entrypoint.sh

#!/bin/sh
# ── entrypoint.sh ─────────────────────────────────────────────────────────────
# HF Spaces runs as UID 1000 (not root), so the original docker-entrypoint.sh
# fails — it uses `gosu` to drop root→node, which requires being root.
# We skip it entirely and exec the server directly from /app.
#
# Recovery logic:
#   1. Fix permissions and recreate missing empty PostgreSQL dirs
#   2. Remove stale postmaster.pid from unclean shutdowns
#   3. Probe whether PostgreSQL can start cleanly
#   4. If the cluster is unrecoverable, restore from the latest backup
#   5. Start Paperclip
# ──────────────────────────────────────────────────────────────────────────────

set -e

DB_DIR="/data/paperclip/instances/default/db"
BACKUP_DIR="/data/paperclip/instances/default/data/backups"
DB_PORT=54329

# ── 1. Ensure /data/paperclip is writable ────────────────────────────────────
mkdir -p "${DB_DIR}"

# ── 1a. Fix PostgreSQL data dir permissions ───────────────────────────────────
chmod 0700 "${DB_DIR}"

if [ -d "${DB_DIR}/pg_notify" ]; then
  chmod 0700 "${DB_DIR}/pg_notify"
fi

# ── 1b. Remove stale postmaster.pid ──────────────────────────────────────────
# If the Space was killed mid-run, PostgreSQL leaves behind postmaster.pid.
# Paperclip sees it, thinks Postgres is running, tries to connect, and fails.
# Safe to delete — PostgreSQL recreates it fresh on startup.
if [ -f "${DB_DIR}/postmaster.pid" ]; then
  echo "[entrypoint] Removing stale postmaster.pid (leftover from unclean shutdown)..."
  rm -f "${DB_DIR}/postmaster.pid"
  echo "[entrypoint] postmaster.pid removed."
fi

# ── 1c. Recreate missing PostgreSQL dirs ─────────────────────────────────────
# These are empty directories — safe to recreate if missing after unclean shutdown.
if [ -f "${DB_DIR}/PG_VERSION" ]; then
  for dir in pg_notify pg_tblspc pg_replslot pg_twophase pg_snapshots pg_commit_ts pg_logical pg_logical/snapshots pg_logical/mappings; do
    if [ ! -d "${DB_DIR}/${dir}" ]; then
      echo "[entrypoint] ${dir} missing — recreating (cluster data preserved)..."
      mkdir -p "${DB_DIR}/${dir}"
      chmod 0700 "${DB_DIR}/${dir}"
    fi
  done
  echo "[entrypoint] PostgreSQL directories verified."
fi

# ── 1d. Probe whether PostgreSQL can start — restore from backup if not ───────
restore_from_backup() {
  echo "[entrypoint] ⚠️  Live PostgreSQL cluster is unrecoverable."
  echo "[entrypoint] Looking for latest backup in ${BACKUP_DIR}..."

  LATEST_BACKUP=$(ls -t "${BACKUP_DIR}"/*.tar.gz 2>/dev/null | head -n 1 || true)

  if [ -z "${LATEST_BACKUP}" ]; then
    echo "[entrypoint] ❌ No backup found in ${BACKUP_DIR}. Starting fresh — all previous data is lost."
    rm -rf "${DB_DIR}"
    mkdir -p "${DB_DIR}"
    chmod 0700 "${DB_DIR}"
    return
  fi

  echo "[entrypoint] ✅ Found backup: ${LATEST_BACKUP}"
  echo "[entrypoint] Wiping corrupted cluster and restoring..."
  rm -rf "${DB_DIR}"
  mkdir -p "${DB_DIR}"
  chmod 0700 "${DB_DIR}"
  tar -xzf "${LATEST_BACKUP}" -C "${DB_DIR}" --strip-components=1
  chmod 0700 "${DB_DIR}"
  echo "[entrypoint] ✅ Restore complete from: $(basename ${LATEST_BACKUP})"
  echo "[entrypoint] Note: work done after this backup was taken will be missing."
}

if [ -f "${DB_DIR}/PG_VERSION" ]; then
  echo "[entrypoint] Probing PostgreSQL cluster health..."
  PG_CTL=$(find /usr -name "pg_ctl" 2>/dev/null | head -n 1 || true)

  if [ -n "${PG_CTL}" ]; then
    if ! "${PG_CTL}" -D "${DB_DIR}" -o "-p ${DB_PORT}" -w -t 15 start > /tmp/pg_probe.log 2>&1; then
      echo "[entrypoint] PostgreSQL probe failed. Logs:"
      cat /tmp/pg_probe.log
      restore_from_backup
    else
      echo "[entrypoint] PostgreSQL cluster is healthy. Stopping probe instance..."
      "${PG_CTL}" -D "${DB_DIR}" -o "-p ${DB_PORT}" stop > /dev/null 2>&1 || true
    fi
  else
    echo "[entrypoint] pg_ctl not found — skipping probe, Paperclip will handle startup."
  fi
fi

# ── 2. Write opencode provider config ────────────────────────────────────────
OPENCODE_CONFIG_DIR="${HOME}/.config/opencode"
OPENCODE_STATE_DIR="${HOME}/.local/share/opencode"

mkdir -p "${OPENCODE_CONFIG_DIR}"
cp /app/opencode.json "${OPENCODE_CONFIG_DIR}/config.json"

# ── 3. Write opencode auth file ───────────────────────────────────────────────
mkdir -p "${OPENCODE_STATE_DIR}"
cat > "${OPENCODE_STATE_DIR}/auth.json" <<AUTHEOF
{
  "zai-coding-plan": {
    "type": "api",
    "key": "${ZAI_API_KEY}"
  }
}
AUTHEOF

echo "[entrypoint] OpenCode configured → Z.AI Coding Plan (https://api.z.ai/api/coding/paas/v4)"
echo "[entrypoint] Models: glm-5.2 (primary) | glm-4.7 (small/fast)"

# ── 4. Brief pause to let HF storage fully settle before Paperclip starts ────
# Without this, embedded PostgreSQL can get an ECONNREFUSED on the very first
# boot cycle because the storage mount isn't quite ready when Paperclip tries
# its first DB connection (~03:00:54 vs Postgres ready at ~03:01:20).
echo "[entrypoint] Waiting for storage to settle..."
sleep 8

echo "[entrypoint] Starting Paperclip..."

# ── 5. Start the server ───────────────────────────────────────────────────────
cd /app
exec node --import ./server/node_modules/tsx/dist/loader.mjs server/dist/index.js