SAGECODE

Multi-agent code generation system for competitive programming.
Combines a Generator (DeepSeek-chat) + Reviewer (DeepSeek-R1) + ChromaDB memory.

Results

Baseline	Model	pass@1	Cost/problem
B0	deepseek-chat, single agent, no memory	6%	—
B1	deepseek-chat Generator + deepseek-R1 Reviewer + ChromaDB	19%	$0.0098

Dataset: LiveCodeBench 100 problems (Codeforces-style, stdio)

Pipeline

Problem
  ↓
[1] Memory Retrieval    ChromaDB lookup (no API call)
  ↓
[2] Generator           deepseek-chat, with memory context injected
  ↓
[3] Reviewer            deepseek-R1, checks boundary/complexity/format
  ↓
[4] Executor            local subprocess, stdin/stdout comparison
  ↓
[5] Memory Update       write episode → every 10 problems consolidate → every 50 prune

Quick Start

# 1. Clone and set up env
git clone https://github.com/Hugoean/sagecode.git
cd sagecode
pip install openai python-dotenv chromadb datasets

# 2. Configure API key
cp .env.example .env
# Edit .env and fill in DEEPSEEK_API_KEY

# 3. Test API connectivity
python tests/test_api.py

# 4. Run one problem (smoke test)
python main.py demo

# 5. Run a batch
python main.py batch --n 10

# 6. Run full baseline (100 problems, ~2.5h, ~$1)
python main.py baseline --n 100

Commands

python main.py demo                        # 1 problem, verbose
python main.py batch --n 10               # 10 problems, sorted by difficulty
python main.py batch --n 10 --no-sort     # 10 problems, original order
python main.py baseline --n 100           # full baseline, saves JSON to data/

Project Structure

sagecode/
├── agents/generation/
│   ├── generator.py     DeepSeek-chat code generator (with memory injection)
│   └── reviewer.py      DeepSeek-R1 code reviewer (truncation-safe)
├── memory/
│   ├── chroma_store.py  ChromaDB 3-layer store (episodic/semantic/procedural)
│   ├── retriever.py     Memory retrieval → prompt context
│   └── updater.py       store / consolidate (every 10) / prune (every 50)
├── executor/
│   └── stdio_exec.py    Local subprocess executor for stdio problems
├── data/
│   ├── loader.py        LiveCodeBench / BigCodeBench / SWE-Bench loader
│   └── *.json           Baseline result files
├── pipeline.py          Main orchestration + token cost tracking
├── main.py              CLI entry point
└── api_client.py        Unified DeepSeek/OpenAI client factory

Environment Variables

# .env (copy from .env.example)
DEEPSEEK_API_KEY=sk-your-key-here
DEEPSEEK_BASE_URL=https://api.deepseek.com
GENERATOR_MODEL=deepseek-chat      # cheap, fast
REVIEWER_MODEL=deepseek-reasoner   # R1, slow but thorough