feat(code): implement proper cancellation via CancellationToken

Replace AbortHandle with CancellationToken in SessionManager so that
cancel_operation() cooperatively cancels the streaming LLM loop via
tokio::select! instead of hard-aborting the task.

- ongoing_operations now stores CancellationToken instead of AbortHandle
- cancel_operation() calls token.cancel() for cooperative cancellation
- execute_streaming() returns the CancellationToken alongside the handle
- The streaming loop in call_llm uses tokio::select! to exit immediately
  when the token is cancelled

Author: RoyLin

core/src/agent_api.rs

@@ -1569,7 +1569,7 @@ impl AgentSession {
 
         let agent_loop = self.build_agent_loop();
         let result = agent_loop
-            .execute_from_messages(effective_history, None, None)
+            .execute_from_messages(effective_history, None, None, None)
             .await?;
 
         if use_internal {
@@ -1604,7 +1604,7 @@ impl AgentSession {
         let agent_loop = self.build_agent_loop();
         let handle = tokio::spawn(async move {
             let _ = agent_loop
-                .execute_from_messages(effective_history, None, Some(tx))
+                .execute_from_messages(effective_history, None, Some(tx), None)
                 .await;
         });
 

core/src/session/manager.rs

@@ -721,7 +721,7 @@ impl SessionManager {
 
         // Execute with session context
         let result = agent
-            .execute_with_session(&history, &effective_prompt, Some(session_id), None)
+            .execute_with_session(&history, &effective_prompt, Some(session_id), None, None)
             .await?;
 
         // Update session