lib/llm/src/http/service/openai.rs (2)

426-431: Chat completions: metrics plumbed correctly; also hit by rustfmt check.

Same note: formatting check failed around this hunk; cargo fmt should fix.

---

662-663: Responses: intentional unused stream handle; format to pass CI.

Using _stream_handle is fine for unary. The rustfmt check flagged this area; please run cargo fmt.

lib/llm/src/http/service/metrics.rs (2)

21-21: Prefer a Counter (and “_total” suffix) over a Gauge for disconnect events; consider labels.

Client disconnects are monotonically increasing events; Prometheus best practice is an IntCounter (optionally with labels like endpoint/request_type) named {prefix}_client_disconnects_total. Current IntGauge works but is semantically off.

Proposed change (minimal, no labels):

-    client_disconnect_gauge: prometheus::IntGauge,
+    client_disconnect_counter: prometheus::IntCounter,

-        let client_disconnect_gauge = prometheus::IntGauge::new(
-            frontend_metric_name("client_disconnects"),
-            "Number of connections dropped by clients",
-        )
-        .unwrap();
+        let client_disconnect_counter = prometheus::IntCounter::new(
+            frontend_metric_name("client_disconnects_total"),
+            "Total number of client-initiated disconnects",
+        )
+        .unwrap();

-            client_disconnect_gauge,
+            client_disconnect_counter,

-    /// Increment the gauge for client disconnections
-    pub fn inc_client_disconnect(&self) {
-        self.client_disconnect_gauge.inc();
-    }
+    /// Increment the counter for client disconnections
+    pub fn inc_client_disconnect(&self) {
+        self.client_disconnect_counter.inc();
+    }

-    /// Get the count of client disconnections
-    pub fn get_client_disconnect_count(&self) -> i64 {
-        self.client_disconnect_gauge.get()
-    }
+    /// Get the total count of client disconnections
+    pub fn get_client_disconnect_count(&self) -> u64 {
+        self.client_disconnect_counter.get()
+    }

-        registry.register(Box::new(self.client_disconnect_gauge.clone()))?;
+        registry.register(Box::new(self.client_disconnect_counter.clone()))?;

Optional follow-up: switch to IntCounterVec with labels endpoint, request_type for better slice-and-dice.

Also applies to: 137-142, 206-209, 274-283, 287-287

---

93-105: Docstring: add the new metric to the list for completeness.

Include the disconnect metric in the documented set so operators discover it.

 /// The following metrics will be created with the configured prefix:
 /// - `{prefix}_requests_total` - IntCounterVec for the total number of requests processed
 /// - `{prefix}_inflight_requests` - IntGaugeVec for the number of inflight requests
+/// - `{prefix}_client_disconnects_total` - IntCounter for client-initiated disconnects
 /// - `{prefix}_request_duration_seconds` - HistogramVec for the duration of requests

lib/llm/src/http/service/disconnect.rs (1)

136-139: Counting policy: increments on both connection and stream phases may double-count per request.

If you want “one disconnect per request,” dedupe increments (e.g., track a single AtomicBool and only inc once). Current behavior counts each unexpectedly dropped handle.

I can sketch a minimal dedupe using an Arc<AtomicBool> shared across both awaits if you want it.

Also applies to: 150-153

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro