feat(auto): AUTO-only priority chain with tiered slots + live source probing; CSV export w/ key names; audit persistence; fix prompt token accounting & deepseek thinking

internal/gateway/server.go

@@ -14,6 +14,8 @@ import (
 	"net/http"
 	"net/url"
 	"strings"
+	"sync"
+	"time"
 
 	"llmsproxy/internal/config"
 	"llmsproxy/internal/core"
@@ -24,9 +26,11 @@ var uiFS embed.FS
 
 // Gateway is the HTTP handler for the OpenAI-compatible endpoint + web UI.
 type Gateway struct {
-	core  *core.Core
-	ui    http.Handler
-	stats *Stats
+	core       *core.Core
+	ui         http.Handler
+	stats      *Stats
+	probeMu    sync.Mutex
+	lastProbe  time.Time
 }
 
 func New(c *core.Core, gatewayKeys []string) (*Gateway, error) {
@@ -34,10 +38,14 @@ func New(c *core.Core, gatewayKeys []string) (*Gateway, error) {
 	if err != nil {
 		return nil, err
 	}
+	st := NewStats(3000)
+	if cfg := c.Config(); cfg != nil && cfg.RuntimeFile != "" {
+		st.LoadAudit(cfg.RuntimeFile + ".audit.jsonl")
+	}
 	return &Gateway{
 		core:  c,
 		ui:    http.FileServer(http.FS(sub)),
-		stats: NewStats(3000),
+		stats: st,
 	}, nil
 }
 
@@ -313,7 +321,21 @@ func (g *Gateway) handleModels(w http.ResponseWriter, r *http.Request) {
 	})
 }
 
+// ensureProbe triggers a live source probe at most once every 30s.
+func (g *Gateway) ensureProbe(ctx context.Context) {
+	g.probeMu.Lock()
+	due := time.Since(g.lastProbe) > 30*time.Second
+	if due {
+		g.lastProbe = time.Now()
+	}
+	g.probeMu.Unlock()
+	if due {
+		g.core.Registry().ProbeAll(ctx)
+	}
+}
+
 func (g *Gateway) handleStatusAPI(w http.ResponseWriter, r *http.Request) {
+	g.ensureProbe(r.Context())
 	host := r.Host
 	if host == "" {
 		host = g.core.Listen()