Changelog¶

Unreleased¶

OpenTelemetry meter bridge (latenzy.otel.OTelBridge), behind the new otel extra (pip install 'latenzy[otel]'). A RecordSink protocol + FanoutSink let the prober and live recorder emit to Prometheus and/or OpenTelemetry through one interface. Instrument names follow the OTel GenAI conventions (gen_ai.client.operation.duration, gen_ai.client.token.usage) plus latenzy-namespaced TTFT/throughput. Enable for latenzy run via a new otel config section (enabled, endpoint). Base wheel stays light — the OpenTelemetry packages are lazy-imported and never pulled in by import latenzy.
Passive live-traffic instrumentation: LiveRecorder / LiveObservation / classify_prompt / measure_stream record real application LLM-call latency into the same metric names under a new source="live" label (prober traffic is source="synthetic"). Label values are charset-validated and token counts bounded, matching the prober's hardening. Recording rules, the Grafana dashboard (new source filter), and examples/demo_live.py updated in lockstep. Metric semantics follow the OpenTelemetry GenAI conventions.

First real release.

Synthetic prober + Prometheus exporter for Anthropic, OpenAI, and Gemini. Per-model TTFT, total duration, streaming tokens/sec, and outcome metrics; fail-closed /metrics server; latenzy run / once / doctor CLI.
Grafana model-comparison dashboard (dashboards/, library ID 25642), hourly/daily recording rules and alert rules (prometheus/), and a docker-compose standalone bundle (deploy/) with prober + Prometheus + Grafana pre-provisioned.
Security-hardened through a four-round audit → red-team loop: bounded probe streaming, fail-closed response parsing, connection-capped exporter, no-default-secret deploy bundle with digest-pinned images. See SECURITY.md.