speculative-multi-draft-synthesizer — vibes-plug Skill

English | Bahasa Indonesia

---

<a name="english"></a>

English

Orchestration & Integration

• Multi-Agent Swarms & Consensus: multi-agent-orchestration, parallel-tool-orchestrator, mcp-sampling-critic
• Frontier Reasoning Models: frontier-ai-models-expert, gemini-agent-booster, test-time-compute-optimizer
• Code AST & Sandboxing: living-codebase-ast-graph, ephemeral-wasm-sandbox-executor
• Quality & Anti-Slop: anti-slop, ai-code-review-autonomous, formal-spec-z3-verifier
• Lifecycle & Architecture: brainstorming, zero-to-prod-orchestrator, session-memory-manager

Description

Production guide for architecting speculative multi-draft code synthesis pipelines. Dispatches high-concurrency, fast-draft subagents (powered by Gemini Flash or lightweight models) to generate competing implementation hypotheses in parallel, then feeds these candidates into a frontier arbiter model (Gemini 4 Pro, Claude Opus 5.5, or GPT Astra 6). The arbiter performs semantic AST reconciliation, extracting optimal algorithms, error boundaries, and type safety into a single superior implementation.

Trigger Conditions

Activate this skill when:

• Solving complex, ambiguous algorithmic challenges where multiple valid architectural approaches exist.
• Designing high-scale microservices or critical algorithms where relying on a single model output risks blind spots.
• Combining strengths of multiple model families (e.g. Gemini long-context structural view + Claude rigorous type precision).
• Synthesizing large refactoring proposals submitted by concurrent swarm subagents.

---

Core Concepts & Patterns

1. Speculative Synthesis Swarm Topology

User Intent / Spec
       │
       ├───────────────────┬───────────────────┐
       ▼                   ▼                   ▼
Draft Agent A       Draft Agent B       Draft Agent C
(Algorithmic Focus) (Error Resilience)  (Type Rigor)
       │                   │                   │
       └───────────────────┼───────────────────┘
                           ▼
                  [FRONTIER ARBITER]
            (Gemini 4 Pro / Claude Opus 5.5)
                           │
                           ▼
              AST Reconciliation & Scoring
                           │
                           ▼
             Optimal Synthesized Solution

2. Multi-Draft Arbiter Engine (TypeScript Implementation)

export interface CandidateDraft {
  agentId: string;
  model: string;
  code: string;
  rationale: string;
}

export interface SynthesisEvaluation {
  selectedAlgorithmSource: string;
  selectedErrorStrategySource: string;
  selectedTypingSource: string;
  synthesizedCode: string;
  synthesisRationale: string;
}

export class SpeculativeDraftSynthesizer {
  public static buildArbiterPrompt(spec: string, drafts: CandidateDraft[]): string {
    const formattedDrafts = drafts
      .map(
        (draft, index) =>
          `### DRAFT ${index + 1} (From ${draft.agentId} using ${draft.model}):\n` +
          `Rationale: ${draft.rationale}\n\n` +
          `\`\`\`typescript\n${draft.code}\n\`\`\``
      )
      .join('\n\n---\n\n');

    return [
      'You are the Master Frontier Arbiter in an autonomous AI swarm.',
      'Your task is to analyze the following candidate code drafts for the given specification, identify their individual strengths, and synthesize a single, flawless, production-ready implementation.',
      '',
      'SPECIFICATION:',
      spec,
      '',
      'CANDIDATE DRAFTS:',
      formattedDrafts,
      '',
      'SYNTHESIS MANDATES:',
      '1. Take the highest-performance algorithmic core from the best candidate.',
      '2. Take the most exhaustive, error-resilient boundary handling from the best candidate.',
      '3. Enforce the strictest, most accurate TypeScript types without using "any".',
      '4. Output 100% complete, working production code without ellipses or placeholders.'
    ].join('\n');
  }

  public static reconcileDrafts(evalResult: SynthesisEvaluation): string {
    return evalResult.synthesizedCode;
  }
}

3. Arbiter Selection Criteria

The arbiter reconciles drafts using four quantitative metrics:

• Type Soundness: Zero type assertions (as unknown), zero any, strict null safety.
• Complexity Bound: Minimal cyclomatic complexity and optimal Big-O time/space bounds.
• Failure Resilience: Complete error coverage, explicit domain error classes, no unhandled rejections.
• Resource Footprint: Minimal heap allocation, zero memory leaks, idiomatic resource cleanup.

---

Best Practices

1. Vary System Prompts Across Drafters: Instruct Draft A to prioritize raw algorithmic speed, Draft B to prioritize error resilience, and Draft C to prioritize minimal memory footprint.
2. Never Simply Pick One Winner: Direct the arbiter to extract and fuse the best aspects of each draft into a unified hybrid AST.
3. Run Ephemeral Sandbox Verification: Before finalizing the synthesis, run the candidate through ephemeral-wasm-sandbox-executor to verify that all unit tests pass.
4. Enforce Complete Implementations: The arbiter must never emit truncated blocks; output must be fully executable on first compilation.

---

Common Pitfalls to Avoid

Anti-Pattern	Operational Risk	Engineering Remedy
Asking one model to draft and critique itself	Shared blind spots and self-affirming hallucination	Spawn diverse model drafts (e.g. Gemini + Claude)
Naive text concatenation of code blocks	Syntax errors, duplicate variables, broken imports	Perform AST-level node reconciliation in arbiter
Discarding failed drafts without inspection	Lost insight into novel edge cases	Allow arbiter to review why certain edge branches were formed

---

<a name="bahasa-indonesia"></a>

Bahasa Indonesia

Integrasi Orkestrasi

• Swarm Multi-Agen & Konsensus: multi-agent-orchestration, parallel-tool-orchestrator, mcp-sampling-critic
• Model Penalaran Frontier: frontier-ai-models-expert, gemini-agent-booster, test-time-compute-optimizer
• AST Kode & Sandboxing: living-codebase-ast-graph, ephemeral-wasm-sandbox-executor
• Kualitas & Anti-Slop: anti-slop, ai-code-review-autonomous, formal-spec-z3-verifier
• Siklus Hidup & Arsitektur: brainstorming, zero-to-prod-orchestrator, session-memory-manager

Deskripsi

Panduan produksi untuk merancang pipeline sintesis kode berbasis draf spekulatif (speculative multi-draft). Menugaskan subagent berkecepatan tinggi secara paralel (seperti Gemini Flash atau model ringan) untuk menyusun beberapa hipotesis implementasi alternatif, kemudian menyerahkan draf-draf tersebut kepada model arbiter frontier (Gemini 4 Pro, Claude Opus 5.5, atau GPT Astra 6). Arbiter melakukan rekonsiliasi AST semantik guna menyatukan algoritma tercepat, penanganan error terkuat, dan ketepatan tipe terbaik menjadi satu solusi unggul.

Kondisi Pemicu

Aktifkan skill ini ketika:

• Menyelesaikan masalah algoritma kompleks atau ambigu yang memiliki beberapa pendekatan arsitektur valid.
• Merancang microservices skala besar yang berisiko jika hanya mengandalkan perspektif satu model AI.
• Menggabungkan keunggulan keluarga model yang berbeda (misalnya konteks luas Gemini dengan ketelitian tipe Claude).
• Menyintesis proposal refaktor besar yang dikirimkan oleh beberapa sub-agen secara bersamaan.

---

Konsep Inti & Pola Praktik

1. Topologi Swarm Sintesis Spekulatif

1. Penyebaran Draf (Fan-Out): Tiga sub-agen menghasilkan draf dengan fokus berbeda (Draf 1: performa algoritma, Draf 2: ketahanan error & edge cases, Draf 3: tipe data dan kesederhanaan arsitektur).
2. Evaluasi Arbiter (Consensus Arbitration): Model frontier membedah kelebihan dan kekurangan dari setiap draf.
3. Penyatuan Hibrida (AST Reconciliation): Arbiter mengambil bagian-bagian terbaik dari setiap kandidat untuk menghasilkan kode akhir yang lengkap.

2. Kriteria Seleksi Arbiter

• Ketelitian Tipe: Nol tipe any, penanganan null/undefined yang ketat.
• Kompleksitas Algoritma: Kompleksitas waktu dan memori (Big-O) yang paling optimal.
• Ketahanan Kegagalan: Penanganan error menyeluruh tanpa unhandled promise rejections.

---

Praktik Terbaik

1. Variasikan Fokus Instruksi Draf: Berikan peran yang berbeda pada tiap pembuat draf (kecepatan vs keamanan vs kesederhanaan).
2. Jangan Hanya Memilih Satu Pemenang: Perintahkan arbiter untuk memadukan modul-modul terbaik dari berbagai draf.
3. Verifikasi dengan Sandbox: Jalankan draf yang disintesis di ephemeral-wasm-sandbox-executor untuk membuktikan fungsionalitasnya.
4. Keluaran 100% Lengkap: Pastikan hasil akhir berupa kode produksi siap pakai tanpa placeholder.

---

Jebakan Umum yang Harus Dihindari

Praktik Buruk	Dampak Buruk	Solusi Rekayasa
Model mengkritik dan memperbaiki kodenya sendiri	Bias yang sama berulang (shared blind spot)	Gunakan model frontier yang berbeda sebagai arbiter independen
Penggabungan teks kode secara mentah	Kerusakan sintaksis dan deklarasi variabel ganda	Lakukan rekonsiliasi berbasis AST oleh arbiter
Membuang draf yang salah total	Kehilangan informasi tentang kemungkinan kegagalan	Berikan draf tersebut ke arbiter sebagai panduan apa yang harus dihindari