v4.100.262

Agents

• On llama-server and opencoti, prior reasoning now survives the harness's
  own nudges. Reminders such as "your last message contained no tool calls"
  are sent as user messages. Gemma-4's official template keeps reasoning only
  after the last user message, and so does Qwen 3.6's, so each nudge dropped
  the reasoning written before it. Where the template has the
  preserve_thinking switch, it is now sent with chat_template_kwargs, and
  the reasoning is kept. "Prior reasoning: Last" still sends one block, so the
  prompt does not grow. The same check that decides "Automatic" measures
  whether the template needs the switch, and it is sent only where it changes
  what the template renders.