InstallerAuditor

Validates installs. Picks the right test tier, fixes what it can, escalates what it can't.

Setup

Use directly by reading this skill. Optionally register as a sub-agent:

sessions_spawn(agentId: "installer_auditor", task: "Validate: [WHAT]. Method: [METHOD]. Expected: [DESCRIPTION].", label: "installer-auditor")

Triggers

Coding agent completes
InstallerTester hands off (Phase 3)
Multi-file edit or config change
"validate this", "test this", "does this work"
/installer_auditor, /build-check
Before reporting "done" on non-trivial work

Input: What, Method, Expected behaviour

Output: `STATUS: done/retry/escalate` + tier results + checklist

Tier Selection

CLI / simple install / config change  → Tier 1: Smoke Test
Skill / MCP server / cron / feature   → Tier 2: Functional
Refactor / system integration          → Tier 3: Integration

Default to Tier 2 if unclear.

Tier 1: Smoke Test

Exists? (on PATH, at expected location, registered)
Runs without errors? (--help, --version)
One real command produces expected output?

Self-heal: one attempt. Obvious fix → re-test. Not obvious → escalate.

Tier 2: Functional

Run Tier 1.
List every capability. Test each with real input.
Test edge cases: bad input, missing data, empty states.
Test the actual use case, not just "does API respond."
For workflow skills: simulate primary use case end-to-end, execute each step.

Self-heal: two attempts. Same error twice → escalate.

Tier 3: Integration

Run Tier 1 + 2.
Test connections to dependent tools/skills/configs.
Regression check: everything working before still works.
If replaced something: confirm clean handoff.

Regressions → immediate escalate, no self-heal.

Self-Heal Rules

Do: reinstall cleanly, fix config value, fix path/permission, re-register. Never: delete existing tools, modify system files, create workarounds, retry same fix twice, make architectural decisions.

Escalation

What broke: one sentence. What you tried: what happened. Options: 1. [option] — [why] | 2. [option] — [trade-off] | 3. [option] My pick: [number] — [why]

Immediate escalate (no self-heal): regressions, security concerns, data risk, scope creep.

Structure & Wiring Check

Location

OpenClaw-standard:

Artifact	Location
Skills	`~/.openclaw/skills/<name>/SKILL.md`
Config	`~/.openclaw/openclaw.json`
Knowledge	`~/.openclaw/workspace/`
Memory	`~/.openclaw/workspace/memory/YYYY-MM-DD.md`

Workspace paths (customise for your setup): docs, scripts, plans.

Wiring

MCP server → registered? Shows in mcporter list?
Skill → folder matches name: in front matter?
CLI → in TOOLS.md?
API keys → stored securely? Wired in config?
Cron → correct target, model, channel?
Cross-references → Skill A calls Skill B, path works?

Completeness

Multi-step workflow skill without a completion checklist → flag and add one.

Discoverability

Reachable from TOOLS.md, AGENTS.md, skills list, or workspace links. Orphan → add reference.

Slash Command

If channel supports slash commands, suggest one.

Show It Working

Open the thing and do a real test run — not just "is it there."

Skill → open + trigger with real prompt
CLI → demo command with real output
MCP → call with actual data
Config → show status, confirm behaviour changed
Cron → show in list, trigger test run

Run it automatically. Not done until there's a live result.

Auto-Log

Write to ~/.openclaw/workspace/memory/YYYY-MM-DD.md: what, tier, result, self-heal actions.

Learn From Fixes

Log to ~/.openclaw/workspace/.learnings/SOLUTIONS.md: what broke, what fixed it, install type.

Completion Checklist

Tier tests executed
Structure check — location, wiring, discoverability
Show it working — live test result
Slash command — suggested or skipped
Auto-log — written

Unchecked = not done.

Constraints

Never report done without checklist complete
Max 2 fix attempts without checking in
Never modify system files without permission
Regressions = immediate escalate

installer-auditor