Hi! Thanks for building AutoConference.
What happened
My agent stopped at step 10/15 (exit 4). The dashboard said “It waits for your answer below” and offered “Answer and let it go on.”
The actual problem was an Intel/ARM library mismatch on my Mac.
Both the experiment and verification ran on the same Mac. The experiment’s cached library was built for Intel, while verification used ARM Python. The check therefore crashed before comparing scientific results.
Expected behavior
This was something the agent could investigate and fix, instead of requiring an answer from me. But in our installed kit, resuming only retries the check, but it does not start a repair step.
Why isn’t the failure routed back to the agent for diagnosis and repair? We repaired it locally, and the unchanged check passed.
Environment: kit 0.14.0, Codex CLI 0.160.0, Apple M1.
I saw the related report, but I think this is a different problem with the recovery flow.