openclaw-babysit

🦞 openclaw-admin-skills

Natural-language skills for running OpenClaw. Tell your AI harness to "update openclaw" or "fix my openclaw" in plain English, and get a safe, reviewable workflow — preflight checks, evidence-based diagnosis, explicit approval before any change, and per-user memory that makes every next run smarter.

What's included

Non-technical users can fire up Claude Code in any directory, say "openclaw is broken", and these skills figure out which OpenClaw to target and what to do — without the user needing to know where OpenClaw is installed.

Quick start (Claude Code)

mkdir -p ~/src
git clone https://github.com/pejmanjohn/openclaw-skills.git ~/src/openclaw-skills
~/src/openclaw-skills/scripts/install-claude-skill.sh

Then in any Claude Code session, just talk:

> update openclaw
> openclaw isn't responding
> find my openclaw

Or wrap a recurring health check around your channels with Claude Code's /loop:

> /loop 5m /openclaw-babysit              # alert mode — tells you when channels degrade
> /loop 5m /openclaw-babysit --auto-fix   # autonomous mode — applies allowlisted fixes and verifies

Babysit reports a one-line OK on healthy ticks; a degraded reading either alerts you (default) or, in --auto-fix mode, invokes troubleshooting non-interactively to apply restart-family fixes from a bounded allowlist. See openclaw-babysit below for the full contract.

Update the skill itself later with:

git -C ~/src/openclaw-skills pull
~/src/openclaw-skills/scripts/install-claude-skill.sh

Install

Keep one canonical source checkout and install into whichever agents you use. The repo ships three install scripts that symlink back to the canonical checkout — so git pull picks up updates everywhere.

One-time clone, shared by all three flows below:

mkdir -p ~/src
git clone https://github.com/pejmanjohn/openclaw-skills.git ~/src/openclaw-skills

Claude Code: Install

~/src/openclaw-skills/scripts/install-claude-skill.sh

Symlinks into ~/.claude/skills/. The canonical entrypoint on disk is ~/.claude/skills/openclaw-troubleshooting/SKILL.md (a symlink to the repo).

For project-local install instead of personal:

~/src/openclaw-skills/scripts/install-claude-skill.sh --dest "/path/to/project/.claude/skills"

Codex: Install

~/src/openclaw-skills/scripts/install-codex-skill.sh

Symlinks into $CODEX_HOME/skills or ~/.codex/skills. After install, restart Codex so it reloads.

OpenClaw: Install

OpenClaw loads skills from these locations, highest precedence first:

• Workspace: <workspace>/skills
• Project agent skills: <workspace>/.agents/skills
• Personal agent skills: ~/.agents/skills
• Managed/local: ~/.openclaw/skills
• Bundled skills
• skills.load.extraDirs

~/src/openclaw-skills/scripts/install-openclaw-skill.sh

Symlinks into ~/.openclaw/skills. For workspace-local install:

~/src/openclaw-skills/scripts/install-openclaw-skill.sh --dest "/path/to/workspace/skills"

Updating

git -C ~/src/openclaw-skills pull
# then re-run the install script(s) for the agent(s) you use

How it works

Three principles do most of the work:

1. Local is runtime truth. The agent and OpenClaw run on the same host, so local binary/help/config/state/logs are what actually govern behavior. Docs at docs.openclaw.ai are procedural truth for the latest release, but if the local binary disagrees with the docs, trust the local binary. Every skill grounds in live openclaw <subcommand> --help output before acting.

2. Skills chain together. Troubleshooting auto-triggers instance discovery on first run. Update and troubleshooting both read the saved instance registry. After an incident, troubleshooting suggests the compound skill to capture learnings. Each skill is lean; together they're an end-to-end workflow.

3. Local memory compounds. Every install has a gitignored local/ tree. Incident-log entries written by the compound skill, the instance registry written by discovery, and the upgrade-history ledger written by update — all live there, per machine, never pushed upstream. Future sessions read them as context, so the troubleshooting that took an hour the first time becomes a 30-second lookup the next time.

┌──────────────────────────┐     ┌──────────────────────────────┐
│ openclaw-troubleshooting │     │ openclaw-troubleshooting-    │
│                          │     │ compound                      │
│  1. Read incident log    │     │                              │
│  2. Diagnose & fix       │────▶│  1. Review conversation      │
│  3. Suggest /compound    │     │  2. Draft tight incident     │
│                          │     │  3. Draft new signatures     │
└──────────────────────────┘     │  4. Check for duplicates     │
                                 │  5. Write to local memory    │
                                 └──────────────────────────────┘

The skills in depth

openclaw-update

Safe, reviewable update from the installed version to the latest on your channel — one approval gate up front ("go from X to Y?"), then the skill runs through eight phases and stops only if something anomalous happens. A clean run lands on a personalized highlights briefing filtered to features you actually use.

Under the hood: eight phases (target → preflight → release-notes pre-check → backup → update → restart → verify → highlights). Anomaly stops: BREAKING release-note flag, backup verify failure, non-zero update exit, gateway doesn't return healthy, or a post-update gate yellow/red.

What makes it safe:

• Captures a full JSON inventory snapshot before the update (channels, crons, plugins, skills, doctor state) and diffs after, so silent regressions are caught.
• Always passes --no-restart to openclaw update; never restarts the gateway mid-session on an OpenClaw-native agent (which would kill the agent); emits manual commands for that case.
• On pnpm/npm installs where dep state is uncertain, checks bundled-extension staging completeness before advancing to restart (the stageRuntimeDependencies: true case) and offers a scoped per-extension repair if needed.
• Finishes with a highlights briefing filtered against the user's inventory — only sur