Adaptive Code AI Agents

This repository combines two related code-agent projects:

• simplecoder-cli/: a ReAct-style CLI coding agent with tool use, optional retrieval, planning, context compaction, and permission-aware file access
• code-rl-posttraining/: a sandboxed reinforcement learning pipeline for code generation with both remote Tinker and local LoRA training backends

The agent side focuses on inference-time interaction and intervention discipline. The training side focuses on improving code generation through execution-based rewards.

Repository layout

adaptive-code-ai-agents/
├── simplecoder-cli/
└── code-rl-posttraining/

Configuration

Do not commit API keys to this repository. Use environment variables instead.

SimpleCoder

Required:

• GEMINI_API_KEY

Optional:

• SIMPLECODER_MODEL (defaults to gemini/gemini-2.5-flash)

PowerShell example:

$env:GEMINI_API_KEY = "your-key-here"
$env:SIMPLECODER_MODEL = "gemini/gemini-2.5-flash"

Code RL post-training

Required for backend=tinker:

• TINKER_API_KEY

Optional:

• SANDBOX_URL (defaults to http://localhost:8080/run_code)
• SANDBOX_MAX_CONCURRENCY
• SANDBOX_CLIENT_TIMEOUT_SECONDS

PowerShell example:

$env:TINKER_API_KEY = "tml-..."
$env:SANDBOX_URL = "http://localhost:8080/run_code"

Quick start

SimpleCoder

cd simplecoder-cli
pip install -r requirements.txt
pip install -e .
simplecoder "create a hello.py file"

Code RL post-training

cd code-rl-posttraining
python -m pip install -r requirements.txt
docker run -it -p 8080:8080 -v ./sandbox_config/local.yaml:/root/sandbox/sandbox/configs/local.yaml volcengine/sandbox-fusion:server-20250609
python train.py

See the README inside each subproject for more detail.