openclaw-agent-onboarding

OpenClaw 个人 AgentOS 初始化向导 / Bootstrapper。Use when a user wants to initialize, diagnose, upgrade, repair, or health-check a new or existing OpenClaw setup; in...

Security Scan

The skill's requirements, instructions, and small diagnostic script are coherent with an OpenClaw AgentOS onboarding/health-checker and do not ask for unrelated credentials or hidden installs, though it performs filesystem checks and recommends package installs that should be confirmed by the user.

Purpose & Capability

The name/description (AgentOS bootstrap, diagnose, install baseline skills, set up memory/KB) matches the actual content: a long SKILL.md describing discovery, install plans, memory layout, safety rules, and a short diagnose.py. Nothing requested (no env vars, no credentials) is out of scope for an onboarding tool.

Instruction Scope

The SKILL.md instructs the agent to inspect workspace paths, installed skills, binaries (git/npm/clawhub), and to create directories/templates and install skills after confirmation. These are consistent with bootstrapping and diagnostics. The doc explicitly forbids destructive actions (deletes/overwrites) and requires confirmation for risky installs/paid actions, which constrains scope. It will read filesystem paths and environment variables (OPENCLAW_WORKSPACE, OPENCLAW_SKILLS) — appropriate for this purpose.

Install Mechanism

There is no automated install spec in the registry (instruction-only). The SKILL.md recommends npm/global installs (e.g., `npm i -g clawhub`, `npx skills add ...`) and references a direct backup zip URL (pub-...r2.dev) for a sample skill. These are expected for a bootstrapper but warrant user confirmation before running network downloads or global package installs. The skill itself does not automatically perform installs without confirmation according to its safety rules.

Credentials

The skill declares no required credentials or config paths. The included script reads OPENCLAW_WORKSPACE and OPENCLAW_SKILLS (with sensible defaults) and checks for presence of binaries; this is proportional to its diagnostic purpose. No API keys, tokens, or unrelated secrets are requested.

Persistence & Privilege

always is false and the skill is user-invocable. It does not request permanent elevated presence or modify other skills' configurations. The SKILL.md emphasizes user confirmation for risky writes and does not claim automatic persistent changes.

Assessment

This skill appears to be what it says: a bootstrapper/diagnostic pack. Before running it: (1) Review and confirm any proposed package installs (npm global installs, npx commands) — these will modify your system environment. (2) Be cautious about any non-GitHub/ClawHub download URLs referenced (a backup zip URL is present in the references); fetch and inspect such archives before extracting. (3) The tool reads and writes inside your OpenClaw workspace (defaults to /Users/mac/.openclaw/... but respects OPENCLAW_WORKSPACE/OPENCLAW_SKILLS env vars) — back up important files first. (4) The skill's docs explicitly require confirmation for risky changes; ensure the agent actually prompts you before making changes. If you want extra caution, run the included scripts in a dry-run mode or in a sandboxed environment first.

OpenClaw AgentOS Onboarding

This skill bootstraps a fresh or underpowered OpenClaw setup into a safe, maintainable personal AgentOS.

Prime directive

Do not merely explain. Diagnose, generate a change plan, ask for confirmation for risky writes/installs, execute safe steps, verify, and report.

Default execution contract:

Preflight → Plan → Confirm risky changes → Execute → Verify → Report → Leave rollback notes

Safety rules

• Never delete user files.
• Never overwrite AGENTS.md, MEMORY.md, SOUL.md, TOOLS.md, USER.md, DREAMS.md.
• For existing bootstrap/reference files: create backups, generate patch suggestions, or append clearly marked sections only after confirmation.
• Do not install unknown-source skills automatically.
• Do not perform paid API/cloud actions without explicit confirmation.
• Do not write private user content into reusable public templates.
• Mark API-key-dependent skills before installing or configuring.
• Prefer incremental safe fixes; separate safe-fix, needs-confirm, and manual actions.

Operating modes

Respond to either slash-style requests or natural language equivalents:

/agentos bootstrap
/agentos diagnose
/agentos init
/agentos install
/agentos repair
/agentos upgrade
/agentos health
/agentos memory-check
/agentos skill-check
/agentos kb-init
/agentos kb-check
/agentos kb-obsidian
/agentos context-clean
/agentos report

Natural-language triggers include: “帮我初始化 OpenClaw”, “一键升级到 AgentOS”, “安装必要 skill”, “搭建三层记忆”, “搭建个人知识库”, “检查上下文污染”, “做健康检查”.

Darwin optimization note

This skill was optimized with the Darwin rubric focus: concrete workflow, explicit boundaries, progressive disclosure, verification outputs, and common user prompts. Typical test prompts are stored in test-prompts.json.

Decision matrix

Choose the narrowest mode that satisfies the user:

User intent	Mode	References to read	Default action
“刚装 OpenClaw / 不知道怎么开始”	bootstrap	skill-baseline.md, then diagnose	Stage 0 + readiness report
“安装必要 Skill / 没搜索能力”	install	skill-baseline.md, safety-policy.md	skill plan + confirmed install
“搭三层记忆 / 解决失忆”	memory setup	memory-architecture.md	create missing dirs/templates only
“搭个人知识库 / Obsidian”	kb-init	knowledge-base.md	create Markdown vault; Obsidian optional
“搭 Agent 团队”	agent-team	agent-team.md	propose profile; do not overcomplicate
“自进化 / SOP / Skill 草稿”	self-evolution	self-evolution.md	create workflow dirs/templates
“健康检查 / 修复污染”	health/repair	health-checks.md, context-hygiene.md	diagnose + classify fixes
“安全/覆盖/安装来源”	safety review	safety-policy.md	block risky action until confirmed

Required output formats

Change plan

Goal:
Current level:
Target level:
Safe changes:
Needs confirmation:
Manual steps:
Files/directories affected:
Skills to install:
Verification commands/checks:
Rollback notes:

Final report

Current level → Target level:
Completed:
Skipped:
Installed/missing skills:
Created files/dirs:
Backups/patches:
Health score:
Risks/Pending confirmations:
Next 3 actions:

Workflow

1. Diagnose first

Check:

OpenClaw/gateway status, workspace path, skills dir, memory dir, docs dir,
AGENTS.md/MEMORY.md/SOUL.md/TOOLS.md/USER.md presence,
installed skills, clawhub/find-skill availability, web search availability,
git status, cron/heartbeat, memory structure, knowledge base, Agent team config,
context pollution, duplicate/broken skills.

Output maturity level:

Level 0 fresh install
Level 1 basic assistant
Level 2 assistant with memory
Level 3 AgentOS with knowledge base/workflows
Level 4 multi-agent + self-evolution + health checks
Level 5 advanced personal AgentOS

2. Bootstrap Stage 0: skill discovery + web search

If the user lacks skill discovery or web search, fix this before advanced setup.

Minimum survival package:

clawhub
find-skill
openclawmp
markdown

If clawhub is missing, recommend or run after confirmation:

npm i -g clawhub

Search/install examples:

clawhub search "web search"
clawhub install find-skill
clawhub install openclawmp
clawhub install markdown

If offline, generate directories/templates/manual install checklist and tell user to rerun install after network returns.

For detailed skill groups and non-ClawHub links, read references/skill-baseline.md.

3. Install baseline skill packages

Use packages from references/skill-baseline.md:

bootstrap-minimal
bootstrap-search
bootstrap-docs
bootstrap-agentos-core
bootstrap-skill-lab
bootstrap-engineering
bootstrap-design
bootstrap-creator

Default standard order:

1. bootstrap-minimal
2. bootstrap-search
3. bootstrap-docs
4. bootstrap-agentos-core
5. bootstrap-skill-lab

Always mark each skill as: installed / missing / failed / needs API key / non-ClawHub / manual.

4. Set up HOT/WARM/COLD memory

Create or propose:

MEMORY.md                 # HOT, ≤150 lines recommended
memory/index.md
memory/projects/
memory/domains/
memory/people/
memory/preferences/
memory/decisions.md
memory/gotchas.md
memory/archive/
memory/logs/
memory/raw/

Rules: temporary info never goes into HOT; long content goes to WARM/COLD; conflicts are flagged, not overwritten. See references/memory-architecture.md.

5. Set up personal knowledge base / Obsidian-friendly vault

Create Markdown-first vault under memory/wiki/; Obsidian is optional.

Recommended structure:

memory/wiki/00 Inbox/
memory/wiki/01 Projects/
memory/wiki/02 Areas/
memory/wiki/03 Resources/
memory/wiki/04 Concepts/
memory/wiki/05 People/
memory/wiki/06 Decisions/
memory/wiki/07 Workflows/
memory/wiki/08 Skills/
memory/wiki/09 Reviews/
memory/wiki/99 Archive/

Knowledge flow:

Capture → Distill → Link → Operationalize → Archive

Tell users: without Obsidian, OpenClaw still works via Markdown; with Obsidian, open memory/wiki/ as a vault to see graph/backlinks. See references/knowledge-base.md.

6. Configure Agent team

Offer three profiles:

single: main-agent
three-agent: architect → executor → auditor
six-agent: pm → architect → reasoner → coder → auditor → monitor

Do not force multi-agent complexity on beginners. See references/agent-team.md.

7. Establish self-evolution workflows

Create or propose:

memory/wiki/07 Workflows/TaskNotes/
memory/wiki/07 Workflows/SOP/
memory/wiki/07 Workflows/SkillDrafts/
memory/wiki/07 Workflows/ContextCaptures/
memory/wiki/07 Workflows/Checkpoints/
memory/wiki/07 Workflows/Evaluations/
memory/wiki/07 Workflows/SecurityIntake/

Flow:

Task → TaskNote → SOP → SkillDraft → vetter → official Skill → index

See references/self-evolution.md.

8. Health checks and repair/upgrade

Health check dimensions:

skills, memory, knowledge base, context hygiene, Agent team, cron/heartbeat,
OpenClaw service, logs, git state, security risks.

Output a score and separate P0/P1/P2 issues. For details, see references/health-checks.md and references/context-hygiene.md.

9. Verify

Before final report, verify what changed:

- directories/files exist
- protected files were not overwritten
- installed skills contain SKILL.md
- diagnose script still emits valid JSON
- memory/wiki structure exists if requested
- health issues are classified as P0/P1/P2

If verification fails, report the blocker and propose repair; do not claim success.

10. Final report

Always finish with:

current level, target level, completed changes, installed/missing skills,
created directories/files, backups, risks, pending confirmations, next steps.

Implementation notes

• For simple advisory requests, do not execute writes; produce a plan.
• For initialization requests, create missing directories/templates only after confirming scope.
• If using scripts, read or run files in scripts/ as needed. Scripts are helpers, not authority; safety rules above win.