#!/usr/bin/env bash
# sap-mcp-bridge entrypoint
#
# Lifecycle:
#   1. (as root) Ensure /workspace exists and is owned by the mcp user,
#      then re-exec ourselves as mcp via gosu. This makes the mount
#      writable even when Docker auto-creates the host path as root.
#   2. (as mcp) Install the signal trap BEFORE forking any children, so
#      a SIGTERM arriving during start-up still tears the whole group
#      down cleanly.
#   3. (as mcp) Launch one supergateway per SAP MCP server on
#      127.0.0.1:9001..9005. They speak Streamable HTTP at /mcp and
#      expose /healthz.
#   4. (as mcp) Wait for every gateway's /healthz to answer (with a
#      cap), then launch nginx in the foreground. This avoids the
#      cold-start window where nginx 502s because upstreams haven't
#      bound yet.
#   5. (as mcp) `wait -n` — exit as soon as ANY direct child (gateway
#      or nginx) dies. Docker restarts the container, so a crashed
#      gateway resets the whole group (acceptable for homelab use).
set -euo pipefail

WORKDIR_HOST="${WORKDIR_HOST:-/workspace}"
APP_UID=10001
APP_GID=65534   # nogroup
APP_USER="mcp"
GATEWAY_PORTS=(9001 9002 9003 9004 9005)

# --- Stage 1: privilege handling -----------------------------------
if [ "$(id -u)" = "0" ]; then
  mkdir -p "$WORKDIR_HOST"
  # Skip the chown if the dir is already owned correctly — recursive
  # chown on a large mounted workspace is slow and runs every restart.
  # The chown can fail on read-only / fuse / overlay mounts; we tolerate
  # failure rather than crash (the mcp user just won't be able to write).
  current_owner="$(stat -c '%u' "$WORKDIR_HOST" 2>/dev/null || echo "")"
  if [ "$current_owner" != "$APP_UID" ]; then
    chown -R "${APP_UID}:${APP_GID}" "$WORKDIR_HOST" 2>/dev/null || true
  fi
  exec gosu "${APP_USER}" "$0" "$@"
fi

# --- Stage 2: install trap BEFORE children -------------------------
# kill 0 sends SIGTERM to every process in our group; tini will reap
# them. Done now so a SIGTERM that lands during start-up still tears
# the whole group down cleanly.
trap 'echo "[sap-mcp-bridge] received signal, shutting down"; kill 0' SIGTERM SIGINT

# --- Stage 3: launch gateways --------------------------------------
cd "$WORKDIR_HOST" 2>/dev/null || cd /home/mcp

start() {
  local name="$1" port="$2" cmd="$3"
  echo "[sap-mcp-bridge] starting ${name} on 127.0.0.1:${port}  ->  ${cmd}"
  # --stateful gives each MCP client its own subprocess (safer for the
  # SAP servers, which keep auth state per-session).
  supergateway \
    --stdio "${cmd}" \
    --outputTransport streamableHttp \
    --port "${port}" \
    --streamableHttpPath /mcp \
    --healthEndpoint /healthz \
    --cors \
    --stateful \
    --logLevel info \
    &
}

start "sap-cap"          9001 "iflow-mcp_cap-js-mcp-server"
start "sap-abap-adt"     9002 "mcp-abap-adt"
start "sap-odata"        9003 "btp-sap-odata-to-mcp-server"
start "ui5-mcp-server"   9004 "ui5mcp"
start "fiori-mcp-server" 9005 "fiori-mcp"

# --- Stage 4: wait for upstreams, then nginx -----------------------
# Poll each gateway's /healthz for up to ~30s. Without this nginx can
# come up before the Node processes finish booting and serve 502s on
# the first requests (and to the Traefik healthcheck).
for port in "${GATEWAY_PORTS[@]}"; do
  for _ in $(seq 30); do
    if curl -fsS -m 1 "http://127.0.0.1:${port}/healthz" >/dev/null 2>&1; then
      echo "[sap-mcp-bridge] upstream :${port} ready"
      break
    fi
    sleep 1
  done
done

echo "[sap-mcp-bridge] starting nginx reverse proxy on :8080"
nginx -g 'daemon off;' &

# --- Stage 5: shut down on first child death -----------------------
wait -n
echo "[sap-mcp-bridge] a child exited, taking the container down"
kill 0 2>/dev/null || true
exit 1