360 Secure SkillHub
360 安全技能中心

sn-update

面向 OpenClaw 或 hermes-agent 环境,从上游 SenseNova-Skills 仓库同步更新已安装的 sn-* Skill:按范围选择全部或指定 skill,用提交 SHA、标签或版本字段比对判定是否陈旧,安装前把旧目录移入时间戳备份桶并只保留最近三份,最后输出更新、已最新与错误的分组报告。

sensenova-skills
sensenova-skills
热度 8.88
最近更新 2026-08-19 10:39:38
分类:自动化依赖:代码执行 / 读写文件 / 第三方 API应用场景:工作流 / DevOps

安装技能

复制下面的话术发给你的 Agent,它会通过 secure-skills 检索并安装该技能。

执行命令安装此 skill: npx secure-skills-cli add sn-update
Metadata
name
sn-update
description
Update SenseNova Skills (the sn-* bundle) inside an OpenClaw or hermes-agent install. ALWAYS use this skill when the user says any of: "update SenseNova skills", "update SN skills", "更新 sensenova skills", "更新 sn skills", "刷新 sn-*", "升级 sn-* skills", or names a specific sn-* skill to update (e.g. "更新 sn-ppt-standard", "refresh sn-image-base"). Default scope is the whole sn-* bundle; if the user names specific skills, update ONLY those.
metadata
{"project":"SenseNova-Skills","tier":1,"category":"meta","user_visible":true}
triggers
["sn-update","更新 sn","更新 sensenova","update sn skills","update SN skills","更新 sn skills","刷新 sn-*","刷新 sn skills","更新 sn-ppt-standard","refresh sn-image-base","update sensenova skills"]

sn-update

Refresh installed sn-* skills from upstream SenseNova-Skills.


Decide the scope

  • No list given → every sn-* skill upstream.
  • Specific skills named → only those. Don't expand.
  • A named skill missing upstream → surface as error and stop.
  • User said "force / 强制" → re-install even when up-to-date.

Decide the target agent

Check which directories exist:

~/.openclaw/skills/ ~/.hermes/skills/ Target
exists absent openclaw
absent exists hermes
exists exists ask the user — never silently dual-write
absent absent no install found, stop

Sync the upstream repo

Persistent cache at ~/.cache/sn-update/repo/. Default URL: https://github.com/OpenSenseNova/SenseNova-Skills.git. User may override with a fork URL.

  • First run: if you want to actually limit blob download, use partial clone with --filter=blob:none --no-checkout, then sparse-checkout only the selected skills/<name> paths before copying them. --filter=blob:none alone does not keep the cache small if the full worktree is checked out; that checkout will still download most or all needed blobs. It still preserves history metadata for SHA queries.
  • Subsequent runs: fetch + hard-reset to the upstream default branch, and re-apply sparse-checkout for only the requested skills/<name> paths before copying. If updating the whole sn-* bundle, expect most/all skill blobs to be downloaded.
  • URL changed: if cache's origin differs from the requested URL, delete and re-clone.

Compare versions per skill (A → B → C)

For each skill, pick the highest-precedence signal present on both sides (installed + upstream); equal → skip, differ → install.

Upstream "version" is the per-subtree commit SHA — using repo HEAD would mark unrelated skills as stale every time:

git -C <cache> log -1 --format=%H -- skills/<skill-name>
  • A — .sn-version marker: one-line file inside the installed skill holding the SHA from its last install.
  • B — .sn-release marker (fallback): one-line file holding the upstream tag name. Compare against git describe --tags --abbrev=0.
  • C — optional version: field in SKILL.md frontmatter: parse from YAML on both sides, but only for forks or skills that explicitly add this field. If either side lacks it, C does not apply.
  • Nothing usable → treat as stale and install.

Always write .sn-version on install so future runs can use A.


Install with backup

For each skill flagged "install":

  1. Move (not copy) any existing <agent-skills>/<skill-name>/ into a single timestamped backup bucket shared by all skills in this run: ~/.<agent>/skills_backup/<UTC-timestamp>/<skill-name>/ (e.g. 2026-04-30T15-29-07Z).
  2. Copy <cache>/skills/<skill-name>/ → <agent-skills>/<skill-name>/. Never symlink (ln -s) from the cache. The cache lives under ~/.cache/ with permissions the agent runtime may not be able to traverse, and some runtimes refuse to load skills resolved through symlinks. Always do a real recursive copy so the installed tree is self-contained and owned by the agent skills dir.
  3. Write .sn-version with the upstream subtree SHA inside the new copy.

If the bucket ends up empty (all targets were fresh installs), remove it.

The backup tree is a sibling of skills/, never a .bak folder inside it — most agent runtimes scan the whole skills/ directory and would pick up stale duplicates.


Enforce backup retention

After every run, prune the per-agent backup root to at most 3 buckets. Timestamps sort lexicographically; keep the newest 3, delete the rest. Run this even when the current run produced no backup of its own.


Report to the user

Group by status, keep it short:

Updated (3): sn-ppt-standard, sn-image-base, sn-deep-research
Already up-to-date (5): sn-ppt-creative, sn-ppt-doctor, ...
Backup: ~/.openclaw/skills_backup/2026-04-30T15-29-07Z/
  • Omit Backup: when nothing was backed up.
  • Show SHA pairs (abc1234 → def5678) only if the user asks for detail.
  • Errors get their own line — never bury them in a success summary.

Edge cases

  • Asks to delete sn- skills* — not this skill's job. Decline; point at ~/.<agent>/skills_backup/ if they want to roll back.
  • User is in the dev repo and asks to "update sn skills" — they mean push to their agent install. Proceed normally; this skill only touches the cache and the agent install dirs, never the dev checkout.
  • sn-update updating itself — fine; the new copy takes effect on the next invocation.