CivBench 2026: Tolok Ukur Long-Horizon & Evaluasi Tool-Mediated Multi-Agent melalui Model Context Protocol (MCP)

// CivBench Model Context Protocol (MCP) Runtime Engine & Proactive Monitoring Interceptor
// Menjamin Kepatuhan Eksekusi Long-Horizon (300+ Turns) & Validasi Komitmen RAG@10
import { EventEmitter } from 'events';

export interface MCPToolDefinition {
  name: string;
  description: string;
  parameters: Record;
  category: 'tactical' | 'economy' | 'diplomacy' | 'strategic_monitoring';
}

export interface StrategicReflectionCommitment {
  id: string;
  statedAtTurn: number;
  commitmentType: 'research' | 'city_production' | 'military_reinforcement' | 'diplomatic_alliance';
  targetEntity: string;
  deadlineTurn: number; // Stated at turn T, must be executed before T + 10 (RAG@10)
  status: 'PENDING' | 'EXECUTED' | 'ABANDONED_DRIFT';
  executionTurn?: number;
}

export interface TurnStateTelemetry {
  turnNumber: number;
  lastVictoryProgressQueryTurn: number;
  proactiveMonitoringIntervalTurns: number;
  activeCommitments: StrategicReflectionCommitment[];
  toolInvocationsThisTurn: string[];
}

export class CivBenchMCPRuntime extends EventEmitter {
  private registeredTools: Map = new Map();
  private commitmentsHistory: StrategicReflectionCommitment[] = [];
  private telemetry: TurnStateTelemetry = {
    turnNumber: 1,
    lastVictoryProgressQueryTurn: 0,
    proactiveMonitoringIntervalTurns: 20, // Strict Playbook Rule: Query every 20 turns
    activeCommitments: [],
    toolInvocationsThisTurn: [],
  };

  constructor() {
    super();
    this.registerStandardCivMCPTools();
  }

  // 1. Inisialisasi 76 MCP Tools Environment
  private registerStandardCivMCPTools(): void {
    const coreTools: MCPToolDefinition[] = [
      { name: 'query_victory_progress', description: 'Query latent strategic victory scores and milestone counters of all civs', parameters: {}, category: 'strategic_monitoring' },
      { name: 'inspect_city_production_queue', description: 'Inspect or reorder city build queues', parameters: { cityId: 'string' }, category: 'economy' },
      { name: 'choose_tech_research', description: 'Commit civil research tree branch', parameters: { techId: 'string' }, category: 'economy' },
      { name: 'dispatch_military_unit', description: 'Move tactical unit on hexagonal coordinate space', parameters: { unitId: 'string', targetHex: 'string' }, category: 'tactical' },
      { name: 'propose_diplomatic_treaty', description: 'Send diplomatic trade or peace delegation', parameters: { targetCiv: 'string', terms: 'object' }, category: 'diplomacy' },
    ];
    for (const t of coreTools) {
      this.registeredTools.set(t.name, t);
    }
  }

  // 2. Proactive Monitoring Interceptor: Mengatasi Blindspot Kegagalan Deteksi Lawan
  public validateTurnCompletion(): { canEndTurn: boolean; warningMessage?: string; requiredTool?: string } {
    const turnsSinceLastVictoryQuery = this.telemetry.turnNumber - this.telemetry.lastVictoryProgressQueryTurn;

    // Evaluasi Proactive Monitoring Rate (PMR) Enforcement
    if (turnsSinceLastVictoryQuery >= this.telemetry.proactiveMonitoringIntervalTurns) {
      return {
        canEndTurn: false,
        warningMessage: `[CIVBENCH_PMR_VIOLATION]: Turn ${this.telemetry.turnNumber} breached 20-turn monitoring lease. Latent state must be queried.`,
        requiredTool: 'query_victory_progress',
      };
    }

    return { canEndTurn: true };
  }

  // 3. RAG@10 Commitment Tracker: Mencegah Silent Drift pada Planning Reflection
  public recordStrategicReflection(
    statedCommitments: Array>
  ): void {
    for (const c of statedCommitments) {
      const commitment: StrategicReflectionCommitment = {
        ...c,
        id: `com-${this.telemetry.turnNumber}-${nanoid(6)}`,
        statedAtTurn: this.telemetry.turnNumber,
        deadlineTurn: this.telemetry.turnNumber + 10,
        status: 'PENDING',
      };
      this.telemetry.activeCommitments.push(commitment);
      this.commitmentsHistory.push(commitment);
    }
  }

  // 4. Intercept Invocations & Resolve Commitments
  public executeToolCall(toolName: string, params: Record): { success: boolean; p99LatencyMs: number } {
    const start = performance.now();
    this.telemetry.toolInvocationsThisTurn.push(toolName);

    if (toolName === 'query_victory_progress') {
      this.telemetry.lastVictoryProgressQueryTurn = this.telemetry.turnNumber;
    }

    // Check apakah tool invocation memenuhi commitment RAG@10
    this.telemetry.activeCommitments.forEach((com) => {
      if (com.status === 'PENDING' && com.targetEntity === params.targetEntity) {
        com.status = 'EXECUTED';
        com.executionTurn = this.telemetry.turnNumber;
      }
    });

    return { success: true, p99LatencyMs: Number((performance.now() - start).toFixed(3)) };
  }

  // 5. Advance Turn & Reconcile RAG@10 Penalties
  public advanceToNextTurn(): { turn: number; expiredCommitments: number; adherenceScore: number } {
    const validation = this.validateTurnCompletion();
    if (!validation.canEndTurn) {
      throw new Error(validation.warningMessage);
    }

    // Evaluasi komitmen yang kadaluarsa (>10 turns tanpa aksi nyata)
    let expired = 0;
    this.telemetry.activeCommitments.forEach((com) => {
      if (com.status === 'PENDING' && this.telemetry.turnNumber >= com.deadlineTurn) {
        com.status = 'ABANDONED_DRIFT';
        expired++;
      }
    });

    // Prune resolved or abandoned commitments dari active state
    this.telemetry.activeCommitments = this.telemetry.activeCommitments.filter((c) => c.status === 'PENDING');
    this.telemetry.toolInvocationsThisTurn = [];
    this.telemetry.turnNumber++;

    const executedTotal = this.commitmentsHistory.filter((c) => c.status === 'EXECUTED').length;
    const totalHistorical = this.commitmentsHistory.filter((c) => c.status !== 'PENDING').length;
    const adherenceScore = totalHistorical > 0 ? Number(((executedTotal / totalHistorical) * 100).toFixed(1)) : 100;

    return {
      turn: this.telemetry.turnNumber,
      expiredCommitments: expired,
      adherenceScore,
    };
  }
}