fix: prevent stuck background updates

scripts/tallynote-update-runner.sh

@@ -32,10 +32,96 @@ if [[ -f "$REQUEST_FILE" && ! -L "$REQUEST_FILE" ]]; then
   [[ "$request_operation" == download || "$request_operation" == apply ]] || request_operation='apply'
 fi
 
+# Capture the request id before any privileged preflight can fail. The
+# request file is an application-owned one-shot marker; removing it on an
+# early runner failure lets the server-side lease reaper release the DB row.
+job_id=''
+if [[ -f "$REQUEST_FILE" && ! -L "$REQUEST_FILE" ]]; then
+  job_id=$(sed -n 's/.*"jobId"[[:space:]]*:[[:space:]]*"\([0-9a-f-]*\)".*/\1/p' "$REQUEST_FILE" | head -n 1)
+fi
+STATE_CREATED=0
+heartbeat_pid=''
+heartbeat_owner=$$
+
+write_recovery_state() {
+  local phase=$1 temporary
+  temporary="$PREFIX/.update-state-$$-${RANDOM}.tmp"
+  [[ ! -e "$temporary" && ! -L "$temporary" ]] || return 1
+  printf 'job_id=%s\nold_target=%s\nphase=%s\n' "$job_id" "$old_target" "$phase" > "$temporary"
+  chmod 600 "$temporary"
+  mv -Tf -- "$temporary" "$STATE_FILE"
+  STATE_CREATED=1
+}
+
+clear_recovery_state() {
+  [[ ! -L "$STATE_FILE" ]] || return 1
+  rm -f -- "$STATE_FILE"
+  STATE_CREATED=0
+}
+
+stop_heartbeat() {
+  if [[ -n "$heartbeat_pid" ]]; then
+    kill "$heartbeat_pid" 2>/dev/null || true
+    wait "$heartbeat_pid" 2>/dev/null || true
+    heartbeat_pid=''
+  fi
+}
+
+heartbeat() {
+  # Keep the lease fresh during long downloads/backups, but stop on a hard
+  # runner kill so an orphaned child cannot keep the recovery marker alive.
+  while kill -0 "$heartbeat_owner" 2>/dev/null; do
+    sleep 10 || exit 0
+    [[ -f "$STATE_FILE" && ! -L "$STATE_FILE" ]] || exit 0
+    touch "$STATE_FILE" 2>/dev/null || exit 0
+  done
+}
+
+start_heartbeat() {
+  stop_heartbeat
+  heartbeat &
+  heartbeat_pid=$!
+}
+
+# This trap covers failures before the normal apply cleanup trap is installed,
+# including a missing runtime, an invalid current link, and a failed service
+# stop. It deliberately does not remove a pre-existing recovery marker.
+preflight_cleanup() {
+  local result=$?
+  stop_heartbeat
+  if (( result != 0 )); then
+    rm -f -- "$REQUEST_FILE" 2>/dev/null || true
+    if (( STATE_CREATED == 1 )); then clear_recovery_state || true; fi
+  fi
+  return "$result"
+}
+trap preflight_cleanup EXIT
+
 # Downloading is intentionally handled while the main service remains up.
 # The CLI persists the validated payload under the root-owned workspace and
 # leaves the job staged for a later apply request.
 if [[ "$request_operation" == download ]]; then
+  # A previous download runner may have been interrupted after creating its
+  # marker. Clear only that download marker and retry the idempotent request.
+  if [[ -f "$STATE_FILE" && ! -L "$STATE_FILE" ]] && grep -q '^phase=download$' "$STATE_FILE"; then
+    clear_recovery_state || die '无法清理上一次下载状态'
+  fi
+  write_recovery_state download || die '无法写入更新恢复状态'
+  cleanup_download() {
+    local result=$?
+    stop_heartbeat
+    if (( result != 0 )); then
+      # The CLI normally records failed itself. If it died before opening the
+      # database, the expired marker/request will be reconciled by the app.
+      rm -f -- "$REQUEST_FILE" 2>/dev/null || true
+    fi
+    clear_recovery_state || true
+    return "$result"
+  }
+  trap cleanup_download EXIT
+  trap 'exit 143' TERM
+  trap 'exit 130' INT
+  start_heartbeat
   node_bin="$CURRENT_LINK/runtime/bin/node"
   [[ -x "$node_bin" ]] || node_bin=$(command -v node || true)
   [[ -n "$node_bin" ]] || die 'node runtime not found'
@@ -69,34 +155,21 @@ if systemctl is-active --quiet "$SERVICE_NAME"; then was_active=1; fi
 # shellcheck disable=SC2329 # invoked indirectly by the EXIT trap below
 restore_initial_service() {
   local result=$?
+  stop_heartbeat
+  if (( result != 0 )); then
+    rm -f -- "$REQUEST_FILE" 2>/dev/null || true
+    if (( STATE_CREATED == 1 )); then clear_recovery_state || true; fi
+  fi
   if (( was_active )); then systemctl start "$SERVICE_NAME" || true; fi
   return "$result"
 }
 trap restore_initial_service EXIT
-systemctl stop "$SERVICE_NAME"
-
-job_id=''
-if [[ -f "$REQUEST_FILE" && ! -L "$REQUEST_FILE" ]]; then
-  job_id=$(sed -n 's/.*"jobId"[[:space:]]*:[[:space:]]*"\([0-9a-f-]*\)".*/\1/p' "$REQUEST_FILE" | head -n 1)
-fi
 old_node="$CURRENT_LINK/runtime/bin/node"
 [[ -x "$old_node" ]] || old_node=$(command -v node || true)
-switched=0
 handled=0
 
-write_update_state() {
-  local phase=$1 temporary
-  temporary="$PREFIX/.update-state-$$-${RANDOM}.tmp"
-  [[ ! -e "$temporary" && ! -L "$temporary" ]] || return 1
-  printf 'job_id=%s\nold_target=%s\nphase=%s\n' "$job_id" "$old_target" "$phase" > "$temporary"
-  chmod 600 "$temporary"
-  mv -Tf -- "$temporary" "$STATE_FILE"
-}
-
-clear_update_state() {
-  [[ ! -L "$STATE_FILE" ]] || return 1
-  rm -f -- "$STATE_FILE"
-}
+write_update_state() { write_recovery_state "$1"; }
+clear_update_state() { clear_recovery_state; }
 
 finalize_state_job() {
   local node=$1 status=$2 state_job=$3
@@ -195,13 +268,24 @@ fi
 
 [[ -f "$REQUEST_FILE" && ! -L "$REQUEST_FILE" ]] || exit 0
 
+# Only create the marker for this invocation after any marker from a previous
+# interrupted run has been reconciled. Otherwise the freshly-created `running`
+# marker is indistinguishable from stale recovery state and the runner can
+# finalize its own queued job as failed before the update CLI starts.
+if [[ ! -e "$STATE_FILE" ]]; then
+  write_recovery_state running || die '无法写入更新恢复状态'
+fi
+start_heartbeat
+if ! systemctl stop "$SERVICE_NAME"; then
+  die '无法停止 TallyNote 服务'
+fi
+
 rollback_current() {
   local current_target rollback_link
   current_target=$(readlink -f -- "$CURRENT_LINK" 2>/dev/null || true)
   if [[ "$current_target" == "$old_target" ]]; then
     # An earlier failure branch may already have restored the link. Keep the
     # marker truthful so the EXIT trap can still finalize the job.
-    switched=0
     return 0
   fi
   rollback_link="$PREFIX/.current-rollback-$$-${RANDOM}.tmp"
@@ -211,7 +295,6 @@ rollback_current() {
     rm -f -- "$rollback_link" 2>/dev/null || true
     return 1
   fi
-  switched=0
 }
 
 finalize_failed_job() {
@@ -237,6 +320,7 @@ finalize_completed_job() {
 # shellcheck disable=SC2329 # invoked indirectly by the EXIT trap below
 cleanup_after_update() {
   local result=$? rollback_ok=1
+  stop_heartbeat
   if (( result != 0 && handled == 0 )); then
     if ! rollback_current; then rollback_ok=0; fi
     # Once the old release is active again, always try to close the job. The
@@ -258,7 +342,6 @@ cleanup_after_update() {
 }
 trap cleanup_after_update EXIT
 
-write_update_state running || exit 1
 node_bin="$CURRENT_LINK/runtime/bin/node"
 [[ -x "$node_bin" ]] || node_bin=$(command -v node || true)
 [[ -n "$node_bin" ]] || die 'node runtime not found'
@@ -273,9 +356,6 @@ if (( update_result != 0 )); then
   exit "$update_result"
 fi
 
-if [[ "$(readlink -f -- "$CURRENT_LINK" 2>/dev/null || true)" != "$old_target" ]]; then
-  switched=1
-fi
 write_update_state health-check || exit 1
 
 systemctl start "$SERVICE_NAME"