• Fixed: a model at a high reasoning effort could spend its whole output limit thinking and be cut before writing anything, which ended the agent turn with response truncated by max_tokens (#66). When the cut response delivered nothing and the request still had the model's full output limit available, the request is now retried, up to twice, with a reminder to reason less and act. Cut text, a truncated tool call and a request the context window had already squeezed fail as before.