# Skill Publisher > Publish a completed SKILL.md folder to one or more public marketplaces (ClawHub, Tencent SkillHub, skills.sh, SkillsMP, LobeHub, Agensi, Coze, ClawMart) — validates frontmatter, optionally drafts a bilingual README, strips credential / build / dot-files, tags git, and ships serially. Use when the user names BOTH the action and the destination — e.g. "publish to ClawHub and skills.sh", "把这个 skill 推到腾讯 SkillHub", "ship a new version of my-skill to all 8 marketplaces". Do NOT trigger on bare "发布 skill" or "ship to marketplace" — those alone are too ambiguous for a destructive pipeline. Side effects (see Security & Privacy): local file read/overwrite, optional deletion of credential files (dry-run by default), git tag push, network calls to marketplace APIs, token auth via env vars. Not for: local-only management, drafting an unfinished skill, or single-marketplace docs. - Skill: `mebusw/skill-publisher` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add mebusw/skill-publisher` - Raw SKILL.md: https://api.skillmd.com/api/skills/mebusw/skill-publisher/raw - Safety review: WARNING (external: skill-scanner WARNING, skillspector WARNING) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: DevOps & Infra - Author: mebusw (https://skillmd.com/u/mebusw) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/mebusw/skill-publisher --- # Skill Publisher — 多市场一键发布 v2.2 Publish a completed skill folder to the public marketplaces you name. > **Stop and read "Security & Privacy" below before running.** This skill > touches local files, git remotes, and external APIs. The defaults are > non-destructive (`--dry-run` is on for cleanup, no token in plaintext by > default), but you must opt in to the rest. *** ## Security & Privacy — what this skill WILL do Run through this list with the user before the first invocation. Every item below has been the source of a real complaint; the defaults now err on the safe side, but the user must still explicitly request the action. | # | Side effect | Default | How to opt in / opt out | |---|------------|---------|-------------------------| | 1 | **Read** `skill_dir` and traverse it | Always | Cannot disable. | | 2 | **Write / overwrite** files in `skill_dir` (templates, version bumps) | Off | Pass `--write` (or any concrete publish flag). | | 3 | **Delete** files matching credential / build / dot-file patterns inside `skill_dir` | **Off — dry-run only** | `python3 scripts/strip_bundle.py ` first, review the list, then re-run without `--dry-run`. The inline `rm -rf` paths in the legacy 2.1 SKILL are **removed**. | | 4 | **Mutate global git config** (`git config --global user.email/name`) | **Removed** | No longer happens. If git complains about missing identity, ask the user to set it themselves or pass `--git-user-email` / `--git-user-name` for this repo only. | | 5 | **Patch installed CLI files on disk** (e.g. `sed -i` on `~/.npm/.../publish.js`) | **Removed** | No longer happens. If a CLI bug blocks publish, the skill prints a clear error and asks the user to upgrade the CLI. | | 6 | **Push git tag to `origin`** | Off | `--push-tag` or `--yes-push`. The skill will list the tag it intends to push and stop. | | 7 | **Network call** to each marketplace API in the `markets` list | Off for any market not listed | Caller passes `markets=clawhub,skillsh,skillhub` etc. Empty `markets` = no network calls. | | 8 | **Auth token** sent to that marketplace | Token must be in an env var (`CLAWHUB_TOKEN`, `SKILLHUB_TOKEN`, etc.) — **not** a `--key` flag in shell history | If only a flag form is available, the skill prints a warning. | | 9 | **`curl ... | bash` from a remote URL** (SkillHub installer) | **Not run automatically** | The skill prints the exact command, a SHA-pinning note, and asks the user to run it themselves. | | 10 | **Auto-generate `README.zh-cn.md`** from English content | **Off** | `--bilingual` flag. The default English-only README respects the user's intent. | | 11 | **Report success** for marketplaces that only print a manual-instruction URL (LobeHub, ClawMart) | Reports "📋 TODO" not "✅" | The skill distinguishes "actually published" from "instructions printed". | If the user does not want any of #2–#10, the skill degrades to a read-only validator that prints a checklist and exits. *** ## When to use | User says | Trigger? | |-----------|----------| | "publish `` to ClawHub and skills.sh" | ✅ | | "把这个 skill 推到腾讯 SkillHub 和 ClawHub" | ✅ | | "ship a new version of `` to all 8 marketplaces" | ✅ | | "给 skill-publisher 打 v2.2 并发到 ClawHub + skills.sh + SkillHub" | ✅ | | "发布 skill" (alone, no target) | ❌ — too ambiguous; ask which marketplaces | | "ship to marketplace" (alone, no target) | ❌ — same | | "我刚写好一个 skill，帮我看看" | ❌ — that's `skill-optimizer`, not publish | | "how do I publish to ClawHub" (asking for docs) | ❌ — read `references/clawhub.md` instead | *** ## Required user-supplied parameters | Parameter | Required for | Source | Notes | |-----------|-------------|--------|-------| | `skill_dir` | All | Prompt | Must contain `SKILL.md`. | | `markets` | All | Prompt or flag | Comma-separated. **No default — caller must list at least one.** | | `CLAWHUB_TOKEN` | `markets=clawhub` | Env var | Never put on the command line. | | `SKILLHUB_TOKEN` | `markets=skillhub` | Env var | Never put on the command line. | | `GITHUB_TOKEN` (optional) | `markets=skillssh,skillsmp,lobehub,agensi` | Env var | Only needed if the repo is private. | | `COZE_TOKEN` | `markets=coze` | Env var | Never put on the command line. | | `AGENSI_KEY` | `markets=agensi` | Env var | Never put on the command line. | | `--bilingual` | Optional | Flag | Generates `README.zh-cn.md` from `README.md`. Off by default. | | `--push-tag` | Optional | Flag | Pushes the new git tag to `origin`. Off by default. | | `--confirm-cleanup` | Optional | Flag | Re-runs the cleanup script **for real** (not dry-run) after the user has reviewed the dry-run output. Off by default. | | `--write` | Optional | Flag | Allows Step 3 / Step 5 to write or delete inside `skill_dir`. Off by default — without this, the skill is read-only. | | `--dry-run` | Optional | Flag | **Default on.** Walks every step, prints what it *would* do, makes zero side effects. | > **Token rule:** if the user pastes a token on the command line, the skill > will refuse and ask them to set the env var instead. Tokens in shell > history leak through screenshots, screen sharing, and process listings. *** ## Execution flow (read top to bottom) Each step lists its side effects. Steps 1–2 are always read-only. Steps 3–6 require `--write` and (for destructive parts) `--confirm-cleanup` or `--push-tag`. ### Step 1 — Pre-flight (read only) ```bash [ ! -f "{skill_dir}/SKILL.md" ] && { echo "❌ SKILL.md 不存在"; exit 1; } [ -z "{markets}" ] && { echo "❌ --markets 必填，逗号分隔，例如 clawhub,skillsh"; exit 1; } for cli in clawhub skillhub; do which "$cli" >/dev/null 2>&1 || echo "⚠️ $cli 未安装（markets=$cli 时需要）" done git --version ``` ### Step 2 — SKILL.md frontmatter 校验 (read only) ```bash python3 - <<'PY' import sys, re, yaml content = open("{skill_dir}/SKILL.md").read() m = re.match(r'^---\n(.*?)\n---', content, re.DOTALL) if not m: sys.exit("❌ frontmatter 缺失") data = yaml.safe_load(m.group(1)) for k in ("name", "description", "version", "author"): v = data.get(k) if v is None: sys.exit(f"❌ {k} 缺失") print(f"✅ {k}: {v}") PY ``` ### Step 3 — 文档生成 (write, but bilingual is opt-in) | File | When created | Notes | |------|-------------|-------| | `README.md` | Always (if missing) | English. Refreshes stale version / install command. | | `README.zh-cn.md` | **Only if `--bilingual` is set** | Generated from `README.md` via LLM. Off by default — the user must request it. | | `RELEASE_NOTES.md` | Always (if missing) | Generated from `CHANGELOG.md` or `git log`. | > **Why opt-in for the Chinese README:** auto-translating documentation > into another language can accidentally leak content the user did not > intend to publish, mistranslate legal / compliance language, or create > obligations they did not agree to. The default English-only README > matches the user's likely intent ("publish my skill") without > overreaching. ```bash [ -z "{bilingual}" ] && echo "ℹ️ README.zh-cn.md 跳过（需要 --bilingual 才生成）" ``` ### Step 4 — Git tag (no global mutation, push is opt-in) ```bash cd {skill_dir_parent} # Read existing identity. Do NOT mutate global config. SKILL_VER=$(grep '^version:' {skill_dir}/SKILL.md | head -1 | cut -d: -f2 | tr -d ' ') TARGET_TAG="v$SKILL_VER" CURRENT_TAG=$(git tag --sort=-v:refname | head -1 2>/dev/null || echo "") if [ "$CURRENT_TAG" = "$TARGET_TAG" ]; then echo "ℹ️ Tag $TARGET_TAG 已存在" else echo "📋 将创建本地 tag: $TARGET_TAG" git tag -a "$TARGET_TAG" -m "Release $TARGET_TAG" if [ -n "{push_tag}" ]; then echo "📋 将推送 tag 到 origin: $TARGET_TAG （--push-tag 已设置）" git push origin "$TARGET_TAG" else echo "ℹ️ tag 未推送。需要推送时重跑并加 --push-tag。" echo " ⚠️ 推送 tag 可能在 GitHub 触发 CI/CD、Release 流程，" echo " 并把版本号广播给所有订阅者。请确认后再加 --push-tag。" fi fi ``` > **Why the old `git config --global user.email` lines are removed:** > mutating the user's global git identity is an unrelated, persistent > change to other repositories. The publisher does not need it. If git > complains about missing identity, the skill surfaces that and asks the > user to set it once. ### Step 5 — Bundle 清理 (DRY-RUN BY DEFAULT, never inline rm -rf) > **⚠️ This step WILL delete files inside `skill_dir` if you opt in.** > The legacy 2.1 SKILL used inline `rm -rf` with broad patterns like > `.env`, `*.key`, `*.pem`, `credentials.*`. That is unsafe by default > and has been replaced with a reviewable Python script that exits > non-zero when it touches credential-looking files. Use the bundled script — it knows the allowlist, prunes build / dot / cache directories, and reports credentials separately. ```bash # Step 5a — always run dry-run first and show the user python3 {skill_dir}/../scripts/strip_bundle.py {skill_dir} --dry-run if [ -n "{confirm_cleanup}" ]; then # Step 5b — only after the user has reviewed the dry-run list python3 {skill_dir}/../scripts/strip_bundle.py {skill_dir} else echo "ℹ️ 清理未执行（需要 --confirm-cleanup 才真正删除）" echo " 上面的 [DRY-RUN] 行就是将删除的内容，请人工确认后再加 --confirm-cleanup。" fi ``` Exit code 1 from the script means credential-looking files were removed; the skill surfaces this and asks the user to review before continuing. ### Step 6 — 串行发布 (per-market, secrets via env) For each market in `--markets`, the skill runs the block below. Markets not listed are skipped entirely — no default-all behavior. #### 6a. ClawHub ```bash # Note: this skill does NOT patch ~/.npm/.../publish.js with sed anymore. # If the CLI rejects the publish because of an "acceptLicenseTerms" issue, # upgrade the CLI: npm i -g clawhub@latest, then re-run. [ -z "$CLAWHUB_TOKEN" ] && { echo "❌ CLAWHUB_TOKEN 未设置（请用环境变量）"; exit 1; } CLAWHUB_TOKEN="$CLAWHUB_TOKEN" \ clawhub skill publish {skill_dir} \ --slug {slug} --name "{display_name}" --owner {owner} \ --version {version} \ --changelog "Release v{version}" --tags "latest" ``` Error handling: 401 → token invalid; 409 → slug taken (suggest suffix); 429 → wait & retry. CLI patches via `sed -i` on installed files are no longer in scope. #### 6b. 腾讯 SkillHub > **⚠️ Step 1 (CLI install) is not run automatically.** The official > installer is `curl -fsSL https://skillhub.cn/install/install.sh | bash`. > Piping a remote script to `bash` executes whatever the server returns > with no integrity check. The skill prints the exact command, the URL, > and asks the user to run it themselves. If the user has already > installed the CLI, jump to `skillhub login`. ```bash # Login uses env var, NOT a --key flag in shell history. [ -z "$SKILLHUB_TOKEN" ] && { echo "❌ SKILLHUB_TOKEN 未设置"; exit 1; } SKILLHUB_TOKEN="$SKILLHUB_TOKEN" skillhub login [ ! -f "{skill_dir}/skillhub.yaml" ] && \ skillhub init --name {slug} --category "效率工具" --dir {skill_dir} cd {skill_dir} skillhub push --message "Release v{version}" skillhub publish --version {version} --changelog "Release v{version}" skillhub search {slug} && echo "✅ SkillHub 完成（审核约 5-10 分钟）" ``` #### 6c. skills.sh skills.sh is **read-only** — it indexes public GitHub repos. There is no publish API call. The "publish" step is `git push` of a tag (covered in Step 4) plus verifying the repo is public. The install command the user will paste elsewhere is `npx skills add https://github.com//`. ```bash GH_REPO="{github_repo}" curl -s "https://api.github.com/repos/$GH_REPO" | python3 -c " import sys, json try: d = json.load(sys.stdin) print('✅', d['full_name'], '| Stars:', d['stargazers_count']) except Exception: print('⚠️ Repo 不存在或未公开') " echo "📋 skills.sh: 等 GitHub 公开 repo 推送后由 skills.sh 抓取" echo " 安装命令: npx skills add https://github.com/$GH_REPO" ``` #### 6d. SkillsMP ```bash echo "📋 SkillsMP: 通过 GitHub repo 聚合，无独立 publish" echo " 注册独立页面: https://agentskills.io/submit" echo " 安装: npx skills add https://github.com/{github_repo}" ``` #### 6e. LobeHub Skills LobeHub has no automated publish CLI; the user must open a PR or run `lobe-cli skill add` themselves. The skill **does not** print `✅ lobe-cli 添加成功` when only a manual-instruction URL was printed — that would be a false success. ```bash # Note: previously this block printed "✅ ..." even when it only printed # a manual-PR link. That misleading output has been removed. echo "📋 LobeHub: 需要手动完成下列之一" echo " 1) PR 到 https://github.com/lobehub/lobe-chat#skills" echo " 2) 或本机运行: npx skills add https://github.com/{github_repo}" echo " 3) 或本机运行: lobe-cli skill add {github_repo} (需先安装 lobe-cli)" ``` #### 6f. Agensi ```bash [ -z "$AGENSI_KEY" ] && { echo "❌ AGENSI_KEY 未设置"; exit 1; } curl -X POST https://api.agensi.ai/v1/skills/publish \ -H "Authorization: Bearer $AGENSI_KEY" \ -H "Content-Type: application/json" \ -d '{ "name": "{slug}", "repo": "{github_repo}", "version": "{version}", "description": "...", "tags": ["productivity"] }' echo "📋 Agensi: 已提交，等人工审核" ``` #### 6g. Coze ```bash [ -z "$COZE_TOKEN" ] && { echo "❌ COZE_TOKEN 未设置"; exit 1; } curl -X POST "https://api.coze.cn/v1/skills" \ -H "Authorization: Bearer $COZE_TOKEN" \ -H "Content-Type: application/json" \ -d '{ "name": "{slug}", "description": "...", "icon_url": "...", "skill_type": "published" }' echo "📋 Coze: 已提交，等人工审核" ``` #### 6h. ClawMart ClawMart is a third-party paid marketplace with no CLI. The skill prints instructions and stops — no false success. ```bash echo "📋 ClawMart: 手动发布到 https://clawmart.com/sell" echo " 参考: https://clawmart.com/publish" ``` ### Step 7 — 结果汇总 ``` ========================================= 📋 多市场发布结果 ✅ = actually published 📋 = manual step required (see above) {markets, e.g.} ✅ ClawHub clawhub.com/skills/{slug} clawhub install {slug} ✅ 腾讯 SkillHub skillhub.tencent.com/skills skillhub install {slug} 📋 skills.sh agentskill.sh npx skills add (等 GitHub 抓取) 📋 SkillsMP agentskills.io npx skills add (聚合) 📋 LobeHub lobehub.com/skills 手动 PR 📋 Agensi agensi.ai 审核中 📋 Coze coze.cn 审核中 📋 ClawMart clawmart.com 手动发布 ========================================= ``` *** ## Global flags (summary) | Flag | Effect | Default | |------|--------|---------| | `--dry-run` | Read-only walk-through, prints actions | **On** | | `--write` | Allow Step 3/5 to mutate files | Off | | `--confirm-cleanup` | Run `strip_bundle.py` for real after dry-run review | Off | | `--push-tag` | `git push origin ` after tagging | Off | | `--bilingual` | Generate `README.zh-cn.md` | Off | | `--markets` | Comma-separated targets | **Required** — no default | *** ## Privacy & data flow | Where data goes | What is sent | When | |----------------|--------------|------| | ClawHub API | `skill_dir` contents + `CLAWHUB_TOKEN` | `--markets=clawhub` | | SkillHub API | `skill_dir` + `SKILLHUB_TOKEN` | `--markets=skillhub` | | GitHub API | Repo read (stars, existence) | `--markets=skillssh,lobehub,skillsmp,agensi` | | Agensi API | Skill metadata + `AGENSI_KEY` | `--markets=agensi` | | Coze API | Skill metadata + `COZE_TOKEN` | `--markets=coze` | | skills.sh / SkillsMP / LobeHub / ClawMart | Nothing at runtime — these are GitHub-indexed or manual | Always off | > The skill does **not** send `skill_dir` to anywhere it was not asked > to. The user picks `markets`; the skill does not pick for them. *** ## Known limits 1. SkillHub / Agensi / Coze have human review (5–10 min). 2. LobeHub requires a manual GitHub PR. 3. ClawMart requires manual web publishing. 4. Version must strictly increase; same-version republish is rejected unless `--force` is passed. 5. The cleanup script is conservative — anything it cannot classify as part of the skill is removed. If a skill genuinely needs a file outside the standard folders, edit `.clawhubignore` (or the script's keep-list) before running.