fix: reset stale remote agent updates after sac-api restart (0.4.3)

Clear zombie running jobs on deploy and add admin API to cancel a stuck job.

Co-authored-by: Cursor <cursoragent@cursor.com>

backend/app/api/v1/hosts.py

@@ -22,6 +22,8 @@ from app.services.agent_update import (
 )
 from app.services.host_remote_actions import (
     RemoteActionAlreadyRunningError,
+    cancel_host_remote_action,
+    clear_stale_running_remote_actions,
     get_remote_action_status,
     run_agent_fallback_action,
     run_ssh_monitor_update_action,
@@ -605,6 +607,21 @@ def get_host_remote_job(
     return HostRemoteActionJobResponse(**get_remote_action_status(host))
 
 
+@router.post("/{host_id}/actions/remote-job/cancel", response_model=HostRemoteActionJobResponse)
+def cancel_host_remote_job(
+    host_id: int,
+    db: Session = Depends(get_db),
+    _user=Depends(require_admin),
+) -> HostRemoteActionJobResponse:
+    host = db.get(Host, host_id)
+    if host is None:
+        raise HTTPException(status_code=404, detail="Host not found")
+    if not cancel_host_remote_action(db, host):
+        raise HTTPException(status_code=409, detail="No running remote action for this host")
+    db.refresh(host)
+    return HostRemoteActionJobResponse(**get_remote_action_status(host))
+
+
 @router.post("/{host_id}/actions/sessions/list", response_model=HostSessionsResponse)
 def list_host_sessions(
     host_id: int,

backend/app/services/host_remote_actions.py

@@ -219,6 +219,59 @@ def run_agent_fallback_action(db: Session, host: Host) -> AgentUpdateFallbackRes
     return execute_agent_update_fallback(db, host)
 
 
+def clear_stale_running_remote_actions(
+    db: Session,
+    *,
+    reason: str = "Прервано перезапуском SAC (фоновый worker не завершил job)",
+    host_ids: list[int] | None = None,
+) -> list[str]:
+    """Сбросить зависшие running после restart sac-api (daemon thread умер, запись в БД осталась)."""
+    from sqlalchemy import select
+
+    stmt = select(Host).where(Host.agent_update_state == "running")
+    if host_ids is not None:
+        stmt = stmt.where(Host.id.in_(host_ids))
+    rows = list(db.scalars(stmt).all())
+    if not rows:
+        return []
+
+    finished = _utcnow().isoformat()
+    hostnames: list[str] = []
+    for host in rows:
+        started = (host.remote_action or {}).get("started_at")
+        payload = dict(host.remote_action or {})
+        payload.update(
+            {
+                "title": payload.get("title") or "Remote action",
+                "status": "failed",
+                "message": reason,
+                "ok": False,
+                "finished_at": finished,
+                "started_at": started,
+                "target": payload.get("target") or host.hostname,
+            }
+        )
+        host.remote_action = payload
+        host.agent_update_state = "failed"
+        host.agent_update_last_error = reason[:2000]
+        hostnames.append(host.hostname or str(host.id))
+
+    db.commit()
+    with _lock:
+        for host in rows:
+            _running.discard(int(host.id))
+    return hostnames
+
+
+def cancel_host_remote_action(db: Session, host: Host, *, reason: str | None = None) -> bool:
+    """Отменить зависший running job вручную (admin)."""
+    if host.agent_update_state != "running":
+        return False
+    msg = reason or "Отменено администратором SAC"
+    cleared = clear_stale_running_remote_actions(db, reason=msg, host_ids=[int(host.id)])
+    return bool(cleared)
+
+
 def get_remote_action_status(host: Host) -> dict[str, Any]:
     payload = dict(host.remote_action or {})
     if not payload:

backend/app/version.py

@@ -1,5 +1,5 @@
 """Единый источник версии SAC (API, health, логи, OpenAPI)."""
 
 APP_NAME = "Security Alert Center"
-APP_VERSION = "0.4.2"
+APP_VERSION = "0.4.3"
 APP_VERSION_LABEL = f"{APP_NAME} v.{APP_VERSION}"

backend/tests/test_agent_update.py

@@ -224,3 +224,45 @@ def test_api_agent_update_fallback_ssh(jwt_headers, client, db_session, monkeypa
     job = wait_remote_job(client, host.id, jwt_headers)
     assert job["ok"] is True
     assert job["product_version"] == "2.1.0-SAC"
+
+
+def test_clear_stale_running_remote_actions(db_session):
+    from app.services.host_remote_actions import clear_stale_running_remote_actions
+
+    host = Host(
+        hostname="stale-win",
+        os_family="windows",
+        product="rdp-login-monitor",
+        agent_update_state="running",
+        remote_action={"status": "running", "message": "Подключение…"},
+    )
+    db_session.add(host)
+    db_session.commit()
+
+    cleared = clear_stale_running_remote_actions(db_session, reason="test reset")
+    assert cleared == ["stale-win"]
+    db_session.refresh(host)
+    assert host.agent_update_state == "failed"
+    assert host.remote_action["status"] == "failed"
+    assert host.remote_action["ok"] is False
+
+
+def test_cancel_host_remote_job_api(jwt_headers, client, db_session):
+    host = Host(
+        hostname="cancel-me",
+        os_family="windows",
+        product="rdp-login-monitor",
+        agent_update_state="running",
+        remote_action={"status": "running", "title": "Обновление через WinRM"},
+    )
+    db_session.add(host)
+    db_session.commit()
+
+    response = client.post(
+        f"/api/v1/hosts/{host.id}/actions/remote-job/cancel",
+        headers=jwt_headers,
+    )
+    assert response.status_code == 200
+    body = response.json()
+    assert body["active"] is False
+    assert body["status"] == "failed"

deploy/sac-deploy.sh

@@ -105,6 +105,25 @@ print('db: OK')
 " || die "PostgreSQL: проверьте DATABASE_URL в ${CONFIG_FILE} и systemctl status postgresql"
 
 log "systemctl restart ${SERVICE_NAME} (краткий 502 в UI возможен ~10 с)"
+log "Сброс зависших remote_action (running без worker после restart)"
+sudo -u "${APP_USER}" bash -c "
+  set -a
+  # shellcheck source=/dev/null
+  source '${CONFIG_FILE}'
+  set +a
+  cd '${APP_ROOT}/backend'
+  '${VENV}/bin/python' -c \"
+from app.database import SessionLocal
+from app.services.host_remote_actions import clear_stale_running_remote_actions
+session = SessionLocal()
+try:
+    names = clear_stale_running_remote_actions(session)
+    for name in names:
+        print(f'stale remote_action reset: {name}')
+finally:
+    session.close()
+\"
+"
 systemctl restart "${SERVICE_NAME}"
 systemctl is-active --quiet "${SERVICE_NAME}" || die "${SERVICE_NAME} не active"
 

frontend/src/version.ts

@@ -1,4 +1,4 @@
 /** Fallback до загрузки /health; при релизе держите в sync с backend/app/version.py */
 export const APP_NAME = "Security Alert Center";
-export const APP_VERSION = "0.4.2";
+export const APP_VERSION = "0.4.3";
 export const APP_VERSION_LABEL = `${APP_NAME} v.${APP_VERSION}`;