Metrics reference¶

Three Prometheus metrics are emitted by the gateway. Names are stable; labels are stable within v1.0.0.

sgw_requests_total¶

Counter. One increment per request that reaches the gateway.

Label	Possible values	Notes
`model`	Real OpenRouter ID, `auto`, `smart-router`, `none`	The model field of the request (alias or `auto`) and the resolved real ID.
`status`	`200`, `400`, `403`, `500`, …	HTTP status code returned to the client.

Example series:

sgw_requests_total{model="deepseek/deepseek-v4-flash-0731",status="200"} 17
sgw_requests_total{model="smart-router",status="403"} 1
sgw_requests_total{model="z-ai/glm-5.2",status="200"} 1

Notes:

The model label uses the real OpenRouter ID for proxied requests, and smart-router for requests that were denied or short-circuited.
A 401 increments with status="401". The model label may be empty if the request did not carry a model field.

sgw_routing_decisions_total¶

Counter. One increment per routing decision inside the policy pipeline.

Label	Possible values	Notes
`action`	`ROUTE`, `SCOUT`, `STOP_FOR_HUMAN`, `FALLBACK_TRY`, `FALLBACK_GLM_INVOKE`, `FALLBACK_GLM_FAIL`	The action emitted by the policy.
`model`	Internal alias key (`deepseek_v4_flash`, `deepseek_v4_pro`, `nemotron_ultra`, `glm_5`) or `none`	The model the decision chose.

Example series:

sgw_routing_decisions_total{action="ROUTE",model="deepseek_v4_flash"} 17
sgw_routing_decisions_total{action="ROUTE",model="deepseek_v4_pro"} 4
sgw_routing_decisions_total{action="SCOUT",model="nemotron_ultra"} 2
sgw_routing_decisions_total{action="FALLBACK_TRY",model="deepseek_v4_pro"} 1
sgw_routing_decisions_total{action="FALLBACK_GLM_INVOKE",model="z-ai/glm-5.2"} 1

The decision counter increments before the upstream call; the request counter increments after. They will not match 1:1 during a failover window.

sgw_request_latency_seconds¶

Histogram. End-to-end request latency for non-streaming requests that completed successfully.

Label	Notes
(none)	—

The histogram uses Prometheus's default buckets (0.005, 0.01, 0.025, …, 2.5, 5, 10, +Inf).

The latency is measured inside the gateway, from the start of the FastAPI handler to the response. nginx and TLS overhead are not included. Streaming requests are not observed in v1.0.0.

Suggested queries¶

Total requests per model over 5 minutes¶

sum by (model) (rate(sgw_requests_total[5m]))

Error rate (any 5xx)¶

sum(rate(sgw_requests_total{status=~"5.."}[5m]))
/
sum(rate(sgw_requests_total[5m]))

95th percentile latency¶

histogram_quantile(0.95, sum(rate(sgw_request_latency_seconds_bucket[5m])) by (le))

Fallback utilization rate¶

sum(rate(sgw_routing_decisions_total{action="FALLBACK_GLM_INVOKE"}[5m]))
/
sum(rate(sgw_routing_decisions_total{action=~"ROUTE|FALLBACK.*"}[5m]))

Privacy-deny rate¶

sum(rate(sgw_requests_total{status="403"}[15m]))

A non-zero rate is expected when clients misuse the free alias. A sudden spike is a sign of a misconfigured client or a malicious attempt.