Summary
During a normal long-running ml-intern CLI session, the run can log:
system: Doom loop detected — injecting corrective prompt
This can be surprising: it reads like a failure or bug, but I think it is intentional behavior from check_for_doom_loop in agent/core/doom_loop.py. Users may interpret this as the agent “breaking” or being stuck, even when the underlying issue is “repetitive but legitimate exploration” or “slow tool calls that look identical until results land.”
Could we remove or clarify this messageing?
Summary
During a normal long-running
ml-internCLI session, the run can log:This can be surprising: it reads like a failure or bug, but I think it is intentional behavior from
check_for_doom_loopinagent/core/doom_loop.py. Users may interpret this as the agent “breaking” or being stuck, even when the underlying issue is “repetitive but legitimate exploration” or “slow tool calls that look identical until results land.”Could we remove or clarify this messageing?