AI-Powered Code Refactoring with Reset Weights & Blueprint Orchestration
Kimi Code Explorer is a production-grade VS Code extension that combines cutting-edge AI patterns for intelligent code refactoring:
| Feature | Description |
|---|---|
| 🔁 Reset Weights | Clean state for every analysis, eliminating context pollution |
| 📋 Blueprint Orchestration | Multi-agent coordination for large-scale projects |
| 🧠 3-Tier Memory | Short-term / Mid-term / Long-term memory management |
| ✅ Verification Pipeline | Syntax, security, and style gates before code changes |
| 🔄 Detach/Reload | Context isolation to prevent memory overflow |
┌─────────────────────────────────────────────────────────────────┐
│ ORCHESTRATOR AGENT │
│ (Blueprint Controller + State Manager) │
├─────────────────────────────────────────────────────────────────┤
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────────────────┐ │
│ │ Blueprint │ │ Checklist │ │ 3-Tier Memory │ │
│ │ Store │ │ Manager │ │ (ST/MT/LT + Vectors) │ │
│ └─────────────┘ └─────────────┘ └─────────────────────────┘ │
└─────────────────────────────────────────────────────────────────┘
│
┌───────────────────┼───────────────────┐
▼ ▼ ▼
┌─────────────────┐ ┌─────────────────┐ ┌─────────────────┐
│ WORKER 1 │ │ WORKER 2 │ │ WORKER N │
│ │ │ │ │ │
│ ┌─────────────┐ │ │ ┌─────────────┐ │ │ ┌─────────────┐ │
│ │ 1. RESET │ │ │ │ 1. RESET │ │ │ │ 1. RESET │ │
│ │ 2. LOAD │ │ │ │ 2. LOAD │ │ │ │ 2. LOAD │ │
│ │ 3. EXECUTE │ │ │ │ 3. EXECUTE │ │ │ │ 3. EXECUTE │ │
│ │ 4. VERIFY │ │ │ │ 4. VERIFY │ │ │ │ 4. VERIFY │ │
│ │ 5. DETACH │ │ │ │ 5. DETACH │ │ │ │ 5. DETACH │ │
│ └─────────────┘ │ │ └─────────────┘ │ │ └─────────────┘ │
└─────────────────┘ └─────────────────┘ └─────────────────┘
│ │ │
└───────────────────┼───────────────────┘
▼
┌─────────────────────────────────────────────────────────────────┐
│ SHARED ARTIFACTS │
│ (Code, Tests, Docs, Verification Results) │
└─────────────────────────────────────────────────────────────────┘
# Clone repository
git clone https://github.com/yourname/kimi-code-explorer.git
cd kimi-code-explorer
# Install dependencies
npm install
# Configure API key
cp .env.example .env
# Edit .env and add KIMI_API_KEY
# Build
npm run build
# Launch in VS Code
press F5- Select code in VS Code
- Press
Ctrl+Shift+K - View alternatives in Kimi Explorer panel
- Click "Apply" to replace code
- Create
blueprint.yamlin project root - Press
Ctrl+Shift+B - Workers process modules in parallel
- Each worker detaches after completion
{
"kimiCodeExplorer.apiKey": "your-api-key",
"kimiCodeExplorer.model": "moonshot-v1-32k",
"kimiCodeExplorer.resetWeights": true,
"kimiCodeExplorer.blueprint.enabled": true,
"kimiCodeExplorer.memory.fusionWeight": 0.15,
"kimiCodeExplorer.verification.syntaxCheck": true,
"kimiCodeExplorer.verification.securityScan": true
}kimi-code-explorer/
├── src/
│ ├── core/
│ │ ├── resetWeights.ts # Reset Weights implementation
│ │ ├── worker.ts # Worker Agent lifecycle
│ │ ├── orchestrator.ts # Blueprint orchestration
│ │ ├── verification.ts # Quality verification
│ │ ├── memoryManager.ts # 3-Tier memory system
│ │ └── checkpoint.ts # State persistence
│ ├── services/
│ │ ├── llmService.ts # Kimi API integration
│ │ └── blueprintService.ts # Blueprint parser
│ ├── providers/
│ │ └── kimiExplorer.ts # VS Code sidebar panel
│ └── extension.ts # Entry point
├── docs/
│ ├── ARCHITECTURE.md
│ ├── BLUEPRINT.md
│ └── MEMORY_SYSTEM.md
├── examples/
│ └── blueprint.example.yaml
├── AGENTS.md # Dynamic agent config
├── package.json
├── tsconfig.json
└── README.md
Ensures each analysis starts from a clean state:
// Before analysis
const cleanContext = await resetManager.reset({
reason: 'code_exploration'
});
// Analysis with no historical baggage
const result = await analyze(code, cleanContext);
// After analysis, context is detached
await detachContext();Benefits:
- ✅ No context pollution from previous analyses
- ✅ Consistent results regardless of history
- ✅ Predictable behavior
- ✅ Easier debugging
┌─────────────────────────────────────────────────────────────────┐
│ 3-TIER MEMORY ARCHITECTURE │
├─────────────────────────────────────────────────────────────────┤
│ │
│ ┌─────────────┐ Short-term (Session Cache) │
│ │ Ephemeral │ • Duration: Current session only │
│ │ Buffer │ • Fusion Weight: ≤0.15 │
│ │ (EB) │ • Auto-clear on reset │
│ └──────┬──────┘ │
│ │ │
│ ┌──────▼──────┐ Mid-term (Validated Knowledge) │
│ │ Working │ • Duration: 48 hours │
│ │ Memory │ • Fusion Weight: 0.15-0.35 │
│ │ (WM) │ • Requires external validation │
│ └──────┬──────┘ │
│ │ │
│ ┌──────▼──────┐ Long-term (Proven Knowledge Base) │
│ │ Long-term │ • Duration: Permanent │
│ │ Memory │ • Fusion Weight: 0.35-0.50 │
│ │ (LTM) │ • Requires multiple validations │
│ └─────────────┘ │
│ │
└─────────────────────────────────────────────────────────────────┘
Define project structure and execution order:
# blueprint.yaml
project:
name: "E-Commerce API"
version: "1.0.0"
modules:
- id: "auth"
name: "Authentication Service"
dependencies: []
output: "src/auth/"
checklist:
- "JWT token generation"
- "Password hashing (bcrypt)"
- "Unit tests > 80%"
- id: "api"
name: "API Gateway"
dependencies: ["auth"]
output: "src/api/"
checklist:
- "Rate limiting middleware"
- "Input validation"
execution:
strategy: "parallel"
maxWorkers: 4Multi-layer quality gates:
| Layer | Check | Description |
|---|---|---|
| 1 | Syntax | Validate code can be parsed |
| 2 | Security | Detect dangerous patterns (eval, hardcoded secrets) |
| 3 | Style | Check line length, whitespace, formatting |
| 4 | Complexity | Analyze loop depth, branch count |
| 5 | Checklist | Verify against module requirements |
Worker lifecycle for context isolation:
RESET → LOAD CONTEXT → EXECUTE → VERIFY → SAVE → DETACH
Critical: DETACH Phase
After every task completion:
- Clear LLM instance
- Clear context manager
- Release temporary variables
- Force garbage collection
Purpose: Prevent context pollution between tasks
| Feature | v1 | v2 (Final) |
|---|---|---|
| Reset Weights | ❌ | ✅ |
| Blueprint Mode | ❌ | ✅ |
| 3-Tier Memory | ❌ | ✅ |
| Context Isolation | ❌ | ✅ |
| Parallel Workers | ❌ | ✅ |
| Checkpoint Recovery | ❌ | ✅ |
| Verification Pipeline | Basic | Full |
MIT License - see LICENSE
Kimi Code Explorer - Production-Grade AI Code Refactoring 🚀