Merge remote-tracking branch 'origin/pr/384' into codex/pr-376-377-383-384-combined

# Conflicts:
#	frontend/src/views/admin/PerformanceAnalysis.vue

apps/aether-gateway/src/handlers/admin/observability/stats/analytics_routes.rs

@@ -1,7 +1,7 @@
 use super::range::{build_comparison_range, parse_bounded_u32};
 use super::resolve_admin_usage_time_range;
 use crate::handlers::admin::request::{AdminAppState, AdminRequestContext};
-use crate::handlers::admin::shared::query_param_value;
+use crate::handlers::admin::shared::{query_param_optional_bool, query_param_value};
 use crate::GatewayError;
 use aether_admin::observability::stats::{
     admin_stats_bad_request_response, admin_stats_comparison_empty_response,
@@ -203,6 +203,14 @@ pub(super) async fn maybe_build_local_admin_stats_analytics_response(
             Ok(value) => value.unwrap_or(8) as usize,
             Err(detail) => return Ok(Some(admin_stats_bad_request_response(detail))),
         };
+        let slow_threshold_ms =
+            match query_param_value(request_context.query_string(), "slow_threshold_ms")
+                .map(|value| parse_bounded_u32("slow_threshold_ms", &value, 1, 600_000))
+                .transpose()
+            {
+                Ok(value) => u64::from(value.unwrap_or(10_000)),
+                Err(detail) => return Ok(Some(admin_stats_bad_request_response(detail))),
+            };
         if !state.has_usage_data_reader() {
             return Ok(Some(admin_stats_provider_performance_empty_response()));
         }
@@ -218,6 +226,16 @@ pub(super) async fn maybe_build_local_admin_stats_analytics_response(
                 granularity,
                 tz_offset_minutes: time_range.tz_offset_minutes,
                 limit,
+                provider_id: query_param_value(request_context.query_string(), "provider_id"),
+                model: query_param_value(request_context.query_string(), "model"),
+                api_format: query_param_value(request_context.query_string(), "api_format"),
+                endpoint_kind: query_param_value(request_context.query_string(), "endpoint_kind"),
+                is_stream: query_param_optional_bool(request_context.query_string(), "is_stream"),
+                has_format_conversion: query_param_optional_bool(
+                    request_context.query_string(),
+                    "has_format_conversion",
+                ),
+                slow_threshold_ms,
             })
             .await?;
         return Ok(Some(build_admin_stats_provider_performance_response(

apps/aether-gateway/src/tests/control/admin/stats.rs

@@ -960,6 +960,9 @@ async fn gateway_handles_admin_stats_provider_performance_locally_with_trusted_a
     assert_eq!(payload["summary"]["avg_output_tps"], 20.17);
     assert_eq!(payload["summary"]["avg_first_byte_time_ms"], 55.0);
     assert_eq!(payload["summary"]["avg_response_time_ms"], 590.91);
+    assert_eq!(payload["summary"]["p99_response_time_ms"], 1000);
+    assert_eq!(payload["summary"]["response_time_sample_count"], 11);
+    assert_eq!(payload["summary"]["slow_request_count"], 0);
 
     assert_eq!(payload["providers"].as_array().map(Vec::len), Some(2));
     assert_eq!(payload["providers"][0]["provider_id"], "provider-1");
@@ -973,9 +976,13 @@ async fn gateway_handles_admin_stats_provider_performance_locally_with_trusted_a
     assert_eq!(payload["providers"][0]["avg_first_byte_time_ms"], 55.0);
     assert_eq!(payload["providers"][0]["avg_response_time_ms"], 550.0);
     assert_eq!(payload["providers"][0]["p90_response_time_ms"], 910);
+    assert_eq!(payload["providers"][0]["p99_response_time_ms"], 991);
     assert_eq!(payload["providers"][0]["p90_first_byte_time_ms"], 91);
+    assert_eq!(payload["providers"][0]["p99_first_byte_time_ms"], 99);
     assert_eq!(payload["providers"][0]["tps_sample_count"], 10);
+    assert_eq!(payload["providers"][0]["response_time_sample_count"], 10);
     assert_eq!(payload["providers"][0]["first_byte_sample_count"], 10);
+    assert_eq!(payload["providers"][0]["slow_request_count"], 0);
 
     assert_eq!(payload["providers"][1]["provider_id"], "provider-2");
     assert_eq!(payload["providers"][1]["provider"], "Anthropic");
@@ -994,6 +1001,7 @@ async fn gateway_handles_admin_stats_provider_performance_locally_with_trusted_a
     assert_eq!(payload["timeline"][0]["provider_id"], "provider-1");
     assert_eq!(payload["timeline"][0]["avg_output_tps"], 20.2);
     assert_eq!(payload["timeline"][0]["success_rate"], 90.91);
+    assert_eq!(payload["timeline"][0]["slow_request_count"], 0);
     assert_eq!(payload["timeline"][1]["provider_id"], "provider-2");
     assert_eq!(
         payload["timeline"][1]["avg_first_byte_time_ms"],

crates/aether-admin/src/observability/stats.rs

@@ -1110,9 +1110,13 @@ pub fn build_admin_stats_provider_performance_response(
                 "avg_first_byte_time_ms": rounded_option(row.avg_first_byte_time_ms, 2),
                 "avg_response_time_ms": rounded_option(row.avg_response_time_ms, 2),
                 "p90_response_time_ms": row.p90_response_time_ms,
+                "p99_response_time_ms": row.p99_response_time_ms,
                 "p90_first_byte_time_ms": row.p90_first_byte_time_ms,
+                "p99_first_byte_time_ms": row.p99_first_byte_time_ms,
                 "tps_sample_count": row.tps_sample_count,
+                "response_time_sample_count": row.response_time_sample_count,
                 "first_byte_sample_count": row.first_byte_sample_count,
+                "slow_request_count": row.slow_request_count,
             })
         })
         .collect::<Vec<_>>();
@@ -1129,6 +1133,7 @@ pub fn build_admin_stats_provider_performance_response(
                 "avg_output_tps": rounded_option(row.avg_output_tps, 2),
                 "avg_first_byte_time_ms": rounded_option(row.avg_first_byte_time_ms, 2),
                 "avg_response_time_ms": rounded_option(row.avg_response_time_ms, 2),
+                "slow_request_count": row.slow_request_count,
                 "success_rate": success_rate(row.request_count, row.success_count),
             })
         })
@@ -1141,6 +1146,14 @@ pub fn build_admin_stats_provider_performance_response(
             "avg_output_tps": rounded_option(summary.avg_output_tps, 2),
             "avg_first_byte_time_ms": rounded_option(summary.avg_first_byte_time_ms, 2),
             "avg_response_time_ms": rounded_option(summary.avg_response_time_ms, 2),
+            "p90_response_time_ms": summary.p90_response_time_ms,
+            "p99_response_time_ms": summary.p99_response_time_ms,
+            "p90_first_byte_time_ms": summary.p90_first_byte_time_ms,
+            "p99_first_byte_time_ms": summary.p99_first_byte_time_ms,
+            "tps_sample_count": summary.tps_sample_count,
+            "response_time_sample_count": summary.response_time_sample_count,
+            "first_byte_sample_count": summary.first_byte_sample_count,
+            "slow_request_count": summary.slow_request_count,
         },
         "providers": providers,
         "timeline": timeline,

crates/aether-data-contracts/src/repository/usage/types.rs

@@ -957,6 +957,13 @@ pub struct UsageProviderPerformanceQuery {
     pub granularity: UsageTimeSeriesGranularity,
     pub tz_offset_minutes: i32,
     pub limit: usize,
+    pub provider_id: Option<String>,
+    pub model: Option<String>,
+    pub api_format: Option<String>,
+    pub endpoint_kind: Option<String>,
+    pub is_stream: Option<bool>,
+    pub has_format_conversion: Option<bool>,
+    pub slow_threshold_ms: u64,
 }
 
 #[derive(Debug, Clone, PartialEq, Default, serde::Serialize, serde::Deserialize)]
@@ -966,6 +973,14 @@ pub struct StoredUsageProviderPerformanceSummary {
     pub avg_output_tps: Option<f64>,
     pub avg_first_byte_time_ms: Option<f64>,
     pub avg_response_time_ms: Option<f64>,
+    pub p90_response_time_ms: Option<u64>,
+    pub p99_response_time_ms: Option<u64>,
+    pub p90_first_byte_time_ms: Option<u64>,
+    pub p99_first_byte_time_ms: Option<u64>,
+    pub tps_sample_count: u64,
+    pub response_time_sample_count: u64,
+    pub first_byte_sample_count: u64,
+    pub slow_request_count: u64,
 }
 
 #[derive(Debug, Clone, PartialEq, Default, serde::Serialize, serde::Deserialize)]
@@ -979,9 +994,13 @@ pub struct StoredUsageProviderPerformanceProviderRow {
     pub avg_first_byte_time_ms: Option<f64>,
     pub avg_response_time_ms: Option<f64>,
     pub p90_response_time_ms: Option<u64>,
+    pub p99_response_time_ms: Option<u64>,
     pub p90_first_byte_time_ms: Option<u64>,
+    pub p99_first_byte_time_ms: Option<u64>,
     pub tps_sample_count: u64,
+    pub response_time_sample_count: u64,
     pub first_byte_sample_count: u64,
+    pub slow_request_count: u64,
 }
 
 #[derive(Debug, Clone, PartialEq, Default, serde::Serialize, serde::Deserialize)]
@@ -995,6 +1014,7 @@ pub struct StoredUsageProviderPerformanceTimelineRow {
     pub avg_output_tps: Option<f64>,
     pub avg_first_byte_time_ms: Option<f64>,
     pub avg_response_time_ms: Option<f64>,
+    pub slow_request_count: u64,
 }
 
 #[derive(Debug, Clone, PartialEq, Default, serde::Serialize, serde::Deserialize)]

crates/aether-data/src/repository/usage/memory.rs

@@ -618,6 +618,36 @@ fn usage_matches_provider_performance_query(
     {
         return None;
     }
+    if let Some(provider_id) = query.provider_id.as_deref() {
+        if item.provider_id.as_deref() != Some(provider_id) {
+            return None;
+        }
+    }
+    if let Some(model) = query.model.as_deref() {
+        if item.model != model {
+            return None;
+        }
+    }
+    if let Some(api_format) = query.api_format.as_deref() {
+        if item.api_format.as_deref() != Some(api_format) {
+            return None;
+        }
+    }
+    if let Some(endpoint_kind) = query.endpoint_kind.as_deref() {
+        if item.endpoint_kind.as_deref() != Some(endpoint_kind) {
+            return None;
+        }
+    }
+    if let Some(is_stream) = query.is_stream {
+        if item.is_stream != is_stream {
+            return None;
+        }
+    }
+    if let Some(has_format_conversion) = query.has_format_conversion {
+        if item.has_format_conversion != has_format_conversion {
+            return None;
+        }
+    }
     usage_provider_performance_identity(item)
 }
 
@@ -1729,12 +1759,19 @@ impl UsageReadRepository for InMemoryUsageReadRepository {
             response_time_sample_count: u64,
             response_times: Vec<u64>,
             first_byte_times: Vec<u64>,
+            slow_request_count: u64,
         }
 
         impl ProviderPerformanceBucket {
-            fn add(&mut self, item: &StoredRequestUsageAudit) {
+            fn add(&mut self, item: &StoredRequestUsageAudit, slow_threshold_ms: u64) {
                 self.request_count = self.request_count.saturating_add(1);
                 self.output_tokens = self.output_tokens.saturating_add(item.output_tokens);
+                if item
+                    .response_time_ms
+                    .is_some_and(|value| value >= slow_threshold_ms)
+                {
+                    self.slow_request_count = self.slow_request_count.saturating_add(1);
+                }
                 if !usage_is_success(item) {
                     return;
                 }
@@ -1799,14 +1836,16 @@ impl UsageReadRepository for InMemoryUsageReadRepository {
             else {
                 continue;
             };
-            summary_bucket.add(item);
+            summary_bucket.add(item, query.slow_threshold_ms);
             let bucket = grouped.entry(provider_id).or_default();
             if bucket.provider.is_empty() {
                 bucket.provider = provider;
             }
-            bucket.add(item);
+            bucket.add(item, query.slow_threshold_ms);
         }
 
+        let mut summary_response_times = summary_bucket.response_times.clone();
+        let mut summary_first_byte_times = summary_bucket.first_byte_times.clone();
         let summary = StoredUsageProviderPerformanceSummary {
             request_count: summary_bucket.request_count,
             success_count: summary_bucket.success_count,
@@ -1822,14 +1861,25 @@ impl UsageReadRepository for InMemoryUsageReadRepository {
                 summary_bucket.response_time_ms_sum,
                 summary_bucket.response_time_sample_count,
             ),
+            p90_response_time_ms: usage_percentile_cont(&mut summary_response_times, 0.9),
+            p99_response_time_ms: usage_percentile_cont(&mut summary_response_times, 0.99),
+            p90_first_byte_time_ms: usage_percentile_cont(&mut summary_first_byte_times, 0.9),
+            p99_first_byte_time_ms: usage_percentile_cont(&mut summary_first_byte_times, 0.99),
+            tps_sample_count: summary_bucket.tps_sample_count,
+            response_time_sample_count: summary_bucket.response_time_sample_count,
+            first_byte_sample_count: summary_bucket.first_byte_sample_count,
+            slow_request_count: summary_bucket.slow_request_count,
         };
 
         let mut providers = grouped
             .into_iter()
             .map(|(provider_id, mut bucket)| {
                 let p90_response_time_ms = usage_percentile_cont(&mut bucket.response_times, 0.9);
+                let p99_response_time_ms = usage_percentile_cont(&mut bucket.response_times, 0.99);
                 let p90_first_byte_time_ms =
                     usage_percentile_cont(&mut bucket.first_byte_times, 0.9);
+                let p99_first_byte_time_ms =
+                    usage_percentile_cont(&mut bucket.first_byte_times, 0.99);
                 StoredUsageProviderPerformanceProviderRow {
                     provider_id,
                     provider: bucket.provider,
@@ -1849,9 +1899,13 @@ impl UsageReadRepository for InMemoryUsageReadRepository {
                         bucket.response_time_sample_count,
                     ),
                     p90_response_time_ms,
+                    p99_response_time_ms,
                     p90_first_byte_time_ms,
+                    p99_first_byte_time_ms,
                     tps_sample_count: bucket.tps_sample_count,
+                    response_time_sample_count: bucket.response_time_sample_count,
                     first_byte_sample_count: bucket.first_byte_sample_count,
+                    slow_request_count: bucket.slow_request_count,
                 }
             })
             .collect::<Vec<_>>();
@@ -1888,7 +1942,7 @@ impl UsageReadRepository for InMemoryUsageReadRepository {
             if bucket.provider.is_empty() {
                 bucket.provider = provider;
             }
-            bucket.add(item);
+            bucket.add(item, query.slow_threshold_ms);
         }
 
         let timeline = timeline_grouped
@@ -1913,6 +1967,7 @@ impl UsageReadRepository for InMemoryUsageReadRepository {
                         bucket.response_time_ms_sum,
                         bucket.response_time_sample_count,
                     ),
+                    slow_request_count: bucket.slow_request_count,
                 },
             )
             .collect();
@@ -4875,6 +4930,13 @@ mod tests {
                 granularity: UsageTimeSeriesGranularity::Hour,
                 tz_offset_minutes: 0,
                 limit: 1,
+                provider_id: None,
+                model: None,
+                api_format: None,
+                endpoint_kind: None,
+                is_stream: None,
+                has_format_conversion: None,
+                slow_threshold_ms: 10_000,
             })
             .await
             .expect("provider performance should summarize");

crates/aether-data/src/repository/usage/postgres/mod.rs

@@ -1045,6 +1045,38 @@ fn decode_usage_provider_performance_summary(
         avg_response_time_ms: row
             .try_get::<Option<f64>, _>("avg_response_time_ms")
             .map_postgres_err()?,
+        p90_response_time_ms: row
+            .try_get::<Option<i64>, _>("p90_response_time_ms")
+            .map_postgres_err()?
+            .map(|value| value.max(0) as u64),
+        p99_response_time_ms: row
+            .try_get::<Option<i64>, _>("p99_response_time_ms")
+            .map_postgres_err()?
+            .map(|value| value.max(0) as u64),
+        p90_first_byte_time_ms: row
+            .try_get::<Option<i64>, _>("p90_first_byte_time_ms")
+            .map_postgres_err()?
+            .map(|value| value.max(0) as u64),
+        p99_first_byte_time_ms: row
+            .try_get::<Option<i64>, _>("p99_first_byte_time_ms")
+            .map_postgres_err()?
+            .map(|value| value.max(0) as u64),
+        tps_sample_count: row
+            .try_get::<i64, _>("tps_sample_count")
+            .map_postgres_err()?
+            .max(0) as u64,
+        response_time_sample_count: row
+            .try_get::<i64, _>("response_time_sample_count")
+            .map_postgres_err()?
+            .max(0) as u64,
+        first_byte_sample_count: row
+            .try_get::<i64, _>("first_byte_sample_count")
+            .map_postgres_err()?
+            .max(0) as u64,
+        slow_request_count: row
+            .try_get::<i64, _>("slow_request_count")
+            .map_postgres_err()?
+            .max(0) as u64,
     })
 }
 
@@ -1079,18 +1111,34 @@ fn decode_usage_provider_performance_provider_row(
             .try_get::<Option<i64>, _>("p90_response_time_ms")
             .map_postgres_err()?
             .map(|value| value.max(0) as u64),
+        p99_response_time_ms: row
+            .try_get::<Option<i64>, _>("p99_response_time_ms")
+            .map_postgres_err()?
+            .map(|value| value.max(0) as u64),
         p90_first_byte_time_ms: row
             .try_get::<Option<i64>, _>("p90_first_byte_time_ms")
             .map_postgres_err()?
             .map(|value| value.max(0) as u64),
+        p99_first_byte_time_ms: row
+            .try_get::<Option<i64>, _>("p99_first_byte_time_ms")
+            .map_postgres_err()?
+            .map(|value| value.max(0) as u64),
         tps_sample_count: row
             .try_get::<i64, _>("tps_sample_count")
             .map_postgres_err()?
             .max(0) as u64,
+        response_time_sample_count: row
+            .try_get::<i64, _>("response_time_sample_count")
+            .map_postgres_err()?
+            .max(0) as u64,
         first_byte_sample_count: row
             .try_get::<i64, _>("first_byte_sample_count")
             .map_postgres_err()?
             .max(0) as u64,
+        slow_request_count: row
+            .try_get::<i64, _>("slow_request_count")
+            .map_postgres_err()?
+            .max(0) as u64,
     })
 }
 
@@ -1122,9 +1170,64 @@ fn decode_usage_provider_performance_timeline_row(
         avg_response_time_ms: row
             .try_get::<Option<f64>, _>("avg_response_time_ms")
             .map_postgres_err()?,
+        slow_request_count: row
+            .try_get::<i64, _>("slow_request_count")
+            .map_postgres_err()?
+            .max(0) as u64,
     })
 }
 
+fn push_usage_provider_performance_text_filter(
+    builder: &mut QueryBuilder<'_, Postgres>,
+    column: &'static str,
+    value: &Option<String>,
+) {
+    let Some(value) = value
+        .as_deref()
+        .map(str::trim)
+        .filter(|value| !value.is_empty())
+    else {
+        return;
+    };
+    builder
+        .push(" AND NULLIF(BTRIM(COALESCE(")
+        .push(column)
+        .push(", '')), '') = ")
+        .push_bind(value.to_string());
+}
+
+fn push_usage_provider_performance_filters(
+    builder: &mut QueryBuilder<'_, Postgres>,
+    query: &UsageProviderPerformanceQuery,
+) {
+    push_usage_provider_performance_text_filter(
+        builder,
+        r#""usage".provider_id"#,
+        &query.provider_id,
+    );
+    push_usage_provider_performance_text_filter(builder, r#""usage".model"#, &query.model);
+    push_usage_provider_performance_text_filter(
+        builder,
+        r#""usage".api_format"#,
+        &query.api_format,
+    );
+    push_usage_provider_performance_text_filter(
+        builder,
+        r#""usage".endpoint_kind"#,
+        &query.endpoint_kind,
+    );
+    if let Some(is_stream) = query.is_stream {
+        builder
+            .push(r#" AND "usage".is_stream = "#)
+            .push_bind(is_stream);
+    }
+    if let Some(has_format_conversion) = query.has_format_conversion {
+        builder
+            .push(r#" AND "usage".has_format_conversion = "#)
+            .push_bind(has_format_conversion);
+    }
+}
+
 fn decode_usage_time_series_bucket_row(
     row: &PgRow,
 ) -> Result<StoredUsageTimeSeriesBucket, DataLayerError> {
@@ -4886,6 +4989,11 @@ WITH filtered_usage AS (
     AND NULLIF(BTRIM(COALESCE("usage".provider_id, '')), '') IS NOT NULL
     AND lower(BTRIM(COALESCE("usage".provider_id, ''))) NOT IN ('unknown', 'pending')
     AND lower(BTRIM(COALESCE("usage".provider_name, ''))) NOT IN ('unknown', 'pending')
+"#,
+        );
+        push_usage_provider_performance_filters(&mut builder, query);
+        builder.push(
+            r#"
 )
 SELECT
   COUNT(*)::BIGINT AS request_count,
@@ -4910,7 +5018,49 @@ SELECT
   AVG(first_byte_time_ms::DOUBLE PRECISION)
     FILTER (WHERE success_flag = 1 AND has_first_byte_time) AS avg_first_byte_time_ms,
   AVG(response_time_ms::DOUBLE PRECISION)
-    FILTER (WHERE success_flag = 1 AND has_response_time) AS avg_response_time_ms
+    FILTER (WHERE success_flag = 1 AND has_response_time) AS avg_response_time_ms,
+  CASE
+    WHEN COUNT(response_time_ms) FILTER (WHERE success_flag = 1 AND has_response_time) >= 10
+    THEN FLOOR(PERCENTILE_CONT(0.9) WITHIN GROUP (ORDER BY response_time_ms)
+      FILTER (WHERE success_flag = 1 AND has_response_time))::BIGINT
+    ELSE NULL
+  END AS p90_response_time_ms,
+  CASE
+    WHEN COUNT(response_time_ms) FILTER (WHERE success_flag = 1 AND has_response_time) >= 10
+    THEN FLOOR(PERCENTILE_CONT(0.99) WITHIN GROUP (ORDER BY response_time_ms)
+      FILTER (WHERE success_flag = 1 AND has_response_time))::BIGINT
+    ELSE NULL
+  END AS p99_response_time_ms,
+  CASE
+    WHEN COUNT(first_byte_time_ms) FILTER (WHERE success_flag = 1 AND has_first_byte_time) >= 10
+    THEN FLOOR(PERCENTILE_CONT(0.9) WITHIN GROUP (ORDER BY first_byte_time_ms)
+      FILTER (WHERE success_flag = 1 AND has_first_byte_time))::BIGINT
+    ELSE NULL
+  END AS p90_first_byte_time_ms,
+  CASE
+    WHEN COUNT(first_byte_time_ms) FILTER (WHERE success_flag = 1 AND has_first_byte_time) >= 10
+    THEN FLOOR(PERCENTILE_CONT(0.99) WITHIN GROUP (ORDER BY first_byte_time_ms)
+      FILTER (WHERE success_flag = 1 AND has_first_byte_time))::BIGINT
+    ELSE NULL
+  END AS p99_first_byte_time_ms,
+  COALESCE(SUM(CASE
+    WHEN success_flag = 1 AND response_time_ms > 0 AND output_tokens > 0
+    THEN 1
+    ELSE 0
+  END), 0)::BIGINT AS tps_sample_count,
+  (COUNT(response_time_ms) FILTER (WHERE success_flag = 1 AND has_response_time))::BIGINT
+    AS response_time_sample_count,
+  (COUNT(first_byte_time_ms) FILTER (WHERE success_flag = 1 AND has_first_byte_time))::BIGINT
+    AS first_byte_sample_count,
+  COALESCE(SUM(CASE
+    WHEN has_response_time AND response_time_ms >= "#,
+        );
+        builder.push_bind(query.slow_threshold_ms as i64);
+        builder.push(
+            r#"
+    THEN 1
+    ELSE 0
+  END), 0)::BIGINT AS slow_request_count
 FROM filtered_usage
 "#,
         );
@@ -4970,6 +5120,11 @@ WITH filtered_usage AS (
     AND NULLIF(BTRIM(COALESCE("usage".provider_id, '')), '') IS NOT NULL
     AND lower(BTRIM(COALESCE("usage".provider_id, ''))) NOT IN ('unknown', 'pending')
     AND lower(BTRIM(COALESCE("usage".provider_name, ''))) NOT IN ('unknown', 'pending')
+"#,
+        );
+        push_usage_provider_performance_filters(&mut builder, query);
+        builder.push(
+            r#"
 )
 SELECT
   provider_id,
@@ -5004,19 +5159,42 @@ SELECT
       FILTER (WHERE success_flag = 1 AND has_response_time))::BIGINT
     ELSE NULL
   END AS p90_response_time_ms,
+  CASE
+    WHEN COUNT(response_time_ms) FILTER (WHERE success_flag = 1 AND has_response_time) >= 10
+    THEN FLOOR(PERCENTILE_CONT(0.99) WITHIN GROUP (ORDER BY response_time_ms)
+      FILTER (WHERE success_flag = 1 AND has_response_time))::BIGINT
+    ELSE NULL
+  END AS p99_response_time_ms,
   CASE
     WHEN COUNT(first_byte_time_ms) FILTER (WHERE success_flag = 1 AND has_first_byte_time) >= 10
     THEN FLOOR(PERCENTILE_CONT(0.9) WITHIN GROUP (ORDER BY first_byte_time_ms)
       FILTER (WHERE success_flag = 1 AND has_first_byte_time))::BIGINT
     ELSE NULL
   END AS p90_first_byte_time_ms,
+  CASE
+    WHEN COUNT(first_byte_time_ms) FILTER (WHERE success_flag = 1 AND has_first_byte_time) >= 10
+    THEN FLOOR(PERCENTILE_CONT(0.99) WITHIN GROUP (ORDER BY first_byte_time_ms)
+      FILTER (WHERE success_flag = 1 AND has_first_byte_time))::BIGINT
+    ELSE NULL
+  END AS p99_first_byte_time_ms,
   COALESCE(SUM(CASE
     WHEN success_flag = 1 AND response_time_ms > 0 AND output_tokens > 0
     THEN 1
     ELSE 0
   END), 0)::BIGINT AS tps_sample_count,
+  (COUNT(response_time_ms) FILTER (WHERE success_flag = 1 AND has_response_time))::BIGINT
+    AS response_time_sample_count,
   (COUNT(first_byte_time_ms) FILTER (WHERE success_flag = 1 AND has_first_byte_time))::BIGINT
-    AS first_byte_sample_count
+    AS first_byte_sample_count,
+  COALESCE(SUM(CASE
+    WHEN has_response_time AND response_time_ms >= "#,
+        );
+        builder.push_bind(query.slow_threshold_ms as i64);
+        builder.push(
+            r#"
+    THEN 1
+    ELSE 0
+  END), 0)::BIGINT AS slow_request_count
 FROM filtered_usage
 GROUP BY provider_id
 ORDER BY request_count DESC, provider_id ASC
@@ -5096,6 +5274,11 @@ ORDER BY request_count DESC, provider_id ASC
     AND NULLIF(BTRIM(COALESCE("usage".provider_id, '')), '') IS NOT NULL
     AND lower(BTRIM(COALESCE("usage".provider_id, ''))) NOT IN ('unknown', 'pending')
     AND lower(BTRIM(COALESCE("usage".provider_name, ''))) NOT IN ('unknown', 'pending')
+"#,
+        );
+        push_usage_provider_performance_filters(&mut builder, query);
+        builder.push(
+            r#"
     AND "usage".provider_id = ANY("#,
         );
         builder.push_bind(provider_ids.to_vec());
@@ -5129,7 +5312,16 @@ SELECT
   AVG(first_byte_time_ms::DOUBLE PRECISION)
     FILTER (WHERE success_flag = 1 AND has_first_byte_time) AS avg_first_byte_time_ms,
   AVG(response_time_ms::DOUBLE PRECISION)
-    FILTER (WHERE success_flag = 1 AND has_response_time) AS avg_response_time_ms
+    FILTER (WHERE success_flag = 1 AND has_response_time) AS avg_response_time_ms,
+  COALESCE(SUM(CASE
+    WHEN has_response_time AND response_time_ms >= "#,
+        );
+        builder.push_bind(query.slow_threshold_ms as i64);
+        builder.push(
+            r#"
+    THEN 1
+    ELSE 0
+  END), 0)::BIGINT AS slow_request_count
 FROM filtered_usage
 GROUP BY date, provider_id
 ORDER BY date ASC, provider_id ASC

frontend/src/api/admin.ts

@@ -453,6 +453,14 @@ export interface ProviderPerformanceSummary {
   avg_output_tps: number | null
   avg_first_byte_time_ms: number | null
   avg_response_time_ms: number | null
+  p90_response_time_ms?: number | null
+  p99_response_time_ms?: number | null
+  p90_first_byte_time_ms?: number | null
+  p99_first_byte_time_ms?: number | null
+  tps_sample_count?: number
+  response_time_sample_count?: number
+  first_byte_sample_count?: number
+  slow_request_count?: number
 }
 
 export interface ProviderPerformanceItem {
@@ -467,9 +475,13 @@ export interface ProviderPerformanceItem {
   avg_first_byte_time_ms: number | null
   avg_response_time_ms: number | null
   p90_response_time_ms: number | null
+  p99_response_time_ms?: number | null
   p90_first_byte_time_ms: number | null
+  p99_first_byte_time_ms?: number | null
   tps_sample_count: number
+  response_time_sample_count?: number
   first_byte_sample_count: number
+  slow_request_count?: number
 }
 
 export interface ProviderPerformanceTimelineItem {
@@ -482,6 +494,7 @@ export interface ProviderPerformanceTimelineItem {
   avg_first_byte_time_ms: number | null
   avg_response_time_ms: number | null
   success_rate: number
+  slow_request_count?: number
 }
 
 export interface ProviderPerformanceResponse {
@@ -983,6 +996,13 @@ export const adminApi = {
     tz_offset_minutes?: number
     granularity?: 'day' | 'hour'
     limit?: number
+    provider_id?: string
+    model?: string
+    api_format?: string
+    endpoint_kind?: string
+    is_stream?: boolean
+    has_format_conversion?: boolean
+    slow_threshold_ms?: number
   }): Promise<ProviderPerformanceResponse> {
     const cacheKey = buildCacheKey('admin:stats:performance:providers', params)
     return cachedRequest(

frontend/src/api/monitoring.ts

@@ -118,8 +118,21 @@ export interface GatewayMetricsSummary {
   distributed: GatewayGateMetrics
   tunnel: {
     proxyConnections: number | null
+    availableProxyConnections: number | null
+    closingProxyConnections: number | null
+    drainingProxyConnections: number | null
+    softAvoidProxyConnections: number | null
     nodes: number | null
     activeStreams: number | null
+    outboundQueueDepthTotal: number | null
+    outboundQueueDepthMax: number | null
+    outboundQueueCapacityTotal: number | null
+    outboundQueueRejectedFullTotal: number | null
+    outboundQueueRejectedClosedTotal: number | null
+    proxyConnectionCongestedTotal: number | null
+    softAvoidSelectionTotal: number | null
+    selectionRetryTotal: number | null
+    selectionUnavailableTotal: number | null
   }
   fallbackTotal: number
   fallbacks: GatewayFallbackMetricSummary[]
@@ -159,8 +172,21 @@ export function buildGatewayMetricsSummary(text: string): GatewayMetricsSummary
     distributed: buildGateMetrics(samples, 'gateway_requests_distributed'),
     tunnel: {
       proxyConnections: findMetricValueNumber(samples, 'tunnel_proxy_connections'),
+      availableProxyConnections: findMetricValueNumber(samples, 'tunnel_proxy_connections_available'),
+      closingProxyConnections: findMetricValueNumber(samples, 'tunnel_proxy_connections_closing'),
+      drainingProxyConnections: findMetricValueNumber(samples, 'tunnel_proxy_connections_draining'),
+      softAvoidProxyConnections: findMetricValueNumber(samples, 'tunnel_proxy_connections_soft_avoid'),
       nodes: findMetricValueNumber(samples, 'tunnel_nodes'),
       activeStreams: findMetricValueNumber(samples, 'tunnel_active_streams'),
+      outboundQueueDepthTotal: findMetricValueNumber(samples, 'tunnel_proxy_outbound_queue_depth_total'),
+      outboundQueueDepthMax: findMetricValueNumber(samples, 'tunnel_proxy_outbound_queue_depth_max'),
+      outboundQueueCapacityTotal: findMetricValueNumber(samples, 'tunnel_proxy_outbound_queue_capacity_total'),
+      outboundQueueRejectedFullTotal: findMetricValueNumber(samples, 'tunnel_proxy_outbound_queue_rejected_full_total'),
+      outboundQueueRejectedClosedTotal: findMetricValueNumber(samples, 'tunnel_proxy_outbound_queue_rejected_closed_total'),
+      proxyConnectionCongestedTotal: findMetricValueNumber(samples, 'tunnel_proxy_connection_congested_total'),
+      softAvoidSelectionTotal: findMetricValueNumber(samples, 'tunnel_proxy_soft_avoid_selection_total'),
+      selectionRetryTotal: findMetricValueNumber(samples, 'tunnel_proxy_selection_retry_total'),
+      selectionUnavailableTotal: findMetricValueNumber(samples, 'tunnel_proxy_selection_unavailable_total'),
     },
     fallbackTotal: fallbacks.reduce((total, item) => total + item.total, 0),
     fallbacks,

frontend/src/views/admin/PerformanceAnalysis.vue

@@ -21,7 +21,10 @@
           title="刷新实时与历史性能数据"
           @click="handleManualRefresh"
         />
-        <TimeRangePicker v-model="timeRange" />
+        <TimeRangePicker
+          v-model="timeRange"
+          :show-granularity="false"
+        />
       </div>
     </div>
 
@@ -30,10 +33,10 @@
         <div class="flex flex-col gap-2 sm:flex-row sm:items-center sm:justify-between">
           <div>
             <h2 class="text-sm font-semibold">
-              实时性能面板
+              实时运行状态
             </h2>
             <p class="text-xs text-muted-foreground">
-              聚合系统状态、并发门、Tunnel 与 fallback 指标
+              聚合系统健康、并发保护、代理通道与降级切换
             </p>
           </div>
           <div class="flex flex-wrap items-center gap-2">
@@ -98,19 +101,30 @@
 
           <div class="grid grid-cols-1 gap-4 xl:grid-cols-3">
             <div class="grid grid-cols-1 gap-4 lg:grid-cols-2 xl:col-span-2">
-              <section class="rounded-xl border border-border/70 bg-card/60 p-4">
+              <section class="rounded-xl border border-border/70 bg-card/60 p-4 lg:col-span-2">
                 <div class="flex items-center justify-between gap-3">
                   <h3 class="text-sm font-semibold">
-                    本地并发门
+                    并发保护
                   </h3>
                   <Badge variant="outline">
-                    gateway_requests
+                    全局 {{ distributedGateText }}
+                  </Badge>
+                </div>
+
+                <div class="mt-4 grid grid-cols-1 gap-3 lg:grid-cols-2">
+                  <div class="rounded-lg border border-border/60 bg-background/50 px-3 py-3">
+                    <div class="flex items-center justify-between gap-3">
+                      <div class="text-sm font-medium">
+                        当前节点
+                      </div>
+                      <Badge variant="outline">
+                        本机
                       </Badge>
                     </div>
                     <div class="mt-4 grid grid-cols-2 gap-3 text-sm">
                       <div>
                         <div class="text-xs text-muted-foreground">
-                      In Flight
+                          处理中
                         </div>
                         <div class="mt-1 text-lg font-semibold">
                           {{ formatMetricNumber(gatewayMetrics?.local.inFlight) }}
@@ -118,7 +132,7 @@
                       </div>
                       <div>
                         <div class="text-xs text-muted-foreground">
-                      Available
+                          可接入
                         </div>
                         <div class="mt-1 text-lg font-semibold">
                           {{ formatMetricNumber(gatewayMetrics?.local.availablePermits) }}
@@ -126,7 +140,7 @@
                       </div>
                       <div>
                         <div class="text-xs text-muted-foreground">
-                      High Watermark
+                          峰值并发
                         </div>
                         <div class="mt-1 text-lg font-semibold">
                           {{ formatMetricNumber(gatewayMetrics?.local.highWatermark) }}
@@ -134,20 +148,20 @@
                       </div>
                       <div>
                         <div class="text-xs text-muted-foreground">
-                      Rejected Total
+                          被限流
                         </div>
                         <div class="mt-1 text-lg font-semibold">
                           {{ formatMetricNumber(gatewayMetrics?.local.rejectedTotal) }}
                         </div>
                       </div>
                     </div>
-              </section>
+                  </div>
 
-              <section class="rounded-xl border border-border/70 bg-card/60 p-4">
+                  <div class="rounded-lg border border-border/60 bg-background/50 px-3 py-3">
                     <div class="flex items-center justify-between gap-3">
-                  <h3 class="text-sm font-semibold">
+                      <div class="text-sm font-medium">
-                    分布式并发门
+                        全局保护
-                  </h3>
+                      </div>
                       <Badge :variant="distributedGateVariant">
                         {{ distributedGateText }}
                       </Badge>
@@ -155,7 +169,7 @@
                     <div class="mt-4 grid grid-cols-2 gap-3 text-sm">
                       <div>
                         <div class="text-xs text-muted-foreground">
-                      In Flight
+                          处理中
                         </div>
                         <div class="mt-1 text-lg font-semibold">
                           {{ formatMetricNumber(gatewayMetrics?.distributed.inFlight) }}
@@ -163,7 +177,7 @@
                       </div>
                       <div>
                         <div class="text-xs text-muted-foreground">
-                      Available
+                          可接入
                         </div>
                         <div class="mt-1 text-lg font-semibold">
                           {{ formatMetricNumber(gatewayMetrics?.distributed.availablePermits) }}
@@ -171,7 +185,7 @@
                       </div>
                       <div>
                         <div class="text-xs text-muted-foreground">
-                      High Watermark
+                          峰值并发
                         </div>
                         <div class="mt-1 text-lg font-semibold">
                           {{ formatMetricNumber(gatewayMetrics?.distributed.highWatermark) }}
@@ -179,25 +193,27 @@
                       </div>
                       <div>
                         <div class="text-xs text-muted-foreground">
-                      Rejected Total
+                          被限流
                         </div>
                         <div class="mt-1 text-lg font-semibold">
                           {{ formatMetricNumber(gatewayMetrics?.distributed.rejectedTotal) }}
                         </div>
                       </div>
                     </div>
+                  </div>
+                </div>
                 <p
                   v-if="gatewayMetrics?.distributed.unavailable"
                   class="mt-3 text-xs text-yellow-700 dark:text-yellow-300"
                 >
-                  Redis 分布式并发快照当前不可用，需要检查 gate 后端连接。
+                  全局并发保护暂不可用，请检查 Redis 连接。
                 </p>
               </section>
 
               <section class="rounded-xl border border-border/70 bg-card/60 p-4">
                 <div class="flex items-center justify-between gap-3">
                   <h3 class="text-sm font-semibold">
-                    Tunnel / 代理
+                    代理通道
                   </h3>
                   <Badge variant="outline">
                     实时连接
@@ -206,7 +222,7 @@
                 <div class="mt-4 grid grid-cols-2 gap-3 text-sm">
                   <div>
                     <div class="text-xs text-muted-foreground">
-                      Nodes
+                      节点数
                     </div>
                     <div class="mt-1 text-lg font-semibold">
                       {{ formatMetricNumber(currentTunnelNodes) }}
@@ -214,15 +230,15 @@
                   </div>
                   <div>
                     <div class="text-xs text-muted-foreground">
-                      Proxy Connections
+                      可用连接
                     </div>
                     <div class="mt-1 text-lg font-semibold">
-                      {{ formatMetricNumber(currentProxyConnections) }}
+                      {{ formatMetricNumber(gatewayMetrics?.tunnel.availableProxyConnections) }}
                     </div>
                   </div>
                   <div>
                     <div class="text-xs text-muted-foreground">
-                      Active Streams
+                      活跃流
                     </div>
                     <div class="mt-1 text-lg font-semibold">
                       {{ formatMetricNumber(currentActiveStreams) }}
@@ -230,10 +246,10 @@
                   </div>
                   <div>
                     <div class="text-xs text-muted-foreground">
-                      Service Up
+                      避让连接
                     </div>
                     <div class="mt-1 text-lg font-semibold">
-                      {{ gatewayMetrics?.serviceUp === 1 ? '在线' : '未知' }}
+                      {{ formatMetricNumber(gatewayMetrics?.tunnel.softAvoidProxyConnections) }}
                     </div>
                   </div>
                 </div>
@@ -242,43 +258,43 @@
               <section class="rounded-xl border border-border/70 bg-card/60 p-4">
                 <div class="flex items-center justify-between gap-3">
                   <h3 class="text-sm font-semibold">
-                    今日请求概况
+                    代理通道压力
                   </h3>
                   <Badge variant="outline">
-                    {{ systemStatus?.internal_gateway.status || 'gateway' }}
+                    排队 {{ tunnelQueueUtilizationText }}
                   </Badge>
                 </div>
                 <div class="mt-4 grid grid-cols-2 gap-3 text-sm">
                   <div>
                     <div class="text-xs text-muted-foreground">
-                      Requests
+                      排队中
                     </div>
                     <div class="mt-1 text-lg font-semibold">
-                      {{ formatMetricNumber(systemStatus?.today_stats.requests) }}
+                      {{ formatMetricNumber(gatewayMetrics?.tunnel.outboundQueueDepthTotal) }}
                     </div>
                   </div>
                   <div>
                     <div class="text-xs text-muted-foreground">
-                      Tokens
+                      峰值排队
                     </div>
                     <div class="mt-1 text-lg font-semibold">
-                      {{ formatTokens(systemStatus?.today_stats.tokens) }}
+                      {{ formatMetricNumber(gatewayMetrics?.tunnel.outboundQueueDepthMax) }}
                     </div>
                   </div>
                   <div>
                     <div class="text-xs text-muted-foreground">
-                      Cost
+                      队列满拒绝
                     </div>
                     <div class="mt-1 text-lg font-semibold">
-                      {{ systemStatus?.today_stats.cost_usd || '-' }}
+                      {{ formatMetricNumber(tunnelQueueRejectedTotal) }}
                     </div>
                   </div>
                   <div>
                     <div class="text-xs text-muted-foreground">
-                      Active Providers / Keys
+                      无可用通道
                     </div>
                     <div class="mt-1 text-lg font-semibold">
-                      {{ providerAndKeySummary }}
+                      {{ formatMetricNumber(tunnelSelectionPressureTotal) }}
                     </div>
                   </div>
                 </div>
@@ -288,7 +304,7 @@
             <section class="rounded-xl border border-border/70 bg-card/60 p-4">
               <div class="flex items-center justify-between gap-3">
                 <h3 class="text-sm font-semibold">
-                  Fallback 统计
+                  降级切换统计
                 </h3>
                 <span class="text-xs text-muted-foreground">
                   总计 {{ formatMetricNumber(gatewayMetrics?.fallbackTotal) }}
@@ -299,7 +315,7 @@
                 v-if="!fallbackRows.length"
                 class="mt-4 rounded-lg border border-dashed border-border/70 px-3 py-4 text-sm text-muted-foreground"
               >
-                当前没有记录到 fallback 计数。
+                当前没有记录到降级切换。
               </div>
 
               <div
@@ -372,7 +388,7 @@
                       HTTP {{ item.context.status_code ?? '-' }}
                     </Badge>
                     <Badge variant="outline">
-                      {{ item.context.provider_name || item.context.provider_id || '未知 Provider' }}
+                      {{ item.context.provider_name || item.context.provider_id || '未知上游' }}
                     </Badge>
                     <Badge variant="outline">
                       {{ item.context.api_format || item.context.model || '未知格式' }}
@@ -495,15 +511,90 @@
       <div class="flex flex-col gap-2 sm:flex-row sm:items-center sm:justify-between">
         <div>
           <h3 class="text-sm font-semibold">
-            Provider 性能
+            上游服务性能
           </h3>
           <p class="text-xs text-muted-foreground">
             {{ providerPerformanceSubtitle }}
           </p>
         </div>
+        <div class="flex items-center gap-2">
           <Badge variant="outline">
             Top {{ providerPerformanceRows.length || 0 }}
           </Badge>
+          <Button
+            v-if="hasProviderPerformanceFilters"
+            variant="ghost"
+            size="sm"
+            class="h-8 gap-1 px-2 text-xs"
+            title="清除上游服务性能筛选"
+            @click="resetProviderPerformanceFilters"
+          >
+            <FilterX class="h-3.5 w-3.5" />
+            清除
+          </Button>
+        </div>
+      </div>
+      <div class="grid grid-cols-1 gap-2 pt-1 sm:grid-cols-2 xl:grid-cols-4 2xl:grid-cols-7">
+        <Input
+          v-model="providerPerformanceProviderId"
+          size="sm"
+          placeholder="上游 ID"
+        />
+        <Input
+          v-model="providerPerformanceModel"
+          size="sm"
+          placeholder="模型"
+        />
+        <Input
+          v-model="providerPerformanceApiFormat"
+          size="sm"
+          placeholder="API 格式"
+        />
+        <Input
+          v-model="providerPerformanceEndpointKind"
+          size="sm"
+          placeholder="端点类型"
+        />
+        <Select v-model="providerPerformanceIsStream">
+          <SelectTrigger class="h-8 text-xs border-border/60">
+            <SelectValue placeholder="流式" />
+          </SelectTrigger>
+          <SelectContent>
+            <SelectItem value="all">
+              全部流式
+            </SelectItem>
+            <SelectItem value="true">
+              仅流式
+            </SelectItem>
+            <SelectItem value="false">
+              非流式
+            </SelectItem>
+          </SelectContent>
+        </Select>
+        <Select v-model="providerPerformanceHasFormatConversion">
+          <SelectTrigger class="h-8 text-xs border-border/60">
+            <SelectValue placeholder="格式转换" />
+          </SelectTrigger>
+          <SelectContent>
+            <SelectItem value="all">
+              全部转换
+            </SelectItem>
+            <SelectItem value="true">
+              仅转换
+            </SelectItem>
+            <SelectItem value="false">
+              不转换
+            </SelectItem>
+          </SelectContent>
+        </Select>
+        <Input
+          v-model="providerPerformanceSlowThresholdMs"
+          size="sm"
+          type="number"
+          min="1"
+          max="600000"
+          placeholder="慢请求阈值 ms"
+        />
       </div>
 
       <div
@@ -548,7 +639,7 @@
             <thead class="bg-muted/30 text-xs text-muted-foreground">
               <tr>
                 <th class="px-3 py-2 text-left font-medium">
-                  Provider
+                  上游服务
                 </th>
                 <th class="px-3 py-2 text-right font-medium">
                   请求
@@ -556,6 +647,9 @@
                 <th class="px-3 py-2 text-right font-medium">
                   成功率
                 </th>
+                <th class="px-3 py-2 text-right font-medium">
+                  错误率
+                </th>
                 <th class="px-3 py-2 text-right font-medium">
                   输出 TPS
                 </th>
@@ -566,10 +660,16 @@
                   平均响应
                 </th>
                 <th class="px-3 py-2 text-right font-medium">
-                  P90 响应 / 首字
+                  P90/P99 响应
                 </th>
                 <th class="px-3 py-2 text-right font-medium">
-                  样本
+                  P90/P99 首字
+                </th>
+                <th class="px-3 py-2 text-right font-medium">
+                  慢请求
+                </th>
+                <th class="px-3 py-2 text-right font-medium">
+                  样本覆盖
                 </th>
               </tr>
             </thead>
@@ -593,6 +693,9 @@
                 <td class="px-3 py-2 text-right">
                   {{ formatProviderPerformanceMetric(provider.success_rate, '%') }}
                 </td>
+                <td class="px-3 py-2 text-right">
+                  {{ formatErrorRate(provider.success_rate) }}
+                </td>
                 <td class="px-3 py-2 text-right">
                   {{ formatProviderPerformanceMetric(provider.avg_output_tps, ' tps') }}
                 </td>
@@ -605,11 +708,18 @@
                 <td class="px-3 py-2 text-right">
                   {{ formatProviderPerformanceMetric(provider.p90_response_time_ms, 'ms', 0) }}
                   /
+                  {{ formatProviderPerformanceMetric(provider.p99_response_time_ms, 'ms', 0) }}
+                </td>
+                <td class="px-3 py-2 text-right">
                   {{ formatProviderPerformanceMetric(provider.p90_first_byte_time_ms, 'ms', 0) }}
+                  /
+                  {{ formatProviderPerformanceMetric(provider.p99_first_byte_time_ms, 'ms', 0) }}
+                </td>
+                <td class="px-3 py-2 text-right">
+                  {{ formatMetricNumber(provider.slow_request_count) }}
                 </td>
                 <td class="px-3 py-2 text-right text-xs text-muted-foreground">
-                  {{ formatMetricNumber(provider.tps_sample_count) }} /
+                  {{ providerSampleCoverageText(provider) }}
-                  {{ formatMetricNumber(provider.first_byte_sample_count) }}
                 </td>
               </tr>
             </tbody>
@@ -620,7 +730,7 @@
           v-else
           class="rounded-lg border border-dashed border-border/70 px-3 py-4 text-sm text-muted-foreground"
         >
-          当前没有 Provider 性能数据。
+          当前没有上游服务性能数据。
         </div>
       </div>
     </Card>
@@ -695,41 +805,6 @@
       </Card>
     </div>
 
-    <Card class="space-y-3 p-4">
-      <h3 class="text-sm font-semibold">
-        提供商健康度
-      </h3>
-      <div
-        v-if="providerLoading"
-        class="p-4"
-      >
-        <LoadingState />
-      </div>
-      <div
-        v-else-if="providerStatus.length"
-        class="grid grid-cols-1 gap-3 text-sm sm:grid-cols-2 lg:grid-cols-3"
-      >
-        <div
-          v-for="provider in providerStatus"
-          :key="provider.name"
-          class="rounded-lg border p-3"
-        >
-          <div class="flex items-center justify-between">
-            <span class="font-medium">{{ provider.name }}</span>
-            <span class="text-xs text-muted-foreground">{{ provider.requests }} 请求</span>
-          </div>
-          <div class="mt-1 text-xs text-muted-foreground">
-            状态: {{ provider.status }}
-          </div>
-        </div>
-      </div>
-      <div
-        v-else
-        class="rounded-lg border border-dashed border-border/70 px-3 py-4 text-sm text-muted-foreground"
-      >
-        当前没有提供商状态数据。
-      </div>
-    </Card>
   </div>
 </template>
 
@@ -740,6 +815,7 @@ import {
   AlertTriangle,
   Cable,
   CheckCircle2,
+  FilterX,
   GitBranch,
   Gauge,
   ShieldCheck,
@@ -751,9 +827,9 @@ import {
   adminApi,
   type ErrorDistributionResponse,
   type PercentileItem,
+  type ProviderPerformanceItem,
   type ProviderPerformanceResponse,
 } from '@/api/admin'
-import { dashboardApi, type ProviderStatus } from '@/api/dashboard'
 import {
   monitoringApi,
   type AdminMonitoringCircuitHistoryItem,
@@ -765,12 +841,19 @@ import LineChart from '@/components/charts/LineChart.vue'
 import { LoadingState, TimeRangePicker } from '@/components/common'
 import { ErrorDistributionChart, PercentileChart } from '@/components/stats'
 import Badge from '@/components/ui/badge.vue'
+import Button from '@/components/ui/button.vue'
 import Card from '@/components/ui/card.vue'
+import Input from '@/components/ui/input.vue'
 import RefreshButton from '@/components/ui/refresh-button.vue'
+import Select from '@/components/ui/select.vue'
+import SelectContent from '@/components/ui/select-content.vue'
+import SelectItem from '@/components/ui/select-item.vue'
+import SelectTrigger from '@/components/ui/select-trigger.vue'
+import SelectValue from '@/components/ui/select-value.vue'
 import { useToast } from '@/composables/useToast'
 import { getDateRangeFromPeriod } from '@/features/usage/composables'
 import type { DateRangeParams } from '@/features/usage/types'
-import { formatDate, formatNumber, formatTokens } from '@/utils/format'
+import { formatDate, formatNumber } from '@/utils/format'
 import { log } from '@/utils/logger'
 import {
   buildProviderPerformanceChartData,
@@ -778,6 +861,10 @@ import {
 } from './performanceAnalysisHelpers'
 
 const LIVE_REFRESH_INTERVAL_MS = 10_000
+const DEFAULT_PROVIDER_PERFORMANCE_SLOW_THRESHOLD_MS = 10_000
+
+type ProviderPerformanceBooleanFilter = 'all' | 'true' | 'false'
+type ProviderPerformanceParams = NonNullable<Parameters<typeof adminApi.getProviderPerformance>[0]>
 
 const timeRange = ref<DateRangeParams>(getDateRangeFromPeriod('last30days'))
 const { error: showError } = useToast()
@@ -789,10 +876,15 @@ const errorDistribution = ref<ErrorDistributionResponse['distribution']>([])
 const errorTrend = ref<ErrorDistributionResponse['trend']>([])
 const errorLoading = ref(false)
 
-const providerStatus = ref<ProviderStatus[]>([])
-const providerLoading = ref(false)
 const providerPerformance = ref<ProviderPerformanceResponse | null>(null)
 const providerPerformanceLoading = ref(false)
+const providerPerformanceProviderId = ref('')
+const providerPerformanceModel = ref('')
+const providerPerformanceApiFormat = ref('')
+const providerPerformanceEndpointKind = ref('')
+const providerPerformanceIsStream = ref<ProviderPerformanceBooleanFilter>('all')
+const providerPerformanceHasFormatConversion = ref<ProviderPerformanceBooleanFilter>('all')
+const providerPerformanceSlowThresholdMs = ref(String(DEFAULT_PROVIDER_PERFORMANCE_SLOW_THRESHOLD_MS))
 
 const systemStatus = ref<AdminMonitoringSystemStatus | null>(null)
 const resilienceStatus = ref<AdminMonitoringResilienceStatus | null>(null)
@@ -806,12 +898,12 @@ const liveLastUpdatedAt = ref<string | null>(null)
 
 let percentilesRequestId = 0
 let errorsRequestId = 0
-let providersRequestId = 0
 let providerPerformanceRequestId = 0
 let liveRequestId = 0
 let loadAllPromise: Promise<void> | null = null
 let hasPendingLoadAll = false
 let loadAllDebounceTimer: ReturnType<typeof setTimeout> | null = null
+let providerPerformanceDebounceTimer: ReturnType<typeof setTimeout> | null = null
 let liveRefreshTimer: ReturnType<typeof setInterval> | null = null
 
 function buildTimeRangeParams() {
@@ -824,6 +916,117 @@ function buildTimeRangeParams() {
   }
 }
 
+function normalizeProviderPerformanceFilter(value: string): string | undefined {
+  const trimmed = value.trim()
+  return trimmed ? trimmed : undefined
+}
+
+function parseProviderPerformanceBooleanFilter(
+  value: ProviderPerformanceBooleanFilter
+): boolean | undefined {
+  if (value === 'true') return true
+  if (value === 'false') return false
+  return undefined
+}
+
+function clampProviderPerformanceSlowThreshold(value: string): number {
+  const parsed = Number(value)
+  if (!Number.isFinite(parsed)) {
+    return DEFAULT_PROVIDER_PERFORMANCE_SLOW_THRESHOLD_MS
+  }
+  return Math.min(600_000, Math.max(1, Math.round(parsed)))
+}
+
+function resolveProviderPerformanceGranularity(): 'day' | 'hour' {
+  if (timeRange.value.preset === 'today' || timeRange.value.preset === 'yesterday') {
+    return 'hour'
+  }
+
+  if (!timeRange.value.preset && timeRange.value.start_date && timeRange.value.end_date) {
+    return timeRange.value.start_date === timeRange.value.end_date ? 'hour' : 'day'
+  }
+
+  return 'day'
+}
+
+const providerPerformanceSlowThresholdValue = computed(() => (
+  clampProviderPerformanceSlowThreshold(providerPerformanceSlowThresholdMs.value)
+))
+
+const providerPerformanceSlowThresholdLabel = computed(() => {
+  const value = providerPerformanceSlowThresholdValue.value
+  if (value >= 1000) {
+    const seconds = value / 1000
+    return `${Number.isInteger(seconds) ? seconds.toFixed(0) : seconds.toFixed(1)}s`
+  }
+  return `${value}ms`
+})
+
+const providerPerformanceActiveFilterCount = computed(() => {
+  const textFilterCount = [
+    providerPerformanceProviderId.value,
+    providerPerformanceModel.value,
+    providerPerformanceApiFormat.value,
+    providerPerformanceEndpointKind.value,
+  ].filter(value => normalizeProviderPerformanceFilter(value)).length
+
+  const booleanFilterCount = [
+    providerPerformanceIsStream.value,
+    providerPerformanceHasFormatConversion.value,
+  ].filter(value => value !== 'all').length
+
+  const thresholdFilterCount = providerPerformanceSlowThresholdValue.value === DEFAULT_PROVIDER_PERFORMANCE_SLOW_THRESHOLD_MS
+    ? 0
+    : 1
+
+  return textFilterCount + booleanFilterCount + thresholdFilterCount
+})
+
+const hasProviderPerformanceFilters = computed(() => providerPerformanceActiveFilterCount.value > 0)
+
+function buildProviderPerformanceParams(): ProviderPerformanceParams {
+  const params: ProviderPerformanceParams = {
+    ...buildTimeRangeParams(),
+    granularity: resolveProviderPerformanceGranularity(),
+    limit: 8,
+    slow_threshold_ms: providerPerformanceSlowThresholdValue.value,
+  }
+
+  const providerId = normalizeProviderPerformanceFilter(providerPerformanceProviderId.value)
+  if (providerId) params.provider_id = providerId
+
+  const model = normalizeProviderPerformanceFilter(providerPerformanceModel.value)
+  if (model) params.model = model
+
+  const apiFormat = normalizeProviderPerformanceFilter(providerPerformanceApiFormat.value)
+  if (apiFormat) params.api_format = apiFormat
+
+  const endpointKind = normalizeProviderPerformanceFilter(providerPerformanceEndpointKind.value)
+  if (endpointKind) params.endpoint_kind = endpointKind
+
+  const isStream = parseProviderPerformanceBooleanFilter(providerPerformanceIsStream.value)
+  if (isStream !== undefined) params.is_stream = isStream
+
+  const hasFormatConversion = parseProviderPerformanceBooleanFilter(
+    providerPerformanceHasFormatConversion.value
+  )
+  if (hasFormatConversion !== undefined) {
+    params.has_format_conversion = hasFormatConversion
+  }
+
+  return params
+}
+
+function resetProviderPerformanceFilters() {
+  providerPerformanceProviderId.value = ''
+  providerPerformanceModel.value = ''
+  providerPerformanceApiFormat.value = ''
+  providerPerformanceEndpointKind.value = ''
+  providerPerformanceIsStream.value = 'all'
+  providerPerformanceHasFormatConversion.value = 'all'
+  providerPerformanceSlowThresholdMs.value = String(DEFAULT_PROVIDER_PERFORMANCE_SLOW_THRESHOLD_MS)
+}
+
 function formatMetricNumber(value: number | null | undefined): string {
   if (value == null || Number.isNaN(value)) {
     return '-'
@@ -836,6 +1039,24 @@ function formatMetricNumber(value: number | null | undefined): string {
   return formatNumber(value)
 }
 
+function formatErrorRate(successRate: number | null | undefined): string {
+  if (successRate == null || Number.isNaN(successRate)) {
+    return '-'
+  }
+  return `${Math.max(0, 100 - successRate).toFixed(2)}%`
+}
+
+function providerSampleCoverageText(provider: ProviderPerformanceItem): string {
+  if (!provider.request_count) {
+    return '-'
+  }
+  const responseSamples = provider.response_time_sample_count ?? 0
+  const firstByteSamples = provider.first_byte_sample_count ?? 0
+  const responseCoverage = Math.round(responseSamples / provider.request_count * 100)
+  const firstByteCoverage = Math.round(firstByteSamples / provider.request_count * 100)
+  return `${responseCoverage}% / ${firstByteCoverage}%`
+}
+
 async function loadPercentiles() {
   const requestId = ++percentilesRequestId
   percentileLoading.value = true
@@ -874,39 +1095,17 @@ async function loadErrors() {
   }
 }
 
-async function loadProviders() {
-  const requestId = ++providersRequestId
-  providerLoading.value = true
-  try {
-    const data = await dashboardApi.getProviderStatus()
-    if (requestId !== providersRequestId) return
-    providerStatus.value = data
-  } catch (error) {
-    if (requestId !== providersRequestId) return
-    providerStatus.value = []
-    log.error('加载提供商状态失败', error)
-  } finally {
-    if (requestId === providersRequestId) {
-      providerLoading.value = false
-    }
-  }
-}
-
 async function loadProviderPerformance() {
   const requestId = ++providerPerformanceRequestId
   providerPerformanceLoading.value = true
   try {
-    const data = await adminApi.getProviderPerformance({
+    const data = await adminApi.getProviderPerformance(buildProviderPerformanceParams())
-      ...buildTimeRangeParams(),
-      granularity: 'day',
-      limit: 8,
-    })
     if (requestId !== providerPerformanceRequestId) return
     providerPerformance.value = data
   } catch (error) {
     if (requestId !== providerPerformanceRequestId) return
     providerPerformance.value = null
-    log.error('加载 Provider 性能统计失败', error)
+    log.error('加载上游服务性能统计失败', error)
   } finally {
     if (requestId === providerPerformanceRequestId) {
       providerPerformanceLoading.value = false
@@ -1034,7 +1233,7 @@ const healthStatusText = computed(() => {
 })
 
 const metricsAvailabilityText = computed(() => (
-  gatewayMetrics.value ? 'Prometheus 在线' : 'Prometheus 暂不可达'
+  gatewayMetrics.value ? '网关指标在线' : '网关指标暂不可达'
 ))
 
 const distributedGateVariant = computed<'warning' | 'outline'>(() => (
@@ -1061,11 +1260,26 @@ const currentTunnelNodes = computed(() => (
   gatewayMetrics.value?.tunnel.nodes ?? systemStatus.value?.tunnel.nodes ?? null
 ))
 
-const providerAndKeySummary = computed(() => {
+const tunnelQueueRejectedTotal = computed(() => {
-  if (!systemStatus.value) {
+  const full = gatewayMetrics.value?.tunnel.outboundQueueRejectedFullTotal ?? 0
+  const closed = gatewayMetrics.value?.tunnel.outboundQueueRejectedClosedTotal ?? 0
+  return full + closed
+})
+
+const tunnelSelectionPressureTotal = computed(() => {
+  const congested = gatewayMetrics.value?.tunnel.proxyConnectionCongestedTotal ?? 0
+  const retry = gatewayMetrics.value?.tunnel.selectionRetryTotal ?? 0
+  const unavailable = gatewayMetrics.value?.tunnel.selectionUnavailableTotal ?? 0
+  return congested + retry + unavailable
+})
+
+const tunnelQueueUtilizationText = computed(() => {
+  const depth = gatewayMetrics.value?.tunnel.outboundQueueDepthTotal
+  const capacity = gatewayMetrics.value?.tunnel.outboundQueueCapacityTotal
+  if (depth == null || capacity == null || capacity <= 0) {
     return '-'
   }
-  return `${systemStatus.value.providers.active}/${systemStatus.value.providers.total} · ${systemStatus.value.api_keys.active}/${systemStatus.value.api_keys.total}`
+  return `${Math.round(depth / capacity * 100)}%`
 })
 
 const fallbackRows = computed(() => {
@@ -1085,39 +1299,69 @@ const providerPerformanceRows = computed(() => providerPerformance.value?.provid
 
 const providerPerformanceSubtitle = computed(() => {
   const requests = providerPerformance.value?.summary.request_count ?? 0
-  return `完成窗口内 ${formatMetricNumber(requests)} 个 Provider 请求样本`
+  const filters = providerPerformanceActiveFilterCount.value
+  const filterText = filters > 0 ? ` · ${filters} 个筛选条件` : ''
+  return `完成窗口内 ${formatMetricNumber(requests)} 个上游请求样本${filterText}`
 })
 
 const providerPerformanceSummaryCards = computed(() => {
   const summary = providerPerformance.value?.summary
   return [
+    {
+      title: '请求样本',
+      value: formatMetricNumber(summary?.request_count),
+      hint: `${formatMetricNumber(providerPerformanceRows.value.length)} 个上游服务`,
+      icon: Activity,
+      iconClass: 'text-blue-500',
+    },
+    {
+      title: '成功率',
+      value: formatProviderPerformanceMetric(summary?.success_rate, '%'),
+      hint: `错误率 ${formatErrorRate(summary?.success_rate)}`,
+      icon: CheckCircle2,
+      iconClass: 'text-emerald-500',
+    },
+    {
+      title: 'P99 响应',
+      value: formatProviderPerformanceMetric(summary?.p99_response_time_ms, 'ms', 0),
+      hint: `P90 ${formatProviderPerformanceMetric(summary?.p90_response_time_ms, 'ms', 0)}`,
+      icon: Gauge,
+      iconClass: 'text-violet-500',
+    },
+    {
+      title: 'P99 首字',
+      value: formatProviderPerformanceMetric(summary?.p99_first_byte_time_ms, 'ms', 0),
+      hint: `P90 ${formatProviderPerformanceMetric(summary?.p90_first_byte_time_ms, 'ms', 0)}`,
+      icon: Timer,
+      iconClass: 'text-sky-500',
+    },
     {
       title: '输出 TPS',
       value: formatProviderPerformanceMetric(summary?.avg_output_tps, ' tps'),
-      hint: `请求 ${formatMetricNumber(summary?.request_count)}`,
+      hint: `TPS 样本 ${formatMetricNumber(summary?.tps_sample_count)}`,
       icon: Zap,
       iconClass: 'text-amber-500',
     },
     {
       title: '平均首字',
       value: formatProviderPerformanceMetric(summary?.avg_first_byte_time_ms, 'ms'),
-      hint: '成功请求首字样本',
+      hint: `首字样本 ${formatMetricNumber(summary?.first_byte_sample_count)}`,
       icon: Timer,
       iconClass: 'text-sky-500',
     },
     {
       title: '平均响应',
       value: formatProviderPerformanceMetric(summary?.avg_response_time_ms, 'ms'),
-      hint: '成功请求响应耗时',
+      hint: `响应样本 ${formatMetricNumber(summary?.response_time_sample_count)}`,
       icon: Gauge,
       iconClass: 'text-violet-500',
     },
     {
-      title: '成功率',
+      title: '慢请求',
-      value: formatProviderPerformanceMetric(summary?.success_rate, '%'),
+      value: formatMetricNumber(summary?.slow_request_count),
-      hint: `${formatMetricNumber(providerPerformanceRows.value.length)} 个 Provider`,
+      hint: `响应耗时 >= ${providerPerformanceSlowThresholdLabel.value}`,
-      icon: CheckCircle2,
+      icon: AlertTriangle,
-      iconClass: 'text-emerald-500',
+      iconClass: 'text-yellow-500',
     },
   ]
 })
@@ -1162,7 +1406,7 @@ const liveSummaryCards = computed(() => [
   {
     title: '系统健康',
     value: resilienceStatus.value ? `${resilienceStatus.value.health_score}/100` : '-',
-    hint: `${healthStatusText.value} · 开路 ${formatMetricNumber(resilienceStatus.value?.error_statistics.open_circuit_breakers)}`,
+    hint: `${healthStatusText.value} · 熔断打开 ${formatMetricNumber(resilienceStatus.value?.error_statistics.open_circuit_breakers)}`,
     icon: ShieldCheck,
     iconClass: 'text-emerald-500',
   },
@@ -1174,34 +1418,34 @@ const liveSummaryCards = computed(() => [
     iconClass: 'text-yellow-500',
   },
   {
-    title: '当前活跃流',
+    title: '代理活跃流',
     value: formatMetricNumber(currentActiveStreams.value),
     hint: `代理连接 ${formatMetricNumber(currentProxyConnections.value)}`,
     icon: Cable,
     iconClass: 'text-sky-500',
   },
   {
-    title: '本地 In Flight',
+    title: '当前节点处理中',
     value: formatMetricNumber(gatewayMetrics.value?.local.inFlight),
-    hint: `剩余 permit ${formatMetricNumber(gatewayMetrics.value?.local.availablePermits)}`,
+    hint: `可接入 ${formatMetricNumber(gatewayMetrics.value?.local.availablePermits)}`,
     icon: Activity,
     iconClass: 'text-blue-500',
   },
   {
-    title: '分布式 In Flight',
+    title: '全局处理中',
     value: gatewayMetrics.value?.distributed.unavailable
       ? '不可用'
       : formatMetricNumber(gatewayMetrics.value?.distributed.inFlight),
     hint: gatewayMetrics.value?.distributed.unavailable
-      ? '检查 Redis gate 状态'
+      ? '检查 Redis 连接'
-      : `剩余 permit ${formatMetricNumber(gatewayMetrics.value?.distributed.availablePermits)}`,
+      : `可接入 ${formatMetricNumber(gatewayMetrics.value?.distributed.availablePermits)}`,
     icon: Workflow,
     iconClass: 'text-violet-500',
   },
   {
-    title: 'Fallback 累计',
+    title: '降级切换',
     value: formatMetricNumber(gatewayMetrics.value?.fallbackTotal),
-    hint: `今日请求 ${formatMetricNumber(systemStatus.value?.today_stats.requests)}`,
+    hint: '当前进程累计',
     icon: GitBranch,
     iconClass: 'text-rose-500',
   },
@@ -1212,7 +1456,6 @@ const isRefreshing = computed(() => (
   liveRefreshing.value ||
   percentileLoading.value ||
   errorLoading.value ||
-  providerLoading.value ||
   providerPerformanceLoading.value
 ))
 
@@ -1225,7 +1468,6 @@ async function loadAll() {
   loadAllPromise = Promise.all([
     loadPercentiles(),
     loadErrors(),
-    loadProviders(),
     loadProviderPerformance(),
   ])
     .then(() => undefined)
@@ -1255,7 +1497,30 @@ function scheduleLoadAll() {
   }, 120)
 }
 
+function scheduleProviderPerformanceLoad() {
+  if (providerPerformanceDebounceTimer) {
+    clearTimeout(providerPerformanceDebounceTimer)
+  }
+
+  providerPerformanceDebounceTimer = setTimeout(() => {
+    providerPerformanceDebounceTimer = null
+    void loadProviderPerformance()
+  }, 180)
+}
+
 watch(timeRange, scheduleLoadAll, { deep: true })
+watch(
+  [
+    providerPerformanceProviderId,
+    providerPerformanceModel,
+    providerPerformanceApiFormat,
+    providerPerformanceEndpointKind,
+    providerPerformanceIsStream,
+    providerPerformanceHasFormatConversion,
+    providerPerformanceSlowThresholdMs,
+  ],
+  scheduleProviderPerformanceLoad
+)
 
 onMounted(() => {
   void loadLiveData()
@@ -1271,6 +1536,11 @@ onUnmounted(() => {
     loadAllDebounceTimer = null
   }
 
+  if (providerPerformanceDebounceTimer) {
+    clearTimeout(providerPerformanceDebounceTimer)
+    providerPerformanceDebounceTimer = null
+  }
+
   if (liveRefreshTimer) {
     clearInterval(liveRefreshTimer)
     liveRefreshTimer = null
@@ -1280,7 +1550,6 @@ onUnmounted(() => {
   loadAllPromise = null
   percentilesRequestId += 1
   errorsRequestId += 1
-  providersRequestId += 1
   providerPerformanceRequestId += 1
   liveRequestId += 1
 })