Run commands from the repository root unless a command explicitly enters another directory. Start with the README quick start.
The local validation environment uses Python 3.13.5, Node.js 20.19.2, PyTorch 2.11.0+cpu, and the frontend package lock. Python packages are pinned in backend/requirements.txt; the recorded upstream environment is in backend/environment-record.json.
For CPU operation, install the pinned Torch CPU wheel first, then the requirements. A compatible CUDA wheel may be selected for separate research work. The default dashboard replay remains a local CPU workflow.
DATA_BUNDLE.json pins the release URL, compressed size, archive SHA-256, and upstream source. DATA_MANIFEST.json pins all 18 extracted files, including the checkpoint, prepared corpus, tokens, saved scores, and SQLite investigation database. The original archive's internal dashboard/data/ paths are mapped into backend/dashboard/data/ without modifying file bytes.
The compressed archive is approximately 262 MiB; extracted files total approximately 1.53 GiB. Downloads are cached in .cache/, which is ignored by Git. Uploads live separately in backend/dashboard/data/graph_uploads/ and are never included in the release or source manifest.
Offline installation:
python scripts/download_local_data.py --archive /path/to/cadets-v3-local-data.tar.gz
python scripts/verify_package.pyThe installer rejects unexpected files, symbolic links, path traversal, duplicate entries, and checksum mismatches. Keep the backend stopped while installing or replacing artifacts. Restart it after rebuilding an investigation index.
python scripts/run_backend.py --host 127.0.0.1 --port 8001 --reloadSet TGDETECT_API_ORIGIN=http://127.0.0.1:8001 in frontend/.env.local, then restart Next.js. The frontend proxy uses the origin at build/runtime configuration time; rebuild production output after changing it.
Frontend development:
cd frontend
npm run devThe frontend opens the landing page at http://localhost:3000/. Use Open workspace or visit /dashboard for the investigation application. npm run dev and npm run build build the landing assets automatically. After editing frontend/landing/ during development, run npm run build:landing from frontend/ to refresh them. No separate landing server is required.
Use Ctrl+C in each service terminal to stop it. Avoid running a development server and a production build against the same .next/ directory simultaneously.
| Symptom | Action |
|---|---|
| Port 3000 or 8000 is occupied | Stop the other service or select another port and adjust the frontend API origin. |
| Landing assets are missing | Run npm run build:landing from frontend/, or rebuild with npm run build. Keep generated public/landing/ beside the Next.js output when deploying. |
| API proxy cannot connect | Start the backend, check http://127.0.0.1:8000/api/cadets/summary, and check the proxy origin. |
| CADETS index unavailable | Run the verified data installer. It includes the prepared investigation index. |
| Checkpoint/source hash mismatch | Restore the pinned research files and data bundle; do not disable validation or modify trained source to bypass it. |
| Replay returns 409 | Another replay owns the model runtime. Wait for it to finish and retry. |
| An upload is rejected | Convert it to the supported endpoint-event schema in the investigation guide. |
/legacy returns missing-artifact errors |
That interface needs independent Sysmon/DEDALE artifacts and is outside the supplied CADETS package. |
Research command from backend/:
PYTHONPATH=research python -m sliding_node_mem_v3.run --helpCloud scripts require a configured Modal account and their original volume names. They are retained for reproducibility and are not launched by normal setup. An ORTHRUS trained checkpoint is not bundled; its official source and recorded comparison outputs are included.
Default servers bind to loopback for local use. Authentication, authorization, TLS, rate limiting, and deployment-specific upload retention must be configured before exposing the service to other users. The current research checkpoint's detection limits remain applicable after deployment.