Claude runs

Run Claude in the runnable environments. Each episode is verified, recorded as an attempt, and on train scenarios followed by a reflection that proposes lessons for you to approve on Lessons. Approve lessons, run again with memory, and compare held-out pass rates to see whether learning transfers. Attempts also feed the MCMC models on Engineer.

No ANTHROPIC_API_KEY in the server environment. Set one (or log in with `ant auth login`) before starting a batch; otherwise episodes will fail with an authentication error.