[Bug]: Cursor drops the request randomly with no response or error

[invarrow]

Before submitting

• I searched existing issues and did not find a duplicate.
• I included enough detail to reproduce or investigate the problem.

Area

apps/desktop

Steps to reproduce

1. Open app
2. connect cursor
3. ask a follow up question

Expected behavior

The thread should resume with the response to the query

Actual behavior

No response is shown for the query, initially when the query is sent the agent works and makes tool calls but after few seconds it drops the request and shows no response

Impact

Blocks work completely

Version or commit

No response

Environment

Ubuntu 24.04.2 LTS

Logs or stack traces

Screenshots, recordings, or supporting files

Screenshot from 2026-07-08 11-40-26.png

Workaround

No response

[hicder]

happens for me as well. I think it just started happening within the past few days, so i'm guessing it has something to do with new cursor cli?

[hicder]

yeah reverting to version 2026.07.01 seems to work consistently.

update: remove ~/.t3 and i haven't seen the problem happening even on latest cursor CLI ✊🪵

[hicder]

ok, AI helped me debug. Here is the bug:

 What happened at 03:19

  1. Session was cold-restarted (initialize → authenticate → session/load), then your prompt ran.
  2. Cursor answered successfully in ~6s:

  │ “Hello. What would you like to work on?”
  3. session/prompt and sendTurn both succeeded. No provider error.

  Why the UI looked empty

  ACP item IDs are assistant:${sessionId}:segment:${N}. On restart, the segment counter resets to 0, so the new reply reused:

  assistant:assistant:0c77ff2c-…:segment:0

  That ID already belonged to the first assistant message from 21:50. Because message_id is the primary key, the new text was appended into that old message instead of creating a new
  one after “hello”.

  From the DB:

  • Your user message hello exists at 03:19:41
  • No new assistant row after it
  • The old segment:0 message was updated at 03:19:50 and now contains the original text plus the new “Hello…” reply
  • Same class of corruption hit the earlier 22:51 turn (segment:1 reuse)

  Bottom line

  Provider path worked. The bug is in ACP assistant item ID stability across session restarts: segment indices reset and collide with persisted messages, so new replies get merged
  into old ones and look missing at the bottom of the chat.

and here is the fix: https://gist.github.com/hicder/19d8e836d9bf4e061558254d91f9dcdd
I don't think it's up to standard (I don't really know JS or TS), this is just AI vibe-coded. Hopefully it aids with the debugging if someone wants to send proper patch to fix it.