diff --git a/session/session.go b/session/session.go
index f9bc1dc..e6f6a99 100644
--- a/session/session.go
+++ b/session/session.go
@@ -35,6 +35,7 @@ type Session struct {
 	prevPrompt []byte // last submitted prompt; overwritten on each new submission
 
 	// Cached model list (expensive API call; refreshed every 24h).
+	modelsMu      sync.Mutex
 	modelsCache   string
 	modelsCacheAt time.Time
 }
@@ -49,30 +50,32 @@ const modelsCacheTTL = 24 * time.Hour
 
 // CachedListModels returns the model list, using a 24h cache to avoid
 // repeated expensive API calls.
+// NOTE: Must NOT be called while holding sess.mu (caller content() holds RLock).
+// Uses its own modelsMu to avoid deadlock.
 func (sess *Session) CachedListModels() string {
-	sess.mu.RLock()
+	sess.modelsMu.Lock()
 	if sess.modelsCache != "" && time.Since(sess.modelsCacheAt) < modelsCacheTTL {
 		result := sess.modelsCache
-		sess.mu.RUnlock()
+		sess.modelsMu.Unlock()
 		return result
 	}
-	sess.mu.RUnlock()
+	sess.modelsMu.Unlock()
 
 	// Cache miss — fetch and store.
 	result := sess.Core.ListModels()
-	sess.mu.Lock()
+	sess.modelsMu.Lock()
 	sess.modelsCache = result
 	sess.modelsCacheAt = time.Now()
-	sess.mu.Unlock()
+	sess.modelsMu.Unlock()
 	return result
 }
 
 // InvalidateModelsCache clears the cached model list.
 func (sess *Session) InvalidateModelsCache() {
-	sess.mu.Lock()
+	sess.modelsMu.Lock()
 	sess.modelsCache = ""
 	sess.modelsCacheAt = time.Time{}
-	sess.mu.Unlock()
+	sess.modelsMu.Unlock()
 }
 
 func (sess *Session) RunnableID() string { return sess.id }