Agent uses `timeout xx command` instead of setting the timeout arg on the bash tool

[deansheather]

This has been a recurring problem for me in many chats on https://github.com/coder/mux-llm-proxy.

In pretty much every chat mux will try to run a command which times out with the default 3s timeout. Then instead of bumping the timeout parameter for the bash tool call, it will just prepend the command with timeout 30 or something similar, which doesn't work.

Maybe we could change the documentation for the bash tool to make it less likely to try this? Or maybe we could detect timeout commands and use that as the timeout?

[ammario]

what model are you using? I haven't experienced this behavior. We could block commands that start with timeout in the same way be block redundant cds. The blocking has the advantage of "teaching" the model to not waste tokens on future calls. We could also parse out the timeout command and use whichever timeout is greater in the bash tool timeout but that seems like masking a more fundamental issue.

[deansheather]

anthropic:claude-sonnet-4-5

I could add it to the AGENTS.md, but it seems like I'd be working around an issue that users would need to also work around.

Another instance today:

[deansheather]

Perhaps we could improve the Command exceeded timeout error we give to the LLM to tell it explicitly to use the timeout argument on the tool call

[deansheather]