(long thread)
there has gotta be a better way for models to handle terminal and bash running
i was working with opus 4.5 over the weekend on having it do some video gassian training on an SSH box i gave it access too
the primary issues (at least in opencode) is that the timeouts for the bash commands and the "max" character limits were the main bottleneck, I would often have to say like "give me the commands to run so I can run async" becuase these jobs would take hours and it would just either truncate or auto kill the running processes
there needs to be a process to like
run command -> watch command <if command takes longer than 30 seconds move it to a background terminal and release the hold on the conversation lock> then the agent can see the last logs before it was released and then it can grep/view last lines etc via like a terminals tool similar to how @cursor_ai does it
the only issue with what cursor is doing is that when they start a terminal it is locked, like no user input which is very unintuitive, I want to be able to (like subagents) switch to that terminal and "observe or interact" with it in a way
this should be pretty possible in implementation, its then just a fact of getting the model to reliably use it


