Both eval checkouts used this connection-only replacement, then restored it. No model instructions, tool results, graders, or case data changed.

Old:
p = tomllib.loads((home / 'config.toml').read_text())['model_providers']['vibeproxy']

Replacement:
c = tomllib.loads((home / 'config.toml').read_text())
p = c.get('model_providers', {}).get('vibeproxy', {'base_url': c['openai_base_url'], 'requires_openai_auth': True})