InstallerTester

No tool gets installed without passing this check.

Requirements

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"
• /installer_tester, /software-check

Phase 0: Docs & Security

1. Run openclaw docs "<tool-name>" (fallback: https://docs.openclaw.ai). Informs Check 3.
2. If target is a skill → activate skill-vetter, 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:

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, claude mcp add).

• 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 installer-auditor 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

• Phase 0 — docs + vetter (if skill)
• Check 1 — one-sentence description
• Check 2 — friction/enhancement or "shiny toy"
• Check 3 — coverage verdict with real checks
• Check 4 — health assessed (if not fully covered)
• Decision — verdict delivered, approval received
• Phase 2 — installed (method recorded)
• Phase 3 — validated
• Phase 4 — documented

Unchecked without a reason = not done.