There is unclear documentation and recent changes regarding how terminal commands run by the agent time out.
Some commands run by the agent can take up to 20 minutes or even 1 hour. It seems that despite adding CRITICAL timeout instructions, some commands still time out way earlier.
We know that our repository does not have timeout issues or dead commands, so it is fine for us to wait, we don’t mind.
Is there any workaround to disable the timeout on the terminal/shell tool?
This is painful because, in the case of long-running test commands, the agent gets stuck.
Is there any quick workaround, or is a feature planned that would allow us to customize this parameter?
Steps to Reproduce
Run a command for more than 5-10min via agent shell
Expected Behavior
Control timeout with a Cursor Setting for terminal/shell tool
Operating System
MacOS
Current Cursor Version (Menu → About Cursor → Copy)
My repository rules specify running local CI with a 25-minute timeout, and the Agent (and OpenAI Codex) seems to be following suit. However, the longest I’ve seen with this builder is 23 minutes, and usually it’s either 2-3 or 9-10 minutes, depending on the changes and the launch method.
The problem is that if the launch takes longer than ~ten minutes, the system assumes the connection with the Agent has been lost and aborts the request.
Same issue here. Running a long test, where there is no benefit to the chat continuing while it runs… there is nothing to do until the test results are available.
I don’t understand why someone thought it was a good idea to add in arbitrary limits without the user option to control them. Better customer empathy, please!
I can confirm that I switched my rules to use Block Until ms. Currently, it seems to work and forces the command to finish even if the process takes 15 minutes. I’m just hoping the Cursor team doesn’t break this feature again, it’s important.
Super important @Artemonim It also seems that tool usage is customized per model. Composer, for example, can’t use the Block Until ms option. I don’t know why. Codex and Claude seem to work properly.
I also had some issues steering Composer correctly with certain tools, so I stopped using it. Now I only use it for search, and honestly it’s too expensive and not very smart compared to Codex 5.2 Medium.
I used to use Grok Code Fast if I needed information from the repository. I have now switched to Gemini 3 Flash. Although the default /explorer-subagent is Composer, and you can only prevent it from being used through rules. But I’ve accepted it
