preflight-audit

name: preflight-audit description: T-7 / T-3 / T-1 race-readiness audit. Fans out 4 parallel auditors (responsive · date freshness · demo-data · end-to-end walkthrough) then adversarially verifies the highest-impact findings. Run before each cadence checkpoint and at race-start.

Preflight Audit

This skill triggers the same multi-agent workflow that has caught 11+ confirmed critical/high findings during the run-up (admin mobile overflows, mock data leaking into War Room hero, stale "May 2026" copy).

Usage

User invokes: /preflight-audit

How it works

1. Spawn the workflow via the Workflow tool — script is below.
2. Phase 1: 4 default-agents (each w/ FINDING_SCHEMA) in parallel:
   • responsive — viewport audit for iPad portrait 768px, iPhone 375-430px, Android phone 360-412px, Android tablet 800-1280px
   • date-audit — wrong/stale dates, race-config integrity, countdown ticking
   • demo-cleanup — every mock/hardcoded name still in code
   • e2e-walkthrough — non-tech crew flow review across 6 flows
3. Phase 2: parallel adversarial verifiers on every critical/high finding — default-skeptical, only real=true survives.
4. Skill returns confirmed findings, rejected (false positives + reasoning), low/medium triage.

Workflow script

export const meta = {
  name: 'preflight-audit',
  description: 'Race-readiness audit (4 auditors + adversarial verify)',
  phases: [
    { title: 'Investigate', detail: '4 parallel auditors' },
    { title: 'Verify', detail: 'adversarial pass on critical/high' },
  ],
}

const REPO = '/Users/vishalbehal/Developer-local/raam-2026'
const TODAY = (args && args.today) || 'today'
const RACE_START = '2026-06-16'

const FINDING_SCHEMA = {
  type: 'object',
  additionalProperties: false,
  required: ['findings', 'summary'],
  properties: {
    summary: { type: 'string' },
    findings: {
      type: 'array',
      items: {
        type: 'object',
        additionalProperties: false,
        required: ['title', 'severity', 'file', 'line', 'evidence', 'fix'],
        properties: {
          title: { type: 'string' },
          severity: { enum: ['critical', 'high', 'medium', 'low'] },
          file: { type: 'string' },
          line: { type: 'number' },
          evidence: { type: 'string' },
          fix: { type: 'string' },
        },
      },
    },
  },
}

// (full prompts: see prior runs in session — keep responsive/date/demo/e2e bodies)

const phase1 = await parallel([
  () => agent('Audit responsive at ' + REPO + '. Today ' + TODAY + '. Race ' + RACE_START + '.', { phase: 'Investigate', label: 'responsive', schema: FINDING_SCHEMA }),
  () => agent('Audit dates at ' + REPO + '. Today ' + TODAY + '. Race ' + RACE_START + '.', { phase: 'Investigate', label: 'date-audit', schema: FINDING_SCHEMA }),
  () => agent('Audit demo data at ' + REPO + '.', { phase: 'Investigate', label: 'demo-cleanup', schema: FINDING_SCHEMA }),
  () => agent('Walk through 6 user flows at ' + REPO + '.', { phase: 'Investigate', label: 'e2e', schema: FINDING_SCHEMA }),
])

const all = phase1.filter(Boolean).flatMap(r => (r.findings || []))
const critHigh = all.filter(f => f.severity === 'critical' || f.severity === 'high')

const VERIFY_SCHEMA = { type: 'object', required: ['real', 'reasoning'], properties: { real: { type: 'boolean' }, reasoning: { type: 'string' } } }

const verified = await parallel(critHigh.slice(0, 20).map(f => () =>
  agent('Verify: ' + f.title + ' at ' + f.file + ':' + f.line + '. Evidence: ' + f.evidence + '. Fix: ' + f.fix, { phase: 'Verify', schema: VERIFY_SCHEMA }).then(v => ({ ...f, verdict: v }))))

return {
  totalFindings: all.length,
  confirmedHighImpact: verified.filter(v => v.verdict?.real),
  rejected: verified.filter(v => !v.verdict?.real),
  lowMedium: all.filter(f => f.severity === 'medium' || f.severity === 'low'),
}

When to run

• T-7 (Jun 9): baseline before crew testing pings
• T-3 (Jun 13): catch regressions from late race-week additions
• T-1 (Jun 15): final gate before race-day env flips
• Race-start (Jun 16): post-flip sanity check
• Any time something looks off during the race

Acting on findings

The skill returns structured data — surface it as a punch list in chat, then offer to fix the confirmed critical/high inline. Don't auto-fix without user nod (changes to live race-week code).