InstallerTester
No tool gets installed without passing this check.
Requirements
| Skill | Role | Required? |
|---|
| Post-install validation | Recommended |
| Security vetting for external skills | Required if installing a skill |
Triggers
- "install X", "add X", "set up X", "should we use X"
- GitHub/npm/brew link shared
- "is X worth installing", "compare X to Y"
- ,
Phase 0: Docs & Security
- Run
openclaw docs "<tool-name>"
- If target is a skill → activate , get ✅ SAFE or ⚠️ CAUTION (with approval) before proceeding. Fallback: manually review SKILL.md for red flags.
Phase 1: Pre-Install Check
Run in order. Stop and report if any check is a clear no.
Check 1 — What does it do?
State what it does in one sentence. Not what the user thinks — what it actually does. Can't determine? Stop.
Check 2 — What would it enhance or fix?
Search error logs, memory, session history, TOOLS.md, skills for:
- Friction: errors, workarounds, slow steps, token waste
- Enhancement: faster, cheaper, more reliable, new capability
Neither found → "Shiny toy — skip."
Check 3 — Already covered?
Check docs results from Phase 0, then:
bash
mcporter list # optional — requires mcporter
which <toolname> 2>/dev/null || echo "not installed"
openclaw skills list 2>&1 | grep -i "<toolname>"
grep -i "<toolname>" ~/.openclaw/workspace/TOOLS.md
Verdict: Not covered → Check 4 | Partially covered → Check 4 | Fully covered → stop, report what covers it.
Check 4 — Better than what we have?
If overlap: quantify improvement (tokens, speed, reliability, capability).
Health check: last commit, stars, issues, bus factor, dependency count, red flags (archived, abandoned, no docs).
Decision Output
One summary line:
Coverage: [not covered/partial/fully covered] | Quality: active/stale/dead | Verdict: Install / Skip / Flag
Wait for approval.
Phase 2: Install
Follow the tool's install instructions (Homebrew, npm, pip,
).
- Dependency missing or install fails → stop, report. Don't improvise.
- API keys → platform secure credential system. Never inline.
- Record install method (needed for rollback).
Phase 3: Validation
If InstallerAuditor is a registered agent:
sessions_spawn(agentId: "installer_auditor", task: "Validate: [WHAT]. Method: [METHOD]. Expected: [DESCRIPTION].", label: "installer-auditor")
Otherwise read the
skill and follow its tiers directly.
Validation fails → uninstall, report.
Phase 4: Document
Log in TOOLS.md: name, purpose, invocation, priority order. Check OpenClaw docs first for native coverage.
Completion Checklist
Unchecked without a reason = not done.