Supersearch — Maximum-Effort Research Mode

export const meta = {
  name: 'supersearch',
  description: 'Exhaustive research: parallel aspect agents, source audit, contradiction hunt, timeline, practitioner reality, devil\'s advocate, synthesis-prep, quality gate.',
  phases: [
    { title: 'Research' },
    { title: 'Audit' },
    { title: 'Contradict' },
    { title: 'Timeline' },
    { title: 'Practitioner' },
    { title: 'Synthesize' },
    { title: 'DevilsAdvocate' },
    { title: 'QualityGate' },
  ],
}

const { topic, intent, highStakes, sourcePref, angle, aspects } = args

// ---- Structured-output schemas (force validated returns) ----
const FINDING = {
  type: 'object',
  required: ['claim', 'url', 'title', 'tier', 'date', 'confidence', 'evidenceType'],
  properties: {
    claim: { type: 'string' },
    url: { type: 'string' },
    title: { type: 'string' },
    tier: { type: 'string', enum: ['S', 'A', 'B', 'C', 'D'] },
    date: { type: 'string' },
    confidence: { type: 'string', enum: ['high', 'medium', 'low'] },
    evidenceType: { type: 'string', enum: ['direct', 'inference'] },
  },
}
const FINDINGS_SCHEMA = {
  type: 'object',
  required: ['aspect', 'findings'],
  properties: {
    aspect: { type: 'string' },
    coverage: { type: 'object', properties: { attempted: { type: 'array', items: { type: 'string' } }, useful: { type: 'array', items: { type: 'string' } }, failed: { type: 'array', items: { type: 'string' } } } },
    findings: { type: 'array', items: FINDING },
    contradictions: { type: 'array', items: { type: 'string' } },
    practitionerReality: { type: 'string' },
    gaps: { type: 'array', items: { type: 'string' } },
    bestSources: { type: 'array', items: { type: 'object', properties: { url: { type: 'string' }, title: { type: 'string' }, why: { type: 'string' } } } },
  },
}
const AUDIT_SCHEMA = { type: 'object', properties: {
  duplicateUrls: { type: 'array', items: { type: 'string' } },
  repeatedDomains: { type: 'array', items: { type: 'string' } },
  weakSources: { type: 'array', items: { type: 'string' } },
  singleSourceClaims: { type: 'array', items: { type: 'string' } },
  outdatedClaims: { type: 'array', items: { type: 'string' } },
  likelyAiGenerated: { type: 'array', items: { type: 'string' } },
  conflictsOfInterest: { type: 'array', items: { type: 'string' } },
  missingPrimarySources: { type: 'array', items: { type: 'string' } },
  unsupportedNumbers: { type: 'array', items: { type: 'string' } },
  citationLoops: { type: 'array', items: { type: 'string' } },
  recommendedRemovals: { type: 'array', items: { type: 'string' } },
  recommendedReplacements: { type: 'array', items: { type: 'string' } },
  report: { type: 'string' },
}, required: ['report'] }
const CONTRA_SCHEMA = { type: 'object', properties: {
  counterclaims: { type: 'array', items: { type: 'object', properties: {
    claim: { type: 'string' }, evidence: { type: 'array', items: { type: 'object', properties: { url: { type: 'string' }, title: { type: 'string' } } } },
    weakensSynthesis: { type: 'boolean' }, strength: { type: 'string', enum: ['high', 'medium', 'low'] } } } },
  unresolvedDebates: { type: 'array', items: { type: 'string' } },
  claimsToSoften: { type: 'array', items: { type: 'string' } },
}, required: ['counterclaims'] }
const TIMELINE_SCHEMA = { type: 'object', properties: {
  milestones: { type: 'array', items: { type: 'object', properties: { date: { type: 'string' }, event: { type: 'string' }, url: { type: 'string' } } } },
  recentChanges: { type: 'array', items: { type: 'object', properties: { date: { type: 'string' }, change: { type: 'string' }, url: { type: 'string' } } } },
  obsoleteClaims: { type: 'array', items: { type: 'string' } },
  upcoming: { type: 'array', items: { type: 'object', properties: { date: { type: 'string' }, event: { type: 'string' } } } },
}, required: ['milestones'] }
const PRACT_SCHEMA = { type: 'object', properties: {
  reports: { type: 'array', items: { type: 'object', properties: { who: { type: 'string' }, url: { type: 'string' }, finding: { type: 'string' }, breadth: { type: 'string', enum: ['anecdote', 'broad'] } } } },
  summary: { type: 'string' },
}, required: ['summary'] }
const SYNTH_SCHEMA = { type: 'object', properties: {
  insights: { type: 'array', items: { type: 'object', properties: { insight: { type: 'string' }, supportingSources: { type: 'array', items: { type: 'string' } }, confidence: { type: 'string', enum: ['high', 'medium', 'low'] } } } },
  patterns: { type: 'array', items: { type: 'string' } },
  causalChains: { type: 'array', items: { type: 'string' } },
  claimSourceMatrix: { type: 'array', items: { type: 'object', properties: { claim: { type: 'string' }, sources: { type: 'array', items: { type: 'string' } }, weak: { type: 'boolean' } } } },
  caveats: { type: 'array', items: { type: 'string' } },
}, required: ['insights'] }
const DEVIL_SCHEMA = { type: 'object', properties: {
  attacks: { type: 'array', items: { type: 'object', properties: { targetClaim: { type: 'string' }, attack: { type: 'string' }, severity: { type: 'string', enum: ['high', 'medium', 'low'] } } } },
  claimsToRemove: { type: 'array', items: { type: 'string' } },
  biases: { type: 'array', items: { type: 'string' } },
}, required: ['attacks'] }
const GATE_SCHEMA = { type: 'object', required: ['verdict', 'scores', 'counts'], properties: {
  scores: { type: 'object', properties: {
    saturation: { type: 'number' }, sourceDiversity: { type: 'number' }, tierQuality: { type: 'number' },
    evidenceDepth: { type: 'number' }, recency: { type: 'number' }, contradictionCoverage: { type: 'number' },
    practicalUsefulness: { type: 'number' }, claimHygiene: { type: 'number' }, uncertaintyHonesty: { type: 'number' } } },
  counts: { type: 'object', properties: { sourcesIncluded: { type: 'number' }, saTierSources: { type: 'number' }, insights: { type: 'number' } } },
  verdict: { type: 'string', enum: ['PASS', 'WARN', 'FAIL'] },
  failures: { type: 'array', items: { type: 'string' } },
  weakestAspects: { type: 'array', items: { type: 'string' } },
  recommendation: { type: 'string' },
} }

const prefLine = sourcePref ? `SOURCE PREFERENCE (obey): ${sourcePref}` : ''
const angleLine = angle ? `ANALYTICAL ANGLE (bias toward, do not lose general coverage): ${angle}` : ''
const stakesLine = highStakes ? 'HIGH-STAKES TOPIC: only load-bearing claims from primary/official/peer-reviewed sources; flag all uncertainty.' : ''

function aspectBrief(a) {
  return [
    'You are a maximum-effort research agent investigating ONE aspect of a larger topic.',
    `\nOVERALL TOPIC:\n${topic}`,
    `\nINTENT TYPE: ${intent}`,
    `\nYOUR ASPECT:\n${a.name} — lens: ${a.lens}`,
    `\nSEARCH QUERIES:\n${(a.queries || []).map(q => '- ' + q).join('\n')}`,
    prefLine, angleLine, stakesLine,
    ASPECT_STANDARD,
  ].filter(Boolean).join('\n')
}

phase('Research')
const aspectResults = (await parallel(
  aspects.map(a => () => agent(aspectBrief(a), { label: `aspect:${a.name}`, phase: 'Research', schema: FINDINGS_SCHEMA }))
)).filter(Boolean)
const allFindings = aspectResults.flatMap(r => r.findings || [])
log(`Research wave done: ${allFindings.length} findings across ${aspectResults.length} aspects`)

const corpus = JSON.stringify(allFindings).slice(0, 120000)

phase('Audit')
const [audit, contradictions, timeline, practitioner] = await parallel([
  () => agent(`Source-auditor. Inspect this finding corpus and produce a source quality report. Flag: duplicate URLs, repeated domains, weak/low-substance sources, single-source claims, outdated claims, likely-AI-generated pages, conflicts of interest, missing primary sources, unsupported numbers, suspiciously repeated claims, and citation loops (sources citing each other). Recommend removals and replacements.\n\nCORPUS:\n${corpus}`, { label: 'source-audit', phase: 'Audit', schema: AUDIT_SCHEMA }),
  () => agent(`Contradiction-hunter for: ${topic}\nActively search AGAINST the emerging consensus. Ask: what would make the main conclusion false? Who disagrees and why? Do benchmarks/claims fail to reproduce? Are vendors exaggerating? Do users report different results than docs? Has it changed recently? Are obsolete claims being repeated? Is there a hidden incentive shaping the narrative? Does it work in demos but fail in practice? Run new searches; do not just re-read the corpus.\n\nCORPUS:\n${corpus}`, { label: 'contradiction-hunt', phase: 'Contradict', schema: CONTRA_SCHEMA }),
  () => agent(`Timeline & change-detection agent for: ${topic}\nIdentify major milestones, version/release changes, regulatory/market shifts, changes in the last 6–18 months, claims once true but now obsolete, and known upcoming events/deadlines/roadmaps. Use EXACT dates where possible (verify, never invent).\n\nCORPUS:\n${corpus}`, { label: 'timeline', phase: 'Timeline', schema: TIMELINE_SCHEMA }),
  () => agent(`Practitioner-reality agent for: ${topic}\nFind grounded real-world reports from maintainers, operators, engineers, researchers, users, customers, and affected communities — issue trackers, postmortems, bug reports, forum threads, case studies. Distinguish one-off anecdote from broad evidence.\n\nCORPUS:\n${corpus}`, { label: 'practitioner-reality', phase: 'Practitioner', schema: PRACT_SCHEMA }),
])

phase('Synthesize')
const synthPrep = await agent(`Synthesis-prep agent. From the corpus + audit + contradictions + timeline + practitioner data, produce: deduped insights (each synthesizing 2–3+ independent findings, with supporting source URLs and a confidence level), recurring patterns, causal chains, a claim→source matrix flagging single-weak-source claims, and caveats the report cannot responsibly exceed.\n\nCORPUS:\n${corpus}\n\nAUDIT:\n${JSON.stringify(audit).slice(0,15000)}\n\nCONTRADICTIONS:\n${JSON.stringify(contradictions).slice(0,15000)}\n\nTIMELINE:\n${JSON.stringify(timeline).slice(0,8000)}\n\nPRACTITIONER:\n${JSON.stringify(practitioner).slice(0,8000)}`, { label: 'synthesis-prep', phase: 'Synthesize', schema: SYNTH_SCHEMA })

phase('DevilsAdvocate')
const devil = await agent(`Devil's-advocate. Attack the report's OWN strongest claims. For each top insight below, mount the best refutation you can, rate severity, name any selection/confirmation/vendor bias, and list claims that should be removed or softened.\n\nINSIGHTS:\n${JSON.stringify(synthPrep.insights || []).slice(0,20000)}`, { label: 'devils-advocate', phase: 'DevilsAdvocate', schema: DEVIL_SCHEMA })

phase('QualityGate')
const gate = await agent(`Quality-gate. Score 0–1 each: saturation, sourceDiversity, tierQuality, evidenceDepth, recency, contradictionCoverage, practicalUsefulness, claimHygiene, uncertaintyHonesty. Count sourcesIncluded, saTierSources, insights. Minimums: ≥45 sources, ≥15 S/A-tier, ≥10 insights, audit present, contradictions present, practitioner present, devil's-advocate present. Verdict PASS (publishable for a serious decision) / WARN (useful but limited) / FAIL (insufficient). If FAIL, list weakestAspects to re-research.\n\nFINDINGS: ${allFindings.length}\nSYNTH: ${JSON.stringify(synthPrep).slice(0,12000)}\nAUDIT: ${JSON.stringify(audit).slice(0,6000)}\nCONTRA: ${JSON.stringify(contradictions).slice(0,6000)}`, { label: 'quality-gate', phase: 'QualityGate', schema: GATE_SCHEMA })

// One automatic targeted re-research pass on FAIL
let reResearch = null
if (gate.verdict === 'FAIL' && (gate.weakestAspects || []).length) {
  phase('Research')
  reResearch = (await parallel(gate.weakestAspects.map(name => () => {
    const a = aspects.find(x => x.name === name) || { name, lens: 'targeted', queries: [] }
    return agent(aspectBrief(a) + '\n\nRE-RESEARCH PASS: prior coverage was insufficient. Find NEW primary sources and fill the named gaps.', { label: `re:${name}`, phase: 'Research', schema: FINDINGS_SCHEMA })
  }))).filter(Boolean)
}

return { aspectResults, audit, contradictions, timeline, practitioner, synthPrep, devil, gate, reResearch }