bash/skill/web_search plus coordination tools;
   implement carries read/read_image/write/edit/glob/grep/bash/skill/web_search plus coordination and
   goal tools;
   search_internal_deep carries the full read-only internal-search surface (glob/grep/read/read_image/bash
@@ -107,7 +108,7 @@ measured calibration; figures are order-of-magnitude estimates):
 |---|---|---|---|
 | Monolithic (status quo, untrimmed) | All 60 tools | **~16.2k** | Paid in full on every request |
 | Orchestrator main agent | 21-entry whitelist (∩ registry) | **~5.5k** | ↓66%; zero delegation latency for internal search |
-| search_external sub-agent | 10 allowed entries | ~3.0k | Paid only when web research actually runs |
+| search_external sub-agent | 11 allowed entries | ~3.0k | Paid only when web research actually runs |
 | design sub-agent | 12 allowed entries | ~3.3k | Paid only when a design task is delegated |
 | implement sub-agent | 17 allowed entries | ~3.8k | Paid only when an implementation task is delegated |
 | search_internal_deep sub-agent | 10 allowed entries | not separately estimated | Paid only when a deep repo search is delegated |
@@ -160,7 +161,7 @@ memory_search / mnemon_* etc.) are "install-time facts" — when the main agent
 `main_agent_extra` in the configuration file `$DSH_HOME/dsh-paoding.config.yml`, the generator injects the
 still-enabled ones into the restrict `config.allow` based on runtime detection, and disabled ones are
 dropped automatically (see [docs/configuration_en.md](configuration_en.md)). The main agent's default surface
-therefore excludes `web_search`, every `mcp__tavily__*`, `write` / `edit` / `str_replace_editor`, `skill`,
+therefore excludes `web_search` / `web_fetch`, every `mcp__tavily__*`, `write` / `edit` / `str_replace_editor`, `skill`,
 `workflow`, `ralph`, and the unfiltered `subagent` / `subagent_fork` — either because they belong to role
 sub-agents or because their cost is deliberately zeroed out.
 
@@ -184,7 +185,7 @@ For example, a `ppt` agent carrying the `html-ppt` skill gives slide-making its
 
 | Delegation tool | Instance ID | Persona in one sentence | Load cost |
 |---|---|---|---|
-| `search_external` | `delegation-search-external` | Web research only (web_search + tavily), never modifies files; returns a structured research summary (findings, sources, open questions) | ~3.0k, only when delegated |
+| `search_external` | `delegation-search-external` | Web research only (web_search + web_fetch + tavily), never modifies files; returns a structured research summary (findings, sources, open questions) | ~3.0k, only when delegated |
 | `design` | `delegation-design` | Produces UI/UX designs, wireframes and frontend specs; may load design skills (frontend-design, html-ppt) via the skill tool; reads reference material first, delivers a spec, does not implement final code | ~3.3k, only when delegated |
 | `implement` | `delegation-implement` | Writes and edits code: read surrounding context → surgical change → verify (build/test/grep); does not redesign architecture | ~3.8k, only when delegated |
 | `search_internal_deep` | `delegation-search-internal-deep` | Repository exploration only: never modifies files, never uses the network; handles tasks too heavy for the main agent's context (whole-repo grep dumps, full-file reads beyond a page, cross-directory symbol/definition tracking); returns condensed summaries only | not separately estimated, only when delegated |
@@ -213,12 +214,12 @@ on disk silently, and fail at runtime with unknown tools at child creation — s
 note for that role below; see
 [docs/configuration_en.md](configuration_en.md)).
 
-**`search_external` (external research, 10 entries)**
+**`search_external` (external research, 11 entries)**
 
 | Field | Value |
 |---|---|
 | Persona job | Web research only, returning a structured research summary (findings, sources, open questions); never modifies files |
-| toolFilter.allow | `web_search`, `mcp__tavily__tavily_search`, `mcp__tavily__tavily_crawl`, `mcp__tavily__tavily_extract`, `mcp__tavily__tavily_map`, `mcp__tavily__tavily_research`, `glob`, `grep`, `read`, `ask_user_question` |
+| toolFilter.allow | `web_search`, `web_fetch`, `mcp__tavily__tavily_search`, `mcp__tavily__tavily_crawl`, `mcp__tavily__tavily_extract`, `mcp__tavily__tavily_map`, `mcp__tavily__tavily_research`, `glob`, `grep`, `read`, `ask_user_question` |
 | Load cost | ~3.0k tokens, incurred only when web research actually runs |
 
 **`design` (design, 12 entries)**

package/docs/configuration.md

@@ -73,6 +73,7 @@ roles:                  # toolName -> { persona?, tools[] }
       You are the external-research agent. Web research only — never modify files.
     tools:
       - web_search
+      - web_fetch
       - mcp__tavily__tavily_search
       - mcp__tavily__tavily_crawl
       - mcp__tavily__tavily_extract
@@ -138,11 +139,14 @@ skills:                 # skill -> 分配给的角色（写进角色 persona 的
     search_external:
       tools:
         - web_search
+        - web_fetch
         - mcp__tavily__tavily_*   # 生成时展开为 search/crawl/extract/map/research（以实际检测为准）
   ```
 
   `main_agent_extra` 与自定义角色的 `tools` 同样支持。配置文件里模式条目原样存取（配置面板回显不受影响），只有生成的 preset 里落的是展开后的具体名。
 
+  > 宿主 0.2.1-rc.2 起可在 **设置 → 网页搜索** 开启 DeepSeek Search / Fetch API：`web_search` / `web_fetch` 的名字不变，后方 provider 自动切换到官方新后端——无需改动庖丁配置，角色 allow 里的这两个名字照常生效。
+
 示例（替换 `implement` 的 persona 与 tools）：下方 tools 与静态默认一致——注意它不含 `str_replace_editor`（原因见 8.2）。显式写出 tools 等于固定意图：将来静态源怎么改都不再影响本角色：
 
 ```yaml
@@ -350,7 +354,7 @@ fork 的语义（如实说明）：
 | 内部搜索（热路径） | `glob` `grep` `read` `read_image` `bash` |
 | 协调 | `todo_write` `ask_user_question` `get_goal` `create_goal` `update_goal` `exit_plan_mode` `job_output` `job_list` `job_kill` |
 
-对照说明：`web_search`、全部 `mcp__tavily__*`、`write` / `edit` / `str_replace_editor`、`skill`、`workflow`、`ralph`、未过滤的 `subagent` 委派等**刻意不在**主 agent 面上（host 工具如 codegraph / `memory_search` / mnemon 同理不硬编码进源码，需要时用 `main_agent_extra` 勾选，见 3.2）。另外 `restrict.mjs` 在注册层对主 agent deny 了 `skill`（目录注入守卫查的是注册面而非模型可见面），主 agent 的技能目录注入随之归零。
+对照说明：`web_search` / `web_fetch`、全部 `mcp__tavily__*`、`write` / `edit` / `str_replace_editor`、`skill`、`workflow`、`ralph`、未过滤的 `subagent` 委派等**刻意不在**主 agent 面上（host 工具如 codegraph / `memory_search` / mnemon 同理不硬编码进源码，需要时用 `main_agent_extra` 勾选，见 3.2）。另外 `restrict.mjs` 在注册层对主 agent deny 了 `skill`（目录注入守卫查的是注册面而非模型可见面），主 agent 的技能目录注入随之归零。
 
 ### 3.2 追加工具：main_agent_extra
 
@@ -735,6 +739,8 @@ workspaces:
 - 落盘即改 mtime，面板的检测状态缓存按对账自动失效，视图随后按新配置重建。
 - 与「保存并应用」共用同一把写盘互斥闸：应用进行中导入（或反过来）会收到 409，稍后重试即可。
 
+**备份与恢复**：凡是往配置文件写盘的动作（保存并应用、导入、恢复备份）都会先把当前文件轮换备份三份——`dsh-paoding.config.yml.bak-1`（最新）到 `.bak-3`，逐份后移、最旧的淘汰；备份与配置文件同目录，首次保存前没有旧值就不备份。回滚走「更多」→「恢复备份…」：列出各备份的生成时间与大小，二次确认后整份覆盖回当前配置（恢复前当前文件还会再自动备份一份，连按恢复也能原路退回）；恢复同样先验格式（解析不出顶层映射的备份拒收），共用写盘互斥闸。没有面板时，手动把 `.bak-N` 拷回主文件名即可。
+
 ## 12. 检查提示（体检）
 
 预览或应用后，若服务端体检发现问题，预览区上方会出现一张**检查提示**卡：逐条列出示意，红（error）/ 琥珀（warn）/ 中性（info）三色分级，悬停可见检查项代号（排障用）。没有问题就没有这张卡，零占位。
@@ -751,6 +757,8 @@ workspaces:
 
 作废口径：**预览**产生的提示随配置草稿的任何改动一并作废（与预览同一逻辑）；**应用**产生的提示描述的是已落盘结果，保留到下一次预览或应用。应用失败时错误 toast 照旧，同时卡片给出每条失败原因；应用成功但存在 warn/error 级提示时，成功 toast 会换成琥珀档「有 N 条检查提示，见下方」。
 
+体检之外，预览还带两样「看得见」的东西：一是**本次变更**——相对上次保存配置的逐项明细（角色增删、各角色工具与 persona、专用模型、启动方式、后台续跑、主 agent 工具增删、`roles_remove` 与 `main_agent_extra` 通道），没有任何改动就一行「与上次保存的配置一致」；二是**工具面瘦身统计**——主 agent 当前 allow 数对照单体全量数的百分比，页头常驻一份（按已保存配置），预览区再示一份（按当前未保存配置），保存前后各看各的。
+
 ## 13. HTTP 接口
 
 插件的 Web 后端随 DSH Web 服务同进程挂载在 `/api/paoding/*` 前缀下（回环地址 + 端口随部署，信任围栏与面板一致），面板能做的配置读写，脚本与自动化工具也能照做。预设对象的完整 CRUD 四条：

package/docs/configuration_en.md

@@ -73,6 +73,7 @@ roles:                  # toolName -> { persona?, tools[] }
       You are the external-research agent. Web research only — never modify files.
     tools:
       - web_search
+      - web_fetch
       - mcp__tavily__tavily_search
       - mcp__tavily__tavily_crawl
       - mcp__tavily__tavily_extract
@@ -138,11 +139,14 @@ Of the four built-in delegation roles, the installer (the `ROLES` constant) mana
     search_external:
       tools:
         - web_search
+        - web_fetch
         - mcp__tavily__tavily_*   # expanded at generation to search/crawl/extract/map/research (per actual detection)
   ```
 
   `main_agent_extra` and custom-role `tools` support them too. Pattern entries are stored and read back verbatim in the config file (the config panel renders them fine); only the generated preset contains the expanded concrete names.
 
+  > Host 0.2.1-rc.2 and later can enable the DeepSeek Search / Fetch API under **Settings → Web Search**: `web_search` / `web_fetch` keep their names while the provider behind them switches to the official new backend — no 庖丁 (Paoding) config change needed, and both names in a role's allow list keep working as-is.
+
 Example — replacing `implement`'s persona and tools. The `tools` below match the static default — note that it does **not** contain `str_replace_editor` (why: see 8.2). Writing `tools` out explicitly freezes the intent: later changes to the static source no longer affect this role:
 
 ```yaml
@@ -350,7 +354,7 @@ The `MAIN_AGENT_ALLOW` constant in `restrict.mjs` is the **base** (21 names) tha
 | Internal search (hot path) | `glob` `grep` `read` `read_image` `bash` |
 | Coordination | `todo_write` `ask_user_question` `get_goal` `create_goal` `update_goal` `exit_plan_mode` `job_output` `job_list` `job_kill` |
 
-For orientation: `web_search`, every `mcp__tavily__*`, `write` / `edit` / `str_replace_editor`, `skill`, `workflow`, `ralph`, and the unfiltered `subagent` delegation are **deliberately absent** from the main-agent surface (host tools such as codegraph / `memory_search` / mnemon are likewise not hard-coded into the source — opt them in via `main_agent_extra`, see 3.2). `restrict.mjs` also denies `skill` at the registry level for the main agent (the catalog-injection guard checks registry visibility, not the model-facing list), which brings the main agent's skill-catalog token tax to zero.
+For orientation: `web_search` / `web_fetch`, every `mcp__tavily__*`, `write` / `edit` / `str_replace_editor`, `skill`, `workflow`, `ralph`, and the unfiltered `subagent` delegation are **deliberately absent** from the main-agent surface (host tools such as codegraph / `memory_search` / mnemon are likewise not hard-coded into the source — opt them in via `main_agent_extra`, see 3.2). `restrict.mjs` also denies `skill` at the registry level for the main agent (the catalog-injection guard checks registry visibility, not the model-facing list), which brings the main agent's skill-catalog token tax to zero.
 
 ### 3.2 Appending tools: main_agent_extra
 
@@ -747,6 +751,8 @@ The header "More" menu offers export and import of the whole config file (the §
 - Writing changes the file's mtime, so the panel's detection-state cache invalidates automatically through the existing reconciliation, and the view rebuilds from the new config.
 - Import shares the same in-flight write gate as "save & apply": applying during an import (or the reverse) answers 409 — retry shortly.
 
+**Backups and restore**: every action that writes the config file (save & apply, import, restore) first rotates the current file through three backups — `dsh-paoding.config.yml.bak-1` (newest) to `.bak-3`, each shifting one slot and the oldest falling off; backups live next to the config file, and a first-ever save with no previous file backs up nothing. Rollback runs through "More" → "Restore backup…": each backup is listed with its time and size, one confirming dialog puts it back in place (the current file is backed up once more first, so even a mistaken restore can be walked back); the restore validates the backup too (a backup that parses to a non-mapping is rejected) and shares the write gate. Without the panel, copy a `.bak-N` back over the main filename by hand.
+
 ## 12. Check messages (health review)
 
 After a preview or an apply, if the server-side health review finds problems, a **check messages** card appears above the preview area: one row per finding, color-coded red (error) / amber (warn) / neutral (info), with the check code available on hover for troubleshooting. No findings — no card, zero footprint.
@@ -763,6 +769,8 @@ Entries come from the structured health review (`tools/lib/health.mjs`); codes a
 
 Invalidation scope: messages from a **preview** are discarded together with the draft on any config edit (same rule as the preview itself); messages from an **apply** describe what has already been written and survive until the next preview or apply. When an apply fails, the error toast stays as usual and the card lists every reason; when an apply succeeds but warn/error-level findings exist, the success toast switches to the amber variant "N check messages, see below".
 
+Beside the checks, a preview carries two more pieces of visibility. One is **changes since the last save** — an itemized diff against the saved config (roles added/removed, each role's tools and persona, dedicated model, start mode, background continuation, main-agent tool changes, the `roles_remove` and `main_agent_extra` channels); with nothing changed it collapses to a single "identical to the last saved config" line. The other is the **tool-surface stat** — the main agent's current allow count against the monolithic total as a percentage: one permanent line in the header (from the saved config) and one in the preview (from the unsaved config), one for each side of a save.
+
 ## 13. HTTP API
 
 The plugin's web backend is mounted under the `/api/paoding/*` prefix inside the DSH web server itself (loopback address + port per your deployment; the same trust fence as the panel). Anything the panel can do to the config, scripts and automation can do too. Full CRUD for preset objects:

package/docs/installation.md

@@ -191,22 +191,30 @@ patch 与配置文件用 **`yaml` 包**解析——通过 `createRequire` 从 `$
 
 ### 数据接口与安全
 
-面板数据走**同源 `/api/paoding/*`**：Node 侧注册前缀路由——`GET /api/paoding/state`（带缓存的检测状态）、`GET /api/paoding/version`（最新版检测，版本/更新卡的数据源）、`GET /api/paoding/models`（宿主模型运行时的模型清单）、`GET` + `POST /api/paoding/workspaces`（工作区列表 / 添加工作区）、`POST /api/paoding/rescan`（强制重检测，含 MCP 握手）、`POST /api/paoding/preview`（不写盘生成）、`POST /api/paoding/apply`（安装并保存配置）、`POST /api/paoding/upgrade`（就地升级：普通 profile 走 `dsh plugin update`，desktop profile 直跑 `pnpm add`，pnpm 按「裸 pnpm → `corepack pnpm` → PNPM_HOME 落位」候选链解析）、`GET /api/paoding/config/export`（配置文件原文导出，下载/备份用）、`POST /api/paoding/config/import`（整份导入：先验格式、与「保存并应用」共用写盘闸）、`GET /api/paoding/client.css`（面板样式静态下发）——复用 `tools/` 的检测与生成管线（`plugins/paoding-config-ui/api-core.mjs`）。路由自带**浏览器信任围栏**：Host 必须回环或落在 `webRuntime.trustedHosts`，并拒绝 cross-site 请求（复刻 DSH `/api` 网关围栏语义，因为 `/api/paoding` 前缀更长会命中本插件而绕开网关）。面板**没有独立服务端/端口**：DSH Web 端口只监听 `127.0.0.1`，不对外网暴露。
+面板数据走**同源 `/api/paoding/*`**：Node 侧注册前缀路由——`GET /api/paoding/state`（带缓存的检测状态）、`GET /api/paoding/version`（最新版检测，版本/更新卡的数据源）、`GET /api/paoding/models`（宿主模型运行时的模型清单）、`GET` + `POST /api/paoding/workspaces`（工作区列表 / 添加工作区）、`POST /api/paoding/rescan`（强制重检测，含 MCP 握手）、`POST /api/paoding/preview`（不写盘生成）、`POST /api/paoding/apply`（安装并保存配置）、`POST /api/paoding/upgrade`（就地升级：普通 profile 走 `dsh plugin update`，desktop profile 直跑 `pnpm add`，pnpm 按「裸 pnpm → `corepack pnpm` → PNPM_HOME 落位」候选链解析）、`GET /api/paoding/config/export`（配置文件原文导出，下载/备份用）、`POST /api/paoding/config/import`（整份导入：先验格式、与「保存并应用」共用写盘闸）、`GET /api/paoding/config/backups`（自动备份清单）、`POST /api/paoding/config/restore`（恢复备份：文件名白名单防穿越，先验可解析再覆盖）、`GET /api/paoding/stats`（委派效果统计：冷读会话存档聚合，days 窗口与 project 过滤）、`GET /api/paoding/client.css`（面板样式静态下发）——复用 `tools/` 的检测与生成管线（`plugins/paoding-config-ui/api-core.mjs`）。路由自带**浏览器信任围栏**：Host 必须回环或落在 `webRuntime.trustedHosts`，并拒绝 cross-site 请求（复刻 DSH `/api` 网关围栏语义，因为 `/api/paoding` 前缀更长会命中本插件而绕开网关）。面板**没有独立服务端/端口**：DSH Web 端口只监听 `127.0.0.1`，不对外网暴露。
 
 ### 面板能力
 
-- **状态与操作**：页头是标题 + 版本/更新卡（新版就地提示与一键升级）+ 副标题，正文直接就是工作区与角色卡，没有单独的检测摘要与 chips 区；操作按钮「**保存并应用**」「预览生成」「重新检测」固定在面板底部悬浮操作坞（sticky bottom），滚动任意位置都可直接点击。
-- **Agent 工具分配卡片**：主 agent 一张固定卡（基础工具取消勾选 = 加入 `main_agent_remove` 从主 agent 剔除；MCP / 插件工具勾选 = 加入 `main_agent_extra`；技能「read 按需」与「内联全文」两组勾选；另有「人设追加」编辑区，可改、清空、恢复默认）＋ `search_external` / `design` / `implement` 三张内置角色卡 ＋ 任意自定义角色卡（带删除按钮）；每张角色卡把工具按「基础工具 / MCP 工具 / 插件工具」三组呈现、可逐项勾选，并可编辑 persona。底部「新建自定义 agent 角色」。
+- **状态与操作**：页头是标题 + 版本/更新卡（新版就地提示与一键升级）+ 副标题，副标题下一行常驻主 agent 工具面瘦身统计（当前 allow 数 / 单体全量数与瘦身百分比，配完第一次能看到省了多少），正文直接就是工作区与角色卡，没有单独的检测摘要与 chips 区；操作按钮「**保存并应用**」「预览生成」「重新检测」固定在面板底部悬浮操作坞（sticky bottom），滚动任意位置都可直接点击。
+- **Agent 工具分配卡片**：主 agent 一张固定卡（基础工具取消勾选 = 加入 `main_agent_remove` 从主 agent 剔除；MCP / 插件工具勾选 = 加入 `main_agent_extra`；技能「read 按需」与「内联全文」两组勾选；另有「人设追加」编辑区，可改、清空、恢复默认）＋ `search_external` / `design` / `implement` 三张内置角色卡 ＋ 任意自定义角色卡（带删除按钮）；每张角色卡把工具按「基础工具 / MCP 工具 / 插件工具」三组呈现、可逐项勾选（任一组超过 8 项时组顶出现筛选框，输入即滤），并可编辑 persona。底部「新建自定义 agent 角色」。
 - **技能分配**：每个 skill 一行，勾选分配给哪些 agent（写入对应 persona 的 `Available skills` 软引导；角色技能软引导，主 agent 技能另有 read 按需/内联两组）。
-- **预览生成**：不写盘生成 `agent.cordis.yml`，显示每个角色 allow 的 intent → kept 计数与生成全文。
+- **预览生成**：不写盘生成 `agent.cordis.yml`，显示每个角色 allow 的 intent → kept 计数与生成全文；同时列「本次变更」——相对上次保存配置的逐项明细（角色增删、各角色工具与 persona、专用模型、启动方式、后台续跑、主 agent 工具增删），无改动时一行带过；瘦身统计在这里按当前未保存配置另示一行。
 - **保存并应用**：确认后写入 preset（覆盖 `$DSH_HOME/.agent-presets/orchestrator`）并按宿主版本同步登记——≤ 0.1.6 目录扫描即生效，≥ 0.1.7 同步维护 `$DSH_HOME/cordis.patch.yml` 里的声明行托管块（见第 2 节）——再把当前分配保存到 `$DSH_HOME/dsh-paoding.config.yml`；应用成功后提示重启 DSH 或新建会话生效（检测到 patch 比生成结果新时提示重启后生效）。打开面板时会加载既有配置文件作为当前值——手编过的选择直接可见、可改。
 - **场景模板**：页头三档切换（轻量：只留实现角色；调研：只留检索角色；全开：全员到齐、主 agent 额外工具带满检测库存），一键换「班底」；只动 `roles_remove` 与 `main_agent_extra`，各角色精调过的工具面、persona、模型、启动方式分毫不动，切回即原样复原。
 - **检查提示卡**：每次「预览生成 / 保存并应用」自动过一遍体检（配置了未挂载的工具、零工具角色、查无此名的模型、预设落后于配置），分级提示以卡片形式贴在页头下方，处理完再保存不迟。
-- **「更多」菜单**：配置导出（下载 `dsh-paoding.config.yml` 原文）/ 导入（先验格式，整份覆盖含工作区段，需二次确认）与面板语言切换（中英自动识别、手动选择优先并记住）。
+- **备份与回滚**：保存、导入、恢复配置前都会自动把当前配置轮换备份三份（`config.yml.bak-1..3`，新压旧）；「更多」→「恢复备份…」列时间与大小、二次确认后回退，恢复前当前配置还会再自动备份一份——误操作有后悔药。
+- **委派效果统计**：页头「委派效果」卡按角色统计被委派的调用数、平均耗时与失败数（近 30 天，`days` 参数可调）。数据来自宿主会话存档的冷读——经宿主 `sessionPersistence` 服务逐会话抽摘要、本地聚合，不出机、不带负载；失败记的是子会话零轮次或轮次中断（spawn 后即死、中途终止两种形态）。宿主缺会话服务时整卡不渲染。
+- **「更多」菜单**：配置导出（下载 `dsh-paoding.config.yml` 原文）/ 导入（先验格式，整份覆盖含工作区段，需二次确认）/ 恢复备份（见上条）与面板语言切换（中英自动识别、手动选择优先并记住）。
 
 ## 7. 卸载
 
-三条命令，删得干干净净：
+先用一条命令清掉庖丁写进宿主目录的东西（托管声明块、`orchestrator*` 预设目录、版本标记；配置文件默认保留，`--purge` 才连配置与备份一起删；非交互终端需加 `--yes`）：
+
+```bash
+npx dsh-paoding@latest uninstall
+```
+
+包本体再按 profile 摘掉，三条命令删得干干净净：
 
 ```bash
 dsh plugin --profile web remove dsh-paoding
@@ -216,7 +224,7 @@ rm -rf "${DSH_HOME:-$HOME/.dsh}/.agent-presets"/orchestrator-*
 
 第一条摘掉插件与侧栏面板（只影响指定 profile），第二条删掉编排预设目录，第三条清掉按工作区配置时生成的各工作区预设目录（`orchestrator-<目录名>`；没用过按工作区配置就没有这类目录，通配不命中、无害）——基础 `orchestrator` 目录已由上一条命令删掉，不含在这个通配里。配置文件 `~/.dsh/dsh-paoding.config.yml` 不随之上删——留着或删掉都不影响 DSH 其它配置，重装时首装自动化会再次按它应用。
 
-宿主为 DSH ≥ 0.1.7 时再多看一眼：`$DSH_HOME/cordis.patch.yml` 里可能还留着 dsh-paoding 托管块（起止标记注释包着 `- insert:` 声明行）。预设目录删掉后，块内声明行就成了指向已删目录的悬空引用，请把整个托管块（从 `# --- dsh-paoding presets (auto-generated; do not edit) ---` 起始行到 end 标记行）一并删去——块外内容不受影响。
+（走了上面的 `uninstall` 就免了这一步和下一眼的托管块手删——它已一并清掉。）宿主为 DSH ≥ 0.1.7 时若手动删目录，再多看一眼：`$DSH_HOME/cordis.patch.yml` 里可能还留着 dsh-paoding 托管块（起止标记注释包着 `- insert:` 声明行）。预设目录删掉后，块内声明行就成了指向已删目录的悬空引用，请把整个托管块（从 `# --- dsh-paoding presets (auto-generated; do not edit) ---` 起始行到 end 标记行）一并删去——块外内容不受影响。
 
 卸载不影响 host 的其它内容（MCP、插件、会话等），与 DSH 源码零耦合；之后新建会话时「编排模式 (Orchestrator)」不再出现在预设选择器里。
 
@@ -232,6 +240,8 @@ npx dsh-paoding@latest doctor          # 或仓库内：node tools/install.mjs d
 
 doctor 不改任何文件，逐项检查并给出 `✔ 正常 / ⚠ 警告 / ✘ 错误 / – 跳过` 四档结论：配置文件能否解析、预设安装轨道与宿主版本是否匹配、托管声明行有无孤儿行或损坏块、`.agent-presets/orchestrator-*` 有没有配置里已不引用的孤儿目录、已装预设与当前配置的再生成结果有无漂移、`.generator-version` 版本标记是否对得上。结尾有汇总；出现 ✘ 时退出码为 1，⚠ 与 – 不影响使用（跳过多见于全新机器，属正常）。`--config` / `--profile` 与安装时同义。
 
+`doctor --fix` 会在这份诊断之后自动修两类可安全处理的问题——孤儿 `orchestrator-*` 预设目录、托管块里的孤儿/过期声明行（写盘前自动备份 `.bak-1`），其余发现只给处理建议不代做（预设漂移就提示跑 `--auto` 重新生成）；输出逐项标注「已修 / 需手动」。非交互终端下 `--fix` 与 `uninstall` 都要显式 `--yes` 才落盘，否则只列计划、退出码 2。
+
 面板侧也有同源的保存时体检：每次「预览生成 / 保存并应用」的响应都带 `warnings` 字段（配置了未挂载工具、零工具角色、角色专用模型不在模型目录、预设落后于配置等），按条目提示处理即可，不用等运行时报错。
 
 **pnpm 缺失**

package/docs/installation_en.md

@@ -191,22 +191,30 @@ If you'd rather not hand-edit `dsh-paoding.config.yml`, open the **庖丁配置
 
 ### Data interface and security
 
-Panel data goes over the **same-origin `/api/paoding/*`**: the Node side registers prefix routes — `GET /api/paoding/state` (cached detection state), `GET /api/paoding/version` (latest-version check, the data source of the version/update card), `GET /api/paoding/models` (model list from the host LLM runtime), `GET` + `POST /api/paoding/workspaces` (list / add workspaces), `POST /api/paoding/rescan` (forced re-detection, MCP handshakes included), `POST /api/paoding/preview` (generate without writing), `POST /api/paoding/apply` (install and save config), `POST /api/paoding/upgrade` (in-place upgrade: `dsh plugin update` for regular profiles, a direct `pnpm add` for the desktop profile — pnpm is resolved through a candidate chain: bare `pnpm` → `corepack pnpm` → PNPM_HOME locations), `GET /api/paoding/config/export` (raw config file for download/backup), `POST /api/paoding/config/import` (whole-file import: validated first, shares the apply write mutex), `GET /api/paoding/client.css` (statically served panel stylesheet) — reusing the detection/generation pipeline in `tools/` (`plugins/paoding-config-ui/api-core.mjs`). The routes carry a **browser trust fence**: the Host must be loopback or listed in `webRuntime.trustedHosts`, and cross-site requests are rejected (mirroring DSH's `/api` gateway fence semantics, since the longer `/api/paoding` prefix would otherwise hit this plugin and bypass the gateway). The panel has **no separate server or port**: the DSH Web port listens on `127.0.0.1` only and is never exposed to the outside.
+Panel data goes over the **same-origin `/api/paoding/*`**: the Node side registers prefix routes — `GET /api/paoding/state` (cached detection state), `GET /api/paoding/version` (latest-version check, the data source of the version/update card), `GET /api/paoding/models` (model list from the host LLM runtime), `GET` + `POST /api/paoding/workspaces` (list / add workspaces), `POST /api/paoding/rescan` (forced re-detection, MCP handshakes included), `POST /api/paoding/preview` (generate without writing), `POST /api/paoding/apply` (install and save config), `POST /api/paoding/upgrade` (in-place upgrade: `dsh plugin update` for regular profiles, a direct `pnpm add` for the desktop profile — pnpm is resolved through a candidate chain: bare `pnpm` → `corepack pnpm` → PNPM_HOME locations), `GET /api/paoding/config/export` (raw config file for download/backup), `POST /api/paoding/config/import` (whole-file import: validated first, shares the apply write mutex), `GET /api/paoding/config/backups` (automatic backup list), `POST /api/paoding/config/restore` (restore a backup: filename whitelist against path traversal, validated for parseability before overwriting), `GET /api/paoding/stats` (delegation stats: cold-read aggregation over the session archive, with a days window and project filter), `GET /api/paoding/client.css` (statically served panel stylesheet) — reusing the detection/generation pipeline in `tools/` (`plugins/paoding-config-ui/api-core.mjs`). The routes carry a **browser trust fence**: the Host must be loopback or listed in `webRuntime.trustedHosts`, and cross-site requests are rejected (mirroring DSH's `/api` gateway fence semantics, since the longer `/api/paoding` prefix would otherwise hit this plugin and bypass the gateway). The panel has **no separate server or port**: the DSH Web port listens on `127.0.0.1` only and is never exposed to the outside.
 
 ### Panel capabilities
 
-- **State and actions**: the page header is a title + version/update card (in-place new-version notice and one-click upgrade) + subtitle, and the body goes straight to the workspace and role cards with no separate detection-summary/chips section; the action buttons「**保存并应用**」(Save & Apply), Preview, and Rescan sit in a sticky bottom dock, clickable from any scroll position.
-- **Agent tool-assignment cards**: one fixed card for the main agent (unchecking a base tool adds it to `main_agent_remove`, dropping it from the main agent; checking an MCP / plugin tool adds it to `main_agent_extra`; skills in two groups, "read on demand" and "inline full text"; plus a "persona extra" editor — edit, clear, or restore the default) ＋ three built-in role cards (`search_external` / `design` / `implement`) ＋ any number of custom role cards (with a delete button); every role card presents its tools in three groups — base / MCP / plugin — with per-item checkboxes, and the persona is editable. "Create a custom agent role" at the bottom.
+- **State and actions**: the page header is a title + version/update card (in-place new-version notice and one-click upgrade) + subtitle, with a permanent main-agent tool-surface line under the subtitle (current allow count vs the monolithic total, plus the slimming percentage — the first visible payoff of configuring), and the body goes straight to the workspace and role cards with no separate detection-summary/chips section; the action buttons「**保存并应用**」(Save & Apply), Preview, and Rescan sit in a sticky bottom dock, clickable from any scroll position.
+- **Agent tool-assignment cards**: one fixed card for the main agent (unchecking a base tool adds it to `main_agent_remove`, dropping it from the main agent; checking an MCP / plugin tool adds it to `main_agent_extra`; skills in two groups, "read on demand" and "inline full text"; plus a "persona extra" editor — edit, clear, or restore the default) ＋ three built-in role cards (`search_external` / `design` / `implement`) ＋ any number of custom role cards (with a delete button); every role card presents its tools in three groups — base / MCP / plugin — with per-item checkboxes (any group over 8 items grows a filter box on top; type to narrow), and the persona is editable. "Create a custom agent role" at the bottom.
 - **Skill assignment**: one row per skill, check which agents it goes to (written into the matching personas as `Available skills` soft guidance; role skills are soft guidance, main-agent skills have the read-on-demand/inline groups).
-- **Preview**: generates `agent.cordis.yml` without writing, showing each role's intent → kept counts plus the full generated text.
+- **Preview**: generates `agent.cordis.yml` without writing, showing each role's intent → kept counts plus the full generated text; it also lists "changes since the last save" — an itemized diff (roles added/removed, each role's tools and persona, dedicated model, start mode, background continuation, main-agent tool changes) reduced to a single line when nothing changed; the tool-surface stat appears here too, computed from the unsaved config.
 - **Save & Apply**: after confirmation, writes the preset (overwriting `$DSH_HOME/.agent-presets/orchestrator`) and registers it per host generation — on ≤ 0.1.6 the directory scan picks it up as-is, on ≥ 0.1.7 the declaration managed block in `$DSH_HOME/cordis.patch.yml` is maintained in the same pass (see §2) — then saves the current assignments to `$DSH_HOME/dsh-paoding.config.yml`; success asks you to restart DSH or start a new session (and flags when a patch is newer than the generated output). Opening the panel loads the existing config file as the current values — hand-edited choices are visible and editable right away.
 - **Scenario templates**: three one-click presets in the header (Lite: implement role only; Research: search roles only; Full: everyone on deck with the main agent carrying the whole detected inventory) swap the roster — they only touch `roles_remove` and `main_agent_extra`; each role's hand-tuned tools, persona, model, and start mode stay untouched and restore as-is when you switch back.
 - **Health-check card**: every Preview / Save & Apply runs a check pass (tools configured but not mounted, zero-tool roles, unknown model names, a preset lagging behind the config); findings show as a severity-tagged card under the header so you can fix them before saving.
-- **"More" menu**: config export (download the raw `dsh-paoding.config.yml`) / import (validated first; overwrites the whole file, workspaces included, with confirmation) and the panel language toggle (English/Chinese auto-detected, manual choice wins and is remembered).
+- **Backups and rollback**: every write to the config (save & apply, import, restore) first rotates the current file through three backups — `config.yml.bak-1` (newest) through `.bak-3`; "More" → "Restore backup…" lists each backup's time and size and, after a confirming dialog, puts it back (the current file is backed up once more before the restore, so even a wrong restore can be undone).
+- **Delegation stats**: the "Delegation" card in the header counts, per role, delegations, average duration, and failures (last 30 days, adjustable via `days`). The data is a cold read of the host session archive — summaries drawn per session through the host `sessionPersistence` service and aggregated locally; nothing leaves the machine and no payloads are read. A failure means the child session ran zero turns or was interrupted mid-turn (died right after spawn, or cut off). When the host lacks the session service the card simply doesn't render.
+- **"More" menu**: config export (download the raw `dsh-paoding.config.yml`) / import (validated first; overwrites the whole file, workspaces included, with confirmation) / restore backup (see above) and the panel language toggle (English/Chinese auto-detected, manual choice wins and is remembered).
 
 ## 7. Uninstalling
 
-Three commands, and it's gone cleanly:
+One command first clears what dsh-paoding wrote into the host directory (the managed declaration block, the `orchestrator*` preset directories, the version marker; the config file is kept unless `--purge`; non-interactive terminals need `--yes`):
+
+```bash
+npx dsh-paoding@latest uninstall
+```
+
+Then lift the package itself off the profile with three commands:Three commands, and it's gone cleanly:
 
 ```bash
 dsh plugin --profile web remove dsh-paoding

package/docs/orchestration.md

@@ -10,7 +10,7 @@
 dsh-paoding 把 DSH 从「单体 agent 全量加载工具」改造成「主 agent 编排 + 角色子 agent 分工」。分工遵循任务纹理：
 
 - **热路径——内部文件问题不委派。** 找代码、查引用、读文件这类问题由主 agent 自己用只读工具解决：`glob` / `grep` / `read` / `read_image` / `bash` 全部在主 agent 白名单内，零委派往返延迟。persona 文案里的 `codegraph` 是 host MCP 工具，仅在安装期检测到 codegraph 服务器并经 `main_agent_extra` 注入后才真正可用（详见[配置](configuration.md)与[安装](installation.md)）。
-- **冷路径——只干自己那活的角色子 agent。** 外部调研委派 `search_external`（只带 `web_search` + tavily 五工具 + 本地只读）、UI/前端设计委派 `design`（读参考 + 设计技能 + 出 spec，不写最终代码）、代码实现委派 `implement`（读上下文 → 精准修改 → 验证）、超大仓库全量探索委派 `search_internal_deep`（第四角色，2026-09 加入，见下节与[架构](architecture.md)）。每个子 agent 只在真正被委派的那一次才加载工具面（按需加载）。
+- **冷路径——只干自己那活的角色子 agent。** 外部调研委派 `search_external`（只带 `web_search` / `web_fetch` + tavily 五工具 + 本地只读）、UI/前端设计委派 `design`（读参考 + 设计技能 + 出 spec，不写最终代码）、代码实现委派 `implement`（读上下文 → 精准修改 → 验证）、超大仓库全量探索委派 `search_internal_deep`（第四角色，2026-09 加入，见下节与[架构](architecture.md)）。每个子 agent 只在真正被委派的那一次才加载工具面（按需加载）。
 - **运行中指挥。** 子 agent 运行期间，主 agent 可用 `send_message` 追加指示、`list_agents` 查看状态、`interrupt_agent` 叫停（三者由 `dsh-tool-subagent-control` 注册，属「子 agent 管理」工具面）。
 - **完成后只集成摘要。** 主 agent 从不把子 agent 的完整上下文倒进自己的上下文——它只收一份摘要（见第 5 节「上下文隔离」）。
 
@@ -22,7 +22,7 @@ dsh-paoding 把 DSH 从「单体 agent 全量加载工具」改造成「主 agen
 | 子 agent 管理 | `send_message` / `list_agents` / `interrupt_agent` |
 | 内部搜索（热路径） | `glob` / `grep` / `read` / `read_image` / `bash` |
 | 协调 | `todo_write` / `ask_user_question` / `get_goal` / `create_goal` / `update_goal` / `exit_plan_mode` / `job_output` / `job_list` / `job_kill` |
-| 主 agent 刻意看不到（节选，全文见 `restrict.mjs` 头注释） | `web_search`、`mcp__tavily__*`、`write` / `edit` / `str_replace_editor`、`skill`、`workflow`、`ralph`、`ssh_*`、`mcp__tablepro__*`、`describe_image`、裸 `subagent` / `subagent_fork` |
+| 主 agent 刻意看不到（节选，全文见 `restrict.mjs` 头注释） | `web_search` / `web_fetch`、`mcp__tavily__*`、`write` / `edit` / `str_replace_editor`、`skill`、`workflow`、`ralph`、`ssh_*`、`mcp__tablepro__*`、`describe_image`、裸 `subagent` / `subagent_fork` |
 
 每请求工具开销（估算，随注册面浮动）：主 agent 全量 ~16.2k tokens → 白名单后 ~5.5k；`search_external` ~3.0k / `design` ~3.3k / `implement` ~3.8k，仅在真正委派时发生（数值来源：`agent.cordis.yml` 头部注释）。
 

package/docs/orchestration_en.md

@@ -10,7 +10,7 @@ Orchestration is the core of dsh-paoding: the main agent (the orchestrator) deco
 dsh-paoding turns DSH from "a single agent that loads every tool" into "a main agent that orchestrates plus role agents that specialize". The split follows the texture of the task:
 
 - **Hot path — internal file questions are never delegated.** Finding code, chasing references, and reading files are solved by the main agent itself with its read-only tools: `glob` / `grep` / `read` / `read_image` / `bash` are all on the main-agent allow list, so there is zero delegation round-trip latency. `codegraph` in the persona copy is a host MCP tool: it is only actually available when the install-time detection finds the codegraph server and it is injected through `main_agent_extra` (see [Configuration](configuration_en.md) and [Installation](installation_en.md)).
-- **Cold path — role agents that carry only their own tools.** External research goes to `search_external` (only `web_search` + the five tavily tools + local reads); UI/frontend design goes to `design` (reads references, loads design skills, produces specs, never writes final code); code implementation goes to `implement` (read context → surgical change → verify); whole-repo deep exploration goes to `search_internal_deep` (a fourth role added 2026-09, see next section and [Architecture](architecture_en.md)). Each subagent is paid for only the run in which it is actually used (pay-per-use).
+- **Cold path — role agents that carry only their own tools.** External research goes to `search_external` (only `web_search` / `web_fetch` + the five tavily tools + local reads); UI/frontend design goes to `design` (reads references, loads design skills, produces specs, never writes final code); code implementation goes to `implement` (read context → surgical change → verify); whole-repo deep exploration goes to `search_internal_deep` (a fourth role added 2026-09, see next section and [Architecture](architecture_en.md)). Each subagent is paid for only the run in which it is actually used (pay-per-use).
 - **Directing a running child.** While a subagent runs, the main agent can append instructions with `send_message`, inspect status with `list_agents`, and stop it with `interrupt_agent` (all three are registered by `dsh-tool-subagent-control` and belong to the "child management" surface).
 - **Integrate summaries only.** The main agent never dumps a child's full context into its own — it receives one summary (see section 5, Context isolation).
 
@@ -22,7 +22,7 @@ The main agent's model-facing surface is narrowed by `restrict.mjs` on the `syst
 | Child management | `send_message` / `list_agents` / `interrupt_agent` |
 | Internal search (hot path) | `glob` / `grep` / `read` / `read_image` / `bash` |
 | Coordination | `todo_write` / `ask_user_question` / `get_goal` / `create_goal` / `update_goal` / `exit_plan_mode` / `job_output` / `job_list` / `job_kill` |
-| Deliberately hidden from the main agent (excerpt; full list in the `restrict.mjs` header comment) | `web_search`, `mcp__tavily__*`, `write` / `edit` / `str_replace_editor`, `skill`, `workflow`, `ralph`, `ssh_*`, `mcp__tablepro__*`, `describe_image`, bare `subagent` / `subagent_fork` |
+| Deliberately hidden from the main agent (excerpt; full list in the `restrict.mjs` header comment) | `web_search` / `web_fetch`, `mcp__tavily__*`, `write` / `edit` / `str_replace_editor`, `skill`, `workflow`, `ralph`, `ssh_*`, `mcp__tablepro__*`, `describe_image`, bare `subagent` / `subagent_fork` |
 
 Per-request tool tax (estimates; they float with the registered surface): full monolithic main agent ~16.2k tokens → ~5.5k after the allow list; `search_external` ~3.0k / `design` ~3.3k / `implement` ~3.8k, incurred only when actually delegated (figures from the header comment of `agent.cordis.yml`).
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-paoding",
-  "version": "0.4.4",
+  "version": "0.4.6",
   "description": "庖丁 — DSH (DeepSeek Harness) 编排模式安装器：host 感知的角色工具过滤 + Web 设置页可视化配置器",
   "type": "module",
   "license": "MIT",
@@ -20,7 +20,8 @@
     "cordis.patch.yml"
   ],
   "engines": {
-    "node": ">=18"
+    "node": ">=18",
+    "dsh": ">=0.1.5 <0.3"
   },
   "peerDependencies": {
     "@deepseek-ai/dsh": ">=0.1.5"