Comment by joegibbs

I don't think this will work. There are too many tools, a lot of them legacy, and it will be an incredible pain to get them all coordinated on a new standard. Instead I think it'd be better to solve this on the agent side by checking every ~5 seconds what the console output is and sending that, along with the amount of time the process has been running (or perhaps every line could be timestamped?), to a smaller model (like Haiku or GPT5-mini). The model could then provide some input, kill the process or do nothing.