consensus-validation

name: consensus-validation description: "Usar cuando una decisión técnica o recomendación necesita validación por panel de jueces." summary: | Panel de 4 jueces: reflection, code-review, business, performance. Cada juez evalua independientemente. Score ponderado 0-1.0. Veto rule: security/GDPR finding anula score. Output: JSON. maturity: stable context: fork agent: consensus-orchestrator context_cost: medium category: "governance" tags: ["consensus", "validation", "multi-judge", "quality"] priority: "high"

Subagent Scope Guard

If you were dispatched as a subagent to execute a specific delegated task, skip this skill's full orchestration workflow. Execute only the assigned task, report result (DONE / DONE_WITH_CONCERNS / BLOCKED), and return. This guard prevents runaway skill activation in nested agent contexts.

Skill: Consensus Validation

Lanza 4 jueces especializados. Cada uno evalúa independientemente. Output: JSON estructurado con verdicts normalizados, score ponderado, dissents. El 4º juez (performance) usa performance-audit skill para detectar hotspots y anti-patterns.

Referencia: @docs/rules/domain/consensus-protocol.md

Decision Checklist

Before invoking consensus, answer sequentially:

1. Was the code-reviewer verdict REJECTED? -> If YES: consensus is mandatory
2. Does the change involve auth, payments, PII, or compliance? -> If YES: consensus is mandatory
3. Is the spec marked ambiguous: true? -> If YES: consensus is mandatory
4. Is this a simple CRUD or cosmetic UI change? -> If YES: consensus unnecessary, standard review suffices
5. Is the risk score < 26 (Low tier)? -> If YES: consensus unnecessary

Abort Conditions

• 2+ judges timeout -> CONDITIONAL verdict (insufficient data)
• Security veto by any judge -> auto-REJECT regardless of score

8-Step Protocol

1. Validar Input

type: spec | pr | decision
ref: file_path or PR_number

2. Formatear por Juez

• reflection-validator: Suposiciones, cadena causal, brechas de lógica
• code-reviewer: Código, diff, reglas SOLID, seguridad
• business-analyst: Reglas negocio, criterios aceptación, impacto
• performance-auditor: N+1 queries, async anti-patterns, complexity hotspots, bundle size

3-5. Invocar 4 Jueces en Paralelo (via dag-scheduling)

Dispatch via dag-scheduling skill — all 4 judges are independent (no deps), run as single parallel cohort. Timeout: 40s por juez (120s total). Cada juez devuelve: verdict + reasoning + confidence (0.0–1.0)

6. Normalizar Verdicts a 0/0.5/1.0

7. Veto Check

if (code_verdict == RECHAZADO) AND (security|gdpr|compliance in reasoning):
  final_verdict = REJECTED; return early
if (perf_verdict == REGRESSION) AND (severity == CRITICAL):
  final_verdict = REJECTED; return early

8. Calcular Score Ponderado

score = (reflection × 0.3) + (code × 0.3) + (business × 0.2) + (performance × 0.2)

if score >= 0.75: verdict = APPROVED
elif score >= 0.50: verdict = CONDITIONAL
else: verdict = REJECTED

8.5. Detectar Dissents

avg = (reflection + code + business + performance) / 4
for judge in [reflection, code, business, performance]:
  if abs(judge_score - avg) > 0.5:
    dissents.append(judge)

If dissents and verdict == APPROVED → downgrade to CONDITIONAL

9. Generar Output JSON

{
  "input": {type, ref, timestamp},
  "judges": [
    {name, verdict, score, reasoning, timeout, elapsed_ms}
  ],
  "veto": {triggered, reason},
  "summary": {
    "weighted_score": 0.62,
    "final_verdict": "CONDITIONAL",
    "dissents": ["business-analyst: ..."],
    "recommended_action": "corrections_required"
  }
}

Escribir a: output/consensus/YYYYMMDD-HHmmss-{type}-{ref}.json

Dissent Rules

Triggered si: abs(judge_score - promedio) > 0.5

Efecto:

• dissents + APPROVED → CONDITIONAL
• dissents + CONDITIONAL → CONDITIONAL
• dissents + REJECTED → REJECTED

Output: listar dissents con razonamiento

Error Handling & Timeline

Errors:

• Judge timeout: usar respuesta parcial (⚠️)
• 2+ timeouts: CONDITIONAL
• Veto triggered: REJECTED (final)

SLA: 120s máximo

Integration & Antipatterns

SDD: opt-in after spec-writer | PR: mandatory if rejected | ADR: opt-in | Audit: output/consensus/ NUNCA: override veto, modificar verdicts post-facto, saltarse jueces