Device Logs

The SDK reports problems — telemetry batches that can't reach the ingest endpoint, registry endpoints failing over, model resolution errors — through each platform's native log system. Nothing crashes when the network is blocked or an API key is wrong; the SDK degrades gracefully and says why in the logs. This page is where to look.

Logging starts automatically the moment your app initializes the SDK (Xybrid.init / Xybrid.initialize — any entry point). There is nothing to enable.

Where logs appear

On mobile the SDK registers the sink itself. In desktop host processes the SDK stays silent unless the host registers a logger (env_logger, tracing subscriber with a log bridge, etc.) — the SDK never claims the global logger out from under an application that owns it.

Reading logs

# Everything from the SDK, live:
adb logcat -s xybrid

# Only warnings and errors:
adb logcat -s xybrid:W

# Simulator:
xcrun simctl spawn booted log stream --predicate 'subsystem == "dev.xybrid.sdk"'

# Physical device (or use Console.app and filter by subsystem):
log stream --device --predicate 'subsystem == "dev.xybrid.sdk"'

What to look for

Telemetry not reaching the dashboard. Each failed export logs the attempt count and the underlying transport error:

W/xybrid: Transport error sending telemetry batch (attempt 1/3): Connection Failed

If you see nothing at all, telemetry isn't enabled — it starts only when an API key is configured at init.

Models failing to download. The registry client logs every endpoint it tries before giving up:

W/xybrid: URL https://registry.xybrid.dev failed: Registry unreachable:
          Failed to resolve model (connection refused or host unreachable), trying next

Both lines above share one common cause worth checking early: networks that block outbound TCP 443 (corporate and guest Wi-Fi are the usual suspects). The device can look "online" — ping works, captive portal passed — while every HTTPS connection from the app times out. Try another network or a hotspot before debugging further.

Model execution. Load and inference steps log at Info level (execution provider selection, preprocessing steps, completion), which makes it easy to confirm the SDK is actually being reached before suspecting anything deeper.

Log levels

The sinks are registered at Info level: warnings and errors always appear; per-request debug detail (HTTP connection traces, cache probing) is compiled in at Debug level and not emitted. There is currently no runtime knob to raise verbosity — if an Info-level log doesn't explain a failure, file an issue with the log lines you have.