screen-run

A Claude Code skill that runs a Python script inside an existing GNU Screen session with output logging.

Useful for launching long-running jobs (e.g. ML training) in persistent terminal sessions, capturing full logs, and getting an immediate sanity check on startup errors.

How it works

1. Resolves the Python binary from the given Conda environment
2. Injects a new window into the target screen session running the script with unbuffered output piped to a log file
3. Waits 8 seconds, then tails the log to surface any startup errors (missing packages, CUDA OOM, etc.)

Installation

Copy SKILL.md into your Claude Code skills directory (typically ~/.claude/skills/).

Usage

/screen-run [session-name] [conda-env] [log-file] [script.py and args...]

Example

/screen-run training myenv /tmp/train.log train.py --epochs 100 --lr 0.001

This will:

• Attach to the screen session named training
• Use the Python binary from the myenv Conda environment
• Log all output (stdout + stderr) to /tmp/train.log
• Report the first few lines of output after startup

Requirements

• GNU Screen
• Conda

License

Apache 2.0