generation-dispatcher

스타10

포크0

업데이트2026년 6월 2일 03:52

Use when evaluating task determinism, dispatching to traditional code generators vs AI generation, enforcing contract consistency gates, and routing for confidence-based human review. This skill unifies D4.5 Generation Dispatch Domain and replaces deterministic-evaluator, code-generator-dispatcher, contract-consistency-gate, and confidence-based-reviewer.

설치

Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.

출처

aether-go

aether-go/skills

GitHub 저장소 열기 Creator 저장소 보기

다운로드

Generation Dispatcher

Overview

The unified skill for all L4→L4.5 generation dispatch work. This skill:

Evaluates the determinism score of component contracts/tasks
Routes tasks to appropriate traditional code generators (zero-token) or AI generation channels
Enforces contract consistency gates (GATE-001) on all generated code
Grades AI-generated code confidence and routes for human review

Replaces (consolidated from 4 skills):

deterministic-evaluator
code-generator-dispatcher
contract-consistency-gate
confidence-based-reviewer

When to Use

Need to route code generation task? ─────────┐
                                             │
Evaluating if task is deterministic? ────────┤
                                             │
Enforcing contract compliance gates? ────────┼─► Use generation-dispatcher
                                             │
Grading AI-generated code confidence? ───────┤
                                             │
Need economic optimization (zero-token)? ────┘

Use when:

Deciding whether to use traditional generator vs AI for a task
Enforcing contract consistency on generated code
Routing AI-generated code for human review based on confidence
Tracking generation cost savings (token economy)
Implementing the L4→L4.5 dispatch layer

Don't use when:

Pure requirement analysis (use requirement-analyzer)
Spec writing (use spec-engineer)
Actual code generation (use code-generator)
Test code generation (use test-planner)

Core Pattern

L4→L4.5 Dispatch Pipeline

Interface Contracts (L4 input from architecture-planner)
    │
    ├──► [1] Determinism Evaluation
    │         │
    │         │  Score: D(t) ∈ [0, 1]
    │         │  Threshold: θ = 0.80
    │         │
    │         ├── if D(t) ≥ 0.80 → Highly Deterministic
    │         │      └──► [2a] Dispatch to Traditional Generator
    │         │              │
    │         │              ▼
    │         │          Generated Code (zero-token)
    │         │
    │         └── if D(t) < 0.80 → Non-Deterministic
    │                  └──► [2b] Route to AI Generation
    │                          │
    │                          ▼
    │                      AI-Generated Code
    │                          │
    │                          ├──► [3] Contract Consistency Gate (GATE-001)
    │                          │         │
    │                          │         ├── PASS → [4] Confidence Grading
    │                          │         │              │
    │                          │         │              ├── ≥ 0.90 → Auto-approve
    │                          │         │              ├── 0.70-0.90 → Suggest confirmation
    │                          │         │              └── < 0.70 → Mandatory review
    │                          │         │
    │                          │         └── FAIL → [5] Iterate / Reject
    │                          │
    │                          ▼
    │                      Final Code (L4.5 output)
    │
    ▼

Determinism Evaluation Dimensions

Dimension	Weight	Description
Contract Completeness	0.30	How fully the contract specifies the task
Pattern Match	0.25	Match against known generator patterns (CRUD, REST, DTO)
Historical Success	0.20	Past success rate of similar tasks with traditional generators
Complexity	0.15	Cyclomatic/cognitive complexity of the task
Context Stability	0.10	Stability of inputs and expected outputs

Score Classification

Range	Classification	Action
0.90 – 1.00	Highly Deterministic	Traditional generation, minimal review
0.80 – 0.89	Deterministic	Traditional generation + lightweight review
0.60 – 0.79	Semi-Deterministic	AI generation + enhanced review
0.40 – 0.59	Non-Deterministic	AI generation + full test loop
0.00 – 0.39	Highly Non-Deterministic	AI generation + mandatory human review

Threshold: θ = 0.80 (configurable)

Traditional Code Generators Registry

Generator	Use Case	Input Format	Output Language
openapi-generator	REST APIs from OpenAPI spec	OpenAPI 3.0/3.1	Go, Java, TS, etc.
sqlc	Type-safe SQL from queries	.sql files	Go
oapi-codegen	Go from OpenAPI	OpenAPI 3.0	Go
protoc	gRPC services	.proto files	Go, Java, etc.
Buf	Protobuf linting/breaking	.proto files	Multiple
Wire	Go dependency injection	Provider configs	Go
Kotlinx.serialization	Kotlin serialization	JsonSchema	Kotlin
Vue Codegen	Vue components from JSON	JSON Schema	Vue + TS
JHipster	Full Spring Boot + Angular app	JDL	Java + TS
Terraformer	IaC from cloud state	Cloud state	Terraform
CUE	Data validation & generation	CUE files	Multiple

Input Format

dispatch_request:
  project: "Aether Platform"
  
  tasks:
    - id: "T-001"
      description: "Generate POST /users endpoint with validation"
      interface_contract: "IC-User-001"
      language: "go"
      framework: "gin"
    - id: "T-002"
      description: "Implement password hashing with bcrypt"
      interface_contract: "IC-Auth-002"
      language: "go"
  
  # Determinism configuration
  determinism_threshold: 0.80
  
  # Generator registry (or use default)
  generators:
    - name: "openapi-generator"
      enabled: true
      version: "7.0.0"
    - name: "sqlc"
      enabled: true
      version: "1.25.0"
  
  # GATE-001 configuration
  gate:
    enabled: true
    checks:
      - "interface_compliance"
      - "type_safety"
      - "behavior_match"
      - "dependency_bounds"
      - "performance_budget"
  
  # Confidence thresholds
  confidence:
    auto_approve: 0.90
    suggest_confirmation: 0.70
    mandatory_review: 0.00
  
  # Economic tracking
  track_costs: true

Output Format

dispatch_result:
  project: "Aether Platform"
  timestamp: "2026-06-02T10:00:00Z"
  
  task_results:
    - task_id: "T-001"
      description: "POST /users endpoint"
      
      # === Step 1: Determinism Evaluation ===
      determinism_evaluation:
        contract_completeness: 0.95
        pattern_match: 0.90
        historical_success: 0.92
        complexity: 0.30  # Low complexity
        context_stability: 0.95
        score: 0.85
        classification: "Deterministic"
        threshold_met: true
      
      # === Step 2a: Traditional Generator Dispatch ===
      dispatch:
        generator: "oapi-codegen"
        generator_version: "1.12.0"
        estimated_tokens: 0
        estimated_time_seconds: 5
        output_path: "internal/api/user_handler.go"
      
      # === Step 3: Contract Consistency Gate ===
      gate:
        result: "PASS"
        checks:
          interface_compliance: PASS
          type_safety: PASS
          behavior_match: PASS
          dependency_bounds: PASS
          performance_budget: PASS
      
      # === Step 4: Confidence Grading (for traditional) ===
      confidence:
        score: 0.99  # Deterministic generators are near-100% reliable
        grade: "high"
        action: "auto_approve"
        human_review_required: false
      
      generated_files:
        - path: "internal/api/user_handler.go"
          lines: 87
          confidence: 0.99
    
    - task_id: "T-002"
      description: "Password hashing with bcrypt"
      
      determinism_evaluation:
        contract_completeness: 0.80
        pattern_match: 0.70
        historical_success: 0.85
        complexity: 0.40
        context_stability: 0.90
        score: 0.73
        classification: "Semi-Deterministic"
        threshold_met: false
      
      # === Step 2b: AI Generation ===
      dispatch:
        channel: "ai_generation"
        model: "claude-opus-4"
        estimated_tokens: 2500
        estimated_cost_usd: 0.075
        output_path: "internal/auth/password.go"
      
      gate:
        result: "PASS"
        checks:
          interface_compliance: PASS
          type_safety: PASS
          behavior_match: PASS
          dependency_bounds: PASS
          performance_budget: PASS
      
      confidence:
        score: 0.87
        grade: "medium"
        action: "suggest_confirmation"
        human_review_required: true
        review_priority: "medium"
        suggested_reviewer: "auth-team-lead"
      
      generated_files:
        - path: "internal/auth/password.go"
          lines: 45
          confidence: 0.87
  
  # === Economic Summary ===
  economics:
    total_tasks: 2
    traditional_generated: 1
    ai_generated: 1
    tokens_saved: 0  # By using traditional for T-001
    ai_tokens_used: 2500
    estimated_cost_usd: 0.075
    generation_coverage_rate: 0.50  # 50% deterministic
    efficiency_ratio: 0.50  # 50% zero-token
  
  # === Gate Summary ===
  gate_summary:
    total_checks: 10
    passed: 10
    failed: 0
    warnings: 0

Implementation

Step 1: Determinism Evaluation

def evaluate_determinism(task: dict) -> dict:
    """Score the determinism of a generation task."""
    scores = {
        "contract_completeness": score_contract_completeness(task),
        "pattern_match": score_pattern_match(task),
        "historical_success": score_historical_success(task),
        "complexity": score_complexity(task),  # Lower complexity = higher determinism
        "context_stability": score_context_stability(task),
    }
    weights = {
        "contract_completeness": 0.30,
        "pattern_match": 0.25,
        "historical_success": 0.20,
        "complexity": 0.15,
        "context_stability": 0.10,
    }
    total = sum(scores[k] * weights[k] for k in scores)
    return {
        "scores": scores,
        "score": total,
        "classification": classify(total),
        "threshold_met": total >= 0.80,
    }

Determinism Indicators (boost score):

Input fully specified by OpenAPI/JSON Schema
Output structure mechanical-mapped from input
No business rule variants
Template-able pattern (CRUD, DTO, REST routes, DB migrations)

Non-Determinism Indicators (lower score):

Complex business rules with variants
State machine with edge cases
Ambiguous error handling strategy
Novel algorithm or pattern

Step 2a: Traditional Generator Dispatch

def dispatch_traditional(task: dict, generators: list) -> dict:
    """Match task to appropriate traditional generator."""
    matches = match_generators(task, generators)
    if not matches:
        return {"generator": None, "fallback": "ai_generation"}
    best = matches[0]
    return {
        "generator": best.name,
        "version": best.version,
        "estimated_tokens": 0,
        "estimated_time_seconds": estimate_time(best, task),
        "command": best.generate_command(task),
    }

Step 2b: AI Generation Routing

def route_ai(task: dict) -> dict:
    """Route non-deterministic task to AI generation."""
    return {
        "channel": "ai_generation",
        "model": select_model(task),  # Based on complexity
        "estimated_tokens": estimate_tokens(task),
        "estimated_cost_usd": estimate_cost(task),
        "context_files": gather_context(task),
    }

Step 3: Contract Consistency Gate (GATE-001)

For all generated code (traditional or AI), verify:

Check	Description	Tool
Interface Compliance	Implements declared interface	go vet, TypeScript compiler
Type Safety	Types match across boundaries	go vet, tsc
Behavior Match	Output matches contract semantics	Integration tests
Dependency Bounds	Within P5 limit (≤5 core deps)	dep check, custom rule
Performance Budget	Within declared NFR budget	Benchmarks, load tests

def enforce_gate(code: str, contract: dict) -> dict:
    """Run GATE-001 checks on generated code."""
    results = {
        "interface_compliance": check_interface(code, contract),
        "type_safety": check_types(code, contract),
        "behavior_match": check_behavior(code, contract),
        "dependency_bounds": check_deps(code, max_deps=5),
        "performance_budget": check_perf(code, contract.perf_budget),
    }
    all_pass = all(r.status == "PASS" for r in results.values())
    return {
        "result": "PASS" if all_pass else "FAIL",
        "checks": results,
    }

Step 4: Confidence-Based Routing

For AI-generated code, grade confidence and route:

Confidence	Action	Human Review
≥ 0.90	Auto-approve	None (audit log only)
0.70 – 0.90	Suggest confirmation	Reviewer confirms or rejects
0.00 – 0.70	Mandatory review	Reviewer must approve before merge

Confidence Sources (weighted):

Model's own confidence score (40%)
Test pass rate on generated code (30%)
Contract gate result (15%)
Historical patterns from similar tasks (10%)
Code complexity metrics (5%)

def grade_confidence(generated_code: str, tests_passed: bool, 
                     gate_result: dict, similar_history: list) -> dict:
    """Grade AI-generated code confidence."""
    model_conf = get_model_confidence(generated_code)
    test_score = 1.0 if tests_passed else 0.5
    gate_score = 1.0 if gate_result["result"] == "PASS" else 0.0
    history_score = avg(similar_history) if similar_history else 0.5
    complexity_score = 1.0 - measure_complexity(generated_code) / 10
    
    total = (model_conf * 0.40 + test_score * 0.30 + 
             gate_score * 0.15 + history_score * 0.10 + 
             complexity_score * 0.05)
    
    return {
        "score": total,
        "grade": "high" if total >= 0.90 else "medium" if total >= 0.70 else "low",
        "action": "auto_approve" if total >= 0.90 
                  else "suggest_confirmation" if total >= 0.70
                  else "mandatory_review",
    }

Economic Tracking

Track cost savings from using traditional generators:

economics:
  total_tasks: N
  traditional_generated: M
  ai_generated: K  # K = N - M
  tokens_saved: sum(ai_tokens_estimate) for traditional tasks
  ai_tokens_used: actual tokens for AI tasks
  estimated_cost_usd: cost of AI tokens
  generation_coverage_rate: M / N  # % of tasks zero-token
  efficiency_ratio: tokens_saved / (tokens_saved + ai_tokens_used)

Validation Rules

✅ All tasks have determinism score before dispatch
✅ All generated code passes GATE-001 (no exceptions)
✅ AI confidence < 0.70 always requires human review (P12)
✅ Traditional generator coverage tracked per iteration
✅ Economic metrics exported to metrics-tracker

Integration with Aether.go Methodology

Input from:
- architecture-planner (interface contracts)
- test-planner (test contracts)
Output to:
- code-generator (actual code generation)
- metrics-tracker (cost/efficiency metrics)
- constitution-validator (P8, P12 enforcement)
Part of: D4.5 Generation Dispatch Domain
Principle alignment:
- P8 Tool-System-Adaptation: Optimal generator selection
- P7 Constraint-Mechanization: GATE-001 enforcement
- P12 Human-AI-Boundary: Confidence-based routing
- P11 Non-Functional-Built-in: Performance gate

Examples

Example 1: Standard CRUD API

tasks: 12
traditional_generated: 9  # CRUD endpoints from OpenAPI
ai_generated: 3  # Complex business logic
tokens_saved: 28000
cost_usd: 0.12
coverage_rate: 0.75
gate_result: 12/12 PASS

Example 2: Mixed System

tasks: 30
traditional_generated: 18  # Schemas, migrations, simple handlers
ai_generated: 12  # Business rules, state machines, complex logic
coverage_rate: 0.60
gates_failed: 0
human_review_required: 4

Migration Notes

This skill consolidates the following previously separate skills:

deterministic-evaluator — Step 1
code-generator-dispatcher — Step 2a
contract-consistency-gate — Step 3
confidence-based-reviewer — Step 4

Invocation parameters:

action: evaluate → Step 1
action: dispatch → Step 2 (a or b)
action: gate → Step 3
action: grade → Step 4
action: full (default) → All steps

name	generation-dispatcher
description	Use when evaluating task determinism, dispatching to traditional code generators vs AI generation, enforcing contract consistency gates, and routing for confidence-based human review. This skill unifies D4.5 Generation Dispatch Domain and replaces deterministic-evaluator, code-generator-dispatcher, contract-consistency-gate, and confidence-based-reviewer.

generation-dispatcher

이 저장소의 다른 Skills

이 저장소의 다른 Skills

Generation Dispatcher

Overview

When to Use

Core Pattern

L4→L4.5 Dispatch Pipeline

Determinism Evaluation Dimensions

Score Classification

Traditional Code Generators Registry

Input Format

Output Format

Implementation

Step 1: Determinism Evaluation

Step 2a: Traditional Generator Dispatch

Step 2b: AI Generation Routing

Step 3: Contract Consistency Gate (GATE-001)

Step 4: Confidence-Based Routing

Economic Tracking

Validation Rules

Integration with Aether.go Methodology

Examples

Example 1: Standard CRUD API

Example 2: Mixed System

Migration Notes

Generation Dispatcher

Overview

When to Use

Core Pattern

L4→L4.5 Dispatch Pipeline

Determinism Evaluation Dimensions

Score Classification

Traditional Code Generators Registry

Input Format

Output Format

Implementation

Step 1: Determinism Evaluation

Step 2a: Traditional Generator Dispatch

Step 2b: AI Generation Routing

Step 3: Contract Consistency Gate (GATE-001)

Step 4: Confidence-Based Routing

Economic Tracking

Validation Rules

Integration with Aether.go Methodology

Examples

Example 1: Standard CRUD API

Example 2: Mixed System

Migration Notes