Desktop development

Desktop shares the Python backend, templates and styles with Web. Tauri supplies the native window, menus and file dialogs. See the installation guide for source setup.

Local process boundary

The backend binds a random loopback port, requires a per-launch cookie, rejects foreign origins, and exits when its parent input closes. The native window cannot navigate outside that origin; external HTTP(S) links open in the system browser. These controls do not replace host access controls or a security review.

Only one desktop instance may own a runtime. Background analyses use subprocesses without Redis, with the same runners and result format as Web. Closing the app interrupts work; interrupted jobs are marked failed rather than rerun. Timeouts stop the analysis process tree.

Supported numerical libraries use one-thread limits; pathway enrichment uses a serial backend. These settings are not an OS-level CPU quota for arbitrary external programs. The custom-script sandbox has a separate one-CPU quota.

Check navigation during a background run

For a controlled one-minute test, select Synthetic cross-sectional demo, then Custom Python / R script (synthetic prototype) in Analysis. This requires Docker and the configured sandbox image; the Inputs checks show their availability. Choose Python and paste this complete script. Input preparation happens when you click Run; generating or copying a template is not required:

import time
from pathlib import Path

print("Background-run diagnostic started", flush=True)
time.sleep(60)
output = Path('/work/output')
output.mkdir(parents=True, exist_ok=True)
(output / "diagnostic.txt").write_text(
    "One-minute background-run diagnostic completed. No scientific analysis performed.\n"
)

Run it, visit Dashboard or Query, then return to Analysis. Check the running status in Jobs and the completed result/files afterwards. Keep the desktop open. The script deliberately waits rather than doing scientific work; it tests UI responsiveness and completion delivery, not computational performance. Keep the delay below the sandbox's 90-second execution limit.