@topmindspace/topmind-skills 4.15.8 → 4.15.10

package/CHANGELOG.md

@@ -4,6 +4,17 @@ topmind-skills 的版本记录。GitHub 只保留最近 2 个 Release，本文
 
 版本真源：`topmind-pack.json`。`package.json`、每个 `SKILL.md` 的 `metadata.version`、`evals/evals.json` 与之一致。
 
+## 4.15.10 — 2026-10-31
+
+- 命令面与引擎对齐为 9 域 / 42 条：补上 `list-category-files`、待办的更新、截止日和删除，以及只读 `creator`（inspect、snapshot、matchLengthReport）。`creator` 不是第九引擎，也不写工作区。
+- 授权与建议改为方向优先：用工作区里已有的笔记和结论判断，缺少的知识先研究。耐久经验只在用户明确要求或接受建议后写入。入口不再复制第二套阶梯。
+
+## 4.15.9 — 2026-10-28
+
+- 写作路由按用户意图直达独立平台技能；通用工作区稿件才走 `topmind-write`，发布动作由对应连接器处理。
+- 安装升级固定在指定项目、全局或 `--dest` 目录；单目录回执不再误判为多宿主安装，多宿主复制安装升级后仍保持复制方式。
+- 技能包先暂存再替换，失败恢复旧文件；拒绝无关同名技能或链接、非法清单路径、自安装与冲突的共享资源；在写入前检查多宿主目录，避免链接别名造成自引用或半完成安装。
+
 ## 4.15.8 — 2026-10-26
 
 - 明确写作包独立发布、Desktop/Obsidian 默认内置，其他 Agent 按需安装；发现宿主能力后路由，不凭核心目录误判。

package/INSTALL.md

@@ -14,9 +14,12 @@ npm：[`@topmindspace/topmind-skills`](https://www.npmjs.com/package/@topmindspa
 # 推荐：装到全局，自动给检测到的助手建链接
 npx @topmindspace/topmind-skills add topmindspace/topmind-skills -g
 
-# 升级
+# 升级全局安装
 npx @topmindspace/topmind-skills update -g
 
+# 升级当前项目安装
+npx @topmindspace/topmind-skills update -p
+
 # 从 GitHub Release 安装（固定版本）
 npx @topmindspace/topmind-skills add release:latest -g
 ```
@@ -70,7 +73,7 @@ npx @topmindspace/topmind-skills paths -g
 npx @topmindspace/topmind-skills doctor --repair
 ```
 
-升级规则：包里版本更新时直接覆盖，不留备份，打印 `Upgraded <技能> <旧> -> <新>`；版本相同跳过，打印 `Already up to date`（加 `--force` 重装）；已装版本比包里新时不降级，加 `--force` 才降级。
+升级规则：`update -g` 只读取全局安装记录，`update -p` 只读取当前项目的安装记录；用 `--dest <目录>` 安装的，请用 `update --dest <同一目录>`。安装器沿用记录中的来源、目标助手与复制/链接方式，不会猜测另一个目录。新版本先完整暂存再替换；失败时恢复旧文件。版本相同跳过并提示 `Already up to date`（加 `--force` 重装）；已装版本比包里新时不降级，加 `--force` 才降级。
 
 私有仓库：先 `export GH_TOKEN=…`（或 `GITHUB_TOKEN`）再执行 `add`。
 

package/README.en.md

@@ -16,7 +16,8 @@ Then say "topmind" or just ask for what you need. To upgrade:
 npx @topmindspace/topmind-skills update -g
 ```
 
-- A newer packaged skill overwrites the installed copy and prints `Upgraded <skill> <old> -> <new>`; the same version prints `Already up to date`; a newer installed version is kept unless you pass `--force`.
+- Use `update -p` for a project install, or `update --dest <same-directory>` for a custom directory. Updates stay within the selected installation scope.
+- A new version is staged before replacement; a failed upgrade restores the prior files. The same version prints `Already up to date`; a newer installed version is kept unless you pass `--force`.
 - To install for some assistants only, add `-a claude-code -a codex` to `add`.
 - The community CLI `npx skills add topmindspace/topmind-skills -g -y` also works, but it leaves out the `shared/` folder, so some rules cannot be read. See [INSTALL.md](./INSTALL.md#社区-cli-npx-skills) for the extra step.
 - Pinned or offline installs, project scope and per-assistant folders are covered in [INSTALL.md](./INSTALL.md).
@@ -43,7 +44,7 @@ The pack has 10 skills: the `topmind` entry, 6 actions, 2 optional connectors an
 
 ### Companion skills (installed separately)
 
-These are not part of this pack. When installed, `topmind` hands matching requests to them; otherwise it falls back to `topmind-write` or explains in the receipt.
+These are not part of this pack. Explicit WeChat, X article/post, cover, or long-image requests go straight to an installed specialist skill. Without one, the assistant completes the feasible general draft and explains any missing delivery capability.
 
 | Need | Skill | Repository |
 |---|---|---|
@@ -79,7 +80,7 @@ topmind-workspace/
 - **Save settings**: `writeback_mode: auto | confirm`. auto writes and returns a receipt; confirm also saves content edits directly and queues deletion/archive for review. Explicitly requested topic or memory changes reuse your authorization; inferred suggestions need acceptance; see [`shared/auto-suggest.md`](./shared/auto-suggest.md).
 - **Files are the only source of truth**: assistant session state must not become topmind content truth.
 - **Desktop is not required, UTR is optional**: without topmind Desktop or the UTR CLI, the assistant uses its own file tools under the same conventions.
-- **Nothing is posted for you**: skills only draft; you publish.
+- **Drafting and posting are separate**: writing skills only prepare drafts. An available connector handles posting when you explicitly request it, subject to its permissions.
 
 ## For maintainers
 

package/README.md

@@ -16,7 +16,8 @@ npx @topmindspace/topmind-skills add topmindspace/topmind-skills -g
 npx @topmindspace/topmind-skills update -g
 ```
 
-- 包里版本更新时直接覆盖，打印 `Upgraded <技能> <旧> -> <新>`；版本相同打印 `Already up to date`；已装版本更新时不降级，要降级加 `--force`。
+- 项目安装用 `update -p`，指定目录安装用 `update --dest <同一目录>`；升级只读对应安装记录，不跳到别的工作区。
+- 新版本先暂存再替换，失败恢复旧文件；版本相同提示 `Already up to date`，已装版本更新时不降级，要降级加 `--force`。
 - 只装给部分助手：在 `add` 后面加 `-a claude-code -a codex`。
 - 社区 CLI `npx skills add topmindspace/topmind-skills -g -y` 也能装，但不带 `shared/` 目录，部分规则会读不到，装完要补一步，见 [INSTALL.md](./INSTALL.md#社区-cli-npx-skills)。
 - 固定版本或离线安装、项目级安装、各助手的目录位置，都在 [INSTALL.md](./INSTALL.md)。
@@ -43,7 +44,7 @@ npx @topmindspace/topmind-skills update -g
 
 ### 配套技能（单独安装）
 
-下面这些不随本包发布，装了以后 `topmind` 会把对应请求转过去，没装时退回 `topmind-write` 或在回执里说明。
+下面这些不随本包发布。明确的公众号、X 长文/短帖、封面和长图需求会直达已安装的专用技能；没装时完成可行的通用稿，说明缺少的交付能力。
 
 | 需求 | 技能 | 仓库 |
 |---|---|---|
@@ -79,7 +80,7 @@ topmind-workspace/
 - **保存设置**：`writeback_mode: auto | confirm`。auto 直接写并给回执；confirm 内容编辑也直接写，仅删除/归档进入审阅。明确要求的建专题、记忆更新直接执行；推测的建议需接受，规则见 [`shared/auto-suggest.md`](./shared/auto-suggest.md)。
 - **文件是唯一的数据来源**：助手自己的会话状态不得成为 topmind 内容真源。
 - **不要求 Desktop，UTR 可选**：topmind Desktop 和 UTR 命令行都不是必需的，没有时用助手自带的文件工具，遵守同一套约定。
-- **不代发**：任何技能都只起草，发布由你操作。
+- **写稿与发布分开**：写作技能只出稿；你明确要求发布时，由实际可用的连接器按其权限执行。
 
 ## 给维护者
 

package/bin/install-skills.mjs

@@ -191,6 +191,15 @@ function samePath(a, b) {
   return path.resolve(a) === path.resolve(b);
 }
 
+async function sameExistingDirectory(a, b) {
+  if (samePath(a, b)) return true;
+  try {
+    return (await fs.realpath(a)) === (await fs.realpath(b));
+  } catch {
+    return false;
+  }
+}
+
 function detectedAgentIds() {
   return Object.values(agentCatalog())
     .filter((a) => a.detect())
@@ -636,6 +645,8 @@ async function listInstallEntries(skillsRoot, skillsFilter) {
     version = pack.version || version;
     name = pack.name || name;
     ids = (pack.skills || []).map((s) => s.id || s.path).filter(Boolean);
+    const invalid = ids.find((id) => typeof id !== "string" || !/^[a-z0-9]+(?:-[a-z0-9]+)*$/u.test(id));
+    if (invalid) fail(`invalid skill id in pack: ${JSON.stringify(invalid)}`);
   } else {
     // bare Agent Skills layout (external pack, e.g. topmind-writing-skills)
     try {
@@ -735,10 +746,17 @@ function compareVersions(a, b) {
   return 0;
 }
 
-/** Existing non-symlink dst is a previous install of this skill (same SKILL.md name) or an installer-owned file. */
+/** Match a previous skill by identity, or a link already pointing at this entry. */
 async function isPreviousInstall(src, dst) {
   const st = await fs.lstat(dst).catch(() => null);
   if (!st) return false;
+  if (st.isSymbolicLink()) {
+    const link = await fs.readlink(dst);
+    if (samePath(path.resolve(path.dirname(dst), link), src)) return true;
+    const a = await readSkillMeta(src);
+    const b = await readSkillMeta(dst);
+    return !!(a && b && a.name && a.name === b.name);
+  }
   const sst = await fs.lstat(src).catch(() => null);
   if (sst && sst.isFile() && st.isFile()) return true;
   if (!st.isDirectory()) return false;
@@ -747,17 +765,49 @@ async function isPreviousInstall(src, dst) {
   return !!(a && b && a.name && a.name === b.name);
 }
 
-async function installCopy(skillsRoot, dest, entries, locale, { force = false } = {}) {
+async function assertSharedNoConflict(sourceDir, targetDir, relative = "") {
+  for (const entry of await fs.readdir(sourceDir, { withFileTypes: true })) {
+    const rel = path.join(relative, entry.name);
+    const src = path.join(sourceDir, entry.name);
+    const dst = path.join(targetDir, entry.name);
+    const existing = await fs.lstat(dst).catch(() => null);
+    if (entry.isDirectory()) {
+      if (existing && !existing.isDirectory()) fail(`shared resource conflict: ${rel} is not a directory`);
+      if (existing) await assertSharedNoConflict(src, dst, rel);
+    } else if (!entry.isFile() || (existing && (!existing.isFile() || !(await fs.readFile(src)).equals(await fs.readFile(dst))))) {
+      fail(`shared resource conflict: ${rel} differs from an installed pack; pass --force to replace it`);
+    }
+  }
+}
+
+async function installCopy(skillsRoot, dest, entries, locale, { force = false, topmindPack = false } = {}) {
   await fs.mkdir(dest, { recursive: true });
-  // Pre-pass: never silently downgrade a skill that is newer than the packaged one.
+  // Validate the whole plan before touching an installed entry.
   const plan = new Map();
   const newer = [];
+  const destReal = await fs.realpath(dest);
   for (const name of entries) {
     const src = path.join(skillsRoot, name);
     const dst = path.join(dest, name);
+    if (!(await pathExists(src))) continue;
+    const srcReal = await fs.realpath(src);
+    const target = path.join(destReal, name);
+    if (target === srcReal || target.startsWith(srcReal + path.sep)) {
+      fail(`refusing to install ${name} into its own source tree: ${target}`);
+    }
     const sm = await readSkillMeta(src);
+    if (name === "shared" && !force) {
+      const installed = await fs.lstat(dst).catch(() => null);
+      if (installed && !installed.isDirectory()) {
+        fail(`${dst} is not a directory owned by this installation; pass --force to replace it`);
+      }
+    }
     if (!sm) continue;
-    const dm = (await fs.lstat(dst).catch(() => null))?.isSymbolicLink() ? null : await readSkillMeta(dst);
+    const existing = await fs.lstat(dst).catch(() => null);
+    const dm = existing ? await readSkillMeta(dst) : null;
+    if (existing && (!dm || dm.name !== sm.name) && !force) {
+      fail(`${dst} is not an installed ${sm.name} skill — pass --force to replace it`);
+    }
     if (!dm) continue;
     const cmp = compareVersions(sm.version, dm.version);
     plan.set(name, { cmp, old: dm.version || "unknown", now: sm.version || "unknown" });
@@ -766,50 +816,100 @@ async function installCopy(skillsRoot, dest, entries, locale, { force = false }
   if (newer.length && !force) {
     fail(`installed version is newer than packaged (${newer.join(", ")}); not downgrading — pass --force to downgrade`);
   }
-  for (const name of entries) {
-    const src = path.join(skillsRoot, name);
-    const dst = path.join(dest, name);
-    if (!(await pathExists(src))) continue;
-    const pv = plan.get(name);
-    if (pv && pv.cmp === 0 && !force) {
-      log(`  = Already up to date: ${name} ${pv.now}`);
-      continue;
-    }
-    const st = await fs.lstat(src);
-    if (await pathExists(dst)) await fs.rm(dst, { recursive: true, force: true });
-    if (st.isDirectory()) {
-      await fs.cp(src, dst, { recursive: true });
-      // Apply locale overlay if exists
-      if (locale) {
-        const overlayDir = path.join(src, "locales", locale);
-        if (await pathExists(overlayDir)) {
-          await applyLocaleOverlay(overlayDir, dst);
-          log(`    locale: ${locale}`);
+  const staging = await fs.mkdtemp(path.join(dest, ".topmind-skills-stage-"));
+  const changes = [];
+  let keepStaging = false;
+  try {
+    // Stage every changed entry first. Preserve sibling packs' shared resources.
+    for (const name of entries) {
+      const src = path.join(skillsRoot, name);
+      if (!(await pathExists(src))) continue;
+      const pv = plan.get(name);
+      if (pv && pv.cmp === 0 && !force) continue;
+      const dst = path.join(dest, name);
+      const next = path.join(staging, "next", name);
+      const backup = path.join(staging, "previous", name);
+      await fs.mkdir(path.dirname(next), { recursive: true });
+      const st = await fs.lstat(src);
+      if (st.isDirectory()) {
+        if (name === "shared" && (await fs.lstat(dst).catch(() => null))?.isDirectory()) {
+          if (!force && !topmindPack) await assertSharedNoConflict(src, dst);
+          await fs.cp(dst, next, { recursive: true });
+          await fs.cp(src, next, { recursive: true, force: true });
+        } else {
+          await fs.cp(src, next, { recursive: true });
         }
-        // Clean up locales/ directory from installed copy (not needed at runtime)
-        const localesInstalledDir = path.join(dst, "locales");
-        if (await pathExists(localesInstalledDir)) {
-          await fs.rm(localesInstalledDir, { recursive: true, force: true });
+        if (locale) {
+          const overlayDir = path.join(src, "locales", locale);
+          if (await pathExists(overlayDir)) await applyLocaleOverlay(overlayDir, next);
+          await fs.rm(path.join(next, "locales"), { recursive: true, force: true });
         }
+      } else {
+        await fs.copyFile(src, next);
       }
-    } else {
-      await fs.copyFile(src, dst);
+      changes.push({ name, dst, next, backup, pv, backedUp: false, installed: false });
+    }
+
+    for (const item of changes) {
+      if (await fs.lstat(item.dst).catch(() => null)) {
+        await fs.mkdir(path.dirname(item.backup), { recursive: true });
+        await fs.rename(item.dst, item.backup);
+        item.backedUp = true;
+      }
+      await fs.rename(item.next, item.dst);
+      item.installed = true;
+    }
+  } catch (error) {
+    let rollbackError = null;
+    for (const item of [...changes].reverse()) {
+      try {
+        if (item.installed) await fs.rm(item.dst, { recursive: true, force: true });
+        if (item.backedUp) await fs.rename(item.backup, item.dst);
+      } catch (restoreError) {
+        rollbackError = restoreError;
+        keepStaging = true;
+      }
+    }
+    if (rollbackError) {
+      throw new Error(`install failed and rollback needs attention; previous files remain at ${staging}: ${rollbackError.message}`, { cause: error });
+    }
+    throw error;
+  } finally {
+    if (!keepStaging) {
+      await fs.rm(staging, { recursive: true, force: true }).catch((error) => {
+        log(`warning: temporary install files remain at ${staging}: ${error.message}`);
+      });
+    }
+  }
+  for (const name of entries) {
+    const pv = plan.get(name);
+    if (pv && pv.cmp === 0 && !force) log(`  = Already up to date: ${name} ${pv.now}`);
+    else if (changes.some((item) => item.name === name)) {
+      if (pv && pv.cmp > 0) log(`  ↑ Upgraded ${name} ${pv.old} -> ${pv.now}`);
+      else if (pv) log(`  + Reinstalled ${name} ${pv.now} (--force)`);
+      else log(`  + ${name}`);
     }
-    if (pv && pv.cmp > 0) log(`  ↑ Upgraded ${name} ${pv.old} -> ${pv.now}`);
-    else if (pv) log(`  + Reinstalled ${name} ${pv.now} (--force)`);
-    else log(`  + ${name}`);
   }
 }
 
 async function installSymlink(skillsRoot, dest, entries, { force }) {
   await fs.mkdir(dest, { recursive: true });
+  const destReal = await fs.realpath(dest);
+  for (const name of entries) {
+    const src = path.join(skillsRoot, name);
+    if (!(await pathExists(src))) continue;
+    const srcReal = await fs.realpath(src);
+    const target = path.join(destReal, name);
+    if (target === srcReal || target.startsWith(srcReal + path.sep)) {
+      fail(`refusing to install ${name} into its own source tree: ${target}`);
+    }
+  }
   for (const name of entries) {
     const src = path.join(skillsRoot, name);
     const dst = path.join(dest, name);
     if (!(await pathExists(src))) continue;
-    if (await pathExists(dst)) {
-      const st = await fs.lstat(dst);
-      if (st.isSymbolicLink() || force || (await isPreviousInstall(src, dst))) await fs.rm(dst, { recursive: true, force: true });
+    if (await fs.lstat(dst).catch(() => null)) {
+      if (force || (await isPreviousInstall(src, dst))) await fs.rm(dst, { recursive: true, force: true });
       else fail(`${dst} is an unrelated folder (no matching SKILL.md name) — pass --force to replace it`);
     }
     await fs.symlink(src, dst, "junction");
@@ -834,6 +934,7 @@ async function writeInstallReceipt(dest, meta) {
     source_raw: meta.sourceRaw,
     repo_path: meta.repoPath,
     mode: meta.mode,
+    install_kind: meta.installKind || "single",
     host: meta.host || null,
     scope: meta.scope || null,
     agents: meta.agents || null,
@@ -856,6 +957,9 @@ async function writeInstallReceipt(dest, meta) {
 /** Relative-symlink each pack entry from an agent root back to the canonical body. */
 async function linkPackEntries(canonicalRoot, agentRoot, entryNames, { force }) {
   await fs.mkdir(agentRoot, { recursive: true });
+  // A host's entire skills directory may alias the canonical directory.
+  // Never replace an entry through that alias with a link to itself.
+  if (await sameExistingDirectory(agentRoot, canonicalRoot)) return;
   for (const name of entryNames) {
     const src = path.join(canonicalRoot, name);
     const dst = path.join(agentRoot, name);
@@ -863,11 +967,9 @@ async function linkPackEntries(canonicalRoot, agentRoot, entryNames, { force })
     if (samePath(agentRoot, canonicalRoot)) continue;
     if ((await pathExists(dst)) || (await isBrokenSymlink(dst))) {
       const st = await fs.lstat(dst).catch(() => null);
-      if (st && st.isSymbolicLink()) await fs.rm(dst, { recursive: true, force: true });
-      else if (force || (await isPreviousInstall(src, dst))) await fs.rm(dst, { recursive: true, force: true });
+      if (st && (force || (await isPreviousInstall(src, dst)))) await fs.rm(dst, { recursive: true, force: true });
       else {
-        log(`  ! ${dst} is an unrelated folder (no matching SKILL.md name) — skipped; pass --force to replace it`);
-        continue;
+        fail(`${dst} changed during install or is an unrelated entry — pass --force to replace it`);
       }
     }
     const rel = path.relative(agentRoot, src);
@@ -914,8 +1016,37 @@ async function installViaMatrix(skillsRoot, plan, opts) {
     return { canonical, targets, plan };
   }
 
+  await fs.mkdir(canonical, { recursive: true });
+  for (const t of targets) {
+    if (!samePath(t.root, canonical)) {
+      try {
+        await fs.mkdir(t.root, { recursive: true });
+        const stat = await fs.stat(t.root);
+        if (!stat.isDirectory()) fail(`${t.root} is not a skills directory`);
+      } catch (error) {
+        fail(`cannot use agent skills directory ${t.root}: ${error.message}`);
+      }
+    }
+    t.sharesCanonical = await sameExistingDirectory(t.root, canonical);
+  }
+
+  if (linkMode === "symlink" && !opts.force) {
+    for (const t of targets) {
+      if (t.sharesCanonical) continue;
+      for (const name of plan.entries) {
+        const dst = path.join(t.root, name);
+        if (!(await fs.lstat(dst).catch(() => null))) continue;
+        const sourceEntry = path.join(skillsRoot, name);
+        const canonicalEntry = path.join(canonical, name);
+        if (!(await isPreviousInstall(sourceEntry, dst)) && !(await isPreviousInstall(canonicalEntry, dst))) {
+          fail(`${dst} is an unrelated entry — pass --force to replace it`);
+        }
+      }
+    }
+  }
+
   // 1) Full pack body → canonical
-  await installCopy(skillsRoot, canonical, plan.entries, opts.locale, opts);
+  await installCopy(skillsRoot, canonical, plan.entries, opts.locale, { ...opts, topmindPack: plan.topmindPack });
 
   // 2) Private agent dirs: symlink entries (default) or copy
   for (const t of targets) {
@@ -924,7 +1055,7 @@ async function installViaMatrix(skillsRoot, plan, opts) {
       continue;
     }
     if (linkMode === "copy") {
-      await installCopy(skillsRoot, t.root, plan.entries, opts.locale, opts);
+      await installCopy(skillsRoot, t.root, plan.entries, opts.locale, { ...opts, topmindPack: plan.topmindPack });
     } else {
       await linkPackEntries(canonical, t.root, plan.entries, opts);
     }
@@ -937,6 +1068,7 @@ async function installViaMatrix(skillsRoot, plan, opts) {
     sourceRaw: opts.source,
     repoPath: opts.repoPath,
     mode: linkMode,
+    installKind: "matrix",
     host: null,
     scope: opts.scope,
     agents: targets.map((t) => t.agent.id),
@@ -948,7 +1080,7 @@ async function installViaMatrix(skillsRoot, plan, opts) {
   log(`done → ${canonical}`);
   if (plan.topmindPack) {
     log(`daily entry: ${path.join(canonical, "topmind", "SKILL.md")}`);
-    log(`update later: node bin/install-skills.mjs update -g`);
+    log(`update later: npx @topmindspace/topmind-skills update ${global ? "-g" : "-p"}`);
   } else {
     log(`update later: re-run add ${opts.source || "<source>"} (update only follows the topmind pack receipt)`);
   }
@@ -1084,7 +1216,7 @@ async function cmdAdd(opts, sourceOverride) {
       }
       await installSymlink(skillsRoot, dest, plan.entries, opts);
     } else {
-      await installCopy(skillsRoot, dest, plan.entries, opts.locale, opts);
+      await installCopy(skillsRoot, dest, plan.entries, opts.locale, { ...opts, topmindPack: plan.topmindPack });
     }
 
     await writeInstallReceipt(dest, {
@@ -1096,7 +1228,8 @@ async function cmdAdd(opts, sourceOverride) {
       mode: opts.mode,
       host: opts.host,
       scope: opts.scope,
-      agents: opts.agents,
+      agents: null,
+      installKind: "single",
       skillIds: plan.skillIds,
       locale: opts.locale,
       topmindPack: plan.topmindPack,
@@ -1115,10 +1248,8 @@ async function cmdAdd(opts, sourceOverride) {
 }
 
 async function cmdUpdate(opts) {
-  // Prefer receipt at matrix canonical; fall back to explicit --dest receipt.
-  const destCandidates = opts.dest
-    ? [opts.dest]
-    : [canonicalBase(opts.global), canonicalBase(!opts.global), HOST_DEFAULTS["claude-code"]()];
+  // A scope is an instruction, not a search hint: never update another scope.
+  const destCandidates = [opts.dest || canonicalBase(opts.global)];
   let dest = null;
   let receipt = null;
   for (const d of destCandidates) {
@@ -1131,8 +1262,8 @@ async function cmdUpdate(opts) {
   if (!receipt) {
     fail(
       `no ${RECEIPT_NAME} in ${destCandidates.join(", ")}.\n` +
-        `  First install: node bin/install-skills.mjs add ${DEFAULT_OWNER_REPO} -g\n` +
-        `  Or community:  npx skills update -g -y`,
+        `  First install: npx @topmindspace/topmind-skills add ${DEFAULT_OWNER_REPO} ${opts.dest ? "--dest <same directory>" : opts.global ? "-g" : "-p"}\n` +
+        `  To update another scope, pass -g, -p or --dest explicitly.`,
     );
   }
   log(`update from receipt: ${receipt.source} @ ${receipt.version} → ${dest}`);
@@ -1142,15 +1273,17 @@ async function cmdUpdate(opts) {
     source: receipt.source_raw || receipt.update?.source || DEFAULT_OWNER_REPO,
     repoPath: receipt.repo_path || receipt.update?.path || DEFAULT_PATH,
     mode: receipt.mode === "symlink" ? "symlink" : "copy",
-    linkMode: receipt.agents ? "symlink" : opts.linkMode,
+    linkMode: receipt.mode === "copy" ? "copy" : "symlink",
     host: receipt.host || opts.host,
     scope: receipt.scope || opts.scope,
-    agents: receipt.agents || opts.agents,
+    agents: Array.isArray(receipt.agents) ? receipt.agents : opts.agents,
     locale: receipt.locale || opts.locale,
   };
-  // Keep matrix mode when the receipt was a matrix install (no single dest host).
-  if (receipt.agents) next.dest = null;
-  else if (!next.dest) next.dest = dest;
+  // Legacy matrix receipts used source="matrix"; old single-dir receipts also
+  // contained agents: [], so array presence alone cannot identify the mode.
+  const matrix = receipt.install_kind === "matrix" ||
+    (!receipt.install_kind && receipt.source === "matrix");
+  next.dest = matrix ? null : dest;
 
   if (next.mode === "symlink" && next.dest) {
     const local = parseSource(next.source, next.repoPath);

package/evals/evals.json

@@ -1,7 +1,7 @@
 {
   "skill_name": "topmind",
   "daily_entry": "topmind",
-  "version": "4.15.8",
+  "version": "4.15.10",
   "evals": [
     {
       "id": 1,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@topmindspace/topmind-skills",
-  "version": "4.15.8",
+  "version": "4.15.10",
   "description": "topmind Agent Skills pack: one daily entry (topmind) routing to capture, organize, write, memory, maintain and loop, plus optional WeRead, X and ledger skills.",
   "type": "module",
   "license": "MIT",

package/shared/auto-suggest.md

@@ -1,5 +1,9 @@
 # 请求授权与主动建议
 
+**方向大于约束**：引导、目标、已有经验与知识经验优先于固定操作剧本；规约只保留底线。用户概念仍 ≤5（记一下 · 动态 · 专题 · 我的情况 · 交付），文件仍是内容真源，不把「沉淀」教成产品 UI 里的第六个动词。硬闸仍在：写回分级、路径围栏、外部发布确认，以及禁止只做口头改写。
+
+按用户目标行动。用工作区里已有的笔记、记忆和结论做判断；缺的知识先研究分析再决定。耐久的知识经验只在用户明确要求或接受建议后写入。推测出的分类、记忆、专题仍只提案。明确请求不重复询问。
+
 先完成用户请求，再按价值决定是否建议延伸。用户不需要学习模块名，不为分类、待办、记忆弹一组选择题。
 
 ## 三种情况

package/shared/capability-degradation.md

@@ -42,14 +42,15 @@ Level 3（最低）: 仅对话
 
 | 需要 | 命令 |
 |------|------|
-| 类别 / 专题 | `list-categories`（WorkspaceModel Descriptor）· `list-topics` · `inspect-topic` · `list-topic-files` · `list-inbox` |
+| 类别 / 专题 / 大类根稿 | `list-categories` · `list-category-files` · `list-topics` · `inspect-topic` · `list-topic-files` · `list-inbox` |
 | 捕获 / 创建 | `create-topic` · `capture-note` · `save-output` |
-| 待办 | `memory.list-todos` · `memory.add-todo` · `memory.toggle-todo` |
-| 记忆 | `memory.append-profile` · `memory.append-topic` · `memory.promote` · `memory.digest` · `memory.retire-profile` · `memory.update-profile` · `memory.compact-history` · `memory.restore-profile` |
-| Inbox 路由 | `plan-inbox-routing`（primary；整理 inbox 前规划） |
+| 待办 | `memory.list-todos` · `memory.add-todo` · `memory.toggle-todo` · `memory.update-todo` · `memory.set-todo-due` · `memory.delete-todo` |
+| 记忆 | `memory.promote` · `memory.digest` · `memory.append-profile` · `memory.append-topic` · `memory.retire-profile` · `memory.update-profile` · `memory.compact-history` · `memory.restore-profile` |
+| 创作只读 | `creator.inspect` · `creator.snapshot` · `creator.matchLengthReport` |
+| Inbox 路由 | `plan-inbox-routing` |
 | 检查 / 维护 | `doctor-workspace` · `archive-topic` · `archive-stream-year` · `restore-safety-receipt` · `contract.validate` · `contract.reseed` |
 
-扩展面见 `TOOLS.md` §Current Command Surface。输入用独立字段 `category` + `topic`。
+上表覆盖注册表里的全部 primary，含 `list-category-files` 与 `creator`。`creator` 是只读投影，不是引擎，也不写工作区。advanced 与 MCP 默认数量不在这里另抄，以引擎仓 `TOOLS.md` 与 `utr/core/contract-registry.mjs` 为准。输入用独立字段 `category` + `topic`。
 
 ## Level 3 — 对话 only
 

package/shared/trigger-disambiguation.md

@@ -8,7 +8,7 @@
 > - 独立技能：`topmind-research`（对外检索与研究报告）、`topmind-presentation`（HTML / PPTX 演示稿）、`topmind-handoff`（跨工具交接包）
 > - 第三方：`last30days`、`TopStream每日精选`
 >
-> 任何技能都不代用户发布：写稿技能只出稿；要发出去走对应连接器（如 `topmind-x`），并且每次都须用户确认。
+> 写稿技能只出稿；明确要发出去时走对应连接器（如 `topmind-x`），遵守其发布前确认与实际权限，不能把出稿说成已发布。
 
 ## 碰撞词裁决
 
@@ -26,7 +26,7 @@
 | 整理 | organize | 「整理 inbox」→ organize + `plan-inbox-routing`。其它 → organize |
 | 复盘 | organize vs loop | 「复盘」→ loop。「复盘要点成笔记」→ organize |
 | 归档 | capture vs maintain | 「归档材料」→ capture。「归档专题 / archive topic」→ maintain |
-| 笔记 / note | capture vs write | 未限定 → capture。「写作 / 起草 / 续写」→ write |
+| 笔记 / note | capture vs write | 未限定 → capture。「写作 / 起草 / 续写」→ write；「新建笔记」创建后打开，交给宿主编辑器 |
 | 记账 / 记一笔 / 花了 / 存入 | ledger vs capture | 「记账 / 记一笔 / 花了 / 存入 / 查看账单 / 账户余额」→ `topmind-ledger`（记忆平面账本）。「记一下」仍 → capture |
 | 研究 / 分析 / 对比 | organize vs topmind-research（外部可选） | 对象是工作区里已存的笔记、专题内材料 → `topmind-organize`。需要对外检索一手资料（论文、技术报告、官方公告、模型或产品横向对比）、AI 前沿深挖 → `topmind-research`（analyze）；未安装 → `topmind-organize`「研究分析」 |
 | AI 动态 / 大厂动态 / AI 周报 / 最近有什么新的 | topmind-research vs last30days vs TopStream每日精选（均外部可选） | 官方公告、技术报告、论文的周度汇总 → `topmind-research`（collect）。社区讨论与口碑（X、Reddit、HN 等）→ `last30days`。每日精选例行 → `TopStream每日精选`。对应技能未安装 → `topmind-organize`（只用工作区已存材料，回执注明）。「整理本周 / 理顺本周动态」指工作区动态类别，仍 → organize「整理本周」 |
@@ -66,7 +66,8 @@
 - 修复 / 清理 / 重建 → maintain
 - 巡检 / 整体 / 断点 → loop
 - 整理 / 分析 / 总结 → organize
-- 写 / 发布 / 出稿 → write
+- 写 / 出稿 / 通用稿件 → write；明确平台稿优先对应独立写作技能
+- 发布 / 发出去 → 对应连接器；写作技能只准备稿件
 - 记忆 / 更新我的情况 / 周期反思 → memory
 - 仅类别名 → capture 到该类别根
 - 全不像 → capture 到 Inbox（buffer）

package/topmind/SKILL.md

@@ -7,14 +7,14 @@ compatibility: >-
   topmind workspace with {NN-Name}/ categories. Host file tools primary; UTR optional.
   Install with shared/ sibling for progressive disclosure links.
 metadata:
-  version: "4.15.8"
+  version: "4.15.10"
   action_category: "router"
   entrypoint: "true"
   triggers: "topmind, 知识库, 工作区, 待办, 有什么要做的, 要做的事, todo, todo list"
   tags: "router, entrypoint, topmind"
   author: "TopMindSpace"
   homepage: "https://github.com/topmindspace/topmind-skills"
-  updated: "2026-10-26"
+  updated: "2026-10-28"
   degradation: "../shared/capability-degradation.md"
 ---
 
@@ -31,15 +31,16 @@ metadata:
 | 记链接、摘录、想法、记一下要做某事 | `topmind-capture`；动态周期本 / 专题 / Inbox，显式待办进 `memory/todo.md` |
 | 待办有哪些、完成/更新某条待办 | 直接读写 `memory/todo.md`，不启动整理流程 |
 | 整理本周、总结/对比已有材料、开/并专题 | `topmind-organize`；临时问答按需回复，持久整理写回 |
-| 起草、润色、成稿、交付 | `topmind-write` |
+| 用工作区素材起草、润色、保存通用稿件 | `topmind-write` |
+| 明确公众号、X 长文/短帖、封面或长图 | 直接使用已安装的独立写作技能；需保存到工作区时沿用 `topmind-write` 的落盘约定 |
 | 记住我、更新我的情况、周期反思 | `topmind-memory` |
 | 结构体检、修复、恢复 | `topmind-maintain` |
 | 全工作区语义巡检、长任务断点恢复 | `topmind-loop` |
 | 微信读书 / X / 记账 | 可选 `topmind-weread` / `topmind-x` / `topmind-ledger` |
 
-外部检索研究、公众号、X 长文/短帖、封面/海报、幻灯片、交接的可选技能与容易混淆的请求，按需读 [路由消歧](../shared/trigger-disambiguation.md)。仅使用宿主实际可用的技能和工具；未安装专用技能时完成通用可行部分并说明限制。
+外部检索研究、平台写作、封面/海报、幻灯片、交接的可选技能与容易混淆的请求，按需读 [路由消歧](../shared/trigger-disambiguation.md)。仅使用宿主实际可用的技能和工具；未安装专用技能时完成通用可行部分并说明限制。不让用户先选择内部技能。
 
-用户一次要求多步，按依赖连续完成到最终交付；不因切换技能重复确认。用户未请求的延伸只给建议，见 [多意图](references/multi-intent.md) 与 [授权及建议](../shared/auto-suggest.md)。
+用户一次要求多步，按依赖连续完成到最终交付；不因切换技能重复确认。怎么判断、何时用工作区经验、何时研究、何时写入耐久经验，只见 [授权及建议](../shared/auto-suggest.md)（方向大于约束）。用户未请求的延伸只给建议。本入口不复制第二套阶梯。多意图边界见 [多意图](references/multi-intent.md)。
 
 ## 产品特有约定
 

package/topmind-capture/SKILL.md

@@ -5,7 +5,7 @@ description: >-
 license: MIT
 compatibility: topmind workspace. Host file tools or optional UTR capture-note.
 metadata:
-  version: "4.15.8"
+  version: "4.15.10"
   action_category: "capture"
   entrypoint: "false"
   triggers: "收集, 收进, 保存, 剪藏, 记一下, 临时存放, 随手记, 想法, 灵感, 暂存沉淀, 记录, 备忘, 存一下, 暂存, 笔记, 抄下来, 摘录, 记一下要做, 别忘了, 要做, capture, note, save, Note it"

package/topmind-ledger/SKILL.md

@@ -5,7 +5,7 @@ description: >-
 license: MIT
 compatibility: Host file tools or Kernel writeback. Desktop optional mini-app. Feishu/lark-cli is not required.
 metadata:
-  version: "4.15.8"
+  version: "4.15.10"
   action_category: "memory"
   entrypoint: "false"
   triggers: "记账, 记一笔, 花了, 存入, 查看账单, 账户余额, bookkeeping, log expense"

package/topmind-loop/DESIGN.md

@@ -4,7 +4,7 @@
 
 ## Why Loop Exists
 
-UTR 提供 **8 域 / 38** 条确定性命令（以 topmind 仓库 `TOOLS.md` 为准）；`doctor-workspace` 做规则健康检查。Loop 在之上增加**语义层**：阅读 `topic.md` 与笔记、对照 skill 契约、再调用 UTR 修复。
+UTR 提供 **9 域 / 42** 条确定性命令（以 topmind 仓库 `TOOLS.md` 为准）；`doctor-workspace` 做规则健康检查。Loop 在之上增加**语义层**：阅读 `topic.md` 与笔记、对照 skill 契约、再调用 UTR 修复。
 
 ## Design Choice — Skill + Agent + State File, Not CLI
 

package/topmind-loop/SKILL.md

@@ -5,7 +5,7 @@ description: >-
 license: MIT
 compatibility: topmind workspace. Progress under .topmind/loop/*. Host LLM runs the walk; UTR optional.
 metadata:
-  version: "4.15.8"
+  version: "4.15.10"
   action_category: "loop"
   entrypoint: "false"
   triggers: "跑一遍 loop, loop 一下, 整体体检, 工作区复盘, 把工作区跑一遍, 巡检, 继续 loop, 从断点继续, 全面检查, 复盘, review, audit, walk"

package/topmind-maintain/SKILL.md

@@ -5,7 +5,7 @@ description: >-
 license: MIT
 compatibility: topmind engine and/or workspace. Prefer read-only diagnostics first.
 metadata:
-  version: "4.15.8"
+  version: "4.15.10"
   action_category: "maintain"
   entrypoint: "false"
   triggers: "快速体检, 体检, 诊断, doctor, 健康检查, 清理, 修复, 快速检查, 检查, 状态, workspace check, scan, fix, maintain"

package/topmind-memory/SKILL.md

@@ -5,7 +5,7 @@ description: >-
 license: MIT
 compatibility: workspace core profile (memory/profile.md) + periodic reflections.
 metadata:
-  version: "4.15.8"
+  version: "4.15.10"
   action_category: "memory"
   entrypoint: "false"
   triggers: "写入专题记忆, 更新我的情况, stable memory, 追加确认结论, 记住这个, 记住我, 加到专题记忆, 沉淀结论, 提炼到记忆, update profile, memory"

package/topmind-organize/SKILL.md

@@ -5,7 +5,7 @@ description: >-
 license: MIT
 compatibility: topmind workspace with stream period notes, topics, or Inbox.
 metadata:
-  version: "4.15.8"
+  version: "4.15.10"
   action_category: "organize"
   entrypoint: "false"
   triggers: "整理本周, 理顺本周, 周复盘, 整理, 组织, 研究, 分析, 对比, 提取, 梳理, 归纳, 证据, 实体, 质量审查, 总结, 概括, 汇总, 提纲, 分类整理, 找关联, 总结成笔记要点, organize, summarize, analyze, weekly review"

package/topmind-pack.json

@@ -4,8 +4,8 @@
   "label": "topmind",
   "description": "topmind skill pack: single daily entry topmind + progressive disclosure (shared/ + references/), Agent Skills spec-valid frontmatter (custom fields under metadata, checked by agentskills validate), capture-first template-driven categories, optional connectors topmind-weread and topmind-x, optional topmind-ledger (记账). None of the optional skills is an extra user concept. WeChat (公众号) and other writing skills live in the external topmind-writing-skills pack (topmind-wechat-post etc.) and are routed only when installed. File-system-first content truth; UTR optional.",
   "daily_entry": "topmind",
-  "version": "4.15.8",
-  "updated": "2026-10-26",
+  "version": "4.15.10",
+  "updated": "2026-10-28",
   "license": "MIT",
   "locales": {
     "default": "zh-CN",
@@ -334,14 +334,16 @@
       "workspace-transform",
       "workspace-maintain",
       "contract",
+      "creator",
       "memory",
       "lifecycle",
       "derived"
     ],
-    "kind_naming_legacy": "Domain names (workspace-read / workspace-write / workspace-transform / workspace-maintain / contract / memory / lifecycle / derived) are stable identifiers. Command names follow the topic-and-category vocabulary.",
+    "kind_naming_legacy": "Domain names (workspace-read / workspace-write / workspace-transform / workspace-maintain / contract / creator / memory / lifecycle / derived) are stable identifiers. Command names follow the topic-and-category vocabulary.",
     "command_vocabulary": {
       "workspace-read": [
         "list-categories",
+        "list-category-files",
         "list-topics",
         "inspect-topic",
         "list-topic-files",
@@ -372,6 +374,11 @@
         "ensure",
         "reseed"
       ],
+      "creator": [
+        "inspect",
+        "snapshot",
+        "matchLengthReport"
+      ],
       "memory": [
         "promote",
         "digest",
@@ -395,11 +402,12 @@
         "rebuild"
       ]
     },
-    "command_count": 38,
-    "mcp_default_count": 29,
+    "command_count": 42,
+    "mcp_default_count": 33,
     "command_exposure": {
       "primary": [
         "workspace-read.list-categories",
+        "workspace-read.list-category-files",
         "workspace-read.list-topics",
         "workspace-read.inspect-topic",
         "workspace-read.list-topic-files",
@@ -410,6 +418,9 @@
         "workspace-transform.plan-inbox-routing",
         "workspace-maintain.doctor-workspace",
         "contract.validate",
+        "creator.inspect",
+        "creator.snapshot",
+        "creator.matchLengthReport",
         "memory.promote",
         "memory.digest",
         "memory.append-profile",

package/topmind-weread/SKILL.md

@@ -5,7 +5,7 @@ description: >-
 license: MIT
 compatibility: Requires WeRead Skill API key (wrk-*). Desktop WereadService preferred; host may call gateway manually.
 metadata:
-  version: "4.15.8"
+  version: "4.15.10"
   action_category: "connector"
   entrypoint: "false"
   triggers: "微信读书, weread, 划线同步, 读书笔记, 阅读统计, 书架, weread sync, sync highlights"

package/topmind-write/SKILL.md

@@ -1,18 +1,18 @@
 ---
 name: topmind-write
 description: >-
-  根据 topmind 素材起草、续写、改稿或交付。Use when 写成稿、润色、write、draft、revise。Do NOT use for capture-only, profile-only, or external posting requests.
+  根据 topmind 工作区素材起草、续写、改稿或保存通用稿件。Use when 写成稿、润色、write、draft、revise。明确平台稿优先已安装的独立写作技能。Do NOT use for capture-only, profile-only, or external posting requests.
 license: MIT
 compatibility: topmind workspace. Delivery role category (often 88-交付 / 88-Delivery).
 metadata:
-  version: "4.15.8"
+  version: "4.15.10"
   action_category: "write"
   entrypoint: "false"
-  triggers: "写, 写作, 起草, 续写, 修订, 润色, 交付, 改写, de-AI, 输出, 出稿, 定稿, 发布, 导出, 排版, write, draft, deliver, publish, export"
+  triggers: "写作, 起草, 续写, 修订, 润色, 交付, 改写, de-AI, 出稿, 定稿, 导出, write, draft, deliver, export"
   tags: "write, draft, revise, polish, deliver"
   author: "TopMindSpace"
   homepage: "https://github.com/topmindspace/topmind-skills"
-  updated: "2026-10-26"
+  updated: "2026-10-28"
   degradation: "../shared/capability-degradation.md"
 ---
 
@@ -36,7 +36,7 @@ metadata:
 
 ## 可选专用能力
 
-写作技能保持独立仓库；Desktop 与 Obsidian 默认内置受验证版本，其他 Agent 按需安装。改稿加载 `topmind-revise`，只改排版加载 `topmind-format`；宿主有 `render_article` 时用它生成平台 HTML/文本，无需猜测或执行 Python 命令。生成后按用户意图落盘，预览/复制不等于发布。
+写作技能保持独立仓库；Desktop 与 Obsidian 默认内置受验证版本，其他 Agent 按需安装。明确指定平台或只改排版时，优先直接使用相应独立技能：改稿 `topmind-revise`，只排版 `topmind-format`；本技能仅补工作区素材读取和落盘，不再要求用户选一次技能。宿主有 `render_article` 时用它生成平台 HTML/文本，无需猜测或执行 Python 命令。生成后按用户意图落盘，预览/复制不等于发布。
 
 公众号 → `topmind-wechat-post`；X 长文 → `topmind-x-article`；干货短文 / 引流短帖 → `topmind-briefs` / `topmind-viral-posts`；幻灯片 → `topmind-presentation`。已安装且匹配时使用，未安装则完成通用稿并说明专用格式限制，不能伪装已生成 HTML/PPTX。消歧详见 [trigger-disambiguation](../shared/trigger-disambiguation.md)。写稿不构成外部发布授权。
 

package/topmind-x/SKILL.md

@@ -5,7 +5,7 @@ description: >-
 license: MIT
 compatibility: Desktop needs Bearer (read) and/or xurl (post). Agent hosts use official X MCP + xurl bridge.
 metadata:
-  version: "4.15.8"
+  version: "4.15.10"
   action_category: "connector"
   entrypoint: "false"
   triggers: "发推, 推特, twitter, x.com, 发帖, 推文, post tweet, x platform"