sync.sh

#!/usr/bin/env bash
# The server holds everything (venv, model weights, data, runs); this folder holds the code and a copy
# of the results.
#   ./sync.sh push    code -> server
#   ./sync.sh pull    results <- server (files over 200 MB stay on the server; newer local files are kept)
# push also sends what is decided locally: hand-grading labels (results/review/labels_*.json), the frozen
# dev/test split (results/split.json) and the experiment freeze marker.
# Remote is read from .remote (gitignored), e.g.:  REMOTE=host  PORT=22  DIR=work/exp
#
# `push` sends an explicit whitelist of code paths and deletes only inside the code directories, so
# server-side state can never be removed by a sync.
set -euo pipefail
cd "$(dirname "$0")"
source .remote
SSH="ssh -p ${PORT:-22}"
case "${1:-push}" in
  push) ssh -p "${PORT:-22}" "$REMOTE" "mkdir -p ${DIR}"
        rsync -az -e "$SSH" --delete --exclude __pycache__ core configs tests "${REMOTE}:${DIR}/"
        rsync -az -e "$SSH" ./*.py ./*.sh requirements.txt README.md "${REMOTE}:${DIR}/"
        if [ -d results/review ]; then
          rsync -az -e "$SSH" --include 'labels_*.json' --exclude '*' results/review/ "${REMOTE}:${DIR}/results/review/"
        fi
        # decisions made here: the frozen dev/test split and the experiment freeze marker
        for f in results/split.json results/experiment_frozen.json; do
          if [ -f "$f" ]; then rsync -az -e "$SSH" "$f" "${REMOTE}:${DIR}/results/"; fi
        done ;;
  pull) rsync -azu -e "$SSH" --max-size=200m "${REMOTE}:${DIR}/results/" results/ ;;
  *) echo "usage: $0 push|pull"; exit 1 ;;
esac