i ask for the agent to do something. it runs a tool. the tool outputs an error. the agent never notices. i have to either move to background, or i have to spend a request to talk to the agent and move it up in priority above the wasted task.
Steps to Reproduce
i ask for the agent to do something. it runs a tool. the tool outputs an error. the agent never notices.
Expected Behavior
notice that the tool failed and react to it appropriately
as you can see it has nothing to do with the amount of response content so it’s not like a ‘more’ or shell interactivity required situation. looks like it’s just the ide chat not recognizing that the command has returned a result.
through a lot of discussion with claude 3.5 systematically trying varying commands with varying output formats and lengths, i figured out it was kubectl command and the PAGER env variable.
i set my environment variable PAGER to cat in windows user level environment variables. this let’s the entire output go out to the console regardless of length. I am not sure why it started to have this behavior out of the blue today but this seems to have fixed it. AND i’m not sure why cursor cannot see the (more) output.
basically, every now and then (not always, seemingly no apparent reason) powershell script will complete ‘show prompt’ but cursor won’t recognize that it’s finished and my only options are to ‘cancel it’ or to ‘move to background’
if i move it to background, it does recognize the work that was done and resumes processing the request. really obnoxious because i have to sit here and watch closely and click ‘move to background’ so it resumes.
