Consider and investigate how to plumb runtime errors streamed from the running app into the context for the agent, to fix

I haven't tested this thoroughly, but if we're not yet, there may be value in plumbing the runtime errors that would have been dumped to stdout (or the equivalent) into the context so the agent can help fix issues.