Releases: catch a release that made your AI app worse

Open Build → Releases (/releases).

Tell Shield360 the release

Shield360 reads each span’s release from service.version. Set it in the SDK option, or with one of these environment variables (checked in this order):

export SHIELD360_SERVICE_VERSION=2026.10.30
export SHIELD360_RELEASE=2026.10.30      # an alias
export OTEL_SERVICE_VERSION=2026.10.30   # the OpenTelemetry variable

If none is set but the span carries a VCS revision, the release is commit: followed by the first 12 characters of the commit.

The releases table

Pick an app and a window. The table lists the app’s releases from the last 30 days (up to 50), each with when it was first and last seen and its traffic.

Regression report

Select a release to compare it with the one before it. Shield360 takes the new release from when it was first seen, and the previous release over the same length of time (24 hours by default, 7 days at most), and compares:

Metric	Worse when
Cost per call (estimate)	higher
Tokens per call	higher
Latency p50 and p95	higher
Error rate	higher
Evaluation score	lower
Guard interventions	higher
Prompt injections	higher
Feedback score	lower

A metric counts as a regression only when it moves past both a percentage and a minimum change, and both releases have enough calls. This keeps a handful of calls from raising a false alarm.

Alert on a regression

Add an alert rule of the kind Release regression under Monitor → Alerts & cases → Alert rules. Choose the percentage, the minimum calls, the window, the metrics and, optionally, one app. Shield360 checks every 15 minutes and raises release.regression once per app, release and metric. There is no default rule.

Permissions

Viewing releases needs View telemetry.