repl

name	repl
description	Persistent Python REPL server that avoids reloading heavy packages (torch, transformers, nnterp, vllm). Use when running multiple Python commands with heavy imports, loading ML models, or doing iterative testing where variables should persist between executions.

Python REPL - Persistent Session

Keeps Python interpreter running between commands, avoiding 20+ second package reload overhead.

When to use this skill

Use the REPL when you're about to:

• Import heavy packages: torch, transformers, nnterp, vllm, diffusers
• Run multiple Python commands that share state
• Do iterative testing/debugging with ML models
• Execute multi-step workflows (load model → compile config → query)

Don't use for one-off simple Python commands.

Setup (once per project)

uv add --editable ~/projects/research-libs/claude-repl

Quick check: Is server running?

echo '{"action":"ping"}' | timeout 2 nc -U /run/user/$(id -u)/claude-repl.sock

Start server

uv run claude-repl server &

Create session with imports

cat > /run/user/$(id -u)/session.json << 'EOF'
{
  "action": "create_session",
  "session_id": "main",
  "python_path": ["./src"],
  "preload_code": "import torch\nfrom mylib import Model\nprint('Ready')"
}
EOF
cat /run/user/$(id -u)/session.json | nc -U /run/user/$(id -u)/claude-repl.sock

Execute code

For simple code:

echo '{"action":"execute","session_id":"main","code":"print(model.num_layers)"}' | nc -U /run/user/$(id -u)/claude-repl.sock

For complex code (avoids shell escaping issues):

cat > /run/user/$(id -u)/code.py << 'EOF'
# Your code here
result = model.forward(x)
print(result.shape)
EOF

uv run python -c "
import json
code = open('/run/user/$(id -u)/code.py').read()
print(json.dumps({'action': 'execute', 'session_id': 'main', 'code': code}))
" | nc -U /run/user/$(id -u)/claude-repl.sock

Export to Jupyter notebook

Convert your session history to a notebook for sharing/reproducibility:

echo '{"action":"export_notebook","session_id":"main","path":"./experiment.ipynb"}' | nc -U /run/user/$(id -u)/claude-repl.sock

Key points

• Variables persist between execute calls
• Single-threaded: long operations block other requests
• Server uses packages from the venv it runs in
• Repo: https://github.com/Butanium/claude-repl