Changeset 0.28.8 (#194)

server/providers/provider.go

@@ -7,6 +7,7 @@ import (
 	"log"
 	"net/http"
 	"net/url"
+	"sync"
 	"time"
 
 	"git.gobha.me/xcaliber/chat-switchboard/models"
@@ -51,12 +52,30 @@ type ProviderConfig struct {
 	ProxyURL      string                 // Only used when ProxyMode == "custom"
 }
 
+// ── HTTP Client Pool ───────────────────────
+//
+// Provider HTTP clients are pooled by (ProxyMode, ProxyURL) tuple.
+// Previously Client() created a new http.Transport per call, leaking
+// idle connections and preventing TCP reuse across requests to the same
+// provider. With pooling, a Venice model sync followed by a completion
+// reuses the same TCP connection to api.venice.ai.
+
+// clientPool holds shared *http.Client instances keyed by proxy config.
+var clientPool sync.Map // map[string]*http.Client
+
 // Client returns an *http.Client configured for this provider's proxy settings.
+// Clients are cached and reused across calls with the same proxy configuration.
 // system = http.ProxyFromEnvironment (Go default), direct = no proxy,
 // custom = explicit proxy URL (http/https/socks5).
 func (c ProviderConfig) Client() *http.Client {
+	key := c.ProxyMode + "|" + c.ProxyURL
+	if v, ok := clientPool.Load(key); ok {
+		return v.(*http.Client)
+	}
+
 	transport := &http.Transport{
 		MaxIdleConns:        100,
+		MaxIdleConnsPerHost: 10,
 		IdleConnTimeout:     90 * time.Second,
 		TLSHandshakeTimeout: 10 * time.Second,
 	}
@@ -81,7 +100,10 @@ func (c ProviderConfig) Client() *http.Client {
 	default: // "system" or empty
 		transport.Proxy = http.ProxyFromEnvironment
 	}
-	return &http.Client{Transport: transport}
+
+	client := &http.Client{Transport: transport}
+	actual, _ := clientPool.LoadOrStore(key, client)
+	return actual.(*http.Client)
 }
 
 // ── Request / Response Types ────────────────