Mcp Sampling Critic

---
name: mcp-sampling-critic
description: "MCP client sampling and critic-in-the-loop multi-model consensus / Sampling MCP client dan validasi konsensus multi-model AI."
author: "Roedy Rustam"
version: "4.8.0"

---

# mcp-sampling-critic — vibes-plug Skill

[English](#english) | [Bahasa Indonesia](#bahasa-indonesia)

---

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

### Orchestration & Integration
- **Protocol & Model Sampling**: `mcp-server-architect`, `ai-llm-integration-expert`, `frontier-ai-models-expert`
- **Adversarial Audit & Quality**: `autonomous-red-teamer`, `ai-code-review-autonomous`, `anti-slop`
- **Multi-Agent Consensus**: `speculative-multi-draft-synthesizer`, `multi-agent-orchestration`, `parallel-tool-orchestrator`
- **Testing & Verification**: `autonomous-tdd-debugger`, `property-mutation-testing-expert`, `formal-spec-z3-verifier`
- **Lifecycle Gate**: `zero-to-prod-orchestrator`, `session-memory-manager`

### Description
Production protocol for leveraging Model Context Protocol (MCP v1.x) Client Sampling (`sampling/createMessage`) to create recursive adversarial quality gates. In standard architectures, tools execute code blindly upon invocation. With MCP Sampling, a specialized Critic engine pauses execution, initiates a nested sampling call to the host LLM (or a secondary frontier model), and asks the Critic model to inspect proposed diffs against the 7 Anti-Slop Pillars, security boundaries, and invariant state machines. Writes are only permitted when the Critic grants an `APPROVED` verdict.

### Trigger Conditions
Activate this skill when:
- Designing autonomous agent loops that modify filesystem files or execute database migrations.
- Implementing the `CRITIC_VALIDATOR` swarm topology where an implementer agent's draft must pass an independent quality audit.
- Eliminating AI-generated placeholders (`// TODO`), silent error suppressions, or SQL injection vectors before pull request creation.
- Enforcing human-in-the-loop or critic-in-the-loop safeguards without requiring manual human clicking.

---

### Critic-in-the-Loop Protocol Architecture

```
PROPOSED CODE DRAFT
        │
        ▼
┌───────────────────────────┐
│ STATIC REGEX & AST FILTER │  Instant rule scan (P1-P7 Anti-Slop, silent catches)
└─────────────┬─────────────┘
              │
  ┌───────────┴───────────┐
  ▼                       ▼
FAIL (Score < 85)     PASS (Score >= 85)
  │                       │
Reject immediately   ┌────▼────────────────────────┐
with suggestions     │ MCP SAMPLING REQUEST        │  sampling/createMessage
                     │ to Host LLM (Independent)   │
                     └────────────┬────────────────┘
                                  │
                                  ▼
                        FINAL CRITIC VERDICT
                     (APPROVED -> Write to disk)
```

#### 1. MCP Sampling Request Construction (TypeScript)

```typescript
export interface SamplingMessagePayload {
  messages: Array<{
    role: 'user' | 'assistant';
    content: { type: 'text'; text: string };
  }>;
  systemPrompt: string;
  maxTokens: number;
  modelPreferences?: {
    hints?: Array<{ name: string }>;
    costPriority?: number;
    speedPriority?: number;
    intelligencePriority?: number;
  };
}

export function buildCriticSamplingPayload(
  targetFile: string,
  proposedCode: string,
  staticFindings: Array<{ ruleName: string; message: string }>
): SamplingMessagePayload {
  const findingsSummary = staticFindings.length > 0
    ? staticFindings.map(f => `- ${f.ruleName}: ${f.message}`).join('\n')
    : 'None detected by initial AST filter.';

  return {
    systemPrompt: `You are the Sovereign Anti-Slop Critic (2026 Edition).
Perform a strict code audit against:
1. Zero placeholder stubs (// TODO, // ...rest of code).
2. Zero silent error suppression (empty catch blocks).
3. 5-state interactive UI contract (hover, focus-visible, active, disabled).
4. SQL injection prevention (parameterized queries only).
Output JSON: { "verdict": "APPROVED" | "REJECTED", "score": 0-100, "rationale": "..." }`,
    maxTokens: 2048,
    modelPreferences: {
      intelligencePriority: 1.0,
      costPriority: 0.2
    },
    messages: [
      {
        role: 'user',
        content: {
          type: 'text',
          text: `Target: ${targetFile}\nPreliminary AST Warnings:\n${findingsSummary}\n\nCode to Audit:\n\`\`\`typescript\n${proposedCode}\n\`\`\``
        }
      }
    ]
  };
}
```

### Hard Guardrails (Declarative Invariants)
- **Recursive Critic Approval Gate**: High-risk mutating tool calls (filesystem writes, database DDL/DML, shell commands) MUST pass an independent `sampling/createMessage` adversarial audit with an explicit `APPROVED` verdict before execution.
- **Static Anti-Slop AST Pre-Filter**: Proposed code diffs MUST pass deterministic AST static filters (checking for empty catch blocks, placeholder stubs, unparameterized SQL) before initiating model sampling to conserve thinking tokens.
- **Model Diversity / Independent Evaluator Invariant**: The Critic sampling request MUST prioritize a distinct model family or separate evaluation prompt instance from the generator agent to prevent confirmation bias.
- **Unambiguous Structured JSON Contract**: Critic sampling responses MUST strictly conform to typed JSON contracts (`{ "verdict": "APPROVED" | "REJECTED", "score": number, "rationale": string }`), rejecting markdown conversational fluff.

### Non-Functional Requirements (NFR)
- **Critic Sampling Latency**: Round-trip sampling evaluation $\le 1,200\text{ms}$.
- **Zero False Approvals**: 0 passes granted to code with Anti-Slop violations or raw SQL injections.
- **Strict Schema Conformance**: 100% adherence to MCP `sampling/createMessage` specification.

### Sovereign Execution, Token Efficiency & Invariants
| Dimension | Sovereign Mandate & Deterministic Contract |
| :--- | :--- |
| **Frontier Model Cascade** | Medium / High (`GPT-5.6 Sol Medium`, `Gemini 4 Argon`, `Claude Sonnet 5.5 High`, `Grok 4.7 High Fast`) |
| **Availability Fallback** | `Gemini 4 Argon/Pro` -> `Gemini 3.8 Flash`; `Claude Opus 5.5` -> `Claude Opus 4.6` (Zero unhandled missing-model errors) |
| **KV-Cache & Token Protocol** | Pin static instructions & schemas at prompt prefix (`kv-cache-prefix-optimizer`); pass surgical `CONTEXT_MAP.md` slices (<250 tokens/call) during Swarm Fan-Out |
| **SaaT CLI & MCP Dispatch** | CLI: `vibes skill exec mcp-sampling-critic <audit\|eval\|verify\|generate\|blueprint>` / MCP: `vibes_skill_dispatch`, `vibes_ast_query`, `vibes_compile_cache`, `vibes_optimize_budget`, `vibes_verify_gate` |
| **NFR & Anti-Slop Invariants** | P99 API Gateway Latency < 85ms, OpenAPI 3.1 / TypeSpec strict schema parity, RFC 9457 Problem Details error contracts, Idempotency-Key enforcement; 100% complete zero-placeholder output verified via `node scripts/check-anti-slop.js --strict` |

---


---

<a name="bahasa-indonesia"></a>
## Bahasa Indonesia

### Integrasi Orkestrasi
- **Sampling Protokol & Model**: `mcp-server-architect`, `ai-llm-integration-expert`, `frontier-ai-models-expert`
- **Audit Adversarial & Kualitas**: `autonomous-red-teamer`, `ai-code-review-autonomous`, `anti-slop`
- **Konsensus Multi-Agen**: `speculative-multi-draft-synthesizer`, `multi-agent-orchestration`, `parallel-tool-orchestrator`
- **Pengujian & Verifikasi**: `autonomous-tdd-debugger`, `property-mutation-testing-expert`, `formal-spec-z3-verifier`
- **Gerbang Siklus Hidup**: `zero-to-prod-orchestrator`, `session-memory-manager`

### Deskripsi
Protokol produksi untuk memanfaatkan *Client Sampling* pada Model Context Protocol (`sampling/createMessage`) guna membangun gerbang kualitas adversarial rekursif. Menggunakan MCP Sampling, server atau subagen menghentikan eksekusi sementara, melakukan panggilan sampling ke LLM host untuk mengaudit usulan perubahan kode terhadap 7 Pilar Anti-Slop, keamanan injeksi SQL, dan penanganan error. Kode hanya ditulis ke disk apabila mendapatkan status `APPROVED`.

### Kondisi Pemicu
- Merancang alur agen otonom yang memodifikasi file repositori atau migrasi basis data.
- Menjalankan topologi swarm `CRITIC_VALIDATOR` di mana draf implementer harus diverifikasi auditor independen.
- Menghapus stub placeholder malas (`// TODO`) dan penelanan error secara otomatis sebelum commit.

---

### Pembuatan Payload Permintaan MCP Sampling (TypeScript)

```typescript
export interface PayloadPesanSampling {
  messages: Array<{
    role: 'user' | 'assistant';
    content: { type: 'text'; text: string };
  }>;
  systemPrompt: string;
  maxTokens: number;
}

export function bangunPayloadSamplingCritic(
  fileTarget: string,
  drafKode: string,
  peringatanAwal: string[]
): PayloadPesanSampling {
  return {
    systemPrompt: `Anda adalah Auditor Anti-Slop Berdaulat (Edisi 2026).
Lakukan audit kode ketat terhadap:
1. Ketiadaan stub placeholder (// TODO, // ...rest of code).
2. Ketiadaan penelanan error diam-diam (empty catch block).
3. Kontrak 5-state UI interaktif (hover, focus-visible, active, disabled).
4. Pencegahan injeksi SQL (wajib parameterized query).
Keluarkan JSON: { "verdict": "APPROVED" | "REJECTED", "score": 0-100, "rationale": "..." }`,
    maxTokens: 2048,
    messages: [
      {
        role: 'user',
        content: {
          type: 'text',
          text: `File: ${fileTarget}\nPeringatan AST Awal: ${peringatanAwal.join(', ')}\n\nKode yang Diaudit:\n\`\`\`typescript\n${drafKode}\n\`\`\``
        }
      }
    ]
  };
}
```

### Hard Guardrails (Invarian Deklaratif)
- **Gerbang Persetujuan Kritik Rekursif**: Pemanggilan tool berisiko tinggi (tulis sistem file, migrasi DDL/DML, perintah shell) WAJIB lolos audit adversarial via `sampling/createMessage` dengan status keputusan `APPROVED` eksplisit sebelum dieksekusi.
- **Pra-Penyaringan AST Statis Anti-Slop**: Usulan diff kode wajib lolos filter aturan statis AST deterministik (cek blok catch kosong, stub placeholder, SQL tidak berparameter) sebelum memicu panggilan sampling model untuk menghemat token penalaran.
- **Invarian Keberagaman Model / Evaluator Independen**: Permintaan sampling Critic wajib memprioritaskan famili model berbeda atau instans prompt evaluasi terpisah dari agen generator guna mencegah bias konfirmasi (*confirmation bias*).
- **Kontrak JSON Terstruktur Tanpa Ambiguitas**: Respons sampling Critic wajib mematuhi skema JSON bertipe (`{ "verdict": "APPROVED" | "REJECTED", "score": number, "rationale": string }`), menolak basa-basi percakapan markdown.

### Non-Functional Requirements (NFR)
- **Latensi Sampling Critic**: Waktu evaluasi round-trip sampling $\le 1.200\text{ms}$.
- **Zero False Approval**: 0 kelolosan yang diberikan pada kode dengan pelanggaran Anti-Slop atau celah injeksi SQL.
- **Kepatuhan Protokol Ketat**: 100% kepatuhan terhadap spesifikasi MCP `sampling/createMessage`.

### Matriks Eksekusi Berdaulat, Efisiensi Token & Invarian
| Dimensi | Mandat Berdaulat & Kontrak Deterministik |
| :--- | :--- |
| **Kaskade Model Frontier** | Medium / High (`GPT-5.6 Sol Medium`, `Gemini 4 Argon`, `Claude Sonnet 5.5 High`, `Grok 4.7 High Fast`) |
| **Invarian Fallback Ketersediaan** | `Gemini 4 Argon/Pro` -> `Gemini 3.8 Flash`; `Claude Opus 5.5` -> `Claude Opus 4.6` (Bebas gagal saat model belum tersedia/kuota penuh) |
| **Protokol Hemat Token & KV-Cache** | Kunci instruksi & skema statis di prefix awal (`kv-cache-prefix-optimizer`); kirim potongan `CONTEXT_MAP.md` (<250 token/panggilan) saat Fan-Out |
| **Eksekusi SaaT CLI & MCP** | CLI: `vibes skill exec mcp-sampling-critic <audit\|eval\|verify\|generate\|blueprint>` / MCP: `vibes_skill_dispatch`, `vibes_ast_query`, `vibes_compile_cache`, `vibes_optimize_budget`, `vibes_verify_gate` |
| **Invarian NFR & Anti-Slop** | Latensi P99 API < 85ms, paritas skema ketat OpenAPI 3.1 / TypeSpec, kontrak error RFC 9457 Problem Details, penegakan Idempotency-Key; kode 100% tuntas tanpa placeholder diverifikasi melalui `node scripts/check-anti-slop.js --strict` |