Merge branch 'main' into feat/help-feedback-entry

Merge pull request #1021 from DingTalk-Real-AI/fix/test-focused-shard-matrix
fix(ci): shard the focused test job instead of one long-lived run
2026-08-16 06:57:18 +08:00 · 2026-08-15 23:33:12 +08:00 · 2026-08-15 22:58:30 +08:00 · 2026-08-15 22:32:39 +08:00 · 2026-08-15 22:10:28 +08:00 · 2026-08-15 16:38:57 +08:00
1027 changed files with 175787 additions and 23230 deletions
@@ -0,0 +1,5 @@
+---
+category: Added
+---
+
+- **Privacy-safe CLI telemetry** (#1009) — reports reviewed command outcomes and profile identity dimensions while excluding command arguments, output, paths, device fingerprints, and automatic system dimensions; `DO_NOT_TRACK=1` disables reporting.
@@ -0,0 +1,5 @@
+---
+category: Added
+---
+
+- **Feedback survey entry in root help** (#1019) — `dws --help` now closes with a Feedback section linking the user-experience survey form.
@@ -0,0 +1,7 @@
+---
+category: Changed
+---
+
+- **Chat IM ID flags** (#954) — standardizes chat command entry points on `--conversation-id` for conversation IDs and `--message-id` for message IDs, so help, Schema, and Agent recommendations use the same canonical flags.
+- **Legacy chat flag compatibility** (#954) — keeps older chat IM ID flags such as `--group`, `--id`, `--chat`, `--open-conversation-id`, `--msg-id`, and `--open-message-id` working as compatibility aliases where applicable, while hiding migrated aliases from recommended help and Schema surfaces.
+- **Chat group bots target flag** (#954) — keeps `dws chat group bots` on the visible `--group` flag; this command does not register `--group-name`, and `--group` accepts either an openConversationId or a uniquely resolved group name.
@@ -0,0 +1,34 @@
+# Release fragments
+
+普通功能、修复和面向用户的行为变更不要再修改根目录 `CHANGELOG.md` 的
+`Unreleased` 区域。每个 PR 在本目录新增一个独立的 Markdown fragment，避免
+并行 PR 争用同一文件。
+
+文件名使用能唯一定位变更的短名，通常是 PR 号，例如
+`1234-chat-reply-mentions.md`。文件名必须匹配
+`^[a-z0-9][a-z0-9._-]*\.md$`，且必须是普通文件，不能是符号链接。本目录顶层
+只接受 `README.md`、`released/` 和符合该规则的 fragment：fragment 一律平铺在
+顶层，不接受任何其它子目录，本目录自身也不能被替换成文件或符号链接。其余条目
+会被 CI 直接拒绝而不是忽略，以免非法条目跳过校验后拖垮下一个 PR。文件格式
+严格如下：
+
+```markdown
+---
+category: Added
+---
+
+- **Chat reply mentions** (#1234) — supports mentioning selected members.
+```
+
+`category` 只能是 `Added`、`Changed`、`Deprecated`、`Removed`、`Fixed` 或
+`Security`。正文至少包含一个 Markdown 列表项，且不得包含 `TODO` 或 `TBD`。
+
+发布 beta 时，`scripts/release/prepare-changelog.sh` 会按分类和文件名稳定排序，
+将未归档 fragments 汇总为唯一的版本章节，并移动到
+`.changes/released/<version>/`。因此 release-seal PR 是唯一会修改
+`CHANGELOG.md` 的 PR；它同时归档已消费的 fragments，供审计追溯。
+归档只能在同一个 release-seal PR 中以原样移动完成；CI 会拒绝直接修改、
+删除或重写已归档文件。
+
+无需面向用户发布说明的改动不添加 fragment。评审者根据改动是否可见来判断该
+例外是否成立。
@@ -0,0 +1,6 @@
+---
+category: Fixed
+---
+
+- **Chat card update evidence** — distinguishes an accepted update request from an independently verified visible update, preserving the real `bizId` and warning callers not to repeat an unverified write.
+- **Chat command guidance** — splits message and group references by task and explains that `--from` is ambiguous between sender and time-range intent.
@@ -0,0 +1,8 @@
+---
+category: Added
+---
+
+- **Agent version and extended context passthrough** (Aone 85384225) — adds
+  validated `DWS_AGENT_VER` and sensitive JSON `DWS_AGENT_EXT` metadata to
+  ordinary non-plugin MCP requests without forwarding it to A2A, OAuth,
+  Discovery, or third-party plugins.
@@ -0,0 +1,5 @@
+---
+category: Changed
+---
+
+- **Chat message send help** - Clarifies Markdown image syntax for inline mixed text and images.
@@ -0,0 +1,5 @@
+---
+category: Added
+---
+
+- **Drive file comments** (#961) — adds `dws drive comment list` and `dws drive comment create` for comments on ordinary preview files.
@@ -0,0 +1,5 @@
+---
+category: Added
+---
+
+- **Chat automatic pagination controls** (#970) — adds bounded `--max-items` and cancellable `--page-delay` support to the core IM list shortcuts, with safe continuation metadata and truncation reporting.
@@ -0,0 +1,5 @@
+---
+category: Changed
+---
+
+- **Doc/drive/wiki routing descriptions** — clarifies the document-space container-vs-content boundary across the doc, drive, and wiki skill descriptions for more predictable first-round Agent selection, without changing CLI behavior.
@@ -0,0 +1,20 @@
+---
+category: Fixed
+---
+
+- **Drive `--latest` refuses incomplete Top-N** (#899) — `dws drive list --latest` used to
+  exit 0 with a "Top-N" computed over a partially scanned tree whenever a directory read
+  failed mid-recursion (permission denied, API error), letting an incomplete set pose as the
+  globally newest files. Truncation at the 2000-item scan cap and mid-recursion directory
+  failures now both fail closed (`LATEST_SCAN_TRUNCATED` / `LATEST_SCAN_INCOMPLETE`), report
+  the first failing folder with its depth and reason, and emit a recovery command that
+  reproduces the original candidate set — query domain, `--folder`, `--pattern`, `--type`,
+  `--start` and `--end` are all carried over. On POSIX shells each user-supplied value is
+  quoted so a URL query string or a shell metacharacter cannot change how the copied command
+  parses. On Windows no quoting form is safe for both `cmd.exe` and PowerShell, so values
+  containing metacharacters are not inlined at all: the command carries a placeholder and the
+  original value is shown on a separate line marked as data rather than an executable command.
+  Unrecoverable errors under `--latest` return the root cause instead of a partial result.
+  Remote-controlled folder names and server error text are stripped of ANSI escapes and
+  control characters before they reach the plain-text stderr message. The internal `sortTime`
+  sort key no longer leaks into `drive list --depth` output on any path.
@@ -0,0 +1,12 @@
+---
+category: Added
+---
+
+- **Drive list type/time filtering** (#942) — `dws drive list` gains `--type
+  file|folder`, `--start`, and `--end` for client-side filtering by node type
+  and modification time on both the pan and workspace routes. Filtering runs
+  a bounded full scan of the target directory (2000-entry cap, reported via
+  `truncated=true`), composes with `--latest`/`--pattern`/`--depth`, and is
+  mutually exclusive with `--versions`/`--cursor`/`--order-by`/`--order`/
+  `--limit`. Time values accept relative forms (`24h`/`7d`/`2w`), RFC 3339,
+  zone-less ISO 8601 (Asia/Shanghai), or a plain date.
@@ -0,0 +1,12 @@
+---
+category: Fixed
+---
+
+- **Drive list pattern filtering** (#942) — `dws drive list --pattern` on the
+  single-layer pan route now filters the returned page by name pattern; the
+  flag was previously accepted but silently ignored.
+
+- **Drive list `--type folder --latest` composition** (#942) — `--latest` now
+  ranks the filtered entries (folders included when `--type folder` is set)
+  instead of unconditionally dropping folders, so the documented combination
+  returns the most recently modified folders rather than an empty list.
@@ -0,0 +1,5 @@
+---
+category: Fixed
+---
+
+- **Chat message time defaults** (#973) — default omitted `chat message list-all` time bounds in `Asia/Shanghai` when emitting timezone-less `yyyy-MM-dd HH:mm:ss` values, matching parsing semantics and rejecting reversed windows.
@@ -0,0 +1,5 @@
+---
+category: Fixed
+---
+
+- **Doc and Drive parameter aliases** — normalizes reviewed identifier, pagination, path, version, and role synonyms while blocking ambiguous values before dispatch.
@@ -0,0 +1,15 @@
+---
+category: Added
+---
+
+- **Drive folder synchronization** — adds `dws drive status`, `dws drive pull`,
+  `dws drive push`, and `dws drive sync` for file-level comparison and transfer
+  between a local folder and a Drive folder. Differences come from exact MD5 by
+  default or from modification time with `--quick`; `status` is read-only, `pull`
+  and `push` are one-directional with `--if-exists skip|smart|overwrite`, and
+  `sync` is bidirectional with `--on-conflict remote-wins|local-wins|keep-both|ask`.
+  Only regular files are transferred — online documents and shortcuts are skipped,
+  neither side deletes extra files, downloads are staged through a temporary file
+  and committed with an atomic rename, and remote names that would escape
+  `--local-folder` are reported as failures instead of being written. Every command
+  prints a structured summary on stdout and exits non-zero when any item fails.
@@ -0,0 +1,5 @@
+---
+category: Added
+---
+
+- **International DingTalk region support** — adds `.io` login and MCP routing, pre-release endpoint overrides, and profile-aware gateway selection while preserving the existing `.com` flow.
@@ -0,0 +1,5 @@
+---
+category: Changed
+---
+
+- **Chat identity routing** — validates explicit `openDingTalkId` inputs and improves name, `userId`, and `openDingTalkId` routing for message shortcuts.
@@ -0,0 +1,9 @@
+---
+category: Added
+---
+
+- **Wiki Shortcut workflows** — publishes 20 reviewed space, member, node, and
+  activity shortcuts with strict collection validation, cursor handling,
+  write-terminal evidence, safe read-backs where the backend supports them,
+  task-oriented routing, and documented backend
+  boundaries.
@@ -19,3 +19,12 @@

 # Cache directory (optional, defaults to ~/.dws/cache)
 # DWS_CACHE_DIR=
+
+# Agent integration metadata (optional; ordinary non-plugin MCP requests only)
+# DWS_AGENT_PRODUCT=example-agent
+# DWS_AGENT_HOST=cloud
+# DWS_AGENT_VER=0.1.5
+# DWS_AGENT_EXT='{"umt":"example-redacted","miniwua":"example-redacted","ua":"ExampleAgent/0.1.5"}'
+# The outer single quotes above are shell syntax and are not part of the value.
+# DWS_AGENT_EXT is sensitive caller-declared JSON (max 8 KiB); never put real
+# tokens in committed files or use this metadata alone for authentication.
@@ -19,8 +19,10 @@ repeat the entire CI suite locally only to fill this checklist: CI expands the
 selected tier from documentation checks, through affected-package tests, to
 the complete high-risk suite.

- [ ] Exact in-place `CHANGELOG.md`-only check (otherwise `N/A`):
-  `./scripts/policy/check-changelog-pr.sh --fast-path "$(git merge-base HEAD origin/main)" HEAD`
+- [ ] Release fragment added for a user-visible behavior/interface change (otherwise `N/A`):
+  `.changes/<unique-name>.md`; ordinary PRs must not edit `CHANGELOG.md`.
+- [ ] Release-seal validation (otherwise `N/A`):
+  `./scripts/policy/check-changelog-pr.sh --content-only "$(git merge-base HEAD origin/main)" HEAD`
 - [ ] Targeted test/check commands and results:
 - [ ] Behavior evidence (test name, CLI output shape, or before/after result):
 - [ ] Documentation links/content/rendering checked (documentation-only, otherwise
@@ -0,0 +1,61 @@
+# /eval 自助触发允许名单
+#
+# 名单内的 GitHub 登录名可对【自己创建的 PR】触发 /eval 评测；
+# 对任意 PR 触发仍需仓库 write/maintain/admin 权限（维护者背书）。
+# 授权读取的始终是默认分支上的本文件，PR 无法修改自身授权。
+#
+# 变更本文件必须走 PR 评审。每行一个 GitHub login，# 开头为注释。
+
+aftersss
+notable-open
+EdgarWang0925
+ayunya
+yutongshe
+qingyang1014
+caiTriumph
+xlb1130
+Anonymity-0
+FuShu-Yang
+guimingyue
+AlwaysLee
+TaoJikun
+zengyoulingzyl-stack
+liyuan333
+huangyoo
+lifeihong
+nitonitori
+cywan1998
+gangwn
+junlonghuo2
+aqruan
+Freda0909
+ShawnWhite777
+PeterGuy326
+abucraft
+pengzhihan47-star
+rainyak8
+gongrongyun
+huangyuanzhuo-coder
+ybcstudy
+bigqy
+liwang-ai
+meng93
+wxianfeng
+Patrick-Star-CN
+rossluo28-hz
+dxy704330469
+gtezg30062
+Neige-Premaire
+zhuoyu20
+avicii-chen
+typefield
+Haofeng0705
+Huwenjiao
+liuzeyang
+maoqxxmm
+FloralTide
+lingyun9833
+dxb121
+C0922
+xiaoji121
+H3java
@@ -0,0 +1,285 @@
+'use strict';
+
+// 评审归属是受保护分支上的声明式规则；未知路径不猜测，交给工作流负载均衡兜底。
+const REVIEWER_POOL = ['wxianfeng', 'typefield', 'haofeng0705', 'hlzjsong'];
+
+const PRODUCT_GROUPS = [
+  {
+    primary: 'wxianfeng',
+    backup: 'typefield',
+    products: ['chat', 'contact', 'ding', 'event', 'mail', 'live', 'conference', 'dev', 'devapp', 'mcp', 'aiapp'],
+  },
+  {
+    primary: 'typefield',
+    backup: 'wxianfeng',
+    products: ['doc', 'drive', 'wiki', 'markdown', 'docparse', 'aidesign', 'devdoc', 'blackboard', 'finance', 'law', 'credit'],
+  },
+  {
+    primary: 'haofeng0705',
+    backup: 'typefield',
+    products: ['minutes', 'sheet', 'aitable', 'calendar', 'todo', 'oa', 'attendance', 'report', 'agoal', 'aisearch', 'yida', 'hrbrain'],
+  },
+];
+
+const pathStartsWith = (prefixes) => (path) => prefixes.some((prefix) => path.startsWith(prefix));
+
+const MODULES = [
+  {
+    id: 'security',
+    label: '登录、认证、权限、安全',
+    primary: 'hlzjsong',
+    backup: 'typefield',
+    requiresSecondary: true,
+    matches: pathStartsWith([
+      'internal/auth/',
+      'internal/keychain/',
+      'internal/audit/',
+      'internal/pat/',
+      'internal/security/',
+      'internal/safety/',
+      'pkg/edition/',
+    ]),
+  },
+  {
+    id: 'delivery',
+    label: 'CI、测试、发布、安装',
+    primary: 'haofeng0705',
+    backup: 'wxianfeng',
+    requiresSecondary: true,
+    matches: (path) =>
+      path.startsWith('.github/') ||
+      path.startsWith('scripts/release/') ||
+      path.startsWith('scripts/policy/') ||
+      path.startsWith('scripts/dev/') ||
+      path.startsWith('scripts/install') ||
+      path.startsWith('Formula/') ||
+      path.startsWith('build/') ||
+      path.startsWith('internal/upgrade/') ||
+      path.startsWith('internal/app/upgrade') ||
+      path.startsWith('test/') ||
+      path.startsWith('verify/') ||
+      path.startsWith('.workflow/') ||
+      path === 'coverage.txt' ||
+      path === 'coverage-base.txt' ||
+      path === '.goreleaser.yaml' ||
+      path === 'package.json' ||
+      path === 'package-lock.json' ||
+      path === 'docs/releasing.md',
+  },
+  {
+    id: 'architecture',
+    label: 'DWS 架构、公共内核',
+    primary: 'wxianfeng',
+    backup: 'typefield',
+    requiresSecondary: true,
+    matches: pathStartsWith([
+      'cmd/',
+      'internal/apiclient/',
+      'internal/app/',
+      'internal/cli/',
+      'internal/cobracmd/',
+      'internal/corecmd/',
+      'internal/errors/',
+      'internal/executor/',
+      'internal/generator/',
+      'internal/i18n/',
+      'internal/interfacesnapshot/',
+      'internal/jsonutil/',
+      'internal/localio/',
+      'internal/logging/',
+      'internal/output/',
+      'internal/pipeline/',
+      'internal/plugin/',
+      'internal/profilectx/',
+      'internal/registry/',
+      'internal/syncdata/',
+      'internal/testseam/',
+      'internal/transport/',
+      'pkg/',
+    ]),
+  },
+  {
+    id: 'compatibility',
+    label: '兼容性',
+    primary: 'wxianfeng',
+    backup: 'typefield',
+    requiresSecondary: true,
+    matches: (path) =>
+      /(?:^|[/_.-])compat(?:ibility)?(?=$|[/_.-])/.test(path) ||
+      path.includes('schema_compat'),
+  },
+];
+
+function productMatches(path, product) {
+  const aliases = product === 'blackboard' ? ['blackboard', 'whiteboard'] : [product];
+  return aliases.some((alias) => new RegExp(`(?:^|[/_.-])${alias}(?=$|[/_.-])`).test(path));
+}
+
+const PRODUCT_MODULES = PRODUCT_GROUPS.flatMap((group) =>
+  group.products.map((product) => ({
+    id: `product:${product}`,
+    label: `产品：${product}`,
+    primary: group.primary,
+    backup: group.backup,
+    requiresSecondary: false,
+    matches: (path) => productMatches(path, product),
+  })),
+);
+
+const ALL_MODULES = [MODULES[0], MODULES[1], ...PRODUCT_MODULES, MODULES[2], MODULES[3]];
+
+function normalizedPaths(file) {
+  return [file?.filename, file?.previous_filename]
+    .filter((path) => typeof path === 'string' && path !== '')
+    .map((path) => path.toLowerCase());
+}
+
+function compareStats(left, right) {
+  return right.files - left.files || left.module.order - right.module.order || left.module.id.localeCompare(right.module.id);
+}
+
+function classifyFiles(files) {
+  const counts = new Map();
+  for (const file of files || []) {
+    const matchingModules = new Set();
+    for (const path of normalizedPaths(file)) {
+      const matches = ALL_MODULES.filter((module) => module.matches(path));
+      const securityOrDelivery = matches.filter(
+        (module) => module.id === 'security' || module.id === 'delivery',
+      );
+      const effectiveMatches = securityOrDelivery.length > 0
+        ? [...securityOrDelivery, ...matches.filter((module) => module.id === 'compatibility')]
+        : matches;
+      for (const match of effectiveMatches) {
+        matchingModules.add(match.id);
+      }
+      if (
+        effectiveMatches.length === 0 &&
+        (path.startsWith('internal/helpers/') || path.startsWith('internal/shortcut/'))
+      ) {
+        matchingModules.add('architecture');
+      }
+    }
+    for (const moduleID of matchingModules) {
+      counts.set(moduleID, (counts.get(moduleID) || 0) + 1);
+    }
+  }
+
+  return [...counts.entries()]
+    .map(([id, files]) => {
+      const index = ALL_MODULES.findIndex((module) => module.id === id);
+      return {module: {...ALL_MODULES[index], order: index}, files};
+    })
+    .sort(compareStats);
+}
+
+function chooseModuleReviewer(module, unavailable) {
+  return [module.primary, module.backup].find(
+    (reviewer) => REVIEWER_POOL.includes(reviewer) && !unavailable.has(reviewer),
+  );
+}
+
+function addReviewer(reviewers, reviewer) {
+  if (reviewer && !reviewers.includes(reviewer)) {
+    reviewers.push(reviewer);
+  }
+}
+
+function reviewerCandidates({preferredReviewers, fallbackReviewers, eligibleReviewers}) {
+  const eligible = new Set(eligibleReviewers.map((reviewer) => reviewer.toLowerCase()));
+  const candidates = [];
+  for (const reviewer of [...preferredReviewers, ...fallbackReviewers]) {
+    if (
+      eligible.has(reviewer.toLowerCase()) &&
+      !candidates.some((candidate) => candidate.toLowerCase() === reviewer.toLowerCase())
+    ) {
+      candidates.push(reviewer);
+    }
+  }
+  return candidates;
+}
+
+async function requestReviewersWithFallback({
+  candidates,
+  requiredReviewers,
+  satisfiedReviewers = [],
+  requestReviewer,
+  onFailure = () => {},
+}) {
+  const alreadySatisfied = new Set(
+    satisfiedReviewers.map((reviewer) => reviewer.toLowerCase()),
+  );
+  const satisfied = new Set();
+  const requested = [];
+
+  for (const reviewer of candidates) {
+    if (satisfied.size >= requiredReviewers) {
+      break;
+    }
+    const normalizedReviewer = reviewer.toLowerCase();
+    if (alreadySatisfied.has(normalizedReviewer)) {
+      satisfied.add(normalizedReviewer);
+      continue;
+    }
+
+    try {
+      const shouldContinue = await requestReviewer(reviewer);
+      if (shouldContinue === false) {
+        return {requested, satisfiedReviewers: [...satisfied], aborted: true};
+      }
+      requested.push(reviewer);
+      satisfied.add(normalizedReviewer);
+    } catch (error) {
+      onFailure(reviewer, error);
+    }
+  }
+
+  return {requested, satisfiedReviewers: [...satisfied], aborted: false};
+}
+
+function resolveReviewRouting({files, author, latestPusher, fallbackReviewers = REVIEWER_POOL}) {
+  const modules = classifyFiles(files);
+  const unavailable = new Set([author, latestPusher].filter(Boolean).map((login) => login.toLowerCase()));
+  const reviewers = [];
+  const primaryModule = modules[0];
+
+  if (!primaryModule) {
+    return {modules: [], reviewers, requiredReviewers: 1, reason: 'unknown_paths'};
+  }
+
+  addReviewer(reviewers, chooseModuleReviewer(primaryModule.module, unavailable));
+
+  const requiresSecondary =
+    modules.length > 1 || modules.some(({module}) => module.requiresSecondary);
+  const secondaryModule = modules.find(({module}) => module.id !== primaryModule.module.id) || primaryModule;
+  if (requiresSecondary) {
+    addReviewer(
+      reviewers,
+      chooseModuleReviewer(secondaryModule.module, new Set([...unavailable, ...reviewers])),
+    );
+  }
+
+  for (const reviewer of fallbackReviewers) {
+    if (reviewers.length >= (requiresSecondary ? 2 : 1)) {
+      break;
+    }
+    if (REVIEWER_POOL.includes(reviewer) && !unavailable.has(reviewer)) {
+      addReviewer(reviewers, reviewer);
+    }
+  }
+
+  return {
+    modules: modules.map(({module, files}) => ({id: module.id, label: module.label, files})),
+    reviewers,
+    requiredReviewers: requiresSecondary ? 2 : 1,
+    reason: requiresSecondary ? 'cross_or_sensitive' : 'single_module',
+  };
+}
+
+module.exports = {
+  REVIEWER_POOL,
+  classifyFiles,
+  requestReviewersWithFallback,
+  resolveReviewRouting,
+  reviewerCandidates,
+};
@@ -0,0 +1,148 @@
+'use strict';
+
+const assert = require('node:assert/strict');
+const {
+  requestReviewersWithFallback,
+  resolveReviewRouting,
+  reviewerCandidates,
+} = require('./reviewer-routing');
+
+function route(files, author = 'author', latestPusher = author) {
+  return resolveReviewRouting({files: files.map((filename) => ({filename})), author, latestPusher});
+}
+
+{
+  const result = route(['internal/helpers/chat_toolbar.go']);
+  assert.deepEqual(result.reviewers, ['wxianfeng']);
+  assert.equal(result.requiredReviewers, 1);
+  assert.deepEqual(result.modules.map((module) => module.id), ['product:chat']);
+}
+
+{
+  const result = route(['internal/helpers/chat_toolbar.go', 'internal/helpers/doc_style.go']);
+  assert.deepEqual(result.reviewers, ['wxianfeng', 'typefield']);
+  assert.equal(result.requiredReviewers, 2);
+}
+
+{
+  const result = route(['.github/workflows/ci.yml']);
+  assert.deepEqual(result.reviewers, ['haofeng0705', 'wxianfeng']);
+  assert.equal(result.requiredReviewers, 2);
+  assert.equal(result.reason, 'cross_or_sensitive');
+}
+
+{
+  const result = route(['internal/auth/login.go'], 'hlzjsong');
+  assert.deepEqual(result.reviewers, ['typefield', 'wxianfeng']);
+  assert.equal(result.requiredReviewers, 2);
+}
+
+{
+  const result = route(['internal/upgrade/downloader.go']);
+  assert.deepEqual(result.reviewers, ['haofeng0705', 'wxianfeng']);
+  assert.equal(result.requiredReviewers, 2);
+}
+
+{
+  const result = route(['internal/app/upgrade.go', 'scripts/dev/test-release.sh']);
+  assert.deepEqual(result.reviewers, ['haofeng0705', 'wxianfeng']);
+  assert.equal(result.requiredReviewers, 2);
+}
+
+{
+  const result = route(['pkg/edition/edition.go']);
+  assert.deepEqual(result.reviewers, ['hlzjsong', 'typefield']);
+  assert.equal(result.requiredReviewers, 2);
+}
+
+{
+  const result = route(['internal/shortcut/chat/compatibility_coverage_test.go']);
+  assert.deepEqual(result.reviewers, ['wxianfeng', 'typefield']);
+  assert.equal(result.requiredReviewers, 2);
+  assert.deepEqual(result.modules.map((module) => module.id), ['product:chat', 'compatibility']);
+}
+
+{
+  const result = route(['internal/helpers/leaf_dispatch.go']);
+  assert.deepEqual(result.reviewers, ['wxianfeng', 'typefield']);
+  assert.equal(result.requiredReviewers, 2);
+}
+
+{
+  const result = route(['docs/unknown-area.md']);
+  assert.deepEqual(result.reviewers, []);
+  assert.equal(result.reason, 'unknown_paths');
+}
+
+async function testSingleReviewerFallback() {
+  const candidates = reviewerCandidates({
+    preferredReviewers: ['wxianfeng'],
+    fallbackReviewers: ['wxianfeng', 'typefield', 'haofeng0705'],
+    eligibleReviewers: ['wxianfeng', 'typefield', 'haofeng0705'],
+  });
+  const attempts = [];
+  const result = await requestReviewersWithFallback({
+    candidates,
+    requiredReviewers: 1,
+    requestReviewer: async (reviewer) => {
+      attempts.push(reviewer);
+      if (reviewer === 'wxianfeng') {
+        throw Object.assign(new Error('cannot request primary'), {status: 422});
+      }
+      return true;
+    },
+  });
+  assert.deepEqual(attempts, ['wxianfeng', 'typefield']);
+  assert.deepEqual(result.requested, ['typefield']);
+  assert.equal(result.satisfiedReviewers.length, 1);
+}
+
+async function testTwoReviewerFallback() {
+  const candidates = reviewerCandidates({
+    preferredReviewers: ['haofeng0705', 'wxianfeng'],
+    fallbackReviewers: ['haofeng0705', 'wxianfeng', 'typefield', 'hlzjsong'],
+    eligibleReviewers: ['haofeng0705', 'wxianfeng', 'typefield', 'hlzjsong'],
+  });
+  const attempts = [];
+  const result = await requestReviewersWithFallback({
+    candidates,
+    requiredReviewers: 2,
+    requestReviewer: async (reviewer) => {
+      attempts.push(reviewer);
+      if (reviewer === 'wxianfeng') {
+        throw Object.assign(new Error('temporary failure'), {status: 503});
+      }
+      return true;
+    },
+  });
+  assert.deepEqual(attempts, ['haofeng0705', 'wxianfeng', 'typefield']);
+  assert.deepEqual(result.requested, ['haofeng0705', 'typefield']);
+  assert.equal(result.satisfiedReviewers.length, 2);
+}
+
+async function testLowerPriorityExistingRequestDoesNotReplaceOwner() {
+  const attempts = [];
+  const result = await requestReviewersWithFallback({
+    candidates: ['wxianfeng', 'typefield'],
+    requiredReviewers: 1,
+    satisfiedReviewers: ['typefield'],
+    requestReviewer: async (reviewer) => {
+      attempts.push(reviewer);
+      return true;
+    },
+  });
+  assert.deepEqual(attempts, ['wxianfeng']);
+  assert.deepEqual(result.requested, ['wxianfeng']);
+  assert.deepEqual(result.satisfiedReviewers, ['wxianfeng']);
+}
+
+Promise.all([
+  testSingleReviewerFallback(),
+  testTwoReviewerFallback(),
+  testLowerPriorityExistingRequestDoesNotReplaceOwner(),
+])
+  .then(() => console.log('reviewer routing policy tests passed'))
+  .catch((error) => {
+    console.error(error);
+    process.exitCode = 1;
+  });
@@ -24,6 +24,7 @@ jobs:
      pull-requests: read
    outputs:
      changelog_only: ${{ steps.classify.outputs.changelog_only }}
+      release_seal_only: ${{ steps.classify.outputs.release_seal_only }}
      changelog_changed: ${{ steps.classify.outputs.changelog_changed }}
      docs_only: ${{ steps.classify.outputs.docs_only }}
      full_suite: ${{ steps.classify.outputs.full_suite }}
@@ -39,6 +40,7 @@ jobs:
        with:
          script: |
            let changelogOnly = false;
+            let releaseSealOnly = false;
            let changelogChanged = false;
            let docsOnly = false;
            let fullSuite = context.eventName === 'push';
@@ -111,6 +113,7 @@ jobs:
                filename.startsWith('internal/app/') ||
                filename.startsWith('internal/cli/') ||
                filename.startsWith('internal/cobracmd/') ||
+                filename.startsWith('internal/corecmd/') ||
                filename.startsWith('internal/helpers/') ||
                filename.startsWith('internal/i18n/') ||
                filename.startsWith('internal/interfacesnapshot/') ||
@@ -147,12 +150,22 @@ jobs:
                filename === '.github/actionlint.yaml' ||
                filename.startsWith('scripts/') ||
                filename.startsWith('verify/') ||
+                filename.startsWith('internal/helpers/') ||
+                // Shortcut declarations feed the live command tree and Schema
+                // assembly. Their reverse dependencies include the expensive
+                // app and generator packages, which must run in separate shards.
+                filename.startsWith('internal/shortcut/') ||
                filename.startsWith('internal/generator/') ||
                filename.startsWith('internal/cli/schema') ||
+                // Parameter aliases are reduced against the live command tree.
+                // Their reverse-dependency set is too large for one focused
+                // race job, so use the existing full-suite shards.
+                filename === 'internal/cli/param_concepts.json' ||
+                filename === 'internal/cli/param_concepts.schema.json' ||
+                filename === 'internal/cli/param_aliases_generated.go' ||
                filename.startsWith('internal/interfacesnapshot/') ||
                filename.startsWith('internal/app/upgrade') ||
                filename.startsWith('internal/transport/') ||
-                filename.startsWith('internal/recovery/') ||
                filename.startsWith('internal/syncdata/') ||
                filename.includes('/testdata/') ||
                filename.startsWith('testdata/') ||
@@ -161,6 +174,43 @@ jobs:
                filename === 'go.mod' ||
                filename === 'go.sum'
              );
+            const isExactReleaseSeal = (candidates) => {
+              const changelog = candidates.filter(
+                ({ filename, status, previous_filename }) =>
+                  filename === 'CHANGELOG.md' &&
+                  status === 'modified' &&
+                  !previous_filename
+              );
+              if (changelog.length !== 1 || candidates.length < 2) {
+                return false;
+              }
+              let version = '';
+              return candidates.every((file) => {
+                if (file.filename === 'CHANGELOG.md') {
+                  return file.status === 'modified' && !file.previous_filename;
+                }
+                if (
+                  file.status !== 'renamed' ||
+                  typeof file.filename !== 'string' ||
+                  typeof file.previous_filename !== 'string' ||
+                  file.additions !== 0 ||
+                  file.deletions !== 0
+                ) {
+                  return false;
+                }
+                const target = file.filename.match(
+                  /^\.changes\/released\/([0-9]+\.[0-9]+\.[0-9]+(?:-beta\.[1-9][0-9]*)?)\/([a-z0-9][a-z0-9._-]*\.md)$/
+                );
+                if (!target || file.previous_filename !== `.changes/${target[2]}`) {
+                  return false;
+                }
+                if (version && version !== target[1]) {
+                  return false;
+                }
+                version = target[1];
+                return true;
+              });
+            };
            const classifyFiles = (complete) => {
              const paths = files.flatMap(({ filename, previous_filename }) =>
                [filename, previous_filename].filter(
@@ -247,19 +297,26 @@ jobs:
                );
              }

-              changelogOnly =
+              const exactChangelogDiff =
                files.length === 1 &&
                files[0].filename === 'CHANGELOG.md' &&
                files[0].status === 'modified' &&
                !files[0].previous_filename;
+              releaseSealOnly = isExactReleaseSeal(files);
+              changelogOnly = exactChangelogDiff || releaseSealOnly;
              changelogChanged = files.some(
                ({ filename, previous_filename }) =>
                  filename === 'CHANGELOG.md' ||
                  previous_filename === 'CHANGELOG.md'
              );
              classifyFiles(true);
+              if (releaseSealOnly) {
+                fullSuite = false;
+              }
              fastPathTrust = changelogOnly
-                ? 'exact pull-request revision and synthetic merge policy'
+                ? releaseSealOnly
+                  ? 'exact release-seal fragment archival and synthetic merge policy'
+                  : 'exact CHANGELOG-only revision and synthetic merge policy'
                : docsOnly
                  ? 'documentation-only focused admission'
                  : fullSuite
@@ -294,7 +351,8 @@ jobs:
                    per_page: 100,
                  });
                files = Array.isArray(comparison.files) ? comparison.files : [];
-                classifyFiles(files.length < 300);
+                const pushFilesComplete = files.length < 300;
+                classifyFiles(pushFilesComplete);
                const linearFromValidatedTip =
                  comparison.status === 'ahead' &&
                  comparison.merge_base_commit?.sha === expectedBefore &&
@@ -306,8 +364,10 @@ jobs:
                  files[0].filename === 'CHANGELOG.md' &&
                  files[0].status === 'modified' &&
                  !files[0].previous_filename;
+                const exactReleaseSealDiff =
+                  pushFilesComplete && isExactReleaseSeal(files);

-                if (linearFromValidatedTip && exactChangelogDiff) {
+                if (linearFromValidatedTip && (exactChangelogDiff || exactReleaseSealDiff)) {
                  const requiredContexts = [
                    'Lint',
                    'Test',
@@ -358,9 +418,15 @@ jobs:

                  if (missing.length === 0 && nonSuccess.length === 0) {
                    changelogOnly = true;
+                    releaseSealOnly = exactReleaseSealDiff;
                    changelogChanged = true;
+                    if (releaseSealOnly) {
+                      fullSuite = false;
+                    }
                    fastPathTrust =
-                      `exact CHANGELOG-only successor of validated ${expectedBefore}`;
+                      releaseSealOnly
+                        ? `exact release-seal successor of validated ${expectedBefore}`
+                        : `exact CHANGELOG-only successor of validated ${expectedBefore}`;
                  } else {
                    fastPathTrust =
                      'predecessor Code Admission is not fully successful; ' +
@@ -375,6 +441,7 @@ jobs:
            }

            core.setOutput('changelog_only', String(changelogOnly));
+            core.setOutput('release_seal_only', String(releaseSealOnly));
            core.setOutput('changelog_changed', String(changelogChanged));
            core.setOutput('docs_only', String(docsOnly));
            core.setOutput('full_suite', String(fullSuite));
@@ -386,7 +453,8 @@ jobs:
            await core.summary
              .addHeading('Code Admission scope')
              .addRaw(`- Event: \`${context.eventName}\`\n`)
-              .addRaw(`- Exact modified CHANGELOG only: \`${changelogOnly}\`\n`)
+              .addRaw(`- Metadata-only fast path: \`${changelogOnly}\`\n`)
+              .addRaw(`- Release-seal fragments only: \`${releaseSealOnly}\`\n`)
              .addRaw(`- CHANGELOG touched: \`${changelogChanged}\`\n`)
              .addRaw(`- Documentation-only: \`${docsOnly}\`\n`)
              .addRaw(`- Full suite: \`${fullSuite}\`\n`)
@@ -401,7 +469,14 @@ jobs:

      - name: Record CHANGELOG-only fast path
        if: steps.classify.outputs.changelog_only == 'true'
-        run: echo "Lint is satisfied by the trusted CHANGELOG-only Policy path." >> "$GITHUB_STEP_SUMMARY"
+        env:
+          RELEASE_SEAL_ONLY: ${{ steps.classify.outputs.release_seal_only }}
+        run: |
+          if [ "$RELEASE_SEAL_ONLY" = true ]; then
+            echo "Lint is satisfied by the trusted release-seal fragment Policy path." >> "$GITHUB_STEP_SUMMARY"
+          else
+            echo "Lint is satisfied by the trusted CHANGELOG-only Policy path." >> "$GITHUB_STEP_SUMMARY"
+          fi

      - name: Record documentation-only fast path
        if: steps.classify.outputs.changelog_only != 'true' && steps.classify.outputs.docs_only == 'true'
@@ -433,12 +508,33 @@ jobs:
        if: steps.classify.outputs.changelog_only != 'true' && steps.classify.outputs.docs_only != 'true'
        run: go run github.com/rhysd/actionlint/cmd/actionlint@v1.7.12

+      - name: Test reviewer routing policy
+        if: steps.classify.outputs.changelog_only != 'true' && steps.classify.outputs.docs_only != 'true'
+        run: node .github/reviewer-routing.test.js
+
  test-focused:
-    name: Test (changed packages)
+    name: "Test (focused: ${{ matrix.shard }})"
    needs: lint
    if: ${{ needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true' && needs.lint.outputs.full_suite != 'true' }}
    runs-on: ubuntu-latest
+    # Each shard owns one bounded slice of the impacted set, so no single job
+    # carries internal/app together with every reverse dependency. The shard
+    # list and per-shard execution below mirror test-race, which runs the same
+    # shards at full-suite scope; release-scripts is included because its
+    # dedicated job only runs at full-suite or release-sensitive scope, and
+    # dropping it here would stop testing test/scripts changes entirely.
    timeout-minutes: 20
+    strategy:
+      fail-fast: false
+      matrix:
+        shard:
+          - app
+          - generators
+          - helpers
+          - cli
+          - smoke
+          - remaining
+          - release-scripts
    steps:
      - name: Check out repository
        uses: actions/checkout@v4
@@ -467,30 +563,85 @@ jobs:
        with:
          go-version-file: go.mod

-      - name: Test changed packages and reverse dependencies
+      - name: Select impacted packages for shard
+        id: select
        shell: bash
        env:
-          DWS_PACKAGE_VERSION: 0.0.0-test
+          TEST_SHARD: ${{ matrix.shard }}
        run: |
          set -euo pipefail
          package_output="$(
            ./scripts/ci/changed-test-packages.sh \
-              list "$TEST_BASE_REF" "$TEST_HEAD_REF"
+              list-shard "$TEST_SHARD" "$TEST_BASE_REF" "$TEST_HEAD_REF"
          )"
          if [ -z "$package_output" ]; then
-            echo "No buildable Go package is affected by this revision." \
+            echo "No buildable Go package in shard $TEST_SHARD is affected by this revision." \
              >> "$GITHUB_STEP_SUMMARY"
+            echo "affected=false" >> "$GITHUB_OUTPUT"
            exit 0
          fi
-          mapfile -t packages <<< "$package_output"
-          go test -v -race -count=1 -timeout=15m "${packages[@]}"
+          # The package list travels through a file rather than a step output:
+          # reading it with `mapfile < file` has unambiguous line semantics,
+          # whereas a here-string over a multi-line output would append an extra
+          # empty element if the value ever carried a trailing newline, and an
+          # empty element would reach go test as an empty package argument.
+          printf '%s\n' "$package_output" > "$RUNNER_TEMP/focused-shard-packages.txt"
+          echo "affected=true" >> "$GITHUB_OUTPUT"
+
+      - name: Build
+        if: ${{ matrix.shard == 'remaining' && steps.select.outputs.affected == 'true' }}
+        run: make build
+
+      - name: Install archive tooling
+        if: ${{ matrix.shard == 'release-scripts' && steps.select.outputs.affected == 'true' }}
+        run: sudo apt-get update && sudo apt-get install -y zip unzip
+
+      - name: Test shard with Race Detection
+        if: ${{ steps.select.outputs.affected == 'true' }}
+        shell: bash
+        env:
+          DWS_PACKAGE_VERSION: 0.0.0-test
+          TEST_SHARD: ${{ matrix.shard }}
+        run: |
+          set -euo pipefail
+          mapfile -t packages < "$RUNNER_TEMP/focused-shard-packages.txt"
+          test "${#packages[@]}" -gt 0
+          for package in "${packages[@]}"; do
+            test -n "$package" || {
+              echo "shard package list contains an empty entry" >&2
+              exit 1
+            }
+          done
+          if [ "$TEST_SHARD" = "app" ]; then
+            # A single long-lived app test process retains every constructed
+            # command tree in framework registries. Isolate Schema assembly and
+            # bounded name ranges so each process releases that state on exit.
+            test "${#packages[@]}" -eq 1
+            ./scripts/ci/run-app-race-tests.sh run "${packages[0]}"
+            exit 0
+          fi
+          if [ "$TEST_SHARD" = "release-scripts" ]; then
+            # Mirror the dedicated release-contract job: these suites shell out
+            # to archive tooling and are not race-instrumented there.
+            go test -v -count=1 -timeout=10m "${packages[@]}"
+            exit 0
+          fi
+          # cli/smoke own heavy NewRootCommand / Schema assembly under -race;
+          # give them a dedicated package timeout on slower hosted runners.
+          timeout_budget=12m
+          if [ "$TEST_SHARD" = "cli" ] ||
+             [ "$TEST_SHARD" = "smoke" ]; then
+            timeout_budget=15m
+          fi
+          go test -v -race -count=1 -timeout="$timeout_budget" "${packages[@]}"

  test-race:
    name: "Test (race: ${{ matrix.shard }})"
    needs: lint
    if: ${{ needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true' && needs.lint.outputs.full_suite == 'true' }}
    runs-on: ubuntu-latest
-    # cli/smoke shards need headroom beyond go test -timeout for setup + assembly.
+    # app runs several independently bounded processes; cli/smoke need headroom
+    # beyond go test -timeout for setup + assembly.
    timeout-minutes: 20
    strategy:
      fail-fast: false
@@ -526,10 +677,19 @@ jobs:
          test -n "$package_output"
          mapfile -t packages <<< "$package_output"
          test "${#packages[@]}" -gt 0
-          # cli/smoke own heavy NewRootCommand / Schema assembly under -race; give
-          # them a dedicated budget so remaining is not SIGTERM'd by OOM/timeout.
+          if [ "$TEST_SHARD" = "app" ]; then
+            # A single long-lived app test process retains every constructed
+            # command tree in framework registries. Isolate Schema assembly and
+            # bounded name ranges so each process releases that state on exit.
+            test "${#packages[@]}" -eq 1
+            ./scripts/ci/run-app-race-tests.sh run "${packages[0]}"
+            exit 0
+          fi
+          # cli/smoke own heavy NewRootCommand / Schema assembly under -race;
+          # give them a dedicated package timeout on slower hosted runners.
          timeout_budget=12m
-          if [ "$TEST_SHARD" = "cli" ] || [ "$TEST_SHARD" = "smoke" ]; then
+          if [ "$TEST_SHARD" = "cli" ] ||
+             [ "$TEST_SHARD" = "smoke" ]; then
            timeout_budget=15m
          fi
          go test -v -race -count=1 -timeout="$timeout_budget" "${packages[@]}"
@@ -626,7 +786,7 @@ jobs:
          failed=0
          if [ "$CHANGELOG_ONLY" = true ] || [ "$DOCS_ONLY" = true ]; then
            for shard in \
-              "changed packages:$FOCUSED_RESULT" \
+              "focused shards:$FOCUSED_RESULT" \
              "race shards:$RACE_RESULT" \
              "release scripts:$RELEASE_SCRIPTS_RESULT" \
              "cross-platform compile:$CROSS_PLATFORM_RESULT" \
@@ -655,7 +815,7 @@ jobs:
            release_expected=success
          fi
          for shard in \
-            "changed packages:$FOCUSED_RESULT:$focused_expected" \
+            "focused shards:$FOCUSED_RESULT:$focused_expected" \
            "race shards:$RACE_RESULT:$race_expected" \
            "release scripts:$RELEASE_SCRIPTS_RESULT:$release_expected" \
            "cross-platform compile:$CROSS_PLATFORM_RESULT:success"
@@ -702,14 +862,11 @@ jobs:
        with:
          go-version-file: go.mod

-      # Full ./internal/app (including schema --all assembly) already runs in the
-      # race:app shard. Keep this job focused on native auth/keychain paths so
-      # heavy Schema completeness tests cannot exhaust the 10m budget.
-      - name: Test macOS auth and Keychain paths with Race Detection
-        run: go test -v -race -count=1 -timeout=10m ./internal/keychain ./internal/auth
+      - name: Test macOS auth and Keychain packages with Race Detection
+        run: go test -v -race -count=1 -timeout=6m ./internal/keychain ./internal/auth

-      - name: Test macOS auth migration and portable auth diagnostics
-        run: go test -v -race -count=1 -timeout=5m ./internal/app -run '^Test(CrossPlatformCoverage)?Auth(MigrateKeychain|StatusDiagnosticReportsCiphertextKeyMismatch|ExportRejectsWindowsDPAPIBackend|ImportRejectsWindowsDPAPIBackend)'
+      - name: Test macOS auth migration, Keychain diagnostics, and upgrade self-heal with Race Detection
+        run: go test -v -race -count=1 -timeout=5m ./internal/app -run '^(TestValidateNewBinary_RecoversFromUnsignedDarwin|Test(CrossPlatformCoverage)?Auth(MigrateKeychain|StatusDiagnosticReportsCiphertextKeyMismatch))'

  test-windows:
    name: Test (Windows)
@@ -820,6 +977,7 @@ jobs:
        run: ./scripts/policy/run-platform-coverage-gate.sh --base-ref "$COVERAGE_BASE_REF" --profile coverage-windows.txt

      - name: Upload Windows coverage artifact
+        if: always()
        uses: actions/upload-artifact@v4
        with:
          name: coverage-windows
@@ -828,7 +986,7 @@ jobs:
  coverage-current:
    name: Coverage (current)
    needs: lint
-    if: ${{ needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true' }}
+    if: ${{ needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true' && needs.lint.outputs.full_suite != 'true' }}
    runs-on: ubuntu-latest
    timeout-minutes: 20
    steps:
@@ -843,10 +1001,6 @@ jobs:
        with:
          go-version-file: go.mod

-      - name: Install archive tooling
-        if: needs.lint.outputs.full_suite == 'true'
-        run: sudo apt-get update && sudo apt-get install -y zip unzip
-
      - name: Resolve authoritative coverage base
        env:
          PR_HEAD_SHA: ${{ github.event.pull_request.head.sha }}
@@ -868,41 +1022,33 @@ jobs:
      - name: Build
        run: make build

-      - name: Run current unit tests with coverage
+      - name: Run scoped unit tests with coverage
        shell: bash
        env:
          DWS_PACKAGE_VERSION: 0.0.0-test
-          FULL_SUITE: ${{ needs.lint.outputs.full_suite }}
        run: |
          set -euo pipefail
-          if [ "$FULL_SUITE" = true ]; then
+          changed_output="$(
+            ./scripts/ci/changed-test-packages.sh \
+              changed "$COVERAGE_BASE_REF" "$COVERAGE_HEAD_REF"
+          )"
+          impacted_output="$(
+            ./scripts/ci/changed-test-packages.sh \
+              list "$COVERAGE_BASE_REF" "$COVERAGE_HEAD_REF"
+          )"
+          if [ -z "$changed_output" ] || [ -z "$impacted_output" ]; then
+            printf 'mode: atomic\n' > coverage.txt
+            echo "No buildable Go package needs scoped coverage." \
+              >> "$GITHUB_STEP_SUMMARY"
+          else
+            mapfile -t changed_packages <<< "$changed_output"
+            mapfile -t impacted_packages <<< "$impacted_output"
+            coverpkg="$(IFS=,; echo "${changed_packages[*]}")"
            go test -count=1 -p 1 \
+              -coverpkg="$coverpkg" \
              -coverprofile=coverage.txt \
              -covermode=atomic \
-              ./ ./cmd/... ./internal/... ./skills/...
-          else
-            changed_output="$(
-              ./scripts/ci/changed-test-packages.sh \
-                changed "$COVERAGE_BASE_REF" "$COVERAGE_HEAD_REF"
-            )"
-            impacted_output="$(
-              ./scripts/ci/changed-test-packages.sh \
-                list "$COVERAGE_BASE_REF" "$COVERAGE_HEAD_REF"
-            )"
-            if [ -z "$changed_output" ] || [ -z "$impacted_output" ]; then
-              printf 'mode: atomic\n' > coverage.txt
-              echo "No buildable Go package needs scoped coverage." \
-                >> "$GITHUB_STEP_SUMMARY"
-            else
-              mapfile -t changed_packages <<< "$changed_output"
-              mapfile -t impacted_packages <<< "$impacted_output"
-              coverpkg="$(IFS=,; echo "${changed_packages[*]}")"
-              go test -count=1 -p 1 \
-                -coverpkg="$coverpkg" \
-                -coverprofile=coverage.txt \
-                -covermode=atomic \
-                "${impacted_packages[@]}"
-            fi
+              "${impacted_packages[@]}"
          fi
          if [ "$(wc -l < coverage.txt)" -gt 1 ]; then
            go tool cover -func=coverage.txt
@@ -915,6 +1061,66 @@ jobs:
          path: coverage.txt
          retention-days: 1

+  coverage-current-full:
+    name: "Coverage (current: ${{ matrix.shard }})"
+    needs: lint
+    if: ${{ needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true' && needs.lint.outputs.full_suite == 'true' }}
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    strategy:
+      fail-fast: false
+      matrix:
+        shard:
+          - app
+          - cli
+          - generators
+          - helpers
+          - remaining
+    steps:
+      - name: Check out repository
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+
+      - name: Set up Go
+        uses: actions/setup-go@v5
+        with:
+          go-version-file: go.mod
+
+      - name: Install archive tooling
+        run: sudo apt-get update && sudo apt-get install -y zip unzip
+
+      - name: Build
+        run: make build
+
+      # Each shard keeps -p 1 so the authoritative measurement stays serial
+      # inside one instrumented process group; shards run on isolated runners,
+      # and scripts/ci/test-packages.sh verify proves the shard union equals
+      # the previous single full-suite package set exactly once.
+      - name: Run current shard tests with coverage
+        shell: bash
+        env:
+          DWS_PACKAGE_VERSION: 0.0.0-test
+          COVERAGE_SHARD: ${{ matrix.shard }}
+        run: |
+          set -euo pipefail
+          package_output="$(./scripts/ci/test-packages.sh list-coverage "$COVERAGE_SHARD")"
+          test -n "$package_output"
+          mapfile -t packages <<< "$package_output"
+          test "${#packages[@]}" -gt 0
+          go test -count=1 -p 1 \
+            -coverprofile="coverage-shard-$COVERAGE_SHARD.txt" \
+            -covermode=atomic \
+            "${packages[@]}"
+          go tool cover -func="coverage-shard-$COVERAGE_SHARD.txt" | tail -n 1
+
+      - name: Upload current shard coverage profile
+        uses: actions/upload-artifact@v4
+        with:
+          name: coverage-current-shard-${{ matrix.shard }}
+          path: coverage-shard-${{ matrix.shard }}.txt
+          retention-days: 1
+
  coverage-supporting:
    name: Coverage (supporting)
    needs: lint
@@ -968,14 +1174,11 @@ jobs:
          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}

      - name: Set up Go
+        id: setup-go
        uses: actions/setup-go@v5
        with:
          go-version-file: go.mod

-      - name: Install archive tooling
-        if: needs.lint.outputs.full_suite == 'true'
-        run: sudo apt-get update && sudo apt-get install -y zip unzip
-
      - name: Resolve authoritative coverage base
        env:
          PR_HEAD_SHA: ${{ github.event.pull_request.head.sha }}
@@ -993,7 +1196,34 @@ jobs:
          git rev-parse --verify "${base_ref}^{commit}" >/dev/null
          echo "COVERAGE_BASE_REF=$base_ref" >> "$GITHUB_ENV"

+      # The merge-base full-suite profile is a pure function of the base
+      # commit. Reuse the profile published by the last green push run of
+      # exactly that commit instead of re-running the whole suite; any key
+      # mismatch falls back to authoritative recomputation. Exact key only,
+      # never prefix fallback: a near-miss profile would compare the
+      # candidate against the wrong commit.
+      - name: Restore cached merge-base coverage profile
+        id: baseline-cache
+        if: needs.lint.outputs.full_suite == 'true'
+        uses: actions/cache/restore@v4
+        with:
+          path: coverage-cache.txt
+          key: dws-coverage-full-v2-${{ env.COVERAGE_BASE_REF }}-go${{ steps.setup-go.outputs.go-version }}
+
+      - name: Materialize cached merge-base coverage profile
+        if: needs.lint.outputs.full_suite == 'true' && steps.baseline-cache.outputs.cache-hit == 'true'
+        run: |
+          set -eu
+          test -s coverage-cache.txt
+          test "$(head -n 1 coverage-cache.txt)" = "mode: atomic"
+          cp coverage-cache.txt coverage-base.txt
+
+      - name: Install archive tooling
+        if: needs.lint.outputs.full_suite == 'true' && steps.baseline-cache.outputs.cache-hit != 'true'
+        run: sudo apt-get update && sudo apt-get install -y zip unzip
+
      - name: Run baseline unit tests with coverage
+        if: steps.baseline-cache.outputs.cache-hit != 'true'
        shell: bash
        env:
          DWS_PACKAGE_VERSION: 0.0.0-test
@@ -1042,6 +1272,21 @@ jobs:
            fi
          )

+      - name: Prepare merge-base coverage profile cache
+        if: needs.lint.outputs.full_suite == 'true' && steps.baseline-cache.outputs.cache-hit != 'true'
+        run: |
+          set -eu
+          test -s coverage-base.txt
+          test "$(head -n 1 coverage-base.txt)" = "mode: atomic"
+          cp coverage-base.txt coverage-cache.txt
+
+      - name: Save merge-base coverage profile cache
+        if: needs.lint.outputs.full_suite == 'true' && steps.baseline-cache.outputs.cache-hit != 'true'
+        uses: actions/cache/save@v4
+        with:
+          path: coverage-cache.txt
+          key: dws-coverage-full-v2-${{ env.COVERAGE_BASE_REF }}-go${{ steps.setup-go.outputs.go-version }}
+
      - name: Upload baseline coverage profile
        uses: actions/upload-artifact@v4
        with:
@@ -1054,6 +1299,7 @@ jobs:
    needs:
      - lint
      - coverage-current
+      - coverage-current-full
      - coverage-supporting
      - coverage-baseline
      - coverage-darwin
@@ -1069,6 +1315,7 @@ jobs:
          FULL_SUITE: ${{ needs.lint.outputs.full_suite }}
          PLATFORM_SENSITIVE: ${{ needs.lint.outputs.platform_sensitive }}
          CURRENT_RESULT: ${{ needs.coverage-current.result }}
+          CURRENT_FULL_RESULT: ${{ needs.coverage-current-full.result }}
          SUPPORTING_RESULT: ${{ needs.coverage-supporting.result }}
          BASELINE_RESULT: ${{ needs.coverage-baseline.result }}
          DARWIN_RESULT: ${{ needs.coverage-darwin.result }}
@@ -1076,6 +1323,7 @@ jobs:
        run: |
          failed=0
          current_expected=success
+          current_full_expected=skipped
          supporting_expected=skipped
          baseline_expected=success
          native_expected=skipped
@@ -1083,6 +1331,8 @@ jobs:
            current_expected=skipped
            baseline_expected=skipped
          elif [ "$FULL_SUITE" = true ]; then
+            current_expected=skipped
+            current_full_expected=success
            supporting_expected=success
          fi
          if [ "$CHANGELOG_ONLY" != true ] &&
@@ -1093,6 +1343,7 @@ jobs:

          for profile in \
            "current:$CURRENT_RESULT:$current_expected" \
+            "current shards:$CURRENT_FULL_RESULT:$current_full_expected" \
            "supporting:$SUPPORTING_RESULT:$supporting_expected" \
            "baseline:$BASELINE_RESULT:$baseline_expected"
          do
@@ -1128,6 +1379,7 @@ jobs:
          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}

      - name: Set up Go
+        id: setup-go
        if: needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true'
        uses: actions/setup-go@v5
        with:
@@ -1151,11 +1403,12 @@ jobs:
          git rev-parse --verify "${base_ref}^{commit}" >/dev/null
          echo "COVERAGE_BASE_REF=$base_ref" >> "$GITHUB_ENV"

-      - name: Download current coverage profile
+      - name: Download current coverage profiles
        if: needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true'
        uses: actions/download-artifact@v4
        with:
-          name: coverage-current-profile
+          pattern: coverage-current-*
+          merge-multiple: true
          path: .

      - name: Download supporting coverage profiles
@@ -1172,13 +1425,33 @@ jobs:
          name: coverage-baseline-profile
          path: .

+      # Shard profiles cover disjoint package sets, so their block-level
+      # concatenation is the same candidate profile one serial run produced.
+      # Every expected shard must be present; a missing shard would silently
+      # shrink the scope-matched overall comparison.
+      - name: Assemble full-suite coverage profile
+        if: needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true' && needs.lint.outputs.full_suite == 'true'
+        shell: bash
+        run: |
+          set -euo pipefail
+          test ! -f coverage.txt
+          for shard in app cli generators helpers remaining; do
+            profile="coverage-shard-$shard.txt"
+            test -f "$profile"
+            test "$(head -n 1 "$profile")" = "mode: atomic"
+          done
+          printf 'mode: atomic\n' > coverage.txt
+          for shard in app cli generators helpers remaining; do
+            tail -n +2 "coverage-shard-$shard.txt" >> coverage.txt
+          done
+
      - name: Enforce coverage gate
        if: needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true'
        env:
          FULL_SUITE: ${{ needs.lint.outputs.full_suite }}
          COVERAGE_TARGET: "100"
          COVERAGE_ENFORCE_OVERALL: "false"
-          COVERAGE_OVERALL_TOLERANCE: "0"
+          COVERAGE_OVERALL_TOLERANCE: "0.1"
        run: |
          policy_profile=coverage-policy.txt
          if [ "$FULL_SUITE" != true ]; then
@@ -1192,6 +1465,26 @@ jobs:
          COVERAGE_ADDITIONAL_DIFF_PROFILE="$additional_profile" \
            make coverage-gate BASE_REF="$COVERAGE_BASE_REF"

+      # Publish this push's full-suite profile as the merge-base cache for
+      # future PRs whose merge-base is exactly this commit. Saved only after
+      # the gate passed so a broken run never becomes a baseline. Both producer
+      # and consumer use coverage-cache.txt because the cache version includes
+      # the configured path as well as the compression tool.
+      - name: Prepare push coverage profile as merge-base cache
+        if: github.event_name == 'push' && needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true' && needs.lint.outputs.full_suite == 'true'
+        run: |
+          set -eu
+          test -s coverage.txt
+          test "$(head -n 1 coverage.txt)" = "mode: atomic"
+          cp coverage.txt coverage-cache.txt
+
+      - name: Save push coverage profile as merge-base cache
+        if: github.event_name == 'push' && needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true' && needs.lint.outputs.full_suite == 'true'
+        uses: actions/cache/save@v4
+        with:
+          path: coverage-cache.txt
+          key: dws-coverage-full-v2-${{ github.sha }}-go${{ steps.setup-go.outputs.go-version }}
+
      - name: Generate coverage report
        if: needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true'
        run: |
@@ -1253,6 +1546,7 @@ jobs:
        env:
          CLASSIFIED_CHANGELOG_CHANGED: ${{ needs.lint.outputs.changelog_changed }}
          CHANGELOG_ONLY: ${{ needs.lint.outputs.changelog_only }}
+          RELEASE_SEAL_ONLY: ${{ needs.lint.outputs.release_seal_only }}
          PR_BASE_SHA: ${{ github.event.pull_request.base.sha }}
        run: |
          set -eu
@@ -1279,31 +1573,53 @@ jobs:
          fi

          mode=--content-only
-          if [ "$CHANGELOG_ONLY" = true ]; then
+          if [ "$CHANGELOG_ONLY" = true ] && [ "$RELEASE_SEAL_ONLY" != true ]; then
            mode=--fast-path
          fi
          ./scripts/policy/check-changelog-pr.sh \
            "$mode" "$PR_BASE_SHA" HEAD

-      - name: Validate trusted main CHANGELOG-only push
+      - name: Validate release fragment lifecycle
+        if: github.event_name == 'pull_request'
+        env:
+          PR_BASE_SHA: ${{ github.event.pull_request.base.sha }}
+        run: ./scripts/policy/check-release-fragments.sh "$PR_BASE_SHA" HEAD
+
+      - name: Validate trusted main metadata-only push
        if: github.event_name == 'push' && needs.lint.outputs.changelog_only == 'true'
        env:
          PUSH_BEFORE_SHA: ${{ github.event.before }}
          PUSH_AFTER_SHA: ${{ github.event.after }}
+          RELEASE_SEAL_ONLY: ${{ needs.lint.outputs.release_seal_only }}
        run: |
          set -eu
          test "$(git rev-parse HEAD)" = "$PUSH_AFTER_SHA" || {
            echo "checked-out push revision does not match event after SHA" >&2
            exit 1
          }
+          mode=--fast-path
+          if [ "$RELEASE_SEAL_ONLY" = true ]; then
+            mode=--content-only
+          fi
          ./scripts/policy/check-changelog-pr.sh \
-            --fast-path "$PUSH_BEFORE_SHA" "$PUSH_AFTER_SHA"
+            "$mode" "$PUSH_BEFORE_SHA" "$PUSH_AFTER_SHA"
+          if [ "$RELEASE_SEAL_ONLY" = true ]; then
+            ./scripts/policy/check-release-fragments.sh \
+              "$PUSH_BEFORE_SHA" "$PUSH_AFTER_SHA"
+          fi

      - name: Record CHANGELOG-only fast path
        if: needs.lint.outputs.changelog_only == 'true'
+        env:
+          RELEASE_SEAL_ONLY: ${{ needs.lint.outputs.release_seal_only }}
        run: |
-          echo "Only the trusted base-equivalent CHANGELOG validator ran; executable sources are unchanged." \
-            >> "$GITHUB_STEP_SUMMARY"
+          if [ "$RELEASE_SEAL_ONLY" = true ]; then
+            echo "Only the trusted release-seal and fragment validators ran; executable sources are unchanged." \
+              >> "$GITHUB_STEP_SUMMARY"
+          else
+            echo "Only the trusted base-equivalent CHANGELOG validator ran; executable sources are unchanged." \
+              >> "$GITHUB_STEP_SUMMARY"
+          fi

      - name: Validate scoped policy
        if: ${{ needs.lint.outputs.changelog_only != 'true' && (needs.lint.outputs.docs_only == 'true' || (needs.lint.outputs.full_suite != 'true' && needs.lint.outputs.interface_sensitive != 'true')) }}
@@ -1368,29 +1684,47 @@ jobs:
          if [ -z "$base_ref" ] || [ "$base_ref" = "0000000000000000000000000000000000000000" ]; then
            base_ref="$(git rev-parse HEAD^)"
          fi
+          candidate_ref="$(git rev-parse 'HEAD^{commit}')"
+          if [ "$GITHUB_EVENT_NAME" = "pull_request" ] && [ "$candidate_ref" != "$PR_HEAD_SHA" ]; then
+            echo "Compatibility checkout $candidate_ref does not match PR head $PR_HEAD_SHA" >&2
+            exit 1
+          fi
          git rev-parse --verify "${base_ref}^{commit}" >/dev/null
-          stable_ref="$(git tag --merged "$base_ref" --list 'v[0-9]*' --sort=-version:refname | awk 'index($0, "-") == 0 { print; exit }')"
+          . ./scripts/release/release-lib.sh
+          stable_ref=""
+          for tag in $(git tag --merged "$base_ref" --list 'v*' --sort=-version:refname); do
+            release_is_stable_version "$tag" || continue
+            if git rev-parse --verify --quiet "refs/tags/withdrawn/$tag" >/dev/null; then
+              continue
+            fi
+            stable_ref="$tag"
+            break
+          done
          if [ -z "$stable_ref" ]; then
            echo "No stable release tag is reachable from compatibility base $base_ref" >&2
            exit 1
          fi
          git rev-parse --verify "${stable_ref}^{commit}" >/dev/null
-          echo "COMPATIBILITY_BASE_REF=$base_ref" >> "$GITHUB_ENV"
-          echo "COMPATIBILITY_STABLE_REF=$stable_ref" >> "$GITHUB_ENV"
+          printf '%s\n' \
+            "COMPATIBILITY_BASE_REF=$base_ref" \
+            "COMPATIBILITY_STABLE_REF=$stable_ref" \
+            "COMPATIBILITY_CANDIDATE_REF=$candidate_ref" >> "$GITHUB_ENV"

-      - name: Check historical commands and help compatibility
+      - name: Check historical commands, help, and complete CLI compatibility
        if: ${{ needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true' && (needs.lint.outputs.full_suite == 'true' || needs.lint.outputs.interface_sensitive == 'true') }}
        run: |
          make authoritative-interface-integrity \
-            BASE_REF="$COMPATIBILITY_BASE_REF"
-          if [ "$(git rev-parse "${COMPATIBILITY_BASE_REF}^{commit}")" != "$(git rev-parse "${COMPATIBILITY_STABLE_REF}^{commit}")" ]; then
-            make authoritative-interface-integrity \
-              BASE_REF="$COMPATIBILITY_STABLE_REF"
-          fi
+            BASE_REF="$COMPATIBILITY_BASE_REF" \
+            STABLE_REF="$COMPATIBILITY_STABLE_REF" \
+            CANDIDATE_REF="$COMPATIBILITY_CANDIDATE_REF"

      - name: Check complete Schema compatibility
        if: ${{ needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true' && (needs.lint.outputs.full_suite == 'true' || needs.lint.outputs.interface_sensitive == 'true') }}
-        run: make schema-compatibility BASE_REF="$COMPATIBILITY_BASE_REF"
+        run: |
+          make schema-compatibility \
+            BASE_REF="$COMPATIBILITY_BASE_REF" \
+            STABLE_REF="$COMPATIBILITY_STABLE_REF" \
+            CANDIDATE_REF="$COMPATIBILITY_CANDIDATE_REF"

      - name: Check skill command references
        if: ${{ needs.lint.outputs.changelog_only != 'true' && needs.lint.outputs.docs_only != 'true' && (needs.lint.outputs.full_suite == 'true' || needs.lint.outputs.interface_sensitive == 'true') }}
@@ -0,0 +1,296 @@
+name: PR Eval Dispatch
+
+# `/eval <products> [sha=<full-head-sha>] [cases=<ref>]` PR 评论 → 生成可验证的评测请求，报告由 bot 回贴。
+# 本 workflow 只在默认分支上下文运行，不 checkout、不执行 PR 代码。
+# 审核 SHA 规则：评测他人 PR 必须显式携带 sha=（审阅背书凭据，验证
+# 其恰为当前 open head）；评测自己创建的 PR 可省略，自动钉住派发时刻
+# 的当前 head（作者自背书，无第三方偷换窗口）；受控评测执行端另以
+# FETCH_HEAD 校验兜底派发后的变更。
+# 授权两级：仓库 write/maintain/admin 可派发任意 PR；默认分支
+# .github/eval-allowlist.txt 名单内的用户仅可派发自己创建的 PR。
+# 触发通道：workflow 先创建占位评论，再上传与本次 run/comment 绑定的
+# 不可变 manifest artifact，最后把 artifact 指针写回同一评论。评论仅是
+# 不可信通知；受控评测服务必须验证成功 run、artifact 与 manifest，并在
+# 触发评测前原子占用 manifest.idempotency_key，重复占用只能 no-op。
+
+on:
+  issue_comment:
+    types:
+      - created
+
+permissions: {}
+
+concurrency:
+  group: eval-dispatch-${{ github.event.issue.number }}
+  cancel-in-progress: false
+
+jobs:
+  dispatch:
+    name: Dispatch internal evaluation
+    if: >-
+      github.event.issue.pull_request &&
+      startsWith(github.event.comment.body, '/eval')
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    permissions:
+      contents: read
+      # 该 job 仅处理 PR；评论写入也限定在 PR Conversation 这一权限域。
+      pull-requests: write
+    steps:
+      - name: Check out default branch tooling
+        uses: actions/checkout@v4
+
+      - name: Verify commenter dispatch authorization
+        env:
+          GH_TOKEN: ${{ github.token }}
+          COMMENTER: ${{ github.event.comment.user.login }}
+          PR_AUTHOR: ${{ github.event.issue.user.login }}
+          EVAL_ALLOWLIST_PATH: .github/eval-allowlist.txt
+        run: |
+          # 不用 --fail：非协作者查权限返回 404 错误体，交由 guard 走名单分支；硬网络错误降级为空对象同样 fail-closed
+          permission_json="$(curl --silent --show-error \
+            -H "Authorization: Bearer ${GH_TOKEN}" \
+            -H "Accept: application/vnd.github+json" \
+            "https://api.github.com/repos/${GITHUB_REPOSITORY}/collaborators/${COMMENTER}/permission")" || permission_json='{}'
+          printf '%s' "$permission_json" | python3 scripts/ci/eval_dispatch_guard.py permission
+
+      - name: Parse /eval command
+        id: parse
+        continue-on-error: true
+        env:
+          COMMENT_BODY: ${{ github.event.comment.body }}
+        run: python3 scripts/ci/eval_comment_parse.py
+
+      - name: Reply usage on parse failure
+        if: steps.parse.outcome == 'failure'
+        env:
+          GH_TOKEN: ${{ github.token }}
+          PR_NUMBER: ${{ github.event.issue.number }}
+          PARSE_ERROR: ${{ steps.parse.outputs.error }}
+        run: |
+          body="❌ /eval 命令解析失败：${PARSE_ERROR}"
+          gh api --method POST \
+            "repos/${GITHUB_REPOSITORY}/issues/${PR_NUMBER}/comments" \
+            --raw-field body="$body" \
+            > /dev/null
+          exit 1
+
+      - name: Verify reviewed PR head
+        id: pr
+        env:
+          GH_TOKEN: ${{ github.token }}
+          PR_NUMBER: ${{ github.event.issue.number }}
+          EXPECTED_PR_NUMBER: ${{ github.event.issue.number }}
+          REVIEWED_SHA: ${{ steps.parse.outputs.reviewed_sha }}
+          COMMENTER: ${{ github.event.comment.user.login }}
+        run: |
+          pr_json="$(curl --fail --silent --show-error \
+            -H "Authorization: Bearer ${GH_TOKEN}" \
+            -H "Accept: application/vnd.github+json" \
+            "https://api.github.com/repos/${GITHUB_REPOSITORY}/pulls/${PR_NUMBER}")"
+          printf '%s' "$pr_json" \
+            | python3 scripts/ci/eval_dispatch_guard.py head \
+            >> "$GITHUB_OUTPUT"
+
+      - name: Create dispatch placeholder
+        id: placeholder
+        env:
+          GH_TOKEN: ${{ github.token }}
+          PR_NUMBER: ${{ github.event.issue.number }}
+        run: |
+          set -euo pipefail
+          placeholder_body="🛰️ /eval 请求已通过权限与版本校验，正在生成可验证的评测请求。"
+          response="$(
+            gh api --method POST \
+              "repos/${GITHUB_REPOSITORY}/issues/${PR_NUMBER}/comments" \
+              --raw-field body="$placeholder_body"
+          )"
+          comment_id="$(
+            printf '%s' "$response" \
+              | jq -er \
+                --arg issue_url "https://api.github.com/repos/${GITHUB_REPOSITORY}/issues/${PR_NUMBER}" \
+                'select(.issue_url == $issue_url) | .id | tostring | select(test("^[1-9][0-9]*$"))'
+          )"
+          printf 'comment_id=%s\n' "$comment_id" >> "$GITHUB_OUTPUT"
+
+      - name: Build dispatch request manifest
+        env:
+          REPOSITORY_ID: '1187709537'
+          REPOSITORY: ${{ github.repository }}
+          WORKFLOW_ID: '331725458'
+          WORKFLOW_PATH: .github/workflows/eval-dispatch.yml
+          RUN_ID: ${{ github.run_id }}
+          RUN_ATTEMPT: ${{ github.run_attempt }}
+          SOURCE_COMMENT_ID: ${{ github.event.comment.id }}
+          DISPATCH_COMMENT_ID: ${{ steps.placeholder.outputs.comment_id }}
+          ACTOR_ID: ${{ github.event.comment.user.id }}
+          ACTOR_LOGIN: ${{ github.event.comment.user.login }}
+          PR_NUMBER: ${{ github.event.issue.number }}
+          PR_HEAD_SHA: ${{ steps.pr.outputs.head_sha }}
+          PRODUCTS: ${{ steps.parse.outputs.products }}
+          CASES_REF: ${{ steps.parse.outputs.cases_ref }}
+          SOURCE_BODY: ${{ github.event.comment.body }}
+          MANIFEST_PATH: ${{ runner.temp }}/eval-dispatch-request.json
+        run: |
+          set -euo pipefail
+          if [ "$REPOSITORY" != "DingTalk-Real-AI/dingtalk-workspace-cli" ]; then
+            echo "unexpected repository: ${REPOSITORY}" >&2
+            exit 1
+          fi
+          for value in \
+            "$REPOSITORY_ID" \
+            "$WORKFLOW_ID" \
+            "$RUN_ID" \
+            "$RUN_ATTEMPT" \
+            "$SOURCE_COMMENT_ID" \
+            "$DISPATCH_COMMENT_ID" \
+            "$ACTOR_ID" \
+            "$PR_NUMBER"; do
+            if [[ ! "$value" =~ ^[1-9][0-9]*$ ]]; then
+              echo "dispatch manifest contains a non-canonical identifier" >&2
+              exit 1
+            fi
+          done
+          if [[ ! "$PR_HEAD_SHA" =~ ^[0-9a-f]{40}$ ]]; then
+            echo "dispatch manifest contains an invalid PR head SHA" >&2
+            exit 1
+          fi
+
+          hash_output="$(printf '%s' "$SOURCE_BODY" | sha256sum)"
+          source_body_sha256="${hash_output%% *}"
+          if [[ ! "$source_body_sha256" =~ ^[0-9a-f]{64}$ ]]; then
+            echo "failed to hash source comment" >&2
+            exit 1
+          fi
+          idempotency_key="${REPOSITORY_ID}:${SOURCE_COMMENT_ID}"
+
+          umask 077
+          jq -n \
+            --arg repository_id "$REPOSITORY_ID" \
+            --arg repository "$REPOSITORY" \
+            --arg workflow_id "$WORKFLOW_ID" \
+            --arg workflow_path "$WORKFLOW_PATH" \
+            --arg run_id "$RUN_ID" \
+            --arg run_attempt "$RUN_ATTEMPT" \
+            --arg source_comment_id "$SOURCE_COMMENT_ID" \
+            --arg dispatch_comment_id "$DISPATCH_COMMENT_ID" \
+            --arg actor_id "$ACTOR_ID" \
+            --arg actor_login "$ACTOR_LOGIN" \
+            --arg pr_number "$PR_NUMBER" \
+            --arg pr_head_sha "$PR_HEAD_SHA" \
+            --arg products "$PRODUCTS" \
+            --arg cases_ref "$CASES_REF" \
+            --arg source_body_sha256 "$source_body_sha256" \
+            --arg idempotency_key "$idempotency_key" \
+            '{
+              schema_version: 1,
+              repository_id: $repository_id,
+              repository: $repository,
+              workflow_id: $workflow_id,
+              workflow_path: $workflow_path,
+              run_id: $run_id,
+              run_attempt: $run_attempt,
+              source_comment_id: $source_comment_id,
+              dispatch_comment_id: $dispatch_comment_id,
+              actor_id: $actor_id,
+              actor_login: $actor_login,
+              pr_number: $pr_number,
+              pr_head_sha: $pr_head_sha,
+              products: $products,
+              cases_ref: $cases_ref,
+              source_body_sha256: $source_body_sha256,
+              idempotency_key: $idempotency_key
+            }' > "$MANIFEST_PATH"
+
+      - name: Upload dispatch request manifest
+        id: artifact
+        uses: actions/upload-artifact@v4
+        with:
+          name: eval-dispatch-request-${{ github.run_id }}-${{ github.run_attempt }}-${{ steps.placeholder.outputs.comment_id }}
+          path: ${{ runner.temp }}/eval-dispatch-request.json
+          if-no-files-found: error
+          retention-days: 1
+          overwrite: false
+
+      - name: Finalize dispatch marker
+        env:
+          GH_TOKEN: ${{ github.token }}
+          DISPATCH_COMMENT_ID: ${{ steps.placeholder.outputs.comment_id }}
+          REPOSITORY_ID: '1187709537'
+          WORKFLOW_ID: '331725458'
+          WORKFLOW_PATH: .github/workflows/eval-dispatch.yml
+          RUN_ID: ${{ github.run_id }}
+          RUN_ATTEMPT: ${{ github.run_attempt }}
+          ARTIFACT_ID: ${{ steps.artifact.outputs.artifact-id }}
+          ARTIFACT_DIGEST: ${{ steps.artifact.outputs.artifact-digest }}
+          PR_HEAD_SHA: ${{ steps.pr.outputs.head_sha }}
+          PRODUCTS: ${{ steps.parse.outputs.products }}
+          CASES_REF: ${{ steps.parse.outputs.cases_ref }}
+        run: |
+          set -euo pipefail
+          if [[ ! "$DISPATCH_COMMENT_ID" =~ ^[1-9][0-9]*$ ]] || \
+             [[ ! "$ARTIFACT_ID" =~ ^[1-9][0-9]*$ ]]; then
+            echo "artifact marker contains a non-canonical identifier" >&2
+            exit 1
+          fi
+
+          artifact_digest="${ARTIFACT_DIGEST,,}"
+          if [[ "$artifact_digest" != sha256:* ]]; then
+            artifact_digest="sha256:${artifact_digest}"
+          fi
+          if [[ ! "$artifact_digest" =~ ^sha256:[0-9a-f]{64}$ ]]; then
+            echo "artifact marker contains an invalid digest" >&2
+            exit 1
+          fi
+
+          marker_json="$(
+            jq -nc \
+              --arg repository_id "$REPOSITORY_ID" \
+              --arg workflow_id "$WORKFLOW_ID" \
+              --arg workflow_path "$WORKFLOW_PATH" \
+              --arg run_id "$RUN_ID" \
+              --arg run_attempt "$RUN_ATTEMPT" \
+              --arg dispatch_comment_id "$DISPATCH_COMMENT_ID" \
+              --arg artifact_id "$ARTIFACT_ID" \
+              --arg artifact_digest "$artifact_digest" \
+              '{
+                schema_version: 1,
+                repository_id: $repository_id,
+                workflow_id: $workflow_id,
+                workflow_path: $workflow_path,
+                run_id: $run_id,
+                run_attempt: $run_attempt,
+                dispatch_comment_id: $dispatch_comment_id,
+                artifact_id: $artifact_id,
+                artifact_digest: $artifact_digest
+              }'
+          )"
+          cases_note=""
+          if [ -n "$CASES_REF" ]; then
+            cases_note="，用例版本 \`${CASES_REF}\`"
+          fi
+          body="<!-- eval-dispatch: ${marker_json} -->"$'\n'"🛰️ /eval 已受理：产品集 \`${PRODUCTS}\`${cases_note}，评测对象 \`${PR_HEAD_SHA}\`。"$'\n'"受控评测服务将在数分钟内处理，完成后由 bot 回贴报告。"
+          response="$(
+            gh api --method PATCH \
+              "repos/${GITHUB_REPOSITORY}/issues/comments/${DISPATCH_COMMENT_ID}" \
+              --raw-field body="$body"
+          )"
+          printf '%s' "$response" \
+            | jq -e \
+              --arg comment_id "$DISPATCH_COMMENT_ID" \
+              --arg body "$body" \
+              '((.id | tostring) == $comment_id) and (.body == $body)' \
+              > /dev/null
+
+      - name: Mark dispatch preparation failure
+        if: ${{ failure() && steps.placeholder.outputs.comment_id != '' }}
+        env:
+          GH_TOKEN: ${{ github.token }}
+          DISPATCH_COMMENT_ID: ${{ steps.placeholder.outputs.comment_id }}
+        run: |
+          failure_body="❌ /eval 请求准备失败，未生成可消费的评测请求。请稍后重试。"
+          gh api --method PATCH \
+            "repos/${GITHUB_REPOSITORY}/issues/comments/${DISPATCH_COMMENT_ID}" \
+            --raw-field body="$failure_body" \
+            > /dev/null \
+            || true
@@ -2645,6 +2645,40 @@ jobs:
          "$GITHUB_WORKSPACE/tmp/trusted-release-tooling/scripts/release/verify-github-tag-authority.sh" \
            "$RELEASE_VERSION" "$RELEASE_COMMIT" "$RELEASE_TAG_OBJECT"

+      # The sealed candidate tag is intentionally visible while its GitHub
+      # authority is checked above. Compatibility must instead discover the
+      # previous delivered stable tag, so hide only this verified candidate
+      # from this isolated runner's local tag namespace.
+      - name: Prepare delivered-stable compatibility ref view
+        if: ${{ matrix.check == 'compatibility' }}
+        env:
+          RELEASE_VERSION: ${{ needs.release-contract.outputs.release_version }}
+          RELEASE_COMMIT: ${{ needs.release-contract.outputs.release_commit }}
+          RELEASE_TAG_OBJECT: ${{ needs.release-contract.outputs.release_tag_object }}
+          PREVIOUS_STABLE: ${{ needs.release-contract.outputs.previous_stable }}
+          PREVIOUS_STABLE_COMMIT: ${{ needs.release-contract.outputs.previous_stable_commit }}
+        run: |
+          set -eu
+          test -n "$RELEASE_VERSION"
+          test -n "$RELEASE_COMMIT"
+          test -n "$RELEASE_TAG_OBJECT"
+          test -n "$PREVIOUS_STABLE"
+          test -n "$PREVIOUS_STABLE_COMMIT"
+          test "$RELEASE_VERSION" != "$PREVIOUS_STABLE"
+          test "$(git rev-parse HEAD)" = "$RELEASE_COMMIT"
+          test "$(git rev-parse --verify "refs/tags/${RELEASE_VERSION}")" = "$RELEASE_TAG_OBJECT"
+          test "$(git rev-parse --verify "refs/tags/${RELEASE_VERSION}^{commit}")" = "$RELEASE_COMMIT"
+          test "$(git rev-parse --verify "${PREVIOUS_STABLE}^{commit}")" = "$PREVIOUS_STABLE_COMMIT"
+
+          git update-ref -d "refs/tags/${RELEASE_VERSION}" "$RELEASE_TAG_OBJECT"
+
+          if git show-ref --verify --quiet "refs/tags/${RELEASE_VERSION}"; then
+            echo "sealed candidate tag is still visible to compatibility baseline discovery" >&2
+            exit 2
+          fi
+          test "$(git rev-parse HEAD)" = "$RELEASE_COMMIT"
+          test "$(git rev-parse --verify "${PREVIOUS_STABLE}^{commit}")" = "$PREVIOUS_STABLE_COMMIT"
+
      - name: Set up Go
        uses: actions/setup-go@v5
        with:
@@ -2793,7 +2827,7 @@ jobs:
            fi
            if test "${{ needs.dispatch-contract.outputs.mode }}" = plan_release; then
              echo
-              echo "Plan only: no tag or package was created. Add the exact \`CHANGELOG.md\` section, merge it to main, then run publish."
+              echo "Plan only: no tag or package was created. Render pending \`.changes/*.md\` fragments into the exact \`CHANGELOG.md\` section, merge the release-seal PR to main, then run publish."
            fi
          } >> "$GITHUB_STEP_SUMMARY"

@@ -21,6 +21,11 @@ jobs:
    runs-on: ubuntu-latest
    timeout-minutes: 5
    steps:
+      - name: Check out trusted routing policy
+        uses: actions/checkout@11bd71901bbe5b1630ceea73d27597364c9af683 # v4.2.2
+        with:
+          ref: ${{ github.event.pull_request.base.sha }}
+          persist-credentials: false
      - name: Route review and enable auto-merge
        uses: actions/github-script@f28e40c7f34bde8b3046d885e986cb6290c5673b # v7.1.0
        with:
@@ -29,12 +34,13 @@ jobs:
            const repo = context.repo.repo;
            const pullNumber = context.payload.pull_request.number;
            const eventHeadSha = context.payload.pull_request.head.sha;
-            const reviewerPool = [
-              'sczheng189',
-              'shangguanxuan633-lab',
-              'audanye-sudo',
-              'wxianfeng',
-            ];
+            const {
+              REVIEWER_POOL,
+              requestReviewersWithFallback,
+              resolveReviewRouting,
+              reviewerCandidates,
+            } = require('./.github/reviewer-routing.js');
+            const reviewerPool = REVIEWER_POOL;

            async function getReadyEventPull(phase) {
              const {data: currentPull} = await github.rest.pulls.get({
@@ -67,6 +73,21 @@ jobs:
                : author;

            async function routeReview() {
+              let changedFiles;
+              try {
+                changedFiles = await github.paginate(github.rest.pulls.listFiles, {
+                  owner,
+                  repo,
+                  pull_number: pullNumber,
+                  per_page: 100,
+                });
+              } catch (error) {
+                core.warning(
+                  `Could not inspect changed files for PR #${pullNumber}; using load-balanced fallback (${error.status || 'unknown status'}).`,
+                );
+                changedFiles = [];
+              }
+
              const eligible = reviewerPool.filter(
                reviewer =>
                  reviewer.toLowerCase() !== author &&
@@ -77,13 +98,9 @@ jobs:
                return;
              }

-              const alreadyRequested =
-                (pullRequest.requested_reviewers || []).length > 0 ||
-                (pullRequest.requested_teams || []).length > 0;
-              if (alreadyRequested) {
-                core.info(`PR #${pullNumber} already has a requested reviewer; leaving it unchanged.`);
-                return;
-              }
+              const existingRequestedReviewers = new Set(
+                (pullRequest.requested_reviewers || []).map(({login}) => login.toLowerCase()),
+              );

              let reviews;
              try {
@@ -116,22 +133,18 @@ jobs:
                  latestDecisionByLogin.set(login, review);
                }
              }
-              const currentHeadDecision = [...latestDecisionByLogin.values()].find(
-                review =>
-                  review.commit_id === headSha &&
-                  eligible.some(
-                    reviewer =>
-                      reviewer.toLowerCase() ===
-                      review.user.login.toLowerCase(),
-                  ) &&
-                  ['APPROVED', 'CHANGES_REQUESTED'].includes(review.state),
+              const currentHeadReviewers = new Set(
+                [...latestDecisionByLogin.values()]
+                  .filter(
+                    review =>
+                      review.commit_id === headSha &&
+                      eligible.some(
+                        reviewer => reviewer.toLowerCase() === review.user.login.toLowerCase(),
+                      ) &&
+                      ['APPROVED', 'CHANGES_REQUESTED'].includes(review.state),
+                  )
+                  .map(review => review.user.login.toLowerCase()),
              );
-              if (currentHeadDecision) {
-                core.info(
-                  `PR #${pullNumber} already has a ${currentHeadDecision.state} review on its current head; leaving review ownership unchanged.`,
-                );
-                return;
-              }

              const loads = new Map(eligible.map(reviewer => [reviewer, 0]));
              try {
@@ -178,20 +191,42 @@ jobs:
                  tieOrder.get(left) - tieOrder.get(right),
              );

-              for (const reviewer of ranked) {
-                try {
+              const routing = resolveReviewRouting({
+                files: changedFiles,
+                author,
+                latestPusher,
+                fallbackReviewers: ranked,
+              });
+              const candidates = reviewerCandidates({
+                preferredReviewers: routing.reviewers,
+                fallbackReviewers: ranked,
+                eligibleReviewers: eligible,
+              });
+              const desiredReviewers = candidates.slice(0, routing.requiredReviewers);
+              if (routing.reason === 'unknown_paths' && currentHeadReviewers.size > 0) {
+                core.info(
+                  `PR #${pullNumber} has a current-head review for unknown paths; leaving manual ownership unchanged.`,
+                );
+                return;
+              }
+              core.info(
+                `PR #${pullNumber} routing: ${routing.reason}; modules=${routing.modules.map(module => module.id).join(',') || 'unknown'}; reviewers=${desiredReviewers.join(',') || 'load-balanced fallback'}.`,
+              );
+
+              const satisfiedReviewers = new Set([
+                ...currentHeadReviewers,
+                ...[...existingRequestedReviewers].filter((reviewer) =>
+                  candidates.some((candidate) => candidate.toLowerCase() === reviewer),
+                ),
+              ]);
+              const requestResult = await requestReviewersWithFallback({
+                candidates,
+                requiredReviewers: routing.requiredReviewers,
+                satisfiedReviewers: [...satisfiedReviewers],
+                requestReviewer: async (reviewer) => {
                  const currentPull = await getReadyEventPull('review request');
                  if (!currentPull) {
-                    return;
-                  }
-                  if (
-                    (currentPull.requested_reviewers || []).length > 0 ||
-                    (currentPull.requested_teams || []).length > 0
-                  ) {
-                    core.info(
-                      `PR #${pullNumber} received a reviewer while routing; leaving it unchanged.`,
-                    );
-                    return;
+                    return false;
                  }
                  await github.rest.pulls.requestReviewers({
                    owner,
@@ -202,15 +237,23 @@ jobs:
                  core.info(
                    `Requested @${reviewer} for PR #${pullNumber} (open request load: ${loads.get(reviewer)}).`,
                  );
-                  return;
-                } catch (error) {
+                  return true;
+                },
+                onFailure: (reviewer, error) => {
                  core.warning(
                    `Could not request @${reviewer} for PR #${pullNumber}; trying the next candidate (${error.status || 'unknown status'}).`,
                  );
-                }
+                },
+              });
+              if (requestResult.aborted) {
+                return;
              }

-              core.warning(`No reviewer request could be created for PR #${pullNumber}.`);
+              if (requestResult.satisfiedReviewers.length < routing.requiredReviewers) {
+                core.warning(
+                  `Only ${requestResult.satisfiedReviewers.length} of ${routing.requiredReviewers} required reviewers could be satisfied for PR #${pullNumber}.`,
+                );
+              }
            }

            async function enableAutoMerge() {
@@ -20,6 +20,10 @@ test/cli_compat/testdata/
 .gitignore
 .worktrees/
 .qoder/
+_logs/
+_docs/
+_output/
+vendor/

 # Secrets & credentials
 .env
@@ -302,9 +302,8 @@ on the leaf:

 ```bash
 dws auth status                 # token_valid should be true
-dws cache refresh               # deprecated no-op: prints a retirement notice (discovery cache is gone; refreshes nothing)
-dws schema <mcp-canonical> -f json
-# or CLI path: dws schema --cli-path "drive copy" -f json
+dws schema <mcp-canonical> --jq '{canonical_path,interface_ref,parameters}' -f json
+# or CLI path: dws schema --cli-path "drive copy" --jq '{canonical_path,interface_ref,parameters}' -f json
 ```

 Resolve MCP identity via declared `interface_ref` when CLI canonical ≠ MCP path
@@ -321,8 +320,10 @@ Skill (evidence only)**.

 Split work by product groups. Each agent must:

- Read Skill, Cobra/`--help`, Runtime confirmation sites, and live `dws schema`
-  for its tools.
+- Read Skill, Cobra/`--help`, Runtime confirmation sites, and live
+  `dws schema <leaf> --compact` for its tools. Mapping/interface/provenance
+  audits may query the full leaf only through a narrow `--jq` / `--fields`
+  projection; do not load an entire full leaf into Agent context.
 - Hand-write selection prose and leaf Contract / ProductDecl declarations;
  forbid wholesale JSON merges from review dumps.
 - Edit only its product’s leaf declarations (and `ProductDecl` when needed).
@@ -463,6 +464,134 @@ Keep CLI confirmation behavior and Schema metadata consistent, and add a
 semantic regression test through the final embedded loader/query delivery
 path; a generator unit test or JSON count alone is insufficient.

+## Unified result Schema and performance
+
+The unified runtime envelope and the per-command Schema result declaration are
+related but distinct contracts:
+
+- Runtime owns the outer machine envelope (`ok`, `outcome`, `data`, `error`,
+  `meta`) and derives it through `internal/output`. Business commands return a
+  `CommandResult`; they must not hand-author the outer JSON shape.
+- A leaf `Contract.Result` / `contract.ResultSpec` describes the reviewed
+  business value inside `data`. It may declare `outcomes`, `data_schema`, and
+  `sensitive_paths`. `Contract.Pagination` is a separate command capability
+  because pagination is emitted under envelope `meta`, not inside `data`.
+- `outcomes` is the set of results a command may produce; it is not the outcome
+  of the current invocation. `data_schema` is a JSON Schema object for business
+  data and must not duplicate the framework envelope.
+- Result declarations are delivered in the full leaf and in the reviewed
+  `--compact` Agent projection. Compact retains the normalized `result` object
+  verbatim but still omits provenance, interface bindings, and other audit-only
+  fields. Product/group summaries remain navigation views and need not repeat
+  every leaf Result. When an Agent needs return-shape facts, query the compact
+  leaf directly; do not load the whole full Catalog.
+- A missing `result` means “no reviewed return-value declaration is published
+  for this leaf.” It does **not** prove that the runtime is legacy, and it must
+  not be filled by inference from examples, MCP samples, or previous command
+  output. Runtime rollout remains an internal per-command fact.
+- The public contract has no `contract_version`, no `--output-contract`, and no
+  Agent-selectable protocol alias. Agents continue to request machine output
+  with `--format json`; migrated commands use the unified result directly and
+  unmigrated commands retain their current legacy output.
+- Existing `dev` / `devapp` pilot coverage is gradual. Active reviewed
+  `devapp` shortcuts are gated on a non-empty Result declaration, while `dev`
+  currently has representative Result coverage. Do not describe that as
+  repository-wide coverage. Any newly activated Agent-visible command should
+  add and test its Result declaration; the remaining pilot gaps should shrink,
+  not expand.
+
+The compact/full leaf `result` object has one stable shape:
+
+```json
+{
+  "result": {
+    "outcomes": ["success", "pending", "partial_failure", "failure"],
+    "data_schema": {
+      "type": "object",
+      "properties": {
+        "items": {
+          "type": "array",
+          "items": {
+            "type": "object",
+            "properties": {
+              "id": {"type": "string", "description": "Stable resource ID"},
+              "name": {"type": "string", "description": "Display name"}
+            }
+          }
+        }
+      }
+    },
+    "sensitive_paths": ["credential.secret"]
+  },
+  "pagination": {
+    "kind": "cursor",
+    "cursor_parameter": "cursor",
+    "meta_path": "meta.pagination",
+    "endpoint_exhausted_path": "meta.pagination.endpoint_exhausted",
+    "next_token_path": "meta.pagination.next_token"
+  }
+}
+```
+
+Field rules:
+
+| Field | Required | Contract |
+|---|---|---|
+| `outcomes` | yes | Non-empty unique subset of `success`, `pending`, `partial_failure`, `failure`; normalization publishes canonical order. |
+| `data_schema` | yes | One recursive JSON Schema **object** describing only the runtime envelope's `data` value. Every named `properties` child must have a non-empty `description`. It must not duplicate `ok`, `outcome`, `error`, or `meta`. |
+| `sensitive_paths` | no | Unique safe dot paths relative to `data`; renderers/redaction consumers must not treat them as shell/JQ expressions. |
+
+Optional members are omitted, never emitted as `null`. A leaf without a
+reviewed Result omits the entire `result` key. Compact must preserve the same
+normalized Result value as the full leaf; it must not summarize, infer, rename,
+or independently rebuild any Result field. Product/group summaries do not
+aggregate child Result objects.
+
+`pagination` is a sibling of `result`, not a child. It declares the canonical
+CLI cursor parameter and the fixed framework paths under `meta.pagination`.
+Product response fields used to derive that metadata remain mapper internals;
+they are not part of `result.data_schema`. Do not execute a second request to
+derive pagination metadata.
+
+Invalid result declarations fail closed during normalization: unknown or
+duplicate outcomes, a non-object/multiple `data_schema`, unsafe or duplicate
+sensitive paths, unsupported pagination kinds, attempts to override framework
+meta paths, and an invalid cursor parameter must be rejected rather than
+silently removed.
+Full-leaf wire round trips must
+preserve the normalized Result exactly. Do not commit generated Schema JSON as
+evidence; tests construct contracts in Go and runtime/CI assemble the Catalog
+from declarations.
+
+### Performance model and rules
+
+- Catalog construction is declaration-driven and cached through the existing
+  lazy `sync.Once` delivery path. Do not reassemble or reopen annotations per
+  command invocation, per leaf lookup, or per renderer.
+- Normalizing one Result declaration is linear in the size of that declaration.
+  Full `schema --all` is linear in tools + parameters + Result schema bytes and
+  is an audit/compatibility export, not the normal Agent discovery path.
+  Overview → compact product/group → compact leaf remains the normal route;
+  only the final leaf carries its Result declaration.
+- Constructing a `CommandResult` defensively clones result data and validates
+  invariants; rendering is buffer-first and then writes once. Both CPU cost and
+  transient memory are O(payload size), with roughly one additional in-memory
+  rendered copy. This buys immutability and prevents partial JSON leakage, but
+  it is not free.
+- Large list/search commands must use bounded pages and publish continuation
+  facts. The current emitter buffers one command result/page before publishing;
+  pagination is the memory bound. Continuous event streams are a separate,
+  command-specific protocol and are not described by `ResultSpec`.
+- A `dual_validate` command must execute the business request exactly once,
+  validate a shadow unified result, and preserve legacy bytes. Never obtain
+  validation by issuing a second network or write request.
+- Filters and alternate formats are render-time work over the same in-memory
+  result. They must not rerun the business operation or rebuild Schema.
+- Performance changes must preserve the one-result, buffer-first, fail-closed,
+  and atomic `--output` guarantees. Do not trade correctness for a microbenchmark
+  improvement. For a material hot-path change, benchmark representative small
+  and page-sized payloads and report allocations/bytes as well as latency.
+
 ## Current Schema boundaries

 - `schema list` remains a progressive overview. `schema --all` is the stable
@@ -471,13 +600,16 @@ path; a generator unit test or JSON count alone is insufficient.
  `parameters` object for commands without flags. Keep it suitable for the #602
  compatibility baseline and fail rather than silently emitting a partial
  export.
- `schema --all` is not normal command discovery. Use overview -> product/group
-  -> leaf for routine Agent work. `--compact` is supported for context-saving
-  projections, but a compact full export is not a complete compatibility
-  baseline.
+- `schema --all` is not normal command discovery. Use overview -> compact
+  product/group -> compact leaf for routine Agent work. `--compact` is the
+  reviewed positive-field allowlist for Agent context: new full/audit fields
+  must not appear there until explicitly reviewed. A compact full export is not
+  a complete compatibility baseline.
 - `dws <path> --help` defines whether Cobra exposes a path and which flags the
-  executable accepts. A leaf Schema defines Agent selection, parameter mapping
-  and constraints, and safety/confirmation semantics. A conflict is contract
-  drift, not permission to guess.
+  executable accepts. A compact leaf defines Agent selection, CLI parameters,
+  constraints, safety/confirmation semantics, and any reviewed `result`
+  contract. Full leaf fields such as `property`, `interface_ref`, and
+  provenance are audit facts. A conflict is contract drift, not permission to
+  guess.
 - Schema and Help describe commands; neither returns DingTalk business data.
  After discovery, execute the real read/search/list command to obtain data.
@@ -6,8 +6,435 @@ The format is inspired by [Keep a Changelog](https://keepachangelog.com/) and th

 ## [Unreleased]

+## [1.0.59-beta.1] - 2026-08-14
+
+### Added
+
+- **Drive list type/time filtering** (#942) — `dws drive list` gains `--type
+  file|folder`, `--start`, and `--end` for client-side filtering by node type
+  and modification time on both the pan and workspace routes. Filtering runs
+  a bounded full scan of the target directory (2000-entry cap, reported via
+  `truncated=true`), composes with `--latest`/`--pattern`/`--depth`, and is
+  mutually exclusive with `--versions`/`--cursor`/`--order-by`/`--order`/
+  `--limit`. Time values accept relative forms (`24h`/`7d`/`2w`), RFC 3339,
+  zone-less ISO 8601 (Asia/Shanghai), or a plain date.
+
+- **Drive folder synchronization** — adds `dws drive status`, `dws drive pull`,
+  `dws drive push`, and `dws drive sync` for file-level comparison and transfer
+  between a local folder and a Drive folder. Differences come from exact MD5 by
+  default or from modification time with `--quick`; `status` is read-only, `pull`
+  and `push` are one-directional with `--if-exists skip|smart|overwrite`, and
+  `sync` is bidirectional with `--on-conflict remote-wins|local-wins|keep-both|ask`.
+  Only regular files are transferred — online documents and shortcuts are skipped,
+  neither side deletes extra files, downloads are staged through a temporary file
+  and committed with an atomic rename, and remote names that would escape
+  `--local-folder` are reported as failures instead of being written. Every command
+  prints a structured summary on stdout and exits non-zero when any item fails.
+
+- **International DingTalk region support** — adds `.io` login and MCP routing, pre-release endpoint overrides, and profile-aware gateway selection while preserving the existing `.com` flow.
+
 ### Changed

+- **Chat identity routing** — validates explicit `openDingTalkId` inputs and improves name, `userId`, and `openDingTalkId` routing for message shortcuts.
+
+### Fixed
+
+- **Drive `--latest` refuses incomplete Top-N** (#899) — `dws drive list --latest` used to
+  exit 0 with a "Top-N" computed over a partially scanned tree whenever a directory read
+  failed mid-recursion (permission denied, API error), letting an incomplete set pose as the
+  globally newest files. Truncation at the 2000-item scan cap and mid-recursion directory
+  failures now both fail closed (`LATEST_SCAN_TRUNCATED` / `LATEST_SCAN_INCOMPLETE`), report
+  the first failing folder with its depth and reason, and emit a recovery command that
+  reproduces the original candidate set — query domain, `--folder`, `--pattern`, `--type`,
+  `--start` and `--end` are all carried over. On POSIX shells each user-supplied value is
+  quoted so a URL query string or a shell metacharacter cannot change how the copied command
+  parses. On Windows no quoting form is safe for both `cmd.exe` and PowerShell, so values
+  containing metacharacters are not inlined at all: the command carries a placeholder and the
+  original value is shown on a separate line marked as data rather than an executable command.
+  Unrecoverable errors under `--latest` return the root cause instead of a partial result.
+  Remote-controlled folder names and server error text are stripped of ANSI escapes and
+  control characters before they reach the plain-text stderr message. The internal `sortTime`
+  sort key no longer leaks into `drive list --depth` output on any path.
+
+- **Drive list pattern filtering** (#942) — `dws drive list --pattern` on the
+  single-layer pan route now filters the returned page by name pattern; the
+  flag was previously accepted but silently ignored.
+
+- **Drive list `--type folder --latest` composition** (#942) — `--latest` now
+  ranks the filtered entries (folders included when `--type folder` is set)
+  instead of unconditionally dropping folders, so the documented combination
+  returns the most recently modified folders rather than an empty list.
+
+- **Chat message time defaults** (#973) — default omitted `chat message list-all` time bounds in `Asia/Shanghai` when emitting timezone-less `yyyy-MM-dd HH:mm:ss` values, matching parsing semantics and rejecting reversed windows.
+
+- **Doc and Drive parameter aliases** — normalizes reviewed identifier, pagination, path, version, and role synonyms while blocking ambiguous values before dispatch.
+
+
+## [1.0.58] - 2026-08-13
+
+This release promotes the sealed `v1.0.58-beta.6` contents to stable.
+
+### Changed
+
+- **Expanded collaborative workflows** — adds full AI Table, Sheet, Minutes,
+  approval-event, Drive-comment, document export, and CSV workflow support,
+  including safer validation, explicit confirmation for writes, and
+  machine-readable completion receipts.
+- **More capable Chat operations** — adds robot image/file messages, toolbar
+  management, streaming-card mentions, automatic pagination controls, and
+  clearer post-send ID, Markdown-image, paging, and result-shape guidance.
+- **Reliable Agent and CLI contracts** — expands Agent-visible Chat and
+  Minutes commands, aligns bundled skills, improves schema/result envelopes,
+  and hardens parameter, pagination, runtime-token, and write-result
+  verification so ambiguous or incomplete operations fail closed.
+- **Multi-skill install and upgrade** — makes the multi-skill layout the
+  default for fresh installs and upgrades while preserving an explicit legacy
+  mono option.
+- **Safer release delivery** — strengthens release-equivalent compatibility,
+  sealing, package verification, and evaluation-dispatch checks for more
+  reliable cross-platform releases.
+
+## [1.0.58-beta.6] - 2026-08-13
+
+### Fixed
+
+- **npm package verification for multi-skill installs** (#991) — aligns the
+  release verifier with the installer’s concrete Agent skill-root selection,
+  preventing valid multi-skill package layouts from failing release delivery.
+
+### Changed
+
+- **Release-seal CI classification** (#987) — recognizes the reviewed
+  CHANGELOG-and-fragment archival shape while retaining release-contract and
+  lifecycle validation, reducing unrelated CI work for release-seal PRs.
+
+## [1.0.58-beta.5] - 2026-08-13
+
+### Added
+
+- **Agent version and extended context passthrough** (Aone 85384225) — adds
+  validated `DWS_AGENT_VER` and sensitive JSON `DWS_AGENT_EXT` metadata to
+  ordinary non-plugin MCP requests without forwarding it to A2A, OAuth,
+  Discovery, or third-party plugins.
+
+- **Drive file comments** (#961) — adds `dws drive comment list` and `dws drive comment create` for comments on ordinary preview files.
+
+- **Chat automatic pagination controls** (#970) — adds bounded `--max-items` and cancellable `--page-delay` support to the core IM list shortcuts, with safe continuation metadata and truncation reporting.
+
+### Changed
+
+- **Chat message send help** - Clarifies Markdown image syntax for inline mixed text and images.
+
+- **Doc/drive/wiki routing descriptions** — clarifies the document-space container-vs-content boundary across the doc, drive, and wiki skill descriptions for more predictable first-round Agent selection, without changing CLI behavior.
+
+
+## [1.0.58-beta.4] - 2026-08-12
+
+### Added
+
+- **Multi-skill installation and upgrade** — fresh installs, `dws skill setup`,
+  and `dws upgrade` now use the multi-skill layout by default. Existing mono
+  installations migrate during upgrade; mono remains an explicit legacy option.
+- **Native streaming-card mentions** — `dws chat message send-card` now accepts
+  `--at-open-dingtalk-ids` and `--at-all` for group cards and forwards them to
+  `create_and_send_card`, matching the existing shortcut behavior without
+  changing single-chat card creation.
+- **Expanded Minutes workflows** — 27 public Minutes shortcuts now cover
+  upload, download, export, recording, analysis, sharing, and recovery flows;
+  every write command keeps an explicit confirmation requirement.
+- **Chat command discovery** — 30 existing typed Chat commands are now
+  available in the runtime Schema and Agent catalog, with sensitive writes
+  carrying their required confirmation metadata.
+
+### Changed
+
+- **Chat read results** — typed commands and shortcuts now expose a consistent
+  top-level `messages` list with stable `messageId` and `text` fields while
+  retaining existing response envelopes and fields.
+- **Wiki feed results** — Wiki feed list output now formats time fields and
+  trims excess fields. Its `--limit` default is 10 and maximum is 20.
+- **Developer command results** — the `dev` and selected `devapp` commands now
+  use the unified result envelope for consistent success, pending, partial,
+  and failure reporting.
+- **Evaluation dispatch hardening** — `/eval` now uses a verifiable polling
+  relay instead of direct access from the hosted runner, binding the workflow,
+  comment, PR head, parameters, and result provenance.
+
+### Fixed
+
+- **Streaming-card update acknowledgement** — accepts the pre-production
+  `success: true` response from `update_streaming_card` as affirmative write
+  evidence while preserving explicit negative, conflicting, and bizId-drift
+  failures, so Agents do not repeat an update that the service already applied.
+- **Text input bounds** — literal input, stdin, and `@file` inputs now all
+  enforce the same byte limit; file reads validate the opened descriptor and
+  cannot exceed the limit after a path replacement or file growth.
+- **Evaluation PR comments** — restores `/eval` PR conversation comments with
+  the least required pull-request write permission and actionable GitHub 403
+  diagnostics.
+
+## [1.0.58-beta.3] - 2026-08-11
+
+### Added
+
+- **Aitable workflow execution and history** — adds `dws aitable workflow run` for confirmed asynchronous execution of scheduled or record-triggered workflows, plus `dws aitable workflow history` for status-, time-, and page-filtered execution records. The commands map directly to `aitable/run_workflow` and `aitable/get_flow_record_list`, validate trigger-specific arguments locally, and document the `executionId` / `instanceId` correlation.
+- **Streaming-card mentions** — `chat +messages-send-card` now accepts
+  `--at-open-dingtalk-ids` and `--at-all` for group cards, passing mention
+  targets to the initial card-creation request and prepending its returned
+  `atTag` to the automatic streaming update.
+- **Personal OA approval events** — personal event consumers now support task
+  creation, completion, redirection, instance start, termination, and
+  completion events, with typed output and matching usage documentation.
+
+### Fixed
+
+- **Machine-readable export and download receipts** — `dws doc export`,
+  `dws drive download`, and `dws drive download --version` now keep progress
+  logs on stderr under `--format json` and emit one JSON result on stdout after
+  a successful local write. The result includes the saved path and byte size;
+  document exports additionally report the node, requested format, job/task
+  ID, and final status.
+- **IM search and card-write safety** — conversation-scoped search now fails
+  closed when the target cannot be verified, and streaming-card updates require
+  business evidence rather than a transport-only success response.
+- **Document shortcut reliability** — document write, readback verification,
+  pagination, template/version discovery, export, media, and local-file
+  workflows now preserve compatibility while rejecting ambiguous write results.
+- **Event runtime-token handoff** — personal `event consume`, `status`,
+  `stop`, and `+listen-im` honor the root `--token` without falling back to a
+  stale OAuth profile. Detached buses negotiate an owner-only, memory-only IPC
+  credential channel; tokens are never placed in child argv, environment,
+  profiles, logs, or run-state files.
+
+### Changed
+
+- **Minutes `permission apply --policy` type** — `--policy` is now declared as
+  an `int` flag and its required check uses `Flags().Changed`, matching the
+  numeric-parameter convention. `--help` reports `int` instead of `string`;
+  accepted values (2/3/4) and gateway behavior are unchanged.
+- **Minutes skill references** — document `permission apply` in both Minutes
+  skill references: list it in the command trees, describe its policy values and
+  how it differs from `permission add`, and add its intent routing.
+- **Chat paging guidance** — typed chat message commands now document
+  `--page-all`, aggregate result shapes, and cursor behavior in CLI Help and
+  Agent selection examples.
+- **Calendar skill parity** — mono and multi Calendar references are aligned to
+  prevent documentation drift without changing CLI behavior.
+- **Release engineering** — CI now shards helper-package changes through the
+  full race suite, widens a flaky stdio idempotency test budget, governs exact
+  reviewed CLI/Schema type migrations, and lets authorized maintainers trigger
+  internal MCP evaluation with a reviewed `/eval` PR comment.
+
+## [1.0.58-beta.2] - 2026-08-10
+
+### Added
+
+- **`dws sheet create-with-data`（新命令）** — 建表并写入初始数据与样式：`--values`（二维数组写默认表）/ `--sheets`（typed table 多工作表，二者必须给一个）/ `--styles`（`cell_styles` / `row_sizes` / `col_sizes` / `cell_merges`，顶层键对齐飞书 snake_case、列表项内字段兼容 camelCase）。所有结构、字段类型与枚举在创建文档之前校验，非法配置不会留下白建的空文档：`--sheets` 按 `table_put` 的输入契约逐字段校验（`columns` 必填且列名非空不重复、`data` 为二维且行宽与 `columns` 一致、单元格仅限字符串/数字/布尔/null、`dtypes`/`formats` 的键须是列名、`mode`/`header`/`allowOverwrite`/`startCell` 类型与取值、单表 30000 单元格上限），并拒绝未知键、snake_case 变体、`{"sheets":"bad"}` 这类畸形包装与 `sheetId`（服务端会静默丢弃写错的键，导致"只写了表头却报成功"的静默丢数据）；`--values` 校验单元格为标量并受 30000 单元格 / 2000000 字符上限约束；`--styles` 的顶层键与列表项内字段同样拒绝未知键，避免样式只应用一半。写入后回读校验按 `startCell` / `header` / `mode` 推算的首个预期非空单元格，而非固定 A1。该命令是多步编排（建文档 → 探活 → 定位默认工作表 → 写数据 → 回读 → 可选样式），因此如实声明为独立叶子 `sheet.create_with_data` + `interface_mode: composite`（附评审 reason，按契约不带 `interface_ref`）；**`dws sheet create` 保持原样不变**——仍是一次 `create_workspace_sheet` 直连（`interface_mode: mcp`），不新增 flag，避免让 Schema 消费者把编排步骤的参数误当成该 RPC 的入参。
+- **`dws sheet export-csv`（新命令）** — 同步导出单个工作表为 RFC4180 CSV，支持 `--sheet-id` 选表、`--range` 限定范围、`--value-render-option` 选取值模式；`--output` 落盘（为目录时按 `sheet-export.csv` 命名，落盘走 `AtomicWrite` 原子替换，写入失败时已有文件保持原样；父目录不存在按错误处理，不会自动创建），不传则把纯 CSV 打到 stdout（警告只走 stderr）。数据超出单次读取上限时默认报错、既不输出也不写文件，需 `--allow-truncated` 显式接受不完整结果。响应缺 `csv` 字段或类型不对一律报错，不会用 0 字节覆盖已有文件。该分支读的是 `get_range_as_csv`、与 xlsx 的异步导出任务毫无关系，因此独立成叶子 `sheet.export_csv` 并如实声明 `interface_mode: mcp` + `interface_ref: get_range_as_csv`；**`dws sheet export` 保持原样不变**——仍只导 xlsx（`interface_ref: submit_export_job`），flag 面仍是 `--node` / `--output`，csv 专属 flag 不会出现在它上面（此前挂在同一条命令上时，漏写 `--export-format csv` 会让 `--range` 被静默丢弃而导出整篇工作簿）。
+- **`sheet update-dimension --size-type`** — `pixel` / `standard`（恢复默认行高列宽）/ `auto`（按内容自适应行高，仅 ROWS）。
+- **`sheet replace --match-formula`** — 在公式文本中查找替换。
+- **`sheet range set-style` 扩展样式维度** — 新增 `--font-style`（斜体）/ `--font-line`（下划线、删除线）/ `--font-family` / `--border-styles-json`（四边边框；每条边只接受 `style` / `color`，未知键与非字符串 `color` 直接报错，不再静默忽略而画出无颜色的边框，`set-style` / `batch-set-style` / `create-with-data --styles` 三条路径同源校验）。
+- **`sheet range batch-set-style --ranges`** — 一组样式刷多个带工作表前缀的区域，组装为一次原子 `batch_update`。
+
+### Changed
+
+- **Chat message post-send ID handoff** (#897) — CLI Help and bundled Skills
+  now document the `send` → `query-send-status` → `edit`/`recall` workflow,
+  so callers can reuse returned task, message, and conversation IDs instead
+  of searching message history by content.
+- **Sheet mono/multi Skill alignment** — replaces the oversized mono Sheet
+  reference with the progressive routing layout, aligns all 20 Sheet topic
+  references across the mono and multi bundles, and adds a content-policy guard
+  that prevents the paired topic trees from drifting again.
+- **`sheet range set-style` 后端切换为 `set_cell_range`** — 样式统一走 cellStyles 路径（仅设样式、保留原值），这是斜体/下划线删除线/字体族/边框唯一可用的通道。`interface_ref` 由 `update_range` 变为 `set_cell_range`，12 个样式 flag 改为 reviewed mapping exclusion。CLI 用法向后兼容、无 flag 删除；schema-compatibility 经 reviewed 豁免判定为兼容（0 changed fields）。
+- **`sheet range batch-set-style` 改为单次原子提交** — 由本地循环多次 `update_range` 改为一次 `batch_update`，任一项失败默认整批回滚；`--continue-on-error` 由本地控制改为透传服务端。新增批量上限：最多 100 个区域且累计不超过 200000 个单元格。
+- **`sheet range batch-clear` / `batch-set-style` 的 `--ranges` 拒绝空白工作表前缀**（用户可见行为变更）— 此前只按原始串里 `!` 的位置判断，`" !A1:B2"` 修剪后工作表名成了空串，操作却照样带着 `sheetId: ""` 提交：服务端要么让整批 `batch_update` 失败，要么更糟——落到默认工作表而不是用户指定的那张表，且命令报成功。现在工作表名与范围都必须在修剪之后仍非空，否则在发起任何请求之前报错。`batch-set-style --batch` 的纯空白 `sheetId` / `range` 同样拒绝（此前只挡空字符串）；`--batch` 下发仍用原值不替用户修剪，因为 `sheetId` 可以是允许带首尾空格的工作表**名**。两条 `--ranges` 路径现在共用同一个拆分器。
+- **`sheet insert-dimension` / `delete-dimension` / `update-dimension` 的 `--length` 严格校验**（用户可见行为变更）— 解析由 `fmt.Sscanf("%d")` 改为 `strconv.Atoi`。此前只消费前缀数字，`--length 2x` / `3foo` 会被静默当成 `2` / `3` 并对错误的行列数执行操作（删除方向不可回滚）；现在整个值必须是合法正整数，否则报错「`--length` 必须为正整数（>= 1）」且不发起任何请求。**升级影响**：原先依赖这种宽松解析、在传畸形 `--length` 的脚本会开始报错，请把参数修正为纯数字。合法数字值行为不变，上限仍为 5000。`add-dimension` 的 `--length` 是 `Int` 类型 flag，一直由 cobra 严格校验，不受影响。
+- **CLI 接口兼容门禁支持 reviewed flag 类型豁免**（无用户可见变更）— `authoritative-interface-integrity` 与 `check-command-compatibility.sh` 此前一律拒绝历史命令的 flag 类型变更，即使新类型只是把同一套校验从 RunE 前移到解析期，也没有任何评审通道。现在两道门禁各带一张精确豁免表：命令路径 + flag 名 + 旧类型 → 新类型四元组全等才命中、方向敏感（`string`→`int` 与 `int`→`string` 是两个不同的键，只有被评审的方向可用），且仅当该 flag 的其他契约（shorthand / required / hidden / no-opt / scope）纹丝不动时才放行，因此豁免夹带不了别的破坏。首条也是目前唯一一条登记的是 `dws minutes permission apply --policy` 的 `string` → `int`（配合 #912）：旧实现在 RunE 里做 `strconv.ParseInt(v, 10, 64)` 再校验 `[2,4]`，新实现由 pflag 以 `strconv.ParseInt(s, 0, 64)` 解析后仍校验 `[2,4]`，**历史上能成功的调用集是新调用集的子集**（base 0 额外接受 `0x3` 这类写法，只放宽不收紧），非法值依然失败、只是报错文案与时机前移；flag 默认值由 `""` 变 `"0"` 是类型的必然结果，两道门禁都不比较默认值，且该 flag 必须显式给出、默认值不可达。两张表必须逐字一致并有守卫测试锚定漂移——重复是被迫的而非选择：`check-authoritative-interface-baselines.sh` 会把整个 `scripts/policy/interface-baseline` 目录复制进检出历史版本的 worktree 再编译，那份拷贝不能 import 本分支新增的包。
+- **Schema 兼容门禁支持 reviewed 参数类型豁免**（无用户可见变更）— 接上一条。`schema-compatibility` 是同一个 `Interface Integrity` job 里排在两道 CLI 接口门禁之后的第三道检查，此前也一律拒绝已发布参数的 `type` 变更。由于前两道先失败、`set -e` 让它从未在 CI 上暴露，上一条豁免只解决了三分之二。现在 `checkParameterCompatibility` 也带一张精确豁免表：`<product>/<tool id>` + 参数名 + 旧类型 + 新类型四元组全等才命中、方向敏感，且仅当该参数**除 `type` 外的全部已发布字段逐字段相等**时才放行。这里刻意用相等性比较而非「没有产生其他兼容性错误」：放宽 `required` / `cli_required`、清空 `required_when`、扩宽 `enum`、清空 `interface_type`、经 reviewed mapping exclusion 清空 `property`——这些变化单独看都是兼容的、根本不产生错误，若以错误列表代替相等性检查，它们就能搭着一次已评审的类型迁移一起蒙混过关。结构体整体比较还意味着将来给 `parameterSchema` 新增字段时会自动纳入守卫，而不是悄悄放宽每一条既有条目。唯一条目是 `minutes/minutes.apply_minutes_permission` 的 `policy` 由 `"string"` 迁移到 `"integer"`（配合 #912）：该 `type` 由 Cobra flag 类型投影而来（provenance `cobra_flag_type`），描述的是 CLI 如何接受取值；消费方据此拼装的是命令行，而 `--policy 4` 在两种声明下是同一个 argv，加引号的 `--policy "4"` 到 pflag 仍是 4，RunE 也仍校验 `[2,4]`——而且该参数映射的 property `policyId` 一直以数字上报，新声明比旧声明更贴近真实请求。表里的类型值必须是 `schemaType` 实际产出的带引号形态（`"string"` 而非裸 `string`），守卫测试用 `schemaType` 复算并校验类型名属于 JSON Schema 的封闭取值集合——`reviewedInterfaceRefRedirect` 曾因键的书写形态错误两次静默失效，这里不重犯。
+
+### Fixed
+
+- **Event runtime-token handoff** — personal `event consume`, `status`, `stop`, and `+listen-im` now honor the existing root `--token` instead of falling back to a stale local OAuth profile. Detached personal-event buses negotiate the credential only after an additive capability handshake, receive and rotate it through owner-only local IPC, and keep it in memory; the token is never forwarded through child argv, environment variables, profiles, logs, or run-state files. Existing OAuth and multi-profile behavior is unchanged when `--token` is absent. A new client refuses to send a runtime token to an older bus and leaves its existing consumers and subscriptions untouched; the recovery message asks users to inspect `event status --as user`, preview `event stop --as user --all --dry-run`, and explicitly confirm `event stop --as user --all --yes` before retrying.
+
+## [1.0.58-beta.1] - 2026-08-07
+
+### Added
+
+- **Robot image and file messages** (#867) — `dws chat message send-by-bot`
+  now supports image URLs and local-file uploads through explicit message
+  types, while retaining Markdown as the default and preserving its existing
+  title and text requirements.
+- **Conversation shortcut-bar management** (#877) — adds `dws chat toolbar`
+  commands to list, add, hide, sort, and manage custom conversation shortcuts,
+  with validation and confirmation for destructive removal.
+- **Complete AI Table Shortcut surface** (#901) — makes all 92 supported
+  AI Table Shortcuts discoverable through Runtime Schema and adds reliable
+  Base, table, record, attachment, view, dashboard, and workflow operations
+  with explicit confirmation and result-verification semantics for writes.
+
+### Changed
+
+- **Doc import upload fallback** — `dws doc import` no longer fails on file
+  formats outside the conversion whitelist (html, pdf, zip, extensionless,
+  and any future format): it now hands the file to the document-space upload
+  chain (the same primitive as `dws drive upload --workspace`), stores the
+  original file at the requested `--folder`/`--workspace` target, and prints
+  an explicit stderr notice with the supported-format list and the
+  convert-to-md alternative. The fallback shares the import file checks
+  (20MB cap, empty-file guard), keeps `--format json` / `--dry-run` output as
+  a single JSON document, and marks the machine-readable result with
+  `fallback: "upload"` and `converted: false` so agents never mistake the
+  stored file for a converted online document. The fallback fails closed
+  unless the commit response parses as JSON and carries a file identity
+  (exposed as `dentry_id`); empty or unverifiable responses surface as
+  errors instead of fabricated success. Importable formats and
+  `dws sheet import` validation are unchanged.
+- **IM natural-target and history alignment** — Chat shortcuts can resolve natural user/group targets before execution, and message-history workflows expose bounded time ranges, ordering, explicit all-page controls, continuation ledgers, safe local export, and thread-reply pagination without treating empty or incomplete reads as successful results. Bundled mono/multi Skills and intent routing now describe the same executable surface.
+- **Sheet CSV formula writes** — `dws sheet csv-put` and batch `csv-put` now expose the service contract that CSV fields beginning with `=` are written as formulas. Prefix the field with an apostrophe to write literal text beginning with `=`; CSV content continues to pass through unchanged.
+- **Release-equivalent PR compatibility gate** (#889) — pull-request
+  admission now runs command-surface compatibility checks against the current
+  release baseline before code reaches `main`.
+- **Reviewer routing governance** (#903) — updates the Reviewer Router pool
+  used for new ready PRs while retaining the existing current-head review and
+  required-check gates.
+
+### Fixed
+
+- **Fail-closed IM pagination and audit evidence** — `+chat-messages`, `+search-msg`, `+thread-replies`, `+at-me`, `+my-groups`, conversation lists, and favorites preserve partial-read failures, reject missing or stalled continuation state, deduplicate page boundaries, and publish completion evidence. The live-audit regression suite now rejects empty projections and incomplete reads instead of promoting them to passing results.
+- **Sheet formula verification** (#873) — `dws sheet formula-verify` now calls
+  the registered remote tool name `verify_formula`; the previous
+  `formula_verify` name failed at gateway dispatch.
+- **CLI and parameter recovery boundaries** (#864) — command and parameter
+  recovery now fail closed when an Agent-provided path or flag cannot be
+  reconciled with the executable CLI surface, reducing unsafe hallucinated
+  retries.
+
+## [1.0.57-beta.4] - 2026-08-06
+
+### Added
+
+- **Expanded open CLI workflows** (#887) — adds calendar event-instance
+  queries, Drive latest-file selection, Markdown diff, Mail calendar/export/
+  share-to-chat workflows, and Minutes hot-word, permission, and audio-memo
+  operations, with matching Schema and cross-platform coverage.
+
+### Changed
+
+- **Multi-skill framework alignment** (#887) — folds long-tail skills into
+  `dingtalk-misc`, renames the shared package to `dingtalk-shared`, removes
+  stale Preview guidance, and reorganizes shared recipes and routing for more
+  predictable Agent selection.
+- **Bounded Agent Schema delivery** (#887) — keeps compact and wire projections
+  focused on executable contract facts, retires stale MCP metadata candidates,
+  and teaches Agents to prefer `dws schema --compact` for bounded context.
+
+### Deprecated
+
+- **Recovery and discovery-cache compatibility surfaces** (#887) — keeps
+  visible Deprecated `dws recovery` and `dws cache` compatibility stubs while
+  retiring their former recovery engine and dynamic discovery-cache behavior.
+  Recovery plan/execute/finalize now return an explicit “不再支持” notice, and
+  Skills no longer teach either retired workflow.
+
+### Fixed
+
+- **Mail share-to-chat confirmation** (#887) — requires explicit confirmation
+  before the first remote write, while preserving the confirmed sign-retry
+  flow and covering both direct-success and retry responses.
+
+## [1.0.57] - 2026-08-06
+
+This stable release promotes the fully delivered `v1.0.57-beta.4` baseline.
+It includes the v1.0.57 beta-line command-contract, document, chat, OA, Wiki,
+and compatibility improvements, plus the multi-skill framework alignment and
+expanded calendar, Drive, Markdown, Mail, and Minutes workflows validated in
+the final prerelease.
+
+- **Promote v1.0.57-beta.4** — publishes the final validated prerelease
+  baseline as stable `v1.0.57` without adding post-beta product changes.
+
+## [1.0.57-beta.3] - 2026-08-06
+
+### Added
+
+- **Reviewed document shortcuts** (#880) — adds public document shortcuts for
+  safe local downloads, content and history, review, media and style, and
+  document access/sharing workflows, while retaining reviewed compatibility
+  identities and confirmation safeguards for writes.
+- **Mentions in chat replies** (#881) — `chat message reply` now supports
+  `--at-open-dingtalk-ids` and `--at-all`, forwarding reply mention fields and
+  adding any required mention placeholders without changing existing send
+  behavior.
+
+## [1.0.57-beta.2] - 2026-08-05
+
+### Fixed
+
+- **Stable Chat command compatibility** (#876) — restores the hidden migration
+  entries for `chat send`, `chat history`, and their `im` aliases, preserving
+  the v1.0.56 command surface while directing callers to the supported
+  `chat message send/list` commands. Legacy flags now reach the same migration
+  hints instead of failing during flag parsing.
+- **Drive download cancellation-test stability** (#876) — replaces a
+  timing-sensitive worker-cancellation coverage test with a deterministic seam,
+  reducing flaky CI without changing download behavior.
+
+## [1.0.57-beta.1] - 2026-08-05
+
+This beta starts the v1.0.57 line on top of v1.0.56. It packages the unified
+command-contract and runtime Schema architecture, complete Multi IM Chat
+coverage, document whiteboard and OA approval workflows, Wiki activity feeds,
+and compatibility and CI reliability fixes.
+
+### Added
+
+- **Contact personal-status updates** (#872) — adds `contact user update-ownness`
+  (alias `set-ownness`) for updating a user's personal status text. The write
+  operation maps reviewed `userId` and `ownnessText` parameters to the service
+  contract and requires confirmation unless `--yes` is explicitly supplied.
+- **Document whiteboard workflows** (#861) — adds `doc whiteboard insert`,
+  `whiteboard query/update`, and `doc media upload`. These commands support
+  confirmed document-embedded whiteboard creation and updates, structured
+  OpenNodes reads, and preparation of node-bound Vector/SVG resources.
+- **Complete Multi IM Chat coverage** (#860) — hardens deterministic group and
+  stable-ID resolution, sending, querying, downloading, pagination, and JSON
+  export. The remaining reviewed Chat Shortcuts enter Schema coverage, with
+  destructive delete and clear operations aligned to confirmation gates.
+- **OA approval form workflows** (#853) — adds OA form-schema lookup,
+  process forecast, and confirmed approval-instance creation, supporting both
+  simple flags and complete `--request` payloads.
+- **Wiki activity-feed queries** (#862) — adds `wiki feed list` to retrieve
+  workspace document activity, with cursor paging and optional file exclusion.
+
+### Changed
+
+- **Unified command and Schema contract framework** (#830) — Leaf commands and
+  Shortcuts now use the shared typed `corecmd` base for flags, constraints,
+  confirmation, Help, and runtime Schema projection. Schema delivery assembles
+  from leaf Contract declarations at runtime; the retired hint overlays,
+  pinned MCP metadata, and committed Catalog artifacts are no longer delivery
+  authorities.
+- **Faster macOS CI without reducing native coverage** (#857) — narrows the
+  macOS race suite to Keychain, codesign, and Darwin-only tests while adding a
+  reachability contract that prevents native-only tests from being silently
+  excluded.
+
+### Fixed
+
+- **Chat media-download JSON compatibility** (#854) — restores parseable
+  `success`, `downloadUrl`, and `output` fields for
+  `chat message download-media --format json` after a successful download,
+  without progress output corrupting JSON stdout.
+
+### Added
+
+- **Document-embedded whiteboard workflows** — adds `doc whiteboard insert` for confirmed creation and part-ID verification, `whiteboard query/update` for structured OpenNodes reads and confirmed writes, and `doc media upload` for preparing node-bound Vector/SVG resources. The public adapter uses an explicit helper-only whiteboard endpoint, validates update envelopes locally, decodes `resultJson`, and publishes the full command, Schema, Skill, and safety contract migrated from `dws-wukong@e2da8ab947c6`.
+- **Robot image and file messages** — extends `chat message send-by-bot` with public image URL delivery through `--msg-type image --image-url` and local-file upload/send through `--msg-type file --file-path`, while preserving Markdown as the default message type.
+
+### Changed
+
+- **Chat reply mentions** — `dws chat message reply` can @ specified group members with `--at-open-dingtalk-ids` or @ everyone with `--at-all`, forwarding the existing `send_personal_message` mention fields and automatically adding missing current-user `<@id>` / `<@all>` placeholders.
 - **Pinned MCP metadata retired** — deletes `internal/cli/schema_mcp_metadata.json` and removes its embed/loader/fallback role from Schema assembly. Catalog now assembles from Contract/ParamDecl/Interface + Cobra only; `make fetch-mcp-metadata` remains an optional diagnostic dump under `artifacts/` and refuses the retired pin path. Policy bans the pin from reappearing.
 - **MCP service review retired** — deletes `schema_mcp_service_review.json` and removes its policy jq / outputguard / test disposition gate (`notify` → `out_of_surface`, snapshot hash pin). No replacement ledger.
 - **Hints retired; ContractDecl is the leaf Schema source** (#830) — `schema_hints/`, Manual/Schema hint overlays, and `schema_agent_metadata/` delivery are removed. Selection, safety, parameters, and interface facts declare on ProductDecl / leaf `Contract` (`corecmd.ContractDecl` + `contract.ParamDecl` / `Safety`). Authoring renamed `SchemaDecl` → `ContractDecl`; nested fields reuse `contract.*` directly.
@@ -16,6 +443,7 @@ The format is inspired by [Keep a Changelog](https://keepachangelog.com/) and th

 ### Fixed

+- **Fail-closed IM pagination and audit evidence** — `+chat-messages`, `+search-msg`, `+thread-replies`, `+at-me`, `+my-groups`, conversation lists, and favorites preserve partial-read failures, reject missing or stalled continuation state, deduplicate page boundaries, and publish completion evidence. The live-audit regression suite now rejects empty projections and incomplete reads instead of promoting them to passing results.
 - **Unified command safety and Shortcut runtime (H0)** — Shortcut leaves now execute through `corecmd.New`, sharing the same typed Safety confirmation gate as Leaf commands. EOF / closed stdin returns `confirmation_required`, and interactive `no` returns the existing non-zero cancellation validation error instead of reporting success for an operation that did not run. Pass `--yes` or `--dry-run` to skip the prompt.
 - **Constraint "provided" for `at_least_one` / `exactly_one` (H0)** — a flag set to an empty string (`--flag ""`) no longer counts as provided; previously bare Cobra `Changed` satisfied the constraint. Pass a non-blank value for a member of the group.
 - **Chat media download JSON compatibility** — `dws chat message download-media --format json` once again returns a clean `{success, downloadUrl, output}` result after the file is saved, preserving the temporary URL and resolved local path without progress text corrupting JSON stdout.
@@ -68,14 +68,25 @@ coverage is additionally selected for platform-sensitive code.
 3. Include both the commands/results and user-visible or contract-level
   behavior evidence in the PR description.
 4. Run `./scripts/policy/check-command-surface.sh --strict` when command
-   paths/flags change. CI also runs
-   `./scripts/policy/check-command-compatibility.sh --base-ref <main-ref> --stable-ref <latest-GA-tag>`
-   against both the target branch and latest stable release.
+   paths/flags change. CI resolves the exact merge-base, latest reachable
+   non-withdrawn stable GA tag, and committed candidate SHA, then enters the single compatibility
+   decision seam through
+   `make authoritative-interface-integrity BASE_REF=<merge-base> STABLE_REF=<latest-GA-tag> CANDIDATE_REF=<candidate-sha>`.
+   The Make target delegates to the authoritative wrapper; CI does not invoke a
+   second comparator or the legacy fixture checker. See
+   [CLI flag compatibility migration governance](docs/cli-interface-flag-migrations.md)
+   for the reviewed two-stage `pending` → `consumed` lifecycle.
+   Agent-visible flag migrations must also run
+   `make schema-compatibility BASE_REF=<merge-base> STABLE_REF=<latest-GA-tag> CANDIDATE_REF=<candidate-sha>`;
+   it consumes the same base-owned ledger rather than a second exception list.
 5. Run `./scripts/policy/check-generated-drift.sh` when generated artifacts may
   change.
 6. Run `./scripts/release/verify-package-managers.sh` when packaging or
   installer surfaces change (run `make package` first).
-7. Update docs and `CHANGELOG.md` for behavior/interface changes.
+7. Update docs and add one `.changes/<unique-name>.md` release fragment for
+   behavior/interface changes. Do not edit `CHANGELOG.md` in an ordinary PR;
+   the release-seal workflow renders and archives fragments into the versioned
+   changelog section.

 ## Submission Flow

@@ -1,33 +1,33 @@
 class DingtalkWorkspaceCliBeta < Formula
  desc "Automate DingTalk workspace tasks from the terminal (beta channel)"
  homepage "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli"
-  version "1.0.56-beta.4"
+  version "1.0.59-beta.1"
  license "Apache-2.0"
  keg_only "it is the beta channel and conflicts with dingtalk-workspace-cli"

  on_macos do
    if Hardware::CPU.arm?
-      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.56-beta.4/dws-darwin-arm64.tar.gz"
-      sha256 "f1f9b6394137edbd0b08d632aab34e92a0f3f81d80107a47de1bec9b384f0515"
+      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.59-beta.1/dws-darwin-arm64.tar.gz"
+      sha256 "36a30f3496e0f759c15c0b09f67dbd23b8ecdfff2eebe572f88125b26485830f"
    else
-      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.56-beta.4/dws-darwin-amd64.tar.gz"
-      sha256 "cd3c64d20723c420e2490405d0bf8eecfd7e2b8fc352f63f23de5847a1d38f55"
+      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.59-beta.1/dws-darwin-amd64.tar.gz"
+      sha256 "e7a04906380efd8da88cd112e6a512bb6470a3956dc370150037ed6e314db445"
    end
  end

  on_linux do
    if Hardware::CPU.arm?
-      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.56-beta.4/dws-linux-arm64.tar.gz"
-      sha256 "910918d88074534e680a2e320d3cb364ad092e96b9c422f9e75d11c9c0815dd8"
+      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.59-beta.1/dws-linux-arm64.tar.gz"
+      sha256 "f59ab055f3e841e4cebc964ae3ef969668475548abaaf8bede44afdca9a3e28d"
    else
-      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.56-beta.4/dws-linux-amd64.tar.gz"
-      sha256 "172fe0d84443be953d0c6f2c2433540e4b972fbe7776cff1417ec9c73723552b"
+      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.59-beta.1/dws-linux-amd64.tar.gz"
+      sha256 "2c8f919489d958c7d49262615e81faac70a9fbcae2d589ab54a0bb3c5700a057"
    end
  end

  resource "skills" do
-    url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.56-beta.4/dws-skills.zip"
-    sha256 "a3457befe858cbf3fe85848428b630bfd3a5f626256ed6b49415267948915152"
+    url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.59-beta.1/dws-skills.zip"
+    sha256 "25f4a7e1d01fa4d771d79201b34b11ee8a24182bdcdc94bfb98d2bd5845bed3b"
  end

  def install
@@ -1,33 +1,33 @@
 class DingtalkWorkspaceCli < Formula
  desc "Automate DingTalk workspace tasks from the terminal"
  homepage "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli"
-  version "1.0.56"
+  version "1.0.58"
  license "Apache-2.0"


  on_macos do
    if Hardware::CPU.arm?
-      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.56/dws-darwin-arm64.tar.gz"
-      sha256 "5c6003fe484aa36cc00820a574186652467b9d075f19c159cf807e57590256ba"
+      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.58/dws-darwin-arm64.tar.gz"
+      sha256 "7d98599f90cae9d42b51ff2863efc87dbfb4a3176ff3c84fc2216110c0157a70"
    else
-      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.56/dws-darwin-amd64.tar.gz"
-      sha256 "969b005a10682c2a1a828fa112165b5b0cd8ceeed8d22110ef7f39402cc36804"
+      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.58/dws-darwin-amd64.tar.gz"
+      sha256 "4c12e35e5bf7e0905812cd42dc94a5345068a2c16e306bb50b13c5c78b5cb95d"
    end
  end

  on_linux do
    if Hardware::CPU.arm?
-      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.56/dws-linux-arm64.tar.gz"
-      sha256 "530c5ea7ddc7de320d9c2471fbd33752a723d00c9665f49321c7580e8392c756"
+      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.58/dws-linux-arm64.tar.gz"
+      sha256 "5ef6bde24bc3db6a11a0f1d0b3343a048956b2cbcf6cd3409a037fb6ba425489"
    else
-      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.56/dws-linux-amd64.tar.gz"
-      sha256 "675fa42727ac9a549c6710b82e1980cd0f795363d71d5116a4e69771b7c5470e"
+      url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.58/dws-linux-amd64.tar.gz"
+      sha256 "3ccadcc6f070a39d2b2ba20429a4fcdc2f21639bf79f34361dc7d16f501bfda6"
    end
  end

  resource "skills" do
-    url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.56/dws-skills.zip"
-    sha256 "3d57794e4660a089209ce3962571d16ca0d46141e973c9993257a301cce0e097"
+    url "https://github.com/DingTalk-Real-AI/dingtalk-workspace-cli/releases/download/v1.0.58/dws-skills.zip"
+    sha256 "2626debc21c3daadfd155b4c167b2219b97e801398fe4441a8b48138960ab264"
  end

  def install
@@ -10,7 +10,7 @@ SCHEMA_META_INDEX_OUTPUT ?= artifacts/schema_meta_index.gob
 POLICY_ENV = DWS_POLICY_TMPDIR="$(DWS_POLICY_TMPDIR)" GOTMPDIR="$(POLICY_GOTMPDIR)"
 GO_SOURCE_LIST = git ls-files -z --cached --others --exclude-standard -- '*.go'

-.PHONY: all help build rebuild test test-plan test-auth-legacy-compat lint format-check fmt policy edition-test interface-integrity authoritative-interface-integrity coverage-gate coverage-gate-platform update-interface-baseline reset-interface-baseline schema-compatibility skill-command-integrity skill-context-budget cli-smoke mock-mcp-smoke test-schema-agent-examples generate-schema fetch-mcp-metadata generate-schema-catalog package release release-pre release-stable changelog-pre changelog-stable publish-homebrew-formula setup-hooks
+.PHONY: all help build rebuild test test-plan test-auth-legacy-compat lint format-check fmt policy edition-test interface-integrity authoritative-interface-integrity coverage-gate coverage-gate-platform update-interface-baseline reset-interface-baseline schema-compatibility skill-command-integrity skill-context-budget multi-im-skill-chain-integrity cli-smoke mock-mcp-smoke test-schema-agent-examples generate-schema fetch-mcp-metadata generate-schema-catalog package release release-pre release-stable changelog-pre changelog-stable publish-homebrew-formula setup-hooks

 all: setup-hooks fmt lint build test rebuild

@@ -18,21 +18,22 @@ help:
 	@printf "Available targets:\n"
 	@printf "  make build         - Build the dws CLI binary\n"
 	@printf "  make test          - Run the Go test suite\n"
-	@printf "  make test-plan     - Verify every default Go package belongs to one CI test shard\n"
+	@printf "  make test-plan     - Verify CI test and full-suite coverage package plans cover their scopes exactly once\n"
 	@printf "  make test-auth-legacy-compat - Run stable legacy authentication compatibility regressions\n"
 	@printf "  make lint          - Run formatting checks, go vet, and staticcheck\n"
 	@printf "  make format-check  - Check all repository Go source files with gofmt\n"
 	@printf "  make fmt           - Format all repository Go source files\n"
 	@printf "  make policy        - Check the built dws plus open-source and Schema policies\n"
-	@printf "  make interface-integrity - Check historical commands and help contracts still work\n"
-	@printf "  make authoritative-interface-integrity BASE_REF=<ref> - Check the Git-owned PR merge-base\n"
+	@printf "  make interface-integrity [BASE_REF=<ref>] [STABLE_REF=<tag>] [CANDIDATE_REF=<ref>] - Check authoritative CLI history\n"
+	@printf "  make authoritative-interface-integrity BASE_REF=<ref> [STABLE_REF=<tag>] [CANDIDATE_REF=<ref>] - Check Git-owned CLI history\n"
 	@printf "  make coverage-gate BASE_REF=<ref> - Enforce overall non-regression and 100%% changed-code coverage\n"
 	@printf "  make coverage-gate-platform BASE_REF=<ref> PROFILE=<file> - Enforce 100%% native changed-code coverage\n"
-	@printf "  make update-interface-baseline - Add new CLI contracts without removing history\n"
-	@printf "  make reset-interface-baseline - DANGEROUS: replace all CLI compatibility history\n"
-	@printf "  make schema-compatibility BASE_REF=<ref> - Check the complete Schema contract against the PR merge-base\n"
+	@printf "  make update-interface-baseline - Update the non-authoritative CLI smoke fixture\n"
+	@printf "  make reset-interface-baseline - DANGEROUS: replace the non-authoritative CLI smoke fixture\n"
+	@printf "  make schema-compatibility BASE_REF=<ref> [STABLE_REF=<tag>] [CANDIDATE_REF=<ref>] - Check the authoritative Schema history\n"
 	@printf "  make skill-command-integrity - Check dws commands referenced by skills exist\n"
 	@printf "  make skill-context-budget - Check generated Skill drift and common-path context budgets\n"
+	@printf "  make multi-im-skill-chain-integrity - Check reviewed IM intents keep one default Skill route\n"
 	@printf "  make cli-smoke     - Verify help for every public top-level command\n"
 	@printf "  make mock-mcp-smoke - Verify HTTP and stdio MCP request/response transport\n"
 	@printf "  make test-schema-agent-examples - Contract-check all Agent examples and dry-run the eligible subset\n"
@@ -87,6 +88,8 @@ policy: test-auth-legacy-compat
 	@mkdir -p "$(POLICY_GOTMPDIR)"
 	@$(POLICY_ENV) ./scripts/policy/check-open-source-assets.sh
 	@$(POLICY_ENV) ./scripts/policy/check-skill-context-budget.sh
+	@$(POLICY_ENV) ./scripts/policy/check-multi-im-skill-chain.sh
+	@python3 scripts/run_chat_shortcut_live_audit_test.py
 	@$(POLICY_ENV) ./scripts/policy/check-command-surface.sh --strict
 	@$(POLICY_ENV) ./scripts/policy/check-generated-drift.sh
 	@$(POLICY_ENV) ./scripts/policy/check-param-concepts.sh
@@ -100,10 +103,22 @@ edition-test:
 	$(GO) test -v -count=1 ./pkg/editiontest/...

 interface-integrity:
-	@./scripts/policy/check-interface-baseline.sh
+	@base_ref="$(BASE_REF)"; \
+	candidate_ref="$(CANDIDATE_REF)"; \
+	if [ -z "$$base_ref" ]; then base_ref="origin/main"; fi; \
+	if [ -z "$$candidate_ref" ]; then candidate_ref="HEAD"; fi; \
+	./scripts/policy/check-authoritative-interface-baselines.sh \
+		--base-ref "$$base_ref" \
+		--stable-ref "$(STABLE_REF)" \
+		--candidate-ref "$$candidate_ref"

 authoritative-interface-integrity:
-	@./scripts/policy/check-authoritative-interface-baselines.sh --base-ref "$(BASE_REF)"
+	@candidate_ref="$(CANDIDATE_REF)"; \
+	if [ -z "$$candidate_ref" ]; then candidate_ref="HEAD"; fi; \
+	./scripts/policy/check-authoritative-interface-baselines.sh \
+		--base-ref "$(BASE_REF)" \
+		--stable-ref "$(STABLE_REF)" \
+		--candidate-ref "$$candidate_ref"

 coverage-gate:
 	@./scripts/policy/check-coverage-gate.sh --base-ref "$(BASE_REF)" --scope-buildable
@@ -118,7 +133,12 @@ reset-interface-baseline:
 	@./scripts/policy/check-interface-baseline.sh --reset

 schema-compatibility:
-	@./scripts/policy/check-authoritative-schema-compatibility.sh --base-ref "$(BASE_REF)"
+	@candidate_ref="$(CANDIDATE_REF)"; \
+	if [ -z "$$candidate_ref" ]; then candidate_ref="HEAD"; fi; \
+	./scripts/policy/check-authoritative-schema-compatibility.sh \
+		--base-ref "$(BASE_REF)" \
+		--stable-ref "$(STABLE_REF)" \
+		--candidate-ref "$$candidate_ref"

 skill-command-integrity:
 	@./scripts/policy/check-skill-commands.sh
@@ -126,6 +146,12 @@ skill-command-integrity:
 skill-context-budget:
 	@./scripts/policy/check-skill-context-budget.sh

+multi-im-skill-chain-integrity:
+	@./scripts/policy/check-multi-im-skill-chain.sh
+
+skill-mono-multi-content:
+	@./scripts/policy/check-mono-multi-skill-content.sh
+
 cli-smoke:
 	@./scripts/policy/check-cli-smoke.sh

@@ -143,9 +169,13 @@ generate-schema:
 	@set -e; \
 	concepts_guard=$$(mktemp); \
 	concepts_schema_guard=$$(mktemp); \
-	trap 'rm -rf "$$concepts_guard" "$$concepts_schema_guard"' EXIT HUP INT TERM; \
+	command_fallbacks_guard=$$(mktemp); \
+	command_fallbacks_schema_guard=$$(mktemp); \
+	trap 'rm -f "$$concepts_guard" "$$concepts_schema_guard" "$$command_fallbacks_guard" "$$command_fallbacks_schema_guard"' EXIT HUP INT TERM; \
 	cp internal/cli/param_concepts.json "$$concepts_guard"; \
 	cp internal/cli/param_concepts.schema.json "$$concepts_schema_guard"; \
+	cp internal/cli/command_path_fallbacks.json "$$command_fallbacks_guard"; \
+	cp internal/cli/command_path_fallbacks.schema.json "$$command_fallbacks_schema_guard"; \
 	$(GO) generate ./internal/cli; \
 	rm -rf internal/cli/schema_agent_metadata internal/cli/schema_agent_metadata_audit.json; \
 	rm -f internal/cli/schema_meta_index.json; \
@@ -161,6 +191,14 @@ generate-schema:
 		printf '%s\n' 'generation modified reviewed input internal/cli/param_concepts.schema.json' >&2; \
 		exit 1; \
 	}; \
+	cmp -s internal/cli/command_path_fallbacks.json "$$command_fallbacks_guard" || { \
+		printf '%s\n' 'generation modified reviewed input internal/cli/command_path_fallbacks.json' >&2; \
+		exit 1; \
+	}; \
+	cmp -s internal/cli/command_path_fallbacks.schema.json "$$command_fallbacks_schema_guard" || { \
+		printf '%s\n' 'generation modified reviewed input internal/cli/command_path_fallbacks.schema.json' >&2; \
+		exit 1; \
+	}; \
 	if [ -e internal/cli/schema_hints ]; then \
 		printf '%s\n' 'retired schema_hints/ must not reappear after generation' >&2; \
 		exit 1; \
@@ -80,7 +80,7 @@ How to pick:
 - **Quick install** (one-liner above): non-interactive, installs `multi`.
 - **TTY install** (download then run): `curl -O .../install.sh && bash install.sh` — prompts `1) multi  2) mono` (default 1).
 - **Override via env**: `DWS_SKILL_MODE=mono curl -fsSL ... | sh`.
- **Switch later**: `dws skill setup --mode mono` (or `--mode multi`) — re-run any time.
+- **Switch later**: `dws skill setup --mode mono` (or `--mode multi`) — review the listed paths and confirm interactively.

 </details>

@@ -210,7 +210,7 @@ The verifier uses isolated directories and does not replace the `dws` on the cur
 The upgrade process follows a two-phase atomic flow to ensure consistency:

 1. **Prepare** — downloads the platform-specific binary and skill packages to a temporary directory, verifies SHA256 checksums, and extracts/validates all files. If any step fails, the upgrade aborts without modifying the existing installation.
-2. **Apply** — only after all preparations succeed, the binary is replaced and skill packages are installed to all detected agent directories (`~/.agents/skills/dws`, `~/.claude/skills/dws`, `~/.cursor/skills/dws`, etc.).
+2. **Apply** — only after all preparations succeed, the binary is replaced and skills are flattened into detected agent-specific roots (for example `~/.codex/skills/dingtalk-chat`). `~/.agents/skills` is used only when no specific Agent is detected; once a specific root is active, older DWS-managed generic copies are backed up and retired so the same Skill is not discovered twice.

 A backup of the current version is automatically created before each upgrade. Use `dws upgrade --rollback` to restore the previous version if needed.

@@ -371,7 +371,7 @@ dws contact user get-self --jq '.result[0].orgEmployeeModel | {name: .orgUserNam
 Use Cobra help and Schema for different parts of the command contract:

 - `dws <path> --help` is the source of truth for whether a command exists and which flags the binary accepts.
- `dws schema "<path>"` is the Agent contract for command selection, parameter mappings and constraints, risk, and confirmation semantics.
+- `dws schema "<path>" --compact` is the normative Agent view for command selection, CLI parameters and constraints, risk, and confirmation; use a full leaf with a narrow `--jq` projection for mapping or provenance audits.
 - If Help and Schema disagree, treat it as contract drift: pass only flags accepted by Cobra and use the more conservative safety semantics.
 - Schema describes commands; it does not read or search DingTalk business data. Execute the real product command after discovery.

@@ -380,14 +380,14 @@ Use Cobra help and Schema for different parts of the command contract:
 dws aitable record query --help

 # Discover within a product, then inspect the selected leaf contract
-dws schema aitable
-dws schema "aitable record query"
+dws schema aitable --compact
+dws schema "aitable record query" --compact

 # Execute the real business query
 dws aitable record query --base-id BASE_ID --table-id TABLE_ID --limit 10
 ```

-`dws schema --all` exports the complete contract for tooling, CI, audits, and compatibility baselines. Agents should prefer product/group discovery followed by a leaf query to avoid loading the full Catalog into context.
+`dws schema --all` exports the complete contract for tooling, CI, audits, and compatibility baselines. Agents should query progressively with `--compact`; its positive field allowlist prevents new full/audit fields from silently expanding Agent context.

 ### Agent Skills

@@ -405,7 +405,7 @@ After installing, AI tools like Claude Code / Cursor can operate DingTalk direct
 curl -fsSL https://raw.githubusercontent.com/DingTalk-Real-AI/dingtalk-workspace-cli/main/scripts/install-skills.sh | sh
 ```

-> `install.sh` installs under `$HOME/.agents/skills/` (global; multi layout is per-product siblings, mono is the `dws/` subdirectory); `install-skills.sh` installs under `./.agents/skills/` (current project).
+> Installers prefer detected agent-specific roots such as `$HOME/.codex/skills/`. They use `.agents/skills/` only as the generic fallback when no specific Agent is detected; multi layout is per-product siblings, while mono uses the `dws/` subdirectory.
 >
 > China users: prefix `DWS_GITEE_REPO` to use the Gitee mirror — see [China mirror](#china-mirror).

@@ -415,22 +415,31 @@ curl -fsSL https://raw.githubusercontent.com/DingTalk-Real-AI/dingtalk-workspace
 # Interactive: prompts for mode + target agents
 dws skill setup

-# Install mono skill to every detected agent home (claude / cursor / codex / opencode / qoder)
-dws skill setup --mode mono --target all --yes
+# Preview the exact directories that mono setup would back up and replace
+dws skill setup --mode mono --target all --dry-run

-# Install multi skills to a single agent home
-dws skill setup --mode multi --target cursor --yes
+# Run interactively and confirm the listed directories
+dws skill setup --mode mono --target all

-# Point at a local source tree (e.g. a fork or work-in-progress)
+# Preview, then install multi skills to a single agent home with interactive confirmation
+dws skill setup --mode multi --target cursor --dry-run
+dws skill setup --mode multi --target cursor
+
+# Point at a local source tree (e.g. a fork or work-in-progress), preview first
+DWS_SKILL_SOURCE=/path/to/skills dws skill setup --mode multi --dry-run
 DWS_SKILL_SOURCE=/path/to/skills dws skill setup --mode multi
 ```

 | Flag | Values | Description |
 |------|--------|-------------|
 | `--mode` | `mono` \| `multi` | Skill layout; defaults to interactive prompt |
-| `--target` | `all` \| `claude` \| `cursor` \| `codex` \| `opencode` \| `qoder` | Where to install; `all` covers every detected agent home |
+| `--target` | `all` \| `claude` \| `cursor` \| `codex` \| `zcode` \| `opencode` \| `qoder` | Where to install; `all` covers every detected agent home, including ZCode at `~/.zcode/skills` |
 | `--source` | path | Local source directory (overrides bundled skills) |
-| `--yes` | — | Skip confirmation prompts |
+| `--yes` | — | Scripting-only: skip the confirmation prompt. Removals are still backed up to `~/.dws/skill-backups/` first |
+
+> The setup command can remove the opposite-mode layout (`dws/` for multi, DWS-managed multi Skills for mono) and stale managed Skills not in the bundle. DWS records ownership, installer version, source, and content digest centrally in `~/.dws/skills-state.json` (or `$DWS_CONFIG_DIR/skills-state.json`). Exact official names shipped before the centralized state remain a frozen migration list. A `dingtalk-*` prefix alone never authorizes cleanup, so other same-prefix market/user Skills are preserved. Every removal is previewed before confirmation and preserved under `~/.dws/skill-backups/<timestamp>/`; a directory that cannot be backed up is never removed. In a non-interactive shell, first run `--dry-run` and inspect its output; only then may the caller explicitly choose the scripting-only confirmation bypass.
+
+After a multi setup or upgrade, DWS stores the official bundle snapshot and centralized ownership metadata in `~/.dws/skills-state.json` (or `$DWS_CONFIG_DIR/skills-state.json`). Every upgrade installs and overwrites the complete bundled Skill set from that release. Deleting or excluding a bundled Skill is not sticky: the next upgrade restores it. `dws upgrade --force` additionally allows reinstalling the current CLI version when no newer version is available.

 Env vars: `DWS_SKILL_MODE=mono|multi` (also honored by `install.sh` / `install.ps1`), `DWS_SKILL_SOURCE=<path>`.

@@ -443,7 +452,6 @@ Env vars: `DWS_SKILL_MODE=mono|multi` (also honored by `install.sh` / `install.p
 | Intent guide | `skills/mono/references/intent-guide.md` | Disambiguation for confusing scenarios (e.g. report vs todo) |
 | Global reference | `skills/mono/references/global-reference.md` | Auth, output formats, global flags |
 | Error codes | `skills/mono/references/error-codes.md` | Error codes + debugging workflows |
-| Recovery guide | `skills/mono/references/recovery-guide.md` | `RECOVERY_EVENT_ID` handling |
 | Ready-made scripts | `skills/mono/scripts/*.py` | 13 batch operation scripts (see below) |

 <details>
@@ -474,7 +482,7 @@ Env vars: `DWS_SKILL_MODE=mono|multi` (also honored by `install.sh` / `install.p
 <details>
 <summary><strong>Personal Event Subscription</strong> — real-time DingTalk messages for event-driven agents</summary>

-`dws event consume` subscribes as the currently logged-in user over a managed Stream WebSocket and emits each event as one NDJSON line on stdout. The public catalog covers scoped and all one-to-one/group messages, specified senders, read/recall/reaction events, and group title/disband lifecycle events.
+`dws event consume` subscribes as the currently logged-in user over a managed Stream WebSocket and emits each event as one NDJSON line on stdout. The public catalog covers scoped and all one-to-one/group messages, specified senders, read/recall/reaction events, group lifecycle events, and six OA approval task/instance events.

 The default `ndjson`, `json`, and `pretty` output preserves the transport envelope (`type`, `event_type`, string `data`, and `headers`) for existing scripts; `compact` retains its existing processor. Add `--flatten` to emit the stable top-level business fields used by Agent workflows. `--format` controls JSON serialization; `--flatten` controls the data structure and cannot be combined with `-f raw` or `--debug-raw-events`.

@@ -484,28 +492,33 @@ For an event-focused installation, use the official convenience installer:

 ```bash
 curl -fsSL https://raw.githubusercontent.com/DingTalk-Real-AI/dingtalk-workspace-cli/main/scripts/install-event.sh | sh
+
+# Or install the standalone multi skill from an existing dws installation
+dws skill setup --mode multi -s event
 ```

 ```bash
 # Inspect the public personal event catalog and schema
 dws event list
 dws event schema user_im_message_receive_o2o --flatten
+dws event list --category oa
+dws event schema user_oa_approval_task_created --flatten

 # Listen for messages that mention the current user
-dws event consume user_im_message_receive_at --flatten -f ndjson
+dws event +listen-im --kind at-me -f ndjson

-# Listen for one-to-one messages with a specified user
-dws event consume user_im_message_receive_o2o --user <userId> --flatten -f ndjson
+# Listen for messages from a specified sender
+dws event +listen-im --kind sender --user <userId> -f ndjson

 # Listen by openDingtalkId (external contact, bot, or cross-organization identity)
-dws event consume user_im_message_receive_o2o --open-dingtalk-id <openDingtalkId> --flatten -f ndjson
+dws event +listen-im --kind sender --open-dingtalk-id <openDingtalkId> -f ndjson

 # Listen for messages in a specified group
-dws event consume user_im_message_receive_group --group <openConversationId> --flatten -f ndjson
+dws event +listen-im --kind group --chat-id <openConversationId> -f ndjson

 # Listen for all one-to-one or all group messages
-dws event consume user_im_message_receive_o2o_all --flatten -f ndjson
-dws event consume user_im_message_receive_group_all --flatten -f ndjson
+dws event +listen-im --kind all-direct -f ndjson
+dws event +listen-im --kind all-group -f ndjson

 # Listen for a specified group's title changes, member changes, or disband event
 dws event consume user_im_group_updated --group <openConversationId> --flatten -f ndjson
@@ -513,14 +526,19 @@ dws event consume user_im_group_member_added --group <openConversationId> --flat
 dws event consume user_im_group_member_exited --group <openConversationId> --flatten -f ndjson
 dws event consume user_im_group_disbanded --group <openConversationId> --flatten -f ndjson

-# Listen for multiple events for the same user in one process
+# Listen for messages, reads, and recalls from the same sender in one process
+dws event +listen-im --kind sender --user <userId> \
+  --events message,read,recall -f ndjson
+
+# Listen for all six public OA approval events in one process
 dws event consume \
-  user_im_message_receive_o2o \
-  user_im_message_read_o2o \
-  user_im_message_recall_o2o \
-  --user <userId> \
-  --flatten \
-  -f ndjson
+  user_oa_approval_task_created \
+  user_oa_approval_task_finished \
+  user_oa_approval_task_redirected \
+  user_oa_approval_instance_started \
+  user_oa_approval_instance_terminated \
+  user_oa_approval_instance_finished \
+  --flatten -f ndjson

 # Inspect local consumers and cancel a subscription
 dws event status
@@ -624,7 +642,7 @@ dws aitable record query --base-id BASE_ID --tabel-id TABLE_ID       # --tabel-i
 ```bash
 # Built-in jq expressions
 dws aitable record query --base-id BASE_ID --table-id TABLE_ID --jq '.invocation.params'
-dws schema "dev app create" --jq '.tool.required'
+dws schema "dev app create" --jq '.parameters'

 # Return only specific fields
 dws aitable record query --base-id BASE_ID --table-id TABLE_ID --fields invocation,response
@@ -636,9 +654,9 @@ dws aitable record query --base-id BASE_ID --table-id TABLE_ID --fields invocati
 <summary><strong>Schema Introspection</strong> — Agent command discovery and execution contracts</summary>

 ```bash
-dws schema aitable                                      # discover product commands
-dws schema "aitable record query"                       # view the selected leaf contract
-dws schema "aitable record query" --jq '.tool.required' # view required fields
+dws schema aitable --compact                            # discover product commands
+dws schema "aitable record query" --compact             # view the selected Agent leaf contract
+dws schema "aitable record query" --jq '[.parameters | to_entries[] | select(.value.required)]' # view required fields
 dws schema --all                                        # full export for CI/audit/baselines
 ```

@@ -719,7 +737,7 @@ See [`docs/robot-quickstart.md`](./docs/robot-quickstart.md) for the full 4-step
 <summary>Coming soon</summary>

 - `conference` (video meetings)
- Multi-skill mode (default) — per-product skills under `skills/multi/`; installs and upgrades default to it, `dws skill setup --mode mono` switches back
+- Multi-skill mode (default) — per-product skills under `skills/multi/`; installs and upgrades default to it, `dws skill setup --mode mono` switches back after interactive confirmation

 </details>

@@ -768,6 +786,7 @@ See [`docs/robot-quickstart.md`](./docs/robot-quickstart.md) for the full 4-step

 ## Reference & Docs

+- [International DingTalk (`.io`) guide](./docs/international-region-guide.md) — international login, domestic/international profile switching, isolated testing, and troubleshooting
 - [Command Index](./docs/command-index.md) — every runtime command with description and when-to-use guidance
 - [Reference](./docs/reference.md) — environment variables, exit codes, output formats, shell completion
 - [Architecture](./docs/architecture.md) — static endpoint pipeline, command surface, transport layer
@@ -80,7 +80,7 @@ irm https://raw.githubusercontent.com/DingTalk-Real-AI/dingtalk-workspace-cli/ma
 - **快速安装**（上方一行 curl）：非交互，默认装 `multi`。
 - **TTY 安装**（先下载再执行）：`curl -O .../install.sh && bash install.sh`，会弹出 `1) multi  2) mono` 选项（默认 1）。
 - **环境变量覆盖**：`DWS_SKILL_MODE=mono curl -fsSL ... | sh`。
- **装完之后再切换**：`dws skill setup --mode mono`（或 `--mode multi`），随时重跑都行。
+- **装完之后再切换**：`dws skill setup --mode mono`（或 `--mode multi`），核对列出的路径后交互确认。

 </details>

@@ -207,7 +207,7 @@ bash verify-all-channels.sh
 升级过程采用两阶段原子流程，确保一致性：

 1. **准备阶段** — 将平台对应的二进制文件和技能包下载到临时目录，校验 SHA256 校验和，解压并验证所有文件。任何步骤失败则立即中止，不会修改现有安装。
-2. **执行阶段** — 仅在所有准备工作成功后，替换二进制文件并将技能包安装到所有已检测到的 Agent 目录（`~/.agents/skills/dws`、`~/.claude/skills/dws`、`~/.cursor/skills/dws` 等）。
+2. **执行阶段** — 仅在所有准备工作成功后，替换二进制文件并将技能包平铺到已检测到的具体 Agent 目录（例如 `~/.codex/skills/dingtalk-chat`、`~/.claude/skills/dingtalk-chat`）。只有未检测到具体 Agent 时才使用 `~/.agents/skills`；检测到具体 Agent 后会备份迁走旧的 DWS 通用副本，避免同一 Skill 被重复发现。

 每次升级前自动备份当前版本，可通过 `dws upgrade --rollback` 随时回滚。

@@ -365,7 +365,7 @@ dws contact user get-self --jq '.result[0].orgEmployeeModel | {name: .orgUserNam
 命令帮助和 Schema 分别负责命令契约的不同部分：

 - `dws <path> --help` 是命令是否存在、当前二进制接受哪些 flags 的事实源。
- `dws schema "<path>"` 是 Agent 选命令、参数映射与约束、风险和确认语义的契约。
+- `dws schema "<path>" --compact` 是 Agent 选命令、CLI 参数与约束、风险和确认语义的规范视图；映射或 provenance 审计使用 full leaf 配合 `--jq` 精确投影。
 - Help 与 Schema 冲突时视为契约漂移：执行只传 Cobra 接受的参数，安全语义取更保守值。
 - Schema 只描述命令，不读取或搜索钉钉业务数据；发现命令后仍需执行真实产品命令。

@@ -374,14 +374,14 @@ dws contact user get-self --jq '.result[0].orgEmployeeModel | {name: .orgUserNam
 dws aitable record query --help

 # 先在产品内发现命令，再查看选中 leaf 的契约
-dws schema aitable
-dws schema "aitable record query"
+dws schema aitable --compact
+dws schema "aitable record query" --compact

 # 执行真实业务查询
 dws aitable record query --base-id BASE_ID --table-id TABLE_ID --limit 10
 ```

-`dws schema --all` 会完整导出命令契约，供工具、CI、审计和兼容性基线使用。Agent 应优先按产品/分组发现后查询 leaf，避免把整个 Catalog 加载进上下文。
+`dws schema --all` 会完整导出命令契约，供工具、CI、审计和兼容性基线使用。Agent 应使用 `--compact` 渐进查询；该视图采用正向字段白名单，full 新增的审计字段不会自动进入 Agent 上下文。

 ### Agent Skills

@@ -399,7 +399,7 @@ Schema 生成的叶子 safety/参数/选型文案由 Go 中的 ProductDecl / Con
 curl -fsSL https://raw.githubusercontent.com/DingTalk-Real-AI/dingtalk-workspace-cli/main/scripts/install-skills.sh | sh
 ```

-> `install.sh` 安装到 `$HOME/.agents/skills/`（全局，multi 为按产品平铺，mono 为 `dws/` 子目录）；`install-skills.sh` 安装到 `./.agents/skills/`（当前项目）。
+> 安装器优先使用检测到的具体 Agent 根目录（如 `$HOME/.codex/skills/`）；仅在未检测到具体 Agent 时回退到 `.agents/skills/`。multi 为按产品平铺，mono 为 `dws/` 子目录。
 >
 > 国内用户加 `DWS_GITEE_REPO` 走 Gitee 镜像，见 [国内加速安装](#国内加速安装)。

@@ -409,22 +409,31 @@ curl -fsSL https://raw.githubusercontent.com/DingTalk-Real-AI/dingtalk-workspace
 # 交互式：提示选模式 + 目标 Agent
 dws skill setup

-# 把 mono skill 铺到所有检测到的 Agent home（claude / cursor / codex / opencode / qoder）
-dws skill setup --mode mono --target all --yes
+# 先预览 mono setup 将备份和替换的精确目录
+dws skill setup --mode mono --target all --dry-run

-# 只装到某一个 Agent home
-dws skill setup --mode multi --target cursor --yes
+# 交互执行并确认列出的目录
+dws skill setup --mode mono --target all

-# 指定本地源目录（比如 fork 或正在改的版本）
+# 先预览，再交互确认装到某一个 Agent home
+dws skill setup --mode multi --target cursor --dry-run
+dws skill setup --mode multi --target cursor
+
+# 指定本地源目录（比如 fork 或正在改的版本），先预览
+DWS_SKILL_SOURCE=/path/to/skills dws skill setup --mode multi --dry-run
 DWS_SKILL_SOURCE=/path/to/skills dws skill setup --mode multi
 ```

 | 参数 | 取值 | 说明 |
 |------|------|------|
 | `--mode` | `mono` \| `multi` | skill 布局，不指定则交互式询问 |
-| `--target` | `all` \| `claude` \| `cursor` \| `codex` \| `opencode` \| `qoder` | 安装目标，`all` 表示铺到所有检测到的 Agent home |
+| `--target` | `all` \| `claude` \| `cursor` \| `codex` \| `zcode` \| `opencode` \| `qoder` | 安装目标；`all` 表示铺到检测到的具体 Agent home（ZCode 为 `~/.zcode/skills`），仅在未检测到具体 Agent 时回退到 `~/.agents/skills` |
 | `--source` | 路径 | 本地源目录（覆盖内置 skills） |
-| `--yes` | — | 跳过确认提示 |
+| `--yes` | — | 仅供脚本使用：跳过确认提示。删除操作仍会先备份到 `~/.dws/skill-backups/` |
+
+> setup 命令可能移除对面模式残留（装 multi 删 `dws/`，装 mono 清理统一状态中登记或属于状态上线前精确官方名称集合的 multi Skill）以及不在 bundle 内的过期受管 Skill。DWS 在 `~/.dws/skills-state.json`（或 `$DWS_CONFIG_DIR/skills-state.json`）集中记录所有权、安装版本、来源和内容摘要。仅有 `dingtalk-*` 前缀不能触发清理，因此其他同前缀市场/用户 Skill 会保留。所有删除都会先列入确认预览，并备份到 `~/.dws/skill-backups/<时间戳>/`；备份失败的目录会保留原样、绝不删除。非交互环境应先用 `--dry-run` 核对输出，再由调用方显式决定是否使用仅供脚本的确认跳过参数。
+
+multi setup 或 upgrade 后，DWS 会把官方 bundle 快照和统一所有权元数据写入 `~/.dws/skills-state.json`（或 `$DWS_CONFIG_DIR/skills-state.json`）。每次 upgrade 都会安装并覆盖该版本的全部预制 Skill；手工删除或通过 setup 排除预制 Skill 不会永久保留，下次 upgrade 会恢复。`dws upgrade --force` 还允许在没有新版本时重装当前 CLI 版本。

 环境变量：`DWS_SKILL_MODE=mono|multi`（`install.sh` / `install.ps1` 也认）、`DWS_SKILL_SOURCE=<路径>`。

@@ -437,7 +446,6 @@ DWS_SKILL_SOURCE=/path/to/skills dws skill setup --mode multi
 | 意图指南 | `skills/mono/references/intent-guide.md` | 易混淆场景消歧（如 report vs todo） |
 | 全局参考 | `skills/mono/references/global-reference.md` | 认证、输出格式、全局 flag |
 | 错误码 | `skills/mono/references/error-codes.md` | 错误码 + 调试流程 |
-| Recovery 指南 | `skills/mono/references/recovery-guide.md` | `RECOVERY_EVENT_ID` 处理 |
 | 现成脚本 | `skills/mono/scripts/*.py` | 13 个批量操作脚本（见下方） |

 <details>
@@ -468,7 +476,7 @@ DWS_SKILL_SOURCE=/path/to/skills dws skill setup --mode multi
 <details>
 <summary><strong>个人事件订阅</strong> — 实时接收钉钉消息，驱动事件触发的 Agent</summary>

-`dws event consume` 使用当前 OAuth 登录用户建立托管的 Stream WebSocket 长连接，并把每条事件以 NDJSON 一行输出到 stdout。当前公开目录覆盖指定范围和全量单聊/群消息、指定发送人、已读/撤回/表情回应，以及群标题变更和群解散事件。
+`dws event consume` 使用当前 OAuth 登录用户建立托管的 Stream WebSocket 长连接，并把每条事件以 NDJSON 一行输出到 stdout。当前公开目录覆盖指定范围和全量单聊/群消息、指定发送人、已读/撤回/表情回应、群生命周期，以及六个 OA 审批任务/实例事件。

 默认 `ndjson`、`json`、`pretty` 输出保留兼容 transport envelope（`type`、`event_type`、字符串 `data`、`headers`），`compact` 继续沿用原 processor。Agent 或新脚本显式加 `--flatten` 后，输出稳定的顶层业务字段。`--format` 控制 JSON 序列化，`--flatten` 控制数据结构，且不能与 `-f raw` 或 `--debug-raw-events` 同时使用。

@@ -478,28 +486,33 @@ DWS_SKILL_SOURCE=/path/to/skills dws skill setup --mode multi

 ```bash
 curl -fsSL https://raw.githubusercontent.com/DingTalk-Real-AI/dingtalk-workspace-cli/main/scripts/install-event.sh | sh
+
+# 或在已有 dws 环境中安装独立的 multi skill
+dws skill setup --mode multi -s event
 ```

 ```bash
 # 查看公开个人事件目录和 schema
 dws event list
 dws event schema user_im_message_receive_o2o --flatten
+dws event list --category oa
+dws event schema user_oa_approval_task_created --flatten

 # 监听当前用户被 @ 的消息
-dws event consume user_im_message_receive_at --flatten -f ndjson
+dws event +listen-im --kind at-me -f ndjson

-# 监听与指定用户的单聊消息
-dws event consume user_im_message_receive_o2o --user <userId> --flatten -f ndjson
+# 监听指定发送人的消息
+dws event +listen-im --kind sender --user <userId> -f ndjson

 # 使用 openDingtalkId 监听外部联系人、机器人或跨组织身份
-dws event consume user_im_message_receive_o2o --open-dingtalk-id <openDingtalkId> --flatten -f ndjson
+dws event +listen-im --kind sender --open-dingtalk-id <openDingtalkId> -f ndjson

 # 监听指定群的消息
-dws event consume user_im_message_receive_group --group <openConversationId> --flatten -f ndjson
+dws event +listen-im --kind group --chat-id <openConversationId> -f ndjson

 # 监听所有单聊或所有群消息
-dws event consume user_im_message_receive_o2o_all --flatten -f ndjson
-dws event consume user_im_message_receive_group_all --flatten -f ndjson
+dws event +listen-im --kind all-direct -f ndjson
+dws event +listen-im --kind all-group -f ndjson

 # 监听指定群标题变更、成员进退群或群解散
 dws event consume user_im_group_updated --group <openConversationId> --flatten -f ndjson
@@ -507,14 +520,19 @@ dws event consume user_im_group_member_added --group <openConversationId> --flat
 dws event consume user_im_group_member_exited --group <openConversationId> --flatten -f ndjson
 dws event consume user_im_group_disbanded --group <openConversationId> --flatten -f ndjson

-# 一个进程监听同一用户的多个事件
+# 一个进程监听同一发送人的消息、已读和撤回
+dws event +listen-im --kind sender --user <userId> \
+  --events message,read,recall -f ndjson
+
+# 一个进程监听全部六个公开 OA 审批事件
 dws event consume \
-  user_im_message_receive_o2o \
-  user_im_message_read_o2o \
-  user_im_message_recall_o2o \
-  --user <userId> \
-  --flatten \
-  -f ndjson
+  user_oa_approval_task_created \
+  user_oa_approval_task_finished \
+  user_oa_approval_task_redirected \
+  user_oa_approval_instance_started \
+  user_oa_approval_instance_terminated \
+  user_oa_approval_instance_finished \
+  --flatten -f ndjson

 # 查看本地 consume，并取消指定订阅
 dws event status
@@ -618,7 +636,7 @@ dws aitable record query --base-id BASE_ID --tabel-id TABLE_ID       # --tabel-i
 ```bash
 # 内置 jq 表达式
 dws aitable record query --base-id BASE_ID --table-id TABLE_ID --jq '.invocation.params'
-dws schema "dev app create" --jq '.tool.required'
+dws schema "dev app create" --jq '.parameters'

 # 只返回指定字段
 dws aitable record query --base-id BASE_ID --table-id TABLE_ID --fields invocation,response
@@ -630,9 +648,9 @@ dws aitable record query --base-id BASE_ID --table-id TABLE_ID --fields invocati
 <summary><strong>Schema 自省</strong> — Agent 命令发现与执行契约</summary>

 ```bash
-dws schema aitable                                      # 发现产品命令
-dws schema "aitable record query"                       # 查看选中 leaf 契约
-dws schema "aitable record query" --jq '.tool.required' # 查看必填字段
+dws schema aitable --compact                            # 发现产品命令
+dws schema "aitable record query" --compact             # 查看 Agent leaf 契约
+dws schema "aitable record query" --jq '[.parameters | to_entries[] | select(.value.required)]' # 定向查看必填字段
 dws schema --all                                        # CI/审计/基线的全量导出
 ```

@@ -708,7 +726,7 @@ dws dev connect --channel auto --robot-client-id <id> --robot-client-secret <sec
 <summary>即将推出</summary>

 - `conference`（视频会议）
- 多 skill 模式（默认）— 每产品一个独立 skill，位于 `skills/multi/`，安装与升级默认启用；`dws skill setup --mode mono` 可切回单 skill
+- 多 skill 模式（默认）— 每产品一个独立 skill，位于 `skills/multi/`，安装与升级默认启用；`dws skill setup --mode mono` 交互确认后可切回单 skill

 </details>

@@ -759,6 +777,7 @@ dws dev connect --channel auto --robot-client-id <id> --robot-client-secret <sec

 ## 参考与文档

+- [国际版（`.io`）使用手册](./docs/international-region-guide.zh-CN.md) — 国际版登录、国内/国际 profile 切换、隔离验证与排障
 - [命令索引](./docs/command-index.md) — 全部运行时命令，带描述与使用场景
 - [参考手册](./docs/reference.md) — 环境变量、退出码、输出格式、Shell 补全
 - [架构设计](./docs/architecture.md) — 静态端点管道、命令面、Transport 层
@@ -3,6 +3,7 @@
 "use strict";

 const fs = require("fs");
+const crypto = require("crypto");
 const os = require("os");
 const path = require("path");
 const childProcess = require("child_process");
@@ -16,6 +17,7 @@ const AGENT_DIRS = [
  ".qoderwork/skills",
  ".gemini/skills",
  ".codex/skills",
+  ".zcode/skills",
  ".github/skills",
  ".windsurf/skills",
  ".augment/skills",
@@ -45,6 +47,58 @@ function ensureCleanDir(dir) {
  fs.mkdirSync(dir, { recursive: true });
 }

+// backupStamp returns the UTC timestamp used for backup directory names,
+// matching the shell installers' `date -u +%Y%m%d-%H%M%S` layout.
+function backupStamp() {
+  const d = new Date();
+  const pad = (n) => String(n).padStart(2, "0");
+  return (
+    `${d.getUTCFullYear()}${pad(d.getUTCMonth() + 1)}${pad(d.getUTCDate())}` +
+    `-${pad(d.getUTCHours())}${pad(d.getUTCMinutes())}${pad(d.getUTCSeconds())}`
+  );
+}
+
+// backupAndRemoveSkillDir moves dir into <homeDir>/.dws/skill-backups/
+// <stamp>/<rel-or-basename> instead of destroying it (non-interactive
+// installs cannot confirm, so removals must stay reversible). Missing paths
+// are a no-op success. On any backup failure the directory is left in place
+// and false is returned so callers skip that target rather than silently
+// deleting data.
+function backupAndRemoveSkillDir(homeDir, dir, backups = null, renameFn = fs.renameSync) {
+  if (!fs.existsSync(dir) || !fs.statSync(dir).isDirectory()) {
+    return true;
+  }
+  const rel = path.relative(homeDir, dir);
+  const name =
+    rel && rel !== "." && !rel.startsWith("..") && !path.isAbsolute(rel)
+      ? rel.split(path.sep).join("-")
+      : path.basename(dir);
+  const stamp = backupStamp();
+  const backupRoot = path.join(homeDir, ".dws", "skill-backups");
+  let targetRoot = path.join(backupRoot, stamp);
+  let target = path.join(targetRoot, name);
+  for (let i = 1; fs.existsSync(target); i++) {
+    if (i > 1000) {
+      console.warn(`⚠️  备份目录冲突，保留原目录 ${dir}`);
+      return false;
+    }
+    targetRoot = path.join(backupRoot, `${stamp}-${i}`);
+    target = path.join(targetRoot, name);
+  }
+  try {
+    fs.mkdirSync(targetRoot, { recursive: true });
+    renameFn(dir, target);
+  } catch (err) {
+    console.warn(`⚠️  备份失败，保留原目录 ${dir}: ${err.message}`);
+    return false;
+  }
+  if (backups) {
+    backups.push({ original: dir, backup: target });
+  }
+  console.log(`  × 已备份并移除 ${dir} → ${target}`);
+  return true;
+}
+
 function findBinary(root) {
  const entries = fs.readdirSync(root, { withFileTypes: true });
  for (const entry of entries) {
@@ -117,34 +171,131 @@ function copyChildren(srcDir, destDir) {
  }
 }

+// publishCacheAtomically prepares a complete sibling tree before replacing a
+// cache. If copying or publishing fails, the previous cache stays available.
+// copyFn is injectable so the failure contract can be tested without relying
+// on platform-specific permission behavior.
+function publishCacheAtomically(sourceDir, cacheDir, copyFn = copyChildren) {
+  const cacheParent = path.dirname(cacheDir);
+  const cacheName = path.basename(cacheDir);
+  fs.mkdirSync(cacheParent, { recursive: true });
+
+  const stagedDir = fs.mkdtempSync(path.join(cacheParent, `.${cacheName}.tmp-`));
+  let rollbackDir = "";
+  let published = false;
+  try {
+    copyFn(sourceDir, stagedDir);
+
+    if (fs.existsSync(cacheDir)) {
+      rollbackDir = fs.mkdtempSync(path.join(cacheParent, `.${cacheName}.old-`));
+      fs.rmSync(rollbackDir, { recursive: true, force: true });
+      fs.renameSync(cacheDir, rollbackDir);
+    }
+
+    try {
+      fs.renameSync(stagedDir, cacheDir);
+      published = true;
+    } catch (publishErr) {
+      if (rollbackDir) {
+        try {
+          fs.renameSync(rollbackDir, cacheDir);
+          rollbackDir = "";
+        } catch (restoreErr) {
+          throw new Error(
+            `failed to publish cache ${cacheDir}: ${publishErr.message}; ` +
+              `failed to restore previous cache from ${rollbackDir}: ${restoreErr.message}`,
+          );
+        }
+      }
+      throw publishErr;
+    }
+
+    if (rollbackDir) {
+      try {
+        fs.rmSync(rollbackDir, { recursive: true, force: true });
+      } catch (cleanupErr) {
+        console.warn(
+          `⚠️  New cache is active, but old cache cleanup failed at ${rollbackDir}: ${cleanupErr.message}`,
+        );
+      }
+      rollbackDir = "";
+    }
+  } finally {
+    if (!published) {
+      fs.rmSync(stagedDir, { recursive: true, force: true });
+    }
+  }
+}
+
 function installSkillsToHomes(skillRoot) {
  const homeDir = os.homedir();
+  const managedNames = readManagedSkillNames(homeDir);
  let installed = 0;
+  let attempted = 0;
+  let failed = 0;
+
+  const specificAgentDirs = AGENT_DIRS.slice(1).filter((agentDir) =>
+    fs.existsSync(path.dirname(path.join(homeDir, agentDir))),
+  );
+
+  const installToBase = (baseDir) => {
+    const victims = [path.join(baseDir, "dws")];
+    if (fs.existsSync(baseDir)) {
+      for (const entry of fs.readdirSync(baseDir, { withFileTypes: true })) {
+        if (entry.isDirectory() && isManagedMultiSkillDir(path.join(baseDir, entry.name), managedNames)) {
+          victims.push(path.join(baseDir, entry.name));
+        }
+      }
+    }
+    try {
+      publishManagedMonoSkillSetAtomically(homeDir, skillRoot, baseDir, victims);
+    } catch (err) {
+      console.warn(`⚠️  跳过 ${baseDir}（mono 集合发布失败，已回滚）: ${err.message}`);
+      return false;
+    }
+    return true;
+  };

  AGENT_DIRS.forEach((agentDir, index) => {
+    if (index === 0 && specificAgentDirs.length > 0) {
+      return;
+    }
    const baseDir = path.join(homeDir, agentDir);
    const parentGate = path.dirname(baseDir);
    if (index > 0 && !fs.existsSync(parentGate)) {
      return;
    }
-    // Mutual exclusion: remove multi leftovers before laying down mono.
-    // Directories only — a stray file named dingtalk-x.md must survive.
-    if (fs.existsSync(baseDir)) {
-      for (const entry of fs.readdirSync(baseDir, { withFileTypes: true })) {
-        if (entry.isDirectory() && (entry.name.startsWith("dingtalk-") || entry.name === "dws-shared")) {
-          fs.rmSync(path.join(baseDir, entry.name), { recursive: true, force: true });
-        }
-      }
+    attempted += 1;
+    if (installToBase(baseDir)) {
+      installed += 1;
+    } else {
+      failed += 1;
    }
-    const destDir = path.join(baseDir, "dws");
-    fs.rmSync(destDir, { recursive: true, force: true });
-    copyChildren(skillRoot, destDir);
-    installed += 1;
  });

-  if (installed === 0) {
-    copyChildren(skillRoot, path.join(homeDir, ".agents", "skills", "dws"));
+  if (specificAgentDirs.length > 0 && installed > 0) {
+    try {
+      retireGenericSkillRoot(homeDir, managedNames);
+    } catch (err) {
+      console.warn(`⚠️  通用 Skill 副本迁移失败: ${err.message}`);
+      failed += 1;
+    }
  }
+
+  if (attempted === 0) {
+    if (installToBase(path.join(homeDir, ".agents", "skills"))) {
+      installed += 1;
+    } else {
+      failed += 1;
+    }
+  }
+  if (installed === 0) {
+    throw new Error("未安装任何 mono Skill：所有检测到的 Agent 目标均失败");
+  }
+  if (failed > 0) {
+    throw new Error(`有 ${failed} 个 Agent 目标安装 mono Skill 失败`);
+  }
+  fs.rmSync(path.join(skillStateDir(homeDir), "skills-state.json"), { force: true });
 }

 // multiTreeHasSkills mirrors multi_tree_has_skills in scripts/install.sh and
@@ -162,10 +313,306 @@ function multiTreeHasSkills(dir) {
    .some((e) => e.isDirectory() && fs.existsSync(path.join(dir, e.name, "SKILL.md")));
 }

+const MANAGED_SKILL_DIGEST_SCOPE = "skill-directory-v1";
+// Frozen exact names shipped before centralized ownership metadata. Retired
+// names stay here so old installs can be migrated without treating every
+// dingtalk-* directory as DWS-owned.
+const LEGACY_OFFICIAL_MULTI_SKILLS = new Set([
+  "dingtalk-agoal", "dingtalk-aiapp", "dingtalk-aisearch", "dingtalk-aitable",
+  "dingtalk-attendance", "dingtalk-calendar", "dingtalk-chat", "dingtalk-contact",
+  "dingtalk-dev", "dingtalk-devapp", "dingtalk-devdoc", "dingtalk-ding",
+  "dingtalk-doc", "dingtalk-drive", "dingtalk-event", "dingtalk-hrbrain",
+  "dingtalk-live", "dingtalk-mail", "dingtalk-markdown", "dingtalk-minutes",
+  "dingtalk-misc", "dingtalk-oa", "dingtalk-pat", "dingtalk-profile",
+  "dingtalk-report", "dingtalk-shared", "dingtalk-sheet", "dingtalk-skill",
+  "dingtalk-todo", "dingtalk-wiki", "dws-shared",
+]);
+
+function skillStateDir(homeDir) {
+  return (process.env.DWS_CONFIG_DIR || "").trim() || path.join(homeDir, ".dws");
+}
+
+function readManagedSkillNames(homeDir) {
+  try {
+    const state = JSON.parse(fs.readFileSync(path.join(skillStateDir(homeDir), "skills-state.json"), "utf8"));
+    return new Set((state.managed_skills || []).map((record) => record.name).filter(Boolean));
+  } catch (_) {
+    return new Set();
+  }
+}
+
+function isManagedMultiSkillDir(dir, managedNames) {
+  const name = path.basename(dir);
+  return LEGACY_OFFICIAL_MULTI_SKILLS.has(name) || managedNames.has(name);
+}
+
+function retireGenericSkillRoot(homeDir, managedNames) {
+  const baseDir = path.join(homeDir, ".agents", "skills");
+  const victims = [path.join(baseDir, "dws")];
+  if (fs.existsSync(baseDir)) {
+    for (const entry of fs.readdirSync(baseDir, { withFileTypes: true })) {
+      if (entry.isDirectory() && isManagedMultiSkillDir(path.join(baseDir, entry.name), managedNames)) {
+        victims.push(path.join(baseDir, entry.name));
+      }
+    }
+  }
+  const backups = [];
+  try {
+    for (const victim of victims) {
+      if (!backupAndRemoveSkillDir(homeDir, victim, backups)) {
+        throw new Error(`failed to back up Skill directory ${victim}`);
+      }
+    }
+  } catch (err) {
+    const restoreErrors = [];
+    for (let i = backups.length - 1; i >= 0; i -= 1) {
+      try {
+        fs.mkdirSync(path.dirname(backups[i].original), { recursive: true });
+        fs.renameSync(backups[i].backup, backups[i].original);
+      } catch (restoreErr) {
+        restoreErrors.push(`${backups[i].original}: ${restoreErr.message}`);
+      }
+    }
+    if (restoreErrors.length > 0) {
+      throw new Error(`${err.message}; generic-root rollback failed: ${restoreErrors.join("; ")}`);
+    }
+    throw err;
+  }
+}
+
+function skillDirectoryDigest(dir) {
+  const files = [];
+  const visit = (current, prefix) => {
+    for (const entry of fs.readdirSync(current, { withFileTypes: true })) {
+      const rel = prefix ? `${prefix}/${entry.name}` : entry.name;
+      const full = path.join(current, entry.name);
+      if (entry.isDirectory()) {
+        visit(full, rel);
+      } else {
+        files.push({ rel, full });
+      }
+    }
+  };
+  visit(dir, "");
+  files.sort((a, b) => Buffer.from(a.rel).compare(Buffer.from(b.rel)));
+  const hash = crypto.createHash("sha256");
+  for (const file of files) {
+    hash.update(file.rel, "utf8");
+    hash.update(Buffer.from([0]));
+    hash.update(fs.readFileSync(file.full));
+    hash.update(Buffer.from([0]));
+  }
+  return `sha256:${hash.digest("hex")}`;
+}
+
+// Publish a complete multi-skill set as one transaction. The entire new set
+// is staged before any Agent-visible directory moves. If a later backup or
+// publish fails, every partial publication is removed and all old directories
+// are restored from their exact backup paths.
+function publishManagedMultiSkillSetAtomically(
+  homeDir,
+  multiRoot,
+  baseDir,
+  skills,
+  victims,
+  options = {},
+) {
+  const copyFn = options.copyFn || copyChildren;
+  const renameFn = options.renameFn || fs.renameSync;
+  const removeFn = options.removeFn || ((dir) => fs.rmSync(dir, { recursive: true, force: true }));
+  fs.mkdirSync(baseDir, { recursive: true });
+  const stageRoot = fs.mkdtempSync(path.join(baseDir, ".dws-multi-set.tmp-"));
+  const staged = [];
+  const backups = [];
+  const published = [];
+
+  const restore = () => {
+    const restoreErrors = [];
+    for (let i = published.length - 1; i >= 0; i -= 1) {
+      try {
+        removeFn(published[i]);
+      } catch (err) {
+        restoreErrors.push(`remove ${published[i]}: ${err.message}`);
+      }
+    }
+    for (let i = backups.length - 1; i >= 0; i -= 1) {
+      const item = backups[i];
+      try {
+        fs.mkdirSync(path.dirname(item.original), { recursive: true });
+        renameFn(item.backup, item.original);
+      } catch (err) {
+        restoreErrors.push(`restore ${item.original} from ${item.backup}: ${err.message}`);
+      }
+    }
+    if (restoreErrors.length > 0) {
+      throw new Error(restoreErrors.join("; "));
+    }
+  };
+
+  try {
+    for (const name of skills) {
+      const stagedDir = path.join(stageRoot, name);
+      copyFn(path.join(multiRoot, name), stagedDir);
+      staged.push({ staged: stagedDir, dest: path.join(baseDir, name) });
+    }
+
+    const seen = new Set();
+    for (const victim of victims) {
+      const normalized = path.resolve(victim);
+      if (seen.has(normalized)) {
+        continue;
+      }
+      seen.add(normalized);
+      if (!backupAndRemoveSkillDir(homeDir, victim, backups, renameFn)) {
+        throw new Error(`failed to back up Skill directory ${victim}`);
+      }
+    }
+
+    for (const item of staged) {
+      renameFn(item.staged, item.dest);
+      published.push(item.dest);
+    }
+  } catch (err) {
+    try {
+      restore();
+    } catch (restoreErr) {
+      throw new Error(`${err.message}; rollback failed: ${restoreErr.message}`);
+    }
+    throw err;
+  } finally {
+    removeFn(stageRoot);
+  }
+}
+
+// Publish mono plus every mutually-exclusive managed multi victim as one
+// transaction. The complete dws/ tree is staged before any live directory is
+// moved; a later backup or publish failure restores the exact previous set.
+function publishManagedMonoSkillSetAtomically(
+  homeDir,
+  monoRoot,
+  baseDir,
+  victims,
+  options = {},
+) {
+  const copyFn = options.copyFn || copyChildren;
+  const renameFn = options.renameFn || fs.renameSync;
+  const removeFn = options.removeFn || ((dir) => fs.rmSync(dir, { recursive: true, force: true }));
+  fs.mkdirSync(baseDir, { recursive: true });
+  const stageRoot = fs.mkdtempSync(path.join(baseDir, ".dws-mono-set.tmp-"));
+  const stagedDir = path.join(stageRoot, "dws");
+  const destDir = path.join(baseDir, "dws");
+  const backups = [];
+  const published = [];
+
+  const restore = () => {
+    const restoreErrors = [];
+    for (let i = published.length - 1; i >= 0; i -= 1) {
+      try {
+        removeFn(published[i]);
+      } catch (err) {
+        restoreErrors.push(`remove ${published[i]}: ${err.message}`);
+      }
+    }
+    for (let i = backups.length - 1; i >= 0; i -= 1) {
+      const item = backups[i];
+      try {
+        fs.mkdirSync(path.dirname(item.original), { recursive: true });
+        renameFn(item.backup, item.original);
+      } catch (err) {
+        restoreErrors.push(`restore ${item.original} from ${item.backup}: ${err.message}`);
+      }
+    }
+    if (restoreErrors.length > 0) {
+      throw new Error(restoreErrors.join("; "));
+    }
+  };
+
+  try {
+    copyFn(monoRoot, stagedDir);
+
+    const seen = new Set();
+    for (const victim of victims) {
+      const normalized = path.resolve(victim);
+      if (seen.has(normalized)) {
+        continue;
+      }
+      seen.add(normalized);
+      if (!backupAndRemoveSkillDir(homeDir, victim, backups, renameFn)) {
+        throw new Error(`failed to back up Skill directory ${victim}`);
+      }
+    }
+
+    published.push(destDir);
+    renameFn(stagedDir, destDir);
+  } catch (err) {
+    try {
+      restore();
+    } catch (restoreErr) {
+      throw new Error(`${err.message}; rollback failed: ${restoreErr.message}`);
+    }
+    throw err;
+  } finally {
+    removeFn(stageRoot);
+  }
+}
+
+function writeSkillsState(homeDir, multiRoot, skills) {
+  const version = process.env.npm_package_version || process.env.DWS_PACKAGE_VERSION || "unknown";
+  const managedSkills = [...skills].sort().map((name) => ({
+    name,
+    version,
+    source: "npm-postinstall",
+    digest: skillDirectoryDigest(path.join(multiRoot, name)),
+    digest_scope: MANAGED_SKILL_DIGEST_SCOPE,
+  }));
+  const state = {
+    version,
+    official_skills: [...skills].sort(),
+    updated_skills: [...skills].sort(),
+    managed_skills: managedSkills,
+    updated_at: new Date().toISOString(),
+  };
+  const stateDir = skillStateDir(homeDir);
+  fs.mkdirSync(stateDir, { recursive: true });
+  const stage = fs.mkdtempSync(path.join(stateDir, ".skills-state.tmp-"));
+  const stagedFile = path.join(stage, "skills-state.json");
+  const statePath = path.join(stateDir, "skills-state.json");
+  const rollbackPath = path.join(stage, "skills-state.previous.json");
+  let movedPrevious = false;
+  let preserveRecovery = false;
+  try {
+    fs.writeFileSync(stagedFile, `${JSON.stringify(state, null, 2)}\n`, "utf8");
+    if (fs.existsSync(statePath)) {
+      fs.renameSync(statePath, rollbackPath);
+      movedPrevious = true;
+    }
+    try {
+      fs.renameSync(stagedFile, statePath);
+    } catch (err) {
+      if (movedPrevious && !fs.existsSync(statePath)) {
+        try {
+          fs.renameSync(rollbackPath, statePath);
+          movedPrevious = false;
+        } catch (restoreErr) {
+          preserveRecovery = true;
+          throw new Error(
+            `publish skills state failed: ${err.message}; restore also failed: ${restoreErr.message}; previous state retained at ${rollbackPath}`,
+          );
+        }
+      }
+      throw err;
+    }
+  } finally {
+    if (!preserveRecovery) {
+      fs.rmSync(stage, { recursive: true, force: true });
+    }
+  }
+}
+
 // installMultiSkillsToHomes mirrors installSkillsToHomes for the multi bundle:
 // every product skill becomes a sibling directory of the agent home. Mutual
-// exclusion: the mono leftover (dws/) and stale dingtalk-* skills not present
-// in the new bundle are removed first.
+// exclusion: the mono leftover (dws/) and stale, proven DWS-managed skills not
+// present in the new bundle are removed first.
 function installMultiSkillsToHomes(multiRoot) {
  const homeDir = os.homedir();
  const skills = fs
@@ -176,40 +623,81 @@ function installMultiSkillsToHomes(multiRoot) {
    throw new Error(`no product skills found under ${multiRoot}`);
  }
  const skillSet = new Set(skills);
+  const managedNames = readManagedSkillNames(homeDir);
  let installed = 0;
+  let attempted = 0;
+  let failed = 0;
+
+  const specificAgentDirs = AGENT_DIRS.slice(1).filter((agentDir) =>
+    fs.existsSync(path.dirname(path.join(homeDir, agentDir))),
+  );

  const installToBase = (baseDir) => {
    fs.mkdirSync(baseDir, { recursive: true });
-    fs.rmSync(path.join(baseDir, "dws"), { recursive: true, force: true });
+    const victims = [path.join(baseDir, "dws")];
+    // Mutual exclusion: include the mono leftover and stale managed skills in
+    // the same transaction as every replaced bundled skill.
    for (const entry of fs.readdirSync(baseDir, { withFileTypes: true })) {
      if (
        entry.isDirectory() &&
-        (entry.name.startsWith("dingtalk-") || entry.name === "dws-shared") &&
+        (LEGACY_OFFICIAL_MULTI_SKILLS.has(entry.name) || managedNames.has(entry.name)) &&
        !skillSet.has(entry.name)
      ) {
-        fs.rmSync(path.join(baseDir, entry.name), { recursive: true, force: true });
+        victims.push(path.join(baseDir, entry.name));
      }
    }
    for (const name of skills) {
-      const destDir = path.join(baseDir, name);
-      fs.rmSync(destDir, { recursive: true, force: true });
-      copyChildren(path.join(multiRoot, name), destDir);
+      victims.push(path.join(baseDir, name));
    }
+    try {
+      publishManagedMultiSkillSetAtomically(homeDir, multiRoot, baseDir, skills, victims);
+    } catch (err) {
+      console.warn(`⚠️  跳过 ${baseDir}（multi 集合发布失败，已回滚）: ${err.message}`);
+      return false;
+    }
+    return true;
  };

  AGENT_DIRS.forEach((agentDir, index) => {
+    if (index === 0 && specificAgentDirs.length > 0) {
+      return;
+    }
    const baseDir = path.join(homeDir, agentDir);
    const parentGate = path.dirname(baseDir);
    if (index > 0 && !fs.existsSync(parentGate)) {
      return;
    }
-    installToBase(baseDir);
-    installed += 1;
+    attempted += 1;
+    if (installToBase(baseDir)) {
+      installed += 1;
+    } else {
+      failed += 1;
+    }
  });

-  if (installed === 0) {
-    installToBase(path.join(homeDir, ".agents", "skills"));
+  if (specificAgentDirs.length > 0 && installed > 0) {
+    try {
+      retireGenericSkillRoot(homeDir, managedNames);
+    } catch (err) {
+      console.warn(`⚠️  通用 Skill 副本迁移失败: ${err.message}`);
+      failed += 1;
+    }
  }
+
+  if (attempted === 0) {
+    if (installToBase(path.join(homeDir, ".agents", "skills"))) {
+      installed += 1;
+    } else {
+      failed += 1;
+    }
+  }
+  if (installed === 0) {
+    throw new Error("未安装任何 multi Skill：所有检测到的 Agent 目标均失败");
+  }
+  if (failed > 0) {
+    throw new Error(`有 ${failed} 个 Agent 目标安装 multi Skill 失败`);
+  }
+  writeSkillsState(homeDir, multiRoot, skills);
 }

 // resolveSkillMode mirrors scripts/install.sh: DWS_SKILL_MODE (mono|multi)
@@ -257,15 +745,13 @@ function cacheUserSkills(extractedSkillsRoot) {
    : extractedSkillsRoot;
  if (fs.existsSync(path.join(monoSource, "SKILL.md"))) {
    const monoCache = path.join(cacheBase, "mono");
-    fs.rmSync(monoCache, { recursive: true, force: true });
-    copyChildren(monoSource, monoCache);
+    publishCacheAtomically(monoSource, monoCache);
  }

  const multiSource = path.join(extractedSkillsRoot, "multi");
  if (multiTreeHasSkills(multiSource)) {
    const multiCache = path.join(cacheBase, "multi");
-    fs.rmSync(multiCache, { recursive: true, force: true });
-    copyChildren(multiSource, multiCache);
+    publishCacheAtomically(multiSource, multiCache);
  }
 }

@@ -321,4 +807,12 @@ function main() {
  cacheUserSkills(skillsStaging);
 }

-main();
+if (require.main === module) {
+  main();
+}
+
+module.exports = {
+  publishCacheAtomically,
+  publishManagedMonoSkillSetAtomically,
+  publishManagedMultiSkillSetAtomically,
+};
@@ -17,18 +17,23 @@
 package main

 import (
+	"bytes"
 	"encoding/json"
 	"flag"
 	"fmt"
 	"io"
 	"os"
 	"path/filepath"
+	"strings"

 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/app"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/i18n"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/interfacesnapshot"
+	"github.com/spf13/cobra"
 )

+var newRootCommand = func() *cobra.Command { return app.NewRootCommand() }
+
 func main() {
 	os.Exit(run(os.Args[1:], os.Stdout, os.Stderr))
 }
@@ -112,7 +117,11 @@ func runGenerate(args []string, stdout, stderr io.Writer) error {
 	defer i18n.SetLang(previousLang)
 	i18n.SetLang("en")

-	snapshot := interfacesnapshot.Capture(app.NewRootCommand())
+	root := newRootCommand()
+	snapshot := interfacesnapshot.Capture(root)
+	if err := validateHelpRendering(root, snapshot); err != nil {
+		return err
+	}
 	if *output == "-" {
 		return interfacesnapshot.Write(stdout, snapshot)
 	}
@@ -138,6 +147,16 @@ func runCompare(args []string, stdout, stderr io.Writer) (bool, error) {
 	currentPath := flags.String("current", "", "candidate snapshot path")
 	basePath := flags.String("base", "", "target main/development baseline snapshot path")
 	stablePath := flags.String("stable", "", "latest stable GA snapshot path")
+	approvedMigrationsPath := flags.String(
+		"approved-flag-migrations",
+		"",
+		"merge-base-owned approved flag migration manifest",
+	)
+	candidateMigrationsPath := flags.String(
+		"candidate-flag-migrations",
+		"",
+		"candidate flag migration manifest",
+	)
 	if err := flags.Parse(args); err != nil {
 		return false, err
 	}
@@ -150,6 +169,14 @@ func runCompare(args []string, stdout, stderr io.Writer) (bool, error) {
 	if *basePath == "" && *stablePath == "" {
 		return false, fmt.Errorf("compare requires --base, --stable, or both")
 	}
+	if (*approvedMigrationsPath == "") != (*candidateMigrationsPath == "") {
+		return false, fmt.Errorf(
+			"--approved-flag-migrations and --candidate-flag-migrations must be provided together",
+		)
+	}
+	if *approvedMigrationsPath != "" && (*basePath == "" || *stablePath == "") {
+		return false, fmt.Errorf("flag migration compare requires both --base and --stable")
+	}

 	current, err := readSnapshot(*currentPath)
 	if err != nil {
@@ -170,6 +197,25 @@ func runCompare(args []string, stdout, stderr io.Writer) (bool, error) {
 	}

 	report := interfacesnapshot.CompareAll(current, references)
+	if *approvedMigrationsPath != "" {
+		approved, readErr := readFlagMigrationManifest(*approvedMigrationsPath)
+		if readErr != nil {
+			return false, fmt.Errorf("read approved flag migrations: %w", readErr)
+		}
+		candidate, readErr := readFlagMigrationManifest(*candidateMigrationsPath)
+		if readErr != nil {
+			return false, fmt.Errorf("read candidate flag migrations: %w", readErr)
+		}
+		report, err = interfacesnapshot.CompareAllWithFlagMigrations(
+			current,
+			references,
+			approved,
+			candidate,
+		)
+		if err != nil {
+			return false, fmt.Errorf("validate flag migration lifecycle: %w", err)
+		}
+	}
 	encoder := json.NewEncoder(stdout)
 	encoder.SetEscapeHTML(false)
 	encoder.SetIndent("", "  ")
@@ -179,6 +225,49 @@ func runCompare(args []string, stdout, stderr io.Writer) (bool, error) {
 	return report.Compatible, nil
 }

+func readFlagMigrationManifest(path string) (interfacesnapshot.FlagMigrationManifest, error) {
+	file, err := os.Open(filepath.Clean(path))
+	if err != nil {
+		return interfacesnapshot.FlagMigrationManifest{}, err
+	}
+	defer file.Close()
+	return interfacesnapshot.ReadFlagMigrationManifest(file)
+}
+
+func validateHelpRendering(root *cobra.Command, snapshot interfacesnapshot.Snapshot) error {
+	for _, command := range snapshot.Commands {
+		path := strings.TrimPrefix(command.Path, "dws")
+		resolved, remaining, err := root.Find(strings.Fields(path))
+		if err != nil || len(remaining) != 0 || resolved == nil {
+			return fmt.Errorf("resolve %q before help rendering: remaining=%v error=%v", command.Path, remaining, err)
+		}
+		if err := renderCommandHelp(resolved); err != nil {
+			return fmt.Errorf("render %q help: %w", command.Path, err)
+		}
+	}
+	return nil
+}
+
+func renderCommandHelp(command *cobra.Command) (err error) {
+	var stdout, stderr bytes.Buffer
+	command.InitDefaultHelpFlag()
+	command.SetOut(&stdout)
+	command.SetErr(&stderr)
+	defer func() {
+		if recovered := recover(); recovered != nil {
+			err = fmt.Errorf("help renderer panicked: %v", recovered)
+		}
+	}()
+	command.HelpFunc()(command, []string{})
+	if stderr.Len() > 0 {
+		return fmt.Errorf("help renderer wrote an error: %s", strings.TrimSpace(stderr.String()))
+	}
+	if stdout.Len() == 0 {
+		return fmt.Errorf("help renderer produced empty output")
+	}
+	return nil
+}
+
 func readSnapshot(path string) (interfacesnapshot.Snapshot, error) {
 	file, err := os.Open(filepath.Clean(path))
 	if err != nil {
@@ -191,5 +280,5 @@ func readSnapshot(path string) (interfacesnapshot.Snapshot, error) {
 func printUsage(w io.Writer) {
 	fmt.Fprintln(w, "usage:")
 	fmt.Fprintln(w, "  interface-snapshot generate [--output FILE]")
-	fmt.Fprintln(w, "  interface-snapshot compare --current FILE [--base FILE] [--stable FILE]")
+	fmt.Fprintln(w, "  interface-snapshot compare --current FILE [--base FILE] [--stable FILE] [--approved-flag-migrations FILE --candidate-flag-migrations FILE]")
 }
@@ -15,11 +15,16 @@ package main

 import (
 	"bytes"
+	"errors"
+	"io"
 	"os"
 	"path/filepath"
+	"strings"
 	"testing"

 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/interfacesnapshot"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/testseam"
+	"github.com/spf13/cobra"
 )

 func TestCrossPlatformCoverageRunGenerateCapturesActualRootOffline(t *testing.T) {
@@ -56,6 +61,24 @@ func TestCrossPlatformCoverageRunGenerateCapturesActualRootOffline(t *testing.T)
 	}
 }

+func TestCrossPlatformCoverageRunGenerateRejectsHelpRenderingFailure(t *testing.T) {
+	testseam.Swap(t, &newRootCommand, func() *cobra.Command {
+		root := &cobra.Command{Use: "dws"}
+		root.SetHelpFunc(func(command *cobra.Command, _ []string) {
+			_, _ = io.WriteString(command.ErrOrStderr(), "injected help failure")
+		})
+		return root
+	})
+
+	var stdout, stderr bytes.Buffer
+	if exitCode := run([]string{"generate"}, &stdout, &stderr); exitCode != 2 {
+		t.Fatalf("run(generate) exit=%d stderr=%s", exitCode, stderr.String())
+	}
+	if !strings.Contains(stderr.String(), "injected help failure") {
+		t.Fatalf("run(generate) stderr=%q", stderr.String())
+	}
+}
+
 func TestCrossPlatformCoverageRunCompareUsesBothSnapshotInputsAndExitCode(t *testing.T) {
 	current := commandSnapshot("dws")
 	mergeBase := commandSnapshot("dws")
@@ -83,6 +106,368 @@ func TestCrossPlatformCoverageRunCompareUsesBothSnapshotInputsAndExitCode(t *tes
 	}
 }

+func TestCrossPlatformCoverageRunCompareEnforcesBaseOwnedFlagMigrationLifecycle(t *testing.T) {
+	dir := t.TempDir()
+	before := flagMigrationSnapshot(false)
+	after := flagMigrationSnapshot(true)
+	currentPath := writeSnapshot(t, dir, "current.json", after)
+	basePath := writeSnapshot(t, dir, "base.json", before)
+	stablePath := writeSnapshot(t, dir, "stable.json", before)
+	approvedPath := writeManifest(t, dir, "approved.json", flagMigrationManifestJSON("pending"))
+	candidatePath := writeManifest(t, dir, "candidate.json", flagMigrationManifestJSON("consumed"))
+
+	var stdout, stderr bytes.Buffer
+	exitCode := run([]string{
+		"compare",
+		"--current", currentPath,
+		"--base", basePath,
+		"--stable", stablePath,
+		"--approved-flag-migrations", approvedPath,
+		"--candidate-flag-migrations", candidatePath,
+	}, &stdout, &stderr)
+	if exitCode != 0 {
+		t.Fatalf("exact base-owned migration exit=%d stderr=%s", exitCode, stderr.String())
+	}
+	if !bytes.Contains(stdout.Bytes(), []byte(`"compatible": true`)) {
+		t.Fatalf("exact migration report is not compatible:\n%s", stdout.String())
+	}
+
+	stdout.Reset()
+	stderr.Reset()
+	emptyApproved := writeManifest(t, dir, "empty-approved.json", `{"version":1,"migrations":[]}`)
+	exitCode = run([]string{
+		"compare",
+		"--current", currentPath,
+		"--base", basePath,
+		"--stable", stablePath,
+		"--approved-flag-migrations", emptyApproved,
+		"--candidate-flag-migrations", candidatePath,
+	}, &stdout, &stderr)
+	if exitCode != 2 || !strings.Contains(stderr.String(), "must start pending") {
+		t.Fatalf("candidate self-approval exit=%d stdout=%s stderr=%s", exitCode, stdout.String(), stderr.String())
+	}
+}
+
+func TestCrossPlatformCoverageRunCompareRequiresBothFlagMigrationInputs(t *testing.T) {
+	dir := t.TempDir()
+	currentPath := writeSnapshot(t, dir, "current.json", commandSnapshot("dws"))
+	basePath := writeSnapshot(t, dir, "base.json", commandSnapshot("dws"))
+	approvedPath := writeManifest(t, dir, "approved.json", `{"version":1,"migrations":[]}`)
+
+	var stdout, stderr bytes.Buffer
+	exitCode := run([]string{
+		"compare",
+		"--current", currentPath,
+		"--base", basePath,
+		"--approved-flag-migrations", approvedPath,
+	}, &stdout, &stderr)
+	if exitCode != 2 || !strings.Contains(stderr.String(), "must be provided together") {
+		t.Fatalf("one-sided migration input exit=%d stdout=%s stderr=%s", exitCode, stdout.String(), stderr.String())
+	}
+}
+
+func TestCrossPlatformCoverageRunCompareRequiresBothReferencesForFlagMigrations(t *testing.T) {
+	dir := t.TempDir()
+	currentPath := writeSnapshot(t, dir, "current.json", commandSnapshot("dws"))
+	basePath := writeSnapshot(t, dir, "base.json", commandSnapshot("dws"))
+	stablePath := writeSnapshot(t, dir, "stable.json", commandSnapshot("dws"))
+	approvedPath := writeManifest(t, dir, "approved.json", `{"version":1,"migrations":[]}`)
+	candidatePath := writeManifest(t, dir, "candidate.json", `{"version":1,"migrations":[]}`)
+
+	tests := []struct {
+		name string
+		args []string
+	}{
+		{
+			name: "missing stable",
+			args: []string{"--base", basePath},
+		},
+		{
+			name: "missing base",
+			args: []string{"--stable", stablePath},
+		},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			args := []string{"compare", "--current", currentPath}
+			args = append(args, test.args...)
+			args = append(args,
+				"--approved-flag-migrations", approvedPath,
+				"--candidate-flag-migrations", candidatePath,
+			)
+			var stdout, stderr bytes.Buffer
+			exitCode := run(args, &stdout, &stderr)
+			if exitCode != 2 || !strings.Contains(stderr.String(), "requires both --base and --stable") {
+				t.Fatalf("one-reference migration compare exit=%d stdout=%s stderr=%s", exitCode, stdout.String(), stderr.String())
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageRunPrintsUsageForMissingAndUnknownCommands(t *testing.T) {
+	tests := []struct {
+		name       string
+		args       []string
+		wantStderr []string
+	}{
+		{
+			name:       "missing command",
+			args:       nil,
+			wantStderr: []string{"usage:", "interface-snapshot generate", "--approved-flag-migrations"},
+		},
+		{
+			name:       "unknown command",
+			args:       []string{"unknown"},
+			wantStderr: []string{`unknown command "unknown"`, "usage:", "interface-snapshot compare"},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var stdout, stderr bytes.Buffer
+			if exitCode := run(test.args, &stdout, &stderr); exitCode != 2 {
+				t.Fatalf("run(%v) exit=%d, want 2", test.args, exitCode)
+			}
+			if stdout.Len() != 0 {
+				t.Fatalf("run(%v) unexpectedly wrote stdout: %s", test.args, stdout.String())
+			}
+			for _, want := range test.wantStderr {
+				if !strings.Contains(stderr.String(), want) {
+					t.Errorf("run(%v) stderr missing %q:\n%s", test.args, want, stderr.String())
+				}
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageRunRejectsInvalidSubcommandArguments(t *testing.T) {
+	tests := []struct {
+		name string
+		args []string
+		want string
+	}{
+		{name: "generate unknown flag", args: []string{"generate", "--unknown"}, want: "flag provided but not defined"},
+		{name: "generate positional", args: []string{"generate", "unexpected"}, want: "generate accepts no positional arguments"},
+		{name: "compare unknown flag", args: []string{"compare", "--unknown"}, want: "flag provided but not defined"},
+		{name: "compare positional", args: []string{"compare", "unexpected"}, want: "compare accepts no positional arguments"},
+		{name: "compare missing current", args: []string{"compare", "--base", "base.json"}, want: "compare requires --current"},
+		{name: "compare missing reference", args: []string{"compare", "--current", "current.json"}, want: "compare requires --base, --stable, or both"},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var stdout, stderr bytes.Buffer
+			if exitCode := run(test.args, &stdout, &stderr); exitCode != 2 {
+				t.Fatalf("run(%v) exit=%d, want 2", test.args, exitCode)
+			}
+			if !strings.Contains(stderr.String(), test.want) {
+				t.Fatalf("run(%v) stderr missing %q:\n%s", test.args, test.want, stderr.String())
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageRunGenerateRejectsUnsafeOutputPath(t *testing.T) {
+	outputDirectory := t.TempDir()
+	var stdout, stderr bytes.Buffer
+	exitCode := run([]string{"generate", "--output", outputDirectory}, &stdout, &stderr)
+	if exitCode != 2 || !strings.Contains(stderr.String(), "create snapshot") {
+		t.Fatalf("directory output exit=%d stdout=%s stderr=%s", exitCode, stdout.String(), stderr.String())
+	}
+}
+
+func TestCrossPlatformCoverageRunGenerateReportsTemporaryDirectoryFailure(t *testing.T) {
+	missingTempRoot := filepath.Join(t.TempDir(), "missing")
+	for _, name := range []string{"TMPDIR", "TMP", "TEMP"} {
+		t.Setenv(name, missingTempRoot)
+	}
+
+	var stdout, stderr bytes.Buffer
+	exitCode := run([]string{"generate"}, &stdout, &stderr)
+	if exitCode != 2 || !strings.Contains(stderr.String(), "create isolated home") {
+		t.Fatalf("invalid temporary root exit=%d stdout=%s stderr=%s", exitCode, stdout.String(), stderr.String())
+	}
+}
+
+func TestCrossPlatformCoverageRunCompareReportsSnapshotReadFailures(t *testing.T) {
+	dir := t.TempDir()
+	validPath := writeSnapshot(t, dir, "valid.json", commandSnapshot("dws"))
+	missingPath := filepath.Join(dir, "missing.json")
+	tests := []struct {
+		name string
+		args []string
+		want string
+	}{
+		{
+			name: "current",
+			args: []string{"compare", "--current", missingPath, "--base", validPath},
+			want: "read current snapshot",
+		},
+		{
+			name: "main",
+			args: []string{"compare", "--current", validPath, "--base", missingPath},
+			want: "read main/development baseline snapshot",
+		},
+		{
+			name: "stable",
+			args: []string{"compare", "--current", validPath, "--stable", missingPath},
+			want: "read stable snapshot",
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var stdout, stderr bytes.Buffer
+			if exitCode := run(test.args, &stdout, &stderr); exitCode != 2 {
+				t.Fatalf("run(compare) exit=%d, want 2", exitCode)
+			}
+			if !strings.Contains(stderr.String(), test.want) {
+				t.Fatalf("stderr missing %q:\n%s", test.want, stderr.String())
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageRunCompareReportsEachManifestReadFailure(t *testing.T) {
+	dir := t.TempDir()
+	snapshotPath := writeSnapshot(t, dir, "snapshot.json", commandSnapshot("dws"))
+	validManifest := writeManifest(t, dir, "valid-manifest.json", `{"version":1,"migrations":[]}`)
+	invalidManifest := writeManifest(t, dir, "invalid-manifest.json", `{`)
+	tests := []struct {
+		name      string
+		approved  string
+		candidate string
+		want      string
+	}{
+		{
+			name:      "approved manifest",
+			approved:  invalidManifest,
+			candidate: validManifest,
+			want:      "read approved flag migrations",
+		},
+		{
+			name:      "candidate manifest",
+			approved:  validManifest,
+			candidate: invalidManifest,
+			want:      "read candidate flag migrations",
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			var stdout, stderr bytes.Buffer
+			exitCode := run([]string{
+				"compare",
+				"--current", snapshotPath,
+				"--base", snapshotPath,
+				"--stable", snapshotPath,
+				"--approved-flag-migrations", test.approved,
+				"--candidate-flag-migrations", test.candidate,
+			}, &stdout, &stderr)
+			if exitCode != 2 || !strings.Contains(stderr.String(), test.want) {
+				t.Fatalf("%s exit=%d stdout=%s stderr=%s", test.name, exitCode, stdout.String(), stderr.String())
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageRunCompareReportsOutputFailure(t *testing.T) {
+	dir := t.TempDir()
+	snapshotPath := writeSnapshot(t, dir, "snapshot.json", commandSnapshot("dws"))
+	var stderr bytes.Buffer
+	exitCode := run([]string{
+		"compare",
+		"--current", snapshotPath,
+		"--base", snapshotPath,
+	}, failingWriter{}, &stderr)
+	if exitCode != 2 || !strings.Contains(stderr.String(), "write comparison report") {
+		t.Fatalf("comparison output failure exit=%d stderr=%s", exitCode, stderr.String())
+	}
+}
+
+func TestCrossPlatformCoverageReadHelpersRejectMissingAndInvalidInputs(t *testing.T) {
+	dir := t.TempDir()
+	missingPath := filepath.Join(dir, "missing.json")
+	invalidPath := filepath.Join(dir, "invalid.json")
+	if err := os.WriteFile(invalidPath, []byte(`{`), 0o600); err != nil {
+		t.Fatalf("write invalid fixture: %v", err)
+	}
+
+	if _, err := readSnapshot(missingPath); err == nil {
+		t.Fatal("readSnapshot(missing) unexpectedly succeeded")
+	}
+	if _, err := readSnapshot(invalidPath); err == nil {
+		t.Fatal("readSnapshot(invalid) unexpectedly succeeded")
+	}
+	if _, err := readFlagMigrationManifest(missingPath); err == nil {
+		t.Fatal("readFlagMigrationManifest(missing) unexpectedly succeeded")
+	}
+	if _, err := readFlagMigrationManifest(invalidPath); err == nil {
+		t.Fatal("readFlagMigrationManifest(invalid) unexpectedly succeeded")
+	}
+}
+
+func TestCrossPlatformCoverageValidateHelpRenderingReportsResolveError(t *testing.T) {
+	root := &cobra.Command{Use: "dws"}
+	err := validateHelpRendering(root, commandSnapshot("dws missing"))
+	if err == nil || !strings.Contains(err.Error(), `resolve "dws missing" before help rendering`) {
+		t.Fatalf("validateHelpRendering resolve error = %v", err)
+	}
+}
+
+func TestCrossPlatformCoverageValidateHelpRenderingReportsTemplateError(t *testing.T) {
+	root := &cobra.Command{Use: "dws"}
+	root.SetHelpTemplate(`{{index .Commands 99}}`)
+	err := validateHelpRendering(root, commandSnapshot("dws"))
+	if err == nil || !strings.Contains(err.Error(), `render "dws" help`) {
+		t.Fatalf("validateHelpRendering template error = %v", err)
+	}
+}
+
+func TestCrossPlatformCoverageValidateHelpRenderingRecoversTemplatePanic(t *testing.T) {
+	root := &cobra.Command{Use: "dws"}
+	root.SetHelpTemplate("{{")
+	err := validateHelpRendering(root, commandSnapshot("dws"))
+	if err == nil || !strings.Contains(err.Error(), `render "dws" help`) {
+		t.Fatalf("validateHelpRendering template panic = %v", err)
+	}
+}
+
+func TestCrossPlatformCoverageValidateHelpRenderingRejectsCustomHelpStderr(t *testing.T) {
+	root := &cobra.Command{Use: "dws"}
+	root.SetHelpFunc(func(command *cobra.Command, _ []string) {
+		_, _ = io.WriteString(command.ErrOrStderr(), "injected help failure")
+	})
+	err := validateHelpRendering(root, commandSnapshot("dws"))
+	if err == nil || !strings.Contains(err.Error(), "injected help failure") {
+		t.Fatalf("validateHelpRendering custom stderr = %v", err)
+	}
+}
+
+func TestCrossPlatformCoverageValidateHelpRenderingRejectsEmptyOutput(t *testing.T) {
+	root := &cobra.Command{Use: "dws"}
+	root.SetHelpFunc(func(*cobra.Command, []string) {})
+	err := validateHelpRendering(root, commandSnapshot("dws"))
+	if err == nil || !strings.Contains(err.Error(), "empty") {
+		t.Fatalf("validateHelpRendering empty output = %v", err)
+	}
+}
+
+func TestCrossPlatformCoverageValidateHelpRenderingAcceptsNormalOutput(t *testing.T) {
+	root := &cobra.Command{Use: "dws", Short: "root command"}
+	if err := validateHelpRendering(root, commandSnapshot("dws")); err != nil {
+		t.Fatalf("validateHelpRendering normal output: %v", err)
+	}
+}
+
+type failingWriter struct{}
+
+func (failingWriter) Write([]byte) (int, error) {
+	return 0, errors.New("injected write failure")
+}
+
+var _ io.Writer = failingWriter{}
+
 func commandSnapshot(paths ...string) interfacesnapshot.Snapshot {
 	commands := make([]interfacesnapshot.Command, 0, len(paths))
 	for _, path := range paths {
@@ -120,6 +505,68 @@ func writeSnapshot(t *testing.T, dir, name string, snapshot interfacesnapshot.Sn
 	return path
 }

+func writeManifest(t *testing.T, dir, name, contents string) string {
+	t.Helper()
+	path := filepath.Join(dir, name)
+	if err := os.WriteFile(path, []byte(contents), 0o600); err != nil {
+		t.Fatalf("write %s: %v", path, err)
+	}
+	return path
+}
+
+func flagMigrationSnapshot(after bool) interfacesnapshot.Snapshot {
+	legacy := interfacesnapshot.Flag{
+		Name:      "legacy-id",
+		Shorthand: "l",
+		Type:      "string",
+		Default:   "",
+		NoOpt:     "auto",
+		Required:  true,
+	}
+	flags := []interfacesnapshot.Flag{legacy}
+	if after {
+		legacy.Required = false
+		legacy.Hidden = true
+		legacy.AliasOf = "message-id"
+		flags = []interfacesnapshot.Flag{
+			legacy,
+			{Name: "message-id", Type: "string", Default: "", Required: true},
+		}
+	}
+	return interfacesnapshot.Snapshot{
+		SchemaVersion: interfacesnapshot.SchemaVersion,
+		Rules: interfacesnapshot.Rules{
+			ExcludedCommandSubtrees: []string{},
+			ExcludedFlags:           []string{},
+		},
+		Commands: []interfacesnapshot.Command{
+			{Path: "dws", Runnable: true, Aliases: []string{}, LocalFlags: []interfacesnapshot.Flag{}, InheritedFlags: []interfacesnapshot.Flag{}},
+			{Path: "dws chat send", Runnable: true, Aliases: []string{}, LocalFlags: flags, InheritedFlags: []interfacesnapshot.Flag{}},
+		},
+	}
+}
+
+func flagMigrationManifestJSON(state string) string {
+	return strings.Replace(`{
+  "version": 1,
+  "migrations": [{
+    "command": "dws chat send",
+    "legacy": {
+      "name": "legacy-id",
+      "before": {"present": true, "type": "string", "required": true, "shorthand": "l", "no_opt": "auto", "scope": "local"},
+      "after": {"present": true, "type": "string", "hidden": true, "shorthand": "l", "no_opt": "auto", "scope": "local", "alias_of": "message-id"}
+    },
+    "canonical": {
+      "name": "message-id",
+      "before": {"present": false},
+      "after": {"present": true, "type": "string", "required": true, "scope": "local"}
+    },
+    "state": "STATE",
+    "reason": "reviewed exact migration"
+  }]
+}`, "STATE", state, 1)
+}
+
 func hasFlag(flags []interfacesnapshot.Flag, name, flagType string) bool {
 	for _, flag := range flags {
 		if flag.Name == name && flag.Type == flagType {
@@ -15,12 +15,76 @@ package main

 import (
 	"os"
+	"strings"

 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/app"
+	"gitlab.alibaba-inc.com/aes/aem-go-sdk/clitrack"
 )

-var exit = os.Exit
+var (
+	appExecute               = app.ExecuteWithTelemetry
+	resolveTelemetryIdentity = app.ResolveTelemetryIdentity
+	trackRun                 = func(cfg clitrack.Config, execute func() error, exitCode func(error) int) {
+		clitrack.New(cfg).Run(execute, exitCode)
+	}
+)
+
+// trackedExitError tells clitrack that the command failed without asking it to
+// print the error a second time. The already-rendered message is published via
+// ExtraFields c5, while app.Execute remains the sole owner of presentation.
+type trackedExitError struct{}
+
+func (trackedExitError) Error() string { return "" }
+
+func trackerConfig(identity app.TelemetryIdentity, commandPath, errorMessage *string) clitrack.Config {
+	return clitrack.Config{
+		PID:                   "wcCRwZ",
+		App:                   "dws",
+		Version:               app.RawVersion(),
+		UID:                   identity.UserID,
+		Username:              identity.UserName,
+		NoCommandLine:         true,
+		NoCwd:                 true,
+		NoAutomaticDimensions: true,
+		ExtraFields: func() map[string]string {
+			fields := map[string]string{"c9": *commandPath}
+			if identity.CorpID != "" {
+				fields["c10"] = identity.CorpID
+			}
+			if *errorMessage != "" {
+				fields["c5"] = *errorMessage
+			}
+			return fields
+		},
+	}
+}
+
+func telemetryOptedOut() bool {
+	return strings.TrimSpace(os.Getenv("DO_NOT_TRACK")) != ""
+}

 func main() {
-	exit(app.Execute())
+	optedOut := telemetryOptedOut()
+	identity := app.TelemetryIdentity{}
+	if !optedOut {
+		identity = resolveTelemetryIdentity(os.Args[1:])
+	}
+	exitCode := 0
+	commandPath := "dws"
+	errorMessage := ""
+	cfg := trackerConfig(identity, &commandPath, &errorMessage)
+	if optedOut {
+		cfg.PID = ""
+	}
+	trackRun(
+		cfg,
+		func() error {
+			exitCode, commandPath, errorMessage = appExecute()
+			if exitCode != 0 {
+				return trackedExitError{}
+			}
+			return nil
+		},
+		func(error) int { return exitCode },
+	)
 }
@@ -1,27 +1,220 @@
 package main

 import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"net/url"
 	"os"
+	"slices"
+	"sort"
+	"strings"
 	"testing"
+	"time"
+
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/app"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/testseam"
+	"gitlab.alibaba-inc.com/aes/aem-go-sdk/clitrack"
 )

-func TestCrossPlatformCoverageMainExitsWithSuccessfulVersionCommand(t *testing.T) {
-	previousExit := exit
-	previousArgs := os.Args
-	t.Cleanup(func() {
-		exit = previousExit
-		os.Args = previousArgs
-	})
+func TestCrossPlatformCoverageMainRunsThroughCLITracker(t *testing.T) {
+	for _, wantCode := range []int{0, 1, 3, 5} {
+		t.Run(fmt.Sprintf("exit_%d", wantCode), func(t *testing.T) {
+			t.Setenv("DO_NOT_TRACK", "")
+			wantError := ""
+			if wantCode != 0 {
+				wantError = "synthetic failure"
+			}
+			testseam.Swap(t, &os.Args, []string{"dws", "sheet", "read", "--profile", "corp-a"})
+			testseam.Swap(t, &resolveTelemetryIdentity, func(args []string) app.TelemetryIdentity {
+				if strings.Join(args, " ") != "sheet read --profile corp-a" {
+					t.Fatalf("telemetry identity args = %#v", args)
+				}
+				return app.TelemetryIdentity{UserID: "user-1", UserName: "Alice", CorpID: "corp-1"}
+			})
+			testseam.Swap(t, &appExecute, func() (int, string, string) { return wantCode, "sheet read", wantError })
+			called := false
+			testseam.Swap(t, &trackRun, func(cfg clitrack.Config, execute func() error, exitCode func(error) int) {
+				called = true
+				if cfg.PID != "wcCRwZ" || cfg.App != "dws" {
+					t.Fatalf("tracker identity = PID %q App %q", cfg.PID, cfg.App)
+				}
+				if cfg.Version != app.RawVersion() {
+					t.Fatalf("tracker Version = %q, want %q", cfg.Version, app.RawVersion())
+				}
+				if !cfg.NoCommandLine || !cfg.NoCwd || !cfg.NoAutomaticDimensions || cfg.CaptureOutput {
+					t.Fatalf("tracker privacy config = NoCommandLine %v NoCwd %v NoAutomaticDimensions %v CaptureOutput %v", cfg.NoCommandLine, cfg.NoCwd, cfg.NoAutomaticDimensions, cfg.CaptureOutput)
+				}
+				if cfg.Env != "" || cfg.EventID != "" || cfg.Endpoint != "" || cfg.FlushTimeout != 0 || cfg.OutputMaxLen != 0 {
+					t.Fatalf("tracker SDK defaults were overridden: %#v", cfg)
+				}
+				if cfg.UID != "user-1" || cfg.Username != "Alice" || cfg.UserType != "" {
+					t.Fatalf("tracker user identity = UID %q Username %q UserType %q", cfg.UID, cfg.Username, cfg.UserType)
+				}

+				err := execute()
+				if wantCode == 0 && err != nil {
+					t.Fatalf("successful tracked execute error = %v", err)
+				}
+				if wantCode != 0 && (err == nil || err.Error() != "") {
+					t.Fatalf("failed tracked execute error = %#v, want empty sentinel", err)
+				}
+				if gotCode := exitCode(err); gotCode != wantCode {
+					t.Fatalf("tracked exit code = %d, want %d", gotCode, wantCode)
+				}
+				fields := cfg.ExtraFields()
+				if fields["c9"] != "sheet read" || fields["c10"] != "corp-1" || fields["c5"] != wantError {
+					t.Fatalf("tracker extra fields = %#v, want command path, corp ID, and error %q", fields, wantError)
+				}
+				if (wantError == "" && len(fields) != 2) || (wantError != "" && len(fields) != 3) {
+					t.Fatalf("tracker extra field count = %d for error %q", len(fields), wantError)
+				}
+			})
+
+			main()
+			if !called {
+				t.Fatalf("trackRun was not called for exit code %d", wantCode)
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageTrackerConfigOmitsEmptyOrganization(t *testing.T) {
+	commandPath := "version"
+	errorMessage := ""
+	cfg := trackerConfig(app.TelemetryIdentity{}, &commandPath, &errorMessage)
+	if cfg.UID != "" {
+		t.Fatalf("empty identity UID = %q", cfg.UID)
+	}
+	if cfg.Username != "" {
+		t.Fatalf("empty identity Username = %q", cfg.Username)
+	}
+	if fields := cfg.ExtraFields(); len(fields) != 1 || fields["c9"] != "version" {
+		t.Fatalf("empty organization fields = %#v", fields)
+	}
+}
+
+func TestCrossPlatformCoverageDefaultTrackRunNoopTracker(t *testing.T) {
 	called := false
-	code := -1
-	exit = func(value int) {
+	trackRun(clitrack.Config{}, func() error {
 		called = true
-		code = value
+		return nil
+	}, nil)
+	if !called {
+		t.Fatal("default tracker did not execute callback")
 	}
-	os.Args = []string{"dws", "version"}
+}
+
+func TestCrossPlatformCoverageMainRespectsDoNotTrack(t *testing.T) {
+	t.Setenv("DO_NOT_TRACK", "1")
+	testseam.Swap(t, &os.Args, []string{"dws", "version"})
+	testseam.Swap(t, &resolveTelemetryIdentity, func([]string) app.TelemetryIdentity {
+		t.Fatal("DO_NOT_TRACK must skip telemetry identity reads")
+		return app.TelemetryIdentity{}
+	})
+	testseam.Swap(t, &appExecute, func() (int, string, string) { return 0, "version", "" })
+	testseam.Swap(t, &trackRun, func(cfg clitrack.Config, execute func() error, exitCode func(error) int) {
+		if cfg.PID != "" || cfg.UID != "" || cfg.Username != "" {
+			t.Fatalf("opted-out tracker config = %#v", cfg)
+		}
+		if err := execute(); err != nil {
+			t.Fatalf("opted-out execution failed: %v", err)
+		}
+		if code := exitCode(nil); code != 0 {
+			t.Fatalf("opted-out exit code = %d, want 0", code)
+		}
+	})
+
 	main()
-	if !called || code != 0 {
-		t.Fatalf("main exit = called %v, code %d", called, code)
+}
+
+func TestCrossPlatformCoverageTrackerPayloadUsesReviewedFieldWhitelist(t *testing.T) {
+	testseam.Protect(t, &os.Args)
+	os.Args = []string{"dws", "sheet", "read", "--access-token", "must-not-leak"}
+	t.Setenv("SHELL", "/bin/zsh")
+	t.Setenv("TERM_SESSION_ID", "stable-session")
+	t.Setenv("TMUX_PANE", "%42")
+	t.Setenv("LANG", "zh_CN.UTF-8")
+	t.Setenv("LC_ALL", "zh_CN.UTF-8")
+	t.Chdir(t.TempDir())
+
+	requestBody := make(chan []byte, 1)
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, req *http.Request) {
+		body, _ := io.ReadAll(req.Body)
+		requestBody <- body
+		w.WriteHeader(http.StatusNoContent)
+	}))
+	defer server.Close()
+
+	commandPath := "sheet read"
+	errorMessage := ""
+	cfg := trackerConfig(app.TelemetryIdentity{UserID: "user-1", UserName: "Alice", CorpID: "corp-1"}, &commandPath, &errorMessage)
+	cfg.Endpoint = server.URL
+	cfg.FlushTimeout = time.Second
+	clitrack.New(cfg).Run(func() error { return nil }, nil)
+
+	var body []byte
+	select {
+	case body = <-requestBody:
+	case <-time.After(time.Second):
+		t.Fatal("timed out waiting for telemetry request")
+	}
+	var envelope map[string]string
+	if err := json.Unmarshal(body, &envelope); err != nil {
+		t.Fatalf("decode telemetry request %q: %v", body, err)
+	}
+	decoded, err := url.QueryUnescape(envelope["gokey"])
+	if err != nil {
+		t.Fatalf("decode gokey: %v", err)
+	}
+	globalFields, err := url.ParseQuery(decoded)
+	if err != nil {
+		t.Fatalf("parse global telemetry fields: %v", err)
+	}
+	eventFields, err := url.ParseQuery(globalFields.Get("msg"))
+	if err != nil {
+		t.Fatalf("parse event telemetry fields: %v", err)
+	}
+
+	assertTelemetryKeys(t, globalFields, []string{"app_name", "app_version", "env", "msg", "pid", "platform", "uid", "username", "version"})
+	assertTelemetryKeys(t, eventFields, []string{"c1", "c10", "c3", "c4", "c9", "p1", "p4", "ts", "type"})
+	for key, want := range map[string]string{
+		"app_name": "dws", "app_version": app.RawVersion(), "env": "prod", "pid": "wcCRwZ",
+		"platform": "cli", "uid": "user-1", "username": "Alice", "version": app.RawVersion(),
+	} {
+		if got := globalFields.Get(key); got != want {
+			t.Fatalf("global telemetry field %s = %q, want %q", key, got, want)
+		}
+	}
+	for key, want := range map[string]string{
+		"type": "event", "p1": "cli.exec", "p4": "SYS", "c1": "dws", "c3": "0", "c9": "sheet read", "c10": "corp-1",
+	} {
+		if got := eventFields.Get(key); got != want {
+			t.Fatalf("event telemetry field %s = %q, want %q", key, got, want)
+		}
+	}
+	for _, key := range []string{"device_id", "ext", "os", "os_version", "pv_id", "sdk_version", "sid", "timezone_offset"} {
+		if globalFields.Has(key) {
+			t.Fatalf("global telemetry leaked %s: %q", key, decoded)
+		}
+	}
+	for _, key := range []string{"c2", "c5", "c6", "c7", "c8"} {
+		if eventFields.Has(key) {
+			t.Fatalf("event telemetry leaked %s: %q", key, globalFields.Get("msg"))
+		}
+	}
+}
+
+func assertTelemetryKeys(t *testing.T, fields url.Values, want []string) {
+	t.Helper()
+	got := make([]string, 0, len(fields))
+	for key := range fields {
+		got = append(got, key)
+	}
+	sort.Strings(got)
+	if !slices.Equal(got, want) {
+		t.Fatalf("telemetry keys = %v, want %v", got, want)
 	}
 }
@@ -40,7 +40,6 @@
 - `internal/output`: response formatting (json, table, raw, pretty)
 - `internal/logging`: structured logging and argument sanitization
 - `internal/tui`: terminal UI helpers
- `internal/recovery`: panic recovery and graceful degradation
 - `pkg/configmeta`: environment variable registry and documentation
 - `pkg/config`: configuration constants and paths
 - `pkg/edition`: edition detection (oss vs enterprise)
@@ -43,7 +43,7 @@ repository root while preserving repo-local guidance for automation.
 - Error message or category issues: inspect `internal/errors`
 - Audit log issues: inspect `internal/audit`
 - Plugin loading or command surface: inspect `internal/plugin`
- Failure or degraded mode: inspect `internal/errors`, `internal/recovery`
+- Failure or degraded mode: inspect `internal/errors`

 ## Policy Checks

@@ -48,8 +48,12 @@ It then runs:
  --fast-path "$PR_BASE_SHA" HEAD
 ```

-Because the verified PR diff contains only `CHANGELOG.md`, the validator and
-its policy dependencies in that merge tree are byte-for-byte the current base
+The exact fast path remains limited to historic one-file maintenance. A
+release-seal PR uses `--content-only`, which permits the generated
+`CHANGELOG.md` change together with archival moves from `.changes/` to
+`.changes/released/`; it receives the normal scoped admission instead of this
+fast path. Ordinary PRs must not modify `CHANGELOG.md`; they add a standalone
+release fragment instead. The validator and its policy dependencies in that merge tree are byte-for-byte the current base
 versions. Validation targets the synthetic merge tree, not the feature-branch
 tree, so a stale branch cannot supply an older validator or combine with newer
 base notes into an invalid final CHANGELOG.
@@ -79,10 +83,12 @@ to the complete main admission suite. A source change can therefore never
 inherit the CHANGELOG-only result.

 Any PR that touches `CHANGELOG.md` but also changes another file runs the same
-content contract in `Policy` with `--content-only`. That mode permits the
-second file but still rejects invalid dates or versions, missing bullets,
-placeholder `TODO`/`TBD`, unmanaged-section changes, and unsafe tree modes.
-Adding a second file therefore cannot bypass CHANGELOG validation.
+content contract in `Policy` with `--content-only`. That mode accepts only
+fragment archival moves (`.changes/<name>.md` to
+`.changes/released/<version>/<name>.md`) alongside the changelog; source and
+documentation changes are rejected. It still rejects invalid dates or
+versions, missing bullets, placeholder `TODO`/`TBD`, unmanaged-section
+changes, and unsafe tree modes.

 ## Risk tiers and downstream boundaries

@@ -157,28 +163,59 @@ expected to repeat every CI job locally:
 ```sh
 make build
 make policy
-make interface-integrity
-make authoritative-interface-integrity BASE_REF=<merge-base>
-make schema-compatibility BASE_REF=<merge-base>
+make interface-integrity BASE_REF=<merge-base> STABLE_REF=<stable-tag> CANDIDATE_REF=<candidate-sha>
+make schema-compatibility BASE_REF=<merge-base> STABLE_REF=<stable-tag> CANDIDATE_REF=<candidate-sha>
 make skill-command-integrity
 make cli-smoke
 make mock-mcp-smoke
 go test -v -count=1 ./pkg/editiontest/...
 ```

-For an exact CHANGELOG-only branch:
+CI 先解析并核对精确的 merge-base、最近可达且未撤回的 stable GA tag 和已提交的 candidate
+SHA，再调用 `make authoritative-interface-integrity`。本地 `make interface-integrity`
+与该 CI target 都只委托给同一个 modern authoritative wrapper，不存在第二个比较
+入口。省略 `BASE_REF` 时本地 target 默认比较 `origin/main`，省略 `STABLE_REF` 时自动
+选择该 base 可达且未撤回的最近 stable GA tag，省略 `CANDIDATE_REF` 时比较已提交的 `HEAD`。
+需要逐字复现某次 CI 时，应显式传入该次运行记录的 merge-base、stable tag 和
+candidate SHA。
+
+`make update-interface-baseline` / `make reset-interface-baseline` 只维护
+`test/fixtures/cli-interface-baseline.txt` 这一份非权威 CLI Smoke fixture。底层旧
+`check-interface-baseline.sh` 不再作为本地或 CI 的兼容性审批入口，也不能用于批准
+flag 迁移。
+
+Schema compatibility 使用同一组 base、stable、candidate refs 和同一份 base-owned flag
+migration ledger。merge-base-owned checker 分别规范化 merge-base 与 stable 的完整
+Schema，并让 candidate 对两份历史 contract 独立执行检查；它只把已通过 Interface
+lifecycle 的 exact rename 规范化到当前历史副本，不会维护第二份 allowlist，也不会
+放宽其他 Schema 历史字段。
+
+For a release-seal branch that archives rendered fragments:

 ```sh
 base_ref=$(git merge-base HEAD origin/main)
-./scripts/policy/check-changelog-pr.sh --fast-path "$base_ref" HEAD
+./scripts/policy/check-changelog-pr.sh --content-only "$base_ref" HEAD
 ```

 `make coverage-gate` is an enforcement step, not a profile generator. For a
 standard PR, CI derives changed packages and their reverse-dependency test
 closure, then generates candidate and merge-base profiles with the same test
 scope and `coverpkg`. High-risk and protected-main runs use the complete
-profiles. Supporting and (when platform-selected) native profiles are
-generated before the aggregate `Coverage` context evaluates them. The
+profiles. The complete candidate profile is produced by disjoint per-shard
+helper jobs (`scripts/ci/test-packages.sh list-coverage`, kept serial with
+`-p 1` inside each shard; `verify` proves the shard union equals the
+full-suite scope exactly once) and concatenated in the aggregate job before
+enforcement. The complete merge-base profile is restored from an exact-key
+cache written by the last green `main` push of that same commit (key:
+merge-base SHA plus resolved Go version); any miss falls back to recomputing
+it in a merge-base worktree. The trusted `main` producer and PR consumer use
+the same dedicated cache profile path because GitHub includes that path in the
+cache version; the runtime-facing candidate and baseline filenames remain
+separate. Near-miss reuse is forbidden — the caches carry no prefix restore
+keys, because a neighbouring commit's profile would compare the candidate
+against the wrong baseline. Supporting and (when
+platform-selected) native profiles are generated before the aggregate
+`Coverage` context evaluates them. The
 aggregate and native gates require 100% coverage for changed executable Go
 statements. Overall coverage remains an unrounded, zero-tolerance,
 scope-matched merge-base non-regression check. Candidate and baseline profiles
@@ -186,11 +223,24 @@ are evaluated by the same block-deduplicating checker; supporting policy and
 shortcut profiles contribute to changed-code coverage only. The checked-in
 badge is presentation only and is never read as a gate input.

-Compatibility checks derive authoritative Interface snapshots from the PR
-merge-base and the latest reachable stable release. The candidate cannot bless
-a breaking change by editing a fixture. Schema additions are allowed;
-historical products, tools, parameters, mappings, positional execution fields,
-constraints, and safety semantics remain protected.
+CLI 兼容检查只使用 modern Interface Snapshot 这一处权威比较 seam，并从 PR
+merge-base 和最近的可达 stable release 生成权威快照。本治理机制合入后，
+merge-base 拥有生成器、比较器和已审批迁移清单，因此 candidate 不能通过修改
+helper、fixture 或在同一 PR 新增 self-approval 记录来放行 breaking change。首次
+bootstrap 仍由 merge-base 已有的 modern helper 做无豁免比较，并只接受 candidate
+提交中的规范空清单；完整边界见下方治理文档。
+
+精确的两阶段 flag 迁移生命周期见
+[CLI flag 兼容迁移治理](cli-interface-flag-migrations.md)。治理 PR 只能在
+surface 未变化时新增 `pending`；后续产品 PR 达到审批的精确 surface 后，才能
+消费 base-owned 记录并改为 `consumed`。在 main 与 stable 都达到 after 状态前
+必须保留该回执，之后再由单独 PR 清理。机制只放行记录中的 legacy
+visible-to-hidden，以及 canonical required 新增或提升；删除、type、scope、
+shorthand、no-opt 和任何无关漂移仍然阻塞。Schema 可以新增；历史 product、
+tool、parameter、mapping、positional execution、constraint 与 safety 语义继续
+受保护。`alias_of` 只是一项由 `FlagSpec.Aliases` 产生的框架关系证据，不是 payload
+等价证明；产品 PR 仍须证明 canonical 与 legacy 的最终运行 payload 等价并在 transport
+前拒绝冲突输入。当前迁移清单为空，不授权 PR #904。

 ## Required GitHub repository settings

@@ -0,0 +1,176 @@
+# CLI flag 兼容迁移治理
+
+本文定义一种受控迁移：保留旧 flag 的可执行兼容性，但把它从 Help 与 Agent Schema 中隐藏，并将新的规范 flag 设为唯一可见入口。迁移必须保持原 flag 的 requiredness：optional 只能迁到 optional，required 只能迁到 required。它只解决这一种精确变更，不是通用 breaking-change 豁免。
+
+同名 flag 的精确类型迁移属于另一类评审机制，只能进入
+`internal/interfacesnapshot/reviewed.go` 与 legacy smoke helper 的镜像表；flag rename
+只能进入本文的 JSON lifecycle ledger。一项迁移不得跨两种机制组合授权。
+
+## 唯一比较入口与信任边界
+
+PR 与本地兼容性审批的唯一权威比较入口是 modern Interface Snapshot：
+
+- `cmd/interface-snapshot` 生成和比较快照；
+- `internal/interfacesnapshot` 实现兼容规则和迁移生命周期；
+- `scripts/policy/check-command-compatibility.sh` 组装 candidate、PR merge-base 和最近可达且未撤回的 stable GA 三份快照；
+- `scripts/policy/check-authoritative-interface-baselines.sh` 只保留为 Makefile 的兼容包装，不再维护第二套判断逻辑。
+
+紧随其后的 Schema compatibility 不是第二份审批清单。它从同一 merge-base-owned
+ledger 和同一组三方 Interface Snapshot 取得已经完成 lifecycle 校验的
+authorization。merge-base-owned checker 会分别规范化 merge-base 与 stable 的完整
+Schema，并让 candidate 对两份历史 contract 独立执行检查；授权的 flag rename 只会
+精确投影到当前被检查的历史副本。candidate 不能为 CLI 与 Schema 分别提供两套例外。
+
+`make interface-integrity` 也调用上述 authoritative wrapper；默认 base 为
+`origin/main`，stable 可由包装脚本自动解析，candidate 默认为已提交的 `HEAD`。旧
+`scripts/policy/check-interface-baseline.sh` 只供
+`make update-interface-baseline` / `make reset-interface-baseline` 维护非权威 CLI
+Smoke fixture，不参与迁移审批。
+
+直接调用 `interface-snapshot compare` 时，只要提供 migration manifest 参数，就必须
+同时提供 `--base` 与 `--stable`；核心 lifecycle 也拒绝缺失 stable 的非空清单，避免
+调用方因漏传历史参考而提前清理 consumed receipt。
+
+PR merge-base 同时拥有快照生成器、比较器和已审批清单。门禁用这套 base-owned helper 检查同一个已提交 candidate revision、merge-base 与 stable，candidate 不能通过修改自己的 Go 比较 helper 来放宽规则。candidate 中的清单只参与迁移状态流转，不能批准同一个 PR 引入的接口变化。首次引入本机制时，merge-base 尚无迁移解析器；bootstrap 会用 merge-base 已有的 modern Interface Snapshot 做不带豁免的普通比较，并只接受 candidate 中逐字匹配的空清单，不会让 candidate 新增的 comparator 决定本 PR 是否兼容。bootstrap 无法让旧 helper 证明新治理实现本身正确，因此本治理 PR 的新 parser、lifecycle、launcher 与 hostile tests 仍是必须由真人评审的受保护策略变更；它们合入后才成为后续 PR 的 base-owned authority。
+
+这条边界保护比较规则和审批数据，不是任意代码沙箱。GitHub workflow / launcher 的变更仍由仓库保护规则和真人评审负责；candidate Cobra 构建也会执行 candidate 代码，因此对同一 runner 上的主动恶意代码，需要独立进程或文件系统隔离，不能把本门禁描述成已经解决。
+
+已审批清单固定为：
+
+```text
+scripts/policy/interface-migrations/approved-flag-migrations-v1.json
+```
+
+清单使用严格 JSON 解析：版本、字段名大小写、JSON 值类型、命令路径和 flag 名都必须精确；拒绝重复键、未知键、scalar `null` 与尾随 JSON 值，`reason` 不能为空；禁止 `*`、`?`、前缀规则或其他 wildcard。当前清单登记了 IM ID rename 的 `pending` 记录；`pending` 只记录已评审计划，候选与 merge-base 仍必须精确匹配 `before`，因此本治理 PR **不授权 PR #904 或任何产品接口变化**。
+
+## 两阶段迁移与回执清理
+
+每条迁移以 `(command, legacy flag, canonical flag)` 为唯一精确键，并经历以下生命周期：
+
+| 阶段 | PR 可以做什么 | 必须满足的快照状态 |
+|---|---|---|
+| 1. 治理审批 | 新增 `state: pending` 的精确记录；不得在同一个 PR 修改产品 surface | candidate 和 merge-base 都与记录中的 `before` 完全一致；该记录不改变 stable 的判断 |
+| 2. 产品迁移 | merge-base 已拥有 `pending` 后，按记录一次性切到精确 `after`，并把记录改为 `state: consumed` | legacy 仍存在但由 visible 变 hidden，且声明 `alias_of`；canonical 的 requiredness 与 legacy 迁移前完全一致 |
+| 3. 保留回执 | 产品 PR 合入后，如果 stable 仍是 `before`，继续保留 `consumed` | merge-base 或 stable 仍有任一份尚未达到 `after` |
+| 4. 单独清理 | 当 merge-base 和 stable 都已经是 `after`，在后续 PR 删除该记录 | 两份参考快照均精确匹配 `after`；继续保留过期回执会被门禁拒绝 |
+
+因此，新增 `pending` 和修改产品 surface 不能发生在同一个 PR；candidate 自己新增的记录不能 self-approve。迁移也不能部分执行：legacy、canonical、`alias_of` 或状态只要有一项不匹配，门禁即失败。
+
+下面只是清单结构示例，不代表已审批命令；实际字段必须从 Interface Snapshot 核对：
+
+```json
+{
+  "version": 1,
+  "migrations": [
+    {
+      "command": "dws chat message recall",
+      "legacy": {
+        "name": "msg-id",
+        "before": {
+          "present": true,
+          "type": "string",
+          "required": true,
+          "scope": "local"
+        },
+        "after": {
+          "present": true,
+          "type": "string",
+          "hidden": true,
+          "scope": "local",
+          "alias_of": "message-id"
+        }
+      },
+      "canonical": {
+        "name": "message-id",
+        "before": { "present": false },
+        "after": {
+          "present": true,
+          "type": "string",
+          "required": true,
+          "scope": "local"
+        }
+      },
+      "state": "pending",
+      "reason": "保留旧 argv 兼容性，并将规范 flag 设为唯一可见入口"
+    }
+  ]
+}
+```
+
+产品迁移 PR 必须保持同一条记录的命令、flag、before/after 和 reason 不变，只把 `pending` 改成 `consumed`。
+
+## `alias_of` 是框架来源的受评审关系证据
+
+`alias_of` 不是 Schema 同义词、参数概念词典或任意文字声明。它只能由 `FlagSpec.Aliases` 写入，并与内部 origin `corecmd.flag_spec_aliases.v1` 成对出现；每次 Interface Integrity 都会在已提交的 detached candidate 上执行源码门禁，禁止其他生产文件写入或复刻这些 evidence token。Interface Snapshot 会验证：
+
+- legacy 与 canonical 位于同一个可执行命令；
+- canonical flag 确实存在；
+- legacy 与 canonical 类型一致；
+- legacy 不是指向自身，也不存在 alias chain；
+- legacy 的 after 状态精确指向该记录中的 canonical flag。
+
+通过命令框架声明 `FlagSpec.Aliases` 时，框架会自动注册隐藏的兼容 flag，并写入两项 relation annotation；仅手写 `alias_of`、伪造 origin、重复值或不精确值都会让快照生成失败。不要用 Schema overlay、迁移清单或手写 Cobra annotation 伪造关系。
+
+这项关系证据只证明 legacy/canonical 经过受控框架路径建立关系，不证明最终 transport payload 等价，也不会替产品代码实现命令特有的值同步。当前框架还禁止把
+`MarkRequired` 与 `FlagSpec.Aliases` 直接组合，因为 Cobra 的 hard-required
+校验只识别 canonical spelling。若产品迁移同时需要 canonical 的 Cobra required
+标记和 legacy spelling，产品 PR 必须提供明确的运行时方案，并通过 canonical / legacy
+最终 payload 等价、同值输入一致、冲突输入在 transport 前失败、legacy 仍可调用但 Help 隐藏等测试；迁移清单和 relation evidence 都不能替代这些证明。
+
+## 豁免边界
+
+一条 base-owned、状态正确且前后快照精确匹配的记录，只会从普通兼容报告中移除以下两类预期 finding：
+
+1. legacy flag 的 `flag_became_hidden`（visible → hidden）；
+2. required legacy 被新增的 required canonical 替代时产生的 `required_flag_added`；如果 canonical 在 before 阶段只是 hidden 占位符，则允许它在转为公开拼写时继承 legacy 的 requiredness。已有的 visible canonical 不允许借 rename 改变 requiredness。
+
+以下变化仍按普通兼容规则阻塞，不能被迁移记录掩盖：
+
+- 删除 legacy、canonical、命令或其他 flag；
+- flag 类型或迁移记录中的 scope、shorthand、`no_opt` 漂移；
+- `alias_of` 缺失、指向变化或 alias chain；
+- 命令路径及任何无关的阻塞性接口变化；
+- 不精确、部分完成、超出记录范围的 surface 变化。
+
+## Schema 投影边界
+
+Agent-visible command 会把 visible Cobra flag 投影为 Schema parameter，因此合法的
+legacy hidden 迁移会同时表现为历史 parameter 消失，constraint member 也可能从
+legacy 名改为 canonical 名。Schema adapter 只接受已经由三方 Interface Snapshot
+判定为 authorized 的迁移，并按 tool 的精确 `primary_cli_path` 绑定：
+
+- reference 仍处于 `before` 且该 flag 有 Schema surface 时，baseline legacy parameter
+  必须存在，candidate legacy parameter 必须消失，candidate canonical parameter 必须存在；
+  如果 baseline 只有 canonical、没有 legacy，则 adapter 不得借 CLI ledger 提升
+  `required` / `cli_required` 或重写 constraint；
+- rename 前后的 `type`、`property`、`interface_type`、default、format、enum 与
+  `required_when` 必须完全一致；
+- `required` / `cli_required` 必须在 rename 前后完全一致，升高或降低都失败；
+- constraint 只允许在同一 tool 内按已枚举的 legacy → canonical map 做 member 替换、
+  排序与去重；group kind、非迁移 member 或 group 增删仍然阻塞；
+- 多个 legacy 指向同一 canonical 时，所有历史 parameter signature 必须一致，否则
+  fail closed。
+
+adapter 先构造经过上述验证的历史 contract 副本，再调用原 Schema checker；它不会按
+错误字符串删除 finding。这样既能处理纯 rename，也能阻止“旧 required 参数改名后意外
+变为 optional”或 property 漂移等伪兼容。`consumed` 回执在 merge-base Schema 已经处于
+canonical-only `after` 状态时不需要再次投影；adapter 保持 baseline 不变，由原 checker
+验证 candidate 是否仍与该 canonical contract 兼容。
+
+## 本地验证
+
+先确保 merge-base 和 stable tag 已在本地，然后运行与 CI 相同的权威门禁：
+
+```sh
+make interface-integrity \
+  BASE_REF=<merge-base> \
+  STABLE_REF=<stable-tag> \
+  CANDIDATE_REF=<candidate-sha>
+
+make schema-compatibility \
+  BASE_REF=<merge-base> \
+  STABLE_REF=<stable-tag> \
+  CANDIDATE_REF=<candidate-sha>
+```
+
+`STABLE_REF` 必须解析到从该 merge-base 可达的最高未撤回 stable GA tag；primary checker 会按 release contract 独立核对，不能用任意 after commit 或已撤回版本提前清理回执。包装脚本可以在省略时自动解析。`CANDIDATE_REF` 省略时固定为命令启动时的已提交 `HEAD`；评审和复现 CI 时应显式传入 candidate SHA，避免 surface 与清单来自不同 revision。
@@ -40,7 +40,7 @@ Every command inherits these flags (documented here once, not repeated per comma
 - [`dws doc` — DingTalk Doc](#dws-doc) · 21 commands
 - [`dws drive` — DingTalk Drive](#dws-drive) · 6 commands
 - [`dws minutes` — AI Minutes](#dws-minutes) · 19 commands
- [`dws oa` — OA Approval](#dws-oa) · 9 commands
+- [`dws oa` — OA Approval](#dws-oa) · 12 commands
 - [`dws report` — Reports](#dws-report) · 7 commands
 - [`dws todo` — Todo Tasks](#dws-todo) · 6 commands

@@ -277,14 +277,17 @@ _AI meeting notes: listing, summary, todos, transcription, recording control, mi

 ## `dws oa` — OA Approval

-_OA approval workflows: list, approve, reject, revoke, records._
+_OA approval workflows: inspect forms, forecast routes, create instances, approve, reject, revoke, and audit records._

-**9 commands**
+**12 commands**

 | Command | Description | When to use |
 |---|---|---|
 | `dws oa approval approve` | Approve a pending approval process instance (task) as the current user. | When the agent acts on a pending approval the user has delegated it to handle. |
+| `dws oa approval create-instance` | Create a real approval process instance from validated form values or a complete request payload. | After the agent has inspected the form Schema, forecast the route, resolved any selectable approvers, and obtained explicit user confirmation. |
 | `dws oa approval detail` | Retrieve full details of an approval process instance, including form fields, attachments, and state. | When the agent needs to read the content of an approval ticket before deciding on it or summarizing it. |
+| `dws oa approval form-schema` | Retrieve the form Schema for an approval template by processCode. | Before collecting or validating values for a new approval instance. |
+| `dws oa approval forecast-process` | Forecast the approval route for a template and its proposed form values. | Before creating an instance, especially when the route contains user-selectable approver or notifier nodes. |
 | `dws oa approval list-forms` | List approval process templates (forms) the current user is allowed to initiate. | When the agent needs to pick the right approval form before submitting a new request. |
 | `dws oa approval list-initiated` | List approval process instances the current user has initiated. | When the agent reviews the status of approvals the user submitted. |
 | `dws oa approval list-pending` | List approval process instances currently awaiting action from the current user. | When the agent surfaces "needs your approval" items in the user's inbox. |
@@ -1,7 +1,7 @@
 # dws dev 命令集 · Agent 人肉手工评测集（10 条复合用例）

 > 性质：**人肉手工评测集**——由测评人逐条手工跑、肉眼核对、人工判分，不是自动化脚本。
-> 用途：评测 agent（加载 `dingtalk-dev` 技能后）能否正确处理开放平台 dev 任务。
+> 用途：评测 agent（加载 `dingtalk-misc` 的 `references/devapp.md` 后）能否正确处理开放平台 dev 任务。
 > 特点：10 条**复合用例**，每条串多个子任务，一条覆盖一类完整场景；10 条合起来覆盖全部 34 个子命令 + 8 类横切行为。
 > 约定：所有命令应带 `--format json`；写操作应先 `--dry-run` 预览、用户确认后再 `--yes`；应用定位只用 `--unified-app-id`。

@@ -6,7 +6,7 @@

 ## 一键安装

-`dws dev` 能力已经合入主干并随正式版发布。专用安装脚本会下载预编译二进制 + `dingtalk-dev` skill，**只需要 curl + tar，不需要 git / go / make**。
+`dws dev` 能力已经合入主干并随正式版发布。专用安装脚本会下载预编译二进制 + `dingtalk-misc` skill（开放平台应用文档落在 misc），**只需要 curl + tar，不需要 git / go / make**。

 ### macOS / Linux

@@ -24,7 +24,7 @@ irm https://raw.githubusercontent.com/DingTalk-Real-AI/dingtalk-workspace-cli/ma

 1. 从 `DingTalk-Real-AI/dingtalk-workspace-cli` 的最新 Release 下载对应平台的预编译二进制。
 2. 安装 `dws` 到默认目录 `~/.local/bin`。
-3. 从 Release 的 skills 包里安装 `dingtalk-dev` skill 到本机已检测到的 Agent 目录。
+3. 从 Release 的 skills 包里安装 `dingtalk-misc` skill 到本机已检测到的 Agent 目录。

 支持这些环境变量（全部可选）：

@@ -33,7 +33,7 @@ irm https://raw.githubusercontent.com/DingTalk-Real-AI/dingtalk-workspace-cli/ma
 | `DEVAPP_REPO` | 覆盖发布仓库，默认 `DingTalk-Real-AI/dingtalk-workspace-cli` |
 | `DEVAPP_VERSION` | 钉某个 release tag，默认取最新 release |
 | `DWS_INSTALL_DIR` | 二进制安装目录，默认 `~/.local/bin` |
-| `DWS_NO_SKILLS` | 设为 `1` 跳过 `dingtalk-dev` skill 安装 |
+| `DWS_NO_SKILLS` | 设为 `1` 跳过 `dingtalk-misc` skill 安装 |

 > `dws dev` 已在正式版里，所以你也可以直接用标准安装脚本 `install.sh`，二者都会带上 `dws dev`。

@@ -0,0 +1,312 @@
+<!doctype html>
+<html lang="zh-CN">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport" content="width=device-width, initial-scale=1">
+  <meta name="description" content="DWS Drive Shortcut 与 lark-cli 的业务能力、真实数据 E2E 证据和平台边界分析。">
+  <title>Drive Shortcut 能力全景｜业务评审版</title>
+  <style>
+    :root {
+      color-scheme: light;
+      --paper: #f4f6f2; --surface: #fffefa; --ink: #17251f; --muted: #66746d;
+      --line: #dce2dc; --forest: #154f3d; --green: #17765a; --mint: #dff4e8;
+      --blue: #265f86; --blue-soft: #e7f1f7; --amber: #8c5a09; --amber-soft: #fff2cf;
+      --red: #a43b32; --red-soft: #fde9e5; --shadow: 0 14px 40px rgba(28, 48, 38, .08);
+    }
+    * { box-sizing: border-box; }
+    html { scroll-behavior: smooth; }
+    body { margin: 0; color: var(--ink); background: var(--paper); font: 15px/1.65 -apple-system, BlinkMacSystemFont, "Segoe UI", "PingFang SC", "Microsoft YaHei", sans-serif; }
+    a { color: inherit; text-decoration: none; }
+    code { padding: .12rem .38rem; border: 1px solid #d6e0da; border-radius: 6px; color: #174f3e; background: #f1f7f3; font: 600 .88em/1.4 ui-monospace, SFMono-Regular, Menlo, Consolas, monospace; white-space: nowrap; }
+    .wrap { width: min(1180px, calc(100% - 40px)); margin: auto; }
+    .hero { position: relative; overflow: hidden; padding: 64px 0 52px; color: #f7fff9; background: linear-gradient(125deg, #102b22 0%, #154c3c 57%, #1c6b54 100%); }
+    .hero::after { position: absolute; inset: -180px -100px auto auto; width: 540px; height: 540px; border: 1px solid rgba(255,255,255,.14); border-radius: 50%; box-shadow: 0 0 0 76px rgba(255,255,255,.035), 0 0 0 152px rgba(255,255,255,.025); content: ""; }
+    .hero-grid { position: relative; z-index: 1; display: grid; grid-template-columns: minmax(0, 1.35fr) minmax(300px, .65fr); gap: 32px; align-items: end; }
+    .eyebrow, .section-kicker { margin: 0 0 9px; color: #9fd6bd; font-size: 11px; font-weight: 900; letter-spacing: .16em; text-transform: uppercase; }
+    h1 { margin: 0; font-size: clamp(40px, 6vw, 68px); line-height: 1.03; letter-spacing: -.045em; }
+    .subtitle { max-width: 760px; margin: 19px 0 0; color: #d3e9de; font-size: 17px; }
+    .meta-row { display: flex; flex-wrap: wrap; gap: 8px; margin-top: 21px; }
+    .meta-pill { padding: 6px 10px; border: 1px solid rgba(255,255,255,.18); border-radius: 999px; color: #d5e9df; background: rgba(255,255,255,.07); font-size: 11px; font-weight: 750; }
+    .meta-pill.good { color: #bff2d5; border-color: rgba(139,232,179,.38); }
+    .hero-stats { display: grid; grid-template-columns: repeat(2, 1fr); gap: 10px; }
+    .hero-stat { padding: 17px 16px; border: 1px solid rgba(255,255,255,.17); border-radius: 15px; background: rgba(255,255,255,.075); backdrop-filter: blur(8px); }
+    .hero-stat strong { display: block; font-size: 30px; line-height: 1; }
+    .hero-stat span { display: block; margin-top: 7px; color: #cce2d7; font-size: 11px; }
+    .nav { position: sticky; top: 0; z-index: 20; border-bottom: 1px solid var(--line); background: rgba(255,254,250,.94); backdrop-filter: blur(12px); }
+    .nav .wrap { display: flex; overflow-x: auto; }
+    .nav a { flex: 0 0 auto; padding: 14px 15px; color: #5b6c64; font-size: 12px; font-weight: 800; }
+    .nav a:hover { color: var(--forest); background: #eaf3ee; }
+    main { padding: 38px 0 74px; }
+    section { margin-top: 50px; scroll-margin-top: 72px; }
+    section:first-child { margin-top: 0; }
+    .section-head { display: flex; justify-content: space-between; gap: 28px; align-items: end; margin-bottom: 19px; }
+    h2 { margin: 0; font-size: clamp(25px, 3.2vw, 36px); line-height: 1.16; letter-spacing: -.025em; }
+    h3 { margin: 0 0 7px; font-size: 18px; }
+    .section-desc { max-width: 660px; margin: 0; color: var(--muted); font-size: 13px; }
+    .callout { padding: 19px 21px; border: 1px solid #bdd8cb; border-left: 4px solid var(--green); border-radius: 13px; background: #ecf7f1; box-shadow: 0 6px 20px rgba(28,48,38,.04); }
+    .callout strong { color: #13513d; }
+    .callout.warn { border-color: #ead29a; border-left-color: #b67508; background: #fff8e7; }
+    .callout.warn strong { color: #784b00; }
+    .callout.danger { border-color: #e9b7b1; border-left-color: var(--red); background: var(--red-soft); }
+    .score-grid, .domain-grid, .evidence-grid, .review-grid { display: grid; gap: 13px; }
+    .score-grid { grid-template-columns: repeat(4, 1fr); margin-top: 16px; }
+    .score, .domain-card, .evidence-card, .review-card { border: 1px solid var(--line); border-radius: 15px; background: var(--surface); box-shadow: var(--shadow); }
+    .score { padding: 19px; }
+    .score strong { display: block; color: var(--forest); font-size: 29px; line-height: 1; }
+    .score span { display: block; margin-top: 8px; color: var(--muted); font-size: 12px; }
+    .domain-grid { grid-template-columns: repeat(4, 1fr); }
+    .domain-card { position: relative; padding: 21px; overflow: hidden; }
+    .domain-card .number { position: absolute; top: 12px; right: 17px; color: #d5e8de; font: 800 42px/1 ui-monospace, monospace; }
+    .domain-card p { min-height: 64px; margin: 8px 0 12px; color: var(--muted); font-size: 13px; }
+    .domain-card small { color: var(--green); font-weight: 800; }
+    .compare { overflow: hidden; border: 1px solid var(--line); border-radius: 16px; background: var(--surface); box-shadow: var(--shadow); }
+    .compare-top { display: grid; grid-template-columns: repeat(3, 1fr); }
+    .compare-column { padding: 21px; border-right: 1px solid var(--line); }
+    .compare-column:last-child { border-right: 0; }
+    .compare-column p, .compare-column li { color: var(--muted); font-size: 13px; }
+    .compare-column ul { margin: 9px 0 0; padding-left: 18px; }
+    .compare-column.covered { border-top: 5px solid var(--green); }
+    .compare-column.partial { border-top: 5px solid #c78b22; }
+    .compare-column.gap { border-top: 5px solid var(--red); }
+    .table-wrap { overflow-x: auto; }
+    table { width: 100%; border-collapse: collapse; }
+    th, td { padding: 12px 14px; border-bottom: 1px solid var(--line); text-align: left; vertical-align: top; }
+    th { color: #617069; background: #f7f8f5; font-size: 11px; font-weight: 900; letter-spacing: .03em; }
+    tr:last-child td { border-bottom: 0; }
+    tbody tr:hover { background: #f8fbf8; }
+    .verdict, .badge { display: inline-flex; align-items: center; padding: 3px 8px; border-radius: 999px; font-size: 10px; font-weight: 900; white-space: nowrap; }
+    .v-covered, .badge.read { color: #116045; background: var(--mint); }
+    .v-ahead, .badge.smart { color: #20577c; background: var(--blue-soft); }
+    .v-partial, .badge.write { color: #7b510a; background: var(--amber-soft); }
+    .v-gap, .badge.high { color: #8f3028; background: var(--red-soft); }
+    .truth-grid { display: grid; grid-template-columns: 1.1fr .9fr; gap: 14px; }
+    .truth-card { padding: 22px; border: 1px solid var(--line); border-radius: 15px; background: var(--surface); box-shadow: var(--shadow); }
+    .truth-step { display: grid; grid-template-columns: 30px 1fr; gap: 11px; margin-top: 13px; }
+    .truth-step b { display: grid; width: 28px; height: 28px; place-items: center; border-radius: 50%; color: #fff; background: var(--forest); font-size: 12px; }
+    .truth-step strong, .truth-step span { display: block; }
+    .truth-step span { color: var(--muted); font-size: 12px; }
+    .toolbar { display: grid; grid-template-columns: minmax(260px, 1fr) 180px 180px auto; gap: 10px; align-items: center; margin: 18px 0; padding: 13px; border: 1px solid var(--line); border-radius: 14px; background: var(--surface); }
+    input, select { width: 100%; min-height: 42px; padding: 9px 11px; border: 1px solid #ccd7d0; border-radius: 9px; color: var(--ink); background: #fff; font: inherit; }
+    input:focus, select:focus { outline: 3px solid rgba(23,118,90,.13); border-color: var(--green); }
+    .result-count { color: var(--muted); font-size: 12px; text-align: right; white-space: nowrap; }
+    .catalog { overflow: hidden; border: 1px solid var(--line); border-radius: 16px; background: var(--surface); box-shadow: var(--shadow); }
+    .shortcut-row { display: grid; grid-template-columns: 215px minmax(0, 1fr) 200px; gap: 16px; align-items: center; padding: 14px 17px; border-bottom: 1px solid var(--line); }
+    .shortcut-row:last-child { border-bottom: 0; }
+    .shortcut-row:hover { background: #f8fbf8; }
+    .command code { font-size: 12px; }
+    .row-main p { margin: 0; font-size: 13px; }
+    .row-main small { color: var(--muted); }
+    .badges { display: flex; justify-content: flex-end; flex-wrap: wrap; gap: 5px; }
+    .hidden-row { display: none; }
+    .evidence-grid { grid-template-columns: repeat(4, 1fr); }
+    .evidence-card { padding: 19px; }
+    .evidence-card strong { display: block; color: var(--forest); font-size: 23px; }
+    .evidence-card p { margin: 7px 0 0; color: var(--muted); font-size: 12px; }
+    .timeline { margin-top: 15px; border-left: 2px solid #bdd7ca; }
+    .event { position: relative; padding: 0 0 17px 22px; }
+    .event::before { position: absolute; left: -7px; top: 5px; width: 12px; height: 12px; border: 3px solid var(--paper); border-radius: 50%; background: var(--green); content: ""; }
+    .event b { display: block; }
+    .event span { color: var(--muted); font-size: 12px; }
+    .review-grid { grid-template-columns: repeat(3, 1fr); }
+    .review-card { position: relative; padding: 20px; }
+    .review-card .review-num { color: #b8d1c4; font: 800 12px/1 ui-monospace, monospace; letter-spacing: .1em; }
+    .review-card p { margin: 7px 0 0; color: var(--muted); font-size: 13px; }
+    footer { margin-top: 52px; padding: 23px 0; border-top: 1px solid var(--line); color: var(--muted); font-size: 11px; }
+    @media (max-width: 900px) { .hero-grid, .truth-grid { grid-template-columns: 1fr; } .score-grid, .domain-grid, .evidence-grid { grid-template-columns: repeat(2, 1fr); } .review-grid { grid-template-columns: 1fr 1fr; } .shortcut-row { grid-template-columns: 180px 1fr; } .badges { grid-column: 1 / -1; justify-content: flex-start; } }
+    @media (max-width: 620px) { .wrap { width: min(100% - 24px, 1180px); } .hero { padding: 44px 0 38px; } .hero-stats, .score-grid, .domain-grid, .evidence-grid, .review-grid, .compare-top { grid-template-columns: 1fr; } .compare-column { border-right: 0; border-bottom: 1px solid var(--line); } .toolbar { grid-template-columns: 1fr; } .result-count { text-align: left; } .shortcut-row { grid-template-columns: 1fr; } }
+    @media print { body { background: #fff; } .hero { color: var(--ink); background: #fff; border-bottom: 2px solid var(--ink); } .subtitle, .meta-pill, .hero-stat span { color: #425249; } .hero-stat { border-color: #aebbb3; } .nav, .toolbar { display: none; } .score, .domain-card, .compare, .truth-card, .catalog, .evidence-card, .review-card { box-shadow: none; break-inside: avoid; } }
+  </style>
+</head>
+<body>
+  <header class="hero">
+    <div class="wrap hero-grid">
+      <div>
+        <p class="eyebrow">Business Review · Drive</p>
+        <h1>Drive Shortcut<br>能力全景</h1>
+        <p class="subtitle">从 lark-cli 对齐出发，但不止于命令名：逐项审查输入、校验、多步编排、失败语义、真实字节和平台边界。</p>
+        <div class="meta-row">
+          <span class="meta-pill good">真实账号 E2E 已执行</span>
+          <span class="meta-pill">28 个公开入口</span>
+          <span class="meta-pill">统一 Result / Pagination</span>
+          <span class="meta-pill">报告已移除 PII / 凭证 / 业务正文</span>
+        </div>
+      </div>
+      <div class="hero-stats" aria-label="关键统计">
+        <div class="hero-stat"><strong>38</strong><span>lark-cli Drive 逐项审查</span></div>
+        <div class="hero-stat"><strong>26</strong><span>已覆盖或跨产品路由</span></div>
+        <div class="hero-stat"><strong>7</strong><span>部分对齐，边界已公开</span></div>
+        <div class="hero-stat"><strong>5</strong><span>客观不可对齐能力</span></div>
+      </div>
+    </div>
+  </header>
+
+  <nav class="nav"><div class="wrap"><a href="#overview">全景</a><a href="#compare">Lark 对齐</a><a href="#ahead">超越项</a><a href="#truth">真实语义</a><a href="#catalog">完整目录</a><a href="#e2e">E2E</a><a href="#review">评审</a></div></nav>
+
+  <main class="wrap">
+    <section id="overview">
+      <div class="section-head"><div><p class="section-kicker">Executive summary</p><h2>28 个公开入口，覆盖文件完整生命周期</h2></div><p class="section-desc">另有 <code>+publish-set</code> 已实现契约和读回逻辑，但真实普通文件与在线文档均被服务端拒绝，因此保持 unavailable，不进入 Agent 公开目录。</p></div>
+      <div class="callout"><strong>结论：</strong>Drive 已从 9 个偏原子入口扩展为 28 个可发现 Shortcut。它不仅补齐 Lark 的核心文件、版本和状态任务，还通过严格响应合同、真实落盘、写后读回、回收恢复和个人收藏形成更可审计的钉盘工作流。</div>
+      <div class="score-grid">
+        <article class="score"><strong>29</strong><span>已审查注册项（含 1 unavailable）</span></article>
+        <article class="score"><strong>25</strong><span>公开主能力 / 语义适配</span></article>
+        <article class="score"><strong>3</strong><span>公开兼容入口</span></article>
+        <article class="score"><strong>3</strong><span>高风险写入口，均需确认</span></article>
+      </div>
+    </section>
+
+    <section>
+      <div class="section-head"><div><p class="section-kicker">Capability map</p><h2>四个业务域</h2></div><p class="section-desc">目录按用户任务组织；兼容命令不重复计为新增能力。</p></div>
+      <div class="domain-grid">
+        <article class="domain-card"><span class="number">09</span><h3>发现与检查</h3><p>严格目录分页、搜索、最近访问，以及元数据、统计和封面聚合检查。</p><small>+list · +search · +recent · +inspect</small></article>
+        <article class="domain-card"><span class="number">09</span><h3>文件生命周期</h3><p>创建目录、上传下载、快捷方式、在线对象复制、移动重命名、删除与恢复。</p><small>+upload · +download · +rename · +recycle-restore</small></article>
+        <article class="domain-card"><span class="number">06</span><h3>个人与公开状态</h3><p>回收站清单、收藏闭环和互联网公开状态的独立安全域。</p><small>+star-list · +star-add · +publish-get</small></article>
+        <article class="domain-card"><span class="number">04</span><h3>历史版本</h3><p>版本列表、精确定位、真实字节下载与高风险回滚读回。</p><small>+version-history · +version-download · +version-revert</small></article>
+      </div>
+    </section>
+
+    <section id="compare">
+      <div class="section-head"><div><p class="section-kicker">Lark alignment</p><h2>对齐业务语义，不追求同名率</h2></div><p class="section-desc">38 项逐项核对。评论、导入导出和成员权限在 DWS 由更成熟的 Doc 或原子权限入口承接，不在 Drive 再复制一套。</p></div>
+      <div class="compare">
+        <div class="compare-top">
+          <article class="compare-column covered"><h3>26 · 已覆盖 / 路由</h3><p>上传下载、目录与快捷方式、版本、移动删除、状态、搜索、评论、导入导出和成员任务均有真实入口。</p></article>
+          <article class="compare-column partial"><h3>7 · 部分对齐</h3><p>预览、resolve/reaction、push/pull、权限申请与 setting 受对象模型或接口粒度约束，明确保留有限语义。</p></article>
+          <article class="compare-column gap"><h3>5 · 客观缺口</h3><p>删除评论恢复、普通文件版本删除、安全标签读写、可靠双向目录同步缺少必要下层接口。</p></article>
+        </div>
+        <div class="table-wrap"><table><thead><tr><th>Lark 任务组</th><th>DWS 主路径</th><th>结论</th><th>关键差异与交付决定</th></tr></thead><tbody>
+          <tr><td>upload / folder / shortcut / download</td><td><code>drive +upload</code> 等</td><td><span class="verdict v-ahead">增强</span></td><td>工作目录边界、OSS PUT、严格 commit、no-clobber、原子落盘、非零字节和读回验证。</td></tr>
+          <tr><td>preview / cover</td><td><code>drive +cover</code></td><td><span class="verdict v-partial">部分</span></td><td>封面/缩略图可读；没有等价的服务端多格式预览转换，不扩大宣称。</td></tr>
+          <tr><td>comments / replies</td><td><code>doc +comment-*</code> / <code>doc +review</code></td><td><span class="verdict v-covered">路由</span></td><td>评论归在线文档协作域；独立 resolve、reaction identity 与删除后恢复仍受接口限制。</td></tr>
+          <tr><td>export / import / task result</td><td><code>doc +export</code> / <code>doc +import</code></td><td><span class="verdict v-ahead">增强</span></td><td>提交、轮询、恢复、安全下载形成类型化闭环，不保留泛化下划线命令。</td></tr>
+          <tr><td>version history / get / revert</td><td><code>drive +version-*</code></td><td><span class="verdict v-ahead">增强</span></td><td>严格分页、精确版本、历史字节落盘、回滚前预检与终态读回；历史版本删除无接口。</td></tr>
+          <tr><td>status / inspect</td><td><code>drive +inspect</code></td><td><span class="verdict v-ahead">超越</span></td><td>元数据为必达结果，统计、公开状态和封面按需 fan-out；可选失败为 partial_success。</td></tr>
+          <tr><td>push / pull / sync</td><td><code>+upload</code> / <code>+download</code> 单文件</td><td><span class="verdict v-partial">部分</span></td><td>不在缺少稳定 hash、rename/delete journal 和冲突向量时制造危险目录同步。</td></tr>
+          <tr><td>member / permission</td><td><code>doc +access-*</code> / <code>drive permission</code></td><td><span class="verdict v-covered">路由</span></td><td>协作者权限与互联网公开是两个安全域；申请权限需真实上下文，未伪装为通用 Shortcut。</td></tr>
+          <tr><td>secure labels</td><td>无等价</td><td><span class="verdict v-gap">缺口</span></td><td>当前 DWS/钉钉下层没有 Drive 安全标签目录和写入接口，不能用普通权限代替。</td></tr>
+          <tr><td>search</td><td><code>drive +search</code> / <code>doc +search</code></td><td><span class="verdict v-ahead">增强</span></td><td>文件与在线文档按域路由；文件搜索严格验证数组、过滤和分页。</td></tr>
+        </tbody></table></div>
+      </div>
+      <div class="callout warn" style="margin-top:14px"><strong>普通文件 copy 边界：</strong>钉钉现有复制接口对普通文件产生 <code>.dlink</code>，不是字节独立副本。因此 <code>+copy</code> 只接受在线对象；普通文件快捷入口用 <code>+create-shortcut</code>，独立副本使用 <code>+download</code> 后 <code>+upload</code>。</div>
+    </section>
+
+    <section id="ahead">
+      <div class="section-head"><div><p class="section-kicker">Beyond parity</p><h2>DWS 可主推的八个差异化点</h2></div><p class="section-desc">价值来自正确性与完整闭环，而不是额外注册同义命令。</p></div>
+      <div class="domain-grid">
+        <article class="domain-card"><h3>严格目录语义</h3><p><code>+list</code> / <code>+recent</code> 只有服务端明确返回数组时才接受空集合。</p><small>缺字段 ≠ 空目录</small></article>
+        <article class="domain-card"><h3>聚合检查</h3><p><code>+inspect</code> 一次汇总身份、统计、公开状态和封面，并保留局部失败。</p><small>partial_success 可审计</small></article>
+        <article class="domain-card"><h3>真实文件传输</h3><p>上传执行完整事务；下载验证受控路径、覆盖策略、原子发布和字节。</p><small>不是只返回临时 URL</small></article>
+        <article class="domain-card"><h3>回收恢复闭环</h3><p>从 <code>recycleItemId</code> 恢复后读取真实节点，证明资源确实回到可访问状态。</p><small>恢复后读回</small></article>
+        <article class="domain-card"><h3>个人收藏闭环</h3><p>收藏、列表、取消收藏覆盖完整用户偏好过程，并保留分页。</p><small>add → list → remove</small></article>
+        <article class="domain-card"><h3>版本真实字节</h3><p>除元数据外可下载任意已知历史版本，并用本地字节核验回滚结果。</p><small>version-download</small></article>
+        <article class="domain-card"><h3>重命名终态</h3><p>处理服务端扩展名规则，再读取节点确认最终名称，避免重复扩展名。</p><small>write → read-back</small></article>
+        <article class="domain-card"><h3>公开域诚实降级</h3><p>查询和关闭可验证；开启在 eligible 节点闭环完成前保持 unavailable。</p><small>不把 notSupported 当成功</small></article>
+      </div>
+    </section>
+
+    <section id="truth">
+      <div class="section-head"><div><p class="section-kicker">Truthful execution</p><h2>空数组不再是“看起来成功”</h2></div><p class="section-desc">合法业务空集合可以成功，但必须先证明响应结构、元素类型和分页语义成立。内部错误、缺字段与坏投影必须失败。</p></div>
+      <div class="truth-grid">
+        <article class="truth-card"><h3>四层成功证据</h3>
+          <div class="truth-step"><b>1</b><div><strong>传输成功</strong><span>进程成功，MCP / HTTP 没有显式错误。</span></div></div>
+          <div class="truth-step"><b>2</b><div><strong>响应合同</strong><span>对象、数组、success 标志和元素类型与命令声明一致。</span></div></div>
+          <div class="truth-step"><b>3</b><div><strong>业务终态</strong><span>写命令必须获得新 ID、终态证据或后续元数据读回。</span></div></div>
+          <div class="truth-step"><b>4</b><div><strong>产物校验</strong><span>下载必须落盘、非零字节；关键链路比较大小和 SHA-256。</span></div></div>
+        </article>
+        <article class="truth-card"><h3>明确失败的情况</h3>
+          <ul><li>空响应、缺少预期集合字段或集合类型错误。</li><li>集合存在坏元素，不能投影时静默丢弃。</li><li><code>success=false</code>、写响应没有 ID 或读回不一致。</li><li>inspect 的可选分支失败却返回整体 success。</li><li>下载得到空文件、越界路径或覆盖既有文件。</li><li>普通文件 copy 返回快捷链接却声称独立副本。</li></ul>
+        </article>
+      </div>
+    </section>
+
+    <section id="catalog">
+      <div class="section-head"><div><p class="section-kicker">Full catalog</p><h2>28 个公开 Shortcut 完整目录</h2></div><p class="section-desc">16 个只读、9 个普通写、3 个高风险写；3 个历史入口保留兼容但不作为新 Agent 主路径。</p></div>
+      <div class="toolbar"><input id="q" type="search" placeholder="搜索命令或用途，例如 版本、回收、+inspect…" aria-label="搜索 Shortcut"><select id="domain"><option value="all">全部业务域</option><option value="discover">发现与检查</option><option value="lifecycle">文件生命周期</option><option value="personal">个人与公开</option><option value="version">历史版本</option></select><select id="risk"><option value="all">全部风险</option><option value="read">只读</option><option value="write">普通写</option><option value="high">高风险写</option></select><span id="result-count" class="result-count">显示 28 / 28</span></div>
+      <div class="catalog">
+        <article class="shortcut-row" data-tool data-domain="discover" data-risk="read"><div class="command"><code>dws drive +list</code></div><div class="row-main"><p>严格分页列出目录，保留游标，区分显式空目录和畸形响应。</p><small>发现与检查</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="discover" data-risk="read"><div class="command"><code>dws drive +inspect</code></div><div class="row-main"><p>聚合元数据与可选统计、公开状态、封面；局部失败如实报告。</p><small>发现与检查</small></div><div class="badges"><span class="badge read">READ</span><span class="badge smart">SMART</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="discover" data-risk="read"><div class="command"><code>dws drive +info</code></div><div class="row-main"><p>历史元数据兼容入口；新场景优先使用 +inspect。</p><small>兼容入口</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="discover" data-risk="read"><div class="command"><code>dws drive +search</code></div><div class="row-main"><p>按关键词、类型、扩展名、创建人、时间和分页搜索钉盘文件。</p><small>发现与检查</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="discover" data-risk="read"><div class="command"><code>dws drive +find-file</code></div><div class="row-main"><p>历史文件搜索兼容入口；新场景优先使用 +search。</p><small>兼容入口</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="discover" data-risk="read"><div class="command"><code>dws drive +search-docs</code></div><div class="row-main"><p>历史跨域搜索入口；新的在线文档搜索路由 doc +search。</p><small>兼容入口</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="discover" data-risk="read"><div class="command"><code>dws drive +recent</code></div><div class="row-main"><p>读取最近访问或编辑列表，支持创建人筛选并保留分页。</p><small>发现与检查</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="discover" data-risk="read"><div class="command"><code>dws drive +stats</code></div><div class="row-main"><p>读取访问、编辑、评论、点赞、预览和下载统计。</p><small>发现与检查</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="discover" data-risk="read"><div class="command"><code>dws drive +cover</code></div><div class="row-main"><p>读取封面或缩略图；不宣称服务端多格式预览。</p><small>发现与检查</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+
+        <article class="shortcut-row" data-tool data-domain="lifecycle" data-risk="write"><div class="command"><code>dws drive +upload</code></div><div class="row-main"><p>上传凭证、OSS PUT、严格提交和远端元数据读回的一体化事务。</p><small>文件生命周期</small></div><div class="badges"><span class="badge write">WRITE · CONFIRM</span><span class="badge smart">SMART</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="lifecycle" data-risk="read"><div class="command"><code>dws drive +download</code></div><div class="row-main"><p>安全落盘、no-clobber、原子发布并验证非零字节。</p><small>文件生命周期</small></div><div class="badges"><span class="badge read">READ</span><span class="badge smart">SMART</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="lifecycle" data-risk="write"><div class="command"><code>dws drive +create-folder</code></div><div class="row-main"><p>创建文件夹后要求新 ID，并读回名称验证。</p><small>文件生命周期</small></div><div class="badges"><span class="badge write">WRITE · CONFIRM</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="lifecycle" data-risk="write"><div class="command"><code>dws drive +create-shortcut</code></div><div class="row-main"><p>创建快捷方式并读回，明确区别于独立副本。</p><small>文件生命周期</small></div><div class="badges"><span class="badge write">WRITE · CONFIRM</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="lifecycle" data-risk="write"><div class="command"><code>dws drive +copy</code></div><div class="row-main"><p>复制在线对象；普通文件预检拒绝，避免把 .dlink 当副本。</p><small>文件生命周期</small></div><div class="badges"><span class="badge write">WRITE · CONFIRM</span><span class="badge smart">SMART</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="lifecycle" data-risk="write"><div class="command"><code>dws drive +move</code></div><div class="row-main"><p>移动到指定文件夹或知识库位置，语义与 copy/shortcut 消歧。</p><small>文件生命周期</small></div><div class="badges"><span class="badge write">WRITE · CONFIRM</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="lifecycle" data-risk="write"><div class="command"><code>dws drive +rename</code></div><div class="row-main"><p>重命名后读取真实节点，验证最终名称和扩展名。</p><small>文件生命周期</small></div><div class="badges"><span class="badge write">WRITE · CONFIRM</span><span class="badge smart">SMART</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="lifecycle" data-risk="high"><div class="command"><code>dws drive +delete</code></div><div class="row-main"><p>将确认过的节点移入回收站，要求 success=true 终态证据。</p><small>文件生命周期</small></div><div class="badges"><span class="badge high">HIGH · CONFIRM</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="lifecycle" data-risk="write"><div class="command"><code>dws drive +recycle-restore</code></div><div class="row-main"><p>按回收项 ID 恢复，并读回恢复后的节点。</p><small>文件生命周期</small></div><div class="badges"><span class="badge write">WRITE · CONFIRM</span><span class="badge smart">SMART</span></div></article>
+
+        <article class="shortcut-row" data-tool data-domain="personal" data-risk="read"><div class="command"><code>dws drive +recycle-list</code></div><div class="row-main"><p>严格分页列出回收项并稳定投影 recycleItemId。</p><small>个人与公开</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="personal" data-risk="read"><div class="command"><code>dws drive +star-list</code></div><div class="row-main"><p>严格分页列出当前用户收藏并保留游标。</p><small>个人与公开</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="personal" data-risk="write"><div class="command"><code>dws drive +star-add</code></div><div class="row-main"><p>以幂等用户偏好语义收藏指定节点。</p><small>个人与公开</small></div><div class="badges"><span class="badge write">WRITE</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="personal" data-risk="write"><div class="command"><code>dws drive +star-remove</code></div><div class="row-main"><p>以幂等用户偏好语义取消收藏指定节点。</p><small>个人与公开</small></div><div class="badges"><span class="badge write">WRITE</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="personal" data-risk="read"><div class="command"><code>dws drive +publish-get</code></div><div class="row-main"><p>只读查询互联网公开状态，不沿用错误的写风险标签。</p><small>个人与公开</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="personal" data-risk="high"><div class="command"><code>dws drive +publish-unset</code></div><div class="row-main"><p>关闭互联网公开并读回验证外链状态。</p><small>个人与公开</small></div><div class="badges"><span class="badge high">HIGH · CONFIRM</span><span class="badge smart">SMART</span></div></article>
+
+        <article class="shortcut-row" data-tool data-domain="version" data-risk="read"><div class="command"><code>dws drive +version-history</code></div><div class="row-main"><p>严格分页列出普通文件历史版本。</p><small>历史版本</small></div><div class="badges"><span class="badge read">READ</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="version" data-risk="read"><div class="command"><code>dws drive +version-get</code></div><div class="row-main"><p>按正整数版本号精确匹配，零命中显式失败。</p><small>历史版本</small></div><div class="badges"><span class="badge read">READ</span><span class="badge smart">SMART</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="version" data-risk="read"><div class="command"><code>dws drive +version-download</code></div><div class="row-main"><p>预检版本后安全下载历史字节，要求非零产物。</p><small>历史版本</small></div><div class="badges"><span class="badge read">READ</span><span class="badge smart">SMART</span></div></article>
+        <article class="shortcut-row" data-tool data-domain="version" data-risk="high"><div class="command"><code>dws drive +version-revert</code></div><div class="row-main"><p>验证版本存在后回滚，并读取当前节点终态。</p><small>历史版本</small></div><div class="badges"><span class="badge high">HIGH · CONFIRM</span><span class="badge smart">SMART</span></div></article>
+      </div>
+      <div class="callout warn" style="margin-top:14px"><strong>未公开入口：</strong><code>+publish-set</code> 的安全契约和 set→get 读回代码已存在，但真实后端返回 <code>operation.notSupported</code>。在找到 eligible 节点并完成 set→get→unset 闭环前，不进入公开 Agent catalog。</div>
+    </section>
+
+    <section id="e2e">
+      <div class="section-head"><div><p class="section-kicker">Real-data E2E</p><h2>真实数据验证，不用空结果证明成功</h2></div><p class="section-desc">测试在隔离目录创建临时资源，覆盖读取、写入、下载、版本、收藏、回收和清理。资源 ID、账号、URL、上传凭证、绝对路径和业务正文均未进入报告。</p></div>
+      <div class="evidence-grid">
+        <article class="evidence-card"><strong>39,838 B</strong><p>真实文件上传后下载字节数；与源文件 SHA-256 完全一致。</p></article>
+        <article class="evidence-card"><strong>2 versions</strong><p>覆盖写入生成两个版本；精确查询、历史下载和回滚全部读回。</p></article>
+        <article class="evidence-card"><strong>4 / 5</strong><p>隔离夹具中搜索命中 4 项、最近列表命中 5 项，证明非空投影链路。</p></article>
+        <article class="evidence-card"><strong>0 remain</strong><p>测试结束后隔离根目录无残留；临时资源进入回收站并完成本地清理。</p></article>
+      </div>
+      <div class="timeline">
+        <div class="event"><b>创建与发现</b><span>创建两个隔离目录并读回；+list 命中真实节点，由此发现并修复 dentryId 与 32 字符 fileId 混用。</span></div>
+        <div class="event"><b>上传与下载</b><span>真实 OSS 上传、远端元数据读回、下载、no-clobber 二次路径、大小与 SHA-256 一致性全部通过。</span></div>
+        <div class="event"><b>检查与个人状态</b><span>+inspect（含 stats / publish / cover）、+stats、+cover、收藏 add→list→remove 通过。</span></div>
+        <div class="event"><b>版本闭环</b><span>覆盖文件产生两个版本；history/get/download/revert 通过，回滚后最新字节与原始内容一致。</span></div>
+        <div class="event"><b>复制、移动与命名</b><span>在线文档 copy 通过；普通文件 .dlink 被修正为预检拒绝；move 往返、rename 扩展名规范化通过。</span></div>
+        <div class="event"><b>删除与恢复</b><span>delete→recycle-list→recycle-restore 通过，真实回收响应字段已按后端形态修正。</span></div>
+        <div class="event"><b>平台负向证据</b><span>publish-set 对普通文件和在线文档均明确返回不支持，因此保持 unavailable；没有把失败改写成空对象成功。</span></div>
+        <div class="event"><b>清理</b><span>隔离目录进入回收站，根目录残留计数为零；本地下载产物删除。</span></div>
+      </div>
+    </section>
+
+    <section id="review">
+      <div class="section-head"><div><p class="section-kicker">Review prompts</p><h2>建议业务评审重点确认</h2></div><p class="section-desc">这些是需要接受的产品边界，不是被空结果遮蔽的实现问题。</p></div>
+      <div class="review-grid">
+        <article class="review-card"><span class="review-num">01</span><h3>是否接受 26 / 7 / 5 结论？</h3><p>按用户任务计覆盖、部分与缺口，不用同名命令数量代替语义保真。</p></article>
+        <article class="review-card"><span class="review-num">02</span><h3>普通文件 copy 是否足够清晰？</h3><p>服务端无法提供原子独立副本；快捷方式与下载后上传两条替代路径已明确。</p></article>
+        <article class="review-card"><span class="review-num">03</span><h3>是否拒绝不可靠目录 sync？</h3><p>缺稳定 hash、删除/重命名日志和冲突向量时，不发布可能覆盖数据的双向同步。</p></article>
+        <article class="review-card"><span class="review-num">04</span><h3>跨产品路由是否合理？</h3><p>评论、导入导出和协作者权限优先复用 Doc 成熟入口，不在 Drive 制造同义表面。</p></article>
+        <article class="review-card"><span class="review-num">05</span><h3>publish-set 是否继续 unavailable？</h3><p>建议维持，直到真实 eligible 节点完成开启、查询、关闭的可恢复闭环。</p></article>
+        <article class="review-card"><span class="review-num">06</span><h3>下一批后端解锁优先级？</h3><p>建议依次评估普通文件原子 copy、同步所需版本信号、安全标签和评论恢复接口。</p></article>
+      </div>
+    </section>
+  </main>
+
+  <footer><div class="wrap">依据：DWS 最终 Shortcut catalog / Schema、Drive 实现与测试、真实账号 E2E、lark-cli Drive registrations 与实现。范围仅含 Drive Shortcut 及必要跨产品路由；不包含原子命令总表。所有业务标识、凭证、签名 URL、用户信息和正文均已脱敏。</div></footer>
+  <script>
+    const q = document.querySelector('#q');
+    const domain = document.querySelector('#domain');
+    const risk = document.querySelector('#risk');
+    const rows = [...document.querySelectorAll('[data-tool]')];
+    const count = document.querySelector('#result-count');
+    function filterTools() {
+      const needle = q.value.trim().toLocaleLowerCase('zh-CN');
+      let visible = 0;
+      rows.forEach((row) => {
+        const show = (!needle || row.textContent.toLocaleLowerCase('zh-CN').includes(needle)) && (domain.value === 'all' || row.dataset.domain === domain.value) && (risk.value === 'all' || row.dataset.risk === risk.value);
+        row.classList.toggle('hidden-row', !show);
+        if (show) visible += 1;
+      });
+      count.textContent = `显示 ${visible} / ${rows.length}`;
+    }
+    [q, domain, risk].forEach((control) => control.addEventListener('input', filterTools));
+  </script>
+</body>
+</html>
@@ -0,0 +1,82 @@
+# Drive Shortcut 对齐与超越 Lark CLI
+
+## 目标与判定口径
+
+本轮以 Lark CLI `drive` 的 38 个 shortcut 为对照，但不把“同名命令数量”当完成标准。对齐按用户任务判定：
+
+1. `drive +...` 有更稳定的 Agent 主入口时，提供 Shortcut，并发布 Selection、Safety、Result 与 Pagination。
+2. 钉钉已经在其他产品提供更成熟入口时，Skill 明确跨产品路由，不在 Drive 重复实现。
+3. 只有原子能力且 Shortcut 不增加校验、编排或投影价值时，保留 Runtime Schema leaf，不制造同义别名。
+4. 下层接口不存在或无法满足相同语义时，明确记录 gap；不得用空数组、空对象或只返回任务提交结果伪装完成。
+
+成功判定统一为：进程成功 + 统一结果 `ok=true/outcome=success` + 必要业务字段 + 真实数据读回或本地字节校验。服务端显式返回空数组可以是合法业务空结果；空响应、缺少数组、数组类型错误、坏元素、`success=false`、写入缺少终态证据都必须失败。
+
+## Lark 38 项映射
+
+| Lark Drive shortcut | DWS 路由 | 结论与原因 |
+|---|---|---|
+| `+upload` | `drive +upload` | 对齐并增强：工作目录边界、OSS PUT、严格 commit、元数据读回。 |
+| `+create-folder` | `drive +create-folder` | 对齐并增强：要求新 fileId 和名称读回。 |
+| `+create-shortcut` | `drive +create-shortcut` | 对齐并增强：明确 shortcut≠copy，创建后读回。 |
+| `+download` | `drive +download` | 对齐并增强：真实落盘、no-clobber、原子发布、非零字节。 |
+| `+preview` | `drive +cover`（有限） | 不完全对齐：钉钉当前只提供封面/缩略图读取，没有等价的服务端多格式预览转换接口。 |
+| `+cover` | `drive +cover` | 对齐：严格读取封面/缩略图对象。 |
+| `+add-comment` | `doc +comment-create` | 用户任务对齐；评论归在线文档协作域，Drive 不复制一套。 |
+| `+list-comments` | `doc +comment-list` | 用户任务对齐；Doc 已有类型、状态与分页。 |
+| `+batch-query-comments` | `doc +review` / `doc +comment-list` | 超越：可聚合未解决评论与确定性正文上下文；跨文档批量仍由调用方按节点编排。 |
+| `+resolve-comment` | `doc +comment-update`（有限） | 部分对齐：DWS 可更新评论，但当前下层未声明独立 resolve 状态接口。 |
+| `+restore-comment` | 无等价 | gap：钉钉当前下层未暴露恢复已删除评论的等价能力。 |
+| `+add-reply` | `doc +comment-reply` | 对齐。 |
+| `+list-replies` | `doc +comment-list` | 用户任务对齐：评论列表返回回复上下文；无独立 Drive reply 目录。 |
+| `+update-reply` | `doc +comment-update` | 对齐到评论/回复统一更新语义。 |
+| `+delete-reply` | `doc +comment-delete` | 对齐到评论/回复统一删除语义，高风险确认。 |
+| `+react-reply` | `doc +comment-reply`（有限） | 部分对齐：支持表情回复；不声称拥有 Lark 的独立 reaction identity。 |
+| `+export` | `doc +export` | 用户任务对齐并增强：提交、轮询、安全下载一体化。 |
+| `+export-download` | `doc +export` / `doc +export-get` | 超越：常规一体化，`+export-get` 仅作中断恢复。 |
+| `+import` | `doc +import` | 用户任务对齐并增强：转换白名单、上传 fallback、轮询终态。 |
+| `+version-history` | `drive +version-history` | 对齐并增强：严格空结果与分页。 |
+| `+version-get` | `drive +version-get` | 对齐并增强：精确版本号，零命中失败。 |
+| `+version-revert` | `drive +version-revert` | 对齐并增强：版本预检、高风险确认、节点读回。 |
+| `+version-delete` | 无等价 | gap：钉钉当前普通文件版本接口没有删除历史版本能力。 |
+| `+move` | `drive +move` | 对齐；与 copy/shortcut 明确消歧并发布确认。 |
+| `+delete` | `drive +delete` | 对齐；移入回收站、高风险确认、终态证据。 |
+| `+status` | `drive +inspect` | 超越：远端身份、统计、公开状态和封面按需聚合；不伪装成本地同步状态。 |
+| `+push` | `drive +upload`（单文件） | 部分对齐：单文件上传可靠；没有可靠的目录 diff、冲突和远端删除传播语义，因此不提供同名批量 push。 |
+| `+pull` | `drive +download`（单文件） | 部分对齐：单文件下载可靠；目录级增量拉取需稳定路径、hash 与冲突策略，当前接口不完整。 |
+| `+sync` | 无等价 | gap：在缺少稳定远端内容 hash、rename/delete journal 和冲突版本向量时，双向同步会有数据覆盖风险。 |
+| `+task_result` | `doc +export-get` / 导入任务恢复入口 | 用户任务对齐；DWS 按任务所属产品提供类型化恢复入口，不保留 Lark 的下划线泛化命令。 |
+| `+apply-permission` | `drive permission apply` raw leaf | 下层能力存在但未提升为 Shortcut：需要真实申请上下文和权限夹具，无法在通用 E2E 中安全创建。 |
+| `+member-add` | `doc +access-grant` | 用户任务对齐并增强：解析接收人、批量 ledger、首次写入前停止。 |
+| `+member-list` | `drive permission list` / `doc +inspect --include-permissions` | 对齐；常规 Agent 场景优先 Doc 聚合检查。 |
+| `+permission-get-setting` | `drive permission list` + `drive +publish-get` | 部分对齐：协作者与互联网公开是两个独立安全域，没有一个与 Lark setting 完全同构的钉钉接口。 |
+| `+secure-label-list` | 无等价 | gap：当前 DWS/钉钉下层没有可声明的 Drive 安全标签目录接口。 |
+| `+secure-label-update` | 无等价 | gap：没有安全标签写接口，不能用普通权限或公开状态替代。 |
+| `+search` | `drive +search` | 对齐并增强：过滤、严格数组和分页；在线文档搜索路由 `doc +search`。 |
+| `+inspect` | `drive +inspect` | 对齐并增强：必达元数据 + 可选聚合，部分失败不伪装成功。 |
+
+## DWS 超出 Lark Drive 的可挖掘能力
+
+- `+list`：严格目录分页，而不是把缺字段当空目录。
+- `+recent`：最近访问/编辑与创建人筛选。
+- `+stats`：阅读、编辑、评论、点赞、预览和下载统计。
+- `+recycle-list` / `+recycle-restore`：显式回收项身份与恢复后读回。
+- `+star-list` / `+star-add` / `+star-remove`：个人收藏完整闭环。
+- `+publish-get` / `+publish-unset`：互联网公开独立安全域与关闭后读回；`+publish-set` 保留为 unavailable 诊断入口。
+- `+version-download`：历史版本真实字节下载与本地 artifact 校验。
+- `+rename`：写后读回验证。
+
+普通钉盘文件的独立 `copy` 是额外确认出的部分 gap：钉钉当前 `doc/copy_document` 对该对象会生成 `.dlink`，不是字节独立副本。`drive +copy` 因此只接受在线对象；普通文件需要快捷入口时用 `+create-shortcut`，需要独立副本时用 `+download` 后 `+upload`。这不是完整的服务端原子 copy，对大文件也不能宣称完全等价。
+
+互联网公开开启也是账号/对象能力 gap：真实普通文件与在线文档夹具都由服务端返回 `operation.notSupported`。`+publish-get` 与关闭语义可验证，但 `+publish-set` 在找到 eligible 节点完成 set→get→unset 闭环前保持 `unavailable` 且不进入公开 Agent catalog。
+
+## 端到端门禁
+
+每个公开 Drive shortcut 必须至少覆盖：
+
+- Cobra 参数、静态确认、Shortcut Execute、MCP 调度和最终输出；
+- 明确业务空集合、空响应、缺字段、错误类型、坏元素、`success=false`；
+- 写入的 ID/终态证据与读回不一致；
+- 下载的本地路径边界、no-clobber、真实字节数；
+- 真实账号数据：读命令必须命中已知非空夹具或明确验证合法空集合；写命令必须创建隔离资源、读回、必要时下载比对字节并清理。
+
+发布前运行 `make build`、完整 Go 测试、Schema 生成/漂移/策略检查，并保存不含账号业务内容的结构化 E2E 汇总。
@@ -4,7 +4,7 @@ Defines the stable `dws event consume` subprocess contract so an
 orchestrator can determine when the consumer is ready, stop it cleanly,
 and machine-read why it exited.

-Scope of this branch: the five **contract** items below. Reconnect
+Scope of this branch: the six **contract** items below. Reconnect
 resilience (keeping the stream alive across a transient upstream drop) is
 tracked separately and intentionally out of scope here.

@@ -159,6 +159,46 @@ marker; reconnecting an established Stream remains a separate mechanism.
  `terminal_hold`, and identity-scoped cleanup; skill/docs tests pin the
  operational recovery instructions.

+### 6. Host runtime-token handoff
+
+When the root command carries an explicit host-supplied `--token`, personal
+event control requests and the foreground Stream use that token with higher
+priority than local OAuth. A detached bus receives it only through the
+owner-only local IPC transport:
+
+1. The child starts in runtime-token mode with non-sensitive identity and
+   ticket metadata only; neither its argv nor environment contains the token.
+2. The consumer sends `Hello` with `credential_mode=runtime_token`.
+3. The bus advertises the additive `runtime_token_v1` capability and its
+   in-memory credential generation in `HelloAck`.
+4. Only after that capability is confirmed does the consumer send a bounded
+   `credential_update` frame. The bus applies it with generation CAS, replies
+   with `credential_update_ack`, and registers the consumer only on success.
+
+The bus blocks ticket acquisition until the first runtime credential arrives.
+A later invocation may rotate Token A to Token B on a compatible existing bus;
+the current WebSocket remains connected and the next ticket request or natural
+reconnect uses B. If a 401 rejects the current runtime token, only an already
+installed newer generation is retried; the runtime path never refreshes or
+falls back to a local OAuth profile and never suggests `dws auth login`.
+
+Clients do not send a token to a bus that lacks the capability, do not stop
+other consumers automatically, and fail before printing the ready marker. With
+no explicit `--token`, the original OAuth, refresh, profile, and old-client to
+new-bus protocol behavior remains unchanged.
+
+**Verification**
+- T6a: a stale local Token A and root Token B produce control and ticket
+  requests authenticated only with B.
+- T6b: compatible bus reuse supports A-to-B rotation and generation conflicts;
+  401 retries only an already-installed newer runtime token.
+- T6c: an old bus receives no credential and remains running; the new consumer
+  exits before its ready marker.
+- T6d: a canary credential is absent from child argv/environment, dry-run,
+  stdout/stderr, `bus.meta`, `bus.log`, run state, and returned errors.
+- T6e: no-token OAuth, refresh, multi-profile, marker/cache, and bus-reuse tests
+  continue to pass.
+
 ## Out of scope (next branch)

 **Reconnect resilience** — today `personal source` retries only
@@ -92,6 +92,7 @@ command/Leaf 不再写 `dws.schema.risk`；SafetySpec 走类型化 Final 载荷
 | `Required` / `MarkRequired` | 非空校验 / cobra 硬必填 | 是（`required`） |
 | `RequiredHint`, `Aliases`, `EnvVar` | 校验提示、隐藏别名、环境回退 | 否（执行细节；别名不上主 parameter 表） |
 | `ArgDefault`, `Bind`, `OmitEmpty`, `Trim`, `Transform` | toolArgs 装配语义 | 否（载荷细节；`Bind` 可进 property 映射，但不另造 flag） |
+| `Input` | 额外取值来源：`@path` 读文件 / `-` 读 stdin，在 required/enum/约束/`Validate` 之前原地解析 | 否（今日：能力由作者写进 `Usage` / `SchemaDescription` 文案，是已声明事实而非推断；不另造 flag。目标形态收敛为类型化投影字段，见 RFC §5.3） |

 #### 1.2.2 编排 / 执行字段（不算声明）

@@ -0,0 +1,154 @@
+# International DingTalk (`.io`) Guide
+
+This guide explains how to log in to the international DingTalk region and run DWS commands against `*.dingtalk.io` services.
+
+## Region behavior
+
+- `dws auth login --intl` creates or refreshes an international login using the `.io` login, OAuth, and MCP services.
+- Omitting `--intl` keeps the existing domestic `.com` behavior.
+- `--intl` is a login option, not a global option for business commands. After login, commands such as `contact`, `calendar`, and `doc` derive the region from the selected Token/profile.
+- Each new Token records its login region. Switching profiles therefore switches the official DingTalk gateway region automatically.
+- `--international` is a compatibility alias. Prefer `--intl` in new scripts.
+
+For the complete Chinese guide, see [DWS 国际版（DingTalk `.io`）使用手册](./international-region-guide.zh-CN.md).
+
+## Check availability
+
+```bash
+dws auth login --help
+```
+
+The help output must include `--intl` and `--international`.
+
+When validating a source checkout, build it first and use `./dws` so an older binary on `PATH` is not invoked accidentally:
+
+```bash
+make build
+./dws auth login --help
+```
+
+## Log in
+
+Browser login:
+
+```bash
+dws auth login --intl
+```
+
+Device flow for SSH, containers, and headless environments:
+
+```bash
+dws auth login --intl --device
+```
+
+User OAuth with custom application credentials:
+
+```bash
+dws auth login --intl \
+  --client-id <APP_KEY> \
+  --client-secret <APP_SECRET>
+```
+
+This mode still requires the user to complete OAuth authorization in a browser; it is not a userless `client_credentials` login. The application must be configured on the international developer platform with the required callback and permissions. Never commit an AppSecret to source control or include it in logs.
+
+## Verify the login
+
+```bash
+dws auth status --format json
+dws profile list --format json
+dws contact user get-self
+```
+
+The last command is a read-only smoke check. If the organization has not enabled CLI access, an organization administrator must enable it or approve the access request on the international developer platform.
+
+## Use domestic and international profiles together
+
+```bash
+# Domestic (.com)
+dws auth login
+
+# International (.io)
+dws auth login --intl
+
+# Find the stable profile selectors
+dws profile list --format json
+```
+
+Persistently switch profiles:
+
+```bash
+dws profile switch <corpId>:<userId>
+```
+
+Toggle back to the previous profile:
+
+```bash
+dws profile switch -
+```
+
+Select a profile for one command without changing the default:
+
+```bash
+dws --profile <corpId>:<userId> contact user get-self
+```
+
+Do not add `--intl` to business commands. DWS routes official endpoints from the selected profile's Token region.
+
+## Isolated smoke testing
+
+Use a separate configuration directory to avoid changing the normal `~/.dws` login state:
+
+```bash
+DWS_CONFIG_DIR=/tmp/dws-intl-smoke ./dws auth login --intl
+DWS_CONFIG_DIR=/tmp/dws-intl-smoke ./dws auth status --format json
+DWS_CONFIG_DIR=/tmp/dws-intl-smoke ./dws contact user get-self
+```
+
+Use the same `DWS_CONFIG_DIR` for every command. Use `./dws` for a source build and `dws` for an installed release.
+
+## Pre-release overrides (maintainers only)
+
+Normal international users need only `--intl`; they should not set `--pre-url` or `--mcp-url`.
+
+Maintainers can test the pre-release login/MCP pair with:
+
+```bash
+dws auth login --intl --pre-url https://pre-login.dingtalk.io
+```
+
+A corresponding `pre-mcp.*` URL is also accepted, and DWS derives the paired `pre-login.*` / `pre-mcp.*` bases. `--mcp-url` explicitly overrides the MCP base URL for that login.
+
+Pre-release services may require internal network access or allowlisted accounts. `--pre-url` is intended primarily for the MCP-managed credential flow. Do not combine it with direct custom `--client-id/--client-secret` mode unless the pre-release API contract explicitly supports that combination.
+
+## Troubleshooting
+
+### The browser still opens a `.com` page
+
+1. Run `dws auth login --help` and confirm `--intl` is present.
+2. For a source checkout, use `./dws` instead of an older installed binary.
+3. Confirm the executed command is `dws auth login --intl`.
+
+### A business command appears to use the wrong region
+
+Run `dws profile list --format json`, then switch with the exact `<corpId>:<userId>` selector or use the global `--profile` option. For a legacy Token created before region metadata existed, reauthorize it with `dws auth login --intl` for an international account or `dws auth login` for a domestic account.
+
+### Login succeeds but the command reports missing permission
+
+This normally means the organization has not enabled CLI access or the application lacks a required permission. It does not by itself indicate a region-routing failure.
+
+### Should I edit `~/.dws/mcp_url` manually?
+
+No. Normal users should establish the login with `dws auth login` or `dws auth login --intl`. DWS then routes official endpoints from the selected Token/profile. Manual configuration is reserved for maintainers who explicitly control the target environment.
+
+## Command reference
+
+| Scenario | Command |
+|---|---|
+| Domestic browser login | `dws auth login` |
+| International browser login | `dws auth login --intl` |
+| International device login | `dws auth login --intl --device` |
+| Check auth state | `dws auth status --format json` |
+| List profiles | `dws profile list --format json` |
+| Persistently switch profile | `dws profile switch <corpId>:<userId>` |
+| Toggle to previous profile | `dws profile switch -` |
+| Select a profile once | `dws --profile <corpId>:<userId> <command>` |
@@ -0,0 +1,185 @@
+# DWS 国际版（DingTalk `.io`）使用手册
+
+本手册适用于使用钉钉国际版账号登录并调用国际站服务的用户。
+
+## 核心规则
+
+- `dws auth login --intl` 创建或刷新国际版登录，使用 `*.dingtalk.io` 登录、鉴权和 MCP 服务。
+- 不传 `--intl` 时仍使用国内钉钉 `*.dingtalk.com`，原有链路保持不变。
+- `--intl` 只用于登录命令。登录完成后，`contact`、`calendar`、`doc` 等业务命令不需要再传该参数。
+- 每个 Token 会记录登录区域。执行业务命令时，DWS 根据当前或 `--profile` 指定的账号自动选择 `.com` 或 `.io` 网关。
+- `--international` 是 `--intl` 的兼容别名；新脚本推荐使用较短的 `--intl`。
+
+## 确认当前版本支持国际版
+
+运行：
+
+```bash
+dws auth login --help
+```
+
+帮助中应包含：
+
+```text
+--intl
+--international
+```
+
+从源码分支验证时，先在仓库根目录构建，并始终使用本次构建的 `./dws`，避免误用系统中已安装的旧版本：
+
+```bash
+make build
+./dws auth login --help
+```
+
+## 国际版登录
+
+### 浏览器登录
+
+```bash
+dws auth login --intl
+```
+
+DWS 会打开国际版登录页面。完成扫码或账号授权后，登录结果会保存为本机 profile。
+
+### 设备码登录
+
+适用于 SSH、容器或没有可用浏览器的环境：
+
+```bash
+dws auth login --intl --device
+```
+
+按照终端提示，在另一台可打开浏览器的设备上完成授权。
+
+### 使用自有应用凭证完成用户 OAuth
+
+```bash
+dws auth login --intl \
+  --client-id <APP_KEY> \
+  --client-secret <APP_SECRET>
+```
+
+该模式仍然需要用户在浏览器中完成 OAuth 授权，不是无用户授权的 `client_credentials` 登录。应用必须在国际版开放平台正确配置回调地址和所需权限。不要在命令历史、日志或 PR 中提交真实的 AppSecret。
+
+## 验证登录和业务调用
+
+查看当前登录状态：
+
+```bash
+dws auth status --format json
+```
+
+列出本机全部账号并找到当前 profile：
+
+```bash
+dws profile list --format json
+```
+
+执行一个只读命令验证国际链路，例如：
+
+```bash
+dws contact user get-self
+```
+
+登录状态正常但业务命令提示组织未开通 CLI 时，需要由国际版组织管理员在国际版开发者平台开启 CLI 访问或完成授权审批。
+
+## 国内版和国际版账号并存
+
+可以在同一台机器上分别登录国内版和国际版账号：
+
+```bash
+# 国内版（.com）
+dws auth login
+
+# 国际版（.io）
+dws auth login --intl
+
+# 查看稳定的 profile 选择器
+dws profile list --format json
+```
+
+持久切换账号：
+
+```bash
+dws profile switch <corpId>:<userId>
+```
+
+切回上一个账号：
+
+```bash
+dws profile switch -
+```
+
+只为单次命令指定账号，不修改默认账号：
+
+```bash
+dws --profile <corpId>:<userId> contact user get-self
+```
+
+DWS 会按照选中 profile 的 Token 区域自动选择 `.com` 或 `.io`，不需要在业务命令上追加 `--intl`。
+
+## 使用独立配置目录进行验证
+
+如果不希望测试登录影响日常使用的 `~/.dws`，可以指定独立配置目录：
+
+```bash
+DWS_CONFIG_DIR=/tmp/dws-intl-smoke ./dws auth login --intl
+DWS_CONFIG_DIR=/tmp/dws-intl-smoke ./dws auth status --format json
+DWS_CONFIG_DIR=/tmp/dws-intl-smoke ./dws contact user get-self
+```
+
+请在三条命令中使用同一个 `DWS_CONFIG_DIR`。验证源码分支时使用 `./dws`；验证已安装版本时可改为 `dws`。
+
+## 预发参数（仅维护者）
+
+普通国际版用户只需要 `--intl`，不要配置 `--pre-url` 或 `--mcp-url`。
+
+维护者验证预发登录/MCP 链路时可以使用：
+
+```bash
+dws auth login --intl --pre-url https://pre-login.dingtalk.io
+```
+
+也可以传入对应的 `pre-mcp.*` 地址；DWS 会推导配套的 `pre-login.*` / `pre-mcp.*` 地址。`--mcp-url` 用于显式覆盖本次登录的 MCP base URL。
+
+预发环境可能只对内网或特定测试账号开放。`--pre-url` 主要服务于 MCP 托管凭证登录流程；除非预发 API 契约已经明确支持，否则不要把它与自有 `--client-id/--client-secret` 直连模式组合使用。
+
+## 常见问题
+
+### 仍然打开 `.com` 登录页面
+
+1. 运行 `dws auth login --help`，确认当前二进制包含 `--intl`。
+2. 从源码验证时使用 `./dws`，不要误用 PATH 中的旧版本。
+3. 确认实际执行的是 `dws auth login --intl`，而不是普通 `dws auth login`。
+
+### 业务命令似乎使用了错误区域
+
+先检查当前账号：
+
+```bash
+dws profile list --format json
+```
+
+然后使用精确的 `<corpId>:<userId>` 切换或通过全局 `--profile` 单次指定。对于在区域字段引入前生成的历史 Token，建议使用正确的登录方式重新授权：国际账号执行 `dws auth login --intl`，国内账号执行 `dws auth login`。
+
+### 登录成功但提示没有权限
+
+这通常是组织 CLI 准入或应用授权问题，不代表区域路由失败。请确认目标组织已开启 CLI 访问，并且当前应用拥有命令所需权限。
+
+### 是否需要手工修改 `~/.dws/mcp_url`
+
+不需要。正常使用应通过 `dws auth login` 或 `dws auth login --intl` 建立登录态；业务命令会根据选中的 Token/profile 自动路由。手工修改配置只适用于明确了解目标环境的维护者调试场景。
+
+## 命令速查
+
+| 场景 | 命令 |
+|---|---|
+| 国内版浏览器登录 | `dws auth login` |
+| 国际版浏览器登录 | `dws auth login --intl` |
+| 国际版设备码登录 | `dws auth login --intl --device` |
+| 查看登录状态 | `dws auth status --format json` |
+| 查看所有账号 | `dws profile list --format json` |
+| 持久切换账号 | `dws profile switch <corpId>:<userId>` |
+| 切回上一个账号 | `dws profile switch -` |
+| 单次指定账号 | `dws --profile <corpId>:<userId> <command>` |
@@ -0,0 +1,134 @@
+# 独立 `meta.pagination` Schema 方案
+
+## 1. 目标结构
+
+业务结果与分页控制信息分层：
+
+```json
+{
+  "ok": true,
+  "outcome": "success",
+  "data": {
+    "items": [{"id": "a"}]
+  },
+  "meta": {
+    "pagination": {
+      "endpoint_exhausted": false,
+      "next_token": "cursor-2"
+    }
+  }
+}
+```
+
+对应 compact/full leaf Schema：
+
+```json
+{
+  "result": {
+    "outcomes": ["success", "failure"],
+    "data_schema": {
+      "type": "object",
+      "properties": {
+        "items": {
+          "type": "array",
+          "description": "当前页业务记录",
+          "items": {"type": "object"}
+        }
+      }
+    }
+  },
+  "pagination": {
+    "kind": "cursor",
+    "cursor_parameter": "cursor",
+    "meta_path": "meta.pagination",
+    "endpoint_exhausted_path": "meta.pagination.endpoint_exhausted",
+    "next_token_path": "meta.pagination.next_token"
+  }
+}
+```
+
+`result` 只描述 `data`；`pagination` 是与 `result` 同级的命令能力声明。
+
+## 2. 分页状态
+
+| 状态 | `endpoint_exhausted` | `next_token` | Agent 行为 |
+|---|---:|---|---|
+| 可续跑 | `false` | 必须非空 | 将 token 传给 `--<cursor_parameter>` |
+| 已耗尽 | `true` | 必须省略 | 停止翻页 |
+
+`endpoint_exhausted:true` 只表示观察到 Endpoint 分页耗尽，不表示搜索索引
+健康、数据全量覆盖或业务对象不存在。
+
+## 3. 映射规则
+
+产品 mapper 可以读取服务端原始 `hasMore/nextCursor`、`has_more/page_token`
+等字段，但统一 CLI 输出只公布 `meta.pagination`：
+
+- 服务端表示还有下一页且 cursor 非空 → `endpoint_exhausted:false` + token。
+- 服务端表示没有下一页 → `endpoint_exhausted:true`，不带 token。
+- 表示还有下一页但 cursor 缺失、类型错误或证据冲突 → typed
+  `pagination_inconsistent`，禁止伪装终页。
+- mapper 使用同一份上游响应构造 `data` 与 `meta`，不得重新请求。
+
+原始分页控制字段不进入新的 `result.data_schema`。未迁移命令保持 legacy；
+已迁移命令按命令独立切换和回滚，不通过 Agent 参数选择协议。
+
+## 4. Schema 规则
+
+- `kind` 当前只允许 `cursor`。
+- `cursor_parameter` 是真实 canonical CLI flag 名，不带 `--`，并必须存在于
+  同一 leaf 的 `parameters`。
+- 三个 meta path 由框架固定生成，产品不能覆盖。
+- compact/full leaf 同时包含相同的 `result` 和 `pagination`。
+- product/group 导航摘要不复制分页对象；Agent 需要时查询具体 compact leaf。
+- 没有 `pagination` 表示该命令尚未发布经评审的分页能力，Agent 不得猜测。
+
+## 5. 渐进接入
+
+1. **legacy_only**：保持原输出，不公布分页声明。
+2. **dual_validate**：业务执行一次；影子构造并校验 `meta.pagination`，外部
+   legacy 字节不变。
+3. **unified_active**：输出独立 `meta.pagination`，Schema 公布同级
+   `pagination` 声明。
+4. **unified_stable**：Skill、示例和 Agent 审计均只读取 meta 分页。
+
+不增加 `contract_version`、`--output-contract` 或分页协议别名。
+
+## 6. 验收
+
+每个分页命令至少验证：
+
+1. 有下一页时 `endpoint_exhausted:false` 且 token 非空。
+2. 终页和空终页为 `endpoint_exhausted:true` 且无 token。
+3. 分页矛盾产生 typed failure，不 panic、不静默停止。
+4. `cursor_parameter` 在 Help/Schema 中真实存在。
+5. compact/full 的 `result`、`pagination` 分别 JSON 等价。
+6. `data_schema` 不包含分页控制字段。
+7. 运行时 `data` 不包含迁移后的分页控制字段。
+8. dual validate 与 active 都只消费一次上游响应。
+9. Agent 逐命令扫描结果进入评测台账；不提交生成 Schema JSON fixture。
+
+### DevApp 首批落地
+
+以下 8 个终结命令已发布独立 `pagination` Schema；运行时统一输出只在
+`meta.pagination` 返回分页控制信息：
+
+- `dev app list`
+- `dev app permission list`
+- `dev app event list`
+- `dev app version list`
+- `devapp +list`
+- `devapp +permission-list`
+- `devapp +event-list`
+- `devapp +version-list`
+
+两套既有命令前缀继续保留。原子命令的业务记录字段为 `data.items`；Shortcut
+保留既有业务投影（例如 `data.apps`、`data.permissions`、`data.events`、
+`data.versions` 以及 `data.count`），但两套入口都不再在业务数据中公布
+`hasMore/nextCursor`。
+
+## 7. 对齐依据
+
+GWS 用请求参数和 response schema 描述分页事实；Lark 在统一输出层维护分页
+元数据。DWS 采用更明确的分层：业务 `data` 保真承载记录，框架 `meta` 承载
+续跑状态，Schema 用独立能力把 token 与下一次 CLI 参数连接起来。
@@ -7,6 +7,8 @@
 | `DWS_CONFIG_DIR` | Override default config directory / 覆盖默认配置目录 |
 | `DWS_AGENT_PRODUCT` | Optional, caller-declared Agent product sent as `x-dws-agent-product` (for example `qwenwork`) for downstream logs/BI and used as the IM `clawType` display label when `--ai-tag` is enabled. `--ai-tag` defaults to `true`, so a configured Product changes the displayed label by default. With `--ai-tag=false`, native `chat message send` / `reply` calls send an empty `clawType`, while shortcut calls omit the argument. Surrounding ASCII spaces/tabs are trimmed; the remaining value must be at most 64 bytes and match `^[A-Za-z0-9][A-Za-z0-9_-]*$`. Unset or empty values omit the Header and use the edition's IM display default. This client never uses Product to change the separate HTTP `claw-type` PAT/routing label. / 可选、由调用方声明的 Agent 产品标识，经校验后作为 `x-dws-agent-product` 发送，并用于 IM 小尾巴；`--ai-tag` 默认为 `true`，因此配置 Product 后默认会改变展示标签。使用 `--ai-tag=false` 时，原生 `chat message send` / `reply` 发送空的 `clawType`，shortcut 调用则省略该参数。未设置时省略请求头且 IM 使用发行版默认值；本客户端不会用 Product 修改独立的 HTTP `claw-type` |
 | `DWS_AGENT_HOST` | Optional, caller-declared Agent runtime form sent as `x-dws-agent-host` (for example `cloud` or `desktop`) for downstream logs/BI. Surrounding ASCII spaces/tabs are trimmed; the remaining value must be at most 64 bytes and match `^[a-z0-9][a-z0-9_-]*$`; unset values are omitted. This client does not use Host for PAT, authentication, Discovery, or MCP endpoint selection. / 可选、由调用方声明的 Agent 运行形态，经校验后作为 `x-dws-agent-host` 发送给下游日志/BI；本客户端不使用该值进行 PAT、鉴权、Discovery 或 MCP 端点选择，未设置时省略 |
+| `DWS_AGENT_VER` | Optional caller-declared Agent version / 可选、由调用方声明的 Agent 版本。After trimming surrounding ASCII spaces/tabs, the value must be at most 64 bytes and match `^[A-Za-z0-9][A-Za-z0-9._+-]*$`; a non-empty valid value is sent as `x-dws-agent-ver`, while unset or empty values omit the Header. / 去除首尾 ASCII 空格和 Tab 后，值不得超过 64 字节且必须匹配上述格式；合法非空值通过 `x-dws-agent-ver` 发送，未设置或空值则省略请求头 |
+| `DWS_AGENT_EXT` | Optional caller-declared Agent extended context / 可选、由调用方声明的 Agent 扩展上下文。The value must be a UTF-8 JSON object no larger than 8 KiB, is compacted before being sent as the sensitive `x-dws-agent-ext` Header, and may use the recommended keys `umt`, `miniwua`, and `ua`; unknown keys remain supported. Unset or empty values omit the Header. / 值必须是 UTF-8 JSON 对象且不得超过 8 KiB，压缩后通过敏感请求头 `x-dws-agent-ext` 发送；推荐使用 `umt`、`miniwua`、`ua`，同时允许未知扩展键。未设置或空值则省略请求头 |
 | `DWS_<PRODUCT>_MCP_URL` | Override a product MCP endpoint for local development / 本地开发时覆盖指定产品 MCP endpoint |
 | `DWS_CLIENT_ID` | OAuth client ID (DingTalk AppKey) |
 | `DWS_CLIENT_SECRET` | OAuth client secret (DingTalk AppSecret) |
@@ -14,6 +16,47 @@
 | `DWS_ALLOW_HTTP_ENDPOINTS` | Set `1` to allow HTTP for loopback during dev / 设为 `1` 允许回环地址 HTTP，仅用于开发调试 |
 | `DWS_DISABLE_KEYCHAIN` | macOS only. Set `1` to skip system Keychain for the encryption key and use file-based storage (same scheme as Linux). For sandboxed runtimes (e.g. Codex App) that block Keychain APIs. Weakens at-rest protection — DEK and ciphertext live in the same directory. / 仅 macOS。设为 `1` 时跳过系统 Keychain，密钥以文件形式存储（与 Linux 一致）。用于 Keychain API 被拦截的沙盒环境（如 Codex App）。代价是 DEK 与密文同目录，保护强度低于默认方案 |

+### Agent Version and Extended Context / Agent 版本与扩展上下文
+
+`DWS_AGENT_VER` and `DWS_AGENT_EXT` are sent only on the CLI's ordinary,
+non-plugin MCP requests. They do not change the standard HTTP `User-Agent` or
+the separate `X-Cli-Version` that identifies the DWS CLI version, and they are
+not forwarded to A2A, OAuth, Discovery, or third-party plugin requests.
+
+`DWS_AGENT_EXT` is one JSON-object Header rather than a set of Headers. The
+recommended keys are `umt`, `miniwua`, and `ua`, but the open-source CLI keeps
+the object extensible and does not enforce a key allowlist. For example, using
+fictional, redacted values:
+
+```bash
+DWS_AGENT_VER=0.1.5
+DWS_AGENT_EXT='{"umt":"example-redacted","miniwua":"example-redacted","ua":"ExampleAgent/0.1.5"}'
+```
+
+The shell's outer single quotes group the JSON and are not part of the
+environment-variable value. The CLI trims surrounding ASCII spaces/tabs,
+omits either Header when its value is empty, and compacts EXT to a single-line
+JSON object. A representative current payload is about 657 bytes, well below
+the 8 KiB limit; integrations must still enforce the limit because values can
+grow. EXT may contain sensitive device or runtime signals: the CLI masks it in
+configuration and logs, and removes it on a cross-host redirect.
+
+Both values are declared by the caller and are therefore forgeable. They can
+support compatibility checks, diagnostics, and observability, but they are not
+credentials or attestations and must never be sufficient on their own to
+authenticate a caller or authorize access.
+
+`DWS_AGENT_VER` 与 `DWS_AGENT_EXT` 仅随 CLI 发起的普通非插件 MCP 请求发送，不会
+改变标准 HTTP `User-Agent`，也不会覆盖标识 DWS CLI 自身版本的 `X-Cli-Version`；
+二者不会进入 A2A、OAuth、Discovery 或第三方插件请求。EXT 使用单个 JSON 对象
+请求头，不拆成多个子请求头；推荐键为 `umt`、`miniwua`、`ua`，但开源 CLI 不限制
+扩展键。Shell 示例中的外层单引号只用于保护 JSON，不属于环境变量值。当前典型负载
+约为 657 字节，远低于 8 KiB 上限，但集成方仍须遵守大小限制。EXT 可能包含敏感的
+设备或运行时信号，配置展示和日志会对其脱敏，跨主机重定向时也会移除该请求头。
+
+这两个值都由调用方自行声明，可以被伪造；它们可用于兼容性判断、诊断和可观测性，
+但不是凭据或可信证明，不能单独用于身份认证或访问授权。
+
 ### Agent Product, Host, and `claw-type` / Agent 产品、运行形态与 `claw-type`

 `DWS_AGENT_PRODUCT` and `DWS_AGENT_HOST` are caller-declared observation
@@ -94,7 +137,7 @@ With `-f json`, error responses include structured payloads: `category`, `reason
 dws contact user search --query "Alice" -f table   # Table (default, human-friendly / 表格，默认)
 dws contact user search --query "Alice" -f json    # JSON (for agents and piping / 适合 agent)
 dws contact user search --query "Alice" -f raw     # Raw API response / 原始响应
-dws schema -f pretty "calendar event create"         # Pretty Agent schema view / Agent Schema 彩色查看
+dws schema -f pretty "calendar event create" --compact # Pretty Agent schema view / Agent Schema 彩色查看
 ```

 ## Dry Run / 试运行
@@ -119,27 +162,28 @@ Schema 的稳定 `canonical_path`、主 CLI 路径和 aliases 收集自命令树

 ```bash
 dws schema                                      # 当前公开产品面的紧凑概览
-dws schema calendar                             # 展开一个产品
-dws schema "calendar event"                     # 展开一个命令分组
-dws schema "calendar event create"              # 按 CLI 空格路径查询工具
-dws schema calendar.create_calendar_event       # 按 canonical path 查询工具
-dws schema --cli-path "calendar event create"   # 显式 CLI path
-dws schema "calendar event create" --compact    # 支持：省略 provenance/debug 字段
+dws schema calendar --compact                   # Agent 产品视图
+dws schema "calendar event" --compact          # Agent 分组视图
+dws schema "calendar event create" --compact   # Agent leaf（CLI 空格路径）
+dws schema calendar.create_calendar_event --compact     # Agent leaf（canonical path）
+dws schema --cli-path "calendar event create" --compact # Agent leaf（显式 CLI path）
+dws schema "calendar event create"             # full leaf，仅用于映射/provenance 审计
 dws schema --all                                # 全部工具的完整 leaf Schema，用于审计/CI/baseline
 ```

 兼容入口 `dws schema list` 等价于根概览。`schema --all` 是完整导出：每个工具都包含完整 leaf 参数、约束和安全语义。它输出很大，只用于明确要求的全量导出、审计、CI 或参数 baseline；普通 Agent 任务应按概览、产品/分组、leaf 渐进查询，不要把 `--all` 直接注入上下文。`schema --all --compact` 虽受支持，但会裁掉 provenance 和接口映射字段，不能作为完整 baseline。

-Leaf 查询、`--all` 中对应工具和 Catalog full tool 均由同一个 resolved `ToolSpec` 投影，内容必须一致；概览、产品/分组和 Catalog summary 也由该 `ToolSpec` 的统一 summary 投影生成。通过 alias 查询时，只允许 `cli_path` 和 `is_alias` 发生视图变化，参数、安全和接口契约不得变化。
+省略 `--compact` 的 full leaf、`--all` 中对应工具和 Catalog full tool 均由同一个 resolved `ToolSpec` 投影，内容必须一致；compact leaf 仅做字段白名单投影，不重新解析语义。概览、产品/分组和 Catalog summary 也来自同一 `ToolSpec`。通过 alias 查询时，只允许路径视图发生变化，参数、安全和接口契约不得变化。

-`--compact` 是 Schema 的展示选项。当前版本支持该 flag；若兼容旧二进制时收到 `unknown_flag: --compact`，用同一个 Schema 查询去掉 `--compact` 重试。这只降低输出裁剪能力，不表示 leaf 不存在，也不能改用 Schema 查询业务数据。
+`--compact` 是 Schema 的稳定 Agent 字段白名单，也是普通 Agent 查询的规范选项。它保留 CLI 参数、组合约束、选择和安全语义，但有意省略 `interface_ref`、参数 `property/interface_type` 与 provenance。检查这些映射/审计字段时，使用 full leaf 并通过 `--jq` / `--fields` 精确投影。若兼容旧二进制时收到 `unknown_flag: --compact`，用同一个 Schema 查询去掉 `--compact` 重试；这只降低输出裁剪能力，不表示 leaf 缺失。

 ### Schema、Help 与业务数据的边界

 | 问题 | 事实源 |
 |------|--------|
 | 命令是否由当前二进制暴露、Cobra 接受哪些 flags | `dws <path> --help` |
-| Agent 选哪个命令、参数映射与组合约束、risk/confirmation | 对应的 leaf `dws schema "<path>"` |
+| Agent 选哪个命令、CLI 参数与组合约束、risk/confirmation | 对应的 Agent leaf `dws schema "<path>" --compact` |
+| CLI↔RPC 参数映射、接口绑定与 provenance | full leaf 配合 `--jq` / `--fields` 精确投影 |
 | 当前钉钉中的文档、文件、日程、消息等业务数据 | 实际执行 `dws doc read`、`dws drive search` 等 read/search/list 命令 |

 Schema 与 Help 冲突表示发布契约漂移，不能静默猜测。执行参数必须以 Cobra 实际接受的 flag 为准；安全语义冲突时采用更保守的处理（例如先确认）或停止执行并报告漂移。完成命令发现后，仍必须执行真实业务命令；`dws schema` 本身不会读取或搜索业务内容。
@@ -17,7 +17,8 @@

 1. 在上述 `Release` 页面选择 `Run workflow`，分支必须是默认分支 `main`。
 2. `release_operation=plan`，选择 `release_channel=beta|stable`；仅在开始新 beta 线时选择 `release_bump=patch|minor|major`。
-3. workflow summary 会给出唯一的下一版本。把对应的精确 `CHANGELOG.md` 章节通过 PR 合入 `main`。
+3. workflow summary 会给出唯一的下一版本。运行 `prepare-changelog.sh` 将已合入的
+   release fragments 汇总成对应的精确 `CHANGELOG.md` 章节，并通过唯一的 release-seal PR 合入 `main`。
 4. 再次运行，改为 `release_operation=publish`。beta 会直接进入自动化发布；stable 会在封 tag 前等待管理员签收。

 `plan` 是纯只读操作，不创建 tag、预留版本号或生成包。CHANGELOG 合入期间若另一个发布先占用了该版本，`publish` 会重新分配并因 CHANGELOG 章节不匹配而拒绝，需要重新 plan。`publish` 会先再次确认 dispatch SHA 仍是当前 `main`、Code Admission 和平台治理均通过，再由唯一的 write job 使用 GitHub API 原子创建 annotated tag；同一次 run 随即进入既有的跨平台构建、GitHub/npm、可选 OSS/Gitee 发布和 Homebrew 直交付 DAG。内置 `GITHUB_TOKEN` 创建的 tag 不依赖第二条 workflow 被再次触发。
@@ -94,7 +95,8 @@ main 上的候选代码 + beta CHANGELOG
 dws-release v1.2.3-beta.1
 ```

-如果 CHANGELOG 尚不存在，该命令只生成模板并停止。补全内容、删除所有 `TODO`，提交后通过 PR 合入 `main`；然后重新运行完全相同的命令，它会执行完整预检：
+如果 CHANGELOG 尚不存在，该命令会从 `.changes/*.md` 生成 beta 章节并归档已消费的
+fragments，然后停止。审阅生成内容并通过唯一的 release-seal PR 合入 `main`；然后重新运行完全相同的命令，它会执行完整预检：

 ```bash
 dws-release v1.2.3-beta.1
@@ -132,6 +134,15 @@ dws-release v1.2.3 --from-beta v1.2.3-beta.1

 正式版使用 `## [1.2.3] - YYYY-MM-DD`。该章节会直接成为 GitHub Release Notes。

+### Release fragments
+
+普通 PR 不修改 `CHANGELOG.md` 的 `Unreleased` 区域。需要面向用户发布说明的改动在
+`.changes/<unique-name>.md` 中增加一个独立 fragment；格式和允许的分类见
+[`.changes/README.md`](../.changes/README.md)。预发封板时
+`scripts/release/prepare-changelog.sh prerelease <version>` 会稳定排序并汇总所有未归档
+fragment，写入唯一版本章节后移动到 `.changes/released/<version>/`。因此并发 PR 不会争用
+`CHANGELOG.md`；唯一的 release-seal PR 同时提交生成的章节与归档移动，供审计复核。
+
 ## CI/CD 保证

 - 只接受 `vX.Y.Z-beta.N` 和 `vX.Y.Z`，且新版本必须高于上一正式版。这里的“上一正式版”必须同时具备公开非草稿 GitHub Release 和同 tag/commit 的成功 Release workflow；只有 tag、没有交付成功的孤儿版本会阻断后续发布，要求走机器核验恢复补齐。云端 tag 会固定 `Release-Run`、requester、commit 和版本分配指纹，交付验证按该精确 run/attempt 及完整 job graph 取证，不接受任意 `workflow_dispatch`。历史版本若曾通过专用 recovery workflow 完成交付，只能使用仓库内 `delivered-stable-recoveries.json` 中精确到 tag、commit、run、workflow SHA 与 attempt 的 reviewed 证据。
@@ -292,7 +292,7 @@ Definition（仅声明；不可编译）

 下列字段**是**框架声明面（经 `corecmd.New` 生效并嵌入 `dws.schema.*`）：

- `Flags`（含 Name/Kind/Default/Required/MarkRequired/Usage 等注册面）
+- `Flags`（含 Name/Kind/Default/Required/MarkRequired/Usage 等注册面；`Input` 是取值来源声明，经 `corecmd.New` 生效但**不**嵌入 `dws.schema.*`，能力靠 `Usage` 文案声明，见 §5.3）
 - `Constraints`
 - **非空** `Risk`（空值 = 运行时当只读确认，且**不**嵌入 `dws.schema.risk`）
 - `ConstParams`（载荷声明；不上用户 flag 表）
@@ -673,6 +673,52 @@ func (k Key[T]) Declare(opts ...FlagOption[T]) FlagSpec
 - 构造时拒绝 `InputSourceInvalid`。
 - 当前没有任何 Shortcut 或 Leaf 声明 `Input`，因此 M1 增加能力且零上线表面变化。让现有命令采用它属于 §9 下的用户可见变更。

+`Input` 的框架能力今日已在 `corecmd` 落地（声明即执行的过渡形态，语义与上文目标一致），使用指南：
+
+**今日声明形态**：`FlagSpec.Input []string`，源常量 `corecmd.InputFile`（`"file"`）/ `corecmd.InputStdin`（`"stdin"`）。`helpers.LeafFlag` 是 `corecmd.FlagSpec` 别名，直接可用；`shortcut.Flag.Input` 同形声明，经 `FromShortcut` 映射到 `FlagSpec`。
+
+```go
+// LeafSpec / helpers
+Flags: []helpers.LeafFlag{
+    {
+        Name:  "content",
+        Usage: "文档内容（支持 @文件路径 或 - 读 stdin）",
+        Bind:  "content",
+        Input: []string{corecmd.InputFile, corecmd.InputStdin},
+    },
+}
+
+// shortcut
+Flags: []shortcut.Flag{
+    {Name: "markdown", Desc: "Markdown 内容（支持 @文件路径 或 -）",
+     Input: []string{"file", "stdin"}},
+}
+```
+
+**运行时语义**（`resolveInputFlags`，在 `runDeclaredPreflight` 内、required/enum/约束/Validate 之前执行，原地改写 cobra flag 值）：
+
+- `--flag @path`：文件内容替换取值；`--flag -`：stdin 内容替换取值。
+- `--flag @@value`：转义为字面 `@value`，不做来源解析。
+- 只解析显式 CLI token（主名或别名）；EnvVar 回落与注册默认值透传不解析。
+- 内容前置剥离 UTF-8 BOM；`Trim` 等既有语义照常作用于解析后的值。
+- 读取失败、源不支持、`@` 后空路径都是类型化校验错误（退出码 3）；同时声明两种源而文件读取失败时附 stdin 引导 hint。
+
+**作者守则**：
+
+- 声明即全部能力：required/enum/约束/Validate 校验的已是解析后的真实内容，`Execute`/`Invoke` 无需任何额外代码。
+- `Usage`/`Desc` 必须写明支持 `@路径`/`-`；框架不自动改写 help 文案，今日也不向 Schema 投影（新增投影字段须先过 homology 评审，避免 catalog drift）。
+- `user_required` 确认的写命令若声明 `InputStdin`：stdin 在校验阶段被消费，交互确认将 fail-closed 为 `confirmation_required`，此类调用必须显式 `--yes`（或 `--dry-run`）。
+- **声明前先确认取值空间不会被前缀吃掉**：声明 `InputFile` 后，任何以 `@` 开头的合法值都会被当成文件路径（本产品尤其常见的是 at 提及类取值，如 `--at-user @zhangsan` 会报读取文件失败），用户只能改用 `@@` 转义；声明 `InputStdin` 后字面值 `-` 不可达（与 curl 等约定一致）。若该 flag 的正常取值可能命中这两种形态，就不要声明对应来源。
+- 声明在构造期校验（fail-closed panic）：仅限 `KindString`；源值必须是 `file`/`stdin` 且不重复。
+
+**今日实现与目标形态的差异**（迁移到本节目标 `FlagSpec` 时收敛）：
+
+| 维度 | 今日 | 目标 |
+|---|---|---|
+| 源类型 | `[]string` 常量 | 类型化 `InputSource` |
+| 路径边界 | 直接本地文件 IO | 复用 §5.5.2 本地文件 effect 边界 |
+| Schema 投影 | 无（靠作者在 Usage 声明） | 声明即最终源，随 Catalog 透传 |
+
 核心 FlagSpec 故意没有：

 - `Bind`；
@@ -0,0 +1,219 @@
+# RFC：DWS 预制 Skill 安装、升级与模式迁移
+
+| 字段 | 内容 |
+|---|---|
+| 状态 | Accepted / as implemented |
+| 生效范围 | DWS CLI、升级器、npm 与平台安装脚本 |
+| 事实源 | 本 RFC 与当前代码；两者冲突时以代码和测试为准 |
+| 关联合同 | [Skill 内容框架](skill-content-framework.md)、[Mono↔Multi 内容质检](skill-mono-multi-qa.md) |
+
+## 1. 背景
+
+DWS 同时通过 CLI、升级器、npm、Shell 和 PowerShell 分发预制 Skill。multi
+成为默认布局后，所有入口必须对安装集合、模式互斥、失败退出、缓存发布和目录
+所有权保持一致。此前分散的调研、迁移计划、阶段性 roadmap 和 rollout 文档容易
+相互冲突；本 RFC 将最终行为收敛为一个长期合同。
+
+## 2. 目标与非目标
+
+### 2.1 目标
+
+- 新装与升级默认使用 multi 布局，mono 在兼容期内保留显式 opt-in。
+- 每次升级使用当前版本的官方清单全量覆盖预制 Skill。
+- 删除或替换任何目录前先创建可恢复备份，备份失败不修改该 Agent 目标。
+- 只清理能够证明由 DWS 管理的目录，不通过名称前缀推断所有权。
+- 所有安装入口对部分失败返回非零状态，不误报整体成功。
+- 安装预览、确认和实际执行使用同一份计划。
+
+### 2.2 非目标
+
+- 不建设独立的 `dws skill mode status|set|rollback` 产品面。
+- 不持久化用户对预制 Skill 的本地删除或排除意图。
+- 不提供跨所有 Agent 目标的事务式回滚。
+- 不把市场 Skill 纳入预制 Skill 的升级和清理范围。
+
+## 3. 业内调研
+
+对主流 CLI 与 Agent Skill 分发方式的公开实现进行归纳后，可以得到以下共性：
+
+| 观察 | 对 DWS 的启示 |
+|---|---|
+| 多个产品能力通常以同级 Skill 目录安装，由 Agent 按目录发现 | multi 使用平铺的产品 Skill，并保留一个共享 Skill 承载公共协议 |
+| CLI 本体安装和 Agent Skill 安装是两个生命周期 | DWS 可以在 CLI 安装、setup 和 upgrade 中触发 Skill 同步，但二者的失败与状态必须分别报告 |
+| 生态安装器通常天然采用 multi，不提供 mono/multi 状态机 | DWS 的模式切换保持为重新执行 setup，不新增长期驻留的 mode lifecycle |
+| 市场 Skill 与 CLI 预制 Skill 可能落在同一 Agent 根目录 | 必须使用统一所有权元数据识别受管目录，名称前缀不能作为删除依据 |
+| 多 Skill 更新常以新清单刷新官方集合 | DWS 使用当前 bundle 官方清单全量覆盖，新增 Skill 自动加入，本地删除不视为持久化排除 |
+| 制品可能需要同时服务无运行时依赖、离线和多镜像环境 | DWS 保留 embed、zip 和平台安装脚本，不把单一生态包管理器设为唯一入口 |
+| 中断的复制和原地覆盖容易破坏最后一个可用版本 | 缓存与 Go upgrade 的 Agent 目标采用 staging publish；发布失败自动恢复该目标的完整旧集合 |
+| Agent 通常以 `SKILL.md` 为入口，其他文件按引用或工具规则按需读取 | 安装元数据使用不被内容引用的隐藏文件，并保证其内容不包含 Agent 指令 |
+
+本节只保留可复用的工程结论，不记录具体产品、仓库、版本或逐项能力对照，也不构成
+DWS 对任何外部实现的持续兼容义务。后续设计以 DWS 自身约束和本 RFC 的行为合同为准。
+
+## 4. 布局合同
+
+| 模式 | Agent 目录布局 | 选择方式 |
+|---|---|---|
+| multi（默认） | `<agent-home>/dingtalk-*/` 与必选 `dingtalk-shared/` | 默认；`dws skill setup --mode multi` |
+| mono（兼容） | `<agent-home>/dws/` | `dws skill setup --mode mono` 或安装器的 mono opt-in |
+
+模式切换通过重新执行 setup 完成。安装 multi 前备份并移除 mono 的 `dws/`；安装
+mono 前只备份并移除能够证明由 DWS 管理的 multi 目录。两个方向都不提供隐式、
+不可恢复的删除。
+
+## 5. 官方集合与升级策略
+
+当前版本 bundle 中的 multi 目录清单是升级集合的唯一权威来源。普通 upgrade 和
+`--force` 都安装并覆盖该版本的全部官方预制 Skill：
+
+- 本地删除的预制 Skill 会在下一次升级恢复；
+- setup 时通过 `--exclude` 暂时排除的 Skill 会在下一次升级恢复；
+- 新版本新增的官方 Skill 会自动安装；
+- 用户对预制 Skill 的本地修改会被官方版本覆盖；
+- `dingtalk-shared` 始终随官方集合安装。
+
+`~/.dws/skills-state.json`（设置 `DWS_CONFIG_DIR` 时位于该目录）不参与安装集合
+求解，也不保存排除策略。它既记录结果快照，也集中记录 multi Skill 的所有权和
+provenance，供安全清理、诊断与后续迁移使用。
+
+## 6. 目录所有权
+
+每次 multi setup 或 upgrade 全部成功后，DWS 在统一的
+`~/.dws/skills-state.json` 中写入：
+
+```json
+{
+  "version": "v0.2.14",
+  "official_skills": ["dingtalk-aitable"],
+  "updated_skills": ["dingtalk-aitable"],
+  "managed_skills": [
+    {
+      "name": "dingtalk-aitable",
+      "version": "v0.2.14",
+      "source": "dws-upgrade",
+      "digest": "sha256:<64 个十六进制字符>",
+      "digest_scope": "skill-directory-v1"
+    }
+  ],
+  "updated_at": "2026-08-11T12:34:56Z"
+}
+```
+
+每条 `managed_skills` 记录代表一个由 DWS 管理的官方 Skill。`version` 记录安装该
+副本的 DWS/发布包版本，`source` 记录安装入口，`digest` 是对 bundle 中 Skill 目录
+全部普通文件按相对路径排序后计算的内容摘要。摘要用于诊断和来源追踪，不作为后续
+升级的完整性门禁；用户修改 Skill 内容后，DWS 仍保有明确管理权并能在下一次升级时
+覆盖恢复。
+
+清理 stale Skill 或切换到 mono 时，只接受以下所有权证据：
+
+1. Skill 名称存在于统一状态的 `managed_skills` 中；
+2. 统一状态上线前曾发布过的官方 Skill 精确名称集合。
+
+历史集合是冻结的迁移清单，包含 `dws-shared` 以及已退役、折叠或仍在发布的旧官方
+目录名。仅有 `dingtalk-*` 前缀不构成所有权证据。因此，市场或用户创建的
+`dingtalk-custom` 等非官方精确名称目录不会被迁走。
+
+### 6.1 对 Agent 的影响
+
+Skill 目录内不再放置 DWS 所有权文件，也不增加非通用 frontmatter 字段。支持的
+Agent 仍只需以 `SKILL.md` 发现和加载 Skill；统一元数据位于 Agent Skill 目录之外，
+不会成为提示词上下文或影响 Agent 行为。
+
+## 7. Setup：Plan → Confirm → Execute
+
+`dws skill setup` 分为三个阶段：
+
+1. **Plan**：只读计算目标、安装集合以及所有待备份路径；
+2. **Confirm**：`--dry-run` 和交互确认渲染同一份计划；
+3. **Execute**：确认后严格执行计划中的备份和安装。
+
+安全要求：
+
+- 非交互环境未传 `--yes` 时拒绝执行；
+- 用户拒绝确认时必须零文件写入；
+- 备份失败时跳过整个 Agent 目标，不开始铺设相反布局；
+- 同一目标先完成所有必要备份，再复制新集合；
+- multi Skill 必须在同级 staging 中完成复制，再原子发布到正式目录；
+- 任意 `skipped > 0` 都返回非零退出码，并且不写入完整成功快照；
+- 一个 Agent 目标失败不阻止其他目标尝试，但最终结果仍为失败。
+
+## 8. Upgrade 与恢复语义
+
+升级器对每个 Agent 目标执行：
+
+- 先探测具体 Agent home；只在没有任何具体 Agent 时使用 `~/.agents/skills` 通用 fallback；
+- 具体 Agent 安装成功后，将 `~/.agents/skills` 中旧的 DWS 受管副本可恢复地迁入备份，避免 Codex 等同时扫描两个根目录时重复发现同名 Skill；
+
+1. 只读计算对面布局、过期受管 Skill 和同名官方 Skill；
+2. 在目标文件系统的 staging 中复制完整新集合；
+3. staging 全部成功后，才将旧集合移入备份目录；
+4. 逐项发布 staging；任一发布失败时删除已发布的新目录，并逆序恢复该目标的全部旧目录；
+5. 仅在没有目标失败且至少一个目标成功时更新状态快照。
+
+Go upgrade 当前提供 **单 Agent 目标级事务恢复**：复制失败发生在旧目录移动前；
+备份中途失败会恢复此前已移动的目录；发布中途失败会恢复该目标的完整旧集合。不同
+Agent 目标仍彼此独立，一个目标失败不会回滚此前已经成功升级的其他目标，这与
+“不提供跨所有 Agent 目标的事务式回滚”非目标保持一致。
+
+## 9. 备份合同
+
+- 路径：`~/.dws/skill-backups/<UTC 时间戳>/...`；
+- 主要操作：同一文件系统内使用 rename 移动；
+- 失败语义：备份失败时原目录保持不变，目标安装失败；
+- 可见性：计划和执行日志显示原路径与备份路径；
+- 保留策略：自动修剪，仅保留最近 5 批。
+
+备份是安装安全机制，不等于独立 rollback 产品。需要切回 mono 时重新运行
+`dws skill setup --mode mono`。
+
+## 10. 缓存与制品
+
+发布制品和二进制内嵌内容同时携带 mono 与 multi 源树。`~/.dws/skills/` 只是
+setup 在未显式指定 `--source` 时的本地回退缓存。
+
+缓存刷新必须采用同级 staging + publish：
+
+1. 在 staging 中完整复制并验证新树；
+2. 发布前保留旧缓存；
+3. 通过 rename 发布新缓存；
+4. 复制或发布失败时保留或恢复旧缓存；
+5. 空、缺失或损坏的 bundle 不能擦除有效缓存。
+
+## 11. 安装入口一致性
+
+以下入口都遵守本 RFC：
+
+| 入口 | 默认模式 | 失败合同 |
+|---|---|---|
+| `dws skill setup` | multi | 部分失败返回非零；不写完整成功状态 |
+| `dws upgrade` | bundle 含 multi 时安装 multi | 目标失败返回失败；下次全量重试 |
+| `scripts/install.sh` | multi | 任一检测到的目标失败则脚本非零 |
+| `scripts/install.ps1` | multi | 任一检测到的目标失败则脚本非零 |
+| `scripts/install-skills.sh` | multi | 任一检测到的目标失败则脚本非零 |
+| npm `install.js` | multi | 任一检测到的目标失败则 postinstall 失败 |
+
+Homebrew 不直接向 Agent home 铺设 Skill；安装 CLI 后由 setup 执行相同流程。
+
+## 12. 验收与回归门禁
+
+合入和后续修改至少覆盖：
+
+- mono → multi、multi → mono 互斥切换；
+- 状态上线前的官方 multi 目录切换 mono 时能够被精确迁移；
+- 未登记的同前缀市场/用户 Skill 在刷新和切换后仍存在；
+- 统一状态中登记的过期官方 Skill 被备份并移除；
+- 备份、复制、统一状态写入、缓存 publish 故障注入；
+- 非交互确认拒绝与显式 `--yes`；
+- 部分失败返回非零且不写错误状态快照；
+- 复制失败不留下 Agent 可见的残缺官方目录；
+- 普通 upgrade 恢复被删除的预制 Skill，并安装新增官方 Skill；
+- Windows、macOS、Linux 的路径和覆盖率门禁；
+- npm、Shell、PowerShell 与包管理器安装冒烟。
+
+## 13. 后续演进
+
+- 收敛各安装入口中的 Agent home 清单，减少跨语言复制；
+- 如确有运维需求，可单独设计备份查看和显式恢复命令；
+- mono 的物理删除必须作为独立变更，在 multi 内容、安装入口和迁移回归稳定后推进；
+- `managed_skills` 字段若演进，必须同步更新所有安装入口和跨平台回归。
@@ -6,7 +6,7 @@

 ## 第一步：安装 dws

-一键脚本会自动下载最新版二进制 + `dingtalk-dev` skill，只需要 curl（无需 go / git）。
+一键脚本会自动下载最新版二进制 + `dingtalk-misc` skill（开放平台应用文档落在 misc），只需要 curl（无需 go / git）。

 ### macOS / Linux

@@ -231,7 +231,8 @@ Cobra hard-required 是独立的 executable fact，并通过 `cli_required`/prov
 | 问题 | 事实源 |
 |---|---|
 | 当前二进制是否暴露命令、Cobra 接受哪些 flags | `dws <path> --help` |
-| Agent 选哪个命令、参数映射/required/约束、risk/confirmation | 对应 leaf `dws schema "<path>"` |
+| Agent 选哪个命令、CLI 参数/required/约束、risk/confirmation | Agent leaf `dws schema "<path>" --compact` |
+| CLI↔RPC 参数映射、接口绑定、provenance | full leaf 配合 `--jq` / `--fields` 精确投影 |
 | 钉钉中的文档、文件、日程、消息等实际数据 | 真正执行 `dws doc read`、`dws drive search` 等 read/search/list 命令 |

 Schema 和 Help 冲突是契约漂移，不能静默猜测：
@@ -246,10 +247,10 @@ Schema 和 Help 冲突是契约漂移，不能静默猜测：

 ```bash
 dws schema                                      # 产品紧凑概览
-dws schema calendar                             # 产品摘要
-dws schema "calendar event"                    # 分组摘要
-dws schema "calendar event create"             # 完整 leaf
-dws schema "calendar event create" --compact   # 支持：裁掉 provenance/debug 字段
+dws schema calendar --compact                   # Agent 产品摘要
+dws schema "calendar event" --compact          # Agent 分组摘要
+dws schema "calendar event create" --compact   # Agent leaf
+dws schema "calendar event create"             # full leaf，仅用于映射/provenance 审计
 dws schema --all                               # 所有工具的完整 leaf 导出
 ```

@@ -257,7 +258,7 @@ dws schema --all                               # 所有工具的完整 leaf 导

 `schema --all` 必须包含最终 `SchemaIndex` 中每个 tool 的完整 leaf 参数、约束和安全语义；无业务参数的命令也要包含空 `parameters` 对象。它用于审计、CI 和参数防丢 baseline，但输出很大，普通 Agent 命令发现不得使用，应按 overview -> product/group -> leaf 渐进查询。

-`--compact` 当前受支持，适合减少常规 leaf 查询上下文。`schema --all --compact` 也可执行，但会移除 provenance/debug 和接口映射字段，不能作为完整兼容性 baseline。
+`--compact` 是普通 Agent 查询的规范视图：通过正向字段白名单保留选参、约束与安全语义，full 新增字段不会自动进入 Agent 上下文。省略它的 leaf 包含参数 property、接口绑定和 provenance，只用于定向审计；`schema --all --compact` 也可执行，但不能作为完整兼容性 baseline。

 兼容旧二进制时，如果 Schema 查询返回 `unknown_flag: --compact`，只去掉 `--compact` 重试同一个查询。这是展示能力降级，不代表 leaf 缺失，也不能改用 Schema 查询业务数据。

@@ -1,187 +0,0 @@
-# lark-cli Shortcut 深度对齐矩阵
-
-> 12 个 agent 逐条深读 lark 每个 shortcut 的智能实现（Validate/DryRun/ID解析/投影/多步/分页），映射钉钉、标注保真度差距。
-
-## 2026-07-13 最新源码复核
-
-对比基线：
-
- DWS：`feature/shortcut@b7c14c1`（已合并 `origin/main@390b611`）
- lark-cli：`main@e96c4fa5`
- lark-cli 本轮更新范围：`f495cbb1..e96c4fa5`
-
-本轮 lark-cli **没有增加或删除生产 shortcut 命令**，变化集中在已有命令的实现保真度：统一 `--json` shorthand、文档分享锚点读取、whiteboard 本地文件安全内联、VC meeting events 的 identity/timeline/NDJSON 投影、Apps DB 环境自动选择、Drive push 错误分类，以及 Wiki token 解析兼容性。因此下方历史 gap 清单的命令面没有因本轮 pull 新增条目，但若要追平体验，以下实现差距需要上调优先级。
-
-### 当前命令面快照
-
-| 指标 | 数量 | 说明 |
-|---|---:|---|
-| DWS built-in shortcut | 366 | 16 个服务；运行时 registry 实测 |
-| lark-cli primary shortcut | 363 | 19 个服务；排除 `_test.go` 与 42 个 `sheets/backward` 隐藏兼容别名 |
-| 双方可映射服务内命令 | DWS 313 / lark 324 | 12 组产品映射，不含平台特有服务 |
-| 同服务同名命令 | 50 | 仅是名称交集，不等于语义等价或保真度一致 |
-| DWS 平台特有 shortcut | 53 | attendance / ding / oa / report 等 |
-| lark 平台特有 shortcut | 39 | okr / vc / slides / markdown / whiteboard / note / event |
-
-双方重叠服务的命令面如下；“同名”只用于定位，能力判断仍需看参数、验证、多步编排、输出投影和 dry-run：
-
-| 产品映射 | DWS | lark | 同名 |
-|---|---:|---:|---:|
-| aitable ↔ base | 82 | 87 | 31 |
-| calendar ↔ calendar | 23 | 10 | 3 |
-| chat ↔ im | 89 | 21 | 2 |
-| contact ↔ contact | 16 | 2 | 1 |
-| devapp ↔ apps | 30 | 63 | 3 |
-| doc ↔ doc | 19 | 14 | 1 |
-| drive ↔ drive | 9 | 26 | 3 |
-| mail ↔ mail | 10 | 21 | 0 |
-| minutes ↔ minutes | 13 | 9 | 1 |
-| sheet ↔ sheets | 2 | 42 | 0 |
-| todo ↔ task | 13 | 17 | 2 |
-| wiki ↔ wiki | 7 | 12 | 3 |
-
-### 最新优先差距
-
-1. **文档与白板资源保真度**：lark `doc +fetch/+update` 已支持分享链接 selection anchor、HTML5 block 资源引用，以及相对路径内的 SVG/Mermaid/PlantUML whiteboard 安全内联。DWS 具备文档读写和媒体原子能力，但缺少统一引用解析、路径门禁和资源回写编排。
-2. **Sheets typed workflow**：lark 的 typed table、批量样式、维度移动/冻结、range copy/fill/sort、workbook import/export 仍是最大可建设缺口。DWS 原生 helper 已有部分底层能力，但 shortcut 层只有 2 个精选命令，缺少跨 sheet 分块写、类型推断和 partial rollback。
-3. **Drive 本地同步体验**：lark `+push/+pull/+sync/+import/+export` 带批量计划、错误分类、路径保护和版本操作；DWS 目前偏原子上传/搜索，缺完整目录同步和可恢复批处理。
-4. **Mail 高保真写链路**：lark 对 send/reply/reply-all/forward 提供模板、签名、HTML lint、线程头、定时和附件编排；DWS 有底层发信/草稿工具，但 smart shortcut 尚未覆盖这些组合体验。
-5. **消息资源与统一搜索**：DWS 已有 `+search-msg/+chat-messages/+thread-replies/+at-me` 等拆分场景，lark `+messages-search` 仍在统一多维过滤、会话上下文富化、reaction/资源下载方面更完整。
-6. **会议事件输出**：lark `vc +meeting-events` 本轮新增当前身份、actor、会议状态推断、timeline 与 NDJSON 元数据。DWS 最新 main 已有更强的实时 event bus 和个人事件订阅，但尚未沉淀成同等级 shortcut 投影；这是“底层能力领先、shortcut UX 未收口”。
-
-### 不建议机械追平
-
- lark Apps DB、Spark 发布、Lark Drive/Wiki 特有对象模型属于平台差异，不应只为同名率复制。
- DWS 的 attendance、DING、OA、report、agoal 和最新 event bus 是钉钉侧差异化能力，应优先做场景化组合，而不是追求 363 vs 366 的数字对齐。
- DWS 已具备按姓名解析、跨产品智能编排、失败回滚和 usage→自定义 shortcut 沉淀闭环，这些能力无法由同名命令统计体现。
-
-> 注：下方“361 条”汇总是上一轮逐条人工分类的历史基线；当前 lark-cli primary shortcut 是 363 条，另有 42 个不应重复计为能力的 Sheets 隐藏兼容别名。历史条目的判断仍可复用，但总量数字不能直接代表本轮最新覆盖率，后续应把新增条目按 covered-1to1 / covered-smart / gap-buildable / no-dingtalk-tool 四类补录。
-
-## 汇总（361 条 lark shortcut）
-
-| dws_status | 数量 | 含义 |
-|---|:---:|---|
-| covered-1to1 | 144 | lark 组合在钉钉塌缩成 1:1，封装层已覆盖 |
-| no-dingtalk-tool | 127 | 钉钉无对应工具，客观不可对齐 |
-| **gap-buildable** | **41** | 钉钉有工具、值得补成智能 shortcut（**建设目标**）；已建 minutes `+detail`/`+replace-batch`、base `+record-share-links`/`+resolve-base`、im `+thread-replies`/`+chat-messages`/`+chat-list`、task `+related-tasks` |
-| covered-smart | 49 | 已建智能 shortcut / 部分覆盖 |
-
-## 🎯 gap-buildable 目标清单（原 49 条，已建 8 → 剩 41，按服务）
-
-> 已落地：minutes `+detail`（✅ smart `+detail`）、minutes `+word-replace`（✅ smart `+replace-batch`，批量+去重）、base `+record-share-link-create`（✅ smart `+record-share-links`，>20 去重+分片+合并）、im `+threads-messages-list`（✅ smart `chat +thread-replies`，list_topic_replies + 投影）、im `+chat-list`（✅ smart `chat +chat-list`）、task `+get-related-tasks`（✅ smart `todo +related-tasks`，三角色并集+去重+投影）。
-
-### im → chat（5）
-
-| lark 命令 | risk | 保真度差距（钉钉有 tool，缺什么智能） |
-|---|---|---|
-| `+chat-list` ✅ | read | **已建 smart `chat +chat-list`**：`list_all_conversations` + 默认仅群聊 + `--types group/p2p` 当前页过滤 + `--exclude-muted` + page-size/page-token 别名 + openConversationId/name/conversationType 投影。剩余未做：sort/sort-type、bot 身份 p2p 剥离（DWS 无对应身份模型） |
-| `+chat-messages-list` ✅ | read | **已建 smart `chat +chat-messages`**：群/单聊 list_conversation_message_v2 / list_individual_chat_message 互斥 + sender/text/time 投影。剩余未做：reactions 富化、资源下载 |
-| `+chat-search` | read | dws 无群名模糊搜索v2对应 tool(search_common_groups/find 语义不同)，缺 query规范化、mode映射、mute过滤、meta投影 |
-| `+messages-resources-download` | write | dws download-media 走 get_resource_download_url 拿URL，缺分片Range下载/重试/扩展名推断/安全落盘路径校验 |
-| `+messages-search` | read | dws 有 search_messages_by_keyword/by_time_range/by_sender/at_me 多个原子 tool，但各自单点，缺统一多维filter编排+mget+chat上下文富化+跨字段Validate |
-| `+threads-messages-list` ✅ | read | **已建 smart `chat +thread-replies`**：list_topic_replies + sender/text/time 投影。剩余未做：reactions 富化、资源下载 |
-
-### task → todo（3）
-
-| lark 命令 | risk | 保真度差距（钉钉有 tool，缺什么智能） |
-|---|---|---|
-| `+reminder` | write | dws 有 add_todo_reminder/reset_todo_reminder 但无 lark 的先查现有再替换编排、相对时间(15m/1h)解析与互斥校验，值得补智能 shortcut |
-| `+get-related-tasks` ✅ | read | **已建 smart `todo +related-tasks`**：creator+executor+participant 三角色并集 + taskId 去重 + 投影。剩余未做：followed-by-me 成员比对、subtask_count/tasklists 富投影 |
-| `+upload-attachment` | write | dws add-attachment 走 init→PUT→commit 三步 MCP 上传(能力更重)，但无 50MB/regular 校验、applink 提取与 dry-run 计划展示；可对齐成更智能 shortcut |
-
-### calendar → calendar（1）
-
-| lark 命令 | risk | 保真度差距（钉钉有 tool，缺什么智能） |
-|---|---|---|
-| `+room-find` | read | dws 有 room search(query_available_meeting_room 按单一时间段+过滤)和 busy search，但无多slot并发room_find聚合、无city/building/floor/capacity维度过滤、无按attendee推荐可用室，值得补成智能 shortcut 但未建 |
-
-### doc (docs) → doc（2）
-
-| lark 命令 | risk | 保真度差距（钉钉有 tool，缺什么智能） |
-|---|---|---|
-| `+media-insert` | write | dws doc media insert 为3步(取凭证→PUT→insert_document_block)无回滚、无selection定位、无剪贴板、无宽高比补算、无wiki解析；可补成带回滚的智能shortcut |
-| `+media-download` | read | dws doc media download 走resourceId→downloadUrl两段,缺whiteboard导图分支、自动扩展名、路径安全、overwrite防护；media分支可对齐，whiteboard无工具 |
-
-### drive → drive（1）
-
-| lark 命令 | risk | 保真度差距（钉钉有 tool，缺什么智能） |
-|---|---|---|
-| `+import` | write | dws drive upload 有 --workspace --convert 可转在线文档，但缺按目标类型(docx/sheet/bitable/slides)导入、缺 target-token 挂载与异步轮询 |
-
-### mail → mail（4）
-
-| lark 命令 | risk | 保真度差距（钉钉有 tool，缺什么智能） |
-|---|---|---|
-| `+reply` | write | dws reply 走 create_reply_draft+send_draft 两步、附件仅上传会话，缺 EML 线程头构造、签名自动注入、模板合并、HTML lint、读回执、send-time 定时、跨字段校验 |
-| `+reply-all` | write | dws reply-all 两步且收件人由服务端决定，缺原文收件人抽取去重排己、线程头、签名/模板/lint/定时等编排保真 |
-| `+send` | write | dws send_email 单步(附件时先 create_draft 再传再 send)，缺签名/模板/lint/日历内嵌/定时发送/发件人profile解析/跨字段校验 |
-| `+forward` | write | dws forward 走 create_forward_draft+send_draft，缺 Fw:主题/引用块/原附件转载 EML 构建、签名/模板/lint/定时保真 |
-
-### wiki → wiki（1）
-
-| lark 命令 | risk | 保真度差距（钉钉有 tool，缺什么智能） |
-|---|---|---|
-| `+node-get` | read | dws 无 get_node 对应 tool(proxy wiki doc read 读的是文档正文而非节点元数据/space解析)；缺 token/obj_token/URL→node 解析、obj_type推断、space交叉校验——是值得补的智能 shortcut 缺口 |
-
-### minutes → minutes（4）
-
-| lark 命令 | risk | 保真度差距（钉钉有 tool，缺什么智能） |
-|---|---|---|
-| `+search` | read | dws list_by_keyword_and_time_range 只按 keyword+时间+归属(created/shared)过滤，缺 owner/participant 的 me 解析与筛选、缺 query 长度与跨字段互斥校验、缺输出投影与去头像 |
-| `+download` | read | dws 只有 query_minutes_audio_url 返回 OSS 地址(相当于 --url-only 单条)，缺真正落盘下载、批量 fanout+限速+去重、文件名推断、SSRF 防护与覆盖保护 |
-| `+word-replace` ✅ | write | **已建 smart `+replace-batch`**：多组 `原文=>替换` 批量替换 + 去重校验 + 逐组结果聚合（补齐 1:1 `+word-replace` 的单组限制）。剩余未做：@file/stdin 输入 |
-| `+detail` ✅ | read | **已建 smart `+detail`**：单命令按 `--artifacts` fanout basic/summary/keywords/transcript/todos + partial-failure 容错 + rt.Output 投影。剩余未做：wait-ready 轮询、transcript 落盘 |
-
-### base → aitable（10）
-
-| lark 命令 | risk | 保真度差距（钉钉有 tool，缺什么智能） |
-|---|---|---|
-| `+title-resolve` ✅ | read | **已建 smart `aitable +resolve-base`**：search_bases 按名解析 baseId + 0/1/多候选消歧投影。剩余未做：Drive doc_wiki 全文搜索 |
-| `+field-create` | write | dws create_fields 支持批量,但缺 formula/lookup guide-ack 门禁与逐字段节流,可补智能 shortcut |
-| `+field-update` | write | dws update_field 缺 formula/lookup guide-ack 保护 |
-| `+record-share-link-create` ✅ | read | **已建 smart `+record-share-links`**：>20 条记录去重 + 分片(≤20/批) + 跨 aitable-helper server fanout + 合并 {recordId,shareUrl}，补齐单批 20 条上限 |
-| `+record-upload-attachment` | write | dws 只有 prepare_attachment_upload(拿上传凭证),缺 分片上传编排+append_attachments 回填单元格的完整链路 |
-| `+dashboard-block-list` | read | dws 仪表盘块是 chart(create/get/update/delete_chart),缺通用 block list,可对齐补 |
-| `+dashboard-block-get` | read | dws get_chart 覆盖 chart 类块,缺通用 block get |
-| `+dashboard-block-create` | write | dws create_chart 覆盖图表块,缺其他 block 类型的通用创建 |
-| `+dashboard-block-update` | write | dws update_chart 覆盖图表块更新 |
-| `+dashboard-block-delete` | high-risk-write | dws delete_chart 覆盖图表块删除 |
-
-### sheets → sheet（14）
-
-| lark 命令 | risk | 保真度差距（钉钉有 tool，缺什么智能） |
-|---|---|---|
-| `+sheet-hide` | write | dws update_sheet可能含hidden属性但未见独立hide命令,需确认 |
-| `+sheet-unhide` | write | 同上,dws无独立unhide命令 |
-| `+sheet-set-tab-color` | write | dws update_sheet或可设tab色但无独立命令 |
-| `+sheet-show-gridline` | write | dws无网格线显隐命令 |
-| `+sheet-hide-gridline` | write | dws无网格线显隐命令 |
-| `+workbook-create` | write | dws有create_workspace_sheet但仅建空表,缺typed一步建表+填充+样式+partial回滚编排 |
-| `+dim-hide` | write | dws update-dimension或含hidden但无独立hide命令 |
-| `+dim-unhide` | write | 同上,dws无独立unhide命令 |
-| `+dim-freeze` | write | dws update-dimension可能含frozen但无独立freeze命令 |
-| `+cells-get` | read | dws range read存在但缺include样式/公式投影统一封装 |
-| `+table-get` | read | dws缺typed table读回+列类型推断+多sheet编排,只有裸csv/range读 |
-| `+table-put` | write | dws有append/set_cell_range但缺typed多sheet分块写+建缺失sheet+样式+partial回滚编排 |
-| `+rows-resize` | write | dws update-dimension可调尺寸但无独立rows-resize+size/type互斥校验 |
-| `+cols-resize` | write | dws update-dimension可调尺寸但无独立cols-resize+互斥校验 |
-
-### apps → devapp（3）
-
-| lark 命令 | risk | 保真度差距（钉钉有 tool，缺什么智能） |
-|---|---|---|
-| `+release-create` | write | dws 有 create_dev_app_version(开放平台版本)可类比，但妙搭 release 是低代码应用发布、语义与产物不同 |
-| `+release-get` | read | dws 有 get_dev_app_version_detail 可类比但产品域(开放平台vs妙搭)不同 |
-| `+release-list` | read | dws 有 list_dev_app_versions 可类比但无 status 枚举过滤且产品域不同 |
-
-## 已建智能 shortcut（covered-smart，48）— 可继续升级保真度
-
- **im**: +chat-members-list +chat-list +messages-send +threads-messages-list
- **task**: +complete +assign +get-my-tasks +get-related-tasks
- **contact**: +search-user
- **calendar**: +agenda +create +update +freebusy +suggestion
- **doc (docs)**: +history-revert
- **drive**: +upload +search +inspect
- **mail**: +triage
- **minutes**: +upload +latest-minutes +action-items +transcript +minutes-search +detail +replace-batch
- **base**: +table-get +table-create +view-create +view-get-filter +view-set-filter +view-get-visible-fields +view-set-visible-fields +view-get-group +view-set-group +view-get-sort +view-set-sort +view-get-timebar +view-set-timebar +view-get-card +view-set-card +record-list +record-search +record-get +record-upsert +base-create +workflow-list +form-create +form-list +form-get +record-share-link-create
@@ -1,272 +0,0 @@
-# Skill 分发/安装架构优化方案
-
-> 目标：把 DWS Agent skill 的"装在哪、怎么装、装完记什么"从 7+ 个安装面的
-> N 份漂移拷贝，收敛为 **Go 侧单一事实源 + 单一安装引擎 + 脚本侧 bootstrap**。
-> 本文基于 feat/skill-mode-migration 工作区（multi 化落地后）的代码级盘点
-> （2026-08-05），所有锚点均可跳转验证。
-> 前置文档：[skill-multi-migration-plan.md](skill-multi-migration-plan.md)（下称《迁移计划》）、
-> [skill-distribution-mechanism.md](skill-distribution-mechanism.md)。
-
-## 1. 问题量化
-
-### 1.1 复制矩阵（能力 × 面）
-
-multi 化之后，同一块逻辑在各安装面的拷贝数（✓ = 独立拷贝一份，数字 = 份数）：
-
-| 能力 \ 面 | install.sh | install.ps1 | npm install.js | install-skills.sh | install-event.sh | install-devapp.sh(+ps1) | Homebrew caveats | `dws skill setup` (Go) | `dws upgrade` (Go) |
-|---|---|---|---|---|---|---|---|---|---|
-| agent home 清单 | ✓×3 | ✓ | ✓ | ✓×2 | ✓ | ✓×2 | — | ✓ | ✓ |
-| 互斥清理（mono↔multi） | ✓×2（其中 1 份死代码） | ✓ | ✓ | ✓ | — | — | — | ✓ | ✓ |
-| mode 解析（env/flag/交互） | ✓ | ✓ | ✓ | — | — | — | — | ✓ | 布局嗅探 |
-| stale `dingtalk-*` 清理 | ✓ | ✓ | ✓ | ✓ | — | — | — | ✗（additive，不清） | ✓ |
-| `~/.dws/skills` 缓存写 | ✓ | ✓ | ✓ | — | ✓ | ✓ | — | —（只读回退） | ✓（仅 multi） |
-| 安装状态（state.json） | 无 | 无 | 无 | 无 | 无 | 无 | — | 无 | 无 |
-
-清单锚点（agent home 清单，全仓共 **15 份**）：
-
-| # | 位置 | 备注 |
-|---|---|---|
-| 1 | `internal/app/skill_setup.go:20-37` | `skillSetupAgentHomes`，16 项，**无 opencode** |
-| 2 | `internal/app/skill_command.go:119-141` | `agentSkillPaths`，17 项，含 opencode（`:129`） |
-| 3 | `internal/upgrade/paths.go:35-52` | `knownSkillDirs`，16 项，**无 opencode** |
-| 4 | `scripts/install.sh:370-386` | mono 安装循环内联清单 |
-| 5 | `scripts/install.sh:439-455` | multi 清单（**死代码**，见 1.2） |
-| 6 | `scripts/install.sh:516-532` | multi 清单（生效） |
-| 7 | `scripts/install.ps1:44-61` | `$AgentDirs` |
-| 8 | `build/npm/install.js:11-28` | `AGENT_DIRS` |
-| 9 | `scripts/install-skills.sh:166-182` | multi 清单 |
-| 10 | `scripts/install-skills.sh:247-261` | mono 清单 |
-| 11 | `scripts/install-event.sh:103-107` | 17 项，**含 `.config/opencode/skills`**（`:107`） |
-| 12 | `scripts/install-devapp.sh:87-91` | 含 opencode（`:91`） |
-| 13 | `scripts/install-devapp.ps1:106` | 含 opencode |
-| 14 | `test/scripts/package_script_test.go` | `expectedPackagedSkillTargets` |
-| 15 | `scripts/release/verify-package-managers.sh:75-76` | `HOME_AGENT_PARENTS` / `HOME_SKILL_TARGETS` |
-
-互斥清理逻辑共 **≥8 份**：`skill_setup.go:570-608`、`paths.go:306-327` +
-`paths.go:163-179`（mono 分支内联）、`install.sh:394-399` + `install.sh:557-570`
-+ 死代码 `install.sh:477-486`、`install.ps1:488-498` + `install.ps1:562-570`、
-`install.js:130-137` + `install.js:165-178`、`install-skills.sh:207-220`。
-
-mode 解析共 **4 份**：`skill_setup.go:343-376`、`install.sh:220-257`、
-`install.ps1:293-332`、`install.js:197-214`（各自实现 env 优先级、非法值报错、
-TTY 交互、非 TTY 默认值，措辞与行为细节已不完全一致）。
-
-stale `dingtalk-*` 清理共 **5 份**：`paths.go:317-325`、`install.sh:561-567`、
-`install.ps1:562-570`、`install.js:168-172`、`install-skills.sh:211-220`。
-
-缓存写共 **6 份**：`install.sh:326-344`（multi）+ `install.sh:349-360`（mono）、
-`install.ps1:446-471`、`install.js:221-237`、`install-event.sh:165-169`、
-`install-devapp.sh:82-84`、`paths.go:290-296`（upgrade 只刷 multi）。
-
-### 1.2 已发生的漂移实例（不是假设，是现状）
-
-1. **opencode 清单漂移**：`dws skill install` 支持 opencode 目标
-   （`skill_command.go:129`），install-event / install-devapp 也往
-   `~/.config/opencode/skills` 装（`install-event.sh:107`、
-   `install-devapp.sh:91`），但 `dws skill setup --target all` 与
-   `dws upgrade` 的清单都不含它（`skill_setup.go:20-37`、
-   `paths.go:35-52`）。结果：opencode 用户能收到 event/dev skill，
-   却永远收不到内置 skill 的安装与升级。
-2. **install.sh 死代码**：`install_multi_skills_to_homes` 定义了两次
-   （`install.sh:434` 与 `install.sh:511`），`_install_multi_to_base`
-   同样两次（`install.sh:473` 与 `install.sh:549`）。bash 后定义覆盖先定义，
-   第一对（434-505）整体是死代码——本次 multi 化自己引入的重复。
-3. **"镜像"注释与实现已不符**：`install.sh:507-510` 注释声称 multi 安装
-   "mirroring `dws skill setup --mode multi`"，但脚本会删除 bundle 里不存在
-   的 stale `dingtalk-*`（`install.sh:561-567`），而 setup 是 additive 语义、
-   只清 `dws/`（`skill_setup.go:587-593`）。注释说镜像，行为已分叉。
-4. **互斥语义自相矛盾**：install-event.sh 把 multi 的 `dingtalk-event` 和
-   mono 的 `dws` **同时**装进同一批 agent home（`install-event.sh:171-172`），
-   与其他所有面"mono/multi 互斥"的语义直接冲突。
-5. **缓存只写不读、版本不可见**：`dws skill setup` 默认走二进制 embed
-   （`skill_setup_embed.go:50-59`），`~/.dws/skills` 只是 legacy 回退候选
-   （`skill_setup.go:442-447`）；6 个写入方没有任何版本戳，upgrade 只刷
-   multi 不刷 mono（`paths.go:292-296`），缓存与 embed 漂移不可见。
-6. **保留前缀规则已双份**：`dingtalk-` / `dws-shared` 的保留约定在
-   `skill_setup.go:199`（`multiSkillPrefix`）、`skill_setup.go:205`
-   （`multiSharedSkill`）与 `paths.go:332-334`（`isMultiSkillDirName`）
-   各写一份，互斥清理依赖"市场 skill 不用该前缀"这一隐性约定。
-
-### 1.3 维护成本论证（本次 multi 化实证）
-
-本次"把 7 个面全部 multi 化"这一个语义变更，工作区改动为 **13 个文件、
-+759/−166 行**（`git diff --stat`），横跨 Go / POSIX sh / PowerShell /
-JavaScript 四种语言：
-
-| 文件 | 改动量 | 改了什么 |
-|---|---|---|
-| `scripts/install.sh` | +256/−… | mode 解析、multi 安装×2（含死代码）、互斥清理、缓存写 |
-| `scripts/install.ps1` | +169 | 同上，PowerShell 再写一遍 |
-| `build/npm/install.js` | +87 | 同上，JavaScript 再写一遍 |
-| `scripts/install-skills.sh` | +101 | multi 安装 + stale 清理 |
-| `internal/upgrade/paths.go` | +206 | upgrade multi 刷新 + 互斥清理 + 缓存 |
-| `internal/app/skill_setup.go` 等 Go 侧 | ~+50 | setup multi 语义 |
-| 测试与文档 | 其余 | 4 个测试文件 + README×2 + SKILL.md |
-
-即：**一个语义变更 = 4 种语言 × 7+ 处同步编辑**，且仍然漏了 opencode、
-制造了死代码、留下了语义分叉（1.2）。新增一个 agent home 今天要改 15 份
-清单中的至少 11 份（另 4 份是测试/校验）。这不是"以后会漂移"，而是
-**每一次改动都在当场制造漂移**。
-
-## 2. 目标架构
-
-### 2.1 分层总览
-
-```text
-┌──────────────────────────────────────────────────────────────────┐
-│ L3 分发面（7 个，全部退化为 bootstrap）                            │
-│   install.sh / install.ps1 / npm install.js / install-skills.sh /│
-│   install-event.sh / install-devapp.sh / Homebrew caveats        │
-│   职责：装二进制 → exec `dws skill setup --mode X --yes`           │
-│         └─ 失败 → 冻结的内嵌 fallback 拷贝（一档，不再演进）        │
-├──────────────────────────────────────────────────────────────────┤
-│ L2 单一安装引擎（Go，唯二入口）                                    │
-│   dws skill setup  ─┐                                            │
-│   dws upgrade      ─┴─► 共用同一 install 实现（含互斥清理/记账）   │
-├──────────────────────────────────────────────────────────────────┤
-│ L1 单一事实源：internal/skillhome（新共享包）                      │
-│   AgentHomes() 清单 + 布局规则（mono→<home>/dws，multi→平铺）      │
-│   + 互斥/ stale 清理规则 + 保留前缀常量 + state.json 读写          │
-├──────────────────────────────────────────────────────────────────┤
-│ L0 事实数据                                                       │
-│   //go:embed skills/{mono,multi}（skills_embed.go:28，默认源）     │
-│   ~/.dws/skills/state.json（权威安装状态）                         │
-│   ~/.dws/skills/{mono,multi}（显式回退源，带版本戳，唯一写方=Go）  │
-└──────────────────────────────────────────────────────────────────┘
-        ▲ 门禁：scripts/policy/check-agent-homes-sync.sh（进 make policy）
-        比对各脚本可解析清单块 ↔ skillhome 导出清单
-```
-
-### 2.2 单一事实源：`internal/skillhome` 共享包
-
-新建 `internal/skillhome`，把今天散在 3 处 Go 代码里的规则合并导出，
-setup / upgrade / 测试共用（对应《迁移计划》P0c-1）：
-
-| 导出物 | 收敛的现有拷贝 |
-|---|---|
-| `AgentHomes()`（含 opencode，带"首项必装/其余父目录门控"元数据） | `skill_setup.go:20-37`、`paths.go:35-52`，并与 `skill_command.go:119-141` 互相断言 |
-| `HomeForMode(base, mode)` 布局规则 | `skill_setup.go:502-507`（`agentHomeForMode`） |
-| `MutualExclusionVictims(home, mode)` / stale 判定 | `skill_setup.go:570-596`、`paths.go:306-327` |
-| `ReservedSkillPrefixes` / `ReservedSkillNames` 常量 | `skill_setup.go:199/205`、`paths.go:332-334` |
-| `State` 读写（state.json，见 2.4） | 新增（《迁移计划》P0b-1） |
-
-脚本侧不生成代码（sh/ps1/js 三语言片段生成成本高、措辞差异大），改为
-**可解析标记块 + policy 比对**：每个脚本的清单包在
-`# DWS-AGENT-HOMES-BEGIN` / `# DWS-AGENT-HOMES-END` 注释块内，新增
-`scripts/policy/check-agent-homes-sync.sh` 提取 7 个脚本块与
-`skillhome` 导出清单（`go run ./internal/skillhome/cmd/dump` 或
-`dws skill setup --print-agent-homes` 之类调试出口）逐一比对，挂进
-`make policy`（Makefile:86-97 现有政策链）。对应《迁移计划》P0c-2。
-
-### 2.3 安装逻辑单引擎化：脚本退化为 bootstrap
-
-**终态形态**：每个安装脚本只做两件事——装二进制、执行
-`dws skill setup --mode <resolved> --yes`（mode 解析仍允许脚本做，因为它
-要处理自己的 env/flag；也可以更进一步把 `DWS_SKILL_MODE` 透传给 setup）。
-执行失败（二进制跑不起来、setup 非 0 退出）时，降级到**冻结的内嵌
-fallback 拷贝逻辑一档**——即今天的拷贝实现原样保留但标记"不再演进"，
-语义变更只改 Go 引擎。
-
-各面降级可行性分析：
-
-| 面 | 二进制可得性 | 可行性 | 风险与对策 |
-|---|---|---|---|
-| install.sh | `main` 先 `install_binary` 后 `install_skills`（`install.sh:830-831`），二进制就在 `$INSTALL_DIR` | ✅ 直接调 `"$INSTALL_DIR/dws" skill setup --mode "$SKILL_MODE" --yes` | `DWS_SKILLS_ONLY=1` 时不装二进制 → 先 `command -v dws`，无则走 fallback。curl 下载不带 quarantine 属性，macOS 可直接执行。风险低 |
-| install.ps1 | 同序（`Install-Binary`:337 → `Install-Skills`:622，main 入口 `:700`） | ✅ 调 `& $installDir\dws.exe skill setup ...` | ExecutionPolicy 约束的是 .ps1 脚本本身，**子进程 dws.exe 不受其限制**；AppLocker/WDAC 环境可能拦未签名二进制——但那种环境 dws 本身也跑不了，fallback 拷贝仍必要。风险低-中 |
-| npm install.js | postinstall 已把平台二进制解到 `vendor/`（`install.js:260`）再装 skill（`:271-280`） | ✅ `execFileSync(path.join(vendorDir,'dws'), ['skill','setup','--mode',m,'--yes'])` | `npm i --ignore-scripts` 时 postinstall 整体不跑（现状如此，非新增风险）；部分 CI 以 root 跑生命周期脚本权限怪异；Windows 用 `dws.exe`。风险中，fallback 必须保留 |
-| install-skills.sh | skills-only 面，不装二进制 | ✅ `command -v dws` 有则调引擎，无则 fallback | 本质是"刷新 skill"路径，有 dws 才谈得上刷新。风险低 |
-| install-event.sh | 装二进制 + 单 skill（dingtalk-event + mono dws） | ⚠️ 有条件 | `EVENT_VERSION` 可与已装二进制版本不同（尤其 `DWS_SKILLS_ONLY=1`），embed 与目标 zip 可能错配；且当前同时装 mono+multi 的语义（`:171-172`）需先按 1.2-4 收敛。**最后迁移**，迁移前先把 dingtalk-event 纳入 embed 并改成 `-s event` 调引擎 |
-| install-devapp.sh(+ps1) | 同上（dingtalk-dev） | ⚠️ 有条件 | 同 install-event。风险中-高，最后迁移 |
-| Homebrew caveats | 只打印提示 | ✅ 已是终态 | `homebrew.rb.tmpl:33-36` 现状就是"Run `dws skill setup`"，无需改，是其他面的样板 |
-
-配套收益：《迁移计划》P2-2（sh/ps1 的 multi 分支真装，1.5d）与 P2-3
-（install.js / install-skills.sh 加 mode 支持，1d）在单引擎方案下**大幅
-缩水**——脚本只需把 mode 透传给引擎，不再在 sh/ps1/js 里写安装逻辑。
-
-### 2.4 状态与缓存
-
- **state.json 权威化**（《迁移计划》P0b 原样采纳）：
-  `~/.dws/skills/state.json` 记录 `schema_version / mode / cli_version /
-  installed / agent_homes / previous`，写方只有 `dws skill setup` 与
-  `dws upgrade`（脚本侧不再各自写状态——它们调引擎，引擎记账）；缺失/损坏
-  时按磁盘形态反推（有 `dws/` → mono；有 `dingtalk-*` → multi；都有 →
-  报 drift 要求显式收敛）。
- **`~/.dws/skills` 缓存收敛**：二选一，本文建议后者——
-  1. ~~被 state 取代，删除缓存~~：激进；embed 之外的"无源码机回退源"场景
-     （`skill_setup.go:442-447` 注释描述的场景）会断。
-  2. **明确为 setup 回退源 + 版本戳**（采纳）：缓存目录写入
-     `cli_version` 戳文件（或并入 state.json 的 `cache_version` 字段）；
-     **写缓存收敛到 Go 侧唯一实现**（setup/upgrade 共用），install.sh /
-     ps1 / js / event / devapp 的 6 份缓存写逻辑（1.1 矩阵）随 P1/P3 删除；
-     upgrade 补齐 mono 缓存刷新（修 `paths.go:292-296` 只刷 multi 的不对称）。
-     mono 下线版本再评估整体废弃缓存。
-
-### 2.5 市场 skill 边界正式化
-
-把"`dingtalk-*` 前缀与 `dws-shared` / `dws` 名称为 DWS 内置保留"从隐性约定
-变成共享常量 + 测试：
-
- `skillhome.ReservedSkillPrefixes = ["dingtalk-"]`、
-  `skillhome.ReservedSkillNames = ["dws-shared", "dws"]`，替换
-  `skill_setup.go:199/205` 与 `paths.go:332-334` 两份私有拷贝；所有互斥/
-  stale 清理只认这两个常量。
- 测试一：扫描 `skills/multi/` 实际目录名，断言全部命中保留规则（防新增
-  产品 skill 破坏前缀约定）。
- 测试二：市场 skill 安装路径（`skill_command.go:488-498` 的目标解析）若
-  产物名命中保留名则拒绝安装——把"市场无同名前缀"从祈祷变成门禁。
-
-## 3. 迁移步骤（渐进、不破坏存量用户）
-
-| 阶段 | 内容 | 风险 |
-|---|---|---|
-| **P0 清单 + policy** | 建 `internal/skillhome`：AgentHomes（**补 opencode**，修 1.2-1）、布局/互斥/保留名常量；setup/upgrade 切到共享包；脚本清单包标记块；新增 `check-agent-homes-sync.sh` 进 `make policy`；删 install.sh 死代码（1.2-2） | 低。纯收敛不改行为；policy 脚本初期可能误报 → 先 warn-only 跑一个版本再转 hard-fail |
-| **P1 setup 单引擎 + 脚本降级** | setup/upgrade 安装实现合一（互斥清理、stale 语义统一为"引擎内一种"，修 1.2-3）；install.sh / install.ps1 / install.js / install-skills.sh 改为 bootstrap + 冻结 fallback；install-event/devapp 暂不动 | 中。bootstrap 首版要在三语言 CI 矩阵上验证"引擎失败→fallback"链路；fallback 标记冻结，防止继续演进出第 9 份拷贝 |
-| **P2 state 权威 + 缓存收敛** | state.json 读写进 skillhome，setup/upgrade 记账；缓存加版本戳、写方收敛到 Go、upgrade 补刷 mono 缓存；install-event/devapp 语义收敛（mono+multi 双装改为引擎语义）后同样 bootstrap 化 | 中。存量机器无 state → 磁盘反推逻辑必须有故障注入测试；event/devapp 版本错配场景需保留 zip 直装逃生门 |
-| **P3 删除遗留拷贝** | 删除：脚本侧 6 份缓存写、install.sh/ps1/js 内被引擎接管的安装函数（保留冻结 fallback 一档）、Go 侧 `mutualExclusionVictims` 等被 skillhome 吸收的私有拷贝；观察一个版本后评估 fallback 是否可再降档 | 低-中。每删一处先确认 policy 与测试不再引用；fallback 删除是独立决策，不与本阶段捆绑 |
-
-与《迁移计划》的先后关系：本文 P0 即计划 P0c，应**最先做**；P1 与计划
-P0a/P0b 并行不冲突（引擎合一会让 P0a 的 mode-aware upgrade 少写一份代码）；
-P2 吸收计划 P0b；P3 在计划 P2 切默认之后执行。
-
-## 4. 明确不做
-
- **不重写成 symlink canonical**（canonical 一份 + 各 agent home 软链）：
-  已随生态分发通道一并评估否决，决策与实测原因见
-  《迁移计划》[§7.3](skill-multi-migration-plan.md)（无版本固定、依赖
-  Node/GitHub 可达、mono 会被一起发现、悟空 bundled/离线预装覆盖不了；
-  其中悟空分发线已于 2026-08-05 下线，该条约束随之消失）。
-  本地 symlink 模式（`setup --link`）同样维持计划 §8.2"可选/后置"定位，
-  不在本方案内。
- **不动 `dws-skills.zip` 产物布局**：zip 根 mono 副本 + `mono/` + `multi/`
-  双树保持不变（`scripts/release/post-goreleaser.sh:220-248`）。
- **不动市场 skill**：`dws skill install` 的安装语义、目标解析
-  （`skill_command.go:488-498`）不变；2.5 只新增保留名拒绝门禁与测试，
-  不改变既有安装行为。
- **不删 fallback**：脚本侧内嵌拷贝逻辑降级为冻结档保留，不追求"脚本零
-  拷贝"的纯净化。
-
-## 5. 收益/成本与任务映射
-
-### 5.1 收益/成本表
-
-| 项 | 现状 | 目标 | 量化收益 |
-|---|---|---|---|
-| agent home 清单 | 15 份拷贝，已漂移（opencode） | 1 份 Go 源 + policy 比对 | 新增 agent 从"改 11 处"变"改 1 处" |
-| 安装/互斥/stale 逻辑 | ≥8 份，4 种语言，语义已分叉 | 1 个 Go 引擎 | 语义变更从 4 语言 × 7 面（本次实证 13 文件 +759 行）变 1 包 |
-| mode 解析 | 4 份，行为细节不一 | 脚本只做透传，引擎一处实现 | 非法值/默认值行为天然一致 |
-| 缓存写 | 6 份，无版本戳，只写不读 | Go 唯一写方 + 版本戳 | 缓存漂移可观测、可判废 |
-| 安装状态 | 无 | state.json 权威 | upgrade 粘性、回滚、drift 检测的前提 |
-| 市场边界 | 隐性约定，2 份前缀拷贝 | 共享常量 + 2 个测试 | 互斥清理误伤市场 skill 的风险归零 |
-| 成本 | — | P0≈1.5d，P1≈3-4d，P2≈2d，P3≈1d | 合计 7.5-8.5 人日；其中 P0/P2 与《迁移计划》P0b/P0c 重叠，净新增约 4-5 人日 |
-
-### 5.2 与《迁移计划》§8 任务 ID 映射
-
-| 本文阶段 | 对应计划任务 | 关系 |
-|---|---|---|
-| P0 清单 + policy | P0c-1（清单下沉共享包+补 opencode）、P0c-2（check-agent-homes-sync.sh） | 原样采纳，包名定为 `internal/skillhome` |
-| P1 引擎合一 | P0a-1（UpgradeSkillLocations mode-aware）的前置简化 | 引擎合一后 P0a-1 不复写互斥/清理逻辑 |
-| P1 脚本 bootstrap | **取代** P2-2（sh/ps1 multi 真装，1.5d）、P2-3（install.js/install-skills.sh mode 支持，1d） | 脚本不再写安装逻辑，两个任务缩水为"透传 mode + 调引擎 + 冻结 fallback"，合计从 2.5d 降至 ~1d |
-| P2 state 权威 | P0b-1（state.json schema+setup 写入）、P0b-2（磁盘反推） | 原样采纳；缓存版本戳与写方收敛为本文新增 |
-| P2 event/devapp 收敛 | 新增（修 1.2-4 语义矛盾） | 依赖 P1 引擎稳定 |
-| P3 删除遗留 | 新增 | 在计划 P2 切默认之后执行 |
-| （不在本文） | P0-3 备份式安装、P0-4 请求头、P1-1 `dws skill mode`、P1-2/P2-1 默认翻转、P2-4 文案 | 仍按计划执行，与本文正交；备份式安装落地时直接写进引擎，天然覆盖所有面 |
@@ -1,231 +0,0 @@
-# Skill 能力缺口分析与补全计划
-
-> 基于 `feat/skill-mode-migration` 工作区代码级盘点（2026-08-05）。本工作区已把
-> install / upgrade / setup 的默认形态翻转为 multi；本文回答两个问题：
-> **管理能力上还缺什么**（Part 1）、**multi 内容是否覆盖 mono**（Part 2），
-> 并给出优先级排序的实施清单（Part 3）。
-> 任务 ID 复用 [skill-multi-migration-plan.md](skill-multi-migration-plan.md) §8.2
-> （P0a/P0b/P0c/P0-3/P0-4/P1-1/P1-2/P2-x/W4）；新增项以 `P1-3`、`C1–C7` 编号。
-> 机制背景见 [skill-distribution-mechanism.md](skill-distribution-mechanism.md)。
-
-## 0. 结论速览
-
- **管理面**：默认翻转已落地（setup/四个安装脚本/upgrade 识 multi），但生命周期
-  能力仍缺 6/8 项：无状态（state.json）、无备份回滚、无卸载、无故障恢复、
-  upgrade 不按已装清单做增量、市场/内置边界靠前缀约定。当前形态 = "能装上
-  multi，但装成什么样、出了事怎么退，全靠运气"。
- **内容面**：multi 对产品参考的覆盖**总体是 mono 超集**（chat/event/dev/sheet
-  均更详细），但 mono 有 **4 块全局能力在 multi 完全缺失**（recovery 闭环、
-  确认门禁协议、Schema 渐进查询教学、LICENSE/NOTICE），另有 1 个脚本
-  （`report_inbox_today.py`）未迁移；multi 自身还有 6 项内部不一致（死链、
-  orphan 脚本、漏改 EXPERIMENTAL 文案等）。
- **工作量**：管理面 P0–P2 剩余 ≈ 10.5–13 人日；内容补全 C1–C7 ≈ 3.5–4 人日；
-  合计 ≈ 14–17 人日，与 plan §8.1 的 13–17 人日口径一致（内容项是新增量）。
-
---
-
-## Part 1 — 管理能力缺口（skill 生命周期管理）
-
-### 1.1 现状盘点（本工作区实际状态）
-
-**`dws skill setup`（内置 skill 安装）**
-
-| 能力 | 状态 | 锚点 |
-|---|---|---|
-| mode 选择 | mono / multi；无 `--mode` 时 TTY 交互（multi 为默认项）、非 TTY 默认 multi | `internal/app/skill_setup.go:343-376` |
-| 按产品挑选 | `-s/--skill`、`-x/--exclude` 互斥；`dws-shared` 强制包含；未点名的已有 `dingtalk-*` 保留（additive） | `skill_setup.go:101-102`、`254-318`、`209-226`、`640-670` |
-| `--dry-run` | 有，预览 mode/来源/目标/子 skill，不写文件（root 持久 flag 注入，`internal/app/flags.go:43`） | `skill_setup.go:156-167` |
-| 源 | `--source` / `DWS_SKILL_SOURCE` 显式覆盖（失败不回退）→ 默认 embed（与二进制同版） | `internal/app/skill_setup_embed.go:50-58`、`skills_embed.go:28` |
-| 目标 | 16 个 agent home，父目录门控；`--target all` 不含 opencode，但命名 target 含（不对称） | `skill_setup.go:20-37`、`509-522` vs `internal/app/skill_command.go:129` |
-| 互斥清理 | 装 mono 删 `dingtalk-*`+`dws-shared`，装 multi 删 `dws/`；**best-effort，失败仅 warning 继续装** | `skill_setup.go:570-608` |
-| 备份/记账 | **无**。`RemoveAll` 直删，不写任何状态 | `skill_setup.go:616`、`655` |
-
-**`dws skill search / get / install`（市场 skill）**
-
- 子命令全集仅 `get / install / search / setup`（+ 隐藏的 find/add 兼容提示），
-  **无 status / mode / remove / rollback**（`skill_command.go:202-209`）。
- `install` 解压到 agent skills **根目录**（非 `dws/` 子目录），无 mode 概念、
-  无记账、无覆盖保护（`skill_command.go:378-443`、`653-672`）。
-
-**`dws upgrade`（skill 刷新）**
-
- `LocateSkillsRoot` **优先 zip 内 `multi/`**（`internal/upgrade/paths.go`，
-  接线于 `internal/app/upgrade.go`）；`UpgradeSkillLocations` **包驱动**：有
-  multi 树则始终 multi 刷新，否则 legacy mono 回退（不做磁盘粘性）。
- multi 路径：平铺刷新 `<agent>/dingtalk-*`+`dws-shared`，清 `dws/` 残留与过期
-  `dingtalk-*`，best-effort 刷 `~/.dws/skills/multi`（sibling `mono/` 存在时
-  亦刷 mono 缓存）。
- **有 multi 包时存量 mono 一次性迁 multi**（2026-08-05 owner：升级不做粘性；
-  非运行时 mode-switch 产品，无确认/备份/state）。
- skill 安装发生在**二进制替换之后**，skill 失败时二进制已换 → 半升级态。
-  `dws upgrade --rollback` 只回滚二进制。
-
-**安装脚本（7 面）**
-
-| 面 | 本工作区状态 |
-|---|---|
-| `scripts/install.sh` | 已默认 multi 且**真装**（`install_multi_skills_to_homes`），mono 为 opt-in；`rm -rf` 直删无备份 |
-| `scripts/install.ps1` | 同上（Windows） |
-| `scripts/install-skills.sh` | 已加 `DWS_SKILL_MODE`（默认 multi）+ multi 安装 |
-| `build/npm/install.js` | 已加 `installMultiSkillsToHomes`（互斥清理 + 平铺） |
-| Homebrew formula | 不铺 agent home，caveats 引导 `dws skill setup` |
-| `scripts/install-event.sh` / `install-devapp.sh` | 单 skill 专项语义，缓存 `multi/dingtalk-event` / `multi/dingtalk-dev` 子集 |
-| 共同缺口 | **均不写 state.json、均无备份**；agent home 清单仍是 sh/ps1/js/Go 多份手写（`paths.go:23-52` 的 keep-in-sync 注释约定，无门禁） |
-
-### 1.2 缺口表（对照完整生命周期）
-
-严重度：🔴 高（可致数据丢失/双份派发/半装态）｜🟡 中（能力缺失但有绕行）｜🟢 低（体验项）
-
-| # | 生命周期项 | 现状 | 严重度 | 补全设计（复用 plan §8 ID） |
-|---|---|---|---|---|
-| 1 | **status 查询** | **无**。无 state.json、无 `dws skill mode/status`；判断 mode 只能人工看磁盘形态（`dws/` vs `dingtalk-*`） | 🔴 | **P0b-1** 新增 `~/.dws/skills/state.json`（schema_version/mode/cli_version/installed/agent_homes/previous），setup 与安装脚本写入；**P0b-2** 缺失/损坏时磁盘形态反推，双形态并存 → drift 报错；**P1-1** `dws skill mode`（status 子命令展示 mode/版本/已装列表/上次切换/备份） |
-| 2 | **切换** | 有（`setup --mode`），但**无状态、无记账**：切完不留 previous，互斥清理失败仅 warning 继续装（`skill_setup.go:600-608`）→ 可能 mono+multi 双份共存、Agent 双份派发 | 🔴 | **P0b** 记账（含 `previous`）；**P0-3** 把「清理失败继续装」改为「失败整体回滚」；**P1-1** `mode set <mode>` 复用 setup 安装实现 + 备份 + 记账，成功后提示重启 AI 工具 |
-| 3 | **回滚** | **无备份**。setup/upgrade/安装脚本全部 `RemoveAll`/`rm -rf` 直删（`skill_setup.go:616,655`、`paths.go:181,249,307`、`install.js` `fs.rmSync`）；`upgrade --rollback` 只回二进制 | 🔴 | **P0-3** 备份式安装：`RemoveAll` → `mv` 到 `~/.dws/skills/backup/<ts>-<mode>/`，保留最近 2 份，任一 home 失败自动恢复并非 0 退出；**P1-1** `mode rollback` 一条命令回到 `state.previous` |
-| 4 | **版本对齐** | embed 天然同版（setup 默认源，`skill_setup_embed.go:50-58`）✓；但 `~/.dws/skills` 缓存**只写不读**（仅 legacy 回退候选，`skill_setup.go:445-447`），upgrade 只刷 multi 缓存不刷 mono（`paths.go:290-296`），漂移不可见 | 🟡 | **P0b-1** state.json 记 `cli_version` 使漂移可见；**P0a-1** upgrade 按 mode 同步刷新对应缓存（或评估废弃缓存，plan §6 风险表末行） |
-| 5 | **卸载** | **无 remove**。skill 子命令仅 get/install/search/setup（`skill_command.go:202-209`）；用户只能手动删目录，且不知道该删哪些（16 个 home × N 个 skill） | 🟡 | **新增 P1-3** `dws skill remove`：按 state.json 的 `agent_homes`×`installed` 精确删除内置 skill（`dingtalk-*`+`dws-shared`+`dws/`），不动市场 skill；`--dry-run` 预览（依赖 P0b） |
-| 6 | **局部更新** | setup 侧 additive 语义完整（`-s/-x`，未点名保留）；但 **upgrade 增量语义未按 `state.installed`**：`UpgradeSkillLocations` 用 zip 内 bundle 全集刷新（`paths.go:135-137`、`339-358`），用户 `-x` 排除过的 skill 会被升级装回来 | 🔴 | plan §8.3-1 已定死语义（以 `state.installed` 为准增量刷新、未装不补装）；**P0b-1** 先落 `installed` 列表，**P0a-1** `UpgradeSkillLocations(dir, mode)` 按其过滤 |
-| 7 | **故障恢复** | **无**。setup 清理/拷贝失败仅 warning 或按 home 跳过，留半装态；upgrade 的 skill 失败发生在二进制替换之后（`upgrade.go:593-611`），半升级态无自动恢复、无修复命令 | 🔴 | **P0-3** 备份式安装 + 失败整体回滚（改变「warning 继续装」语义，需同步改 `skill_setup_full_coverage_test.go` 多处断言，plan §8.3-2）；**P0b-2** drift 检测给显式收敛指令；**P1-1** `mode rollback` |
-| 8 | **市场 vs 内置边界** | **隐性前缀约定**：互斥清理按 `dingtalk-`/`dws-shared` 名称扫描（`paths.go:332-334`、`skill_setup.go:581`），无 SKILL.md frontmatter 校验 —— 市场 skill 若同名前缀会被误删；反向地，市场 `install` 解压无保护，可覆盖内置 `dingtalk-*`（`skill_command.go:653-672`） | 🟡 | plan §6 风险行：清理前校验目录内 SKILL.md frontmatter 属 DWS 产品集并写成测试（落入 **P0-3** 的清理改造）；长期由 state.json 的 `installed` 清单取代前缀扫描（P0b 后续） |
-
-**附：本工作区已完成项**（不再列入缺口）：setup 默认 multi（P2-1）、
-install.sh/ps1 真装 multi（P2-2）、install-skills.sh/install.js mode 支持
-（P2-3）、README×2 与 dingtalk-skill 文案（P2-4 部分）、upgrade 识别并刷新
-multi 包（P0a 的布局识别一半）。**尚未做**：P0c 清单收敛、P0b state.json、
-P0-3 备份、P0a 的 mode-aware 一半、P0-4 请求头、P1-1 mode 命令、P1-2 beta 轨。
-
---
-
-## Part 2 — 内容能力对等（mono vs multi）
-
-### 2.1 树对比总览
-
-| 维度 | mono | multi |
-|---|---|---|
-| 规模 | 152 文件 / ≈2.5 MB | 244 文件 / ≈3.3 MB（19 个 `dingtalk-*` + `dws-shared`） |
-| 入口 | 单 `SKILL.md`（332 行，含全局路由/危险表/Schema 教学） | 每产品一个 `SKILL.md` + `dws-shared` 全局契约（90 行） |
-| 全局参考 | `references/` 10 项（intent-guide、global-reference、url-patterns、error-codes、capability-limits、channel-login、field-rules、recovery-guide、best_practices/、products/） | `dws-shared/references/` 9 项 + 各产品 skill 自带 |
-| 脚本 | `scripts/` 37 个 | 10 个 skill 带 `scripts/` 共 55 个 |
-| 最佳实践 | `best_practices/` 01–11 + lite + `_common` | 按产品分发（01→chat … 11→minutes），`_common` 与 lite 入 `dws-shared` |
-
-### 2.2 产品参考覆盖核对
-
-逐产品比对结论：**multi 覆盖 mono 全部产品参考，且多数为超集**。
-
-| mono 产品参考 | multi 对应物 | 结论 |
-|---|---|---|
-| aitable.md + aitable/（20）+ aitable-record-ops | `dingtalk-aitable`（同 20 子章节 + field-rules + 06-data-analytics） | ✅ 超集 |
-| chat.md + chat-emoji-list | `dingtalk-chat`（+ chat/ 5 个子章节） | ✅ 超集 |
-| calendar / contact / doc+doc/ / drive / mail / minutes / todo / wiki / aisearch / hrbrain / pat / markdown | 同名 `dingtalk-*` skill | ✅ 覆盖（doc 侧 mono 的 doc-file-ops/doc-list/doc-permission/doc-search 四篇已迁移为 `dingtalk-doc/references/doc.md:202-207` 的 drive/wiki 迁移表，属刻意重构） |
-| event.md（222 行） | `dingtalk-event`（event-im.md 399 行 + 完整订阅治理契约） | ✅ 超集 |
-| dev.md（212 行） | `dingtalk-dev`（12 篇 reference 共 584 行） | ✅ 超集（结构重组） |
-| oa / attendance / report / sheet / ding / devdoc / agoal | `dingtalk-misc` 对应 reference（sheet 多 3 篇：comment/formula/version） | ✅ 超集 |
-| simple.md（devdoc+oa 合集 + 意图判断 + 上下文传递表） | 已拆入 `dingtalk-misc/references/oa.md:309,425` 与 `devdoc.md:15,19` | ✅ 覆盖 |
-| 多组织/多账号（SKILL.md:64-70 一节） | `dingtalk-profile` 整个 skill | ✅ 超集 |
-| 意图决策树（SKILL.md:101-123） | `dws-shared/references/intent-guide.md`（536 行 ≥ mono 488 行）+ `routing.md` | ✅ 覆盖 |
-| best_practices 01–11 / lite / _common | 按产品分发 + `dws-shared/references/best_practices/_common/` | ✅ 覆盖 |
-| url-patterns / capability-limits / channel-login / error-codes | `dws-shared/references/` 同名 | ✅ 覆盖 |
-| field-rules.md（mono 全局位） | `dingtalk-aitable/references/field-rules.md` | ✅ 合理下沉（内容即 AI 表格字段规则） |
-
-### 2.3 不对等项清单（mono 有、multi 无）
-
-| # | 缺失项 | mono 锚点 | multi 现状证据 | 严重度 | 补齐方式（承载方） |
-|---|---|---|---|---|---|
-| M1 | **Recovery 闭环**（recovery-guide.md + global-reference §Recovery + 错误处理第 2 步） | `skills/mono/SKILL.md:296,311`、`references/recovery-guide.md`、`global-reference.md:62` | multi 全树 `RECOVERY_EVENT_ID` 零引用；`dws-shared/SKILL.md:83-89` 错误最短路径无 recovery；`dingtalk-dev/SKILL.md:124` 指向「root dws / dws-shared 的错误处理」→ **断链** | 🔴 | **dws-shared**：新增 `references/recovery-guide.md`（从 mono 移植）、SKILL.md 错误最短路径加 recovery 步、`global-reference.md` 补 Recovery 节（→ C1） |
-| M2 | **确认门禁协议 + 全局危险操作表** | `skills/mono/SKILL.md:137-187`（危险操作表 + 确认流程 + `confirmation_required` 识别与重试协议） | multi 仅 aitable/doc/misc 三个 SKILL.md 有产品级危险表；`confirmation_required` / 「确认门禁」全树零命中；`dws-shared/SKILL.md:37-38` 只有一行泛化规则 | 🔴 | **dws-shared**：全局确认门禁协议（识别 `confirmation_required`、原始命令追加 `--yes` 重试、`--dry-run` 预览、禁止管道喂答案）+ 危险操作索引指向各产品表（→ C2） |
-| M3 | **Schema 渐进查询教学** | `skills/mono/SKILL.md:198-292`（≈95 行：四层查询、`--compact`/`--all` 边界、字段速查、Schema/Help/业务数据边界表、漂移处理） | 各产品 SKILL.md 仅点状提及 leaf Schema（17 处）；`dws-shared/references/global-reference.md:79-89`「命令自省」只有 `--help` | 🟡 | **dws-shared**：新增 Schema 渐进查询章节（改写为 multi 语境，链入渐进加载表）（→ C3） |
-| M4 | **scripts/report_inbox_today.py** | `skills/mono/scripts/report_inbox_today.py` | multi 无（misc 仅有 `report_received_today.py`）；mono 文档也未引用它 | 🟢 | **dingtalk-misc**：先验证脚本仍可用 → 迁入 `scripts/` 并在 `report.md` 引用；不可用则连同 mono 侧一起删（→ C5） |
-| M5 | **LICENSE / NOTICE** | `skills/mono/LICENSE`、`NOTICE` | multi 20 个 skill 均无 | 🟡 | 每个 multi skill 根复制两份（或 release 打包期注入，`scripts/release/post-goreleaser.sh`）（→ C7） |
-| M6 | **aiapp 意图路由** | `skills/mono/SKILL.md:76,101`（产品表 + 决策树有 aiapp 行，但目标 `aiapp.md` 不存在 —— mono 自身死链） | multi 全树无 aiapp 路由/文档；仅 orphan 脚本 `dingtalk-misc/scripts/aiapp_create_and_poll.py` | 🟡 | 决策：**dws-shared/routing.md** + **dingtalk-misc** 产品索引补 aiapp 行并新建 reference，或明确下线该能力并清掉 mono 死链与 orphan 脚本（→ C5） |
-
-### 2.4 multi 自身不一致项（不阻塞对等结论，但阻塞「multi 可独当一面」）
-
-| # | 问题 | 证据 | 处理 |
-|---|---|---|---|
-| X1 | 16 个 orphan 脚本无任何文档引用（yida×13、finance×2、aiapp×1）；`dws-shared/references/routing.md` 把「宜搭」路由到 dingtalk-misc，但 misc 产品索引无 yida 行、无 yida reference | `skills/multi/dingtalk-misc/scripts/`；`dingtalk-misc/SKILL.md:20-32` | 补文档（misc 产品索引 + reference）或移出发布包（→ C5） |
-| X2 | 死链：`dingtalk-chat/SKILL.md:131` 引用 `scripts/extract_media_id.py`，文件不存在（mono 也无） | 同上 | 补脚本或删引用（→ C4） |
-| X3 | 死链：`dws-shared/references/routing.md:22` 指向 `dingtalk-misc/references/markdown.md`，实际在 `dingtalk-markdown` | 同上 | 改指 `../../dingtalk-markdown/SKILL.md`（→ C4） |
-| X4 | `dingtalk-event/SKILL.md` 缺 `dws-shared` PREREQUISITE 与 `metadata:` 块（其余 19 个 skill 均有） | `skills/multi/dingtalk-event/SKILL.md:1-4` | 补齐（→ C4） |
-| X5 | 4 个 skill 仍是 🧪 EXPERIMENTAL + 「生产优先 mono」文案，与本工作区已翻转的默认矛盾（dingtalk-skill 已改，这 4 个漏改） | `dingtalk-profile/SKILL.md:15`、`dingtalk-hrbrain:15`、`dingtalk-markdown:15`、`dingtalk-pat:15` | 统一下调文案（→ C4，即 plan P2-4 剩余量） |
-| X6 | `<!-- SAFETY_PREAMBLE_INJECT -->` 标记存在于 5 个 SKILL.md，但仓库内无注入器 | `dingtalk-{pat,hrbrain,markdown,skill,profile}/SKILL.md` | 明确注入方（仓外流程则写注释）或移除标记（→ C4） |
-
-### 2.5 测试与政策门覆盖
-
-| 门面 | 覆盖 | 缺口 |
-|---|---|---|
-| `test/skill_tests.md` 覆盖表（40-54 行） | 13 产品 ≈256 用例 | 12/13 行引用 **mono 路径**（`references/products/...`），仅 dev 指 multi；`workbench` 行指向不存在的 `workbench.md`（54 行，mono/multi 均无）；`devdoc` 行指 `simple.md`（47 行，multi 无此文件）；未覆盖 doc/drive/mail/minutes/oa/sheet/wiki/aisearch/hrbrain/markdown/pat/profile/skill。multi 默认后需按 multi 路径重写并补产品（→ C6） |
-| `scripts/policy/check-skill-context-budget.sh` | 锁 `dingtalk-chat/SKILL.md` ≤14000B + shortcut 区块不膨胀 + mono SKILL.md 不含「充分阅读产品参考文件」回归（9-38 行）；`gen_skill_shortcut_sections.py --check` 同时写 mono 与 multi 的 shortcut 区块 | mono 下线判据（plan §5-4）要求先替代对 `skills/mono/SKILL.md` 的依赖 |
-| `make skill-command-integrity`（`check-skill-commands.sh` → `test/skill_static/skill_static_test.go:119-133`） | 静态校验 mono+multi 两树文档中的命令与 flag | 覆盖 OK；X2/X3 类 reference 死链不在其校验面 |
-
-**Part 2 结论**：multi **没有**覆盖 mono 的全部能力 —— 产品参考层面是超集，
-但全局能力缺 M1–M3 三块（recovery / 确认门禁 / Schema 教学），加 M4–M6 三个
-小项；另有 X1–X6 六项 multi 内部不一致。补齐全部落在 `dws-shared`（M1/M2/M3）、
-`dingtalk-misc`（M4/M6/X1）、各产品 skill（X2/X4/X5）与打包流程（M5）。
-
---
-
-## Part 3 — 优先级排序实施清单
-
-> 估时以 1 名熟悉本仓库的工程师计（人日），口径同 plan §8。
-> 「状态」列：✅ 本工作区已完成｜🚧 部分完成｜⬜ 未做。
-
-### 3.1 P0 — 先堵会丢数据/双份派发的洞（≈6.5–8.5d）
-
-| ID | 状态 | 任务 | 文件级改动点 | 估时 | 依赖 |
-|---|---|---|---|---|---|
-| P0c-1 | ⬜ | agent home 清单下沉共享包，补 opencode 不对称 | 新 `internal/skillhome`（或 `internal/upgrade` 导出）：合并 `paths.go:35-52` `knownSkillDirs` + `skill_setup.go:20-37` `skillSetupAgentHomes` + `skill_command.go:119-141` `agentSkillPaths`；setup/upgrade/测试共用 | 0.5d | — |
-| P0c-2 | ⬜ | 清单同步门禁 | 新 `scripts/policy/check-agent-homes-sync.sh`（进 `make policy`）；install.sh:ps1:install.js:install-skills.sh 清单改可解析块 | 1d | P0c-1 |
-| P0b-1 | ❌ CANCELLED | state.json schema + 写入 | 2026-08-05：无运行时切换，不写 state.json | — | — |
-| P0b-2 | ❌ CANCELLED | 磁盘形态反推作 state 兜底 | 无 sticky / 无 state；upgrade 按包刷 multi | — | — |
-| P0-3 | ❌ CANCELLED | 备份式安装 + 失败整体回滚 | 2026-08-05：随切换产品线取消 | — | — |
-| P0a-1 | ✅ | upgrade 包驱动 multi | `UpgradeSkillLocations`：有 multi→始终 multi（含 mono 盘一次性迁移）；legacy 无 multi→mono；不读 state | — | — |
-| P0a-2 | ✅ | force-multi 集成 / E2E | `paths_multi_test.go`（`MonoDiskMigratesToMulti` 等）+ `upgrade_skill_multi_e2e_test.go` | — | — |
-| P0-4 | ❌ CANCELLED | `x-dws-skill-mode` 请求头 | owner 决策移除 | — | — |
-
-### 3.2 P1 — 生命周期命令 + 内容补全（≈7–7.5d，两条线可并行）
-
-**管理命令线**
-
-| ID | 状态 | 任务 | 文件级改动点 | 估时 | 依赖 |
-|---|---|---|---|---|---|
-| P1-1 | ❌ CANCELLED | `dws skill mode` status/set/rollback/--dry-run | 2026-08-05：无运行时模式切换产品 | — | — |
-| P1-3 | ⬜ | `dws skill remove`（新增，缺口 #5） | 按磁盘/约定前缀精确删内置 skill（**不**依赖已取消的 state.json）；`--dry-run` 预览 | 1d | — |
-
-**内容补全线**（对应 Part 2 编号）
-
-| ID | 任务 | 文件级改动点 | 估时 | 依赖 |
-|---|---|---|---|---|
-| C1 | recovery 闭环进 multi（M1） | 新 `skills/multi/dws-shared/references/recovery-guide.md`；`dws-shared/SKILL.md:83-89` 错误最短路径加 recovery 步；`dws-shared/references/global-reference.md` 补 Recovery 节；`dingtalk-dev/SKILL.md:124` 断链改指 | 0.5d | — |
-| C2 | 确认门禁协议 + 危险操作索引（M2） | `dws-shared/SKILL.md` 增「确认门禁」节（移植 mono `SKILL.md:170-187` 协议）+ 危险操作索引表指向各产品 SKILL.md | 0.5d | — |
-| C3 | Schema 渐进查询教学（M3） | `dws-shared/SKILL.md` 渐进加载表加一行 + 新 `references/schema-usage.md`（改写 mono `SKILL.md:198-292`） | 0.5d | — |
-| C4 | multi 一致性修复（X2/X3/X4/X5/X6 = plan P2-4 剩余量） | `dws-shared/references/routing.md:22` 改指 dingtalk-markdown；`dingtalk-chat/SKILL.md:131` 死链处置；`dingtalk-event/SKILL.md` 补 PREREQUISITE+metadata；`dingtalk-{profile,hrbrain,markdown,pat}/SKILL.md:15` EXPERIMENTAL 文案下调；SAFETY_PREAMBLE_INJECT 标记处置 | 0.5d | — |
-| C5 | orphan 脚本与缺失产品处置（X1/M4/M6） | `dingtalk-misc/SKILL.md:20-32` 产品索引补 yida/finance/aiapp 行 + 新 reference（或把 16 个脚本移出发布包）；`report_inbox_today.py` 验证后迁入或删除；aiapp 路由决策落 `dws-shared/references/routing.md` | 0.5–1d | — |
-| C6 | skill_tests.md multi 化 + 扩产品 | `test/skill_tests.md:40-54` 覆盖表改 multi 路径；修 workbench（54 行）/devdoc（47 行）死链；补 doc/drive/mail/minutes/oa/sheet/wiki 用例 | 1d | C1–C5（路径稳定后） |
-| C7 | LICENSE/NOTICE 进 multi（M5） | 20 个 `skills/multi/*/LICENSE|NOTICE`（或 `scripts/release/post-goreleaser.sh` 打包期注入） | 0.25d | — |
-
-### 3.3 P2 / W4 — 收尾（≈1d）
-
-| ID | 状态 | 任务 | 估时 | 依赖 |
-|---|---|---|---|---|
-| P1-2 | ⬜ | beta 轨默认切 multi（版本门控；本工作区已全量切，可选择保留全量或回退为 beta 先行） | 0.5–1d | P1-1 |
-| P2-1 ~ P2-3 | ✅ | setup / install.sh / install.ps1 / install-skills.sh / install.js 默认 multi | — | 已完成 |
-| P2-4 | 🚧 | 文案翻转：README×2、dingtalk-skill 已改 ✅；4 个 EXPERIMENTAL 漏改 → 并入 C4 | — | — |
-| W4 | ⬜ | mono deprecation 警告（不删代码）+ mono 下线判据第 4 条（替代 `check-skill-context-budget.sh:10,34-38` 对 mono 的依赖） | 0.5d | P2、C6 |
-
-### 3.4 依赖图与排期建议
-
-```text
-P0c-1 ─► P0c-2
-P0b-1 ─► P0b-2 ─┐
-P0-3 ───────────┼─► P0a-1 ─► P0a-2 ─► P1-1 ─► P1-3
-P0b-1 ──────────┴─► P0-4
-C1…C5（互相独立，可与 P0 并行）─► C6
-```
-
- **W1**：P0c-1 → P0b-1/P0b-2 → P0-3 → P0a-1/P0a-2（管理面止血）；并行 C1–C4（内容高危项）。
- **W2**：P0-4、P1-1、P1-3；并行 C5、C7。
- **W3**：C6（内容面收口）+ P1-2 beta 决策；`make policy` 全绿。
- **W4**：mono deprecation + 观察（plan §4 原节奏不变）。
-
-合计：P0 ≈ 6.5–8.5d + P1 管理 ≈ 3–3.5d + 内容 C1–C7 ≈ 3.5–4d + P2/W4 ≈ 1d
-= **14–17 人日**。砍法同 plan §8.4：内容线可砍 C3/C7（教学与法律文件可后置），
-管理线不可砍 P0b/P0-3/P0a（缺了就是现在这副「能装不能管」的样子）。
@@ -0,0 +1,95 @@
+# DWS Skill 内容框架合同
+
+> 本分支权威合同：`skills/mono` / `skills/multi` 的**内容组织**与 zip 内容树形状。  
+> 不做安装/升级行为约定。质检见 [skill-mono-multi-qa.md](skill-mono-multi-qa.md)。  
+> 安装、升级与模式迁移见
+> [DWS 预制 Skill 安装、升级与模式迁移 RFC](rfc-skill-installation-and-upgrade.md)。
+
+## 1. 两棵内容树
+
+| 树 | 路径 | 角色 |
+|---|---|---|
+| **mono**（单 skill） | `skills/mono/` | 单一 `SKILL.md` 入口 + `references/products/*` 产品面 + 全局协议 |
+| **multi**（多 skill） | `skills/multi/` | 平铺 `dingtalk-*` 产品 skill + 必选 `dingtalk-shared` |
+
+Agent / 安装面选哪棵树由**行为分支**决定；本文件只规定树内合同。
+
+## 2. Multi 目录合同（如何新增一个产品 skill）
+
+新建 `skills/multi/<name>/` 时必须满足：
+
+1. **命名**
+   - 产品 skill：`dingtalk-<product>`（小写、连字符）
+   - 共享 skill：仅允许 `dingtalk-shared`
+2. **根文件**
+   - 必有 `SKILL.md`（YAML frontmatter + 正文）
+   - `references/` 推荐；无 reference 的 skill（如极简 profile）须在质检 omit 表登记
+   - `scripts/` 可选；脚本须被本 skill 树内某 `.md` 引用，或进入 orphan allowlist
+3. **Frontmatter 最小集**（产品 / shared）
+   - `name`：与目录名一致
+   - `description`：非空，含触发意图与边界
+   - `metadata.category`：`product` 或 `shared`（允许历史写法把 `cli_version` 放在 frontmatter 顶层）
+   - `metadata.requires.bins`：含 `dws`
+4. **契约块**
+   - 产品 skill 推荐内嵌 `<!-- DWS_RUNTIME_CONTRACT_START -->…END -->` **或** 明确 PREREQUISITE 指向 `dingtalk-shared`
+   - `dingtalk-shared` 承载跨产品路由与全局协议落点
+5. **与 mono 映射**
+   - 每个 mono `references/products/<stem>`（文件或目录）必须在
+     `skills/content-qa/mono-multi-coverage.yaml` 有 `coverage` 或 `omit_coverage` 行
+
+### 2.1 推荐骨架
+
+```text
+skills/multi/dingtalk-example/
+├── SKILL.md
+├── references/
+│   ├── example.md          # 主产品面
+│   └── …                   # 子章节 / 意图表
+└── scripts/                # 可选；须被 md 引用
+    └── example_helper.py
+```
+
+## 3. Mono 目录合同（质检对照基准）
+
+```text
+skills/mono/
+├── SKILL.md
+├── references/
+│   ├── products/           # 覆盖质检主源
+│   ├── error-codes.md      # 全局协议示例
+│   ├── error-codes.md
+│   └── …
+└── scripts/
+```
+
+- `references/products/` 下每个顶层 stem（`.md` 去后缀或子目录名）计入覆盖索引。
+- 同 stem 的 `.md` + 子目录视为同一产品面（如 `doc.md` + `doc/`）。
+
+## 4. 共享内容（`dingtalk-shared`）
+
+| 职责 | 落点 |
+|---|---|
+| 跨产品路由 / 工作流 | `references/routing.md`、`workflow-routing.md`、`intent-guide.md` |
+| 运行时最小契约长文 | `references/runtime-contract.md`（受 context-budget 约束） |
+| 全局协议（确认门禁 / Schema 教学等） | `references/`；见质检基线 |
+| 与 mono 全局文同名迁移 | `error-codes`、`url-patterns`、`capability-limits`、`channel-login`、`global-reference`、`recipes/`（`conventions.md`、`meta.md`、`lite-catalog.md`） |
+
+产品专属规则（如 AI 表格 `field-rules`）允许下沉到对应 `dingtalk-*`，须在覆盖表注明。
+
+## 5. Zip 内容布局合同（形状，非安装默认）
+
+发布物 `dws-skills.zip`（及 embed 同源）内容树形状：
+
+| Zip 路径 | 含义 |
+|---|---|
+| `<root>/` | mono 内容副本（兼容旧面） |
+| `<root>/mono/` | 与 `skills/mono/` 同构 |
+| `<root>/multi/` | 与 `skills/multi/` 同构 |
+
+质检可断言源树形状；**不**断言安装器默认解压哪棵。
+
+## 6. 变更流程
+
+1. 改 / 增内容 → 更新 `skills/content-qa/mono-multi-coverage.yaml`（coverage 或 omit）  
+2. 跑 `make skill-mono-multi-content`（该独立门禁不包含在默认 `make policy` 中）
+3. 失败则修内容或更新 reviewed omit（disposition + 原因），**禁止**用安装默认值绕过  
@@ -1,146 +0,0 @@
-# Skill 分发与消费机制调研（含 lark-cli 对标）
-
-> 配套方案：[skill-multi-migration-plan.md](skill-multi-migration-plan.md)
-> 调研日期：2026-08-04，基于 `feat/skill-mode-migration` 工作区代码级梳理 +
-> lark-cli（larksuite/cli@main）公开仓库调研。
-
-## 1. DWS 分发链路（源树 → 制品 → 渠道）
-
-```
-skills/mono/ ─┬─ go build ──► embed.FS（skills_embed.go:28）
-skills/multi/ ┘                    │  dws skill setup 默认源
-              │
-              └─ post-goreleaser.sh:220-248 ──► dws-skills.zip
-                 布局：zip 根 = mono 副本（向后兼容）+ mono/ + multi/
-                        │
-        ┌───────┬───────┼────────┬─────────┬──────────┐
-     GitHub   Gitee    OSS     npm      Homebrew   专项脚本
-     Release  (镜像)  (只发    tarball  (cellar    install-event/
-        │       │     不读)     │       不铺agent) install-devapp
-        │       │              │
-   install.sh/ps1          postinstall
-   install-skills.sh       install.js
-        │                      │
-        ├─► 各 agent home ◄────┤   永远 mono：<agent>/dws/
-        └─► ~/.dws/skills/{mono,multi} 缓存
-```
-
-渠道清单（7 个安装/分发面）：
-
-| 面 | skill 行为 | mode 概念 |
-|---|---|---|
-| `scripts/install.sh` | 装 mono 到 agent homes + 双缓存；选 multi **只打印提示、连缓存都跳过** | 有（半残） |
-| `scripts/install.ps1` | 同 install.sh（Windows） | 有（半残） |
-| `scripts/install-skills.sh` | 只装 mono，缓存 mono+multi | 无 |
-| `build/npm/install.js` | 永远 mono；缓存 mono+multi | 无 |
-| Homebrew formula | 整包 zip 进 cellar，不铺 agent、不写缓存，caveats 提示手动 setup | 无 |
-| `scripts/install-event.sh` | 装 mono + 缓存 `multi/dingtalk-event`；多 `.config/opencode/skills` | 无 |
-| `scripts/install-devapp.sh` | 缓存 `multi/dingtalk-dev` | 无 |
-
-产物事实：`dws-skills.zip` 已含 mono/multi 双树，**切 multi 不需要改 release 产物**。
-
-## 2. DWS 消费链路（源 → agent 目录）
-
-### 2.1 `dws skill setup`（`internal/app/skill_setup.go`）
-
- 源优先级：`--source` / `DWS_SKILL_SOURCE`（失败不回退）→ **embed 默认**
-  （`skill_setup_embed.go:50-58`）；legacy 候选（exe 旁/cwd/`~/.dws/skills`）
-  仅在绕过 wrapper 直连时才走。
- 目标：`skillSetupAgentHomes` 16 个 agent home，父目录门控（i=0 `.agents`
-  无条件，其余需 `~/.claude` 这类父目录存在）。`--target all` **不含
-  opencode**，但 `agentSkillPaths` 命名 target 含（不对称，测试只锁单向）。
- 布局：mono → `<agent-home>/dws/`；multi → `<agent-home>/` 平铺兄弟目录。
- 互斥清理：装 mono 删 `dingtalk-*`、装 multi 删 `dws/`；**best-effort，
-  失败仅 warning 继续装**（`cleanupMutualExclusion:611-620`），无备份。
- multi 过滤：`-s/--skill` 与 `-x/--exclude` 互斥；`dws-shared` 强制包含；
-  未点名的已有 `dingtalk-*` 保留（additive）。
-
-### 2.2 `dws upgrade`（`internal/upgrade/paths.go`）
-
- `LocateSkillMD` 命中 zip 根 mono → `UpgradeSkillLocations` 只写
-  `<agent>/dws/`：**不识 multi、不清理 `dingtalk-*`、不更新 `~/.dws/skills`**。
- 后果：multi 用户升级后 **mono + multi 共存**，Agent 双份派发。
- `--rollback` 只回滚二进制，不回滚 skill。
-
-### 2.3 市场 skill（`dws skill install`）
-
-解压到 agent skills **根目录**（非 `dws/` 子目录），无 mode 概念；互斥清理按
-`dingtalk-*` 前缀扫描，依赖"市场无同名前缀"隐性约定。
-
-### 2.4 状态与缓存
-
- **无任何已安装模式状态**：`~/.dws/` 有 auth/backups/cache，无 install
-  manifest；判断 mode 只能看磁盘形态。
- `~/.dws/skills` 缓存事实**只写不读**（默认 setup 走 embed），upgrade 不更新，
-  版本漂移不可见。
-
-## 3. lark-cli 分发机制（larksuite/cli@main）
-
-npm 包 `@larksuite/cli` 是薄壳（`files` 仅 install.js / install-wizard.js /
-run.js / checksums.txt）：
-
- **二进制**：postinstall 按平台从 GitHub Releases 下载，SHA256 校验
-  （checksums.txt 随 npm 包发）；镜像链 GitHub → 用户 registry 派生镜像 →
-  npmmirror 兜底；host allowlist + checksum 双保险；`run.js` 缺二进制自动补下，
-  Windows 有 `.old` 崩溃恢复。
- **Skills**：不进 npm 包、不进二进制。repo 根 `skills/` 即事实源，由生态
-  安装器 `npx skills add larksuite/cli -y -g`（vercel-labs/skills）安装；
-  wizard 首选 `https://open.feishu.cn` 直链，GitHub shorthand 兜底。
- **一键向导** `npx @larksuite/cli@latest install`：run.js 拦截 `install` →
-  install-wizard.js 串联 4 步（npm 全局装/升级 → skills → config init →
-  auth login），每步幂等（`skills ls -g` 检测 `lark-*` 已装则跳过）；
-  非 TTY 降级为"装完打印后续命令"。
-
-生态安装器 `skills` CLI 提供的能力：
-
- 76 个 agent 目录清单生态维护，自动探测；project/global 双 scope；
- **symlink canonical（推荐）或 copy**；symlink 下升级 = 更新一处；
- `list / find / update / remove` 全生命周期；**skill 升级由
-  `npx skills update` 承担，lark-cli 自己不写 skill 刷新逻辑**；
- 发现约定兼容 catalog 布局 `skills/<catalog>/<name>/SKILL.md`。
-
-## 4. 对标：DWS vs lark-cli
-
-| 维度 | lark-cli | DWS 现状 |
-|---|---|---|
-| 分发单元 | repo `skills/`（源码即事实源） | zip + embed + 5 处拷贝缓存 |
-| 安装器 | 1 个（生态工具） | 7 个自维护脚本，已漂移 |
-| 落盘 | symlink 单点更新 | 全量 copy，升级重写 16 home |
-| skill 升级 | `npx skills update`（生态承担） | `dws upgrade` 自写，不识 multi |
-| mono/multi | 不存在此问题（天生多 skill） | 互斥/切换/状态全自建 |
-| npm 包 | 薄壳运行时下载 | 全平台 archive + zip 全打进 tarball |
-| 大陆镜像 | registry 派生 + npmmirror | Gitee fallback + OSS 只发不读 |
-
-## 5. "分发外包给生态，自己只维护源码目录"是什么
-
-职责切分：skill 的分发/安装/升级/卸载交给生态标准化工具（`npx skills`，
-"agent skill 界的 npm"），DWS 只保证 repo 里 `skills/` 符合 agentskills.io
-规范。它消掉的正是 DWS 现在自维护的四块问题：
-
-1. **N 份事实源**（zip 三拷贝 + embed + 缓存 + 16 home）→ 只剩 repo 目录一份；
-2. **7 个自写安装器** → 零个，agent 清单别人维护（新 agent 自动支持）；
-3. **自建升级/切换/回滚**（UpgradeSkillLocations、互斥清理、方案中的
-   state.json/备份/`dws skill mode`）→ symlink 模式更新 canonical 一处；
-4. **lark-cli 因此根本没有 mono/multi 之争**，也没有 P0 要修的那些 bug。
-
-### 代价与前提（不是免费午餐）
-
-| 风险 | 说明 | 对策 |
-|---|---|---|
-| 版本错配 | DWS skill 从 Cobra 树生成，与 CLI 版本强耦合；embed 天然同版，生态安装从主分支拉可能错配。**已实测**：`skills add`（v1.5.21）无 tag/ref 版本固定参数，`@` 后接的是 skill 名而非 git ref；唯一的可复现机制是 project 级 `skills-lock.json`（experimental_install） | 短期：发布说明引导"skill 随 CLI 升级（`skills update`）"；中期：release 时推一个 `release/vX.Y.Z` 镜像分支或专用 skills 镜像仓供按版本安装；或接受错配（skill 内容为文档，错配成本=提到不存在的命令） |
-| 网络可达 | 依赖 npx + GitHub；大陆/内网现靠 Gitee fallback | `skills` CLI 支持任意 git URL，Gitee 镜像仓兜底，链路需验证 |
-| 离线/打包场景 | 悟空 bundled-skills、企业预装镜像生态通道覆盖不了（2026-08-05 注：悟空分发线已于当日下线，「悟空 bundled-skills」一项不再适用；企业预装镜像约束仍在） | 自维护打包保留一条 |
-| 生态工具策略漂移 | 清单/发现约定/默认值被动跟随 | 作为增量通道而非唯一通道，保留 embed 兜底 |
-
-## 6. 结论
-
- **生态分发通道已否决**（2026-08-04）：无版本固定（`skills add` 不支持
-  tag/ref，实测 v1.5.21）、依赖 Node + GitHub 可达、mono 会被一起发现、
-  悟空/离线场景覆盖不了。分发维持全自维护。（2026-08-05 注：悟空分发线
-  已于当日下线，「悟空场景覆盖不了」一条随之失效；其余否决理由与结论
-  不变。）
- 一个月内：按方案 P0–P2 修自维护通道并切 multi 默认（zip + embed +
-  安装脚本，产物布局不变）。
- 终态即"自维护通道修好之后"的形态，不再向 lark-cli 的生态外包形态收敛。
- 调研保留备查：`npx -y skills add . --list` 实测可发现全部 21 个 skill
-  （若未来生态工具补齐版本固定能力，可重新评估本决策）。
@@ -0,0 +1,74 @@
+# Mono↔Multi Skill 内容质检规格
+
+> 对照基准：`skills/mono`（单 skill）。被测主体：`skills/multi`。  
+> 机读合同：`skills/content-qa/mono-multi-coverage.yaml`。  
+> 执行：`make skill-mono-multi-content`（独立门禁；默认 `make policy` 按设计不包含该检查）。
+> 安装、升级与模式迁移见
+> [DWS 预制 Skill 安装、升级与模式迁移 RFC](rfc-skill-installation-and-upgrade.md)。
+
+## 1. 质检矩阵
+
+| ID | 类型 | 输入 | 通过准则 |
+|---|---|---|---|
+| **G1 形状** | 结构 | `skills/multi/*` | 仅 `dingtalk-*`（含必选 `dingtalk-shared`）；每目录有 `SKILL.md` |
+| **G2 结构** | 结构 | 各 `SKILL.md` frontmatter | `name`==目录名；非空 `description`；`category`∈{product,shared}；`requires.bins` 含 `dws` |
+| **G3 覆盖** | 覆盖 | mono `references/products/*` 顶层 stem | 每 stem ∈ `coverage` 或 `omit_coverage`；coverage 目标 skill/refs 存在 |
+| **G4 漂移** | 漂移 | scripts、成对文件、全局协议 | orphan 脚本 ∈ allowlist；paired 内容一致（允许合同声明的布局链接替换）；全局协议存在或 ∈ `omit_global` |
+| **G5 链接** | 可达性 | 合同覆盖的 paired Markdown | 内联相对链接目标文件或目录存在且不逃出仓库；外链、纯锚点和锚点内容不在检查范围 |
+
+已有门禁（继续复用，不替代本矩阵）：`check-skill-commands`、`check-skill-context-budget`、`check-multi-im-skill-chain`、`skill_docs_policy`、whiteboard 成对测试。
+
+## 2. 有意省略 / 延期登记格式
+
+YAML（见 coverage 文件）：
+
+```yaml
+omit_coverage:
+  - mono: simple
+    disposition: covered_by   # covered_by | defer | wontfix
+    via: dingtalk-misc        # optional
+    reason: "拆入 oa/devdoc…"
+
+omit_global:
+  - id: field-rules-global
+    mono_path: references/field-rules.md
+    expected_multi: dingtalk-aitable/references/field-rules.md
+    disposition: covered_by
+    reason: "AI 表格字段规则已下沉到 dingtalk-aitable；G3/coverage 不强制全局同名"
+
+orphan_scripts_allowlist:
+  - path: dingtalk-misc/scripts/report_received_today.py
+    disposition: defer
+    reason: "pending report.md reference"
+
+paired_files:
+  - mono: references/products/sheet.md
+    multi: dingtalk-misc/references/sheet.md
+    mode: link-normalized
+    link_substitutions:
+      - mono: "../url-patterns.md"
+        multi: "../../dingtalk-shared/references/url-patterns.md"
+      - mono: "../intent-guide.md"
+        multi: "sheet-intent-guide.md"
+```
+
+**处置原则**：质检失败 → 修**内容**或更新 reviewed omit；**不**改安装/升级默认。
+
+## 3. 缺口基线（相对 mono）
+
+| ID | 项 | disposition | 说明 |
+|---|---|---|---|
+| M1 | recovery-guide / RECOVERY_EVENT_ID 闭环 | **removed** | 已从 mono/multi skill 文档删除；不做移植 |
+| M2 | confirmation_required 全局协议 | **done** | `dingtalk-shared/references/confirmation.md` + SKILL 导航 |
+| M3 | Schema 渐进查询教学 | **done** | `dingtalk-shared/references/schema-usage.md` |
+| M4 | `report_inbox_today.py` | `defer` / orphan 侧 | 验证后迁 misc 或删 |
+| M5 | multi LICENSE/NOTICE | `defer` | 内容或打包注入 |
+| M6 | aiapp 路由 vs orphan 脚本 | **done（标明未产品化）** | mono 死链移除；`unsupported-scripts.md` |
+| X1 | yida/finance/aiapp orphan scripts | **done（登记）** | 由 unsupported-scripts 具名引用 |
+| X2 | chat 死链 `extract_media_id.py` | n/a | 现仅为反模式提及 |
+| X3 | routing → markdown 错路径 | **done** | 已指 `dingtalk-misc/references/markdown.md`；drive 尾链已修 |
+| X4 | event 缺 metadata | **done** | |
+| X5 | multi skill 横幅 /「优先 mono」文案 | **done** | 横幅已全部移除 |
+| X6 | SAFETY_PREAMBLE_INJECT 无注入器 | **done** | 标记已移除 |
+
+产品面覆盖：见 YAML `coverage`——mono products 均有 multi 承接（misc 聚合 attendance/oa/sheet/…）。
@@ -1,307 +0,0 @@
-# Skill 多 skill（multi）切换方案
-
-> 目标：一个月内把 DWS 的 Agent skill 默认安装形态从 mono（单 skill）切换为
-> multi（按产品拆分），两套并行期后下掉 mono。
-> 本文基于对分发/消费链路的代码级梳理（2026-08-04），所有锚点均可跳转验证。
-> 机制调研与 lark-cli 对标细节：[skill-distribution-mechanism.md](skill-distribution-mechanism.md)
-
-## 1. 目标与约束
-
-| 项 | 内容 |
-|---|---|
-| 终态 | 新装/升级默认铺 multi（`<agent-home>/dingtalk-*/`、`dws-shared/`）；mono 仅 opt-in；最终物理删除 mono |
-| 并行期 | 约一个月，mono 保留可切换、可回退 |
-| 硬约束 | DWS 无服务端灰度/远程配置；skill 是本地文件分发；安装面至少 7 个且已有漂移 |
-| 前置原则 | 先修已存在的 upgrade×multi 双份 bug，再谈切默认 |
-
-## 2. 现状关键事实（梳理结论）
-
-分发侧：
-
- `dws-skills.zip` 布局：zip 根 = mono 副本（向后兼容）+ `mono/` + `multi/`
-  （`scripts/release/post-goreleaser.sh:220-248`）。**产物无需改动**。
- 二进制 `//go:embed all:skills/mono all:skills/multi`（`skills_embed.go:28`），
-  `dws skill setup` 默认源就是 embed，**与分发渠道无关**。
- 7 个安装/分发面：install.sh、install.ps1、install-skills.sh、npm install.js、
-  Homebrew formula、install-event.sh、install-devapp.sh。
- OSS 只发不读；大陆链路靠 Gitee fallback。
-
-消费侧：
-
- `dws skill setup`：源优先级 `--source`/env → embed；目标 16 个 agent home
-  （父目录门控）；互斥清理 best-effort、失败仅 warning、无备份无回滚
-  （`internal/app/skill_setup.go`）。
- `dws upgrade`：`LocateSkillMD` 命中 zip 根 mono → `UpgradeSkillLocations`
-  只写 `<agent>/dws/`、不清理 `dingtalk-*`、不更新 `~/.dws/skills`
-  （`internal/upgrade/paths.go:119-172`）。**multi 用户升级后 mono+multi 共存**。
- `dws upgrade --rollback` 只回滚二进制，不回滚 skill。
- **无任何已安装模式状态**：`~/.dws/` 无 install manifest。
- `~/.dws/skills` 缓存事实上只写不读（默认 setup 走 embed），且 upgrade 不更新，
-  版本漂移不可见。
- `skill setup --target all` 不含 opencode，但 `agentSkillPaths` 含
-  （`skill_command.go:119-141` vs `skill_setup.go:20-37`）。
- 市场 skill（`dws skill install`）解压到 agent skills 根，与内置 skill 无 mode
-  概念；互斥清理按 `dingtalk-*` 前缀扫描，依赖"市场无同名前缀"这一隐性约定。
-
-## 3. 总体设计
-
-### 3.1 单一事实源收敛（P0c）
-
-agent home 清单目前在 5+ 处各写一份（注释约定 keep in sync，无门禁）。
-
- Go 侧：`skillSetupAgentHomes` 与 `knownSkillDirs` 合并为一个导出列表
-  （放在 `internal/upgrade` 或新 `internal/skillhome` 包），补 opencode，
-  setup/upgrade/测试共用。
- 脚本侧：install.sh / install.ps1 / install.js / install-skills.sh /
-  install-event.sh / install-devapp.sh 的清单由同一个 JSON 生成或政策脚本
-  比对（新增 `scripts/policy/check-agent-homes-sync.sh`，进 `make policy`）。
-
-### 3.2 安装状态文件（P0b）— ❌ CANCELLED（2026-08-05）
-
-原计划新增 `~/.dws/skills/state.json`（mode / cli_version / installed /
-previous / backup）。**已取消**：owner 决策不做运行时模式切换产品；upgrade
-有 multi 包时直接刷 multi，不再需要状态文件作为正确性或切换前提。
-
-### 3.3 备份式安装与真回滚（P0a）— ❌ CANCELLED（2026-08-05）
-
-原计划把 `RemoveAll` 改为 `mv` 到 `~/.dws/skills/backup/<ts>-<mode>/` 并
-支持失败整体回滚 / `dws skill mode rollback`。**已取消**：无运行时切换则
-不交付备份回滚产品面。安装/清理仍可为直接删除；mono retirement 版本若做
-一次性迁移，届时再单独评估临时备份，不预建 switch UX。
-
-### 3.4 upgrade 包驱动 multi 刷新（P0a，已落地）
-
-> 取代「按 state.json mode 刷新」与「磁盘粘性」：有 multi 包时一次性刷成 multi。
-
- `UpgradeSkillLocations(extractedDir)`：**不**推断磁盘布局。
-  - 包内有 multi 技能树 → 始终 multi 刷新（清 `dws/`、刷产品 skill、刷
-    `~/.dws/skills/multi`）；存量 mono 在日常 upgrade 上一并迁走。
-  - 无 multi 树的 legacy 包 → mono 刷新回退。
- 这是包驱动的一次性迁移，**不是**运行时 mode-switch 产品。
- 互斥清理在 multi 路径内执行，避免双布局长期共存。
-
-### 3.5 模式切换命令（P1）— ❌ CANCELLED（2026-08-05）
-
-原计划 `dws skill mode status|set|rollback|--dry-run`。**已取消**。
-
-用户若需改布局：重新走安装入口（`dws skill setup --mode <mono|multi> --yes`
-或安装脚本 `DWS_SKILL_MODE=`），不是 lifecycle 切换命令。
-
-### 3.6 默认切换（P2）
-
-改默认值="multi 默认、mono opt-in"，七个面一起改 + 门禁锁一致性：
-
-| 面 | 改动 |
-|---|---|
-| `dws skill setup` | 无 `--mode` 时默认 multi（交互选项顺序反转，mono 标 legacy） |
-| install.sh / install.ps1 | multi 分支**真正安装**（现在是打印提示跳过）；`DWS_SKILL_MODE=mono` opt-in；TTY 默认项改 multi |
-| install-skills.sh / npm install.js | 从零加 mode 支持（env `DWS_SKILL_MODE` / `--skill-mode`），默认 multi |
-| Homebrew formula | caveats 改提示 `dws skill setup`（默认即 multi），无需改资源 |
-| install-event.sh / install-devapp.sh | 维持单 skill 语义，但改走共享 install 函数 |
-
-文档同步：README/README_zh（`README_zh.md:81` "默认 1"）、
-`skills/multi/dingtalk-skill/SKILL.md` 的 🧪 EXPERIMENTAL 措辞下调、
-install.sh 内 multi 警告文案、AGENTS.md"生产优先 mono"表述。
-
-### 3.7 灰度与止血（无服务端能力下的替代）
-
- **L1 渠道灰度（先行）**：beta 轨（GitHub prerelease / npm `beta` dist-tag /
-  `dws upgrade --beta`）先切默认 multi，stable 保持 mono。零新增代码。
- **L2 确定性分桶（可选增强）**：随 release 发 `rollout.json`
-  （GitHub asset + OSS 同步），安装/升级时 `hash(machine-id) % 100 < pct` 决策
-  并粘入 `state.rollout`。规则：本地显式 env/flag 永远优先；拉取失败 fail-safe
-  mono（并行期）/ 保持现状（切默认后）；存量机器不被 rollout 改模式。
- **Kill switch**：`rollout.json` pct=0（若上 L2，已砍）+ beta 撤回 +
-  公告引导重装 `dws skill setup --mode mono --yes`（**无** `skill mode`
-  命令；备份回滚产品已随 D5 取消）。
- **可观测**：原 `x-dws-skill-mode` 请求头方案 **CANCELLED**（2026-08-05
-  owner）；灰度与下线判断改 issue 反馈 + 主动回访。
-
-### 3.8 安装与升级的语义矩阵（两条路径都要支持 multi）
-
-| 场景 | 时期 | 行为 |
-|---|---|---|
-| 新装 | 并行期（切默认前） | mono 默认，`DWS_SKILL_MODE=multi` / 交互可选 multi |
-| 新装 | 切默认后 | **multi 默认**，`DWS_SKILL_MODE=mono` / `--mode mono` / TTY opt-in（**仅安装时**） |
-| 升级（存量 mono） | 含 `multi/` 的包 | **一次性刷成 multi**（清 `dws/`）；无 switch 提示命令 |
-| 升级（存量 multi） | 含 `multi/` 的包 | multi 刷新，清过期 skill，刷 multi 缓存 |
-| 升级（无安装） | 含 `multi/` 的包 | 安装 multi（与安装默认一致） |
-| 升级（任意磁盘） | legacy 无 multi 树的包 | mono 刷新回退 |
-| `dws upgrade --rollback` | 任意 | 只回滚二进制（现状）；skill 无独立 rollback 命令 |
-
-原则：**升级不做磁盘粘性**；有 multi 包时一律刷 multi（含存量 mono 一次性
-迁移）。默认翻转与安装 opt-in 仍只影响新装/重装。无运行时 `dws skill mode`
-产品。
-
-## 4. 阶段与时间线（4 周）
-
-| 周 | 内容 | 出口标准 |
-|---|---|---|
-| W1 | P0a upgrade mode-aware + P0b state.json + P0c 清单收敛与门禁 + 备份式安装 | upgrade×multi 集成测试（装 multi → upgrade → 无 `dws/` 残留）；备份回滚测试（模拟中途失败）；`make policy` 含 homes 同步检查 |
-| W2 | P1 `dws skill mode`（status/set/rollback/dry-run）；beta 轨默认切 multi（L1）；`x-dws-skill-mode` 头 | 双向切换 + 中断恢复手工验收；beta 轨冒烟 |
-| W3 | P2 stable 默认切 multi、mono 降为 opt-in；文案翻转；（可选 L2 分桶 5%→20%） | 七面默认行为一致（政策脚本）；issue/请求头占比观察 |
-| W4 | （L2 则 50%→100%）；mono 打 deprecation 警告，**不删代码** | 连续 7 天无 multi 相关 P1 |
-
-## 5. mono 下线判据（不满足则不删）
-
-> 判据更新（2026-08-05）：原判据 3「悟空 bundled skill 分发线（dws_res →
-> bundled-skills，本仓库外）已切 multi」作废——悟空分发线已于当日决策
-> 下线（见 [skill-multi-roadmap.md](skill-multi-roadmap.md)
-> 「悟空线下线的影响（2026-08-05）」），无仓外 mono 依赖，mono 下线不再
-> 有仓外节奏闸门。判据重新编号如下（原 4/5 顺延为 3/4）。
-
-1. `x-dws-skill-mode=multi` 请求占比 ≥ 90%；
-2.  mono 主动 opt-in 率 ≤ 2%（install 脚本/命令埋点）；
-3. 连续两周无 multi P1；
-4. 已有等价政策门替代 `scripts/policy/check-skill-context-budget.sh` 对
-   `skills/mono/SKILL.md` 的依赖；`skills_embed.go` 去掉 `all:skills/mono`；
-   存量 `<agent>/dws` 目录有"遇到即迁移清理"逻辑。
-
-满足后单独一个版本窗口物理删除 `skills/mono/`，install/setup/upgrade 中 mono
-分支改为报错并指向 `dws skill mode set multi`。
-
-## 6. 风险与对策
-
-| 风险 | 对策 |
-|---|---|
-| 半装状态（清理成功、拷贝失败） | 3.3 备份式安装，失败整体回滚 |
-| multi 用户被 upgrade 塞回 mono | 3.4 mode-aware，P0a 先修 |
-| Agent 目录清单继续漂移 | 3.1 单一事实源 + policy 门禁 |
-| 互斥清理误伤市场 `dingtalk-*` skill | 清理前校验目录内 SKILL.md frontmatter 属 DWS 产品集，写成测试 |
-| 无灰度全切翻车 | beta 轨先行 + 备份回滚 + kill switch 公告命令 |
-| 用户不重启 AI 工具读到旧 skill | mode set / setup / install 输出统一提示重启 |
-| `~/.dws/skills` 缓存与 embed 版本漂移 | upgrade 时同步刷新对应 mode 缓存；长期可评估废弃缓存 |
-
-## 7. 分发机制对标 lark-cli（2026-08-04 调研）
-
-### 7.1 lark-cli 的实际分发结构
-
-npm 包 `@larksuite/cli` 是一个**薄壳**（package.json `files` 只有
-`install.js` / `install-wizard.js` / `run.js` / `checksums.txt`）：
-
- **二进制**：postinstall 时按平台从 GitHub Releases 下载，SHA256 校验
-  （checksums.txt 随 npm 包发布）；镜像链 = GitHub → 用户 registry 派生镜像 →
-  npmmirror 兜底，host allowlist + checksum 双保险。`run.js` 在二进制缺失时
-  自动补下载；Windows 有 `.old` 崩溃恢复。
- **Skills**：**不进 npm 包、不进二进制**。repo 根 `skills/` 目录即事实源，
-  由生态通用安装器 `npx skills add larksuite/cli -y -g`（vercel-labs/skills）
-  安装；wizard 首选 `https://open.feishu.cn` 直链、GitHub shorthand 兜底。
- **一键向导** `npx @larksuite/cli@latest install`：run.js 拦截 `install`
-  子命令 → install-wizard.js 串联 4 步（npm 全局装/升级 → skills 安装 →
-  config init → auth login），每步幂等可跳过（`skills ls -g` 检测 `lark-*`
-  已装则跳过）；非 TTY 自动降级为"装完打印后续命令"。
-
-生态安装器 `skills` CLI 的能力（lark-cli 免费获得的）：
-
- **76 个 agent 的目录清单由生态维护**，自动探测已装 agent；project/global
-  两种 scope；
- **symlink 到 canonical 副本（推荐）或 copy** —— symlink 模式下升级
-  = 更新 canonical 一份，所有 agent 即时生效；
- `list / find / update / remove` 全套生命周期命令，skill 升级由
-  `npx skills update` 承担，**lark-cli 自己不写 skill 刷新逻辑**；
- 发现约定兼容 catalog 布局 `skills/<catalog>/<name>/SKILL.md`。
-
-### 7.2 与 DWS 的关键差异
-
-| 维度 | lark-cli | DWS 现状 |
-|---|---|---|
-| 分发单元 | repo `skills/` 目录（源码即事实源） | zip + 二进制 embed + 5 处拷贝缓存 |
-| 安装器数量 | **1 个**（生态工具，76 agent 清单别人维护） | **7 个**自维护脚本，清单已漂移 |
-| 落盘方式 | symlink（canonical 单点更新） | 全量 copy，升级要重写 16 个 home |
-| skill 升级 | `npx skills update`（生态承担） | `dws upgrade` 自写，且不识 multi |
-| mono/multi 问题 | **不存在** —— 天生按目录多 skill | 互斥清理/模式切换/状态全是自建 |
-| npm 包体积 | 薄壳（运行时下载单平台二进制） | 全平台 archive + skills.zip 全打进 tarball |
-| 大陆镜像 | registry 派生 + npmmirror | Gitee fallback + OSS（只发不读） |
-
-### 7.3 结论：生态分发通道**已否决**（2026-08-04）
-
-~~借生态安装器做分发通道~~ 方向经评估后**放弃**。否决原因（均为实测）：
-
-1. **无版本固定**：`skills add`（v1.5.21）不支持 tag/ref 安装，而 DWS skill
-   从 Cobra 树生成、与 CLI 版本强耦合，错配不可接受；唯一可复现机制
-   `skills-lock.json` 仍是 experimental。
-2. **依赖 Node + GitHub 可达**：curl|sh / Homebrew / 大陆 Gitee fallback 的
-   用户环境大量无 Node，生态通道在这些场景是断的。
-3. **mono 会被一起发现**：发布即制造双份，须等 mono 下线才能发布，节奏不合。
-4. **悟空 bundled / 离线预装**生态通道永远覆盖不了。（2026-08-05 注：
-   悟空分发线已下线，此条约束随之消失；前 3 条否决理由仍成立。）
-
-**决策：分发维持全自维护通道**（zip + embed + 安装脚本），按 P0–P2 落地；
-不从 lark-cli 借鉴分发架构。可保留的借鉴点只剩两个本地语义，与生态无关：
-
-1. **wizard 式幂等安装**：`install.sh` 的 multi 分支从"打印提示跳过"改为
-   检测 state 可重入的真正安装（参考 install-wizard 的步骤化幂等）。
-2. **npm 下载校验**：install.js 如后续薄壳化，可参考其 host allowlist +
-   checksum 双保险与镜像链（GitHub → registry 派生 → npmmirror）设计。
-   与本次 multi 迁移解耦，不单列任务。
-
-## 8. 工作量评估与任务拆解
-
-> 以 1 名熟悉本仓库的工程师计（人日）。规模基线：涉及生产代码约 5.0k 行
-> （Go 3.0k + shell/ps1/js 2.0k），已有测试 1.2k 行可复用。
-
-### 8.1 总体判断
-
-**核心路径 13–17 人日，一人一个月可行但偏紧**。风险不在 Go 而在"shell /
-ps1 / js 三语言 × 七面同步"和对应的测试矩阵。可选项（rollout 分桶、npm
-薄壳化、symlink、生态通道发布）全部可砍可后置，砍后**最小可行集 8–10
-人日**（见 8.4）。
-
-### 8.2 任务拆解（带依赖）
-
-```
-P0c 清单收敛 ──┐
-P0b state.json ─┼─► P0a upgrade mode-aware ─► P1 skill mode ─► P2 切默认
-备份式安装 ────┘         │                        │
-                         └─► x-dws-skill-mode 头   └─► beta 轨先切（P2 预演）
-```
-
-| # | 任务 | 改动面 | 估时 | 依赖 |
-|---|---|---|---|---|
-| P0c-1 | agent home 清单下沉共享包（setup/upgrade 共用），补 opencode | `internal/upgrade/paths.go` 或新 `internal/skillhome`（~80 行新代码） | 0.5d | — |
-| P0c-2 | `scripts/policy/check-agent-homes-sync.sh`：比对 sh/ps1/js/sh 专项清单与 Go 清单，进 `make policy` | 新脚本 ~120 行 + 各脚本清单改成可解析块 | 1d | P0c-1 |
-| P0b-1 | state.json schema + setup 写入（含 installed 列表、agent_homes、previous） | `internal/app/skill_setup.go` +新文件 ~200 行 | 1d | — |
-| P0b-2 | 磁盘形态反推（dws/ → mono；dingtalk-* → multi；都有 → drift 报错）+ 单测 | ~120 行 | 0.5–1d | P0b-1 schema |
-| P0-3 | 备份式安装：RemoveAll→mv backup、失败自动回滚、保留最近 2 份、改 warning 语义 | setup install 两函数重写 ~150 行 + 测试 | 1.5–2d | — |
-| P0a-1 | `UpgradeSkillLocations(dir, mode)`：mono 现状+清残留；multi 从 zip `multi/` 平铺刷新+清 `dws/`+刷缓存 | `internal/upgrade/paths.go` ~150 行 | 1.5d | P0c-1、P0b |
-| P0a-2 | upgrade×multi 集成测试（装 multi→upgrade→无 dws/ 残留、dingtalk-* 已刷新） | 测试 ~200 行 | 1d | P0a-1 |
-| P0-4 | `x-dws-skill-mode` 请求头（仿 `x-dws-channel`） | `internal/auth/oauth_helpers.go` 附近 ~30 行 | 0.5d | P0b |
-| P1-1 | `dws skill mode` status/set/rollback/--dry-run | 新文件 ~350 行 + 测试 | 2–2.5d | P0b、P0-3 |
-| P1-2 | beta 轨默认切 multi（版本门控的默认值翻转，stable 不变） | setup/install.sh 默认值逻辑 ~40 行 | 0.5–1d | P1-1 |
-| P2-1 | setup 默认翻转 + 交互选项反转 + mono 标 legacy | ~30 行 + 测试更新 | 0.5d | P1 |
-| P2-2 | install.sh / install.ps1 的 multi 分支**真装**（幂等、读 state 跳过） | 两个脚本各 ~80 行 | 1.5d | P0b（脚本侧写 state） |
-| P2-3 | install.js / install-skills.sh 加 mode 支持（env + flag，默认 multi） | 各 ~60 行 | 1d | P2-2 同批 |
-| P2-4 | 文案翻转：README×2、SKILL.md EXPERIMENTAL 下调、install 脚本提示、AGENTS.md | 纯文档 | 0.5d | — |
-| W4 | mono deprecation 警告（不删代码）+ 观察 | ~20 行 | 0.5d | P2 |
-
-**小计：核心 13–17 人日**（P0 ≈ 6.5–8.5，P1 ≈ 3–3.5，P2 ≈ 3.5，W4 0.5）。
-
-可选/后置：rollout.json 分桶 2–3d；`setup --link` symlink 1–2d（需逐 agent
-验证）；npm 薄壳化 2–3d（独立立项）。
-
-### 8.3 风险最高的两处（先动）
-
-1. **P0a upgrade multi 刷新语义**：additive 安装 vs upgrade 全量刷新之间存在
-   一个真实设计题 —— 用户手动 `-x` 排除过的 skill，upgrade 要不要装回来？
-   答案：以 state.json 的 `installed` 为准做增量刷新，未装的不得补装
-   （否则违背 additive 语义）。这条必须在 P0a 开工前定死。
-2. **P0-3 备份回滚改语义**：现有测试断言"清理失败继续装"，改语义会动
-   `skill_setup_full_coverage_test.go` 多处；回滚恢复顺序（先恢复再报错）
-   要用故障注入测试覆盖。
-
-### 8.4 一个月做不完时的砍法
-
-按价值/成本比从后往前砍：rollout 分桶（L1 beta 轨已够）→ install.js /
-install-skills.sh mode 支持（npm 渠道用户量小，可先只改 sh/ps1）。砍后**最小可行集 8–10 人日**：
-
-> P0c-1 + P0b + P0-3 + P0a + P1-1 + P2-1 + P2-2（仅 sh/ps1）+ 文案
-
-即：upgrade 不再制造双份、有状态可回滚、setup 与主流安装脚本默认 multi。
-npm/install-skills.sh 维持 mono 显式行为并在输出中标注即将切换。
-
-## 9. 明确不做
-
- 不做服务端远程配置/灰度平台（用 beta 轨 + rollout.json 替代）。
- 不动 `dws-skills.zip` 产物布局（已含 mono/multi 双树）。
- 不动市场 skill（`dws skill install`）的安装语义。
- 并行期内不删 mono 代码与产物，只降级为 opt-in。
@@ -1,153 +0,0 @@
-# Skill multi 迁移：技术方案（as-implemented）与 Roadmap
-
-> 本文是 multi 迁移的当前事实源：第一部分记录**已落地实现**（代码级锚点，
-> 均可跳转验证），第二部分是带实时状态的 roadmap。
-> 原始方案 [skill-multi-migration-plan.md](skill-multi-migration-plan.md) 的
-> 若干"待做"描述已被后续决策取代（见决策记录 D1/D5）；分发机制调研结论见
-> [skill-distribution-mechanism.md](skill-distribution-mechanism.md)。
-> 代码快照：`feat/skill-mode-migration` @ `402429ac` + 工作区
-> upgrade-force-multi 调整（2026-08-05）。
-
-## 状态速览
-
-| 项 | 内容 |
-|---|---|
-| 当前阶段 | **阶段 1 / 1.5 ✅**（安装/升级默认 multi 已落地）；**阶段 2 运行时切换产品线 ❌ CANCELLED**（2026-08-05 owner 决策） |
-| 硬 deadline | **2026-08-30**：安装/升级默认 multi（✅）+ upgrade **有 multi 包时一次性刷成 multi**（不做磁盘粘性）+ mono 仅安装时 opt-in；mono **物理下线**仍可在独立 retirement 版本推进，但**不再**依赖 `dws skill mode` / 备份回滚产品 |
-| 已完成 | 五面默认 multi、`dws skill setup` 默认 multi、互斥清理对称、文案翻转；upgrade 有 `multi/` 时一律刷新 multi（存量 mono 一次性迁移） |
-| 下一步 | beta/L1 版本门控与观察（靠 issue/回访，**无** `x-dws-skill-mode` 埋点）；可选 agent-home 清单门禁；内容 C 线并行 |
-| 终态 | mono 仅安装时 opt-in；日常 upgrade（含 multi 的包）刷 multi；mono 物理删除仍可在 dedicated retirement 版本收尾（非用户切换命令） |
-
-### 简化设计（2026-08-05 起生效）
-
-1. **安装时一次决定**：默认 multi；`DWS_SKILL_MODE=mono` / `--mode mono` / 安装器 TTY 选 mono 为唯一 opt-in。
-2. **装完无运行时切换产品**：不做 `dws skill mode set/rollback`，不做备份式安装 / `state.json` 记账产品面。
-3. **升级不做粘性**：release zip 含 `multi/` 时 **一律** 刷新 multi（清 `dws/`、刷产品 skill + 缓存）；仅 legacy 无 multi 树的包回退 mono 路径。存量 mono 在日常 upgrade 上一次性迁到 multi。
-4. **mono 下线**：安装侧仍可 opt-in；upgrade 已承担「有 multi 包即迁走」；物理删 mono 树可另议 retirement。
-5. **可观测**：`x-dws-skill-mode` 请求头已按 owner 决策移除；灰度靠 issue + 回访。
-
---
-
-# 第一部分：技术方案（as-implemented）
-
-## 1. 升级：包驱动 multi 刷新（`dws upgrade`）
-
-核心语义：**升级不做磁盘粘性**；产物有 `multi/` 时一次性刷成 multi。
-无需 `state.json`，也无运行时切换命令。
-
- `LocateSkillsRoot` 优先返回 zip 内 `multi/`（`internal/upgrade/paths.go`）。
- `UpgradeSkillLocations`：包内有 multi 技能树 → **始终** `upgradeMultiSkillLocations`
-  （平铺 `dingtalk-*` + `dws-shared`，删 mono 残留 `dws/`，清过期 multi skill，
-  刷 `~/.dws/skills/multi`）；无 multi 树的 legacy 包才走 mono 刷新。
- 这是 upgrade 上的一次性迁移，不是 `dws skill mode` 产品。
-
-测试：`internal/upgrade/paths_multi_test.go`（含
-`TestUpgradeSkillLocationsMonoDiskMigratesToMulti`）+
-`internal/app/upgrade_skill_multi_e2e_test.go`。
-
-## 2. 安装默认 multi（四个脚本面）
-
-四个脚本安装面默认值全部为 multi，`DWS_SKILL_MODE=mono` 为统一 opt-in，
-互斥清理双向对称。详见阶段 1 落地说明（`scripts/install.sh` /
-`install.ps1` / `build/npm/install.js` / `scripts/install-skills.sh`）。
-
-## 3. `dws skill setup` 默认 multi
-
- 非交互未指定 `--mode` 时默认 multi。
- 交互选项 multi 在前（默认）、mono 标 legacy。
- 仍可用 `dws skill setup --mode mono --yes` **重装**到 mono（这是安装入口，
-  不是 lifecycle 切换产品；无备份/state/rollback 命令）。
-
-## 4. 文案翻转
-
- `README.md` / `README_zh.md`：multi 默认、mono legacy。
- `skills/multi/dingtalk-skill/SKILL.md`：去掉 EXPERIMENTAL。
-
-## 5. 已验证
-
- `go test ./internal/upgrade ./internal/app ./test/scripts`（阶段 1 基线）+
-  upgrade-force-multi 单测 / fake-HOME E2E。
- 脚本面契约测试暴露 `DWS_SKILL_MODE` 与 mono/multi 选项。
-
-## 6. 决策记录
-
- **D1 升级不依赖 state.json**（仍成立）。不读状态文件；布局由包内容驱动。
- **D2 无服务端灰度**。L1 beta 轨 + issue/回访；L2 `rollout.json` 已砍；
-  kill switch = beta 撤回 / 重装 `--mode mono`（无 `skill mode` 命令）。
- **D3 生态分发通道已否决**（`npx skills add` 等）。见
-  [skill-distribution-mechanism.md](skill-distribution-mechanism.md)。
- **D4 互斥前缀约定**。`dingtalk-*` / `dws-shared` 属 DWS 产品 skill。
- **D5 无运行时模式切换（2026-08-05）**。取消阶段 2 的备份式安装、
-  `state.json`、`dws skill mode`（status/set/rollback）、`x-dws-skill-mode`
-  请求头。Mode 只在安装时决定。
- **D6 升级不做粘性（2026-08-05）**。有 multi 包时 upgrade 一次性刷成 multi
-  （含存量 mono）；legacy 无 multi 树才回退 mono 路径。
-
---
-
-# 第二部分：Roadmap
-
-## ✅ 阶段 1 / 1.5（已完成，2026-08-05）
-
-安装/升级默认 multi、五面互斥清理、文案翻转、1.5 review 修复与实机 9/9。
-HEAD：`402429ac`。
-
-## ❌ 阶段 2（原切换/状态/备份产品线）— CANCELLED（2026-08-05）
-
-| 原任务 | 状态 | 说明 |
-|---|---|---|
-| 备份式安装（`~/.dws/skills/backup/...`） | ❌ CANCELLED | 不做运行时切换，无需备份回滚产品 |
-| `~/.dws/skills/state.json` | ❌ CANCELLED | 无切换产品；upgrade 按包刷 multi，不需要状态文件 |
-| `dws skill mode` status/set/rollback/--dry-run | ❌ CANCELLED | 无运行时切换 UX |
-| `x-dws-skill-mode` 请求头 | ❌ CANCELLED | owner 决策移除；观测改 issue/回访 |
-| agent home 清单门禁 | ⬜ 可选 | 与切换产品无关，仍可作工程质量项 |
-| `upgrade --dry-run` multi 文案 | ⬜ 可选 | 可随 force-multi 语义轻量对齐 |
-
-## 重构后的 8/30 目标
-
-**底线**：新装默认 multi；含 `multi/` 的 release 上 `dws upgrade` **一次性刷成
-multi**（含存量 mono）；仅需保持 mono 的用户用安装入口 opt-in 后勿升级到
-含 multi 的包，或 retirement 前用 setup 重装；mono 物理删除可另议。
-
-### 建议关键路径（简化）
-
-```text
-阶段1默认multi ✅ → upgrade force-multi ✅ → beta/L1（可选）→ 观察(issue/回访)
-→ stable 默认已是 multi → 可选 mono retirement（删 mono 树 / 安装入口报错）
-```
-
-### mono retirement（原阶段 4，重框）
-
- **不再**提供 `dws skill mode rollback` 作为用户出口。
- 日常 upgrade 已在有 multi 包时迁走存量 mono；retirement 版本可进一步移除
-  install/setup 的 mono 分支与 zip 内 mono 树。
- 下线判据改为：issue/回访无系统性 multi P1、mono opt-in 可接受、政策门
-  不再依赖 `skills/mono/SKILL.md`。（原请求头占比判据作废。）
-
-## 悟空线下线的影响（2026-08-05）
-
-悟空 bundled-skill 分发线已下线：mono 下线无仓外节奏闸门；`skills/multi/`
-为唯一 multi 事实源。设计资产留档
-[skill-wukong-comparison.md](skill-wukong-comparison.md)。
-
-## 明确不做
-
- ❌ 运行时模式切换（`dws skill mode set/rollback`）。
- ❌ `state.json` / 备份式安装产品面（随 D5 取消）。
- ❌ `x-dws-skill-mode` 请求头。
- ❌ 服务端远程配置 / L2 `rollout.json`。
- ❌ 运行时「模式切换」产品（含 sticky 伪切换）；upgrade 有 multi 时刷 multi
-  是包驱动的一次性迁移，不是 switch UX。
- 不动 `dws-skills.zip` 双树布局（仍可含 mono 副本供安装 opt-in）；不动市场
-  skill（`dws skill install`）。
-
---
-
-## 风险表
-
-| 风险 | 现状 | 缓解 |
-|---|---|---|
-| 半装无备份 | 安装/清理仍 `RemoveAll` | 接受为非切换产品下的已知限制；重装可收敛 |
-| mono/multi 漂移 | 无 state；upgrade 有 multi 即刷 multi | 升级后收敛为 multi；安装互斥清理 |
-| 用户想换模式 | 无 switch 命令 | 文档引导：`dws skill setup --mode <mono\|multi> --yes` 重装 |
-| 8/30 滑期 | 切换产品线已砍，关键路径缩短 | 聚焦 force-multi upgrade + 默认 multi 稳定 |
@@ -1,290 +0,0 @@
-# DWS Skill mono→multi 灰度能力设计（rollout capability）
-
-> 本文回答一个问题：**在没有服务端远程配置/灰度平台的前提下，mono→multi
-> 默认翻转如何灰度发布、如何观测、如何止血**。关联文档：
-> [skill-multi-roadmap.md](skill-multi-roadmap.md)（迁移事实源，本文展开其
-> 阶段 3「灰度切流」）、[skill-multi-migration-plan.md](skill-multi-migration-plan.md)
-> §3.7（灰度与止血的原始设计）。代码锚点快照：`feat/skill-mode-migration`
-> 工作区未 commit 变更（2026-08-05）。
-
---
-
-## TL;DR 推荐路线
-
-| 步 | 动作 | 层级 | 前置 |
-|---|---|---|---|
-| 1 | 把「五面默认 multi」拆成**版本门控默认**（beta→multi / stable 观察），同一份代码发 beta 轨先吃 | L1 | 无（本文 §2.1） |
-| 2 | ~~阶段 2 四件套（备份 / state.json / `dws skill mode` / 请求头）~~ | — | **❌ CANCELLED（2026-08-05）**：无运行时模式切换；upgrade 有 multi 时刷 multi（不做粘性） |
-| 3 | beta 轨观察：issue 流入 + 主动回访（**无**请求头占比） | 人工 | 步 1 |
-| 4 | stable：默认 multi 已在阶段 1 落地；L2 `rollout.json` 已砍 | — | — |
-| 5 | kill switch：beta 撤回（已有）/ 公告重装 `dws skill setup --mode mono --yes` | — | 无备份回滚产品 |
-
---
-
-## 1. 现状盘点：今天可用于灰度的全部旋钮
-
-### 1.1 旋钮总表
-
-| # | 旋钮 | 代码锚点 | 生效范围 | 盲区（谁够不着） |
-|---|---|---|---|---|
-| K1 | GitHub Release 双轨（stable / prerelease） | `internal/upgrade/github.go:100-106`（`ReleaseTrack`）；`internal/app/upgrade.go:870-875`（`upgradeTrack`）；release.yml 强制版本→轨映射（`release.yml:547` 含 `-beta.` → prerelease） | `dws upgrade --beta` / `--version vX.Y.Z-beta.N` 的二进制+skill 升级 | 不用 `dws upgrade` 的人；Gitee/OSS 镜像用户（见 K3/K4） |
-| K2 | npm dist-tag 双轨（`latest` / `beta`） | `release.yml:1761-1762`（prerelease→`beta` tag）；发布防倒退 `release.yml:1828-1835`；撤回脚本 `scripts/release/withdraw-release.sh:652-665`（dist-tag 回拨+deprecate） | `npm i dingtalk-workspace-cli@beta` 的新装/重装 | npm 默认安装（`@latest`）用户无感；npm 装完即走、不再 `npm i` 的存量 |
-| K3 | Gitee 镜像（代码 + release 资产） | main 代码镜像 `.github/workflows/mirror-to-gitee.yml:42-82`；release 资产镜像 `release.yml:1957-1965`（`sync-to-gitee.sh`）；安装脚本侧 `DWS_GITEE_REPO` 解析 `scripts/install.sh:18-19`、`scripts/install-skills.sh:21-24` | curl\|sh / install-skills.sh 的国内用户（显式 env 或 GitHub 不可达自动回退） | **`dws upgrade` 不到 Gitee**：upgrade client 只打 GitHub API（`internal/app/upgrade.go:48` → `internal/upgrade/github.go`，包内无任何 Gitee 引用）；Gitee release 是否 prerelease 由镜像脚本原样搬运，无独立轨控 |
-| K4 | OSS 镜像（ossutil 同步） | `scripts/release/sync-to-oss.sh:9-14`（`download/<version>/` + `latest.txt`/`beta.txt` 指针）；`release.yml:1897-1908` | 今天：**只写不读**——脚本注释明示「repository installers currently resolve GitHub/Gitee and do not consume these OSS pointers directly」（`sync-to-oss.sh:5-7`） | 所有人（指针无人消费）；但 `latest.txt`/`beta.txt` 是天然的**可变 channel 指针**（见 §2.2 设计复用） |
-| K5 | 安装脚本 env / flag | `DWS_SKILL_MODE`：`scripts/install.sh:17`（解析 `install.sh:220-256`）、`scripts/install.ps1:293-332`、`scripts/install-skills.sh:29-33`；npm `--skill-mode` / env `build/npm/install.js:197-214`；`DWS_VERSION` 指定 beta 版 `install.sh:38` | 新装时的逐台显式控制（CI、内推灰度名单） | 只对**执行安装那一刻**生效；装完无持久化（无 state.json），事后无法得知当初怎么装的；`DWS_VERSION=latest` 在 GitHub 侧只解析 stable（`/releases/latest` 永不指向 prerelease，`install.sh:190-198`），beta 必须显式给版本号 |
-| K6 | 版本门控默认值（构建期注入） | goreleaser ldflags 注入版本 `.goreleaser.yaml:22`（`internal/app.version=v{{.Version}}`）；`prerelease: auto` `.goreleaser.yaml:68` | 同一 commit，beta build 与 stable build 可表现不同默认值（§2.1 切法 B 的机制） | 脚本面拿不到 Go 变量，需各自从「解析出的版本号」重推导（§2.1）；homebrew formula 只搬 zip 根（mono 布局）到 pkgshare（`build/homebrew.rb.tmpl:26-31`），formula 本身无 mode 概念 |
-| K7 | 本地遥测（opt-in） | `internal/shortcut/usage/recorder.go:82-88`（`DWS_USAGE_TRACKING=1`，默认关）；只写本地 `~/.dws/usage.jsonl`（`recorder.go:91`） | 高频命令形状挖掘（shortcut P2） | **不上传任何服务端**：对灰度占比测量零贡献；默认关意味着即使上传也无统计意义 |
-| K8 | 请求头通道（已有上行链路） | MCP 请求统一注入点 `internal/app/runner.go:953-1008`（`resolveIdentityHeaders`，接线于 `runner.go:140/240/597`、`internal/app/recovery_command.go:271/337`）；登录权限检查 `internal/auth/oauth_helpers.go:1424-1428`；`x-dws-channel`（`DWS_CHANNEL`）先例 `runner.go:1006-1008`、`oauth_helpers.go:1425-1426` | 服务端（MCP 网关）已能按 header 聚合：`x-dws-agent-id`、`x-dingtalk-dws-agent-code`、`X-Cli-Version` 均在线 | 只有「已登录且发 MCP 请求」的用户可被观测；纯安装未使用、auth 失败前的用户在分子里缺席（占比偏高估，§2.3） |
-
-### 1.2 结构性盲区（任何旋钮都够不着）
-
-| 盲区 | 说明 | 出处 |
-|---|---|---|
-| ~~悟空 bundled skill 分发线~~（盲区已移除） | 悟空分发线（dws_res → Wukong.app bundled-skills）已于 2026-08-05 决策下线，本盲区随之移除；历史上该线在**本仓库外**、仅有 main CI 成功后的下游触发（`.github/workflows/notify-wukong.yml:13-38`），本仓库默认值翻转管不到它 | 原 roadmap 风险表「悟空线外挂」行（已解除）、阶段 4 原判据 3（已作废） |
-| homebrew 用户 | formula 只把 zip 根（mono 副本）stage 进 `pkgshare/skills/dws`（`build/homebrew.rb.tmpl:26-31`），caveats 指向 `dws skill setup`（`:33-38`）；multi 源不在包内，`setup --mode multi` 只能靠 `~/.dws/skills/multi` 缓存 | 同上，K6 |
-| 永不升级的存量 mono 用户 | 所有旋钮都作用于「新装/升级/重装」三个时点；不动作的用户一切照旧（这正是灰度的天然保护层） | — |
-| 安装后不再运行 `dws` 的用户 | K8 观测不到，占比分母缺失 | §2.3 |
-
---
-
-## 2. 方案设计（分层）
-
-### 2.1 L1 渠道灰度：beta 轨先吃 multi 默认
-
-目标：**同一代码、不同 release 轨不同默认值**，stable 用户在观察期内完全无感。
-
-#### 切法 A：纯流程（零新增代码，不推荐单独使用）
-
-当前工作区五面已无条件翻转 multi；直接发 beta 即完成「beta 先吃」。
-问题：main 上的默认值已是 multi，**下一个 stable 无处可躲**——stable 发布
-窗口一到就必须全切，观察期长短不由人；且中途想给 stable 出补丁版（hotfix）
-会被迫带上 multi 默认。仅适合「beta 观察期确定短、stable 窗口确定远」的情形。
-
-#### 切法 B：版本门控默认值（推荐）
-
-把五面的默认值从「无条件 multi」改为「beta 版本默认 multi，stable 版本默认
-mono」。判据统一用版本号是否含 `-beta.`（release.yml 已强制版本→轨唯一映射，
-`release.yml:547`、`.goreleaser.yaml:68`）：
-
-| 面 | 门控取值来源 | 落点 |
-|---|---|---|
-| `dws skill setup` / `dws upgrade`（Go） | ldflags 注入的 `internal/app.version`（`.goreleaser.yaml:22`），`strings.Contains(version, "-beta.")` | `internal/app/skill_setup.go:354-357`（非交互默认）与 `:364-368`（交互默认项排序） |
-| install.sh / install.ps1 / install-skills.sh | 脚本自己解析出的 `$VERSION`（`install.sh:177-198`；Gitee 侧 `install.sh:180-188`）——`case "$VERSION" in *-beta.*)` | `install.sh:220-256`、`install.ps1:293-332`、`install-skills.sh:29` |
-| npm install.js | 包内 `package.json` 的 `version`（staging 时由 `stage-npm-package.sh` 写入 release 版本） | `build/npm/install.js:197-214` |
-
-**关键发现——upgrade 路径（2026-08-05 更新）。** skill 升级语义是「跟着 zip
-产物布局走、不做磁盘粘性」：`LocateSkillsRoot` 恒优先 `multi/`，
-`UpgradeSkillLocations` 在包内有 multi 时**始终**刷 multi（含存量 mono
-一次性迁移，清 `dws/`）。若仍要做「stable 观察期不迁 mono」，门控必须落在
-**是否发布含 multi 的 zip / 是否走 upgrade skill 刷新**，而不是磁盘粘性分支
-（粘性方案已否决）。当前产品默认接受：含 multi 的 release 上 upgrade = 迁
-multi。
-
-切法 B 的安装默认门控（beta→multi / stable→mono）仍可独立存在；与 upgrade
-force-multi 正交——安装 opt-in mono 的用户一旦升级含 multi 的包会被迁走。
-
-#### 风险与回退
-
-| 风险 | 说明 | 回退 |
-|---|---|---|
-| 门控不可见 | 默认值随版本号变化，review/测试容易漏 | 每面补契约测试（beta→multi / stable→mono），`test/scripts` 已有同构先例（`test/scripts/install_script_test.go:529-576`） |
-| 升级迁走 mono | **接受为产品语义**：有 multi 包时 upgrade 一次性刷 multi；需 mono 则 `dws skill setup --mode mono --yes` 重装（装完后再 upgrade 仍可能被迁回） | S1 撤回含 multi 的坏包；S3 公告重装 |
-| beta 轨整体有毒 | 二进制或 skill 包级事故 | 现有撤回链：`scripts/release/withdraw-release.sh`（GitHub release 撤回 + npm deprecate + dist-tag 回滚，`withdraw-release.sh:652-665`）；stable 轨不受影响 |
-| 版本字符串被仿造 | 本地 `go build` 无版本注入时 `version=""`，门控落 stable 分支（保守方向，正确） | — |
-
-### 2.2 L2 确定性分桶：随 release 发 `rollout.json`
-
-L1 的粒度是「轨」：beta 全吃、stable 全不吃。stable 切流若要 5%→100% 的
-渐进，需要机器级分桶。无服务端，用**随版本分发的只读配置 + 客户端确定性
-哈希**替代。
-
-#### rollout.json schema 与发布链路
-
-作为 release 资产随每个版本发出（进 `dist/`）：
-
-```json
-{
-  "skill_mode": {
-    "pct": 20,
-    "salt": "skill-mode-2026h2",
-    "note": "mono->multi default rollout for stable track"
-  }
-}
-```
-
- `pct`：0–100，`bucket < pct` 的机器默认 multi。
- `salt`：分桶盐，换盐=重新洗牌（默认不换，保证跨版本粘性可比）。
- 发布链路改动：`scripts/release/post-goreleaser.sh` 生成进 `dist/`；
-  **资产命名空间是精确集合**（`scripts/release/verify-release-artifacts.sh:12-38`，
-  「public release assets must contain exactly the supported files」），必须把
-  `rollout.json` 加进 EXPECTED_ASSETS；stable 晋升门会比较 beta 资产集
-  （`release.yml:833-846`），所以引入该资产的那个 beta 起两轨必须同时带。
- checksums.txt 由 dist 自动生成，镜像脚本（Gitee `release.yml:1957-1965`、
-  OSS `sync-to-oss.sh:9-14`）整目录搬运，**rollout.json 自动随资产集流到
-  Gitee/OSS，无需额外接线**。
-
-#### machine-id 来源：读现成，不新建
-
-`internal/auth/identity.go` 已有稳定 per-install UUID v4 `machineId`
-（`identity.go:19-20`、结构体 `:70-76`），持久化在 `~/.dws/identity.json`
-（`identity.go:51` + `pkg/config/constants.go:177-188`），惰性生成
-（`EnsureExists` `:115-133`），v1 文件透明迁移（`:98-113`）。分桶直接复用：
-
-```
-bucket = int(sha256(machineId + "|" + salt)[:8], 16) % 100
-```
-
-边界情况：`identity.json` 首建于首次 MCP 请求链路（`runner.go:954`）。灰度
-决策发生在 setup/upgrade 时，可能早于任何 MCP 请求——此时按 `EnsureExists`
-同款语义**就地惰性创建**（best-effort 持久化，失败则用进程内随机值且当次
-不记账，下次重决）。不引入第二套 `~/.dws/install-id`，避免双事实源。
-
-脚本面（sh/ps1）做 sha256 分桶要读 JSON + 哈希，复杂且易错；npm install.js
-用 node crypto 是一行。**范围划定：L2 分桶只在 Go 面（`dws upgrade` /
-`dws skill setup` / 未来 `dws skill mode`）与 npm install.js 实现**；sh/ps1
-停在 L1 版本门控（curl 用户全是新装，渠道轨已够；百分比分桶的主战场是存量
-升级，而升级必过 Go 二进制）。
-
-#### state.json 的 rollout 字段
-
-依赖阶段 2 的 `~/.dws/skills/state.json`（P0b，未实现，roadmap 阶段 2）：
-
-```json
-{
-  "mode": "multi",
-  "rollout": {
-    "bucket": 37,
-    "pct": 20,
-    "salt": "skill-mode-2026h2",
-    "decision": "multi",
-    "decided_at": "2026-08-20T08:00:00Z",
-    "decided_by": "rollout.json@v1.4.2",
-    "explicit": false
-  }
-}
-```
-
-决策顺序（每台机器只决策一次，粘性）：
-
-1. 本地显式（`DWS_SKILL_MODE` / `--mode` / `--skill-mode` / `dws skill mode set`）
-   → 用之，`explicit=true`；
-2. `state.json` 已有 `mode` 或磁盘形态可反推（roadmap 阶段 2 既定兜底）
-   → 保持现状，不参与分桶；
-3. `state.rollout.decision` 已存在 → 复用（pct 后续变化不翻案）；
-4. 拉取 `rollout.json`（Go 面：upgrade 已下载本版资产，同 release 再取一个
-   小文件；npm：包内自带）→ 算 bucket 决策并记账；
-5. 任一步失败 → 当期默认（L1 版本门控结果）。
-
-#### 三条硬规则
-
-| 规则 | 内容 | 理由 |
-|---|---|---|
-| R1 本地显式优先 | env/flag/命令任何时候压过 rollout 决策；显式选择落 `explicit=true` 后 rollout 永不改它 | 灰度不能覆盖用户意志；也是 kill switch 的用户侧出口 |
-| R2 拉取失败 fail-safe | 拉不到/解析失败/字段越界 → 保持现状（存量）或当期默认（新装），**绝不因拉取失败翻模式** | 无服务端下网络面即故障面，故障必须倒向保守侧 |
-| R3 存量不被改模式 | 已有 mode（state 或磁盘可推）的机器不参与分桶；pct 只影响「未决策」机器 | pct 从 20 降到 0 不能把已进 multi 的 20% 弹回 mono（那需要 kill switch，不是分桶语义） |
-
-#### 与 Gitee / immutable release 的兼容
-
- **Gitee**：`dws upgrade` 不读 Gitee（§1.1-K3），rollout.json 经
-  reconcile/sync 脚本随资产集镜像到 Gitee release；Gitee 侧安装脚本如需消费，
-  走与 `dws-skills.zip` 相同的 Gitee API 资产枚举（`install.sh:180-188` 同
-  模式）。一期不消费、只保证镜像不缺失。
- **immutable release 约束**：官方仓已开启不可变 release（`release.yml:802`
-  「Immutable releases must be enabled before publishing」），**资产发布后不可
-  替换**——调 pct = 发一个新补丁版（beta 线 release.yml 支持连续 beta：
-  `release_bump` 在连续 beta 线时被忽略，`release.yml:25-27`）。这决定了
-  L2 的调参时延 = 一次发版；追求更快止血见 §2.4。
- （可选远期）OSS `latest.txt`/`beta.txt` 是现成的**可变**指针
-  （`sync-to-oss.sh:13-14`），若未来 installer 学会读 OSS，可把
-  `rollout-current.json` 放 OSS 变指针后面实现「不发版调 pct」；今天无消费
-  方，不建。
-
-### 2.3 L3 可观测性：`x-dws-skill-mode` 请求头 — ❌ CANCELLED
-
-**2026-08-05 owner 决策：不实现该请求头**（与运行时模式切换 / state.json
-一并取消）。原设计（注入 `resolveIdentityHeaders`、按 state/磁盘上报
-`mono|multi|unknown`）仅作历史记录，不进入排期。
-
-观察手段改为：**issue 反馈 + 主动回访**；不再有请求级 multi 占比判据。
-
-### 2.4 Kill switch：四层止血
-
-| 层 | 手段 | 时延 | 现状/依赖 |
-|---|---|---|---|
-| S1 beta 轨整体撤回 | `scripts/release/withdraw-release.sh`：GitHub release 撤回 + npm deprecate + `beta` dist-tag 回拨（`:652-665`） | 分钟级 | **今天可用** |
-| S2 rollout.json `pct=0` | 发补丁版把 pct 打 0（immutable release 不允许原地改资产，§2.2） | 一次发版（小时级） | 依赖 L2 落地 |
-| S3 公告命令 | 公告用户重装 `dws skill setup --mode mono --yes`（安装入口，非 switch 产品） | 用户触达时延 | **可用**（无 `dws skill mode`；阶段 2 切换命令已 CANCELLED） |
-| S4 备份回滚 | ~~`dws skill mode rollback` + 备份式安装~~ | — | **❌ CANCELLED（2026-08-05）** 与运行时切换一并取消 |
-
-二进制侧另有既有的 `dws upgrade --rollback`（`internal/upgrade/rollback.go`，
-备份在 `~/.dws/data/backups`、保留 5 份 `:16/:54-63`），但只回滚二进制不回滚
-skill 布局——skill 止血靠 S1 + S3（重装 mono），**无** S4。
-
-**结论（2026-08-05）：** kill switch = S1（beta 撤回）+ S3（公告重装
-`--mode mono`）。S4 已取消；日常 upgrade 有 multi 包时**一次性刷 multi**
-（不做粘性），故「装完 mono 再 upgrade」会迁走——止血靠撤回坏包或重装。
-
---
-
-## 3. 对标（简要）
-
-**npm dist-tag 双轨（lark-cli 类企业内部 CLI 的通行形态）。** 以 npm 仓
-registry 为唯一分发面时，灰度即 dist-tag：`latest` 稳态、`beta`/`next` 先行
-（`next@canary`、`typescript@beta` 同款模式），安装侧 `npm i pkg@beta` 或
-CI 指定 tag 即完成分群；撤回即 `npm dist-tag add pkg@<prev> latest` +
-`npm deprecate`。DWS 已完整具备此形态（§1.1-K2），lark-cli 等内部 CLI 在
-集团内网 registry 上亦按同一范式运作——差别只在内部 registry 可附带按
-员工/部门灰度的下发规则，那是「registry 有服务端」的红利，DWS 面向公网
-npm 没有这一层，故需 L2 补齐。
-
-**安装时下载器内版本选择（deno / rustup 模式）。** `curl | sh` 安装器不显式
-给版本时，先拉一个**可变 channel 指针文件**（如 deno 的
-`dl.deno.land/release-latest.txt`、rustup 的 channel manifest），再按指针下载
-真实产物——指针一改全量新装即转向，**不发版即可调流**。DWS 的 OSS
-`latest.txt`/`beta.txt`（`sync-to-oss.sh:13-14`）已是同构物，只差安装脚本消费
-它；install.sh 今天直接打 GitHub `/releases/latest` 重定向（`install.sh:190-198`），
-等价于把 GitHub 当不可调指针用。此模式是指针级灰度，做不到机器级百分比，
-需与 L2 分桶叠加。
-
-**双产物并行（VS Code Stable / Insiders 模式）。** 两个渠道各发各的包、用户
-自选安装，灰度靠「渠道人口结构」自然形成，无需任何运行时门控。DWS 的
-GitHub prerelease + npm `@beta` 已是它的轻量版（同包不同 tag 而非两个包名），
-L1 切法 B 的版本门控默认正是把「渠道差异」从纯流程下沉为可测试的代码事实。
-
---
-
-## 4. 推荐路线与改动点清单
-
-### 4.1 路线（与 roadmap 阶段 2/3 对齐后的排序）
-
-```
-L1 版本门控拆分（本工作区之上叠加，先合入）
-  → ~~阶段 2 四件套~~ ❌ CANCELLED（无运行时切换；upgrade force-multi）
-  → beta 轨发版先吃（L1 自动生效）+ L3 观察 ≥2 周
-  → stable 切流：小步直接 100%（删门控）；若要求渐进再上 L2 分桶
-  → mono retirement 判据（roadmap：issue/回访，**无**请求头占比）
-```
-
-### 4.2 改动点清单（文件级）
-
-| 项 | 文件 | 改动 | 估时 |
-|---|---|---|---|
-| L1-a Go 门控函数 | `internal/app/skill_setup.go`（或新 `internal/upgrade/track.go` 下沉共享） | `defaultSkillModeForVersion(version)`：含 `-beta.`→multi 否则 mono；替换非交互默认与交互排序 | 0.5d |
-| L1-b upgrade force-multi（已落地） | `internal/upgrade/paths.go` `UpgradeSkillLocations` | 有 multi→始终 multi（含 mono 盘迁移）；legacy 无 multi→mono | ✅ |
-| L1-c 脚本面门控 | `scripts/install.sh` / `install.ps1` / `install-skills.sh` / `build/npm/install.js` | 默认值解析加 `*-beta.*` 分支（若仍做 L1） | 0.5d |
-| L1-d 契约测试 | `test/scripts` / `internal/upgrade` / `internal/app` | 每面断言 beta→multi；upgrade mono→multi E2E | 0.5–1d |
-| L3 请求头 | — | **❌ CANCELLED** | — |
-| L2-a/b/c rollout.json | — | **已砍**（roadmap） | — |
-| S3/S4 | — | S3=重装 mono（可用）；S4 备份/rollback **❌ CANCELLED** | — |
-
-合计：L1 ≈ 2–2.5d；L3 ≈ 0.5d；L2 ≈ 2.5–3d（在 state.json 之后）。
-L1+L3 是进入 beta 观察期的最小集；L2 只在 stable 需要渐进切流时才启动，
-否则删门控一步到位即可。
-
-### 4.3 明确不做（沿用 roadmap/D2，本文补充）
-
- 不建服务端远程配置/灰度平台；不为灰度单独引入可变配置下发通道（OSS 变
-  指针仅作远期可选，今天无消费方）。
- 不动 `dws-skills.zip` 产物布局（D1）；不按轨发不同 zip——轨差异全部落在
-  版本门控的客户端行为上。
- 不用遥测做灰度测量（K7 本地 opt-in 无统计意义），观测只走 K8 请求头。
@@ -1,340 +0,0 @@
-# DWS multi-skill vs 悟空（dws-wukong）分发线对比
-
-> ⚠️ **留档注记（2026-08-05）**：悟空（dws-wukong）bundled-skill 分发线
-> 已于 2026-08-05 决策下线。本文自此仅作**历史调研留档**，不再作为任何
-> 对齐依据——文中的「判据 #3」「对齐清单（§7.1）」「待确认问题（§7.2）」
-> 等均随悟空线下线而 MOOT。其中 bundle 的自描述打包设计
-> （`manifest.json` + `scripts/_install.sh`）与 symlink 提升消费方式
-> 可作为未来打包方案参考保留。
->
-> 撰写日期：2026-08-05。聚焦 **multi-skill** 主题：DWS 侧（本工作区
-> `feat/skill-mode-migration`）已把 multi 翻转为全通道默认，而 mono 下线
-> 原判据 #3 曾要求"悟空 bundled skill 分发线（dws_res → bundled-skills）
-> 已切 multi"（[skill-multi-roadmap.md](skill-multi-roadmap.md) 阶段 4、
-> [skill-multi-migration-plan.md](skill-multi-migration-plan.md) §5；
-> 该判据已于 2026-08-05 随悟空线下线作废）。
-> 本文盘清悟空线现状、两边差异、切换影响与对齐清单。
->
-> 配套阅读：[skill-multi-roadmap.md](skill-multi-roadmap.md)（DWS 侧
-> as-implemented 事实源）、[skill-distribution-mechanism.md](skill-distribution-mechanism.md)
-> （DWS 分发/消费链路调研）。
-
-## 0. 摘要（TL;DR）
-
- **DWS 侧**：multi（`skills/multi/`，19 个 `dingtalk-*` 产品 skill +
-  `dws-shared`）已是安装（4 脚本面）、升级、`dws skill setup` 五面默认；
-  产物 `dws-skills.zip` 恒含"根 mono 副本 + `mono/` + `multi/`"三树，
-  二进制 embed 双树（`skills_embed.go:28`）。**产物零改动即完成默认翻转**。
- **悟空侧**：multi 打包能力**已合入 dws-wukong `develop`**（merge
-  `9eb801e5`，"DWS MultiSkill 与 Qwen Work Cloud 六平台打包"），但
-  **正式发版 target `make real-platform` 仍只打 mono**
-  `dingtalk-workspace.zip`；multi 以 `dingtalk-workspace-bundle.zip` 双包
-  形态存在（`bundle-platform` / `package-dual`），本地有 2026-07-03 的双包
-  实测产物，但**未随已发布版本出门**（`release/0.2.97`–`0.2.99` 均不含该
-  merge）。
- **关键缺口在客户端**：RewindDesktop（悟空桌面端）构建期
-  `download_binary.py` 只认 `dingtalk-workspace.zip` 单 zip；运行时
-  `dws_update.rs` 灰度更新也只 upsert 单个 `dingtalk-workspace` skill；
-  全仓 grep 无 `dingtalk-workspace-bundle` / T4b 处理。**端内尚无消费
-  multi bundle 的代码路径**。
- **悟空线的 multi 与 DWS 的 multi 是两套独立维护的树**（dws-wukong
-  `dingtalk-skills/` 12 产品 + `dws-shared`，由本仓 mono 机械派生；DWS
-  `skills/multi/` 19 产品 + `dws-shared`），内容靠 SOP 人工对齐，无自动
-  同源。判据 #3 的"切 multi"首先要解决的是**打包形态 + 端内加载**，
-  内容同源是紧随其后的问题。
- 结论（历史）：判据 #3 远未满足。对齐需要 dws-wukong 仓（发版 target）、
-  RewindDesktop 仓（构建期 + 运行时两条加载路径）两侧改动，详见 §7 清单。
-  **（2026-08-05 MOOT：悟空线下线，判据 #3 已作废——见
-  [skill-multi-roadmap.md](skill-multi-roadmap.md) 阶段 4 判据更新；
-  上述对齐工作不再需要。）**
-
-## 1. 证据源与版本快照
-
-本地仓库（均为真实磁盘证据，非仅凭文档）：
-
-| 仓库 | 本地路径 | 核查时状态 |
-|---|---|---|
-| DWS（本仓） | `~/GolandProjects/open-source/dws-skill-mode-migration` | `feat/skill-mode-migration`，含未 commit 的阶段 1 变更 |
-| dws-wukong 主仓 | `~/GolandProjects/open-source/dws-wukong` | checkout `codex/deploy-qwenwork-dev`（落后 develop 602 commits）；本文 Makefile/脚本结论均以 `develop` 分支内容（`git show develop:...`）为准 |
-| dws-wukong multiSkill 工作区 | `~/GolandProjects/open-source/dws-wukong-multiSkill` | detached @ `9eb801e5`（multiSkill merge 本体），`target/` 有 2026-07 实测产物 |
-| RewindDesktop | `~/IdeaProjects/RewindDesktop` | checkout `dws/0.2.98`；`develop` 上 `DEFAULT_DWS_RES_URL` = pod `0.2.96` |
-
-关键版本事实：
-
-| 事实 | 证据 |
-|---|---|
-| multiSkill merge `9eb801e5` 已入 `develop` 与本地 `release/0.2.100` | `git branch --contains 9eb801e5` |
-| `release/0.2.97` / `0.2.98` / `0.2.99` **不含** multiSkill merge；其 `real-platform` 不打 bundle | `git merge-base --is-ancestor` + `git show origin/release/0.2.99:Makefile` |
-| 当前发给悟空的 pod 包为 mono：`dws_res_mac.zip` 内仅 `dingtalk-workspace.zip`（单 skill，`SKILL.md`+`references/products/*`）+ 双架构二进制 | 主仓 `target/dws_res_mac.zip`（2026-06-01）`unzip -l` 实测 |
-| 双包形态已实测：`dws_res_mac/` 同时含 `dingtalk-workspace.zip`（924K）与 `dingtalk-workspace-bundle.zip`（1.27M） | multiSkill 工作区 `target/dws_res_mac.zip`（2026-07-03）实测 |
-| multi bundle 内部布局：`manifest.json` + `scripts/_install.sh` + `skills/<name>/SKILL.md...` 平铺目录 | multiSkill 工作区 `target/dingtalk-workspace.zip`（2026-07-24）实测 |
-
-**本地未能验证**（详见 §7.2 问题清单）：pod 线上当前包内容（需内网
-SSO）；`/Applications/Wukong.app` 未安装在本机（bundled-skills 目录不存在，
-无法核对在端真实 zip）；RewindDesktop 端"T4b 二选一"灰度逻辑（全仓无匹配，
-疑似未开发或在平台侧）；Qwen Work Cloud 六平台打包的实际发布状态。
-
-## 2. 分发链路对比
-
-### 2.1 链路全景
-
-DWS 侧（本仓，multi 已默认）：
-
-```text
-skills/mono/ ─┬─ go:embed all:skills/mono all:skills/multi (skills_embed.go:28)
-skills/multi/ ┘      │  `dws skill setup` 默认源（embed 优先）
-              │
-              └─ scripts/release/post-goreleaser.sh:220-248 ──► dws-skills.zip
-                 （根 = mono 副本 + mono/ + multi/，三树恒含）
-                        │
-        GitHub Release / Gitee / OSS / npm tarball / Homebrew / 专项脚本
-                        │
-   install.sh · install.ps1 · install-skills.sh · npm install.js（四面默认 multi）
-   + `dws skill setup`（第五面默认 multi）+ `dws upgrade`（布局探测→multi 刷新）
-                        │
-              各 agent home 平铺 dingtalk-*/（互斥清理 dws/ 与过期 skill）
-              + ~/.dws/skills/{mono,multi} 双缓存
-```
-
-悟空侧（dws-wukong + RewindDesktop，发版线仍 mono）：
-
-```text
-上游 CLI 仓 ../dingtalk-workspace-cli（go.mod replace，sync-upstream 按
-CLI_UPSTREAM_TAG 重建 release 分支）──► 只提供 Go 代码，不提供 skill 内容
-
-dws-wukong 仓内 skill 内容（自维护）：
-  dingtalk-workspace/（mono 源，含 overlays/real）
-      ├─ build-workspace-zip ──► dingtalk-workspace.zip（mono，单 skill）
-      └─ scripts/sync-monolith-to-multiskill.py ──► dingtalk-skills/（multi 派生树）
-             └─ scripts/build-bundle.sh ──► dingtalk-workspace-bundle.zip
-                  （manifest.json + scripts/_install.sh + skills/<name>/）
-
-make real-platform（发版默认）──► dws_res_{mac,win}.zip
-      = dws 二进制 + dingtalk-workspace.zip（仅 mono）
-make bundle-platform / package-dual（已合入 develop，未用于正式发版）
-      = dws 二进制 + dingtalk-workspace.zip（mono 兜底）+ dingtalk-workspace-bundle.zip
-                        │
-        pod.alibaba-inc.com zipUpload（SSO 浏览器上传，版本号独立递增）
-                        │
-        RewindDesktop scripts/download_binary.py（DEFAULT_DWS_RES_URL 手工对齐）
-                        │
-        Wukong.app Contents/Resources/resources/
-            ├─ dws/bin/dws（二进制）
-            └─ bundled-skills/dingtalk-workspace.zip（原样拷贝的单 zip）
-                        │
-        运行时两条路：
-        a) 启动同步 initialize_bundled_skills_from_resources
-           → 解 zip 到中央技能库 ~/.real/.skills/bundled/dingtalk-workspace
-           （及 ~/.real/users/*/.skills/bundled）
-        b) 灰度自更新 dws_update.rs：Gaea 开关 wukong/dws_auto_update_enabled_v2
-           → LWP /r/Adaptor/DwsGrayI/getLatest 取 {version,url,sha256}
-           → 下载 dws_res → 换 seed 二进制 + 换 bundled zip + upsert 单 skill
-```
-
-### 2.2 链路对照表
-
-| 环节 | DWS（本仓） | 悟空线 | 锚点（悟空侧） |
-|---|---|---|---|
-| skill 事实源 | `skills/mono` + `skills/multi`（同仓双树，multi 19 产品 + dws-shared） | dws-wukong 仓 `dingtalk-workspace/`（mono 源）→ 派生 `dingtalk-skills/`（12 产品 + dws-shared）；**与上游 skills/ 无自动同步** | `scripts/sync-monolith-to-multiskill.py` docstring |
-| 二进制与 skill 的版本耦合 | embed 进二进制，天然同版 | 二进制来自上游 tag（`go.mod:58` replace + `sync-upstream` pin `CLI_UPSTREAM_TAG`）；skill 在 dws-wukong 仓随 `VERSION`/`main.go` 双写发版 | dws-wukong `Makefile` `sync-upstream` |
-| 打包产物 | `dws-skills.zip`：根 mono 副本 + `mono/` + `multi/`（`post-goreleaser.sh:220-248`）；embed 双树 | `dws_res_{mac,win}.zip`：二进制 + `dingtalk-workspace.zip`（mono）；双包 target 已存在但未上发版线 | dws-wukong `Makefile` `real-platform` / `bundle-platform` / `package-mac-dual` |
-| 渠道 | GitHub/Gitee/OSS/npm/Homebrew/专项脚本，7 个安装面 | pod zipUpload（SSO）→ RewindDesktop `download_binary.py` → 客户端 bundle | release skill 文档 + `download_binary.py:72-84` |
-| 端内安装 | 4 脚本 + `dws skill setup` 平铺到 16 个 agent home（父目录门控） | 构建期拷贝 zip 进 app 资源；启动时解到 `~/.real/.skills/bundled/` | `startup.rs:486-554` |
-| 运行时更新 | `dws upgrade`（布局探测→multi 刷新 + 互斥清理 + 缓存刷新） | 客户端灰度自更新（Gaea + LWP），整包替换 seed 二进制 + skill zip | `dws_update.rs:107,27-28,440-545` |
-| 灰度能力 | 无服务端：beta 轨（L1）+ 可选 `rollout.json` 分桶（L2）+ 公告 kill switch（决策 D2） | 有服务端：Gaea 开关 + LWP getLatest + pod 版本号 | 决策记录 D2 vs `dws_update.rs` |
-
-## 3. skill 布局对比
-
-### 3.1 三种形态
-
-| 形态 | 布局 | 消费方 |
-|---|---|---|
-| DWS mono | 单 skill：`SKILL.md` + `references/` + `scripts/`，装进 `<agent-home>/dws/` | DWS legacy 安装面；悟空 `dingtalk-workspace.zip` 同构（多 `plugins/`、real overlay） |
-| DWS multi | 平铺目录树：`multi/dingtalk-<product>/{SKILL.md,references,scripts}` + `multi/dws-shared/`，无 manifest、无安装器，拷贝即平铺到 agent home | DWS 五面默认；`dws upgrade` 探测 `multi/` 树（`internal/upgrade/paths.go:363-369`） |
-| 悟空 multi bundle | zip 内 `manifest.json`（`{"version":...}`）+ `scripts/_install.sh` + `skills/<name>/` 平铺目录（含 `dws-shared`） | **Qwen Work Cloud 已消费**（`_install.sh` 把 `skills/*` 以 symlink 提升到一层 skills 根，供 Codex/OpenCode 扫描；`.dws-multiskill-current` + `.dws-multiskill-links` 记账）；**Wukong.app 尚未消费** |
-
-注意两种 multi 的"平铺"语义不同：DWS multi 是**裸目录树**，由安装面自己
-拷贝到各 agent home；悟空 bundle 是**自描述包**（manifest + 安装脚本），
-由消费方解包后提升。`dws-skills.zip` 的 `multi/` 树与
-`dingtalk-workspace-bundle.zip` 的 `skills/` 树**布局同构但内容不同源**
-（见 §3.3）。
-
-### 3.2 悟空客户端加载约定（RewindDesktop 实测）
-
-构建期（`scripts/download_binary.py`）：
-
- `DWS_RES_WORKSPACE_ZIP = "dingtalk-workspace.zip"`（`:103`），
-  `resolve_dws_res_contents`（`:1366-1380`）强制 dws_res 内必须同时有
-  平台二进制和**这个文件名的 zip**——`dingtalk-workspace-bundle.zip` 会被
-  原样忽略（不报错，但也不使用）。
- `sync_dingtalk_workspace_bundle`（`:1422-1432`）把该 zip **原样拷贝**到
-  `tauri-app/src-tauri/resources/bundled-skills/dingtalk-workspace.zip`，不解包。
-
-运行时（`tauri-app/src-tauri/src/skills/startup.rs`）：
-
- `collect_bundled_skill_sources`（`:486-506`）扫描 `resources/bundled-skills/`：
-  **每个子目录或每个 `.zip` = 一个 skill**，`skill_id = 文件主干名`
-  （`dingtalk-workspace.zip` → skill id `dingtalk-workspace`）。
-  该扫描**天然支持多 skill 平铺**（放 `dingtalk-mail.zip`、`dingtalk-doc.zip`
-  就会被分别注册），但**不认识嵌套 bundle**：若把
-  `dingtalk-workspace-bundle.zip` 丢进去，只会被当成一个名叫
-  `dingtalk-workspace-bundle` 的单 skill 解开（内容是 manifest+skills/ 目录，
-  不会被提升）。
- `sync_bundled_skill_source`（`:533-554`）解 zip 拷贝进中央技能库；
-  `cleanup_removed_bundled_skills`（`:508-531`）会删除 bundled-skills 里
-  已不存在的 bundled skill（mono→multi 切换时可自动清掉旧
-  `dingtalk-workspace`，前提是 store 记录完好）。
- 灰度自更新 `dws_update.rs:440-545` 硬编码单 skill：
-  `upsert_bundled_skill_from_source(store, DWS_WORKSPACE_SKILL_ID, …)`
-  （`DWS_WORKSPACE_SKILL_ID = "dingtalk-workspace"`），换包 = 替换
-  `bundled-skills/dingtalk-workspace.zip` + 重 upsert 这一个 skill。
-
-结论：端内**构建期与运行时两条路都按"单 zip 单 skill"接线**；要支持
-multi，二选一：(a) 端内学会解 bundle（manifest + 提升子 skill），或
-(b) 打包侧把每个子 skill 打成独立 zip 平铺进 dws_res，复用现有平铺扫描
-（仅需把 `dws_update.rs` 的单 skill upsert 改为遍历）。详见 §6.2。
-
-### 3.3 两套 multi 树的集合差异
-
-| | DWS `skills/multi/` | dws-wukong `dingtalk-skills/`（develop） |
-|---|---|---|
-| 产品 skill 数 | 19 | 12 |
-| 共有（12 个） | aisearch, aitable, calendar, chat, contact, doc, drive, mail, minutes, misc, todo, wiki | 同左 |
-| 仅 DWS 有（7 个） | dev, event, hrbrain, markdown, pat, profile, skill | — |
-| 共享层 | `dws-shared` | `dws-shared`（内容独立维护） |
-| 内容来源 | 本仓直接维护 | 由本仓 mono `dingtalk-workspace/` 经 `sync-monolith-to-multiskill.py` 机械派生（链接改写 + misc 桶归并） |
-| 场景 skill | 不涉及 | `dingtalk-products-skills/` 23 个 scenario skill **仅保留源码，不进统一发布包**（`build-bundle.sh` 头注释） |
-
-含义：即使悟空线明天切到 multi 形态，其 multi **内容**与 DWS multi 也不
-一致（少 7 个产品、各自演化）。判据 #3 只要求"分发线切 multi"（形态），
-但长期看内容同源（或明确的子集契约）需要一并决策。
-**（2026-08-05 MOOT：悟空线下线，"DWS skills/multi(19) vs 悟空
-dingtalk-skills(12) 分歧"的长期统一问题随之作废，无需统一；DWS
-`skills/multi/` 成为唯一 multi 事实源。）**
-
-## 4. 版本对齐对比
-
-| 维度 | DWS | 悟空线 |
-|---|---|---|
-| skill↔CLI 耦合 | embed 进二进制，`dws skill setup` 装的就是本二进制版本（`skills_embed.go` + `skill_setup_embed.go`） | 二进制版本 = 上游 tag（`sync-upstream` pin）；skill 版本 = dws-wukong `Makefile VERSION` + `main.go version` 双写；**两者只通过"同一次发版动作"对齐，无结构性强约束** |
-| 产物版本 | `dws-skills.zip` 随 goreleaser 与二进制同 tag 发布 | pod 版本号独立于 dws 版本（右most 段 +1 递增，mac/win 各自一条线，如 mac `0.2.28.0`、win `0.2.2`）；RewindDesktop `DEFAULT_DWS_RES_URL` 手工改指 |
-| 端内版本事实源 | — | 客户端以 `dws --version` 输出为准（`get_dws_version_sync`，失败回退 `versions.json`）；skill zip 无独立版本概念（mono zip 内无 manifest） |
-| multi 包版本 | 无 manifest；版本 = 所属 zip/二进制版本 | bundle 内 `manifest.json` 带 `version`（`build-bundle.sh` 由 `$(VERSION)` 写入）——**multi 形态反而第一次给 skill 包带来了显式版本号** |
-| 升级时的布局兼容 | 新 zip 恒含 `multi/`，`LocateSkillsRoot` 优先 multi；老 zip 自然落回 mono（决策 D1，产物零改动） | dws_res 布局固定（二进制 + workspace zip）；双包形态下 mono zip 保留作兜底（`package-mac-dual` 注释："端内 validate 必含，bundle 缺失时兜底"） |
-
-## 5. 更新 / 回滚对比
-
-| 维度 | DWS | 悟空线 |
-|---|---|---|
-| 更新触发 | 用户主动 `dws upgrade` / 重装脚本 | 两条：(a) 随客户端版本更新（app 内嵌资源替换 + 启动同步）；(b) 运行时灰度自更新（Gaea `dws_auto_update_enabled_v2` + LWP getLatest → 下载整包 → 换 seed 二进制 + 换 bundled zip + upsert skill） |
-| skill 刷新语义 | 布局探测（multi 优先）+ 按 home 互斥清理（删 `dws/`、过期 `dingtalk-*`/`dws-shared`），清理失败则该 home 不装，杜绝共存（`internal/upgrade/paths.go:217-299`） | 启动同步按 bundled-skills 现状全量对账（新增拷贝、缺失删除，`startup.rs:508-554`）；自更新路径是单 zip 替换 + 单 skill upsert（`dws_update.rs:462-513`） |
-| 回滚 | `dws upgrade --rollback` 只回二进制不回 skill；备份式安装 + `dws skill mode rollback` 是阶段 2 P0，**尚未落地** | 无显式回滚命令；事实回滚 = Gaea 开关关闭/改指旧 pod 版本重新下发，或客户端版本回退；`replace-wukong-skill.sh`（仅存在于 `codex/deploy-qwenwork-dev` 分支）提供人工替换 + 重启 |
-| 半装保护 | 现状 `RemoveAll` 直删、失败仅 warning（风险表已列，阶段 2 改备份式） | zip 整体替换 + preflight 可写性检查，失败提示重启；粒度为整个 skill 包，无子 skill 级半装概念 |
-| 离线/内网 | embed 兜底（无网可 setup） | app 内嵌 zip 兜底；灰度通道依赖内网 LWP/pod 可达 |
-
-## 6. 切换影响分析
-
-### 6.1 DWS 切 multi 默认 / 最终删 mono 对悟空线的影响点
-
-| # | 影响点 | 评估 |
-|---|---|---|
-| 1 | DWS 删 `skills/mono` 后上游 embed 只剩 multi；悟空二进制由上游 tag 构建，`dws skill setup` 行为随之变 | **低**。悟空客户端不从 embed 装 skill（走 bundled zip）；但悟空用户在端内手动跑 `dws skill setup` 时会得到 multi——行为变化需在悟空侧公告 |
-| 2 | `dws-skills.zip` 布局（根 mono + mono/ + multi/） | **零影响**。悟空线不消费 `dws-skills.zip`；DWS 侧也已承诺不动该产物（决策 D1、"明确不做"） |
-| 3 | dws-wukong 的 mono 内容源 `dingtalk-workspace/` 与上游 `skills/mono` 本就各自维护 | **低（但需注意）**。上游删 mono 不会直接打破 dws-wukong 构建；但两边 mono 的"内容漂移对照基准"消失，dws-wukong mono 将彻底成为孤儿副本，加速与上游 CLI 能力的文档漂移 |
-| 4 | mono 下线判据 #3 反向卡住 DWS 侧进度 | **高（流程性）**。悟空线一天不切 multi，DWS 就不能物理删 `skills/mono/`（roadmap 风险表"悟空线外挂"行）。**（已解除 2026-08-05：悟空线下线，判据 #3 作废，DWS 侧进度不再受仓外闸门约束）** |
-
-### 6.2 悟空线"吃 multi"的改造选项
-
-**选项 A：端内解 bundle（dws-wukong 现有 bundle 产物直接被消费）**
-
- dws-wukong 侧：`real-platform` 改打（或加打）`dingtalk-workspace-bundle.zip`
-  ——`bundle-platform` / `package-dual` 已就绪，基本零新开发。
- RewindDesktop 侧（主要工作量）：
-  - `download_binary.py`：`resolve_dws_res_contents` 接受/校验
-    `dingtalk-workspace-bundle.zip`，同步进 `resources/bundled-skills/`；
-  - `startup.rs`：识别 bundle（`manifest.json` + `skills/*`），把每个子
-    skill 注册为独立 bundled skill（逻辑等价于 `_install.sh` 的提升，
-    但落在中央技能库）；
-  - `dws_update.rs`：灰度自更新从"单 skill upsert"改为"bundle 全量对账"
-    （可复用启动同步的对账逻辑）。
- 优点：与 Qwen Work Cloud 已消费的包形态一致，一份产物两个端；bundle 自
-  带 `manifest.json` 版本号。
- 缺点：端内要新增 bundle 解析/提升代码与测试；`skills/` 嵌套布局与现有
-  "一 zip 一 skill"约定不同，需谨慎处理迁移期（旧 mono skill 清理）。
-
-**选项 B：打包侧拆 zip（端内零新格式）**
-
- dws-wukong 侧：新增 target 把 `dingtalk-skills/` 每个子 skill 打成独立
-  zip（`dingtalk-mail.zip` … `dws-shared.zip`）平铺进 dws_res。
- RewindDesktop 侧：`download_binary.py` 改为同步多个 zip；
-  `startup.rs` 现有平铺扫描**零改动**（自动注册每个 zip）；
-  `dws_update.rs` 仍需从单 skill upsert 改为遍历。
- 优点：复用端内现有"一 zip 一 skill"约定，启动路径几乎不动。
- 缺点：与 Qwen Work 的 bundle 形态分叉（一份内容两种包）；dws_res 内文件
-  数膨胀；`dws-shared` 作为独立 skill id 出现在用户可见列表里需要确认
-  端内展示策略。
-
-**选项 C（过渡态，事实已在用）**：双包并存——mono zip 兜底 + bundle 灰度，
-端内按灰度二选一。`package-dual` 的注释已写明此意图（"端内 T4b 二选一，
-monolith 端内 validate 必含，bundle 缺失时兜底"），但**端内 T4b/灰度选择
-逻辑在 RewindDesktop 尚未找到实现**，当前双包发出去也只会用 mono。
-
-### 6.3 mono 下线判据 #3 的建议验收方式
-
-> **（2026-08-05 MOOT：判据 #3 已随悟空线下线作废，本节验收清单不再
-> 适用，仅留档。）**
-
-判据原文："悟空 bundled skill 分发线（dws_res → bundled-skills，本仓库
-外）已切 multi"（原 [skill-multi-migration-plan.md](skill-multi-migration-plan.md)
-§5-3；该判据已于 2026-08-05 作废，plan §5 已重新编号）。建议按以下
-可核查项验收（全绿才算满足）：
-
-1. **发版**：dws-wukong 正式 release 流程（release skill 文档中的
-   `make real-platform` 路径）产出的 dws_res 内含 multi 形态包（bundle 或
-   平铺 zip 集），且 pod 上当前版本即为该形态。
-2. **构建期消费**：RewindDesktop `develop` 的 `download_binary.py` 把 multi
-   形态同步进 `bundled-skills/`（不再是只认 `dingtalk-workspace.zip`）。
-3. **运行时消费**：悟空端启动同步后，`~/.real/.skills/bundled/` 下出现
-   `dingtalk-*` 多 skill（而非单个 `dingtalk-workspace`）；灰度自更新路径
-   同样支持多 skill 对账。
-4. **实机回归**：全新安装悟空 → 技能列表出现各 `dingtalk-*` skill 且
-   路由正常；从 mono 旧版升级 → 旧 `dingtalk-workspace` 被清理、无双份
-   派发；灰度通道下发一次 multi 包 → 更新后无残留。
-5. **回退预案**：悟空侧保留 mono 兜底产物或快速重发能力，直至 DWS 删
-   mono 窗口关闭。
-
-## 7. 结论
-
-### 7.1 对齐清单（悟空侧待办，按优先级）
-
-> **（2026-08-05 MOOT：悟空线下线，本清单整体不再需要执行，仅留档。）**
-
-| 优先级 | 事项 | 仓库/位置 | 备注 |
-|---|---|---|---|
-| P0 | 决策端内 multi 消费方案（选项 A 解 bundle vs 选项 B 平铺 zip） | RewindDesktop + dws-wukong 联合 | 建议 A：与 Qwen Work 已消费形态一致，且 bundle 带版本 manifest |
-| P0 | `download_binary.py` 支持 multi 形态同步进 bundled-skills | RewindDesktop `scripts/download_binary.py:103,1366-1432` | 选项 A 下识别 `dingtalk-workspace-bundle.zip` |
-| P0 | 启动同步/技能库支持 bundle 解包与子 skill 注册 | RewindDesktop `tauri-app/src-tauri/src/skills/startup.rs:486-554` | 含旧 mono skill 的迁移清理（现有 `cleanup_removed_bundled_skills` 可复用语义） |
-| P0 | 灰度自更新支持 multi（单 skill upsert → 多 skill 对账） | RewindDesktop `.../dws_update.rs:440-545` | 否则运行时更新会把 multi 打回 mono |
-| P1 | 正式发版 target 切 multi（`real-platform` 改打/加打 bundle，或改用 `bundle-platform`/`package-dual`） | dws-wukong `Makefile` | 打包能力已在 develop，缺的是设为默认 + release skill 文档同步更新 |
-| P1 | 端内灰度选择逻辑落地（双包二选一/T4b，或确认直接全量切） | RewindDesktop（未找到现有实现） | 若选选项 C 过渡则必须 |
-| P1 | 两套 multi 树的内容同源策略：dws-wukong `dingtalk-skills/`（12 产品）vs DWS `skills/multi/`（19 产品） | dws-wukong + 本仓 | **MOOT（2026-08-05）**：悟空线下线，无需统一；原备注：至少明确"悟空子集"契约与同步 SOP 的归属；7 个缺失产品（dev/event/hrbrain/markdown/pat/profile/skill）是否需要进悟空 |
-| P2 | `replace-wukong-skill.sh` 等运维脚本支持 bundle 形态 | dws-wukong `scripts/deploy/` | 当前只在特性分支且只处理单 zip |
-| P2 | 悟空侧公告：端内 `dws skill setup` 行为随上游 embed 变化 | dws-wukong 发版流程 | 对应 §6.1 影响点 1 |
-| P2 | 判据 #3 验收清单（§6.3）写入 DWS roadmap 并跟踪 | 本仓 `docs/skill-multi-roadmap.md` | 阶段 3 期间启动 |
-
-### 7.2 待确认问题（本地无法闭环，需找人/仓库确认）
-
-| # | 问题 | 建议确认方 |
-|---|---|---|
-| 1 | pod 线上当前 `dws_res_mac/win` 的版本与内部构成（是否已有人发过双包） | pod.alibaba-inc.com（需内网 SSO）/ 悟空发版 owner |
-| 2 | "端内 T4b 二选一"灰度逻辑是否已存在（在哪个仓库/平台），还是仅写在 Makefile 注释里的规划 | RewindDesktop 团队 / 悟空端内灰度平台 owner |
-| 3 | LWP `/r/Adaptor/DwsGrayI/getLatest` 服务端返回的下载 URL 指向何处（pod？另一制品库？），multi 包下发是否需要服务端配合改造 | Adaptor/DwsGrayI 服务端 owner |
-| 4 | Qwen Work Cloud 六平台打包的发布状态与其对 bundle 的消费方式是否可作为悟空端改造的直接参照 | dws-wukong 仓 owner（merge `9eb801e5` 提交者） |
-| 5 | 悟空端内是否允许 `dws-shared` 作为独立 bundled skill 暴露（名称/展示/路由策略），还是应内联进各产品 skill | RewindDesktop 技能库 owner |
-| 6 | dws-wukong `dingtalk-skills/` 与上游 `skills/multi/` 的长期关系：保持派生自本仓 mono，还是改为从上游 multi 同步（**MOOT 2026-08-05**：悟空线下线，无需统一） | dws-wukong + DWS 双侧 owner 联合决策 |
-| 7 | 悟空线切换的目标时间窗（决定 DWS 阶段 4 判据 #3 的最早可满足点）（**MOOT 2026-08-05**：判据 #3 已作废） | 悟空发版 owner |
-
---
-
-*本文所有"已验证"结论均可按 §1 的仓库路径与文中锚点复查；未能本地验证
-的项集中在 §7.2。*
@@ -0,0 +1,108 @@
+# DWS 统一命令框架设计概要
+
+> 状态：Framework core 已实现，dingtalk-dev/devapp 首批命令渐进接入中。本文定义框架能力、集成边界和首批 pilot 的发布纪律；其余产品命令迁移、Skill 更新和真实服务复验继续由后续 PR 独立完成。
+
+## 1. 产品裁决
+
+1. 不公开 `--output-contract`，也不增加任何等价别名。
+2. Agent 继续只使用既有 `--format json`。
+3. 每条 terminal command 在一个 release 中只有一个 active wire contract：已迁移命令直接使用统一结果，未迁移命令保持 legacy。
+4. contract 不由用户参数、环境变量、会话能力协商或 Agent 选择。
+5. 回滚是命令声明与发布行为，不改变消费者 argv。
+6. 本 PR 只迁移完成命令级兼容审计的 dingtalk-dev/devapp pilot；其他命令路径、参数和输出保持不变。
+
+## 2. 渐进迁移
+
+内部状态机：
+
+```text
+legacy_only -> dual_validate -> unified_active -> unified_stable -> unified_only
+```
+
+- `legacy_only`：只构造、输出 legacy。
+- `dual_validate`：业务只执行一次；外部仍逐字输出 legacy；同一内存结果 shadow-build 统一结果并严格校验。
+- `unified_active`：`--format json` 直接返回统一结果信封，可按发布声明回退。
+- `unified_stable`：完成真实 Agent 消费观察和兼容窗口。
+- `unified_only`：清理仅服务 legacy 的产品 renderer。
+
+状态是每条 terminal command 的内部发布元数据。Help、Skill、Agent Schema 不展示迁移状态，也不让消费者选择协议。
+
+## 3. 统一结果
+
+统一命令框架表达四类结果：
+
+```text
+success          请求完成且命令认为操作已完成
+pending          请求被受理，但异步操作尚未终结
+partial_failure  批量操作有成功项，也有失败或未知项
+failure          请求或操作失败
+```
+
+JSON 基本形态：
+
+```json
+{
+  "ok": true,
+  "outcome": "success",
+  "data": {}
+}
+```
+
+硬不变量：
+
+```text
+ok == (outcome in {success, pending})
+process rc == 0 <=> ok == true
+top-level error present <=> outcome == failure
+one invocation emits exactly one primary result
+```
+
+框架负责 L1 request outcome 和 L2 operation outcome 的统一表达；L3 verification 必须由产品命令基于业务事实实现，框架不得自动推断 `changed/verified`。
+
+## 4. 输出与错误纪律
+
+- 统一 JSON primary result 写 stdout；stderr 只写诊断。普通命令不把
+  NDJSON 作为通用结果契约；持续事件流若需要逐事件输出，由 event 命令
+  自己声明专用流协议。
+- 分页统一输出到信封 `meta.pagination`，并在命令 Schema 中作为与 `result`
+  同级的 `pagination` 能力声明；`result.data_schema` 只描述业务 data，不再
+  混入分页控制字段。
+- 日志不得污染 stdout。
+- `ok`、`retryable`、`dry_run` 等必须是 JSON boolean。
+- 失败由框架根据 typed error 映射退出码；产品代码不能自报任意 rc。
+- `partial_failure` 保留 `succeeded[]/failed[]/unknown[]`，使用非零 rc 7。
+- `pending` 必须提供 operation id、state 和可执行的 `next_command`。
+- `endpoint_exhausted` 只表示观察到当前 endpoint 分页耗尽；false 必须带 `next_token`，不得扩大成索引健康或业务数据完整。
+- dry-run 是已经完成的无副作用预览，表达为 `success + dry_run:true`，不是 `pending`。
+
+## 5. 重试与超时边界
+
+- 框架只统一表达 `retryable`、`retry_after_seconds` 和 `execution_started`，不自动决定业务操作能否安全重放。
+- 写调用的模糊失败、HTTP timeout 和异步等待预算属于 transport/产品集成范围，不在本 PR 改动。
+- 产品迁移必须证明其重试声明与幂等性、安全等级一致。
+
+## 6. 集成范围
+
+- 产品命令通过 `corecmd.ResultInvoke` 构造 `CommandResult`，由 root 单一出口渲染。
+- 首批 dingtalk-dev/devapp 命令用于验证原子命令与 shortcut 的接入缝；未进入 pilot 的 shortcut、长连接、批量写和异步任务各自需要独立集成 PR。框架 core 不替产品推断 success、pending、partial 或分页事实。
+- 每条 terminal command 独立 rollout；不能整域一次切换，也不能通过 Agent 参数选择协议。
+- 已有命令在进入 `unified_active` 前必须保留 legacy byte golden，并完成真实 Agent 语义扫描。
+
+## 7. 对齐原则
+
+- 对齐 Lark CLI：统一 envelope/emitter、typed error、partial、pending、分页窄语义和强类型结果。
+- 对齐 GWS：机器结果稳定结构化、日志与数据分流、消费者不协商协议版本。
+- DWS 保留差异：声明式 Agent Schema、安全门禁、静态命令与 shortcut 共存，以及四 outcome 模型。
+
+## 8. 发布门禁
+
+命令晋级 `unified_active` 前至少满足：
+
+1. success/failure/dry-run golden；批量或异步命令另有 partial/pending golden。
+2. 业务请求 exactly once；dual validation 不得二次调用服务端。
+3. legacy 命令 stdout/stderr/rc 字节级回归不变。
+4. Help、Schema 和全仓示例不存在协议选择参数。
+5. `--format json` 输出单个合法统一结果文档，stdout 无日志污染。
+6. typed error、进程 rc 与信封 `error.exit_code` 一致。
+7. 安全声明、确认门禁与 dry-run 运行时行为同源。
+8. Agent 语义扫描记录命令级迁移证据；发布回滚无需修改 Agent argv。
@@ -0,0 +1,117 @@
+<!doctype html>
+<html lang="zh-CN">
+<head>
+  <meta charset="utf-8">
+  <meta name="viewport" content="width=device-width,initial-scale=1">
+  <title>DWS Wiki Shortcut 全景评审</title>
+  <style>
+    :root{--ink:#14213d;--muted:#5c677d;--line:#dbe4f0;--paper:#fff;--bg:#f3f7fb;--blue:#1769e0;--cyan:#00a6a6;--green:#178746;--amber:#a45b00;--red:#b42318;--shadow:0 14px 34px rgba(20,33,61,.08)}
+    *{box-sizing:border-box}body{margin:0;overflow-x:hidden;background:linear-gradient(150deg,#edf5ff 0,#f8fbff 45%,#eef8f5 100%);color:var(--ink);font:15px/1.65 -apple-system,BlinkMacSystemFont,"Segoe UI","PingFang SC",sans-serif}
+    main,.card,.two>*{min-width:0}main{width:min(1180px,calc(100% - 32px));margin:28px auto 72px}.hero,.card{background:rgba(255,255,255,.96);border:1px solid var(--line);border-radius:22px;box-shadow:var(--shadow)}
+    .hero{padding:38px;background:radial-gradient(circle at 95% 0,#dff8f3,transparent 36%),linear-gradient(135deg,#fff,#f5f9ff)}h1{font-size:34px;line-height:1.2;margin:0 0 10px}.lead{font-size:17px;color:var(--muted);max-width:900px}.meta{display:flex;gap:10px;flex-wrap:wrap;margin-top:20px}.pill{border:1px solid #cbd9ea;border-radius:999px;padding:5px 11px;background:#fff;font-size:13px}
+    .grid{display:grid;grid-template-columns:repeat(4,1fr);gap:14px;margin:18px 0}.metric{padding:20px}.metric b{display:block;font-size:31px;color:var(--blue)}.metric span{color:var(--muted)}
+    section{margin-top:22px}.card{padding:26px}h2{font-size:23px;margin:0 0 14px}h3{font-size:17px;margin:22px 0 8px}.callout{border-left:4px solid var(--blue);background:#f2f7ff;padding:14px 16px;border-radius:8px}.warn{border-color:var(--amber);background:#fff8eb}.ok{border-color:var(--green);background:#effbf4}
+    table{width:100%;border-collapse:collapse;font-size:14px}th,td{text-align:left;vertical-align:top;border-bottom:1px solid var(--line);padding:11px 9px}th{color:#41516b;background:#f7f9fc;position:sticky;top:0}code{background:#edf2f8;border-radius:5px;padding:2px 5px;color:#24466e}.tag{display:inline-block;border-radius:999px;padding:2px 8px;font-size:12px;font-weight:650;white-space:nowrap}.full{background:#e6f6ec;color:#116436}.partial{background:#fff0d5;color:#875000}.extra{background:#e8f1ff;color:#1854a5}.fixed{background:#f1eaff;color:#6338a5}
+    .toolbar{display:flex;flex-wrap:wrap;gap:10px;margin:12px 0}.toolbar input,.toolbar select{border:1px solid #bdcada;border-radius:10px;padding:9px 11px;background:#fff;min-width:min(220px,100%);max-width:100%;flex:1 1 220px}.matrix{max-height:620px;overflow:auto;border:1px solid var(--line);border-radius:12px}.two{display:grid;grid-template-columns:1fr 1fr;gap:18px}.small{color:var(--muted);font-size:13px}ul{padding-left:20px}.footer{color:var(--muted);text-align:center;margin-top:22px}@media(max-width:850px){.grid,.two{grid-template-columns:1fr 1fr}.hero{padding:25px}}@media(max-width:560px){.grid,.two{grid-template-columns:1fr}main{width:min(100% - 18px,1180px)}.card{padding:18px}h1{font-size:28px}}
+  </style>
+</head>
+<body><main>
+  <header class="hero">
+    <h1>DWS Wiki Shortcut 全景评审</h1>
+    <p class="lead">以 13 项成熟 Wiki 用户任务为基线，重新审视 DWS 的空间、成员、节点与动态能力。本次不是按命令名凑数：每个入口都要求真实业务证据，缺失数组、畸形响应、空确认或读回不一致一律失败。</p>
+    <div class="meta"><span class="pill">评审日期 2026-08-14</span><span class="pill">独立 worktree / 独立分支</span><span class="pill">真实组织数据 E2E 28/28</span><span class="pill">报告已去标识化</span></div>
+  </header>
+
+  <div class="grid">
+    <div class="card metric"><b>20</b><span>公开 Wiki Shortcuts</span></div>
+    <div class="card metric"><b>13/13</b><span>基线用户任务有对应路径</span></div>
+    <div class="card metric"><b>7</b><span>DWS 额外场景</span></div>
+    <div class="card metric"><b>20/20</b><span>真实数据能力已触达</span></div>
+  </div>
+
+  <section class="card">
+    <h2>结论先行</h2>
+    <div class="callout ok"><strong>DWS 已形成比“API 快捷别名”更完整的 Wiki 任务层。</strong> 基线中的 13 个用户任务均有对应入口；DWS 还提供空间搜索/详情/唯一解析、成员角色更新、库内节点搜索、协作动态和按空间名新建文档。创建、复制、移动等关键写能力从“请求发出”升级为“终态 + ID + 读回”成功标准。</div>
+    <div class="callout warn" style="margin-top:12px"><strong>能力边界必须诚实表达。</strong> DingTalk 成员接口不提供游标，单次真实上限是 50，因此不能实现成员 <code>--page-all</code>；成员身份只接受同组织可用的 userId，无法提供 email/open_id 等多种身份模式；节点创建也没有等价的 origin/shortcut 模式。这些差异保留为明确边界，而不是用本地循环或空结果伪装。</div>
+  </section>
+
+  <section class="card">
+    <h2>13 项基线任务逐条映射</h2>
+    <div class="matrix"><table><thead><tr><th>基线任务</th><th>DWS 主入口</th><th>结论</th><th>DWS 视角与边界</th></tr></thead><tbody>
+      <tr><td><code>+space-list</code></td><td><code>wiki +space-list</code></td><td><span class="tag full">完整对齐</span></td><td>严格空集合、游标续传、自动翻页、停滞检测；支持组织/我的知识库。</td></tr>
+      <tr><td><code>+space-create</code></td><td><code>wiki +space-create</code></td><td><span class="tag full">超过</span></td><td>公开真实 32 字符名称上限；创建后按 workspaceId 读回。</td></tr>
+      <tr><td><code>+delete-space</code></td><td><code>wiki +delete-space</code></td><td><span class="tag full">超过</span></td><td>预读目标、高风险确认、只接受 <code>success=true</code>；兼容 <code>+space-delete</code>。</td></tr>
+      <tr><td><code>+member-add</code></td><td><code>wiki +member-add</code></td><td><span class="tag partial">任务对齐</span></td><td>支持 1–30 个 userId 与四种角色；以写接口终态作为成功证据，不把最多 50 条的名单误作精确读回。</td></tr>
+      <tr><td><code>+member-list</code></td><td><code>wiki +member-list</code></td><td><span class="tag partial">任务对齐</span></td><td>严格成员数组、角色过滤、真实上限 50；后端无游标，不能提供诚实的 page-all。</td></tr>
+      <tr><td><code>+member-remove</code></td><td><code>wiki +member-remove</code></td><td><span class="tag partial">任务对齐</span></td><td>支持批量 userId；只接受写接口明确终态，并公开无法进行精确成员读回的边界。</td></tr>
+      <tr><td><code>+node-list</code></td><td><code>wiki +node-list</code></td><td><span class="tag full">完整对齐</span></td><td>正确跨域路由 doc/list_nodes，严格空目录、分页与自动翻页。</td></tr>
+      <tr><td><code>+node-get</code></td><td><code>wiki +node-get</code></td><td><span class="tag partial">任务对齐</span></td><td>支持 DingTalk 节点 ID/在线文档 URL 并返回文档域元数据；不接受跨平台专用的 token/type 组合。</td></tr>
+      <tr><td><code>+node-create</code></td><td><code>wiki +node-create</code></td><td><span class="tag partial">任务对齐</span></td><td>支持 adoc/axls/able/appt/adraw/amind/folder 并读回；无 origin/shortcut 等价接口。</td></tr>
+      <tr><td><code>+node-copy</code></td><td><code>wiki +node-copy</code></td><td><span class="tag full">超过</span></td><td>确认后要求新 nodeId 并读取副本；底层面向在线节点，不把 .dlink 当独立副本。</td></tr>
+      <tr><td><code>+move</code></td><td><code>wiki +move</code></td><td><span class="tag partial">任务对齐</span></td><td>同一入口支持 Wiki 内移动和“我的文档”在线节点入 Wiki，读回 workspace/folder；底层接口没有 apply 权限迁移开关。</td></tr>
+      <tr><td><code>+move-to-drive</code></td><td><code>wiki +move-to-drive</code></td><td><span class="tag full">超过</span></td><td>DWS 当前接口同步完成并读回 workspace 变化，无需暴露异步 task 轮询。</td></tr>
+      <tr><td><code>+node-delete</code></td><td><code>wiki +node-delete</code></td><td><span class="tag full">超过</span></td><td>预读并核对 workspace，高风险确认，要求删除终态。</td></tr>
+    </tbody></table></div>
+  </section>
+
+  <section class="card">
+    <h2>DWS 可挖掘的 7 个额外场景</h2>
+    <div class="two">
+      <div><h3>定位与创建链</h3><ul><li><code>+space-search</code>：严格关键词搜索。</li><li><code>+space-get</code>：空间详情与 workspaceId 证据。</li><li><code>+resolve-space</code>：唯一命中直出 ID，多命中拒绝猜测。</li><li><code>+wiki-new-doc</code>：空间名解析 → 创建 → 文档读回。</li></ul></div>
+      <div><h3>治理与巡检链</h3><ul><li><code>+member-update</code>：角色变更终态与不可精确读回声明。</li><li><code>+node-search</code>：库内关键词/扩展名搜索，严格零命中。</li><li><code>+feed-list</code>：知识库动态时间线与服务端 exclude-file 过滤。</li></ul></div>
+    </div>
+  </section>
+
+  <section class="card">
+    <h2>隐藏问题与修复</h2>
+    <table><thead><tr><th>原问题</th><th>错误风险</th><th>本次修复</th></tr></thead><tbody>
+      <tr><td>5 个旧 Wiki Shortcut 可直接执行，但只有 1 个进入公开目录。</td><td>Help、Schema、Skill 发现链与运行面漂移。</td><td><span class="tag fixed">20 项统一评审</span> 全部具备 Contract/Safety/Result 与语义目录记录。</td></tr>
+      <tr><td>列表投影找不到数组或遇到坏元素时返回空 slice。</td><td>把内部错误、字段漂移误报为“没有数据”。</td><td><span class="tag fixed">失败关闭</span> 只有响应中真实存在的 <code>[]</code> 才是合法空集合。</td></tr>
+      <tr><td>节点列表 Shortcut 调错 Wiki MCP 服务。</td><td>真实后端 <code>success=false</code>，Mock/静态检查看不出。</td><td><span class="tag fixed">跨域路由</span> 明确调用 doc/list_nodes，并纳入真实 E2E。</td></tr>
+      <tr><td>成员帮助宣称最大 200。</td><td>真实接口超过 50 直接参数错误。</td><td><span class="tag fixed">真实上限</span> Shortcut 与原子 Help 均改为 50，并在本地提前拒绝。</td></tr>
+      <tr><td>成员写操作从最多 50 条、不可分页的名单推断成员存在或缺失。</td><td>目标在截断部分时会误报写失败，或把未验证的移除报告为已读回。</td><td><span class="tag fixed">终态证据</span> 只接受写接口 <code>success=true</code>，并在结果中明确 <code>readbackAvailable=false</code>。</td></tr>
+      <tr><td>空间搜索的稳定工作流属性名与实际请求属性名不同。</td><td>直接改写已发布的 <code>query/limit</code> 会造成无版本 Schema 破坏；继续隐式转换又会让审计者误以为请求同名透传。</td><td><span class="tag fixed">显式复合适配</span> 最终 Schema 保留兼容属性并明确声明转换为 <code>keyword/pageSize</code>；回归测试同时锁定最终交付和精确请求参数。</td></tr>
+      <tr><td>知识库名称帮助宣称最大 100。</td><td>真实接口超过 32 失败。</td><td><span class="tag fixed">真实上限</span> Help 与 Shortcut 校验统一为 32。</td></tr>
+      <tr><td>复制/移动/创建只把无异常视为成功。</td><td>空确认、未知远端效果或移动未到目标仍可能被接受。</td><td><span class="tag fixed">读回证明</span> 在后端具备精确查询能力时检查 success、业务 ID、workspace/folder 等最终状态。</td></tr>
+    </tbody></table>
+  </section>
+
+  <section class="card">
+    <h2>真实数据 E2E 证据矩阵</h2>
+    <p class="small">28 项业务断言全部通过。测试使用一次性空知识库、临时在线文档与一名同组织内部测试成员；所有对象在 finally 清理。报告不保存对象 ID、成员身份、组织信息、URL、trace 或原始响应。</p>
+    <div class="toolbar"><input id="q" placeholder="筛选命令或证据"><select id="g"><option value="">全部分组</option><option>空间</option><option>成员</option><option>节点</option><option>动态</option></select></div>
+    <div class="matrix"><table id="catalog"><thead><tr><th>分组</th><th>Shortcut</th><th>实际业务断言</th><th>状态</th></tr></thead><tbody>
+      <tr><td>空间</td><td><code>+space-list</code></td><td>真实 count、hasMore、nextCursor；自动翻页返回两页结果。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>空间</td><td><code>+space-search</code></td><td>等待搜索索引后命中一次性 workspaceId。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>空间</td><td><code>+space-get</code></td><td>读回 workspaceId 与创建结果一致。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>空间</td><td><code>+resolve-space</code></td><td>唯一名称解析为同一 workspaceId。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>空间</td><td><code>+space-create</code></td><td>success=true、workspaceId 非空、详情读回一致。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>空间</td><td><code>+delete-space</code></td><td>目标预读、确认、success=true；兼容别名执行 finally 清理。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>成员</td><td><code>+member-list</code></td><td>真实 owner 条目与显式 members 数组，limit=50。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>成员</td><td><code>+member-add</code></td><td>命令只报告写终态；一次性小规模空间另行确认名单完整且角色为 READER。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>成员</td><td><code>+member-update</code></td><td>命令只报告写终态；一次性小规模空间另行确认角色变为 EDITOR。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>成员</td><td><code>+member-remove</code></td><td>命令只报告写终态；一次性小规模空间另行确认完整名单中不存在该 userId。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>节点</td><td><code>+node-list</code></td><td>空库返回真实 nodes:[]；有数据时验证游标与自动翻页。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>节点</td><td><code>+node-get</code></td><td>读回 nodeId 与请求一致。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>节点</td><td><code>+node-search</code></td><td>等待索引后按标题命中真实 nodeId。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>节点</td><td><code>+node-create</code></td><td>分别创建 folder/adoc，均取得 nodeId 和元数据读回。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>节点</td><td><code>+node-copy</code></td><td>取得不同的新 nodeId，副本元数据可读。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>节点</td><td><code>+move</code></td><td>读回 workspaceId 与 folderId 均等于目标；兼容 +node-move。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>节点</td><td><code>+move-to-drive</code></td><td>移动后读回 workspace 发生变化，再通过 +move 移回。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>节点</td><td><code>+node-delete</code></td><td>目标预读与 workspace 核对后收到 success=true。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>节点</td><td><code>+wiki-new-doc</code></td><td>按唯一空间名创建，nodeId 与文档详情读回一致。</td><td><span class="tag full">PASS</span></td></tr>
+      <tr><td>动态</td><td><code>+feed-list</code></td><td>创建/移动操作后返回真实 feeds 数组，缺字段不会被接受。</td><td><span class="tag full">PASS</span></td></tr>
+    </tbody></table></div>
+    <p class="small">可复跑入口：<code>make build</code> 后设置临时 <code>DWS_WIKI_E2E_MEMBER_ID</code>，在交互终端运行 <code>./scripts/dev/wiki-shortcut-e2e.py</code>。脚本只输出能力标签，不输出业务对象；受保护操作及最终清理均由命令逐项获取终端确认，非交互环境会在创建测试数据前拒绝运行。</p>
+  </section>
+
+  <section class="card">
+    <h2>成功判定与发布门</h2>
+    <div class="two"><div><h3>运行时证据层</h3><ol><li>传输/MCP 调用成功。</li><li>响应契约存在且类型正确。</li><li>写操作必须有 <code>success=true</code>；创建类操作还必须有业务 ID。</li><li>后端具备精确查询时必须读回；不具备时明确发布不可读回，而非从截断集合推断。</li><li>集合只有显式数组才允许为空。</li></ol></div><div><h3>交付门</h3><ol><li>20/20 语义目录与注册面精确覆盖。</li><li>Contract、Safety、Result、统一输出完整。</li><li>生成漂移、Schema、确认真值、全量 Go 测试。</li><li>独立真实数据 E2E 与 finally 清理。</li><li>diff PII/密钥/本地绝对路径扫描。</li></ol></div></div>
+  </section>
+  <p class="footer">DWS Wiki Shortcut business review · sanitized engineering artifact</p>
+</main>
+<script>
+  const q=document.querySelector('#q'),g=document.querySelector('#g'),rows=[...document.querySelectorAll('#catalog tbody tr')];
+  function filter(){const text=q.value.trim().toLowerCase(),group=g.value;rows.forEach(r=>{const okText=!text||r.textContent.toLowerCase().includes(text),okGroup=!group||r.children[0].textContent===group;r.style.display=okText&&okGroup?'':'none'})}q.addEventListener('input',filter);g.addEventListener('change',filter);
+</script></body></html>
@@ -2,6 +2,8 @@ module github.com/DingTalk-Real-AI/dingtalk-workspace-cli

 go 1.25.9

+replace gitlab.alibaba-inc.com/aes/aem-go-sdk => ./third_party/aem-go-sdk
+
 require (
 	github.com/Microsoft/go-winio v0.6.2
 	github.com/RealAlexandreAI/json-repair v0.0.15
@@ -14,9 +16,10 @@ require (
 	github.com/itchyny/gojq v0.12.18
 	github.com/mattn/go-isatty v0.0.20
 	github.com/muesli/termenv v0.16.0
-	github.com/open-dingtalk/dingtalk-stream-sdk-go v0.9.2-0.20260705041131-325e7c1049ad
+	github.com/open-dingtalk/dingtalk-stream-sdk-go v0.9.2-beta.1
 	github.com/spf13/cobra v1.10.2
 	github.com/zalando/go-keyring v0.2.8
+	gitlab.alibaba-inc.com/aes/aem-go-sdk v0.3.0
 	golang.org/x/crypto v0.49.0
 	golang.org/x/sys v0.42.0
 	golang.org/x/text v0.35.0
@@ -88,8 +88,8 @@ github.com/muesli/cancelreader v0.2.2 h1:3I4Kt4BQjOR54NavqnDogx/MIoWBFa0StPA8ELU
 github.com/muesli/cancelreader v0.2.2/go.mod h1:3XuTXfFS2VjM+HTLZY9Ak0l6eUKfijIfMUZ4EgX0QYo=
 github.com/muesli/termenv v0.16.0 h1:S5AlUN9dENB57rsbnkPyfdGuWIlkmzJjbFf0Tf5FWUc=
 github.com/muesli/termenv v0.16.0/go.mod h1:ZRfOIKPFDYQoDFF4Olj7/QJbW60Ol/kL1pU3VfY/Cnk=
-github.com/open-dingtalk/dingtalk-stream-sdk-go v0.9.2-0.20260705041131-325e7c1049ad h1:Bb4I+suYd+ehQ8e22aimLLze+5XTN3+WTc/x2LafmH8=
-github.com/open-dingtalk/dingtalk-stream-sdk-go v0.9.2-0.20260705041131-325e7c1049ad/go.mod h1:ln3IqPYYocZbYvl9TAOrG/cxGR9xcn4pnZRLdCTEGEU=
+github.com/open-dingtalk/dingtalk-stream-sdk-go v0.9.2-beta.1 h1:5WwR5TV6A12taXMH7SggT8yCMMJMF9jWE7Wj+4AuHck=
+github.com/open-dingtalk/dingtalk-stream-sdk-go v0.9.2-beta.1/go.mod h1:ln3IqPYYocZbYvl9TAOrG/cxGR9xcn4pnZRLdCTEGEU=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/rivo/uniseg v0.4.7 h1:WUdvkW8uEhrYfLC4ZzdpI2ztxP1I582+49Oc5Mq64VQ=
@@ -65,12 +65,16 @@ func TestCrossPlatformCoverageTokenManagerCachesUntilMarkerRevisionChanges(t *te
 	token := "token-a"
 	installTokenManagerFakes(t, func() (*authpkg.TokenData, error) {
 		calls.Add(1)
-		return &authpkg.TokenData{AccessToken: token, ExpiresAt: time.Now().Add(time.Hour)}, nil
+		return &authpkg.TokenData{
+			AccessToken: token,
+			ExpiresAt:   time.Now().Add(time.Hour),
+			LoginRegion: string(authpkg.LoginRegionInternational),
+		}, nil
 	})

 	manager := NewTokenManager()
 	first, err := manager.Get(context.Background(), configDir, "")
-	if err != nil || first.AccessToken != "token-a" {
+	if err != nil || first.AccessToken != "token-a" || first.LoginRegion != authpkg.LoginRegionInternational || !first.LoginRegionKnown {
 		t.Fatalf("first token = %#v, %v", first, err)
 	}
 	second, err := manager.Get(context.Background(), configDir, "")
@@ -41,9 +41,11 @@ type accessTokenSnapshotGetter interface {
 // AccessTokenSnapshot is the minimal bearer view needed by the process cache.
 // Refresh-token material never leaves the auth package.
 type AccessTokenSnapshot struct {
-	AccessToken string
-	ExpiresAt   time.Time
-	Source      string
+	AccessToken      string
+	ExpiresAt        time.Time
+	Source           string
+	LoginRegion      authpkg.LoginRegion
+	LoginRegionKnown bool
 }

 type tokenManagerKey struct {
@@ -223,9 +225,11 @@ func resolveAccessTokenSnapshotFromDir(ctx context.Context, configDir, profile s
 		data, err := snapshotProvider.GetTokenSnapshot(ctx)
 		if err == nil && data != nil && strings.TrimSpace(data.AccessToken) != "" {
 			return AccessTokenSnapshot{
-				AccessToken: strings.TrimSpace(data.AccessToken),
-				ExpiresAt:   data.ExpiresAt,
-				Source:      "oauth",
+				AccessToken:      strings.TrimSpace(data.AccessToken),
+				ExpiresAt:        data.ExpiresAt,
+				Source:           "oauth",
+				LoginRegion:      authpkg.LoginRegion(strings.TrimSpace(data.LoginRegion)),
+				LoginRegionKnown: true,
 			}, nil
 		}
 		if err != nil && !errors.Is(err, authpkg.ErrTokenDataNotFound) {
@@ -165,7 +165,7 @@ func TestResolveIdentityHeadersOmitsAbsentOrInvalidAgentHost(t *testing.T) {
 	}
 }

-func TestRootRejectsInvalidAgentHostBeforeEditionHook(t *testing.T) {
+func TestCrossPlatformCoverageRootRejectsInvalidAgentHostBeforeEditionHook(t *testing.T) {
 	t.Setenv("DWS_CONFIG_DIR", t.TempDir())
 	const invalidValue = "DO_NOT_ECHO"
 	t.Setenv(envDWSAgentHost, invalidValue)
@@ -0,0 +1,248 @@
+// Copyright 2026 Alibaba Group
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package app
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"os"
+	"regexp"
+	"strings"
+	"unicode"
+	"unicode/utf8"
+
+	apperrors "github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/errors"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/executor"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/transport"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/pkg/configmeta"
+)
+
+const (
+	envDWSAgentVersion     = "DWS_AGENT_VER"
+	envDWSAgentExt         = "DWS_AGENT_EXT"
+	maxAgentVersionBytes   = 64
+	maxAgentExtensionBytes = 8 * 1024
+)
+
+var agentVersionPattern = regexp.MustCompile(`^[A-Za-z0-9][A-Za-z0-9._+-]*$`)
+
+type agentMetadataSnapshot struct {
+	version    string
+	ext        string
+	versionErr error
+	extErr     error
+}
+
+type agentMetadataSnapshotContextKey struct{}
+
+func (snapshot agentMetadataSnapshot) validationError() error {
+	if snapshot.versionErr != nil {
+		return snapshot.versionErr
+	}
+	return snapshot.extErr
+}
+
+func contextWithAgentMetadataSnapshot(ctx context.Context, snapshot agentMetadataSnapshot) context.Context {
+	return context.WithValue(ctx, agentMetadataSnapshotContextKey{}, snapshot)
+}
+
+func agentMetadataSnapshotFromContext(ctx context.Context) (agentMetadataSnapshot, bool) {
+	if ctx == nil {
+		return agentMetadataSnapshot{}, false
+	}
+	snapshot, ok := ctx.Value(agentMetadataSnapshotContextKey{}).(agentMetadataSnapshot)
+	return snapshot, ok
+}
+
+func init() {
+	configmeta.Register(configmeta.ConfigItem{
+		Name:        envDWSAgentVersion,
+		Category:    configmeta.CategoryExternal,
+		Description: "调用 DWS 的 Agent 版本；仅作为 x-dws-agent-ver 透传到非插件 MCP 请求",
+		Example:     "1.2.3-beta.1+build.7",
+	})
+	configmeta.Register(configmeta.ConfigItem{
+		Name:        envDWSAgentExt,
+		Category:    configmeta.CategoryExternal,
+		Description: "调用 DWS 的 Agent 扩展上下文 JSON；仅作为 x-dws-agent-ext 透传到非插件 MCP 请求",
+		Example:     `{"umt":"<token>","miniwua":"<token>","ua":"agent/1.0"}`,
+		Sensitive:   true,
+	})
+}
+
+// parseAgentVersion normalizes and validates the caller-declared Agent
+// version. Only surrounding ASCII spaces and tabs are trimmed. An unset or
+// ASCII-whitespace-only value means "do not emit".
+func parseAgentVersion(raw string) (string, error) {
+	value := strings.Trim(raw, " \t")
+	if value == "" {
+		return "", nil
+	}
+	if len(value) > maxAgentVersionBytes || !agentVersionPattern.MatchString(value) {
+		return "", invalidAgentVersionError()
+	}
+	return value, nil
+}
+
+// parseAgentExt validates one generic JSON object and returns its compact
+// one-line representation. Raw control characters other than horizontal tab
+// are rejected before JSON parsing; escaped JSON control characters remain
+// valid because they are safe on the HTTP header wire.
+func parseAgentExt(raw string) (string, error) {
+	if len(raw) > maxAgentExtensionBytes || !utf8.ValidString(raw) {
+		return "", invalidAgentExtError()
+	}
+	for _, r := range raw {
+		if unicode.IsControl(r) && r != '\t' {
+			return "", invalidAgentExtError()
+		}
+	}
+
+	value := strings.Trim(raw, " \t")
+	if value == "" {
+		return "", nil
+	}
+
+	var compact bytes.Buffer
+	if err := json.Compact(&compact, []byte(value)); err != nil {
+		return "", invalidAgentExtError()
+	}
+	compactBytes := compact.Bytes()
+	if len(compactBytes) > maxAgentExtensionBytes || len(compactBytes) < 2 || compactBytes[0] != '{' {
+		return "", invalidAgentExtError()
+	}
+	return compact.String(), nil
+}
+
+func invalidAgentVersionError() error {
+	return apperrors.NewValidation(
+		"DWS_AGENT_VER must be at most 64 bytes and match ^[A-Za-z0-9][A-Za-z0-9._+-]*$",
+		apperrors.WithReason("invalid_agent_version"),
+	)
+}
+
+func invalidAgentExtError() error {
+	return apperrors.NewValidation(
+		"DWS_AGENT_EXT must be a UTF-8 JSON object of at most 8192 bytes without raw control characters",
+		apperrors.WithReason("invalid_agent_ext"),
+	)
+}
+
+// readAgentMetadataSnapshot reads both environment variables from one
+// os.Environ snapshot, then parses them once. Normal CLI execution retains the
+// validated result through the invocation so hooks and transport observe the
+// same pair even in an embedding process that mutates its environment.
+func readAgentMetadataSnapshot() agentMetadataSnapshot {
+	var rawVersion, rawExt string
+	for _, entry := range os.Environ() {
+		key, value, _ := strings.Cut(entry, "=")
+		switch key {
+		case envDWSAgentVersion:
+			rawVersion = value
+		case envDWSAgentExt:
+			rawExt = value
+		}
+	}
+	version, versionErr := parseAgentVersion(rawVersion)
+	ext, extErr := parseAgentExt(rawExt)
+	return agentMetadataSnapshot{
+		version:    version,
+		ext:        ext,
+		versionErr: versionErr,
+		extErr:     extErr,
+	}
+}
+
+// removeAgentMetadataHeaders removes every case variant so edition or
+// credential hooks cannot smuggle MCP-only metadata into shared transports.
+func removeAgentMetadataHeaders(headers map[string]string) {
+	for key := range headers {
+		if strings.EqualFold(key, transport.HeaderAgentVersion) ||
+			strings.EqualFold(key, transport.HeaderAgentExt) {
+			delete(headers, key)
+		}
+	}
+}
+
+// applyAgentMetadataHeaders applies validated environment values as the final
+// authority for non-plugin MCP requests. Invalid values are omitted on
+// library paths that bypass root validation; normal CLI execution rejects
+// them before hooks or network access.
+func applyAgentMetadataHeaders(headers map[string]string) map[string]string {
+	return applyAgentMetadataSnapshot(headers, readAgentMetadataSnapshot())
+}
+
+func applyAgentMetadataSnapshot(headers map[string]string, snapshot agentMetadataSnapshot) map[string]string {
+	removeAgentMetadataHeaders(headers)
+
+	if (snapshot.versionErr != nil || snapshot.version == "") && (snapshot.extErr != nil || snapshot.ext == "") {
+		return headers
+	}
+	if headers == nil {
+		headers = make(map[string]string)
+	}
+	if snapshot.versionErr == nil && snapshot.version != "" {
+		headers[transport.HeaderAgentVersion] = snapshot.version
+	}
+	if snapshot.extErr == nil && snapshot.ext != "" {
+		headers[transport.HeaderAgentExt] = snapshot.ext
+	}
+	return headers
+}
+
+// resolveMCPRequestHeaders adds Agent version and extension metadata only to
+// the built-in DingTalk MCP request path. Shared identity consumers (notably
+// A2A) continue to use resolveIdentityHeaders and never receive these fields.
+func resolveMCPRequestHeaders() map[string]string {
+	return resolveMCPRequestHeadersWithSnapshot(readAgentMetadataSnapshot())
+}
+
+func resolveMCPRequestHeadersWithSnapshot(snapshot agentMetadataSnapshot) map[string]string {
+	return applyAgentMetadataSnapshot(resolveIdentityHeaders(), snapshot)
+}
+
+// resolveMCPRequestHeadersForInvocation resolves one immutable Header snapshot
+// for an invocation. The helper-only mcp-meta server performs endpoint
+// discovery rather than an ordinary MCP product call, so caller-declared
+// Agent metadata must not cross that boundary.
+func resolveMCPRequestHeadersForInvocation(invocation executor.Invocation, snapshots ...agentMetadataSnapshot) map[string]string {
+	headers := resolveIdentityHeaders()
+	if strings.EqualFold(strings.TrimSpace(invocation.CanonicalProduct), mcpMetaServerID) {
+		return headers
+	}
+	snapshot := readAgentMetadataSnapshot()
+	if len(snapshots) > 0 {
+		snapshot = snapshots[0]
+	}
+	return applyAgentMetadataSnapshot(headers, snapshot)
+}
+
+// pluginRequestHeaders returns a private, sanitized copy of plugin-owned
+// Headers. Third-party plugins never receive DWS-owned Agent metadata, even if
+// their manifest tries to declare the reserved Header names itself.
+func pluginRequestHeaders(pluginAuth *PluginAuth) map[string]string {
+	if pluginAuth == nil || len(pluginAuth.ExtraHeaders) == 0 {
+		return nil
+	}
+	headers := make(map[string]string, len(pluginAuth.ExtraHeaders))
+	for key, value := range pluginAuth.ExtraHeaders {
+		headers[key] = value
+	}
+	removeAgentMetadataHeaders(headers)
+	if len(headers) == 0 {
+		return nil
+	}
+	return headers
+}
@@ -0,0 +1,743 @@
+// Copyright 2026 Alibaba Group
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package app
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"io"
+	"maps"
+	"os"
+	"strings"
+	"testing"
+
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/audit"
+	apperrors "github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/errors"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/executor"
+	outputpkg "github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/output"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/pipeline"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/testseam"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/transport"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/pkg/agentproduct"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/pkg/configmeta"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/pkg/edition"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/pkg/mcptypes"
+	"github.com/spf13/cobra"
+)
+
+func TestCrossPlatformCoverageParseAgentVersion(t *testing.T) {
+	var nilContext context.Context
+	if _, ok := agentMetadataSnapshotFromContext(nilContext); ok {
+		t.Fatal("nil context unexpectedly contained Agent metadata")
+	}
+	wantSnapshot := agentMetadataSnapshot{version: "context-version", ext: "{}"}
+	if got, ok := agentMetadataSnapshotFromContext(contextWithAgentMetadataSnapshot(context.Background(), wantSnapshot)); !ok || got != wantSnapshot {
+		t.Fatalf("context Agent metadata = %#v, %v; want %#v", got, ok, wantSnapshot)
+	}
+
+	valid := []struct {
+		name string
+		raw  string
+		want string
+	}{
+		{name: "unset", raw: "", want: ""},
+		{name: "ASCII whitespace only", raw: " \t ", want: ""},
+		{name: "semantic version", raw: "1.2.3", want: "1.2.3"},
+		{name: "pre-release and build", raw: " v1.2.3-rc.1+build_7 ", want: "v1.2.3-rc.1+build_7"},
+		{name: "maximum length", raw: strings.Repeat("a", maxAgentVersionBytes), want: strings.Repeat("a", maxAgentVersionBytes)},
+	}
+	for _, tc := range valid {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := parseAgentVersion(tc.raw)
+			if err != nil {
+				t.Fatalf("parseAgentVersion() error = %v", err)
+			}
+			if got != tc.want {
+				t.Fatalf("parseAgentVersion() = %q, want %q", got, tc.want)
+			}
+		})
+	}
+
+	invalid := []struct {
+		name string
+		raw  string
+	}{
+		{name: "leading punctuation", raw: "-1.2.3"},
+		{name: "internal space", raw: "1.2 3"},
+		{name: "slash", raw: "1.2/3"},
+		{name: "line feed", raw: "1.2.3\n"},
+		{name: "carriage return", raw: "1.2.3\r"},
+		{name: "NUL", raw: "1.2\x003"},
+		{name: "Unicode", raw: "版本1"},
+		{name: "too long", raw: strings.Repeat("a", maxAgentVersionBytes+1)},
+	}
+	for _, tc := range invalid {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := parseAgentVersion(tc.raw)
+			if err == nil || got != "" {
+				t.Fatalf("parseAgentVersion(%q) = %q, %v; want validation error", tc.raw, got, err)
+			}
+			assertAgentMetadataValidationError(t, err, "invalid_agent_version", tc.raw)
+		})
+	}
+}
+
+func TestCrossPlatformCoverageParseAgentExt(t *testing.T) {
+	boundary := `{"x":"` + strings.Repeat("a", maxAgentExtensionBytes-8) + `"}`
+	if len(boundary) != maxAgentExtensionBytes {
+		t.Fatalf("invalid boundary fixture size: %d", len(boundary))
+	}
+
+	valid := []struct {
+		name string
+		raw  string
+		want string
+	}{
+		{name: "unset", raw: "", want: ""},
+		{name: "ASCII whitespace only", raw: " \t ", want: ""},
+		{name: "empty object", raw: "{}", want: "{}"},
+		{name: "compact generic object", raw: " \t{ \"umt\": \"masked\",\t \"nested\": { \"ok\": true }, \"unknown\": [1, 2] }\t ", want: `{"umt":"masked","nested":{"ok":true},"unknown":[1,2]}`},
+		{name: "Unicode value", raw: `{"ua":"千问办公/1.0"}`, want: `{"ua":"千问办公/1.0"}`},
+		{name: "escaped control remains safe", raw: `{"ua":"line\nnext"}`, want: `{"ua":"line\nnext"}`},
+		{name: "maximum length", raw: boundary, want: boundary},
+	}
+	for _, tc := range valid {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := parseAgentExt(tc.raw)
+			if err != nil {
+				t.Fatalf("parseAgentExt() error = %v", err)
+			}
+			if got != tc.want {
+				t.Fatalf("parseAgentExt() = %q, want %q", got, tc.want)
+			}
+		})
+	}
+
+	invalidUTF8 := string([]byte{'{', '"', 'x', '"', ':', '"', 0xff, '"', '}'})
+	invalid := []struct {
+		name string
+		raw  string
+	}{
+		{name: "too long raw input", raw: strings.Repeat(" ", maxAgentExtensionBytes+1)},
+		{name: "invalid UTF-8", raw: invalidUTF8},
+		{name: "array", raw: `[]`},
+		{name: "string", raw: `"value"`},
+		{name: "number", raw: `1`},
+		{name: "boolean", raw: `true`},
+		{name: "null", raw: `null`},
+		{name: "malformed object", raw: `{"secret":"DO_NOT_ECHO"`},
+		{name: "trailing value", raw: `{} {}`},
+		{name: "line feed", raw: "{\n}"},
+		{name: "carriage return", raw: "{\r}"},
+		{name: "NUL", raw: "{\x00}"},
+		{name: "vertical tab", raw: "{\v}"},
+		{name: "form feed", raw: "{\f}"},
+		{name: "DEL", raw: "{\x7f}"},
+		{name: "C1 control", raw: "{\u0085}"},
+	}
+	for _, tc := range invalid {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := parseAgentExt(tc.raw)
+			if err == nil || got != "" {
+				t.Fatalf("parseAgentExt() = %q, %v; want validation error", got, err)
+			}
+			assertAgentMetadataValidationError(t, err, "invalid_agent_ext", tc.raw)
+		})
+	}
+}
+
+func assertAgentMetadataValidationError(t *testing.T, err error, reason, raw string) {
+	t.Helper()
+	var appErr *apperrors.Error
+	if !errors.As(err, &appErr) {
+		t.Fatalf("error type = %T, want *errors.Error", err)
+	}
+	if appErr.Category != apperrors.CategoryValidation || appErr.Reason != reason {
+		t.Fatalf("error = category %q reason %q, want validation/%s", appErr.Category, appErr.Reason, reason)
+	}
+	if strings.Contains(raw, "DO_NOT_ECHO") && strings.Contains(err.Error(), "DO_NOT_ECHO") {
+		t.Fatalf("error must not echo invalid value: %v", err)
+	}
+}
+
+func TestCrossPlatformCoverageAgentMetadataConfigRegistrationAndMasking(t *testing.T) {
+	items := configmeta.All()
+	var versionItem, extItem *configmeta.ConfigItem
+	for i := range items {
+		switch items[i].Name {
+		case envDWSAgentVersion:
+			versionItem = &items[i]
+		case envDWSAgentExt:
+			extItem = &items[i]
+		}
+	}
+	if versionItem == nil || extItem == nil {
+		t.Fatalf("Agent metadata config registration missing: version=%v ext=%v", versionItem != nil, extItem != nil)
+	}
+	if versionItem.Category != configmeta.CategoryExternal || versionItem.Sensitive {
+		t.Fatalf("version config metadata = %#v", *versionItem)
+	}
+	if extItem.Category != configmeta.CategoryExternal || !extItem.Sensitive {
+		t.Fatalf("extension config metadata = %#v", *extItem)
+	}
+
+	const canary = `{"umt":"SENSITIVE_CANARY"}`
+	t.Setenv(envDWSAgentExt, canary)
+	got, ok := configmeta.Resolve(envDWSAgentExt)
+	if !ok || got == "" || strings.Contains(got, "SENSITIVE_CANARY") || got == canary {
+		t.Fatalf("sensitive extension was not masked: value=%q ok=%v", got, ok)
+	}
+
+	t.Setenv(envDWSAgentVersion, "9.8.7")
+	command := newConfigListCommand()
+	var output strings.Builder
+	command.SetOut(&output)
+	command.SetArgs([]string{"--category", string(configmeta.CategoryExternal), "--show-values", "--json"})
+	if err := command.Execute(); err != nil {
+		t.Fatalf("config list failed: %v", err)
+	}
+	rawOutput := output.String()
+	if !json.Valid([]byte(rawOutput)) {
+		t.Fatalf("config list emitted invalid JSON: %q", rawOutput)
+	}
+	if !strings.Contains(rawOutput, envDWSAgentVersion) || !strings.Contains(rawOutput, envDWSAgentExt) {
+		t.Fatalf("config list omitted Agent metadata variables: %s", rawOutput)
+	}
+	if strings.Contains(rawOutput, "SENSITIVE_CANARY") || strings.Contains(rawOutput, canary) {
+		t.Fatalf("config list leaked Agent extension: %s", rawOutput)
+	}
+}
+
+func TestCrossPlatformCoverageResolveMCPRequestHeadersScopesAndFinalizesAgentMetadata(t *testing.T) {
+	t.Setenv("DWS_CONFIG_DIR", t.TempDir())
+	t.Setenv(envDWSAgentHost, "")
+	t.Setenv(agentproduct.EnvName, "")
+	t.Setenv(envDWSAgentVersion, " 1.2.3-rc.1 ")
+	t.Setenv(envDWSAgentExt, " { \"umt\": \"masked\", \"unknown\": true } ")
+
+	oldEdition := edition.Get()
+	t.Cleanup(func() { edition.Override(oldEdition) })
+	edition.Override(&edition.Hooks{
+		MergeHeaders: func(headers map[string]string) map[string]string {
+			headers["X-Dws-Agent-Ver"] = "merge-must-not-win"
+			headers["X-Dws-Agent-Ext"] = `{"source":"merge"}`
+			return headers
+		},
+		EnterpriseCredentialHeaders: func(headers map[string]string) map[string]string {
+			headers[transport.HeaderAgentVersion] = "credential-must-not-win"
+			headers[transport.HeaderAgentExt] = `{"source":"credential"}`
+			return headers
+		},
+	})
+
+	for name, headers := range map[string]map[string]string{
+		"shared identity": resolveIdentityHeaders(),
+		"A2A export":      MCPIdentityHeaders(),
+	} {
+		if hasHeaderFold(headers, transport.HeaderAgentVersion) || hasHeaderFold(headers, transport.HeaderAgentExt) {
+			t.Fatalf("%s leaked MCP-only metadata: %#v", name, headers)
+		}
+	}
+
+	headers := resolveMCPRequestHeaders()
+	if got := headers[transport.HeaderAgentVersion]; got != "1.2.3-rc.1" {
+		t.Fatalf("%s = %q, want 1.2.3-rc.1", transport.HeaderAgentVersion, got)
+	}
+	if got := headers[transport.HeaderAgentExt]; got != `{"umt":"masked","unknown":true}` {
+		t.Fatalf("%s = %q", transport.HeaderAgentExt, got)
+	}
+	if got := headers[transport.HeaderVersion]; got != version {
+		t.Fatalf("%s = %q, want CLI version %q", transport.HeaderVersion, got, version)
+	}
+	if _, ok := headers["User-Agent"]; ok {
+		t.Fatal("Agent extension must not create or replace the standard User-Agent header")
+	}
+	for _, key := range []string{"umt", "miniwua", "ua", "x-dws-agent-umt", "x-dws-agent-miniwua", "x-dws-agent-ua"} {
+		if hasHeaderFold(headers, key) {
+			t.Fatalf("Agent extension was split into an extra header %q: %#v", key, headers)
+		}
+	}
+
+	// Library paths are best-effort: one invalid value is omitted without
+	// suppressing the other valid field or preserving hook-injected values.
+	t.Setenv(envDWSAgentExt, `{"secret":"DO_NOT_ECHO"`)
+	headers = resolveMCPRequestHeaders()
+	if got := headers[transport.HeaderAgentVersion]; got != "1.2.3-rc.1" {
+		t.Fatalf("valid version was suppressed: %q", got)
+	}
+	if hasHeaderFold(headers, transport.HeaderAgentExt) {
+		t.Fatalf("invalid extension or hook value leaked: %#v", headers)
+	}
+
+	// Exercise the nil-map and empty-input library paths. An absent environment
+	// must not allocate a map, while an EXT-only value must allocate one and
+	// remain a single compact Header.
+	t.Setenv(envDWSAgentVersion, "")
+	t.Setenv(envDWSAgentExt, "")
+	if got := applyAgentMetadataHeaders(nil); got != nil {
+		t.Fatalf("empty metadata allocated headers: %#v", got)
+	}
+	t.Setenv(envDWSAgentExt, " { } ")
+	headers = applyAgentMetadataHeaders(nil)
+	if got := headers[transport.HeaderAgentExt]; got != "{}" {
+		t.Fatalf("EXT-only metadata = %q, want {}", got)
+	}
+}
+
+func TestCrossPlatformCoverageRootRejectsInvalidAgentMetadataBeforeEditionHook(t *testing.T) {
+	oldEdition := edition.Get()
+	t.Cleanup(func() { edition.Override(oldEdition) })
+
+	tests := []struct {
+		name   string
+		env    string
+		value  string
+		reason string
+	}{
+		{name: "version", env: envDWSAgentVersion, value: "DO_NOT ECHO", reason: "invalid_agent_version"},
+		{name: "extension", env: envDWSAgentExt, value: `{"secret":"DO_NOT_ECHO"`, reason: "invalid_agent_ext"},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			t.Setenv("DWS_CONFIG_DIR", t.TempDir())
+			t.Setenv(envDWSAgentHost, "")
+			t.Setenv(agentproduct.EnvName, "")
+			t.Setenv(envDWSAgentVersion, "")
+			t.Setenv(envDWSAgentExt, "")
+			t.Setenv(tc.env, tc.value)
+
+			headerHookCalled := false
+			afterHookCalled := false
+			edition.Override(&edition.Hooks{
+				MergeHeaders: func(headers map[string]string) map[string]string {
+					headerHookCalled = true
+					return headers
+				},
+				EnterpriseCredentialHeaders: func(headers map[string]string) map[string]string {
+					headerHookCalled = true
+					return headers
+				},
+				AfterPersistentPreRun: func(_ *cobra.Command, _ []string) error {
+					afterHookCalled = true
+					return nil
+				},
+			})
+
+			root := NewRootCommand()
+			root.SetOut(io.Discard)
+			root.SetErr(io.Discard)
+			root.SetArgs([]string{"version"})
+			err := root.Execute()
+			if err == nil {
+				t.Fatalf("root command accepted invalid %s", tc.env)
+			}
+			if headerHookCalled || afterHookCalled {
+				t.Fatalf("edition hook ran before %s validation", tc.env)
+			}
+			assertAgentMetadataValidationError(t, err, tc.reason, tc.value)
+		})
+	}
+}
+
+func TestCrossPlatformCoverageAgentMetadataProcessEntryValidationPrecedesRootConstruction(t *testing.T) {
+	for _, tc := range []struct {
+		name string
+		args []string
+		want bool
+	}{
+		{name: "default JSON", args: []string{"version"}, want: true},
+		{name: "long JSON", args: []string{"version", "--format", "JSON"}, want: true},
+		{name: "long table", args: []string{"--format=table", "version"}, want: false},
+		{name: "short attached JSON", args: []string{"version", "-fjson"}, want: true},
+		{name: "short table", args: []string{"version", "-f", "table"}, want: false},
+		{name: "last wins", args: []string{"--format", "table", "version", "-f=json"}, want: true},
+		{name: "terminator", args: []string{"version", "--format", "table", "--", "--format", "json"}, want: false},
+		{name: "missing value", args: []string{"version", "--format"}, want: false},
+	} {
+		t.Run("presentation/"+tc.name, func(t *testing.T) {
+			if got := processArgsRequestJSON(tc.args); got != tc.want {
+				t.Fatalf("processArgsRequestJSON(%q) = %v, want %v", tc.args, got, tc.want)
+			}
+		})
+	}
+
+	t.Setenv("DWS_CONFIG_DIR", t.TempDir())
+	t.Setenv(envDWSAgentHost, "")
+	t.Setenv(agentproduct.EnvName, "")
+	t.Setenv(envDWSAgentVersion, "")
+	sensitiveRaw := "{\"umt\":\"must-not-leak\"}\n"
+	t.Setenv(envDWSAgentExt, sensitiveRaw)
+	oldEdition := edition.Get()
+	t.Cleanup(func() { edition.Override(oldEdition) })
+	extensionHookCalls := 0
+	edition.Override(&edition.Hooks{
+		Name: "presentation-test",
+		RegisterExtraCommands: func(*cobra.Command, edition.ToolCaller) {
+			extensionHookCalls++
+		},
+		VisibleProducts: func() []string {
+			extensionHookCalls++
+			return nil
+		},
+		StaticServers: func() []edition.ServerInfo {
+			extensionHookCalls++
+			return nil
+		},
+	})
+
+	oldArgs := os.Args
+	os.Args = []string{"dws", "version"}
+	t.Cleanup(func() { os.Args = oldArgs })
+
+	rootConstructed := false
+	preParseCalled := false
+	testseam.Swap(t, &rootNewRootCommandWithEngine, func(context.Context, *pipeline.Engine) *cobra.Command {
+		rootConstructed = true
+		return &cobra.Command{Use: "dws"}
+	})
+	testseam.Swap(t, &rootRunPreParse, func(*cobra.Command, *pipeline.Engine) error {
+		preParseCalled = true
+		return nil
+	})
+
+	stderrFile, err := os.CreateTemp(t.TempDir(), "agent-metadata-stderr-*")
+	if err != nil {
+		t.Fatalf("create stderr capture: %v", err)
+	}
+	oldStderr := os.Stderr
+	os.Stderr = stderrFile
+	t.Cleanup(func() {
+		os.Stderr = oldStderr
+		_ = stderrFile.Close()
+	})
+
+	if code := Execute(); code == 0 {
+		t.Fatal("process entry accepted invalid Agent metadata")
+	}
+	if rootConstructed || preParseCalled {
+		t.Fatalf("invalid Agent metadata reached root hooks: constructed=%v preParse=%v", rootConstructed, preParseCalled)
+	}
+	if extensionHookCalls != 0 {
+		t.Fatalf("invalid Agent metadata executed %d extension hooks", extensionHookCalls)
+	}
+	if err := stderrFile.Sync(); err != nil {
+		t.Fatalf("sync stderr capture: %v", err)
+	}
+	stderrOutput, err := os.ReadFile(stderrFile.Name())
+	if err != nil {
+		t.Fatalf("read stderr capture: %v", err)
+	}
+	if strings.Contains(string(stderrOutput), "must-not-leak") || strings.Contains(string(stderrOutput), sensitiveRaw) {
+		t.Fatalf("process validation error leaked raw EXT: %q", stderrOutput)
+	}
+	if !json.Valid(stderrOutput) || !strings.Contains(string(stderrOutput), `"reason": "invalid_agent_ext"`) {
+		t.Fatalf("default JSON error presentation = %q", stderrOutput)
+	}
+
+	stdoutFile, err := os.CreateTemp(t.TempDir(), "agent-metadata-stdout-*")
+	if err != nil {
+		t.Fatalf("create stdout capture: %v", err)
+	}
+	oldStdout := os.Stdout
+	os.Stdout = stdoutFile
+	t.Cleanup(func() {
+		os.Stdout = oldStdout
+		_ = stdoutFile.Close()
+	})
+	emitEarlyAgentMetadataValidationError(invalidAgentExtError(), []string{"drive", "+list", "--format", "json"})
+	if err := stdoutFile.Sync(); err != nil {
+		t.Fatalf("sync stdout capture: %v", err)
+	}
+	unifiedOutput, err := os.ReadFile(stdoutFile.Name())
+	if err != nil {
+		t.Fatalf("read stdout capture: %v", err)
+	}
+	if !json.Valid(unifiedOutput) || !strings.Contains(string(unifiedOutput), `"outcome": "failure"`) ||
+		!strings.Contains(string(unifiedOutput), `"subtype": "invalid_agent_ext"`) {
+		t.Fatalf("unified JSON error presentation = %q", unifiedOutput)
+	}
+	if extensionHookCalls != 0 {
+		t.Fatalf("presentation-only root executed %d extension hooks", extensionHookCalls)
+	}
+
+	if err := stderrFile.Truncate(0); err != nil {
+		t.Fatalf("truncate fallback stderr capture: %v", err)
+	}
+	if _, err := stderrFile.Seek(0, io.SeekStart); err != nil {
+		t.Fatalf("rewind fallback stderr capture: %v", err)
+	}
+	testseam.Swap(t, &rootEmitResult, func(*cobra.Command, outputpkg.CommandResult) (int, error) {
+		return 0, errors.New("injected result emission failure")
+	})
+	emitEarlyAgentMetadataValidationError(invalidAgentExtError(), []string{"drive", "+list", "--format", "json"})
+	if err := stderrFile.Sync(); err != nil {
+		t.Fatalf("sync fallback stderr capture: %v", err)
+	}
+	fallbackOutput, err := os.ReadFile(stderrFile.Name())
+	if err != nil {
+		t.Fatalf("read fallback stderr capture: %v", err)
+	}
+	if !json.Valid(fallbackOutput) || !strings.Contains(string(fallbackOutput), `"reason": "invalid_agent_ext"`) ||
+		strings.Contains(string(fallbackOutput), "must-not-leak") {
+		t.Fatalf("fallback validation error presentation = %q", fallbackOutput)
+	}
+
+	if err := stderrFile.Truncate(0); err != nil {
+		t.Fatalf("truncate stderr capture: %v", err)
+	}
+	if _, err := stderrFile.Seek(0, io.SeekStart); err != nil {
+		t.Fatalf("rewind stderr capture: %v", err)
+	}
+	emitEarlyAgentMetadataValidationError(invalidAgentExtError(), []string{"version", "--format", "table"})
+	if err := stderrFile.Sync(); err != nil {
+		t.Fatalf("sync human stderr capture: %v", err)
+	}
+	humanOutput, err := os.ReadFile(stderrFile.Name())
+	if err != nil {
+		t.Fatalf("read human stderr capture: %v", err)
+	}
+	if json.Valid(humanOutput) || !strings.Contains(string(humanOutput), "DWS_AGENT_EXT") ||
+		strings.Contains(string(humanOutput), "must-not-leak") {
+		t.Fatalf("human validation error presentation = %q", humanOutput)
+	}
+
+	var capturedRunner *runtimeRunner
+	testseam.Swap(t, &rootNewCommandRunnerWithFlags, func(flags *GlobalFlags) executor.Runner {
+		capturedRunner = newCommandRunnerWithFlags(flags).(*runtimeRunner)
+		return capturedRunner
+	})
+	cachedSnapshot := agentMetadataSnapshot{version: "9.8.7", ext: `{"ua":"cached"}`}
+	_ = newRootCommandWithMode(
+		contextWithAgentMetadataSnapshot(context.Background(), cachedSnapshot),
+		nil,
+		false,
+		true,
+		true,
+	)
+	if capturedRunner == nil || capturedRunner.agentMetadata == nil || *capturedRunner.agentMetadata != cachedSnapshot {
+		t.Fatalf("root runner Agent metadata = %#v, want %#v", capturedRunner, cachedSnapshot)
+	}
+}
+
+func TestCrossPlatformCoverageAgentMetadataExcludedFromServiceDiscovery(t *testing.T) {
+	t.Setenv("DWS_CONFIG_DIR", t.TempDir())
+	t.Setenv(envDWSAgentHost, "")
+	t.Setenv(agentproduct.EnvName, "")
+	t.Setenv(envDWSAgentVersion, "3.0.0")
+	t.Setenv(envDWSAgentExt, `{"umt":"test-value"}`)
+
+	headers := resolveMCPRequestHeadersForInvocation(executor.Invocation{
+		CanonicalProduct: mcpMetaServerID,
+		Tool:             mcpMetaURLTool,
+	})
+	if hasHeaderFold(headers, transport.HeaderAgentVersion) || hasHeaderFold(headers, transport.HeaderAgentExt) {
+		t.Fatalf("service-discovery request leaked Agent metadata: %#v", headers)
+	}
+
+	headers = resolveMCPRequestHeadersForInvocation(executor.Invocation{CanonicalProduct: "doc", Tool: "read"})
+	if headers[transport.HeaderAgentVersion] != "3.0.0" || headers[transport.HeaderAgentExt] == "" {
+		t.Fatalf("ordinary MCP request omitted Agent metadata: %#v", headers)
+	}
+	cached := agentMetadataSnapshot{version: "3.1.0", ext: "{}"}
+	headers = resolveMCPRequestHeadersForInvocation(executor.Invocation{CanonicalProduct: "doc", Tool: "read"}, cached)
+	if headers[transport.HeaderAgentVersion] != "3.1.0" || headers[transport.HeaderAgentExt] != "{}" {
+		t.Fatalf("ordinary MCP request ignored its validated snapshot: %#v", headers)
+	}
+}
+
+func TestCrossPlatformCoverageAgentMetadataMCPAndPluginScoping(t *testing.T) {
+	t.Setenv("DWS_CONFIG_DIR", t.TempDir())
+	t.Setenv(envDWSAgentHost, "")
+	t.Setenv(agentproduct.EnvName, "")
+	t.Setenv(envDWSAgentVersion, "2.0.0")
+	t.Setenv(envDWSAgentExt, `{"ua":"test-agent/2.0"}`)
+
+	oldEdition := edition.Get()
+	t.Cleanup(func() { edition.Override(oldEdition) })
+	edition.Override(&edition.Hooks{})
+
+	pluginAuthMu.Lock()
+	oldPluginRegistry := pluginAuthRegistry
+	pluginAuthRegistry = make(map[string]*PluginAuth)
+	pluginAuthMu.Unlock()
+	t.Cleanup(func() {
+		pluginAuthMu.Lock()
+		pluginAuthRegistry = oldPluginRegistry
+		pluginAuthMu.Unlock()
+	})
+	dynamicMu.Lock()
+	oldDynamicEndpoints := dynamicEndpoints
+	oldDynamicProducts := dynamicProducts
+	oldDynamicAliases := dynamicAliases
+	oldDynamicToolEndpoints := dynamicToolEndpoints
+	dynamicEndpoints = nil
+	dynamicProducts = nil
+	dynamicAliases = nil
+	dynamicToolEndpoints = nil
+	dynamicMu.Unlock()
+	t.Cleanup(func() {
+		dynamicMu.Lock()
+		dynamicEndpoints = oldDynamicEndpoints
+		dynamicProducts = oldDynamicProducts
+		dynamicAliases = oldDynamicAliases
+		dynamicToolEndpoints = oldDynamicToolEndpoints
+		dynamicMu.Unlock()
+	})
+
+	testseam.Swap(t, &runnerPreflightDocDownload, func(*runtimeRunner, context.Context, *transport.Client, string, executor.Invocation) error {
+		return nil
+	})
+	type capturedRequest struct {
+		headers map[string]string
+		token   string
+	}
+	var captured []capturedRequest
+	testseam.Swap(t, &runnerCallTool, func(client *transport.Client, _ context.Context, _, _ string, _ map[string]any) (transport.ToolCallResult, error) {
+		copyHeaders := make(map[string]string, len(client.ExtraHeaders))
+		for key, value := range client.ExtraHeaders {
+			copyHeaders[key] = value
+		}
+		captured = append(captured, capturedRequest{headers: copyHeaders, token: client.AuthToken})
+		return transport.ToolCallResult{Content: map[string]any{"value": "ok"}}, nil
+	})
+
+	created := newCommandRunnerWithFlags(&GlobalFlags{}).(*runtimeRunner)
+	if hasHeaderFold(created.transport.ExtraHeaders, transport.HeaderAgentVersion) ||
+		hasHeaderFold(created.transport.ExtraHeaders, transport.HeaderAgentExt) {
+		t.Fatalf("new runner resolved Agent metadata before invocation validation: %#v", created.transport.ExtraHeaders)
+	}
+
+	// runSingle must not cache ambient MCP metadata on the shared base transport.
+	// Use mock mode to exercise the path without authentication or network I/O.
+	t.Setenv(envDWSAgentVersion, "2.0.1")
+	refreshRunner := &runtimeRunner{
+		transport:   transport.NewClient(nil),
+		globalFlags: &GlobalFlags{Mock: true},
+		auditSink:   audit.NopSink{},
+	}
+	refreshInvocation := executor.Invocation{CanonicalProduct: "refresh", Tool: "tool", Params: map[string]any{}}
+	if _, err := refreshRunner.runSingle(context.Background(), refreshInvocation, false); err != nil {
+		t.Fatalf("mock runSingle failed: %v", err)
+	}
+	if hasHeaderFold(refreshRunner.transport.ExtraHeaders, transport.HeaderAgentVersion) {
+		t.Fatalf("runSingle mutated the shared transport Header map: %#v", refreshRunner.transport.ExtraHeaders)
+	}
+	t.Setenv(envDWSAgentVersion, "2.0.0")
+
+	r := &runtimeRunner{
+		transport:   transport.NewClient(nil),
+		globalFlags: &GlobalFlags{Token: "test-token"},
+		auditSink:   audit.NopSink{},
+		agentMetadata: &agentMetadataSnapshot{
+			version: "2.0.0",
+			ext:     `{"ua":"test-agent/2.0"}`,
+		},
+	}
+	builtIn := executor.Invocation{CanonicalProduct: "built-in", Tool: "tool", Params: map[string]any{}}
+	if _, err := r.executeInvocation(context.Background(), "https://example.test", builtIn); err != nil {
+		t.Fatalf("built-in invocation failed: %v", err)
+	}
+
+	pluginDescriptor := mcptypes.ServerDescriptor{
+		Key:      "third-party",
+		Endpoint: "https://plugin.example.test",
+		CLI:      mcptypes.CLIOverlay{ID: "third-party"},
+		AuthHeaders: map[string]string{
+			"X-Plugin":        "yes",
+			"X-Dws-Agent-Ver": "plugin-must-not-forge-version",
+			"X-Dws-Agent-Ext": `{"source":"plugin"}`,
+		},
+	}
+	registerPluginHTTPServer(pluginDescriptor)
+	registeredPlugin, pluginOwned := LookupPluginAuth("third-party")
+	if !pluginOwned || registeredPlugin == nil || registeredPlugin.Token != "" {
+		t.Fatalf("anonymous HTTP plugin ownership = %#v, %v", registeredPlugin, pluginOwned)
+	}
+	registerPluginHTTPServer(mcptypes.ServerDescriptor{
+		Key:      "anonymous-empty",
+		Endpoint: "https://anonymous.example.test",
+		CLI:      mcptypes.CLIOverlay{ID: "anonymous-empty"},
+	})
+	if emptyPlugin, owned := LookupPluginAuth("anonymous-empty"); !owned || emptyPlugin == nil || emptyPlugin.Token != "" || len(emptyPlugin.ExtraHeaders) != 0 {
+		t.Fatalf("headerless HTTP plugin ownership = %#v, %v", emptyPlugin, owned)
+	}
+	originalPluginHeaders := maps.Clone(registeredPlugin.ExtraHeaders)
+	pluginInvocation := executor.Invocation{CanonicalProduct: "third-party", Tool: "tool", Params: map[string]any{}}
+	if _, err := r.executeInvocation(context.Background(), "https://plugin.example.test", pluginInvocation); err != nil {
+		t.Fatalf("plugin invocation failed: %v", err)
+	}
+
+	if len(captured) != 2 {
+		t.Fatalf("captured %d calls, want 2", len(captured))
+	}
+	if captured[0].headers[transport.HeaderAgentVersion] != "2.0.0" || captured[0].headers[transport.HeaderAgentExt] != `{"ua":"test-agent/2.0"}` {
+		t.Fatalf("built-in MCP metadata = %#v", captured[0].headers)
+	}
+	if hasHeaderFold(captured[1].headers, transport.HeaderAgentVersion) || hasHeaderFold(captured[1].headers, transport.HeaderAgentExt) {
+		t.Fatalf("plugin request leaked Agent metadata: %#v", captured[1].headers)
+	}
+	if got := captured[1].headers["X-Plugin"]; got != "yes" {
+		t.Fatalf("plugin-owned header = %q, want yes", got)
+	}
+	if captured[1].token != "" {
+		t.Fatalf("anonymous plugin unexpectedly received default OAuth token")
+	}
+	if !maps.Equal(registeredPlugin.ExtraHeaders, originalPluginHeaders) {
+		t.Fatalf("plugin Header sanitization mutated registry state: got %#v want %#v", registeredPlugin.ExtraHeaders, originalPluginHeaders)
+	}
+	if got := pluginRequestHeaders(nil); got != nil {
+		t.Fatalf("nil plugin auth produced Headers: %#v", got)
+	}
+	if got := pluginRequestHeaders(&PluginAuth{ExtraHeaders: map[string]string{
+		"X-DWS-AGENT-VER": "forged",
+		"X-DWS-AGENT-EXT": `{"forged":true}`,
+	}}); got != nil {
+		t.Fatalf("reserved-only plugin Headers survived sanitization: %#v", got)
+	}
+
+	// Keep the execution-boundary auth guard independently testable: even if a
+	// future token provider returns an empty token without an error, built-in MCP
+	// calls must fail before preflight or transport while anonymous plugins remain
+	// valid above.
+	resolveCalled := false
+	testseam.Swap(t, &runnerResolveAuthSnapshot, func(*runtimeRunner, context.Context) (AccessTokenSnapshot, error) {
+		resolveCalled = true
+		return AccessTokenSnapshot{}, nil
+	})
+	callsBefore := len(captured)
+	unauthenticated := &runtimeRunner{
+		transport:   transport.NewClient(nil),
+		globalFlags: &GlobalFlags{},
+		auditSink:   audit.NopSink{},
+	}
+	if _, err := unauthenticated.executeInvocation(context.Background(), "https://example.test", executor.Invocation{CanonicalProduct: "built-in-unauthenticated", Tool: "tool"}); err == nil || !isAuthError(err) {
+		t.Fatalf("unauthenticated built-in request = %v, want auth error", err)
+	}
+	if !resolveCalled {
+		t.Fatal("unauthenticated request did not exercise the token resolver")
+	}
+	if len(captured) != callsBefore {
+		t.Fatalf("unauthenticated built-in request reached transport: calls %d -> %d", callsBefore, len(captured))
+	}
+}
+
+func hasHeaderFold(headers map[string]string, want string) bool {
+	for key := range headers {
+		if strings.EqualFold(key, want) {
+			return true
+		}
+	}
+	return false
+}
@@ -158,7 +158,7 @@ func TestApplyAgentProductHeader(t *testing.T) {
 	}
 }

-func TestRootRejectsInvalidAgentProductBeforeEditionHook(t *testing.T) {
+func TestCrossPlatformCoverageRootRejectsInvalidAgentProductBeforeEditionHook(t *testing.T) {
 	t.Setenv("DWS_CONFIG_DIR", t.TempDir())
 	const invalidValue = "DO_NOT ECHO"
 	t.Setenv(agentproduct.EnvName, invalidValue)
@@ -335,12 +335,6 @@ func TestCrossPlatformCoverageOverlayRecoveryHostAndHelperRemainingCoverage(t *t
 	edition.Override(&edition.Hooks{ConfigDir: func() string { return "" }})
 	captureRuntimeFailure(executor.Invocation{}, nil, nil)
 	captureRuntimeFailure(executor.Invocation{}, errors.New("raw"), nil)
-	oldArgs := os.Args
-	os.Args = []string{"dws", "doc", "download", "--node", "n"}
-	if got := runtimeCommandPath(executor.Invocation{}); len(got) != 2 {
-		t.Fatalf("runtime command path = %#v", got)
-	}
-	os.Args = oldArgs

 	t.Setenv(authpkg.AgentCodeEnv, "")
 	if hostControlProviderFromEnv() != "" {
@@ -360,6 +354,10 @@ func TestCrossPlatformCoverageOverlayRecoveryHostAndHelperRemainingCoverage(t *t

 func TestCrossPlatformCoverageConfigAndCacheCommandRemainingCoverage(t *testing.T) {
 	for _, command := range []*cobra.Command{newConfigCommand(), newCacheCommand()} {
+		command.SetOut(io.Discard)
+		rootWrap := &cobra.Command{Use: "dws"}
+		rootWrap.PersistentFlags().String("format", "json", "")
+		rootWrap.AddCommand(command)
 		command.SetOut(io.Discard)
 		if err := command.RunE(command, nil); err != nil {
 			t.Fatal(err)
@@ -389,18 +387,18 @@ func TestCrossPlatformCoverageConfigAndCacheCommandRemainingCoverage(t *testing.
 	for _, format := range []string{"json", "pretty", "table"} {
 		_ = cacheRoot.PersistentFlags().Set("format", format)
 		cacheCmd.SetOut(io.Discard)
-		if err := printCacheCompatNotice(cacheCmd, "status"); err != nil {
+		if err := printCacheCompatNotice(cacheCmd, "dws cache status"); err != nil {
 			t.Fatal(err)
 		}
 	}
 	fail := errors.New("write")
 	cacheCmd.SetOut(appFailWriter{err: fail})
 	_ = cacheRoot.PersistentFlags().Set("format", "pretty")
-	if err := printCacheCompatNotice(cacheCmd, "status"); !errors.Is(err, fail) {
+	if err := printCacheCompatNotice(cacheCmd, "dws cache status"); !errors.Is(err, fail) {
 		t.Fatalf("pretty write error = %v", err)
 	}
 	_ = cacheRoot.PersistentFlags().Set("format", "table")
-	if err := printCacheCompatNotice(cacheCmd, "status"); !errors.Is(err, fail) {
+	if err := printCacheCompatNotice(cacheCmd, "dws cache status"); !errors.Is(err, fail) {
 		t.Fatalf("table write error = %v", err)
 	}
 }
@@ -20,6 +20,8 @@ import (
 	"encoding/json"
 	"fmt"
 	"io"
+	"net"
+	"net/url"
 	"os"
 	"path/filepath"
 	"runtime"
@@ -48,8 +50,24 @@ type authLoginConfig struct {
 	TargetCorpID                   string
 	HistoryProfileSelector         string
 	HistoryProfileSelectorExplicit bool
+	International                  bool
+	PreURL                         string
+	MCPURL                         string
 }

+type authLoginEndpointOverrides struct {
+	LoginURL string
+	MCPURL   string
+}
+
+type authLoginMCPPersistence uint8
+
+const (
+	authLoginMCPUseDefault authLoginMCPPersistence = iota
+	authLoginMCPUseManagedRegion
+	authLoginMCPUseExplicitOverride
+)
+
 type authLoginGuideAction string

 const (
@@ -103,12 +121,17 @@ func newAuthLoginCommand(patCaller edition.ToolCaller) *cobra.Command {
 支持的登录方式:
  - OAuth Loopback 流 (默认): 本机自动起 127.0.0.1 监听接收回调，浏览器授权后自动完成
  - OAuth 设备流 (--device): 显示 user_code + 短 URL，适合 SSH 远程 / 容器 / 无头环境
+  - 自有应用 OAuth (--client-id/--client-secret): 使用指定应用完成用户授权
  - 直接提供 Token (--token): 跳过授权，使用已有 token

 不支持的登录方式:
  - 邮箱/密码登录
  - 手机号/验证码登录
-  - 应用凭证 (AppKey/AppSecret) 直接登录
+  - 无用户授权的纯应用凭证 (client_credentials) 登录
+
+区域:
+  - 默认使用国内钉钉 .com 登录与服务端点
+  - --intl（或 --international）使用国际版 .io 登录；后续业务命令按所选 profile 自动路由

 注意: SSH 远程或无头环境（无本地浏览器可访问远端的 127.0.0.1）请使用 --device，
      否则 OAuth 回调会跳到本机不可达的 127.0.0.1 链接，授权完成后无法回写 token。
@@ -116,6 +139,9 @@ func newAuthLoginCommand(patCaller edition.ToolCaller) *cobra.Command {
 示例:
  dws auth login              # 本机登录并新增/刷新一个组织 profile
  dws auth login --profile <corpId>  # 指定本次授权目标组织，不持久切换当前组织
+  dws auth login --intl       # 使用钉钉国际版 .io 登录入口
+  dws auth login --intl --pre-url https://pre-login.dingtalk.io
+  dws auth login --intl --pre-url https://pre-mcp.dingtalk.io
  dws auth login --recommend  # 无交互批量授权服务端推荐权限
  dws auth login --device     # SSH 远程 / 无头环境登录 (设备流)
  dws auth login --force      # 兼容保留；login 默认已忽略缓存并进入授权流程
@@ -126,6 +152,22 @@ func newAuthLoginCommand(patCaller edition.ToolCaller) *cobra.Command {
 			if err != nil {
 				return err
 			}
+			var preOverrides authLoginEndpointOverrides
+			if cfg.PreURL != "" {
+				var err error
+				preOverrides, err = authLoginEndpointOverridesForPreURL(cfg.PreURL)
+				if err != nil {
+					return err
+				}
+				restoreLoginBaseURL := authpkg.PushLoginBaseURLOverride(preOverrides.LoginURL)
+				defer restoreLoginBaseURL()
+			}
+			mcpBaseURL, mcpPersistence, err := authLoginMCPBaseURLForConfig(cfg, preOverrides)
+			if err != nil {
+				return err
+			}
+			restoreMCPBaseURL := authpkg.PushMCPBaseURLOverride(mcpBaseURL)
+			defer restoreMCPBaseURL()
 			configDir := defaultConfigDir()
 			var tokenData *authpkg.TokenData
 			format, _ := cmd.Root().PersistentFlags().GetString("format")
@@ -139,6 +181,9 @@ func newAuthLoginCommand(patCaller edition.ToolCaller) *cobra.Command {
 					AccessToken: cfg.Token,
 					ExpiresAt:   time.Now().Add(config.ManualTokenExpiry),
 				}
+				if cfg.International {
+					tokenData.LoginRegion = string(authpkg.LoginRegionInternational)
+				}
 				if err := authSaveTokenData(configDir, tokenData); err != nil {
 					return apperrors.NewInternal(fmt.Sprintf("failed to persist auth token: %v", err))
 				}
@@ -149,6 +194,9 @@ func newAuthLoginCommand(patCaller edition.ToolCaller) *cobra.Command {
 				provider := authpkg.NewDeviceFlowProvider(configDir, nil)
 				provider.Output = cmd.ErrOrStderr()
 				provider.NoBrowser, _ = cmd.Flags().GetBool("no-browser")
+				if cfg.International {
+					provider.SetLoginRegion(authpkg.LoginRegionInternational)
+				}
 				provider.IdentityEnricher = func(ctx context.Context, data *authpkg.TokenData) error {
 					return enrichAuthLoginProfileFromContact(ctx, configDir, patCaller, data, authLoginHistoryHint{
 						Selector: cfg.HistoryProfileSelector,
@@ -167,6 +215,9 @@ func newAuthLoginCommand(patCaller edition.ToolCaller) *cobra.Command {
 				provider.Output = cmd.ErrOrStderr()
 				provider.NoBrowser, _ = cmd.Flags().GetBool("no-browser")
 				provider.TargetCorpID = cfg.TargetCorpID
+				if cfg.International {
+					provider.LoginRegion = authpkg.LoginRegionInternational
+				}
 				provider.IdentityEnricher = func(ctx context.Context, data *authpkg.TokenData) error {
 					return enrichAuthLoginProfileFromContact(ctx, configDir, patCaller, data, authLoginHistoryHint{
 						Selector: cfg.HistoryProfileSelector,
@@ -180,6 +231,11 @@ func newAuthLoginCommand(patCaller edition.ToolCaller) *cobra.Command {
 				}
 			}

+			if tokenData != nil {
+				if err := persistAuthLoginMCPBaseURL(configDir, mcpBaseURL, mcpPersistence); err != nil {
+					return apperrors.NewInternal(fmt.Sprintf("failed to persist MCP URL: %v", err))
+				}
+			}
 			ResetRuntimeTokenCache()
 			clearCompatCache()
 			w := cmd.OutOrStdout()
@@ -278,6 +334,10 @@ func newAuthLoginCommand(patCaller edition.ToolCaller) *cobra.Command {
 	}
 	cmd.Flags().String("token", "", "Access token")
 	cmd.Flags().Bool("device", false, "Use device authorization flow")
+	cmd.Flags().Bool("intl", false, "Use DingTalk international (.io) login and service endpoints")
+	cmd.Flags().Bool("international", false, "Use DingTalk international (.io) login and service endpoints")
+	cmd.Flags().String("pre-url", "", "Override pre-release login/MCP base URL for this login")
+	cmd.Flags().String("mcp-url", "", "Override MCP base URL for this login")
 	cmd.Flags().Bool("force", false, "兼容保留；login 默认已忽略缓存并进入授权流程")
 	cmd.Flags().Bool("recommend", false, "登录成功后无交互批量授权服务端推荐权限")
 	// Hidden compatibility flags
@@ -967,6 +1027,7 @@ func newAuthResetCommand() *cobra.Command {
 				return apperrors.NewInternal(fmt.Sprintf("failed to reset token data: %v", err))
 			}
 			_ = authRemove(filepath.Join(configDir, "mcp_url"))
+			_ = authRemove(filepath.Join(configDir, config.ManagedMCPURLRegionFileName))
 			_ = authRemove(filepath.Join(configDir, "token"))
 			_ = authDeleteAppConfig(configDir)
 			ResetRuntimeTokenCache()
@@ -1225,6 +1286,14 @@ func resolveAuthLoginConfig(cmd *cobra.Command) (authLoginConfig, error) {
 	if err != nil {
 		return authLoginConfig{}, apperrors.NewInternal("failed to read --device")
 	}
+	intl, err := cmd.Flags().GetBool("intl")
+	if err != nil {
+		return authLoginConfig{}, apperrors.NewInternal("failed to read --intl")
+	}
+	international, err := cmd.Flags().GetBool("international")
+	if err != nil {
+		return authLoginConfig{}, apperrors.NewInternal("failed to read --international")
+	}
 	force, err := cmd.Flags().GetBool("force")
 	if err != nil {
 		return authLoginConfig{}, apperrors.NewInternal("failed to read --force")
@@ -1233,6 +1302,14 @@ func resolveAuthLoginConfig(cmd *cobra.Command) (authLoginConfig, error) {
 	if err != nil {
 		return authLoginConfig{}, apperrors.NewInternal("failed to read --recommend")
 	}
+	preURL, err := cmd.Flags().GetString("pre-url")
+	if err != nil {
+		return authLoginConfig{}, apperrors.NewInternal("failed to read --pre-url")
+	}
+	mcpURL, err := cmd.Flags().GetString("mcp-url")
+	if err != nil {
+		return authLoginConfig{}, apperrors.NewInternal("failed to read --mcp-url")
+	}
 	yes := false
 	profileSelector := ""
 	if cmd.Root() != nil {
@@ -1266,9 +1343,172 @@ func resolveAuthLoginConfig(cmd *cobra.Command) (authLoginConfig, error) {
 		TargetCorpID:                   targetCorpID,
 		HistoryProfileSelector:         historyProfileSelector,
 		HistoryProfileSelectorExplicit: historyProfileSelectorExplicit,
+		International:                  intl || international,
+		PreURL:                         strings.TrimSpace(preURL),
+		MCPURL:                         strings.TrimSpace(mcpURL),
 	}, nil
 }

+func authLoginEndpointOverridesForPreURL(raw string) (authLoginEndpointOverrides, error) {
+	parsed, normalized, err := normalizeAuthLoginBaseURL(raw, "--pre-url")
+	if err != nil {
+		return authLoginEndpointOverrides{}, err
+	}
+	host := strings.ToLower(parsed.Hostname())
+	switch {
+	case strings.HasPrefix(host, "pre-login."):
+		return authLoginEndpointOverrides{
+			LoginURL: normalized,
+			MCPURL:   authLoginURLWithHost(parsed, "pre-mcp."+strings.TrimPrefix(host, "pre-login.")),
+		}, nil
+	case strings.HasPrefix(host, "pre-mcp."):
+		return authLoginEndpointOverrides{
+			LoginURL: authLoginURLWithHost(parsed, "pre-login."+strings.TrimPrefix(host, "pre-mcp.")),
+			MCPURL:   normalized,
+		}, nil
+	default:
+		return authLoginEndpointOverrides{}, apperrors.NewValidation("--pre-url must be a pre-login.* or pre-mcp.* URL")
+	}
+}
+
+func authLoginMCPBaseURLForConfig(cfg authLoginConfig, preOverrides authLoginEndpointOverrides) (string, authLoginMCPPersistence, error) {
+	if cfg.MCPURL != "" {
+		_, normalized, err := normalizeAuthLoginBaseURL(cfg.MCPURL, "--mcp-url")
+		if err != nil {
+			return "", authLoginMCPUseDefault, err
+		}
+		return normalized, authLoginMCPUseExplicitOverride, nil
+	}
+	if cfg.PreURL != "" {
+		if preOverrides.MCPURL == "" {
+			var err error
+			preOverrides, err = authLoginEndpointOverridesForPreURL(cfg.PreURL)
+			if err != nil {
+				return "", authLoginMCPUseDefault, err
+			}
+		}
+		return preOverrides.MCPURL, authLoginMCPUseExplicitOverride, nil
+	}
+	if cfg.International {
+		return authpkg.InternationalMCPBaseURL, authLoginMCPUseManagedRegion, nil
+	}
+	return authpkg.DefaultMCPBaseURL, authLoginMCPUseDefault, nil
+}
+
+func persistAuthLoginMCPBaseURL(configDir, mcpBaseURL string, persistence authLoginMCPPersistence) error {
+	mcpURLPath := filepath.Join(configDir, "mcp_url")
+	managedRegionPath := filepath.Join(configDir, config.ManagedMCPURLRegionFileName)
+
+	switch persistence {
+	case authLoginMCPUseExplicitOverride:
+		if err := removeAuthLoginManagedMCPRegion(managedRegionPath); err != nil {
+			return fmt.Errorf("clear managed MCP region: %w", err)
+		}
+		if err := authAtomicWrite(mcpURLPath, []byte(mcpBaseURL), config.FilePerm); err != nil {
+			return fmt.Errorf("save explicit MCP URL: %w", err)
+		}
+		return nil
+	case authLoginMCPUseManagedRegion:
+		managedURL, managedErr := authReadFile(managedRegionPath)
+		if managedErr != nil && !os.IsNotExist(managedErr) {
+			return fmt.Errorf("read managed MCP region: %w", managedErr)
+		}
+		currentURL, currentErr := authReadFile(mcpURLPath)
+		switch {
+		case currentErr == nil && os.IsNotExist(managedErr):
+			return nil
+		case currentErr == nil && strings.TrimSpace(string(currentURL)) != strings.TrimSpace(string(managedURL)):
+			return removeAuthLoginManagedMCPRegion(managedRegionPath)
+		case currentErr != nil && !os.IsNotExist(currentErr):
+			return fmt.Errorf("read MCP URL: %w", currentErr)
+		}
+		if err := authAtomicWrite(managedRegionPath, []byte(mcpBaseURL), config.FilePerm); err != nil {
+			return fmt.Errorf("save managed MCP region: %w", err)
+		}
+		if err := authAtomicWrite(mcpURLPath, []byte(mcpBaseURL), config.FilePerm); err != nil {
+			_ = authRemove(managedRegionPath)
+			return fmt.Errorf("save managed MCP URL: %w", err)
+		}
+		return nil
+	case authLoginMCPUseDefault:
+		managedURL, err := authReadFile(managedRegionPath)
+		if os.IsNotExist(err) {
+			return nil
+		}
+		if err != nil {
+			return fmt.Errorf("read managed MCP region: %w", err)
+		}
+		currentURL, err := authReadFile(mcpURLPath)
+		if os.IsNotExist(err) {
+			return removeAuthLoginManagedMCPRegion(managedRegionPath)
+		}
+		if err != nil {
+			return fmt.Errorf("read MCP URL: %w", err)
+		}
+		if strings.TrimSpace(string(currentURL)) != strings.TrimSpace(string(managedURL)) {
+			return removeAuthLoginManagedMCPRegion(managedRegionPath)
+		}
+		if err := authAtomicWrite(mcpURLPath, []byte(mcpBaseURL), config.FilePerm); err != nil {
+			return fmt.Errorf("restore default MCP URL: %w", err)
+		}
+		return removeAuthLoginManagedMCPRegion(managedRegionPath)
+	default:
+		return fmt.Errorf("unsupported MCP persistence mode %d", persistence)
+	}
+}
+
+func removeAuthLoginManagedMCPRegion(path string) error {
+	if err := authRemove(path); err != nil && !os.IsNotExist(err) {
+		return err
+	}
+	return nil
+}
+
+func normalizeAuthLoginBaseURL(raw, flagName string) (*url.URL, string, error) {
+	value := strings.TrimSpace(raw)
+	if value == "" {
+		return nil, "", apperrors.NewValidation(flagName + " cannot be empty")
+	}
+	if !strings.Contains(value, "://") {
+		value = "https://" + value
+	}
+	parsed, err := url.Parse(value)
+	if err != nil {
+		return nil, "", apperrors.NewValidation(fmt.Sprintf("invalid %s: %v", flagName, err))
+	}
+	if parsed.Scheme != "http" && parsed.Scheme != "https" {
+		return nil, "", apperrors.NewValidation(flagName + " must use http or https")
+	}
+	if parsed.Hostname() == "" {
+		return nil, "", apperrors.NewValidation(flagName + " must include a host")
+	}
+	if parsed.Scheme == "http" && !isAuthLoginLoopbackHost(parsed.Hostname()) {
+		return nil, "", apperrors.NewValidation(flagName + " must use HTTPS, except for a loopback HTTP test endpoint")
+	}
+	parsed.RawQuery = ""
+	parsed.Fragment = ""
+	parsed.Path = strings.TrimRight(parsed.Path, "/")
+	return parsed, strings.TrimRight(parsed.String(), "/"), nil
+}
+
+func isAuthLoginLoopbackHost(host string) bool {
+	if strings.EqualFold(strings.TrimSpace(host), "localhost") {
+		return true
+	}
+	ip := net.ParseIP(strings.TrimSpace(host))
+	return ip != nil && ip.IsLoopback()
+}
+
+func authLoginURLWithHost(parsed *url.URL, host string) string {
+	copyURL := *parsed
+	if port := parsed.Port(); port != "" {
+		copyURL.Host = net.JoinHostPort(host, port)
+	} else {
+		copyURL.Host = host
+	}
+	return strings.TrimRight(copyURL.String(), "/")
+}
+
 func authLoginForcesAuthorization(_ authLoginConfig) bool {
 	return true
 }
@@ -7,6 +7,7 @@ import (
 	"fmt"
 	"io"
 	"os"
+	"path/filepath"
 	"strings"
 	"testing"
 	"time"
@@ -214,7 +215,8 @@ func TestCrossPlatformCoverageAuthCoverageFormsParentAndTargets(t *testing.T) {
 }

 func TestCrossPlatformCoverageAuthCoverageLoginFlows(t *testing.T) {
-	t.Setenv("DWS_CONFIG_DIR", t.TempDir())
+	configDir := t.TempDir()
+	t.Setenv("DWS_CONFIG_DIR", configDir)
 	oldSave := authSaveTokenData
 	oldDevice := authDeviceLogin
 	oldOAuth := authOAuthLogin
@@ -255,6 +257,15 @@ func TestCrossPlatformCoverageAuthCoverageLoginFlows(t *testing.T) {
 	if out, _, err := authCoverageRunLogin(t, nil, "json", true, map[string]string{"token": "token"}); err != nil || !strings.Contains(out, `"token_valid": true`) {
 		t.Fatalf("json token login = %q, %v", out, err)
 	}
+	if _, _, err := authCoverageRunLogin(t, nil, "table", true, map[string]string{"token": "token", "pre-url": "https://example.com"}); err == nil {
+		t.Fatal("invalid pre-release host should fail")
+	}
+	if _, _, err := authCoverageRunLogin(t, nil, "table", true, map[string]string{"token": "token", "mcp-url": "http://remote.example.com"}); err == nil {
+		t.Fatal("remote plaintext MCP URL should fail")
+	}
+	if _, _, err := authCoverageRunLogin(t, nil, "table", true, map[string]string{"token": "token", "pre-url": "https://pre-login.dingtalk.io"}); err != nil {
+		t.Fatalf("pre-release token login = %v", err)
+	}

 	authDeviceLogin = func(*authpkg.DeviceFlowProvider, context.Context) (*authpkg.TokenData, error) {
 		return nil, errors.New("device")
@@ -271,6 +282,15 @@ func TestCrossPlatformCoverageAuthCoverageLoginFlows(t *testing.T) {
 	if _, _, err := authCoverageRunLogin(t, nil, "table", true, map[string]string{"device": "true", "no-browser": "true"}); err != nil {
 		t.Fatal(err)
 	}
+	authDeviceLogin = func(provider *authpkg.DeviceFlowProvider, _ context.Context) (*authpkg.TokenData, error) {
+		if provider.LoginRegion != authpkg.LoginRegionInternational {
+			t.Errorf("device login region = %q, want international", provider.LoginRegion)
+		}
+		return &authpkg.TokenData{AccessToken: "a", ExpiresAt: time.Now().Add(time.Hour)}, nil
+	}
+	if _, _, err := authCoverageRunLogin(t, nil, "table", true, map[string]string{"device": "true", "intl": "true"}); err != nil {
+		t.Fatalf("international device login = %v", err)
+	}

 	authOAuthLogin = func(*authpkg.OAuthProvider, context.Context, bool) (*authpkg.TokenData, error) {
 		return nil, errors.New("oauth")
@@ -291,6 +311,25 @@ func TestCrossPlatformCoverageAuthCoverageLoginFlows(t *testing.T) {
 	if out, _, err := authCoverageRunLogin(t, caller, "table", true, map[string]string{"no-browser": "true"}); err != nil || !strings.Contains(out, "Corp") {
 		t.Fatalf("oauth success = %q, %v", out, err)
 	}
+	authOAuthLogin = func(provider *authpkg.OAuthProvider, _ context.Context, _ bool) (*authpkg.TokenData, error) {
+		if provider.LoginRegion != authpkg.LoginRegionInternational {
+			t.Errorf("OAuth login region = %q, want international", provider.LoginRegion)
+		}
+		return &authpkg.TokenData{AccessToken: "a", ExpiresAt: time.Now().Add(time.Hour)}, nil
+	}
+	if _, _, err := authCoverageRunLogin(t, nil, "table", true, map[string]string{"intl": "true"}); err != nil {
+		t.Fatalf("international OAuth login = %v", err)
+	}
+
+	blockedConfigDir := t.TempDir()
+	if err := os.Mkdir(filepath.Join(blockedConfigDir, "mcp_url"), 0o700); err != nil {
+		t.Fatal(err)
+	}
+	t.Setenv("DWS_CONFIG_DIR", blockedConfigDir)
+	if _, _, err := authCoverageRunLogin(t, nil, "table", true, map[string]string{"token": "token", "intl": "true"}); err == nil || !strings.Contains(err.Error(), "failed to persist MCP URL") {
+		t.Fatalf("MCP URL persist failure = %v", err)
+	}
+	t.Setenv("DWS_CONFIG_DIR", configDir)

 	authRunLoginRecommend = func(context.Context, edition.ToolCaller, io.Writer, pat.LoginRecommendOptions) error {
 		return errors.New("recommend")
@@ -1418,7 +1457,7 @@ func TestCrossPlatformCoverageAuthCoveragePortableExchangeAndReset(t *testing.T)
 	authRemove = func(string) error { removed++; return errors.New("ignored") }
 	authDeleteAppConfig = func(string) error { removed++; return errors.New("ignored") }
 	edition.Override(&edition.Hooks{})
-	if err := reset.RunE(reset, nil); err != nil || removed != 3 || !strings.Contains(out.String(), "重新登录") {
+	if err := reset.RunE(reset, nil); err != nil || removed != 4 || !strings.Contains(out.String(), "重新登录") {
 		t.Fatalf("reset = %q, %v, removed=%d", out.String(), err, removed)
 	}
 	edition.Override(&edition.Hooks{IsEmbedded: true})
@@ -33,6 +33,8 @@ import (
 	apperrors "github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/errors"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/keychain"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/pat"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/testseam"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/pkg/config"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/pkg/edition"
 	"github.com/spf13/cobra"
 )
@@ -1033,8 +1035,12 @@ func TestResolveAuthLoginConfigReadsInheritedYes(t *testing.T) {
 	login := &cobra.Command{Use: "login"}
 	login.Flags().String("token", "", "")
 	login.Flags().Bool("device", false, "")
+	login.Flags().Bool("intl", false, "")
+	login.Flags().Bool("international", false, "")
 	login.Flags().Bool("force", false, "")
 	login.Flags().Bool("recommend", false, "")
+	login.Flags().String("pre-url", "", "")
+	login.Flags().String("mcp-url", "", "")
 	root.AddCommand(login)

 	if err := root.PersistentFlags().Set("yes", "true"); err != nil {
@@ -1061,6 +1067,560 @@ func TestResolveAuthLoginConfigReadsInheritedYes(t *testing.T) {
 	}
 }

+func TestCrossPlatformCoverageResolveAuthLoginConfigReadsInternationalAliases(t *testing.T) {
+	for _, flag := range []string{"intl", "international"} {
+		t.Run(flag, func(t *testing.T) {
+			root := &cobra.Command{Use: "dws"}
+			root.PersistentFlags().Bool("yes", false, "")
+			login := &cobra.Command{Use: "login"}
+			login.Flags().String("token", "", "")
+			login.Flags().Bool("device", false, "")
+			login.Flags().Bool("intl", false, "")
+			login.Flags().Bool("international", false, "")
+			login.Flags().Bool("force", false, "")
+			login.Flags().Bool("recommend", false, "")
+			login.Flags().String("pre-url", "", "")
+			login.Flags().String("mcp-url", "", "")
+			root.AddCommand(login)
+
+			if err := login.Flags().Set(flag, "true"); err != nil {
+				t.Fatalf("set %s: %v", flag, err)
+			}
+
+			cfg, err := resolveAuthLoginConfig(login)
+			if err != nil {
+				t.Fatalf("resolveAuthLoginConfig error = %v", err)
+			}
+			if !cfg.International {
+				t.Fatalf("International = false for --%s, want true", flag)
+			}
+		})
+	}
+
+	for _, tc := range []struct {
+		name  string
+		setup func(*cobra.Command)
+	}{
+		{
+			name: "missing intl",
+			setup: func(cmd *cobra.Command) {
+				cmd.Flags().String("token", "", "")
+				cmd.Flags().Bool("device", false, "")
+			},
+		},
+		{
+			name: "missing international",
+			setup: func(cmd *cobra.Command) {
+				cmd.Flags().String("token", "", "")
+				cmd.Flags().Bool("device", false, "")
+				cmd.Flags().Bool("intl", false, "")
+			},
+		},
+		{
+			name: "missing pre url",
+			setup: func(cmd *cobra.Command) {
+				cmd.Flags().String("token", "", "")
+				cmd.Flags().Bool("device", false, "")
+				cmd.Flags().Bool("intl", false, "")
+				cmd.Flags().Bool("international", false, "")
+				cmd.Flags().Bool("force", false, "")
+				cmd.Flags().Bool("recommend", false, "")
+			},
+		},
+		{
+			name: "missing mcp url",
+			setup: func(cmd *cobra.Command) {
+				cmd.Flags().String("token", "", "")
+				cmd.Flags().Bool("device", false, "")
+				cmd.Flags().Bool("intl", false, "")
+				cmd.Flags().Bool("international", false, "")
+				cmd.Flags().Bool("force", false, "")
+				cmd.Flags().Bool("recommend", false, "")
+				cmd.Flags().String("pre-url", "", "")
+			},
+		},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			cmd := &cobra.Command{Use: "login"}
+			tc.setup(cmd)
+			if _, err := resolveAuthLoginConfig(cmd); err == nil {
+				t.Fatal("resolveAuthLoginConfig succeeded with an incomplete flag set")
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageResolveAuthLoginConfigReadsMCPURL(t *testing.T) {
+	root := &cobra.Command{Use: "dws"}
+	root.PersistentFlags().Bool("yes", false, "")
+	login := &cobra.Command{Use: "login"}
+	login.Flags().String("token", "", "")
+	login.Flags().Bool("device", false, "")
+	login.Flags().Bool("intl", false, "")
+	login.Flags().Bool("international", false, "")
+	login.Flags().Bool("force", false, "")
+	login.Flags().Bool("recommend", false, "")
+	login.Flags().String("pre-url", "", "")
+	login.Flags().String("mcp-url", "", "")
+	root.AddCommand(login)
+
+	if err := login.Flags().Set("mcp-url", " https://pre-mcp.dingtalk.io/ "); err != nil {
+		t.Fatalf("set mcp-url: %v", err)
+	}
+
+	cfg, err := resolveAuthLoginConfig(login)
+	if err != nil {
+		t.Fatalf("resolveAuthLoginConfig error = %v", err)
+	}
+	if cfg.MCPURL != "https://pre-mcp.dingtalk.io/" {
+		t.Fatalf("MCPURL = %q, want trimmed flag value", cfg.MCPURL)
+	}
+}
+
+func TestCrossPlatformCoverageResolveAuthLoginConfigReadsPreURL(t *testing.T) {
+	root := &cobra.Command{Use: "dws"}
+	root.PersistentFlags().Bool("yes", false, "")
+	login := &cobra.Command{Use: "login"}
+	login.Flags().String("token", "", "")
+	login.Flags().Bool("device", false, "")
+	login.Flags().Bool("intl", false, "")
+	login.Flags().Bool("international", false, "")
+	login.Flags().Bool("force", false, "")
+	login.Flags().Bool("recommend", false, "")
+	login.Flags().String("pre-url", "", "")
+	login.Flags().String("mcp-url", "", "")
+	root.AddCommand(login)
+
+	if err := login.Flags().Set("pre-url", " pre-login.dingtalk.io "); err != nil {
+		t.Fatalf("set pre-url: %v", err)
+	}
+
+	cfg, err := resolveAuthLoginConfig(login)
+	if err != nil {
+		t.Fatalf("resolveAuthLoginConfig error = %v", err)
+	}
+	if cfg.PreURL != "pre-login.dingtalk.io" {
+		t.Fatalf("PreURL = %q, want trimmed flag value", cfg.PreURL)
+	}
+}
+
+func TestCrossPlatformCoverageAuthLoginEndpointOverridesForPreURL(t *testing.T) {
+	tests := []struct {
+		name      string
+		raw       string
+		wantLogin string
+		wantMCP   string
+	}{
+		{
+			name:      "pre login",
+			raw:       "https://pre-login.dingtalk.io/",
+			wantLogin: "https://pre-login.dingtalk.io",
+			wantMCP:   "https://pre-mcp.dingtalk.io",
+		},
+		{
+			name:      "pre mcp",
+			raw:       "pre-mcp.dingtalk.io",
+			wantLogin: "https://pre-login.dingtalk.io",
+			wantMCP:   "https://pre-mcp.dingtalk.io",
+		},
+		{
+			name:      "pre login with port",
+			raw:       "https://pre-login.dingtalk.io:8443/path/",
+			wantLogin: "https://pre-login.dingtalk.io:8443/path",
+			wantMCP:   "https://pre-mcp.dingtalk.io:8443/path",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := authLoginEndpointOverridesForPreURL(tc.raw)
+			if err != nil {
+				t.Fatalf("authLoginEndpointOverridesForPreURL error = %v", err)
+			}
+			if got.LoginURL != tc.wantLogin || got.MCPURL != tc.wantMCP {
+				t.Fatalf("overrides = %#v, want login %q mcp %q", got, tc.wantLogin, tc.wantMCP)
+			}
+		})
+	}
+
+	for _, raw := range []string{"https://example.com", "http://pre-login.example.com"} {
+		if _, err := authLoginEndpointOverridesForPreURL(raw); err == nil {
+			t.Fatalf("authLoginEndpointOverridesForPreURL(%q) succeeded", raw)
+		}
+	}
+}
+
+func TestCrossPlatformCoverageAuthLoginMCPBaseURLForConfig(t *testing.T) {
+	tests := []struct {
+		name            string
+		cfg             authLoginConfig
+		preOverride     authLoginEndpointOverrides
+		wantURL         string
+		wantPersistence authLoginMCPPersistence
+	}{
+		{
+			name:            "default center login uses com and resets only managed region",
+			cfg:             authLoginConfig{},
+			wantURL:         authpkg.DefaultMCPBaseURL,
+			wantPersistence: authLoginMCPUseDefault,
+		},
+		{
+			name:            "international login persists managed io",
+			cfg:             authLoginConfig{International: true},
+			wantURL:         authpkg.InternationalMCPBaseURL,
+			wantPersistence: authLoginMCPUseManagedRegion,
+		},
+		{
+			name: "pre login persists mapped pre mcp",
+			cfg:  authLoginConfig{PreURL: "pre-login.dingtalk.io"},
+			preOverride: authLoginEndpointOverrides{
+				LoginURL: "https://pre-login.dingtalk.io",
+				MCPURL:   "https://pre-mcp.dingtalk.io",
+			},
+			wantURL:         "https://pre-mcp.dingtalk.io",
+			wantPersistence: authLoginMCPUseExplicitOverride,
+		},
+		{
+			name: "explicit mcp url wins over pre url",
+			cfg: authLoginConfig{
+				PreURL: "pre-login.dingtalk.io",
+				MCPURL: " https://custom-mcp.example.com/ ",
+			},
+			preOverride: authLoginEndpointOverrides{
+				LoginURL: "https://pre-login.dingtalk.io",
+				MCPURL:   "https://pre-mcp.dingtalk.io",
+			},
+			wantURL:         "https://custom-mcp.example.com",
+			wantPersistence: authLoginMCPUseExplicitOverride,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			gotURL, gotPersistence, err := authLoginMCPBaseURLForConfig(tc.cfg, tc.preOverride)
+			if err != nil {
+				t.Fatalf("authLoginMCPBaseURLForConfig error = %v", err)
+			}
+			if gotURL != tc.wantURL || gotPersistence != tc.wantPersistence {
+				t.Fatalf("got url=%q persistence=%v, want url=%q persistence=%v", gotURL, gotPersistence, tc.wantURL, tc.wantPersistence)
+			}
+		})
+	}
+
+	for _, cfg := range []authLoginConfig{
+		{MCPURL: "http://remote.example.com"},
+		{PreURL: "https://example.com"},
+	} {
+		if _, _, err := authLoginMCPBaseURLForConfig(cfg, authLoginEndpointOverrides{}); err == nil {
+			t.Fatalf("authLoginMCPBaseURLForConfig(%#v) succeeded", cfg)
+		}
+	}
+}
+
+func TestCrossPlatformCoveragePersistAuthLoginMCPBaseURL(t *testing.T) {
+	t.Run("persists selected io mcp url", func(t *testing.T) {
+		configDir := t.TempDir()
+		if err := persistAuthLoginMCPBaseURL(configDir, authpkg.InternationalMCPBaseURL, authLoginMCPUseManagedRegion); err != nil {
+			t.Fatalf("persistAuthLoginMCPBaseURL error = %v", err)
+		}
+		data, err := os.ReadFile(filepath.Join(configDir, "mcp_url"))
+		if err != nil {
+			t.Fatalf("ReadFile(mcp_url) error = %v", err)
+		}
+		if string(data) != authpkg.InternationalMCPBaseURL {
+			t.Fatalf("mcp_url = %q, want %q", string(data), authpkg.InternationalMCPBaseURL)
+		}
+		managed, err := os.ReadFile(filepath.Join(configDir, config.ManagedMCPURLRegionFileName))
+		if err != nil || string(managed) != authpkg.InternationalMCPBaseURL {
+			t.Fatalf("managed MCP region = %q, %v", string(managed), err)
+		}
+	})
+
+	t.Run("center login preserves previous persisted override", func(t *testing.T) {
+		configDir := t.TempDir()
+		mcpURLPath := filepath.Join(configDir, "mcp_url")
+		const customURL = "https://custom-mcp.example.com"
+		if err := os.WriteFile(mcpURLPath, []byte(customURL), 0o600); err != nil {
+			t.Fatalf("WriteFile(mcp_url) error = %v", err)
+		}
+		if err := persistAuthLoginMCPBaseURL(configDir, authpkg.DefaultMCPBaseURL, authLoginMCPUseDefault); err != nil {
+			t.Fatalf("persistAuthLoginMCPBaseURL error = %v", err)
+		}
+		data, err := os.ReadFile(mcpURLPath)
+		if err != nil {
+			t.Fatalf("ReadFile(mcp_url) error = %v", err)
+		}
+		if string(data) != customURL {
+			t.Fatalf("mcp_url = %q, want preserved override %q", string(data), customURL)
+		}
+	})
+
+	t.Run("international login preserves an unmanaged explicit override", func(t *testing.T) {
+		configDir := t.TempDir()
+		mcpURLPath := filepath.Join(configDir, "mcp_url")
+		const customURL = "https://private-mcp.example.com"
+		if err := os.WriteFile(mcpURLPath, []byte(customURL), config.FilePerm); err != nil {
+			t.Fatal(err)
+		}
+		if err := persistAuthLoginMCPBaseURL(configDir, authpkg.InternationalMCPBaseURL, authLoginMCPUseManagedRegion); err != nil {
+			t.Fatal(err)
+		}
+		data, err := os.ReadFile(mcpURLPath)
+		if err != nil || string(data) != customURL {
+			t.Fatalf("explicit mcp_url = %q, %v; want preserved custom URL", string(data), err)
+		}
+		if _, err := os.Stat(filepath.Join(configDir, config.ManagedMCPURLRegionFileName)); !os.IsNotExist(err) {
+			t.Fatalf("unmanaged override acquired a managed marker: %v", err)
+		}
+	})
+
+	t.Run("center login resets a managed international URL", func(t *testing.T) {
+		configDir := t.TempDir()
+		if err := persistAuthLoginMCPBaseURL(configDir, authpkg.InternationalMCPBaseURL, authLoginMCPUseManagedRegion); err != nil {
+			t.Fatal(err)
+		}
+		if err := persistAuthLoginMCPBaseURL(configDir, authpkg.DefaultMCPBaseURL, authLoginMCPUseDefault); err != nil {
+			t.Fatal(err)
+		}
+		data, err := os.ReadFile(filepath.Join(configDir, "mcp_url"))
+		if err != nil || string(data) != authpkg.DefaultMCPBaseURL {
+			t.Fatalf("mcp_url = %q, %v; want domestic default", string(data), err)
+		}
+		if _, err := os.Stat(filepath.Join(configDir, config.ManagedMCPURLRegionFileName)); !os.IsNotExist(err) {
+			t.Fatalf("managed region marker remains after domestic login: %v", err)
+		}
+	})
+
+	t.Run("explicit override clears region ownership", func(t *testing.T) {
+		configDir := t.TempDir()
+		if err := persistAuthLoginMCPBaseURL(configDir, authpkg.InternationalMCPBaseURL, authLoginMCPUseManagedRegion); err != nil {
+			t.Fatal(err)
+		}
+		const customURL = "https://custom-mcp.example.com"
+		if err := persistAuthLoginMCPBaseURL(configDir, customURL, authLoginMCPUseExplicitOverride); err != nil {
+			t.Fatal(err)
+		}
+		if err := persistAuthLoginMCPBaseURL(configDir, authpkg.DefaultMCPBaseURL, authLoginMCPUseDefault); err != nil {
+			t.Fatal(err)
+		}
+		data, err := os.ReadFile(filepath.Join(configDir, "mcp_url"))
+		if err != nil || string(data) != customURL {
+			t.Fatalf("explicit mcp_url = %q, %v; want preserved custom URL", string(data), err)
+		}
+	})
+
+	t.Run("stale marker never deletes a different custom URL", func(t *testing.T) {
+		configDir := t.TempDir()
+		if err := os.WriteFile(filepath.Join(configDir, "mcp_url"), []byte("https://custom.example.com"), config.FilePerm); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(filepath.Join(configDir, config.ManagedMCPURLRegionFileName), []byte(authpkg.InternationalMCPBaseURL), config.FilePerm); err != nil {
+			t.Fatal(err)
+		}
+		if err := persistAuthLoginMCPBaseURL(configDir, authpkg.DefaultMCPBaseURL, authLoginMCPUseDefault); err != nil {
+			t.Fatal(err)
+		}
+		data, err := os.ReadFile(filepath.Join(configDir, "mcp_url"))
+		if err != nil || string(data) != "https://custom.example.com" {
+			t.Fatalf("custom mcp_url = %q, %v", string(data), err)
+		}
+	})
+
+	t.Run("international login clears a stale marker without changing a custom URL", func(t *testing.T) {
+		configDir := t.TempDir()
+		mcpURLPath := filepath.Join(configDir, "mcp_url")
+		managedRegionPath := filepath.Join(configDir, config.ManagedMCPURLRegionFileName)
+		const customURL = "https://private-mcp.example.com"
+		if err := os.WriteFile(mcpURLPath, []byte(customURL), config.FilePerm); err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(managedRegionPath, []byte(authpkg.DefaultMCPBaseURL), config.FilePerm); err != nil {
+			t.Fatal(err)
+		}
+		if err := persistAuthLoginMCPBaseURL(configDir, authpkg.InternationalMCPBaseURL, authLoginMCPUseManagedRegion); err != nil {
+			t.Fatal(err)
+		}
+		data, err := os.ReadFile(mcpURLPath)
+		if err != nil || string(data) != customURL {
+			t.Fatalf("custom mcp_url = %q, %v", string(data), err)
+		}
+		if _, err := os.Stat(managedRegionPath); !os.IsNotExist(err) {
+			t.Fatalf("stale managed marker remains: %v", err)
+		}
+	})
+}
+
+func TestCrossPlatformCoveragePersistAuthLoginMCPBaseURLErrors(t *testing.T) {
+	fail := errors.New("persist failure")
+
+	t.Run("explicit marker cleanup", func(t *testing.T) {
+		testseam.Swap(t, &authRemove, func(string) error { return fail })
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), "https://custom.example.com", authLoginMCPUseExplicitOverride); !errors.Is(err, fail) {
+			t.Fatalf("explicit marker cleanup error = %v", err)
+		}
+	})
+
+	t.Run("explicit URL write", func(t *testing.T) {
+		testseam.Swap(t, &authAtomicWrite, func(string, []byte, os.FileMode) error { return fail })
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), "https://custom.example.com", authLoginMCPUseExplicitOverride); !errors.Is(err, fail) {
+			t.Fatalf("explicit URL write error = %v", err)
+		}
+	})
+
+	t.Run("managed marker write", func(t *testing.T) {
+		testseam.Swap(t, &authAtomicWrite, func(string, []byte, os.FileMode) error { return fail })
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), authpkg.InternationalMCPBaseURL, authLoginMCPUseManagedRegion); !errors.Is(err, fail) {
+			t.Fatalf("managed marker write error = %v", err)
+		}
+	})
+
+	t.Run("managed marker read", func(t *testing.T) {
+		testseam.Swap(t, &authReadFile, func(string) ([]byte, error) { return nil, fail })
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), authpkg.InternationalMCPBaseURL, authLoginMCPUseManagedRegion); !errors.Is(err, fail) {
+			t.Fatalf("managed marker read error = %v", err)
+		}
+	})
+
+	t.Run("managed MCP URL read", func(t *testing.T) {
+		reads := 0
+		testseam.Swap(t, &authReadFile, func(string) ([]byte, error) {
+			reads++
+			if reads == 1 {
+				return nil, os.ErrNotExist
+			}
+			return nil, fail
+		})
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), authpkg.InternationalMCPBaseURL, authLoginMCPUseManagedRegion); !errors.Is(err, fail) {
+			t.Fatalf("managed MCP URL read error = %v", err)
+		}
+	})
+
+	t.Run("managed stale marker cleanup", func(t *testing.T) {
+		reads := 0
+		testseam.Swap(t, &authReadFile, func(string) ([]byte, error) {
+			reads++
+			if reads == 1 {
+				return []byte(authpkg.DefaultMCPBaseURL), nil
+			}
+			return []byte("https://private-mcp.example.com"), nil
+		})
+		testseam.Swap(t, &authRemove, func(string) error { return fail })
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), authpkg.InternationalMCPBaseURL, authLoginMCPUseManagedRegion); !errors.Is(err, fail) {
+			t.Fatalf("managed stale marker cleanup error = %v", err)
+		}
+	})
+
+	t.Run("managed URL write cleans marker", func(t *testing.T) {
+		writes := 0
+		removed := false
+		testseam.Swap(t, &authAtomicWrite, func(string, []byte, os.FileMode) error {
+			writes++
+			if writes == 2 {
+				return fail
+			}
+			return nil
+		})
+		testseam.Swap(t, &authRemove, func(string) error { removed = true; return nil })
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), authpkg.InternationalMCPBaseURL, authLoginMCPUseManagedRegion); !errors.Is(err, fail) || !removed {
+			t.Fatalf("managed URL write error = %v, marker removed=%v", err, removed)
+		}
+	})
+
+	t.Run("default managed marker read", func(t *testing.T) {
+		testseam.Swap(t, &authReadFile, func(string) ([]byte, error) { return nil, fail })
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), authpkg.DefaultMCPBaseURL, authLoginMCPUseDefault); !errors.Is(err, fail) {
+			t.Fatalf("managed marker read error = %v", err)
+		}
+	})
+
+	t.Run("missing MCP URL cleans marker", func(t *testing.T) {
+		reads := 0
+		testseam.Swap(t, &authReadFile, func(string) ([]byte, error) {
+			reads++
+			if reads == 1 {
+				return []byte(authpkg.InternationalMCPBaseURL), nil
+			}
+			return nil, os.ErrNotExist
+		})
+		testseam.Swap(t, &authRemove, func(string) error { return nil })
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), authpkg.DefaultMCPBaseURL, authLoginMCPUseDefault); err != nil {
+			t.Fatalf("missing MCP URL cleanup error = %v", err)
+		}
+	})
+
+	t.Run("MCP URL read", func(t *testing.T) {
+		reads := 0
+		testseam.Swap(t, &authReadFile, func(string) ([]byte, error) {
+			reads++
+			if reads == 1 {
+				return []byte(authpkg.InternationalMCPBaseURL), nil
+			}
+			return nil, fail
+		})
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), authpkg.DefaultMCPBaseURL, authLoginMCPUseDefault); !errors.Is(err, fail) {
+			t.Fatalf("MCP URL read error = %v", err)
+		}
+	})
+
+	t.Run("default URL write", func(t *testing.T) {
+		testseam.Swap(t, &authReadFile, func(string) ([]byte, error) { return []byte(authpkg.InternationalMCPBaseURL), nil })
+		testseam.Swap(t, &authAtomicWrite, func(string, []byte, os.FileMode) error { return fail })
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), authpkg.DefaultMCPBaseURL, authLoginMCPUseDefault); !errors.Is(err, fail) {
+			t.Fatalf("default URL write error = %v", err)
+		}
+	})
+
+	t.Run("final marker cleanup", func(t *testing.T) {
+		testseam.Swap(t, &authReadFile, func(string) ([]byte, error) { return []byte(authpkg.InternationalMCPBaseURL), nil })
+		testseam.Swap(t, &authAtomicWrite, func(string, []byte, os.FileMode) error { return nil })
+		testseam.Swap(t, &authRemove, func(string) error { return fail })
+		if err := persistAuthLoginMCPBaseURL(t.TempDir(), authpkg.DefaultMCPBaseURL, authLoginMCPUseDefault); !errors.Is(err, fail) {
+			t.Fatalf("final marker cleanup error = %v", err)
+		}
+	})
+
+	if err := persistAuthLoginMCPBaseURL(t.TempDir(), authpkg.DefaultMCPBaseURL, authLoginMCPPersistence(255)); err == nil {
+		t.Fatal("unsupported MCP persistence mode succeeded")
+	}
+}
+
+func TestCrossPlatformCoverageNormalizeAuthLoginBaseURLTransportSecurity(t *testing.T) {
+	tests := []struct {
+		name    string
+		raw     string
+		wantURL string
+		wantErr string
+	}{
+		{name: "https remote", raw: "https://pre-mcp.example.com/path/?secret=drop#fragment", wantURL: "https://pre-mcp.example.com/path"},
+		{name: "http localhost", raw: "http://localhost:8080/", wantURL: "http://localhost:8080"},
+		{name: "http IPv4 loopback", raw: "http://127.0.0.1:8080", wantURL: "http://127.0.0.1:8080"},
+		{name: "http IPv6 loopback", raw: "http://[::1]:8080", wantURL: "http://[::1]:8080"},
+		{name: "http remote", raw: "http://pre-mcp.example.com", wantErr: "must use HTTPS"},
+		{name: "empty", raw: " ", wantErr: "cannot be empty"},
+		{name: "invalid URL", raw: "https://%", wantErr: "invalid --mcp-url"},
+		{name: "invalid scheme", raw: "ftp://pre-mcp.example.com", wantErr: "must use http or https"},
+		{name: "missing host", raw: "https:///path", wantErr: "must include a host"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			_, got, err := normalizeAuthLoginBaseURL(tc.raw, "--mcp-url")
+			if tc.wantErr != "" {
+				if err == nil || !strings.Contains(err.Error(), tc.wantErr) {
+					t.Fatalf("normalizeAuthLoginBaseURL error = %v, want containing %q", err, tc.wantErr)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("normalizeAuthLoginBaseURL error = %v", err)
+			}
+			if got != tc.wantURL {
+				t.Fatalf("normalized URL = %q, want %q", got, tc.wantURL)
+			}
+		})
+	}
+}
+
 func TestAuthLoginForcesAuthorizationByDefault(t *testing.T) {
 	if !authLoginForcesAuthorization(authLoginConfig{}) {
 		t.Fatal("auth login should force authorization by default so each login can add an organization profile")
@@ -15,11 +15,10 @@ package app

 import "sync"

-// PluginAuth holds authentication credentials for a plugin-owned
-// streamable-http MCP server. Each server is keyed by its canonical
-// product ID (CLI.ID) so that different servers can use independent
-// tokens without interfering with each other or with the default
-// DingTalk OAuth token.
+// PluginAuth marks ownership of a plugin-owned streamable-http MCP server and
+// holds its optional authentication credentials. Every accepted HTTP plugin,
+// including an anonymous one, has a non-nil record keyed by canonical product
+// ID (CLI.ID) so execution never falls back to built-in DingTalk OAuth.
 type PluginAuth struct {
 	// Token is the Bearer token extracted from the plugin's
 	// "Authorization" header (e.g. a third-party API key).
@@ -39,27 +38,25 @@ var (
 	pluginAuthRegistry = make(map[string]*PluginAuth)
 )

-// RegisterPluginAuth stores authentication credentials for a plugin
-// server keyed by its canonical product ID. The runner looks up these
-// credentials at execution time to inject the correct Bearer token
-// instead of the default DingTalk OAuth token.
+// RegisterPluginAuth stores ownership and optional authentication credentials
+// for a plugin server keyed by its canonical product ID.
 func RegisterPluginAuth(productID string, auth *PluginAuth) {
 	pluginAuthMu.Lock()
 	defer pluginAuthMu.Unlock()
 	pluginAuthRegistry[productID] = auth
 }

-// ClearPluginAuth removes credentials for a plugin product. Registration uses
-// this before applying an accepted descriptor so a descriptor without custom
-// auth cannot inherit stale credentials from an earlier root construction.
+// ClearPluginAuth removes the ownership and credential record for a plugin
+// product.
 func ClearPluginAuth(productID string) {
 	pluginAuthMu.Lock()
 	defer pluginAuthMu.Unlock()
 	delete(pluginAuthRegistry, productID)
 }

-// LookupPluginAuth returns the authentication credentials registered
-// for the given product ID, or nil if none exists.
+// LookupPluginAuth returns plugin ownership and optional authentication
+// credentials for the product ID. The bool denotes ownership, not whether a
+// Bearer token is present.
 func LookupPluginAuth(productID string) (*PluginAuth, bool) {
 	pluginAuthMu.RLock()
 	defer pluginAuthMu.RUnlock()
@@ -21,6 +21,10 @@ import (
 	"github.com/spf13/cobra"
 )

+const cacheUnsupportedMessage = "dws cache 不再支持：服务发现已下线，当前版本使用编译期静态端点目录；dws cache 仅保留为兼容入口，不会刷新端点。"
+
+const cacheReplacementHint = "如遇 endpoint_not_resolved，请先执行 dws upgrade 获取包含最新 internal/syncdata 端点的版本；仍失败时检查 internal/syncdata.StaticServers() 是否覆盖目标 product/server。"
+
 type cacheCompatNotice struct {
 	Status      string `json:"status"`
 	Command     string `json:"command"`
@@ -28,24 +32,32 @@ type cacheCompatNotice struct {
 	Replacement string `json:"replacement,omitempty"`
 }

+// newCacheCommand keeps a visible Deprecated compatibility surface for
+// historical argv (refresh/status/clean). Behavior is a successful no-op notice.
+// Skills must not teach this path. Deprecated leaves are excluded from Schema
+// via cobra.IsAvailableCommand() — do not add schema_command_exclusions entries.
 func newCacheCommand() *cobra.Command {
 	cmd := &cobra.Command{
 		Use:               "cache",
-		Short:             "服务发现缓存兼容入口（静态端点模式已弃用）",
-		Hidden:            true,
+		Short:             "不再支持：服务发现缓存兼容入口",
+		Long:              "此命令组仅为历史 argv 兼容保留。静态端点模式下无需服务发现缓存；Skill / Agent 请勿引导此路径。",
+		Deprecated:        "不再支持；" + cacheUnsupportedMessage,
+		Args:              cobra.NoArgs,
 		DisableAutoGenTag: true,
 		RunE: func(cmd *cobra.Command, args []string) error {
-			return cmd.Help()
+			return printCacheCompatNotice(cmd, "dws cache")
 		},
 	}
 	for _, name := range []string{"refresh", "status", "clean"} {
+		subName := name
 		sub := &cobra.Command{
-			Use:               name,
-			Short:             "已弃用：静态端点模式无需服务发现缓存",
+			Use:               subName,
+			Short:             "不再支持：静态端点模式无需服务发现缓存",
+			Deprecated:        "不再支持；" + cacheUnsupportedMessage,
 			Args:              cobra.NoArgs,
 			DisableAutoGenTag: true,
 			RunE: func(cmd *cobra.Command, args []string) error {
-				return printCacheCompatNotice(cmd, name)
+				return printCacheCompatNotice(cmd, "dws cache "+subName)
 			},
 		}
 		cmd.AddCommand(sub)
@@ -56,9 +68,9 @@ func newCacheCommand() *cobra.Command {
 func printCacheCompatNotice(cmd *cobra.Command, command string) error {
 	notice := cacheCompatNotice{
 		Status:      "deprecated",
-		Command:     "dws cache " + command,
-		Message:     "服务发现已下线，当前版本使用编译期静态端点目录；dws cache 仅保留为兼容入口，不会刷新端点。",
-		Replacement: "如遇 endpoint_not_resolved，请先执行 dws upgrade 获取包含最新 internal/syncdata 端点的版本；仍失败时检查 internal/syncdata.StaticServers() 是否覆盖目标 product/server。",
+		Command:     command,
+		Message:     cacheUnsupportedMessage,
+		Replacement: cacheReplacementHint,
 	}
 	format, _ := cmd.Root().PersistentFlags().GetString("format")
 	switch strings.ToLower(strings.TrimSpace(format)) {
@@ -66,8 +78,7 @@ func printCacheCompatNotice(cmd *cobra.Command, command string) error {
 		return json.NewEncoder(cmd.OutOrStdout()).Encode(notice)
 	case "pretty":
 		data, _ := json.MarshalIndent(notice, "", "  ")
-		var err error
-		_, err = fmt.Fprintln(cmd.OutOrStdout(), string(data))
+		_, err := fmt.Fprintln(cmd.OutOrStdout(), string(data))
 		return err
 	default:
 		_, err := fmt.Fprintf(cmd.OutOrStdout(), "%s: %s\n%s\n", notice.Command, notice.Message, notice.Replacement)
@@ -0,0 +1,111 @@
+// Copyright 2026 Alibaba Group
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package app
+
+import (
+	"bytes"
+	"strings"
+	"testing"
+
+	"github.com/spf13/cobra"
+)
+
+func TestCrossPlatformCoverageCacheDeprecatedCompatShim(t *testing.T) {
+	root := NewRootCommand()
+	group := mustFindCommand(t, root, "cache")
+	if group.Hidden || group.Deprecated == "" || !group.Runnable() {
+		t.Fatalf("cache group contract: hidden=%v deprecated=%q runnable=%v", group.Hidden, group.Deprecated, group.Runnable())
+	}
+	if group.IsAvailableCommand() {
+		t.Fatal("deprecated cache group must not be IsAvailableCommand")
+	}
+
+	for _, leaf := range []string{"refresh", "status", "clean"} {
+		cmd := mustFindCommand(t, root, "cache", leaf)
+		if cmd.Hidden || cmd.Deprecated == "" || !cmd.Runnable() {
+			t.Fatalf("cache %s contract: hidden=%v deprecated=%q runnable=%v", leaf, cmd.Hidden, cmd.Deprecated, cmd.Runnable())
+		}
+		if cmd.IsAvailableCommand() {
+			t.Fatalf("deprecated cache %s must not be IsAvailableCommand", leaf)
+		}
+	}
+
+	var out bytes.Buffer
+	cmd := NewRootCommand()
+	cmd.SetOut(&out)
+	cmd.SetErr(&out)
+	cmd.SetArgs([]string{"cache", "refresh", "--format", "json"})
+	if err := cmd.Execute(); err != nil {
+		t.Fatalf("cache refresh compatibility stub: %v\n%s", err, out.String())
+	}
+	got := out.String()
+	for _, want := range []string{`"status":"deprecated"`, `"command":"dws cache refresh"`, "不再支持", "服务发现已下线"} {
+		if !strings.Contains(got, want) {
+			t.Fatalf("cache refresh output missing %q:\n%s", want, got)
+		}
+	}
+
+	for _, format := range []string{"", "json", "pretty", "table"} {
+		var buf bytes.Buffer
+		parent := &cobra.Command{Use: "dws"}
+		parent.PersistentFlags().String("format", format, "")
+		parent.SetOut(&buf)
+		sub := &cobra.Command{Use: "cache"}
+		parent.AddCommand(sub)
+		if err := printCacheCompatNotice(sub, "dws cache status"); err != nil {
+			t.Fatalf("format=%q: %v", format, err)
+		}
+		text := buf.String()
+		if !strings.Contains(text, "不再支持") && !strings.Contains(text, "服务发现已下线") {
+			t.Fatalf("format=%q missing notice:\n%s", format, text)
+		}
+		if format == "" || format == "json" || format == "pretty" {
+			if !strings.Contains(text, `"status":"deprecated"`) && !strings.Contains(text, `"status": "deprecated"`) {
+				t.Fatalf("format=%q missing deprecated JSON status:\n%s", format, text)
+			}
+		}
+	}
+
+	for _, format := range []string{"pretty", "table"} {
+		parent := &cobra.Command{Use: "dws"}
+		parent.PersistentFlags().String("format", format, "")
+		parent.SetOut(failWriter{})
+		sub := &cobra.Command{Use: "cache"}
+		parent.AddCommand(sub)
+		if err := printCacheCompatNotice(sub, "dws cache clean"); err == nil || !strings.Contains(err.Error(), "write failed") {
+			t.Fatalf("format=%q write failure = %v, want write failed", format, err)
+		}
+	}
+
+	parent := newCacheCommand()
+	var parentOut bytes.Buffer
+	rootWrap := &cobra.Command{Use: "dws"}
+	rootWrap.PersistentFlags().String("format", "json", "")
+	rootWrap.SetOut(&parentOut)
+	rootWrap.AddCommand(parent)
+	parent.SetOut(&parentOut)
+	if err := parent.RunE(parent, nil); err != nil {
+		t.Fatalf("cache parent RunE = %v, want nil success", err)
+	}
+	if !strings.Contains(parentOut.String(), `"command":"dws cache"`) {
+		t.Fatalf("cache parent notice missing command:\n%s", parentOut.String())
+	}
+
+	cache := newCacheCommand()
+	cache.SetOut(&bytes.Buffer{})
+	cache.SetArgs([]string{"status"})
+	if err := cache.Execute(); err != nil {
+		t.Fatal(err)
+	}
+}
@@ -0,0 +1,84 @@
+// Copyright 2026 Alibaba Group
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+
+package app
+
+import (
+	"strings"
+	"testing"
+
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/shortcut/chatmsg"
+)
+
+func TestCrossPlatformCoverageChatMessageReceiptActionsBindToRunnableCommands(t *testing.T) {
+	testCases := []struct {
+		name    string
+		payload map[string]any
+	}{
+		{
+			name: "send receipt awaiting status",
+			payload: chatmsg.ProjectMessageSendReceipt(map[string]any{
+				"openTaskId": "task-pending",
+			}),
+		},
+		{
+			name: "send receipt ready for message actions",
+			payload: chatmsg.ProjectMessageSendReceipt(map[string]any{
+				"openTaskId":         "task-ready",
+				"openMessageId":      "message-ready",
+				"openConversationId": "conversation-ready",
+			}),
+		},
+		{
+			name: "send status awaiting message reference",
+			payload: chatmsg.ProjectMessageSendStatus(map[string]any{
+				"status": "PENDING",
+			}, "task-pending"),
+		},
+		{
+			name: "send status ready for message actions",
+			payload: chatmsg.ProjectMessageSendStatus(map[string]any{
+				"openTaskId":         "task-ready",
+				"openMessageId":      "message-ready",
+				"openConversationId": "conversation-ready",
+				"status":             "SUCCESS",
+			}, "task-ready"),
+		},
+	}
+
+	root := NewRootCommand()
+	for _, testCase := range testCases {
+		t.Run(testCase.name, func(t *testing.T) {
+			actions, ok := testCase.payload["nextActions"].([]map[string]any)
+			if !ok || len(actions) == 0 {
+				t.Fatalf("nextActions = %#v, want non-empty []map[string]any", testCase.payload["nextActions"])
+			}
+
+			for index, action := range actions {
+				cliPath, ok := action["cliPath"].(string)
+				if !ok || strings.TrimSpace(cliPath) == "" {
+					t.Fatalf("nextActions[%d].cliPath = %#v, want non-empty string", index, action["cliPath"])
+				}
+
+				command, remaining, err := root.Find(strings.Fields(cliPath))
+				if err != nil {
+					t.Fatalf("nextActions[%d].cliPath %q does not bind: %v", index, cliPath, err)
+				}
+				if command == nil || len(remaining) != 0 || !command.Runnable() {
+					t.Fatalf("nextActions[%d].cliPath %q resolved to command=%v remaining=%v runnable=%v", index, cliPath, command, remaining, command != nil && command.Runnable())
+				}
+
+				arguments, ok := action["arguments"].(map[string]any)
+				if !ok {
+					t.Fatalf("nextActions[%d].arguments = %#v, want map[string]any", index, action["arguments"])
+				}
+				for name := range arguments {
+					if command.Flags().Lookup(name) == nil && command.InheritedFlags().Lookup(name) == nil {
+						t.Errorf("nextActions[%d] argument %q is not a flag of runnable command %q", index, name, cliPath)
+					}
+				}
+			}
+		})
+	}
+}
@@ -0,0 +1,301 @@
+// Copyright 2026 Alibaba Group
+// Licensed under the Apache License, Version 2.0
+
+package app
+
+import (
+	_ "embed"
+	"encoding/json"
+	stderrors "errors"
+	"fmt"
+	"io"
+	"reflect"
+	"strings"
+	"testing"
+
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/cli"
+	apperrors "github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/errors"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/pipeline"
+)
+
+//go:embed testdata/shortcut_hallucination_badcases_20260728.json
+var shortcutHallucinationBadcases20260728JSON []byte
+
+type shortcutHallucinationReplayCorpus struct {
+	SourceReport    string                               `json:"source_report"`
+	SourceFile      string                               `json:"source_file"`
+	SourceDWSCommit string                               `json:"source_dws_commit"`
+	Model           string                               `json:"model"`
+	ExpectedCount   int                                  `json:"expected_count"`
+	Badcases        []shortcutHallucinationReplayBadcase `json:"badcases"`
+}
+
+type shortcutHallucinationReplayBadcase struct {
+	ID               string `json:"id"`
+	CaseID           string `json:"case_id"`
+	Run              int    `json:"run"`
+	CommandIndex     int    `json:"command_index"`
+	Turn             int    `json:"turn"`
+	SourcePath       string `json:"source_path"`
+	ExpectedOutcome  string `json:"expected_outcome"`
+	Raw              string `json:"raw"`
+	OriginalExitCode int    `json:"original_exit_code"`
+	IsHelp           bool   `json:"is_help"`
+}
+
+func TestCrossPlatformCoverageShortcutHallucinationBadcases20260728Replay(t *testing.T) {
+	var corpus shortcutHallucinationReplayCorpus
+	if err := json.Unmarshal(shortcutHallucinationBadcases20260728JSON, &corpus); err != nil {
+		t.Fatalf("decode historical shortcut badcase corpus: %v", err)
+	}
+	if corpus.SourceReport != "param_hallucination_20260728_120943" ||
+		corpus.SourceFile != "raw_dashscope_qwen3_7-max_20260728_120943.json" ||
+		corpus.SourceDWSCommit != "a8e83e5" || corpus.Model != "dashscope/qwen3.7-max" {
+		t.Fatalf("unexpected historical corpus provenance: %#v", corpus)
+	}
+	if corpus.ExpectedCount != 52 || len(corpus.Badcases) != corpus.ExpectedCount {
+		t.Fatalf("historical shortcut badcases = %d, expected_count=%d; want 52", len(corpus.Badcases), corpus.ExpectedCount)
+	}
+
+	wantOutcomeCounts := map[string]int{
+		"rewrite":   8,
+		"ambiguous": 44,
+	}
+	wantSourceCounts := map[string]int{
+		"chat +group-member-list": 1,
+		"chat +group-send-text":   1,
+		"chat +list-group-bots":   1,
+		"chat +list-robot":        1,
+		"chat +list-robots":       1,
+		"chat +members":           3,
+		"chat +message-list":      1,
+		"chat +read-single":       6,
+		"chat +rename-group":      1,
+		"chat +send":              4,
+		"chat +send-by-bot":       3,
+		"chat +send-dm":           2,
+		"chat +send-file":         15,
+		"chat +send-image":        3,
+		"chat +send-media":        2,
+		"chat +send-message":      3,
+		"chat +send-single":       1,
+		"chat +send-text":         2,
+		"chat +send-to":           1,
+	}
+	rewriteTargets := map[string]string{
+		"chat +members":           "chat +group-members",
+		"chat +group-member-list": "chat +group-members",
+		"chat +list-group-bots":   "chat +chat-bots",
+		"chat +list-robot":        "chat +chat-bots",
+		"chat +list-robots":       "chat +chat-bots",
+		"chat +rename-group":      "chat +chat-update",
+	}
+
+	seenIDs := make(map[string]bool, len(corpus.Badcases))
+	gotOutcomeCounts := make(map[string]int)
+	gotSourceCounts := make(map[string]int)
+	for _, badcase := range corpus.Badcases {
+		badcase := badcase
+		t.Run(badcase.ID, func(t *testing.T) {
+			if badcase.ID == "" || seenIDs[badcase.ID] {
+				t.Fatalf("missing or duplicate historical badcase id %q", badcase.ID)
+			}
+			seenIDs[badcase.ID] = true
+			gotOutcomeCounts[badcase.ExpectedOutcome]++
+			gotSourceCounts[badcase.SourcePath]++
+
+			args := historicalShortcutBadcaseArgv(t, badcase.Raw)
+			if len(args) < 2 || strings.Join(args[:2], " ") != badcase.SourcePath {
+				t.Fatalf("raw argv source = %v, fixture source_path=%q", args, badcase.SourcePath)
+			}
+
+			root := NewSchemaSourceRootCommand()
+			root.SetOut(io.Discard)
+			root.SetErr(io.Discard)
+			root.SetArgs(args)
+			ctx, err := pipeline.RunPreParseArgs(root, newPipelineEngine(), args)
+			switch badcase.ExpectedOutcome {
+			case "rewrite":
+				assertHistoricalShortcutRewrite(t, badcase, rewriteTargets[badcase.SourcePath], ctx, err)
+			case "ambiguous":
+				assertHistoricalShortcutReason(t, badcase, ctx, err, "ambiguous_command_fallback")
+				entry, ok := cli.LookupCommandPathFallback(badcase.SourcePath)
+				if !ok || entry.Mode != cli.CommandPathFallbackAmbiguous || len(entry.Candidates) < 2 {
+					t.Fatalf("ambiguous fallback table entry = %#v, %v", entry, ok)
+				}
+			case "unknown_shortcut":
+				assertHistoricalShortcutReason(t, badcase, ctx, err, "unknown_shortcut")
+				if entry, ok := cli.LookupCommandPathFallback(badcase.SourcePath); ok {
+					t.Fatalf("contract-incomplete path unexpectedly entered fallback table: %#v", entry)
+				}
+			default:
+				t.Fatalf("unsupported expected_outcome %q", badcase.ExpectedOutcome)
+			}
+		})
+	}
+	if !reflect.DeepEqual(gotOutcomeCounts, wantOutcomeCounts) {
+		t.Errorf("historical outcome counts = %v, want %v", gotOutcomeCounts, wantOutcomeCounts)
+	}
+	if !reflect.DeepEqual(gotSourceCounts, wantSourceCounts) {
+		t.Errorf("historical source counts = %v, want %v", gotSourceCounts, wantSourceCounts)
+	}
+}
+
+func TestCrossPlatformCoverageShortcutHallucinationMerged20260720Replay(t *testing.T) {
+	tests := []struct {
+		id          string
+		raw         string
+		occurrences int
+		outcome     string
+		target      string
+	}{
+		{
+			id:          "dws_im_v2_0013-search-group",
+			raw:         `dws chat +search-group --name "dws测试群02" --format json`,
+			occurrences: 1,
+			outcome:     "official_alias",
+			target:      "chat +chat-search",
+		},
+		{
+			id:          "dws_oa_0017-list-processes",
+			raw:         `dws oa +list-processes --format json --limit 50`,
+			occurrences: 2,
+			outcome:     "ambiguous",
+		},
+	}
+	totalOccurrences := 0
+	for _, test := range tests {
+		test := test
+		t.Run(test.id, func(t *testing.T) {
+			totalOccurrences += test.occurrences
+			args := historicalShortcutBadcaseArgv(t, test.raw)
+			root := NewSchemaSourceRootCommand()
+			root.SetOut(io.Discard)
+			root.SetErr(io.Discard)
+			root.SetArgs(args)
+			ctx, err := pipeline.RunPreParseArgs(root, newPipelineEngine(), args)
+			badcase := shortcutHallucinationReplayBadcase{
+				ID:              test.id,
+				SourcePath:      strings.Join(args[:2], " "),
+				ExpectedOutcome: test.outcome,
+				Raw:             test.raw,
+			}
+			switch test.outcome {
+			case "rewrite":
+				assertHistoricalShortcutRewrite(t, badcase, test.target, ctx, err)
+			case "official_alias":
+				assertHistoricalOfficialAlias(t, badcase, test.target, ctx, err)
+			case "ambiguous":
+				assertHistoricalShortcutReason(t, badcase, ctx, err, "ambiguous_command_fallback")
+			default:
+				t.Fatalf("unsupported merged expected outcome %q", test.outcome)
+			}
+		})
+	}
+	if totalOccurrences != 3 {
+		t.Fatalf("merged shortcut hallucination occurrences = %d, want 3", totalOccurrences)
+	}
+}
+
+func assertHistoricalOfficialAlias(
+	t *testing.T,
+	badcase shortcutHallucinationReplayBadcase,
+	wantTarget string,
+	ctx *pipeline.Context,
+	err error,
+) {
+	t.Helper()
+	if err != nil {
+		t.Fatalf("historical official alias %q returned error: %v", badcase.SourcePath, err)
+	}
+	if wantTarget == "" || ctx == nil || ctx.Command != "dws "+wantTarget {
+		t.Fatalf("historical official alias context = %#v; want command dws %s", ctx, wantTarget)
+	}
+	for _, correction := range ctx.Corrections {
+		if correction.Handler == "command-path-fallback" {
+			t.Fatalf("historical official alias received fallback correction: %#v", ctx.Corrections)
+		}
+	}
+	if entry, ok := cli.LookupCommandPathFallback(badcase.SourcePath); ok {
+		t.Fatalf("official alias unexpectedly remains in fallback table: %#v", entry)
+	}
+}
+
+func historicalShortcutBadcaseArgv(t *testing.T, raw string) []string {
+	t.Helper()
+	command := strings.TrimSpace(raw)
+	for _, suffix := range []string{" 2>&1 | head -50", " 2>&1"} {
+		command = strings.TrimSuffix(command, suffix)
+	}
+	argv, err := cli.ParseAgentExampleArgv(command)
+	if err != nil {
+		t.Fatalf("parse historical raw command %q without a shell: %v", raw, err)
+	}
+	if len(argv) < 3 || argv[0] != "dws" {
+		t.Fatalf("historical raw command did not produce dws argv: %q => %v", raw, argv)
+	}
+	return append([]string(nil), argv[1:]...)
+}
+
+func assertHistoricalShortcutRewrite(
+	t *testing.T,
+	badcase shortcutHallucinationReplayBadcase,
+	wantTarget string,
+	ctx *pipeline.Context,
+	err error,
+) {
+	t.Helper()
+	if wantTarget == "" {
+		t.Fatalf("historical rewrite %q has no reviewed target", badcase.SourcePath)
+	}
+	if ctx == nil || ctx.Command != "dws "+wantTarget {
+		t.Fatalf("historical rewrite context = %#v, error=%v; want command dws %s", ctx, err, wantTarget)
+	}
+	found := false
+	for _, correction := range ctx.Corrections {
+		if correction.Handler == "command-path-fallback" && correction.Original == badcase.SourcePath && correction.Corrected == wantTarget {
+			found = true
+			break
+		}
+	}
+	if !found {
+		t.Fatalf("historical rewrite corrections = %#v; want %q -> %q", ctx.Corrections, badcase.SourcePath, wantTarget)
+	}
+	wantPrefix := strings.Fields(wantTarget)
+	if len(ctx.Args) < len(wantPrefix) || !reflect.DeepEqual(ctx.Args[:len(wantPrefix)], wantPrefix) {
+		t.Fatalf("historical rewrite argv = %v, want prefix %v", ctx.Args, wantPrefix)
+	}
+	if reason := structuredShortcutReplayReason(err); reason == "unknown_shortcut" || reason == "ambiguous_command_fallback" {
+		t.Fatalf("historical rewrite returned command-resolution reason %q: %v", reason, err)
+	}
+}
+
+func assertHistoricalShortcutReason(
+	t *testing.T,
+	badcase shortcutHallucinationReplayBadcase,
+	ctx *pipeline.Context,
+	err error,
+	wantReason string,
+) {
+	t.Helper()
+	if ctx == nil {
+		t.Fatalf("historical %s badcase returned nil context: %v", wantReason, err)
+	}
+	if got := structuredShortcutReplayReason(err); got != wantReason {
+		t.Fatalf("historical badcase %q reason = %q, error=%v; want %q (original exit=%d help=%v)", badcase.Raw, got, err, wantReason, badcase.OriginalExitCode, badcase.IsHelp)
+	}
+	if strings.Contains(strings.ToLower(fmt.Sprint(err)), "unknown flag") {
+		t.Fatalf("historical badcase %q regressed to unknown flag: %v", badcase.Raw, err)
+	}
+}
+
+func structuredShortcutReplayReason(err error) string {
+	if err == nil {
+		return ""
+	}
+	var structured *apperrors.Error
+	if stderrors.As(err, &structured) {
+		return structured.Reason
+	}
+	return ""
+}
@@ -0,0 +1,603 @@
+// Copyright 2026 Alibaba Group
+// Licensed under the Apache License, Version 2.0
+
+package app
+
+import (
+	"bytes"
+	"context"
+	stderrors "errors"
+	"io"
+	"os"
+	"os/exec"
+	"path/filepath"
+	"reflect"
+	"strings"
+	"testing"
+
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/cli"
+	apperrors "github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/errors"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/pipeline"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/shortcut"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/pkg/cmdutil"
+	"github.com/spf13/cobra"
+)
+
+const (
+	runtimeShortcutFallbackChildEnv  = "DWS_RUNTIME_SHORTCUT_FALLBACK_CHILD"
+	runtimeShortcutFallbackConfigEnv = "DWS_RUNTIME_SHORTCUT_FALLBACK_CONFIG_DIR"
+)
+
+func TestCrossPlatformCoverageRuntimeUserShortcutTakesPrecedenceOverReviewedFallback(t *testing.T) {
+	if os.Getenv(runtimeShortcutFallbackChildEnv) == "1" {
+		t.Setenv("DWS_CONFIG_DIR", os.Getenv(runtimeShortcutFallbackConfigEnv))
+		engine := newPipelineEngine()
+		root := NewRootCommandWithEngine(context.Background(), engine)
+		runtimeCommand := exactAppCommand(root, "chat +members")
+		if runtimeCommand == nil || !runtimeCommand.Runnable() || cmdutil.IsHintOnlyCommand(runtimeCommand) {
+			var userDefined []string
+			for _, candidate := range shortcut.All() {
+				if candidate.UserDefined {
+					userDefined = append(userDefined, candidate.Service+" "+candidate.Command)
+				}
+			}
+			t.Fatalf("runtime shortcut was not mounted as an executable command: command=%#v user_defined=%v config=%q", runtimeCommand, userDefined, os.Getenv("DWS_CONFIG_DIR"))
+		}
+		root.SetOut(io.Discard)
+		root.SetErr(io.Discard)
+		args := []string{"chat", "+members", "--sentinel", "fixture", "--mock", "--format", "json"}
+		root.SetArgs(args)
+		ctx, err := pipeline.RunPreParseArgs(root, engine, args)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if ctx == nil || ctx.Command != "dws chat +members" {
+			t.Fatalf("runtime shortcut context = %#v", ctx)
+		}
+		assertNoCommandPathFallbackCorrection(t, ctx)
+		if err := root.Execute(); err != nil {
+			t.Fatalf("runtime shortcut execute: %v", err)
+		}
+		return
+	}
+
+	configDir := t.TempDir()
+	shortcutDir := filepath.Join(configDir, "shortcuts")
+	if err := os.MkdirAll(shortcutDir, 0o700); err != nil {
+		t.Fatal(err)
+	}
+	definition := `version: 1
+service: chat
+command: "+members"
+product: chat
+description: runtime members fixture
+execute:
+  tool: fixture_user_members
+  bind:
+    sentinel: "${sentinel}"
+flags:
+  - name: sentinel
+    type: string
+    required: true
+    desc: fixture sentinel
+`
+	if err := os.WriteFile(filepath.Join(shortcutDir, "chat.members.yaml"), []byte(definition), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	command := exec.Command(os.Args[0], "-test.run=^TestCrossPlatformCoverageRuntimeUserShortcutTakesPrecedenceOverReviewedFallback$", "-test.count=1")
+	command.Env = append(os.Environ(), runtimeShortcutFallbackConfigEnv+"="+configDir, runtimeShortcutFallbackChildEnv+"=1")
+	output, err := command.CombinedOutput()
+	if err != nil {
+		t.Fatalf("runtime shortcut child failed: %v\n%s", err, strings.TrimSpace(string(output)))
+	}
+	if !strings.Contains(string(output), `"_tool": "fixture_user_members"`) {
+		t.Fatalf("runtime shortcut child used the wrong command:\n%s", strings.TrimSpace(string(output)))
+	}
+}
+
+func TestCrossPlatformCoverageReviewedCommandFallbacksReachCanonicalDryRunPayload(t *testing.T) {
+	_, canonicalQuery, canonicalQueryAttempts, err := executeParamAliasDryRunE2E(t,
+		"chat", "+chat-search", "--query", "project", "--dry-run",
+	)
+	if err != nil || len(canonicalQueryAttempts) != 0 {
+		t.Fatalf("canonical query preview = %#v, attempts=%v, error=%v", canonicalQuery, canonicalQueryAttempts, err)
+	}
+	tests := []struct {
+		name        string
+		args        []string
+		wantCommand string
+		wantPreview paramAliasDryRunPreview
+	}{
+		{
+			name:        "group search query",
+			args:        []string{"chat", "+group-search", "--query", "project", "--dry-run"},
+			wantCommand: "dws chat +chat-search",
+			wantPreview: canonicalQuery,
+		},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			ctx, preview, attempts, executeErr := executeParamAliasDryRunE2E(t, test.args...)
+			if executeErr != nil {
+				t.Fatalf("fallback execute error = %v", executeErr)
+			}
+			if len(attempts) != 0 {
+				t.Fatalf("fallback crossed dry-run dispatch boundary: %#v", attempts)
+			}
+			if !reflect.DeepEqual(preview, test.wantPreview) {
+				t.Fatalf("fallback preview = %#v, want canonical %#v", preview, test.wantPreview)
+			}
+			if ctx == nil || ctx.Command != test.wantCommand || len(ctx.Corrections) == 0 {
+				t.Fatalf("fallback context = %#v, want command %q", ctx, test.wantCommand)
+			}
+			correction := ctx.Corrections[0]
+			if correction.Handler != "command-path-fallback" || correction.Kind != "reviewed-fallback" {
+				t.Fatalf("fallback correction = %#v", correction)
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageOfficialCommandAliasesBypassFallbackAndReachEquivalentPayload(t *testing.T) {
+	_, canonicalQuery, canonicalQueryAttempts, err := executeParamAliasDryRunE2E(t,
+		"chat", "+chat-search", "--query", "project", "--dry-run",
+	)
+	if err != nil || len(canonicalQueryAttempts) != 0 {
+		t.Fatalf("canonical query preview = %#v, attempts=%v, error=%v", canonicalQuery, canonicalQueryAttempts, err)
+	}
+	_, canonicalKeyword, canonicalKeywordAttempts, err := executeParamAliasDryRunE2E(t,
+		"chat", "+chat-search", "--keyword", "project", "--dry-run",
+	)
+	if err != nil || len(canonicalKeywordAttempts) != 0 {
+		t.Fatalf("canonical keyword preview = %#v, attempts=%v, error=%v", canonicalKeyword, canonicalKeywordAttempts, err)
+	}
+	_, nativeCanonical, nativeCanonicalAttempts, err := executeParamAliasDryRunE2E(t,
+		"chat", "search", "--query", "project", "--dry-run",
+	)
+	if err != nil || len(nativeCanonicalAttempts) != 0 {
+		t.Fatalf("native canonical preview = %#v, attempts=%v, error=%v", nativeCanonical, nativeCanonicalAttempts, err)
+	}
+
+	tests := []struct {
+		name        string
+		args        []string
+		wantCommand string
+		wantPreview paramAliasDryRunPreview
+	}{
+		{
+			name:        "search group shortcut alias",
+			args:        []string{"chat", "+search-group", "--keyword", "project", "--dry-run"},
+			wantCommand: "dws chat +chat-search",
+			wantPreview: canonicalKeyword,
+		},
+		{
+			name:        "chat group search shortcut alias",
+			args:        []string{"chat", "+chat-group-search", "--query", "project", "--dry-run"},
+			wantCommand: "dws chat +chat-search",
+			wantPreview: canonicalQuery,
+		},
+		{
+			name:        "hidden native compatibility leaf",
+			args:        []string{"chat", "group", "search", "--query", "project", "--dry-run"},
+			wantCommand: "dws chat group search",
+			wantPreview: nativeCanonical,
+		},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			ctx, preview, attempts, executeErr := executeParamAliasDryRunE2E(t, test.args...)
+			if executeErr != nil {
+				t.Fatalf("official alias execute error = %v", executeErr)
+			}
+			if len(attempts) != 0 {
+				t.Fatalf("official alias crossed dry-run dispatch boundary: %#v", attempts)
+			}
+			if !reflect.DeepEqual(preview, test.wantPreview) {
+				t.Fatalf("official alias preview = %#v, want canonical %#v", preview, test.wantPreview)
+			}
+			if ctx == nil || ctx.Command != test.wantCommand {
+				t.Fatalf("official alias context = %#v, want command %q", ctx, test.wantCommand)
+			}
+			assertNoCommandPathFallbackCorrection(t, ctx)
+		})
+	}
+}
+
+func TestCrossPlatformCoverageReviewedReadFallbacksResolveCanonicalLeafBeforeParameterValidation(t *testing.T) {
+	tests := []struct {
+		name   string
+		args   []string
+		target string
+	}{
+		{
+			name:   "members",
+			args:   []string{"chat", "+members", "--group", "Fixture Group"},
+			target: "chat +group-members",
+		},
+		{
+			name:   "group member list",
+			args:   []string{"chat", "+group-member-list", "--group-name", "Fixture Group"},
+			target: "chat +group-members",
+		},
+		{
+			name:   "list group bots",
+			args:   []string{"chat", "+list-group-bots", "--group", "cid-fixture"},
+			target: "chat +chat-bots",
+		},
+		{
+			name:   "list robot",
+			args:   []string{"chat", "+list-robot", "--group", "cid-fixture"},
+			target: "chat +chat-bots",
+		},
+		{
+			name:   "list robots",
+			args:   []string{"chat", "+list-robots", "--group", "cid-fixture"},
+			target: "chat +chat-bots",
+		},
+		{
+			name:   "bot list",
+			args:   []string{"chat", "+bot-list", "--group", "cid-fixture"},
+			target: "chat +chat-bots",
+		},
+		{
+			name:   "conversation detail",
+			args:   []string{"chat", "+conversation-detail", "--group", "cid-fixture"},
+			target: "chat +conversation-info",
+		},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			root := NewSchemaSourceRootCommand()
+			root.SetOut(io.Discard)
+			root.SetErr(io.Discard)
+			root.SetArgs(test.args)
+			ctx, err := pipeline.RunPreParseArgs(root, newPipelineEngine(), test.args)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if ctx == nil || ctx.Command != "dws "+test.target || len(ctx.Corrections) == 0 {
+				t.Fatalf("read fallback context = %#v", ctx)
+			}
+			wantPrefix := strings.Fields(test.target)
+			if len(ctx.Args) < len(wantPrefix) || !reflect.DeepEqual(ctx.Args[:len(wantPrefix)], wantPrefix) {
+				t.Fatalf("read fallback args = %v, want prefix %v", ctx.Args, wantPrefix)
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageOfficialGroupMembersAliasBypassesCommandFallback(t *testing.T) {
+	args := []string{"chat", "+chat-group-members", "--conversation-id", "cid-fixture", "--member-types", "user"}
+	root := NewSchemaSourceRootCommand()
+	root.SetOut(io.Discard)
+	root.SetErr(io.Discard)
+	root.SetArgs(args)
+	ctx, err := pipeline.RunPreParseArgs(root, newPipelineEngine(), args)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if ctx == nil || ctx.Command != "dws chat +chat-members-list" {
+		t.Fatalf("official group-members alias context = %#v", ctx)
+	}
+	assertNoCommandPathFallbackCorrection(t, ctx)
+}
+
+func TestCrossPlatformCoverageReviewedRenameFallbackResolvesCanonicalLeafBeforeParameterValidation(t *testing.T) {
+	args := []string{"chat", "+rename-group", "--id", "cid-fixture", "--name", "Fixture Group"}
+	root := NewSchemaSourceRootCommand()
+	root.SetOut(io.Discard)
+	root.SetErr(io.Discard)
+	root.SetArgs(args)
+	ctx, err := pipeline.RunPreParseArgs(root, newPipelineEngine(), args)
+	if err != nil {
+		t.Fatal(err)
+	}
+	if ctx == nil || ctx.Command != "dws chat +chat-update" || len(ctx.Corrections) == 0 {
+		t.Fatalf("rename fallback context = %#v", ctx)
+	}
+	wantPrefix := []string{"chat", "+chat-update"}
+	if len(ctx.Args) < len(wantPrefix) || !reflect.DeepEqual(ctx.Args[:len(wantPrefix)], wantPrefix) {
+		t.Fatalf("rename fallback args = %v, want prefix %v", ctx.Args, wantPrefix)
+	}
+}
+
+func TestCrossPlatformCoverageReviewedDocHistorySaveFallbacksPreserveArguments(t *testing.T) {
+	sources := []string{"+create-version", "+save-version", "+snapshot", "+version-create"}
+	for _, source := range sources {
+		t.Run(source, func(t *testing.T) {
+			args := []string{"doc", source, "--node", "node-fixture", "--mock", "--format", "json"}
+			root := NewSchemaSourceRootCommand()
+			root.SetOut(io.Discard)
+			root.SetErr(io.Discard)
+			root.SetArgs(args)
+			ctx, err := pipeline.RunPreParseArgs(root, newPipelineEngine(), args)
+			if err != nil {
+				t.Fatal(err)
+			}
+			if ctx == nil || ctx.Command != "dws doc +history-save" || len(ctx.Corrections) == 0 {
+				t.Fatalf("history-save fallback context = %#v", ctx)
+			}
+			wantArgs := []string{"doc", "+history-save", "--node", "node-fixture", "--mock", "--format", "json"}
+			if !reflect.DeepEqual(ctx.Args, wantArgs) {
+				t.Fatalf("history-save fallback args = %v, want %v", ctx.Args, wantArgs)
+			}
+			correction := ctx.Corrections[0]
+			if correction.Handler != "command-path-fallback" || correction.Kind != "reviewed-fallback" {
+				t.Fatalf("history-save correction = %#v", correction)
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageDocExportPDFRemainsUnknownShortcut(t *testing.T) {
+	if entry, ok := cli.LookupCommandPathFallback("doc +export-pdf"); ok {
+		t.Fatalf("+export-pdf must not receive a path-only fallback: %#v", entry)
+	}
+	root := NewSchemaSourceRootCommand()
+	root.SetOut(io.Discard)
+	root.SetErr(io.Discard)
+	args := []string{"doc", "+export-pdf", "--node", "node-fixture", "--mock", "--format", "json"}
+	root.SetArgs(args)
+	ctx, err := pipeline.RunPreParseArgs(root, newPipelineEngine(), args)
+	if ctx == nil {
+		t.Fatal("+export-pdf returned nil context")
+	}
+	var structured *apperrors.Error
+	if !stderrors.As(err, &structured) || structured.Reason != "unknown_shortcut" {
+		t.Fatalf("+export-pdf preparse error = %T %#v", err, err)
+	}
+	if len(ctx.Corrections) != 0 {
+		t.Fatalf("+export-pdf unexpectedly received corrections: %#v", ctx.Corrections)
+	}
+}
+
+func TestCrossPlatformCoverageReviewedAmbiguousCommandFallbackNeverDispatches(t *testing.T) {
+	tests := []struct {
+		path       string
+		candidates []string
+	}{
+		{path: "chat +group-send-text", candidates: []string{"chat +send-to-group", "chat +messages-send"}},
+		{path: "chat +message-list", candidates: []string{"chat +chat-messages", "chat +messages-list-direct", "chat +search-msg", "chat +unread-chats"}},
+		{path: "chat +read-single", candidates: []string{"chat +messages-list-direct", "chat +chat-messages"}},
+		{path: "chat +send", candidates: []string{"chat +messages-send", "chat +dm", "chat +send-to-group"}},
+		{path: "chat +send-by-bot", candidates: []string{"chat +messages-send", "chat message send-by-bot"}},
+		{path: "chat +send-dm", candidates: []string{"chat +dm", "chat +messages-send"}},
+		{path: "chat +send-message", candidates: []string{"chat +messages-send", "chat +dm", "chat +send-to-group"}},
+		{path: "chat +send-single", candidates: []string{"chat +dm", "chat +messages-send"}},
+		{path: "chat +send-text", candidates: []string{"chat +messages-send", "chat +dm", "chat +send-to-group"}},
+		{path: "chat +send-to", candidates: []string{"chat +messages-send", "chat +dm", "chat +send-to-group"}},
+		{path: "chat +send-file", candidates: []string{"chat +messages-send", "chat message send"}},
+		{path: "chat +send-image", candidates: []string{"chat +messages-send", "chat message send"}},
+		{path: "chat +send-media", candidates: []string{"chat +messages-send", "chat message send"}},
+		{path: "chat +conversation-category-list", candidates: []string{"chat +category-list", "chat +category-list-conversations"}},
+		{path: "chat +conversation-group-list", candidates: []string{"chat +category-list-conversations", "chat +conversation-list"}},
+		{path: "chat +list-my-groups", candidates: []string{"chat +my-groups", "chat +chat-list-mine", "chat +chat-list"}},
+		{path: "oa +list-processes", candidates: []string{"oa +list-forms", "oa +my-initiated", "oa approval list-initiated"}},
+	}
+	for _, test := range tests {
+		t.Run(test.path, func(t *testing.T) {
+			caller := &paramAliasCaptureCaller{}
+			args := append(strings.Fields(test.path), "--format", "json")
+			ctx, err := executeParamAliasE2E(t, caller, args...)
+			if ctx == nil {
+				t.Fatal("ambiguous command fallback returned nil context")
+			}
+			var structured *apperrors.Error
+			if !stderrors.As(err, &structured) || structured.Reason != "ambiguous_command_fallback" || structured.ExitCode() != 3 {
+				t.Fatalf("ambiguous error = %T %#v", err, err)
+			}
+			if len(structured.Actions) != len(test.candidates) || len(caller.calls) != 0 {
+				t.Fatalf("ambiguous actions=%v calls=%#v", structured.Actions, caller.calls)
+			}
+			for _, candidate := range test.candidates {
+				if !strings.Contains(structured.Hint, "dws "+candidate) {
+					t.Errorf("ambiguous hint %q missing candidate %q", structured.Hint, candidate)
+				}
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageCanonicalShortcutBadFlagDoesNotEnterCommandFallback(t *testing.T) {
+	root := NewSchemaSourceRootCommand()
+	root.SetOut(io.Discard)
+	root.SetErr(io.Discard)
+	args := []string{"chat", "+chat-search", "--definitely-not-a-real-flag", "project"}
+	root.SetArgs(args)
+	ctx, err := pipeline.RunPreParseArgs(root, newPipelineEngine(), args)
+	if err != nil {
+		t.Fatalf("preparse error = %v", err)
+	}
+	if ctx == nil || ctx.Command != "dws chat +chat-search" {
+		t.Fatalf("canonical context = %#v", ctx)
+	}
+	for _, correction := range ctx.Corrections {
+		if correction.Handler == "command-path-fallback" {
+			t.Fatalf("canonical command received fallback correction: %#v", ctx.Corrections)
+		}
+	}
+	if err := root.Execute(); err == nil || !strings.Contains(err.Error(), "unknown flag") {
+		t.Fatalf("canonical bad flag error = %v", err)
+	}
+}
+
+func TestCrossPlatformCoverageRewrittenShortcutStillUsesCanonicalParameterErrors(t *testing.T) {
+	tests := []struct {
+		name string
+		args []string
+		want string
+	}{
+		{
+			name: "missing required query",
+			args: []string{"chat", "+group-search", "--dry-run"},
+			want: "请至少指定 --query、--keyword 之一",
+		},
+		{
+			name: "unknown canonical flag",
+			args: []string{"chat", "+group-search", "--definitely-not-a-real-flag", "project"},
+			want: "unknown flag",
+		},
+	}
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			root := NewSchemaSourceRootCommand()
+			root.SetOut(io.Discard)
+			root.SetErr(io.Discard)
+			root.SetArgs(test.args)
+			ctx, err := pipeline.RunPreParseArgs(root, newPipelineEngine(), test.args)
+			if err != nil {
+				t.Fatalf("preparse error = %v", err)
+			}
+			if ctx == nil || ctx.Command != "dws chat +chat-search" || len(ctx.Corrections) == 0 ||
+				ctx.Corrections[0].Handler != "command-path-fallback" {
+				t.Fatalf("rewritten context = %#v", ctx)
+			}
+			if err := root.Execute(); err == nil || !strings.Contains(err.Error(), test.want) {
+				t.Fatalf("canonical parameter error = %v, want containing %q", err, test.want)
+			}
+		})
+	}
+}
+
+func TestCrossPlatformCoverageCommandFallbackNamesStayOutOfHelpSchemaAndShortcutCatalog(t *testing.T) {
+	invalidShortcuts := map[string]bool{
+		"+bot-list":                   true,
+		"+conversation-detail":        true,
+		"+conversation-category-list": true,
+		"+conversation-group-list":    true,
+		"+list-my-groups":             true,
+		"+group-search":               true,
+		"+members":                    true,
+		"+group-member-list":          true,
+		"+list-group-bots":            true,
+		"+list-robot":                 true,
+		"+list-robots":                true,
+		"+message-list":               true,
+		"+read-single":                true,
+		"+rename-group":               true,
+		"+send":                       true,
+		"+send-by-bot":                true,
+		"+send-dm":                    true,
+		"+send-message":               true,
+		"+send-single":                true,
+		"+send-text":                  true,
+		"+send-to":                    true,
+		"+send-file":                  true,
+		"+send-image":                 true,
+		"+send-media":                 true,
+		"+group-send-text":            true,
+	}
+	root := NewSchemaSourceRootCommand()
+	chat := exactAppCommand(root, "chat")
+	if chat == nil {
+		t.Fatal("chat command missing")
+	}
+	for _, child := range chat.Commands() {
+		if invalidShortcuts[child.Name()] {
+			t.Fatalf("fallback name %q became a real command", child.Name())
+		}
+		for _, alias := range child.Aliases {
+			if invalidShortcuts[alias] {
+				t.Fatalf("fallback name %q became a Cobra alias", alias)
+			}
+		}
+	}
+	for path := range invalidShortcuts {
+		if _, ok := cli.ResolveMeta("chat " + path); ok {
+			t.Fatalf("fallback path %q leaked into embedded Schema", path)
+		}
+	}
+	for _, declared := range shortcut.All() {
+		if declared.Service == "chat" && invalidShortcuts[declared.Command] {
+			t.Fatalf("fallback name %q leaked into shortcut catalog", declared.Command)
+		}
+	}
+	if _, ok := cli.ResolveMeta("oa +list-processes"); ok {
+		t.Fatal("fallback path oa +list-processes leaked into embedded Schema")
+	}
+	for _, declared := range shortcut.All() {
+		if declared.Service == "oa" && declared.Command == "+list-processes" {
+			t.Fatal("fallback name +list-processes leaked into shortcut catalog")
+		}
+	}
+
+	group := exactAppCommand(root, "chat group")
+	searchCompatibility := exactAppCommand(root, "chat group search")
+	if group == nil || searchCompatibility == nil || !searchCompatibility.Hidden || cmdutil.IsHintOnlyCommand(searchCompatibility) {
+		t.Fatalf("chat group search must be a hidden executable compatibility leaf: group=%v search=%v", group, searchCompatibility)
+	}
+	var help bytes.Buffer
+	group.SetOut(&help)
+	if err := group.Help(); err != nil {
+		t.Fatal(err)
+	}
+	if strings.Contains(help.String(), "\n  search ") {
+		t.Fatalf("hidden fallback source leaked into group help:\n%s", help.String())
+	}
+}
+
+func TestCrossPlatformCoverageDocCommandFallbackNamesStayOutOfHelpSchemaAndShortcutCatalog(t *testing.T) {
+	invalidShortcuts := map[string]bool{
+		"+create-version": true,
+		"+save-version":   true,
+		"+snapshot":       true,
+		"+version-create": true,
+		"+export-pdf":     true,
+	}
+	root := NewSchemaSourceRootCommand()
+	doc := exactAppCommand(root, "doc")
+	if doc == nil {
+		t.Fatal("doc command missing")
+	}
+	for _, child := range doc.Commands() {
+		if invalidShortcuts[child.Name()] {
+			t.Fatalf("invalid shortcut %q became a real command", child.Name())
+		}
+		for _, alias := range child.Aliases {
+			if invalidShortcuts[alias] {
+				t.Fatalf("invalid shortcut %q became a Cobra alias", alias)
+			}
+		}
+	}
+	for path := range invalidShortcuts {
+		if _, ok := cli.ResolveMeta("doc " + path); ok {
+			t.Fatalf("invalid path %q leaked into embedded Schema", path)
+		}
+	}
+	for _, declared := range shortcut.All() {
+		if declared.Service == "doc" && invalidShortcuts[declared.Command] {
+			t.Fatalf("invalid shortcut %q leaked into shortcut catalog", declared.Command)
+		}
+	}
+}
+
+func assertNoCommandPathFallbackCorrection(t *testing.T, ctx *pipeline.Context) {
+	t.Helper()
+	for _, correction := range ctx.Corrections {
+		if correction.Handler == "command-path-fallback" {
+			t.Fatalf("official command unexpectedly received fallback correction: %#v", ctx.Corrections)
+		}
+	}
+}
+
+func exactAppCommand(root *cobra.Command, path string) *cobra.Command {
+	current := root
+	parts := strings.Fields(path)
+	if len(parts) > 0 && parts[0] == root.Name() {
+		parts = parts[1:]
+	}
+	for _, part := range parts {
+		var next *cobra.Command
+		for _, child := range current.Commands() {
+			if child.Name() == part {
+				next = child
+				break
+			}
+		}
+		if next == nil {
+			return nil
+		}
+		current = next
+	}
+	return current
+}
@@ -32,10 +32,11 @@ import (
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/keychain"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/pat"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/plugin"
-	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/recovery"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/safety"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/skillstate"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/transport"
 	upgradepkg "github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/upgrade"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/pkg/config"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/pkg/edition"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/pkg/mcptypes"
 	tea "github.com/charmbracelet/bubbletea"
@@ -227,115 +228,6 @@ func TestCrossPlatformCoverageDocDownloadPureCoverage(t *testing.T) {
 	}
 }

-func TestCrossPlatformCoverageRecoveryPureCoverage(t *testing.T) {
-	if _, err := decodeRecoveryAttempts(nil, nil, "", ""); err != nil {
-		t.Fatal(err)
-	}
-	if _, err := decodeRecoveryAttempts(json.RawMessage("null"), nil, "", ""); err != nil {
-		t.Fatal(err)
-	}
-	if got, err := decodeRecoveryAttempts(json.RawMessage(`[{"command_summary":"one"}]`), nil, "", ""); err != nil || len(got) != 1 {
-		t.Fatalf("array attempts = %#v %v", got, err)
-	}
-	if _, err := decodeRecoveryAttempts(json.RawMessage("{"), nil, "", ""); err == nil {
-		t.Fatal("malformed attempts succeeded")
-	}
-	if got, err := decodeRecoveryAttempts(json.RawMessage("2"), []string{"a"}, "ok", ""); err != nil || len(got) != 2 {
-		t.Fatalf("legacy attempts = %#v %v", got, err)
-	}
-	if legacyRecoveryAttempts(0, nil, "", "") != nil || len(legacyRecoveryAttempts(1, nil, "", "")) != 1 {
-		t.Fatal("legacy attempts edge mismatch")
-	}
-
-	for _, args := range [][]string{
-		{"dws", "--debug", "doc", "get", "--node", "n"},
-		{"dws", "--format=json", "doc", "--", "ignored"},
-		{"dws", "--unknown", "value", "doc"},
-	} {
-		old := os.Args
-		os.Args = args
-		_ = currentCommandPath()
-		os.Args = old
-	}
-	for _, inv := range []executor.Invocation{
-		{LegacyPath: "legacy path"},
-		{CanonicalProduct: "doc", Tool: "get"},
-		{CanonicalProduct: "doc"},
-		{},
-	} {
-		old := os.Args
-		os.Args = []string{"dws"}
-		_ = runtimeCommandPath(inv)
-		os.Args = old
-	}
-	if cloneRecoveryArgs(nil) != nil {
-		t.Fatal("empty recovery args should clone to nil")
-	}
-	original := map[string]any{"x": 1}
-	clone := cloneRecoveryArgs(original)
-	clone["x"] = 2
-	if original["x"] != 1 {
-		t.Fatal("recovery args were not cloned")
-	}
-
-	if got, _ := (*recoveryRuntime)(nil).Search(context.Background(), "query", recovery.RecoveryContext{}); got.DocSearch.Status != "skipped" {
-		t.Fatalf("nil recovery search = %#v", got)
-	}
-	if got, _ := (&recoveryRuntime{}).Search(context.Background(), " ", recovery.RecoveryContext{}); got.DocSearch.Status != "skipped" {
-		t.Fatalf("blank recovery search = %#v", got)
-	}
-	if _, err := (*recoveryRuntime)(nil).CallToolDirect(context.Background(), "x", "y", nil); err == nil {
-		t.Fatal("nil recovery runtime call succeeded")
-	}
-	if _, err := (&recoveryRuntime{}).resolveEndpoint(context.Background(), "missing", "tool"); err == nil || !strings.Contains(err.Error(), `endpoint not resolved for product "missing" (tool "tool")`) {
-		t.Fatalf("missing recovery endpoint error = %v", err)
-	} else {
-		var apiErr *apperrors.Error
-		if !errors.As(err, &apiErr) || apiErr.Category != apperrors.CategoryAPI || apiErr.Operation != "discovery.resolve" || apiErr.Reason != "endpoint_not_resolved" {
-			t.Fatalf("missing recovery endpoint classification = %#v", err)
-		}
-	}
-	t.Setenv("DINGTALK_OK_MCP_URL", " https://catalog.test ")
-	runtime := &recoveryRuntime{}
-	if got, err := runtime.resolveEndpoint(context.Background(), "ok", "tool"); err != nil || got != "https://catalog.test" {
-		t.Fatalf("recovery endpoint override = %q %v", got, err)
-	}
-	if recoveryRuntimeToken(nil) != "" || recoveryRuntimeToken(&GlobalFlags{Token: " token "}) != "token" {
-		t.Fatal("recovery token mismatch")
-	}
-	if toRecoveryToolResponse(nil) != nil {
-		t.Fatal("nil recovery response should stay nil")
-	}
-	response := toRecoveryToolResponse(&transport.ToolCallResult{IsError: true, Blocks: []transport.ContentBlock{{Type: "text", Text: "body"}}})
-	if response == nil || !response.IsError || len(response.Content) != 1 {
-		t.Fatalf("recovery response = %#v", response)
-	}
-
-	items := []any{map[string]any{"title": "A", "url": "u", "desc": "d"}, "skip", map[string]any{}}
-	for _, payload := range []map[string]any{
-		nil,
-		{"items": items},
-		{"data": map[string]any{"items": items}},
-		{"result": map[string]any{"items": items}},
-	} {
-		_ = parseDocSearchItemsFromMap(payload)
-	}
-	if toDocSearchItems("bad") != nil {
-		t.Fatal("non-list doc items accepted")
-	}
-	result := &transport.ToolCallResult{Content: map[string]any{}, Blocks: []transport.ContentBlock{{Text: "{"}, {Text: `{"items":[{"title":"B"}]}`}}}
-	if got := parseDocSearchItems(result); len(got) != 1 {
-		t.Fatalf("block doc items = %#v", got)
-	}
-	if parseDocSearchItems(nil) != nil {
-		t.Fatal("nil doc result should be nil")
-	}
-	searchItems := []recovery.DocSearchItem{{Title: "query", URL: "u"}, {Title: "other"}, {Title: "third"}, {Title: "fourth"}}
-	if len(rerankDocSearchHits("query", recovery.RecoveryContext{ToolName: "tool", CommandPath: []string{"doc"}}, searchItems)) != 3 || rerankDocSearchHits("", recovery.RecoveryContext{}, nil) != nil {
-		t.Fatal("doc search reranking mismatch")
-	}
-}
-
 func TestCrossPlatformCoverageSmallAppRegistryAndRootCoverage(t *testing.T) {
 	RegisterPluginAuth("coverage-registry", &PluginAuth{Token: "token"})
 	t.Cleanup(func() {
@@ -435,6 +327,29 @@ func TestCrossPlatformCoverageSmallAppRegistryAndRootCoverage(t *testing.T) {
 func TestCrossPlatformCoverageDirectRuntimeCoverage(t *testing.T) {
 	oldEdition := edition.Get()
 	t.Cleanup(func() { edition.Override(oldEdition); SetDynamicServers(nil) })
+	for _, tc := range []struct {
+		raw    string
+		region authpkg.LoginRegion
+		want   string
+	}{
+		{raw: "%", want: "%"},
+		{raw: "https://dingtalk.io/path", want: "https://dingtalk.com/path"},
+		{raw: "https://mcp.dingtalk.com:8443/path", region: authpkg.LoginRegionInternational, want: "https://mcp.dingtalk.io:8443/path"},
+	} {
+		if got := mcpBaseURLForLoginRegion(tc.raw, tc.region); got != tc.want {
+			t.Fatalf("mcpBaseURLForLoginRegion(%q, %q) = %q, want %q", tc.raw, tc.region, got, tc.want)
+		}
+	}
+	if hasDirectRuntimeEndpointOverride("") {
+		t.Fatal("blank product unexpectedly has an endpoint override")
+	}
+	t.Setenv("DINGTALK_COVERAGE_PRODUCT_MCP_URL", "https://override.test")
+	if !hasDirectRuntimeEndpointOverride("coverage-product") {
+		t.Fatal("configured product endpoint override was not detected")
+	}
+	if got := activeDingTalkGatewayEndpointWithBase("https://mcp-gw.dingtalk.com/server/contact", "%"); got != "https://mcp-gw.dingtalk.com/server/contact" {
+		t.Fatalf("invalid gateway base rewrote endpoint to %q", got)
+	}
 	server := mcptypes.ServerDescriptor{
 		Endpoint: "https://one.test",
 		CLI: mcptypes.CLIOverlay{
@@ -511,59 +426,6 @@ func TestCrossPlatformCoverageDirectRuntimeCoverage(t *testing.T) {
 	_ = defaultPATMCPEndpoint()
 }

-func TestCrossPlatformCoverageRecoveryLoadExecutionCoverage(t *testing.T) {
-	if _, err := loadRecoveryExecution(filepath.Join(t.TempDir(), "missing")); err == nil {
-		t.Fatal("missing recovery execution succeeded")
-	}
-	path := filepath.Join(t.TempDir(), "execution.json")
-	if err := os.WriteFile(path, []byte("{"), 0o600); err != nil {
-		t.Fatal(err)
-	}
-	if _, err := loadRecoveryExecution(path); err == nil {
-		t.Fatal("malformed recovery execution succeeded")
-	}
-	for name, body := range map[string]string{
-		"legacy": `{"action":" one ","attempt":2,"result":" ok ","error":" bad "}`,
-		"modern": `{"actions":["one"],"attempts":[{"command_summary":"one"}],"error_summary":"bad"}`,
-	} {
-		t.Run(name, func(t *testing.T) {
-			if err := os.WriteFile(path, []byte(body), 0o600); err != nil {
-				t.Fatal(err)
-			}
-			if got, err := loadRecoveryExecution(path); err != nil || len(got.Actions) != 1 || len(got.Attempts) == 0 {
-				t.Fatalf("loaded execution = %#v %v", got, err)
-			}
-		})
-	}
-}
-
-func TestCrossPlatformCoverageRecoveryRuntimeHTTP(t *testing.T) {
-	var result map[string]any = map[string]any{"content": []map[string]any{{"type": "text", "text": `{"items":[{"title":"query result","url":"u"}]}`}}}
-	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
-		var req struct {
-			ID int `json:"id"`
-		}
-		_ = json.NewDecoder(r.Body).Decode(&req)
-		_ = json.NewEncoder(w).Encode(map[string]any{"jsonrpc": "2.0", "id": req.ID, "result": result})
-	}))
-	defer server.Close()
-	SetDynamicServers([]mcptypes.ServerDescriptor{{Endpoint: server.URL, CLI: mcptypes.CLIOverlay{ID: "devdoc", Tools: []mcptypes.CLITool{{Name: "search_open_platform_docs_rag"}}}}})
-	t.Cleanup(func() { SetDynamicServers(nil) })
-	runtime := &recoveryRuntime{transport: transport.NewClient(server.Client()), flags: &GlobalFlags{Token: "token"}}
-	got, err := runtime.Search(context.Background(), "query", recovery.RecoveryContext{ToolName: "search"})
-	if err != nil || got.DocSearch.Status != "success" || len(got.KBHits) == 0 {
-		t.Fatalf("recovery search = %#v %v", got, err)
-	}
-	result = map[string]any{"isError": true, "content": []map[string]any{{"type": "text", "text": "failed"}}}
-	if _, err := runtime.CallToolDirect(context.Background(), "devdoc", "search_open_platform_docs_rag", nil); err == nil {
-		t.Fatal("recovery MCP error succeeded")
-	}
-	server.Close()
-	if _, err := runtime.CallToolDirect(context.Background(), "devdoc", "search_open_platform_docs_rag", nil); err == nil {
-		t.Fatal("recovery network error succeeded")
-	}
-}
-
 func TestCrossPlatformCoverageEventCommandPureCoverage(t *testing.T) {
 	oldEdition := edition.Get()
 	t.Cleanup(func() { edition.Override(oldEdition) })
@@ -760,7 +622,7 @@ func TestCrossPlatformCoverageVersionCacheCompletionCoverage(t *testing.T) {
 		child := &cobra.Command{Use: "child"}
 		root.AddCommand(child)
 		child.SetOut(io.Discard)
-		if err := printCacheCompatNotice(child, "status"); err != nil {
+		if err := printCacheCompatNotice(child, "dws cache status"); err != nil {
 			t.Fatalf("cache %s: %v", format, err)
 		}
 		root.RemoveCommand(child)
@@ -1075,12 +937,21 @@ func TestCrossPlatformCoverageAuthCommandPureCoverage(t *testing.T) {
 }

 func TestCrossPlatformCoverageAuthLoginTokenCommandCoverage(t *testing.T) {
+	t.Setenv(keychain.DisableKeychainEnv, "1")
+	t.Setenv(keychain.StorageDirEnv, t.TempDir())
 	oldInteractive := authLoginInteractiveTerminal
 	authLoginInteractiveTerminal = func() bool { return false }
 	t.Cleanup(func() { authLoginInteractiveTerminal = oldInteractive; authpkg.SetRuntimeProfile("") })
-	for _, format := range []string{"table", "json"} {
-		t.Run(format, func(t *testing.T) {
-			t.Setenv("DWS_CONFIG_DIR", t.TempDir())
+	for _, tc := range []struct {
+		format        string
+		international bool
+	}{
+		{format: "table"},
+		{format: "json", international: true},
+	} {
+		t.Run(tc.format, func(t *testing.T) {
+			configDir := t.TempDir()
+			t.Setenv("DWS_CONFIG_DIR", configDir)
 			root := &cobra.Command{Use: "dws"}
 			root.PersistentFlags().String("format", "table", "")
 			root.PersistentFlags().Bool("yes", false, "")
@@ -1091,16 +962,90 @@ func TestCrossPlatformCoverageAuthLoginTokenCommandCoverage(t *testing.T) {
 			root.SetOut(&output)
 			root.SetErr(io.Discard)
 			args := []string{"login", "--token", "manual-token", "--yes"}
-			if format == "json" {
+			if tc.international {
+				args = append(args, "--intl")
+			}
+			if tc.format == "json" {
 				args = append(args, "--format", "json")
 			}
 			root.SetArgs(args)
 			if err := root.Execute(); err != nil || output.Len() == 0 {
 				t.Fatalf("token login = %q %v", output.String(), err)
 			}
+			data, err := authpkg.LoadTokenData(configDir)
+			if err != nil {
+				t.Fatalf("LoadTokenData error = %v", err)
+			}
+			wantRegion := ""
+			if tc.international {
+				wantRegion = string(authpkg.LoginRegionInternational)
+			}
+			if data.LoginRegion != wantRegion {
+				t.Fatalf("LoginRegion = %q, want %q", data.LoginRegion, wantRegion)
+			}
+			if tc.international {
+				snapshot, err := resolveAccessTokenSnapshotFromDir(context.Background(), configDir, "")
+				if err != nil {
+					t.Fatalf("resolveAccessTokenSnapshotFromDir error = %v", err)
+				}
+				gotEndpoint := activeDingTalkGatewayEndpointForLoginRegion(
+					"https://mcp-gw.dingtalk.com/server/contact",
+					snapshot.LoginRegion,
+				)
+				if wantEndpoint := "https://mcp-gw.dingtalk.io/server/contact"; gotEndpoint != wantEndpoint {
+					t.Fatalf("international manual-token endpoint = %q, want %q", gotEndpoint, wantEndpoint)
+				}
+			}
 		})
 	}

+	t.Run("international then domestic login restores domestic MCP URL", func(t *testing.T) {
+		configDir := t.TempDir()
+		t.Setenv("DWS_CONFIG_DIR", configDir)
+		runLogin := func(international bool) {
+			t.Helper()
+			root := &cobra.Command{Use: "dws"}
+			root.PersistentFlags().String("format", "table", "")
+			root.PersistentFlags().Bool("yes", false, "")
+			root.PersistentFlags().String("profile", "", "")
+			root.AddCommand(newAuthLoginCommand(nil))
+			args := []string{"login", "--token", "manual-token", "--yes"}
+			if international {
+				args = append(args, "--intl")
+			}
+			root.SetArgs(args)
+			root.SetOut(io.Discard)
+			root.SetErr(io.Discard)
+			if err := root.Execute(); err != nil {
+				t.Fatalf("international=%v login error = %v", international, err)
+			}
+		}
+
+		runLogin(true)
+		if data, err := os.ReadFile(filepath.Join(configDir, "mcp_url")); err != nil || string(data) != authpkg.InternationalMCPBaseURL {
+			t.Fatalf("international mcp_url = %q, %v", string(data), err)
+		}
+		runLogin(false)
+		if data, err := os.ReadFile(filepath.Join(configDir, "mcp_url")); err != nil || string(data) != authpkg.DefaultMCPBaseURL {
+			t.Fatalf("domestic mcp_url = %q, %v", string(data), err)
+		}
+		if _, err := os.Stat(filepath.Join(configDir, config.ManagedMCPURLRegionFileName)); !os.IsNotExist(err) {
+			t.Fatalf("managed MCP region marker remains: %v", err)
+		}
+
+		const customURL = "https://private-mcp.example.com"
+		if err := os.WriteFile(filepath.Join(configDir, "mcp_url"), []byte(customURL), config.FilePerm); err != nil {
+			t.Fatal(err)
+		}
+		runLogin(true)
+		if data, err := os.ReadFile(filepath.Join(configDir, "mcp_url")); err != nil || string(data) != customURL {
+			t.Fatalf("explicit mcp_url after international login = %q, %v; want preserved custom URL", string(data), err)
+		}
+		if _, err := os.Stat(filepath.Join(configDir, config.ManagedMCPURLRegionFileName)); !os.IsNotExist(err) {
+			t.Fatalf("explicit mcp_url acquired a managed marker: %v", err)
+		}
+	})
+
 	for _, hidden := range []bool{false, true} {
 		old := edition.Get()
 		edition.Override(&edition.Hooks{HideAuthLogin: hidden})
@@ -1747,9 +1692,6 @@ func TestCrossPlatformCoverageDoctorCommandCoverage(t *testing.T) {
 	}

 	for _, jsonOut := range []bool{false, true} {
-		if got := doctorCheckCache(io.Discard, jsonOut); got.Status != statusPass {
-			t.Fatal("cache check failed")
-		}
 		if got := doctorCheckPerf(io.Discard, jsonOut); got.Status != statusPass {
 			t.Fatalf("perf check = %#v", got)
 		}
@@ -2168,7 +2110,7 @@ func TestCrossPlatformCoverageSkillSetupRuntimeCoverage(t *testing.T) {
 	}
 	_ = os.Symlink(filepath.Join(mono, "SKILL.md"), filepath.Join(mono, "linked.md"))
 	multi := filepath.Join(t.TempDir(), "multi")
-	for _, name := range []string{"dws-shared", "dingtalk-a", "dingtalk-b"} {
+	for _, name := range []string{"dingtalk-shared", "dingtalk-a", "dingtalk-b"} {
 		dir := filepath.Join(multi, name)
 		if err := os.MkdirAll(dir, 0o755); err != nil {
 			t.Fatal(err)
@@ -2196,9 +2138,13 @@ func TestCrossPlatformCoverageSkillSetupRuntimeCoverage(t *testing.T) {
 	if output, _, err := run("--mode", "multi", "--source", multi, "--target", "agents", "--yes", "--skill", "a"); err != nil || !strings.Contains(output, "installed=2") {
 		t.Fatalf("multi setup = %q, %v", output, err)
 	}
-	if _, err := os.Stat(filepath.Join(home, ".agents", "skills", "dws-shared", "SKILL.md")); err != nil {
+	if _, err := os.Stat(filepath.Join(home, ".agents", "skills", "dingtalk-shared", "SKILL.md")); err != nil {
 		t.Fatal(err)
 	}
+	state, readable, err := skillstate.Read(home)
+	if err != nil || !readable || len(state.OfficialSkills) != 3 || len(state.UpdatedSkills) != 2 {
+		t.Fatalf("setup state = %#v, readable=%v, err=%v", state, readable, err)
+	}
 	if output, _, err := run("--mode", "multi", "--source", multi, "--target", "agents", "--yes", "--dry-run", "--exclude", "b"); err != nil || !strings.Contains(output, "DRY-RUN") {
 		t.Fatalf("multi dry run = %q, %v", output, err)
 	}
@@ -2223,7 +2169,7 @@ func TestCrossPlatformCoverageSkillSetupRuntimeCoverage(t *testing.T) {
 }

 func TestCrossPlatformCoverageSkillSetupPureCoverage(t *testing.T) {
-	all := []string{"dingtalk-a", "dingtalk-b", "dws-shared"}
+	all := []string{"dingtalk-a", "dingtalk-b", "dingtalk-shared"}
 	for _, tc := range []struct {
 		include []string
 		exclude []string
@@ -2241,7 +2187,7 @@ func TestCrossPlatformCoverageSkillSetupPureCoverage(t *testing.T) {
 			t.Errorf("filter %#v/%#v = %v", tc.include, tc.exclude, err)
 		}
 	}
-	for _, selected := range [][]string{nil, {"dws-shared"}, {"dingtalk-a"}} {
+	for _, selected := range [][]string{nil, {"dingtalk-shared"}, {"dingtalk-a"}} {
 		_ = ensureMandatorySharedSkill(selected, all)
 	}
 	_ = ensureMandatorySharedSkill([]string{"dingtalk-a"}, []string{"dingtalk-a"})
@@ -2286,7 +2232,7 @@ func TestCrossPlatformCoverageSkillSetupPureCoverage(t *testing.T) {
 	for _, tc := range []struct{ path, mode string }{{"", skillSetupModeMono}, {mono, skillSetupModeMono}, {filepath.Dir(multi), skillSetupModeMulti}, {root, "bad"}} {
 		_ = isSkillSourceRoot(tc.path, tc.mode)
 	}
-	t.Setenv("HOME", t.TempDir())
+	setTestHome(t, t.TempDir())
 	for _, tc := range []struct{ target, mode string }{{"agents", skillSetupModeMono}, {"agents", skillSetupModeMulti}, {"all", skillSetupModeMono}, {"missing", skillSetupModeMono}} {
 		_, _ = resolveSkillSetupTargets(tc.target, tc.mode)
 	}
@@ -61,11 +61,16 @@ func appRPCServer(t *testing.T, initOK, listOK bool) *httptest.Server {
 }

 func TestCrossPlatformCoveragePluginAuthCoverage(t *testing.T) {
-	registerPluginAuthFromHeaders(mcptypes.ServerDescriptor{Key: "fallback", Endpoint: "%", AuthHeaders: map[string]string{"Authorization": "token"}})
-	registerPluginAuthFromHeaders(mcptypes.ServerDescriptor{Key: "server", Endpoint: "https://x.test", CLI: mcptypes.CLIOverlay{ID: "cli"}, AuthHeaders: map[string]string{"Authorization": "Bearer token", "X": "Y"}})
-	registerPluginAuthFromHeaders(mcptypes.ServerDescriptor{Key: "none"})
-	if got, ok := LookupPluginAuth("cli"); !ok || got == nil || got.Token != "token" {
-		t.Fatalf("registered plugin auth = %#v, %v", got, ok)
+	fallback := pluginAuthFromServerDescriptor(mcptypes.ServerDescriptor{Key: "fallback", Endpoint: "%", AuthHeaders: map[string]string{"Authorization": "token"}})
+	if fallback == nil || fallback.Token != "token" || len(fallback.TrustedDomains) != 0 {
+		t.Fatalf("fallback plugin auth = %#v", fallback)
+	}
+	got := pluginAuthFromServerDescriptor(mcptypes.ServerDescriptor{Key: "server", Endpoint: "https://x.test", CLI: mcptypes.CLIOverlay{ID: "cli"}, AuthHeaders: map[string]string{"Authorization": "Bearer token", "X": "Y"}})
+	if got == nil || got.Token != "token" || got.ExtraHeaders["X"] != "Y" || len(got.TrustedDomains) != 2 {
+		t.Fatalf("plugin auth = %#v", got)
+	}
+	if anonymous := pluginAuthFromServerDescriptor(mcptypes.ServerDescriptor{Key: "none"}); anonymous == nil || anonymous.Token != "" {
+		t.Fatalf("anonymous plugin ownership = %#v", anonymous)
 	}
 }

@@ -302,7 +307,6 @@ func TestCrossPlatformCoverageRootUtilityAndTimingCoverage(t *testing.T) {
 	_ = newConfigCommand()
 	_ = newCacheCommand()
 	_ = newVersionCommand()
-	_ = newRecoveryCommand(&GlobalFlags{})
 	_ = newAPICommand(&GlobalFlags{})
 	_ = NewRootCommand(context.Background())
 }
@@ -74,6 +74,20 @@ func defaultPATMCPEndpoint() string {

 func defaultPATGatewayBaseURL() string {
 	raw := strings.TrimSpace(authpkg.GetMCPBaseURL())
+	return mcpGatewayBaseURL(raw)
+}
+
+func defaultPATGatewayBaseURLForLoginRegion(region authpkg.LoginRegion) string {
+	raw := strings.TrimSpace(authpkg.GetMCPBaseURL())
+	if override := authpkg.MCPBaseURLOverride(); override != "" {
+		raw = override
+	} else {
+		raw = mcpBaseURLForLoginRegion(raw, region)
+	}
+	return mcpGatewayBaseURL(raw)
+}
+
+func mcpGatewayBaseURL(raw string) string {
 	parsed, err := url.Parse(raw)
 	if err != nil || parsed.Scheme == "" || parsed.Host == "" {
 		return strings.TrimRight(raw, "/")
@@ -100,6 +114,33 @@ func defaultPATGatewayBaseURL() string {
 	return strings.TrimRight(parsed.String(), "/")
 }

+func mcpBaseURLForLoginRegion(raw string, region authpkg.LoginRegion) string {
+	parsed, err := url.Parse(strings.TrimSpace(raw))
+	if err != nil || parsed.Scheme == "" || parsed.Host == "" {
+		return raw
+	}
+	host := strings.ToLower(parsed.Hostname())
+	fromSuffix, toSuffix := ".dingtalk.io", ".dingtalk.com"
+	if region.IsInternational() {
+		fromSuffix, toSuffix = toSuffix, fromSuffix
+	}
+	bareFrom := strings.TrimPrefix(fromSuffix, ".")
+	if host != bareFrom && !strings.HasSuffix(host, fromSuffix) {
+		return raw
+	}
+	if host == bareFrom {
+		host = strings.TrimPrefix(toSuffix, ".")
+	} else {
+		host = strings.TrimSuffix(host, fromSuffix) + toSuffix
+	}
+	if port := parsed.Port(); port != "" {
+		parsed.Host = net.JoinHostPort(host, port)
+	} else {
+		parsed.Host = host
+	}
+	return parsed.String()
+}
+
 // SetDynamicServers injects server data discovered from servers.json.
 // All product endpoints are resolved dynamically from this data.
 func SetDynamicServers(servers []mcptypes.ServerDescriptor) {
@@ -131,7 +172,7 @@ func registerDynamicServer(server mcptypes.ServerDescriptor, endpoints map[strin
 		return
 	}
 	id := strings.TrimSpace(server.CLI.ID)
-	endpoint := strings.TrimSpace(server.Endpoint)
+	endpoint := activeDingTalkGatewayEndpoint(server.Endpoint)
 	if id != "" && endpoint != "" {
 		endpoints[id] = endpoint
 		products[id] = true
@@ -262,6 +303,19 @@ func directRuntimeEndpoint(productID, toolName string) (string, bool) {
 	return "", false
 }

+func hasDirectRuntimeEndpointOverride(productID string) bool {
+	normalized := normalizeDirectRuntimeProductID(productID)
+	for _, candidate := range []string{strings.TrimSpace(productID), normalized} {
+		if candidate == "" {
+			continue
+		}
+		if _, ok := productEndpointOverride(candidate); ok {
+			return true
+		}
+	}
+	return false
+}
+
 func editionServerEndpoint(productID string) (string, bool) {
 	productID = strings.TrimSpace(productID)
 	if productID == "" {
@@ -282,7 +336,7 @@ func endpointFromEditionServers(productID string, fn func() []edition.ServerInfo
 		return "", false
 	}
 	for _, server := range fn() {
-		endpoint := strings.TrimSpace(server.Endpoint)
+		endpoint := activeDingTalkGatewayEndpoint(server.Endpoint)
 		if endpoint == "" {
 			continue
 		}
@@ -298,6 +352,46 @@ func endpointFromEditionServers(productID string, fn func() []edition.ServerInfo
 	return "", false
 }

+func activeDingTalkGatewayEndpoint(endpoint string) string {
+	return activeDingTalkGatewayEndpointWithBase(endpoint, defaultPATGatewayBaseURL())
+}
+
+func activeDingTalkGatewayEndpointForLoginRegion(endpoint string, region authpkg.LoginRegion) string {
+	return activeDingTalkGatewayEndpointWithBase(endpoint, defaultPATGatewayBaseURLForLoginRegion(region))
+}
+
+func activeDingTalkGatewayEndpointWithBase(endpoint, gatewayBaseURL string) string {
+	endpoint = strings.TrimSpace(endpoint)
+	parsed, err := url.Parse(endpoint)
+	if err != nil || parsed.Scheme == "" || parsed.Host == "" {
+		return endpoint
+	}
+	if !isDingTalkMCPGatewayHost(parsed.Hostname()) {
+		return endpoint
+	}
+	base, err := url.Parse(gatewayBaseURL)
+	if err != nil || base.Scheme == "" || base.Host == "" {
+		return endpoint
+	}
+	parsed.Scheme = base.Scheme
+	parsed.Host = base.Host
+	return strings.TrimRight(parsed.String(), "/")
+}
+
+func isDingTalkMCPGatewayHost(host string) bool {
+	switch strings.ToLower(strings.TrimSpace(host)) {
+	case "mcp-gw.dingtalk.com", "pre-mcp-gw.dingtalk.com", "mcp-gw.dingtalk.io", "pre-mcp-gw.dingtalk.io":
+		return true
+	default:
+		return false
+	}
+}
+
+func isDingTalkMCPGatewayEndpoint(endpoint string) bool {
+	parsed, err := url.Parse(strings.TrimSpace(endpoint))
+	return err == nil && isDingTalkMCPGatewayHost(parsed.Hostname())
+}
+
 // DirectRuntimeProductIDs returns product IDs that should stay visible for
 // direct runtime execution. Dynamic products come from MCP discovery/plugin
 // registration; built-in helper products such as devapp resolve their endpoint
@@ -61,7 +61,7 @@ func newDoctorCommand() *cobra.Command {
 	cmd := &cobra.Command{
 		Use:               "doctor",
 		Short:             "环境健康检查",
-		Long:              "一键检查登录态、网络连通性、缓存状态和版本更新，快速定位常见问题。",
+		Long:              "一键检查登录态、网络连通性和版本更新，快速定位常见问题。",
 		Args:              cobra.NoArgs,
 		DisableAutoGenTag: true,
 		RunE:              runDoctor,
@@ -92,9 +92,6 @@ func runDoctor(cmd *cobra.Command, _ []string) error {
 	networkResult := doctorCheckNetwork(cmd.Context(), w, jsonOut, networkTimeout)
 	checks = append(checks, networkResult)

-	cacheResult := doctorCheckCache(w, jsonOut)
-	checks = append(checks, cacheResult)
-
 	versionResult := doctorCheckVersion(w, jsonOut, networkTimeout)
 	checks = append(checks, versionResult)

@@ -297,24 +294,6 @@ func doctorCheckNetwork(ctx context.Context, w io.Writer, jsonOut bool, timeout
 	return r
 }

-// ── Cache check ─────────────────────────────────────────────────────────
-
-func doctorCheckCache(w io.Writer, jsonOut bool) checkResult {
-	if !jsonOut {
-		fmt.Fprint(w, tui.Dim("检查缓存状态...       "))
-	}
-
-	r := checkResult{
-		Name:    "cache",
-		Status:  statusPass,
-		Message: "静态端点模式, 无需缓存",
-	}
-	if !jsonOut {
-		printCheckResult(w, r)
-	}
-	return r
-}
-
 // ── Version check ───────────────────────────────────────────────────────

 func doctorCheckVersion(w io.Writer, jsonOut bool, timeout time.Duration) checkResult {
@@ -109,31 +109,6 @@ func TestPrintCheckResultNoHint(t *testing.T) {
 	}
 }

-func TestDoctorCheckCacheEmpty(t *testing.T) {
-	t.Setenv("DWS_CACHE_DIR", t.TempDir())
-
-	var buf bytes.Buffer
-	r := doctorCheckCache(&buf, false)
-
-	if r.Status != statusPass {
-		t.Errorf("expected pass for static endpoint mode, got %s", r.Status)
-	}
-}
-
-func TestDoctorCheckCacheEmptyJSON(t *testing.T) {
-	t.Setenv("DWS_CACHE_DIR", t.TempDir())
-
-	var buf bytes.Buffer
-	r := doctorCheckCache(&buf, true)
-
-	if r.Status != statusPass {
-		t.Errorf("expected pass for static endpoint mode, got %s", r.Status)
-	}
-	if buf.Len() != 0 {
-		t.Error("expected no output in JSON mode")
-	}
-}
-
 func TestDoctorCheckAuthReportsKeychainUnavailable(t *testing.T) {
 	t.Setenv("DWS_CONFIG_DIR", filepath.Join(t.TempDir(), "config"))

@@ -38,6 +38,7 @@ import (
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/event/bus"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/event/busctl"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/event/consume"
+	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/event/personal"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/event/registry"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/event/source"
 	"github.com/DingTalk-Real-AI/dingtalk-workspace-cli/internal/event/transport"
@@ -74,18 +75,18 @@ var (

 // newEventCommand returns the `event` parent command and all its subcommands.
 // Wired into root.go's utilityCommands list.
-func newEventCommand() *cobra.Command {
+func newEventCommand(globalFlags ...*GlobalFlags) *cobra.Command {
 	// Product-level Agent routing Decl (migrated from selection/event.json
 	// products.event). Catalog assembly stamps provenance contract_final.
 	contract.RegisterProductDecl(contract.ProductDecl{
 		ID: "event",
 		Selection: contract.ProductSelectionDecl{
-			AgentSummary: "订阅/消费个人消息、动作与群生命周期事件，并管理订阅生命周期",
+			AgentSummary: "实时监听当前用户相关的个人 IM 与 OA 审批事件，并管理订阅生命周期",
 			UseWhen: []string{
-				"需要实时监听个人消息接收、全量消息、已读、撤回、表情回应或群生命周期事件，或管理个人事件订阅生命周期",
+				"需要实时监听未来发生的个人消息、消息动作、群生命周期或 OA 审批任务/实例事件，或管理个人事件订阅生命周期",
 			},
 			AvoidWhen: []string{
-				"查历史聊天或主动发消息分别用 chat 查询/发送命令",
+				"查历史聊天或主动发消息用 chat；查询或处理审批实例/任务用 oa；配置开放平台应用事件回调用 dev app event",
 			},
 		},
 	})
@@ -99,11 +100,12 @@ func newEventCommand() *cobra.Command {
 		RunE:              func(c *cobra.Command, _ []string) error { return c.Help() },
 	}
 	cmd.AddCommand(
-		newEventConsumeCommand(),
+		newEventListenIMCommand(globalFlags...),
+		newEventConsumeCommand(globalFlags...),
 		newEventListCommand(),
 		newEventSchemaCommand(),
-		newEventStatusCommand(),
-		newEventStopCommand(),
+		newEventStatusCommandWithFlags(globalFlags...),
+		newEventStopCommandWithFlags(globalFlags...),
 		newEventBusCommand(),
 	)
 	return cmd
@@ -113,7 +115,7 @@ func newEventCommand() *cobra.Command {
 //  event consume
 // ─────────────────────────────────────────────────────────────────────

-func newEventConsumeCommand() *cobra.Command {
+func newEventConsumeCommand(globalFlags ...*GlobalFlags) *cobra.Command {
 	var (
 		eventTypes   []string
 		filter       string
@@ -169,6 +171,8 @@ SIGTERM、关 stdin，或先用 dws event stop <subscribe_id> --dry-run 预览
 				return err
 			}
 			if as == "user" {
+				personalOpts.ExplicitToken = eventExplicitToken(globalFlags)
+				personalOpts.ClientIDOverride = eventExplicitClientID(globalFlags)
 				personalOpts.EventKeys = dedupePersonalEventKeys(args)
 				personalOpts.EventKey = firstArg(personalOpts.EventKeys)
 				personalOpts.Flatten = flatten
@@ -332,7 +336,7 @@ SIGTERM、关 stdin，或先用 dws event stop <subscribe_id> --dry-run 预览
 	f.BoolVar(&force, "force", false,
 		"仅 --foreground 模式生效：跳过单实例锁 (慎用：会让云事件被随机切分)")
 	f.BoolVar(&dryRun, "dry-run", false,
-		"仅打印解析后的配置，不连接 bus / 云端")
+		"仅打印解析后的配置；不创建订阅、不连接 bus；复用 --subscribe-id 时会只读查询控制面")
 	f.BoolVar(&foreground, "foreground", false,
 		"当前进程直接跑 bus 服务、不 fork、不打印事件（给 systemd/k8s 托管用）；读事件不要用它")
 	f.StringVar(&personalOpts.SubscribeID, "subscribe-id", "",
@@ -397,22 +401,21 @@ SIGTERM、关 stdin，或先用 dws event stop <subscribe_id> --dry-run 预览
 				Reason:       "Reviewed composite workflow: the command creates or reuses a remote personal-event subscription and coordinates the local event bus and Stream consumer; no single pinned RPC represents the workflow.",
 			},
 			Selection: contract.SelectionSpec{
-				AgentSummary: "订阅并持续消费一个或多个兼容的个人事件；Agent 使用 --flatten 输出顶层业务 NDJSON",
+				AgentSummary: "消费 OA、群生命周期或需要底层控制的个人事件流；Agent 通常使用 --flatten 输出 NDJSON",
 				UseWhen: []string{
-					"需要实时监听 @我、指定单聊、指定群或指定发送人的后续消息事件",
-					"用户明确要求监听当前身份的所有单聊或所有群消息",
-					"需要监听指定单聊或群聊中的消息已读、撤回或表情回应事件",
+					"需要监听六个公开 OA 审批任务/实例 EventKey 中的一个或多个事件",
 					"需要监听指定群的标题变更、成员进退群或群解散事件",
-					"监听机器人、外部联系人等以 openDingtalkId 标识的单聊目标",
-					"同一目标、同一过滤条件需要同时监听多个兼容事件",
+					"用户显式给出原始 EventKey、Filter DSL、subscribe_id，要求原始 transport envelope，或需要普通 IM facade 不提供的高级多事件控制",
 				},
 				AvoidWhen: []string{
+					"普通 @我、指定发送人/群、全部单聊/群聊及 message/reaction/read/recall 监听优先使用 event +listen-im",
 					"只查历史聊天记录时用 chat 查询命令",
+					"查询、同意、拒绝、转交、撤销或发起审批时用 oa；配置应用事件回调时用 dev app event",
 					"只看事件目录/字段时用 event list / event schema",
 				},
 				Examples: []string{
-					"dws event consume user_im_message_receive_user --open-dingtalk-id open-example --flatten --max-events 1 --format ndjson",
-					"dws event consume user_im_message_receive_o2o user_im_message_read_o2o --user test-user-001 --flatten --max-events 2 --format ndjson",
+					"dws event consume user_oa_approval_task_created user_oa_approval_instance_finished --flatten --duration 10m --format ndjson",
+					"dws event consume user_im_group_member_added --group cid-example --flatten --max-events 1 --format ndjson",
 				},
 			},
 		},
@@ -563,6 +566,8 @@ func newEventBusCommand() *cobra.Command {
 		clientIDOverride string
 		idleTimeout      time.Duration
 		sourceKindRaw    string
+		runtimeTokenMode bool
+		identityHashFlag string
 		streamOpts       eventStreamTicketOptions
 	)
 	cmd := &cobra.Command{
@@ -599,23 +604,45 @@ func newEventBusCommand() *cobra.Command {
 				sourceKind = dwsevent.SourceKindAppStream
 			}
 			if sourceKind == dwsevent.SourceKindPersonalStream {
-				identity, err := eventResolvePersonal(ctx, configDir, streamOpts.SourceID)
-				if err != nil {
-					return failEarly(fmt.Errorf("event _bus: %w", err))
+				var (
+					identity     personal.Identity
+					identityHash string
+				)
+				if runtimeTokenMode {
+					identityHash = strings.TrimSpace(identityHashFlag)
+					if !validPersonalIdentityHash(identityHash) {
+						return failEarly(errors.New("event _bus: --identity-hash must be a 16-character hexadecimal identity hash in runtime token mode"))
+					}
+					if strings.TrimSpace(clientIDOverride) == "" {
+						return failEarly(errors.New("event _bus: --client-id is required in runtime token mode"))
+					}
+					identity = personal.Identity{
+						ClientID: strings.TrimSpace(clientIDOverride),
+						SourceID: personalEventStreamSourceID(streamOpts.SourceID),
+					}
+				} else {
+					var err error
+					identity, err = eventResolvePersonal(ctx, configDir, streamOpts.SourceID)
+					if err != nil {
+						return failEarly(fmt.Errorf("event _bus: %w", err))
+					}
+					if clientIDOverride != "" {
+						identity.ClientID = clientIDOverride
+					}
+					identityHash = dwsevent.IdentityHash(identity.Key())
 				}
-				if clientIDOverride != "" {
-					identity.ClientID = clientIDOverride
-				}
-				identityHash := dwsevent.IdentityHash(identity.Key())
 				editionName := editionNameOrDefault()
 				workDir := eventWorkDir(configDir, editionName, dwsevent.SourceKindPersonalStream, identityHash)
 				endpoint := defaultIPCEndpoint(workDir, editionName, dwsevent.SourceKindPersonalStream, identityHash)
+				credentialBroker := newPersonalCredentialBroker(configDir, runtimeTokenMode, runtimeTokenMode)
 				src, err := eventNewPersonalSource(ctx, personalStreamSourceOptions{
 					ConfigDir:        configDir,
 					Identity:         identity,
 					TicketMode:       streamOpts.Mode,
 					TicketURL:        streamOpts.TicketURL,
 					ClientIDOverride: clientIDOverride,
+					CredentialBroker: credentialBroker,
+					RuntimeTokenMode: runtimeTokenMode,
 				})
 				if err != nil {
 					return failEarly(err)
@@ -628,17 +655,18 @@ func newEventBusCommand() *cobra.Command {
 					}
 				}
 				busCfg := bus.Config{
-					WorkDir:      workDir,
-					IPCEndpoint:  endpoint,
-					ClientID:     identity.ClientID,
-					Edition:      editionName,
-					SourceKind:   dwsevent.SourceKindPersonalStream,
-					IdentityHash: identityHash,
-					SourceID:     identity.SourceID,
-					Source:       src,
-					IdleTimeout:  idleTimeout,
-					ReadyPipe:    readyPipe,
-					Logger:       slog.Default(),
+					WorkDir:          workDir,
+					IPCEndpoint:      endpoint,
+					ClientID:         identity.ClientID,
+					Edition:          editionName,
+					SourceKind:       dwsevent.SourceKindPersonalStream,
+					IdentityHash:     identityHash,
+					SourceID:         identity.SourceID,
+					Source:           src,
+					IdleTimeout:      idleTimeout,
+					ReadyPipe:        readyPipe,
+					Logger:           slog.Default(),
+					CredentialBroker: credentialBroker,
 				}
 				bus.ApplyEnvTuning(&busCfg)
 				return eventBusRun(ctx, busCfg)
@@ -698,12 +726,18 @@ func newEventBusCommand() *cobra.Command {
 		"exit after this long with zero consumers (0 = disabled)")
 	cmd.Flags().StringVar(&sourceKindRaw, "source-kind", string(dwsevent.SourceKindAppStream),
 		"event source kind: app_stream|personal_stream")
+	cmd.Flags().BoolVar(&runtimeTokenMode, "runtime-token-mode", false,
+		"use an owner-injected in-memory runtime credential")
+	cmd.Flags().StringVar(&identityHashFlag, "identity-hash", "",
+		"pre-resolved non-sensitive personal identity hash")
 	cmd.Flags().StringVar(&streamOpts.Mode, "stream-ticket-mode", strings.TrimSpace(os.Getenv("DWS_STREAM_TICKET_MODE")),
 		"用户 Stream 建联模式：空=SDK app credential；normal/custom=portal 取票")
 	cmd.Flags().StringVar(&streamOpts.SourceID, "stream-source-id", strings.TrimSpace(os.Getenv("DWS_STREAM_SOURCE_ID")),
 		"用户 Stream sourceId；personal_stream 开源版默认 open")
 	cmd.Flags().StringVar(&streamOpts.TicketURL, "stream-ticket-url", strings.TrimSpace(os.Getenv("DWS_STREAM_TICKET_URL")),
 		"用户 Stream 取票 URL；personal_stream 默认由 MCP base URL 派生")
+	_ = cmd.Flags().MarkHidden("runtime-token-mode")
+	_ = cmd.Flags().MarkHidden("identity-hash")
 	return cmd
 }

@@ -822,6 +856,10 @@ func newEventListCommand() *cobra.Command {
 // ─────────────────────────────────────────────────────────────────────

 func newEventStatusCommand() *cobra.Command {
+	return newEventStatusCommandWithFlags()
+}
+
+func newEventStatusCommandWithFlags(globalFlags ...*GlobalFlags) *cobra.Command {
 	var (
 		all          bool
 		allEditions  bool
@@ -847,6 +885,8 @@ func newEventStatusCommand() *cobra.Command {
 					return fmt.Errorf("event status: %w", err)
 				}
 				personalOpts.Format = formatRaw
+				personalOpts.ExplicitToken = eventExplicitToken(globalFlags)
+				personalOpts.ClientIDOverride = eventExplicitClientID(globalFlags)
 				return eventRunPersonalStatus(c, personalOpts)
 			}
 			if err := rejectChangedFlags(c, "user", "event", "status", "subscribe-id", "personal-event-base-url", "stream-source-id"); err != nil {
@@ -1138,6 +1178,10 @@ func renderStatusBlock(w io.Writer, qs busctl.EntryStatus) {
 }

 func newEventStopCommand() *cobra.Command {
+	return newEventStopCommandWithFlags()
+}
+
+func newEventStopCommandWithFlags(globalFlags ...*GlobalFlags) *cobra.Command {
 	var asIdentity string
 	var opts personalStopOptions
 	cmd := &cobra.Command{
@@ -1158,6 +1202,8 @@ func newEventStopCommand() *cobra.Command {
 			}
 			if as == "user" {
 				opts.SubscribeID = firstArg(args)
+				opts.ExplicitToken = eventExplicitToken(globalFlags)
+				opts.ClientIDOverride = eventExplicitClientID(globalFlags)
 				if eventStopDryRun(c) {
 					return writeEventStopDryRun(c, as, opts)
 				}
@@ -1261,6 +1307,20 @@ func eventStopDryRun(cmd *cobra.Command) bool {
 	return value
 }

+func eventExplicitToken(globalFlags []*GlobalFlags) string {
+	if len(globalFlags) == 0 || globalFlags[0] == nil {
+		return ""
+	}
+	return strings.TrimSpace(globalFlags[0].Token)
+}
+
+func eventExplicitClientID(globalFlags []*GlobalFlags) string {
+	if len(globalFlags) == 0 || globalFlags[0] == nil {
+		return ""
+	}
+	return strings.TrimSpace(globalFlags[0].ClientID)
+}
+
 func writeEventStopDryRun(cmd *cobra.Command, identity string, opts personalStopOptions) error {
 	payload := map[string]any{
 		"dry_run":  true,
--- a/Show More
+++ b/Show More