Paid latest (45.4) version of Cursor. I gave it a task (agentic mode), it worked for a while, it but it didn’t finish all the tasks. It just displays “generating…” and then timeout. I clicked on try again (or retry or whatever it is) it then added something again, but soon after timeout again. I tried to switch model from Sonnet to GPT4o ,but timeout occured again. I have almost all of my fast requests still available, because i am not using Cursor much, because it’s always this same problem.
Guys, it is impossible to work like this with Cursor, please fix this!
This should be more stable now, but if you’re still getting timeouts try starting a new composer session - sometimes the model can get stuck not running tools correctly. Let me know if you’re still having issues after trying that
I have the same issue. As long I use chat window it works fine. But if I will go for a break or switch to different task and back after few minutes I am not able to make any request. I have to use the ancient method turn off/on Cursor to get back to work
Any ideas where is the problem?
Why do you give us the ability to change the timeout duration for tool use ? A quick win would be a general setting but something better would be different timeout for different tools.
For example, when asking it to build my C++ code, it timeouts after 20s consistently which means that it’s a hardcoded value on your part. C++ code will never compile in 20s, so I d be able to make it wait longer.
Is this possible ?
We are a small team trying it cursor entreprise version in order to roll it out to ~100 developers.
Have you got a request ID from one of these failures? Ideally we’d like to figure out why it’s taking so long, vs just increasing the length of the timeout!
I think I may have been a bit broad in my request so I’ll narrow it a bit.
I’m talking specifically about the terminal tool. Meaning cursor using command line tools and scripts inside the terminal by itself.
In order to build my code, it uses a command like “ninja ”. But 20s after the command started, it interrupts it using ctrl + c.
Here’s an example request id: 7a2c8c27-facb-44d5-ba6a-98e9e6b217f3
You can see that it interrupted the command before it finished. This command would have spent about 20s configuring and about a couple of minutes building. But it stopped it right before it finished configuring.
Does this make sense ? Should I give more examples ? I’m pretty sure nearly everyone who uses the agent would like it to build their code, run the tests and iterate on the code from the feedback of the build and the tests.
It doesn’t matter if it uses more requests/tokens, as long as I can set a limit on it.
Hey, I can’t actually see that example as you have privacy mode enabled, but you can usually stop the CTRL + C issue by adding a rule for the AI to follow telling it to never prematurely end a terminal which it is running.
Alternatively, you can tell it to never run commands on its own, only suggest them, and let the user run them. Then, once it’s run to completion, you can @ the terminal so it can see the output.
I think you are right. I’ve tried with sleep 60 and other long commands and it doesn’t stop them. It seems to have an issue with CMake configuration though.
When configuring the project using CMake (which is a tool to help configure the build but it’s not important here), it prints:
-- Configuring done (13.3s)
-- Generating done (17.8s)
-- Build files have been written to: /path/t/build
What happens is when the model runs it, it interrupts just after the ‘Configuring done’ log.
Here’s an example in this screenshot. I’ve removed some sensitive parts due to legal reasons but I basically run a command which runs cmake to configure the project and the build it. screenshot.pdf (201.3 KB)
I’ve tried multiple ways to prompt it and it doesn’t seem to follow the instructions. (Even using upper case which works in most cases).
The weird thing is that the model says that the command has completed. Is it a hallucination or is there an extra tool which parses the output before feeding it to the model. Maybe that processing step is reading ‘configuration done’ and interrupting the command ?
Yeah it’s a strange behaviour. If you are still seeing this, let me know and I’ll flag to the team. I think we have some terminal improvements rolling out in v0.49, but if that doesn’t help we’ll dig into this a bit further!
