feat: always run classifier, default to local for simple tasks

Classifier (free local model) now runs on every message by default.
classifier_min_length=0 means always classify, <0 means disabled.
Short messages no longer skip to primary — they get classified too.
Images route to local model first if it supports vision.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: Alexey Panfilov

config/config.yaml

@@ -100,7 +100,7 @@ routing:
   multimodal: gemini-flash     # vision/audio fallback (used when primary lacks vision)
   compaction_model: deepseek
   classifier: ollama           # gemma3:1b rates complexity 1/2/3
-  classifier_min_length: 100
+  classifier_min_length: 0    # 0 = always classify; >0 = min chars; <0 = disabled
 
 tool_filter:
   top_k: 20

internal/llm/router.go

@@ -343,8 +343,12 @@ func (r *Router) pick(ctx context.Context, messages []Message) Provider {
 		}
 	}
 
-	// Multimodal routing — use dedicated model only if the active provider can't handle images
+	// Multimodal routing — prefer local/primary if they support vision
 	if multimodal {
+		if p := r.get(cfg.Local); p != nil && supportsVision(p) {
+			slog.Info("routing", "reason", "local+vision", "provider", p.Name())
+			return p
+		}
 		if p := r.get(cfg.Primary); p != nil && supportsVision(p) {
 			slog.Info("routing", "reason", "primary+vision", "provider", p.Name())
 			return p
@@ -355,25 +359,27 @@ func (r *Router) pick(ctx context.Context, messages []Message) Provider {
 		}
 	}
 
-	// Classifier-based three-level routing: 1=local, 2=primary, 3=reasoner
-	if cfg.ClassifierMinLen > 0 {
-		if text := lastUserText(messages); len([]rune(text)) >= cfg.ClassifierMinLen {
+	// Classifier-based three-level routing: 1=local, 2=primary, 3=reasoner.
+	// ClassifierMinLen > 0: only classify messages longer than threshold.
+	// ClassifierMinLen == 0: always classify (classifier is a free local model).
+	// ClassifierMinLen < 0: disabled entirely.
+	if cfg.ClassifierMinLen >= 0 && r.get(cfg.Classifier) != nil {
+		text := lastUserText(messages)
+		if cfg.ClassifierMinLen == 0 || len([]rune(text)) >= cfg.ClassifierMinLen {
 			level := r.classify(ctx, text)
 			switch level {
 			case 1:
 				if p := r.get(cfg.Local); p != nil {
 					slog.Info("routing", "reason", "classifier→local", "level", 1, "provider", p.Name())
 					return p
 				}
-				// local not configured — fall through to primary
 			case 3:
 				if p := r.get(cfg.Reasoner); p != nil {
 					slog.Info("routing", "reason", "classifier→reasoner", "level", 3, "provider", p.Name())
 					return p
 				}
-				// reasoner not configured — fall through to primary
 			}
-			// level 2 or fallback from 1/3 → use primary
+			// level 2 or fallback → primary
 		}
 	}
 

internal/llm/router_test.go

@@ -254,23 +254,49 @@ func TestRouter_ClassifierSkippedForShortMessages(t *testing.T) {
 	}
 }
 
-// TestRouter_ClassifierDisabled: minLen=0 → classifier никогда не вызывается.
+// TestRouter_ClassifierDisabled: minLen<0 → classifier никогда не вызывается.
 func TestRouter_ClassifierDisabled(t *testing.T) {
 	classifier := &mockProvider{name: "classifier"}
 	primary := &mockProvider{name: "primary", resp: Response{Content: "ok"}}
 
 	r := newTestRouter(RouterConfig{
 		Primary:          "primary",
 		Classifier:       "classifier",
-		ClassifierMinLen: 0, // отключён
+		ClassifierMinLen: -1, // отключён
 	}, map[string]Provider{
 		"primary":    primary,
 		"classifier": classifier,
 	})
 
 	_, _ = r.Chat(context.Background(), []Message{{Role: "user", Content: "long message that would normally trigger classifier"}}, "", nil)
 	if classifier.calls != 0 {
-		t.Errorf("classifier should be disabled when minLen=0, got %d calls", classifier.calls)
+		t.Errorf("classifier should be disabled when minLen<0, got %d calls", classifier.calls)
+	}
+}
+
+// TestRouter_ClassifierAlwaysWhenZero: minLen=0 → classifier вызывается для любого сообщения.
+func TestRouter_ClassifierAlwaysWhenZero(t *testing.T) {
+	classifier := &mockProvider{name: "classifier", resp: Response{Content: "1"}}
+	local := &mockProvider{name: "local", resp: Response{Content: "local"}}
+	primary := &mockProvider{name: "primary", resp: Response{Content: "cloud"}}
+
+	r := newTestRouter(RouterConfig{
+		Local:            "local",
+		Primary:          "primary",
+		Classifier:       "classifier",
+		ClassifierMinLen: 0, // всегда классифицировать
+	}, map[string]Provider{
+		"local":      local,
+		"primary":    primary,
+		"classifier": classifier,
+	})
+
+	resp, _ := r.Chat(context.Background(), []Message{{Role: "user", Content: "hi"}}, "", nil)
+	if classifier.calls != 1 {
+		t.Errorf("classifier should be called for any message when minLen=0, got %d calls", classifier.calls)
+	}
+	if resp.Content != "local" {
+		t.Errorf("expected local, got %q", resp.Content)
 	}
 }