maomaodev commented on issue #7529:
URL: https://github.com/apache/kyuubi/issues/7529#issuecomment-5311122818

   Hi @wangzhigang1999 , Thanks for your great work on Data Agent!
   
   While trying out Data Agent, I ran into a few small things — some of them 
may not even be real issues. I'd like to hear your take on them.
   
   1. Clicking **Stop** in the Web UI only affects the front-end (SSE 
disconnects, the button resets); the cancel signal never reaches the Data Agent 
engine, so the backend keeps running at every blocking point, for example:
      * The pending human-approval `future.get(300s)` is not woken up.
      * The in-flight LLM streaming call is not aborted.
   
      **Steps to reproduce:**
      * Start `kyuubi-data-agent-engine` and ask the agent to run a destructive 
statement, e.g. `INSERT INTO t VALUES (...)`.
      * When the approval card appears, click **Stop** in the Web UI.
      * Then send an `Approve` for the same `requestId`.
   
      **Expected:** Stop cancels the pending approval; the late Approve is a 
no-op; no SQL is executed.
   
      **Actual:** `resolve(reqId, true)` returns `true`, the blocked agent 
thread receives `approved=true`, and the destructive SQL is submitted to JDBC.
   
   2. `CompactionMiddleware` does not include the summarization call's tokens 
in the accumulated total. This can cause token accounting to diverge from the 
LLM provider's actual billing. Should the compaction tokens be counted into the 
total?
   
   3. When `ToolResultOffloadMiddleware` kicks in and spills data to disk, 
there is no truncation or disk-space check. The prompt does tell the LLM to add 
its own `LIMIT`, but that constraint is unreliable. Should we add a disk-side 
guardrail?
   
   4. In the agent reply view, the Web UI exposes two interactive controls at 
once: a standalone **Stop** button at the top-right of the chat header, and the 
(greyed-out) send button in the input bar. The user's focus has to jump between 
the header and the input bar, which doesn't match the interaction pattern of 
mainstream LLM chat UIs. Should these two controls be merged?
   


-- 
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

To unsubscribe, e-mail: [email protected]

For queries about this service, please contact Infrastructure at:
[email protected]


---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to