fix(claude): guard thinking tool-call fallback when final text exists

- only parse tool_calls from thinking when finalText is empty

- apply the same guard in stream runtime finalizer

- add regression tests for non-stream and stream paths

internal/format/claude/render.go

@@ -9,7 +9,7 @@ import (
 
 func BuildMessageResponse(messageID, model string, normalizedMessages []any, finalThinking, finalText string, toolNames []string) map[string]any {
 	detected := util.ParseToolCalls(finalText, toolNames)
-	if len(detected) == 0 && finalThinking != "" {
+	if len(detected) == 0 && finalText == "" && finalThinking != "" {
 		detected = util.ParseToolCalls(finalThinking, toolNames)
 	}
 	content := make([]map[string]any, 0, 4)

internal/format/claude/render_test.go

@@ -27,3 +27,36 @@ func TestBuildMessageResponseDetectsToolCallsFromThinkingFallback(t *testing.T)
 		t.Fatalf("expected tool name search, got=%#v", last["name"])
 	}
 }
+
+func TestBuildMessageResponseSkipsThinkingFallbackWhenFinalTextExists(t *testing.T) {
+	resp := BuildMessageResponse(
+		"msg_1",
+		"claude-sonnet-4-5",
+		[]any{map[string]any{"role": "user", "content": "hi"}},
+		`{"tool_calls":[{"name":"search","input":{"q":"go"}}]}`,
+		"normal answer",
+		[]string{"search"},
+	)
+
+	if resp["stop_reason"] != "end_turn" {
+		t.Fatalf("expected stop_reason=end_turn, got=%#v", resp["stop_reason"])
+	}
+
+	content, _ := resp["content"].([]map[string]any)
+	foundText := false
+	foundTool := false
+	for _, block := range content {
+		if block["type"] == "text" && block["text"] == "normal answer" {
+			foundText = true
+		}
+		if block["type"] == "tool_use" {
+			foundTool = true
+		}
+	}
+	if !foundText {
+		t.Fatalf("expected text block with finalText, got=%#v", resp["content"])
+	}
+	if foundTool {
+		t.Fatalf("unexpected tool_use block when finalText exists, got=%#v", resp["content"])
+	}
+}