Claude always-thinking models no longer return 400s or empty answers

Requests to Claude models that always think, such as Opus 5.5, now behave predictably:

• A request that ends on an assistant turn (prefill) fails fast with a configuration error that names the fix, instead of a provider 400 and a retry cycle.
• An explicit none reasoning effort maps to low on models where thinking cannot be turned off. Before, the driver sent nothing and the API ran at its own default.
• Each request in a session replays the previous one and only appends to it, so preserved thinking stays valid across turns.

See Anthropic. PRs #3782, #3783.