feat: inject user comments into agent conversation in real-time (try_recv between LLM calls)

src/agent.rs

@@ -683,6 +683,15 @@ pub async fn run_step_loop(
     let step_order = step.order;
 
     for iteration in 0..50 {
+        // Drain any pending user comments and inject as feedback
+        while let Ok(event) = event_rx.try_recv() {
+            if let AgentEvent::Comment { content, .. } = event {
+                tracing::info!("[workflow {}] Step {} received user feedback: {}", workflow_id, step_order, &content[..content.len().min(100)]);
+                step_chat_history.push(ChatMessage::user(&format!("【用户反馈】{}", content)));
+                send_execution(update_tx, workflow_id, step_order, "user_feedback", &content, "已注入到对话上下文", "done").await;
+            }
+        }
+
         // Build messages: system + user context + chat history
         let mut messages = vec![
             ChatMessage::system(&system_prompt),
@@ -1087,6 +1096,14 @@ pub async fn run_agent_loop(
             break;
         }
 
+        // Drain pending user comments
+        while let Ok(event) = event_rx.try_recv() {
+            if let AgentEvent::Comment { content, .. } = event {
+                tracing::info!("[workflow {}] Planning received user feedback: {}", workflow_id, &content[..content.len().min(100)]);
+                state.current_step_chat_history.push(ChatMessage::user(&format!("【用户反馈】{}", content)));
+            }
+        }
+
         let system_prompt = build_planning_prompt(project_id, instructions);
         let messages = state.build_messages(&system_prompt, requirement);
         let msg_count = messages.len() as i32;