Always stops before complete result

Support Bug Reports
1

I have a pro plan, with enough usage left. However, I always has a termination before full complete response generated. Very rarely the response is complete.
I tried different models, modes and settings. still the same.
Here is a message id:
869a95fa-1fd0-4a01-a067-b2f608baca67

Any similar problem would also be appreciated.

5

Hey, we checked the request ID. On our side, all 9 sub-calls finished successfully. There are no errors and no stream disconnects. Your request history for the last week is also all success, which usually means the model decided to stop on its own (typical early stopping), not that the client or network cut the response.

This is a known pattern, we’re tracking it, but there’s no ETA for a fix yet.

To confirm whether this is the same issue or something else, can you send:

  • Which feature it happens in: Agent, Ask, Edit, Plan mode?
  • A screenshot of what the “cut off” answer looks like. Does it stop mid-sentence, on a tool call, or as an empty message?
  • An example prompt or task where it happens reliably
  • Does it reproduce on short or simple requests, or only on long tasks?

Also, based on the logs it looks like you’re in Japan. If you’re using a VPN or proxy, try turning it off, then go to Cursor Settings > Network > Run Diagnostics. Sometimes an unstable connection can look similar, even if the server marks the request as successful.

6

  1. actually, i have tried ask/agent mode with auto/gpt models several times. It doesn’t occur every time, but seems in consecutive chats or resend the same prompt would continuously respond very shot response, and no result, looking like cut at some very first point.

  2. A sample screen shot is:

    image
    image695Ă—227 10.2 KB

    other failure may have shorter response, but all seem like medium process content, definitely not part of the final result. It looks like a cutoff at the beginning as I described above.

  3. the prompt for the screenshot:

    Now, after several tests, here are two new problems:

    1. after user says “stop riding”, the focus does not resume after tts played response.

    log is :

    [2026-05-15 11:55:26.312] [ASRManager] asr stop by silero (trailing silence=885ms)
    …(similar log skipped for saving space)

    [2026-05-15 11:55:33.328] [TTSManager] sent FINISH_SESSION signal

    2. the focus is got once the text shown on in the aichat page bubble, consider to be closer to PLAY_START

    chat request id is 7a4d93b1-a16e-40fb-9d4e-a4534019d847

  4. I think i tried some short requests, but not certain, with the same cutoff like response.

  5. Yes, i have tried Run Diagnostics, ping is good, ping is good. I will try your suggestions.

7

Thanks for the screenshot and details, it’s clear now. On the screenshot you can see the classic early stopping pattern: the agent writes a plan (“I’ll trace the voice action and TTS focus flow…”), does 1 or 2 searches or file reads, then stops without a final result. This matches what the backend shows, all sub-calls are success, no errors and no disconnect. So the model is deciding to end the turn on its own.

We’re tracking the issue, but I can’t share an ETA for a fix yet.

What usually helps as a workaround:

  • Start a new chat instead of resend in the same thread. When context builds up, especially after a few “stop” or “continue” messages, the model is more likely to stop early.
  • In the prompt, clearly specify the output format: “Implement the fix end-to-end and return the full patch, do not stop after exploration.”
  • For big tasks, use Plan Mode first to lock in the steps, then run the Agent using that plan.
  • Try a different model for the same task. Claude Sonnet or GPT-5 Codex are usually less likely to stop early on long tasks.

If it also happens on a short, simple prompt in a new chat, send me the Request ID from that case and I’ll take a look.