fix(agent): empty multimodal content becomes None

src/crates/assembly/core/src/agentic/core/message.rs

@@ -316,9 +316,19 @@ impl From<Message> for AIMessage {
                     content.push(']');
                 }
 
+                let content = if content.trim().is_empty() {
+                    // Empty multimodal content (no text and no images) must not reach
+                    // the provider as a bare empty string. Return None so downstream
+                    // converters apply their empty-content handling instead of billing
+                    // an empty request round.
+                    None
+                } else {
+                    Some(content)
+                };
+
                 Self {
                     role: "user".to_string(),
-                    content: Some(content),
+                    content,
                     reasoning_content: None,
                     thinking_signature: None,
                     tool_calls: None,
@@ -776,6 +786,53 @@ mod tests {
     use bitfun_agent_stream::ToolArgumentRepairKind;
     use bitfun_core_types::{ModelResponseReplay, ModelResponseReplayItem};
     use serde_json::json;
+    use crate::agentic::image_analysis::ImageContextData;
+
+    #[test]
+    fn empty_text_empty_images_multimodal_becomes_none() {
+        let ai_msg = AIMessage::from(Message::user_multimodal(String::new(), vec![]));
+
+        // An empty multimodal message (no text and no images) must not be sent to
+        // the provider as a bare empty string; it should become None so downstream
+        // converters apply their empty-content handling.
+        assert_eq!(ai_msg.content.as_deref(), None);
+    }
+
+    #[test]
+    fn non_empty_text_multimodal_preserves_content() {
+        let ai_msg = AIMessage::from(Message::user_multimodal("hi".to_string(), vec![]));
+
+        assert_eq!(ai_msg.content.as_deref(), Some("hi"));
+    }
+
+    #[test]
+    fn multimodal_with_images_kept_when_text_empty() {
+        let image = ImageContextData {
+            id: "img_1".to_string(),
+            image_path: Some("/tmp/photo.png".to_string()),
+            data_url: None,
+            mime_type: "image/png".to_string(),
+            metadata: Some(json!({ "name": "photo" })),
+        };
+
+        let ai_msg = AIMessage::from(Message::user_multimodal(String::new(), vec![image]));
+
+        // Image-only multimodal content (no text) is valid and must be preserved,
+        // not treated as an empty content value.
+        let content = ai_msg.content.expect("image-attached content should be present");
+        assert!(content.contains("[Attached image(s):"));
+        assert!(content.contains("photo"));
+    }
+
+    #[test]
+    fn user_multimodal_constructor_feeds_conversion() {
+        let ai_msg = AIMessage::from(Message::user_multimodal(String::new(), vec![]));
+
+        // Empty multimodal content built through the constructor still flows
+        // through the conversion arm and becomes None for the provider.
+        assert_eq!(ai_msg.role, "user");
+        assert_eq!(ai_msg.content.as_deref(), None);
+    }
 
     #[test]
     fn preserves_empty_reasoning_content_for_provider_replay() {