Changelog
Every change, in public
We ship most weeks. Everything that touches your traces, evals or guardrails lands here first.
2.4.0
New
Guardrails 2.0
Inline PII, injection and policy checks now run in under six milliseconds at p95. Observe, warn and block modes are configurable per route.
Compiled rule engine replaces the classifier service for the default check suite.
New observe mode records what would have fired without affecting responses.
Per-route configuration, so you can block on support replies and only observe on internal tools.
2.3.2
Improved
Faster trace search
Full-text search across 90 days of traces is roughly eleven times faster, and filters now compose without a reload.
Search latency at p95 dropped from 4.1s to 340ms on large workspaces.
Filters for tool, model, user and error class can now be combined freely.
Saved searches are shareable with a permalink.
2.3.0
New
Eval suites in CI
Run your full eval suite as a pull-request check and block the merge when a rubric regresses beyond your threshold.
GitHub and GitLab checks with an inline score diff on the pull request.
Thresholds configurable per rubric rather than a single overall gate.
Golden datasets can be built directly from production traces.
2.2.4
Improved
Retention controls
Choose retention per project instead of per workspace, and redact fields on ingest rather than after the fact.
Per-project retention from 7 to 365 days.
Field-level redaction rules applied before storage.
Export any project to OpenTelemetry JSON on demand.
2.2.1
Fixed
Fixed span ordering on retries
Retried spans occasionally rendered out of order in the waterfall when a parent span finished before its child.
Waterfall now orders by start time with a stable tiebreak on span id.
Retry attempts are grouped under the original call.
Fixed a related off-by-one in the duration summary.
2.2.0
New
Cost budgets
Set a spend ceiling per agent, per customer or per route. Axon throttles or downgrades the model before you get a surprise invoice.
Budgets in USD or tokens, evaluated per run.
Automatic model downgrade as a soft limit action.
Alerts to Slack and PagerDuty at 80 and 100 percent.