v4.100.261

Agents

• "Prior reasoning: Automatic" now sends Gemma-4 its reasoning in the
  template's own format. The check that decides what to send asked the
  server with the reasoning placed before a second user message. Google's
  official Gemma-4 template keeps reasoning only after the last user message,
  so it was measured as dropping reasoning it actually keeps. The check now
  asks with the layout an agent run sends: the reasoning on a tool call, then
  the tool's result. Gemma-4 is now sent its latest reasoning in the reasoning
  field, and the template writes it as its own thinking block, in the right
  place.
• Reasoning is no longer folded into a message that calls a tool. On a
  template that really drops the reasoning field, the latest reasoning used to
  be written into the assistant's text as <think>…</think>. Gemma-4's
  template puts that text after the tool's result and closes the turn there,
  so the next turn started without its header. A message that calls a tool now
  carries no folded-in reasoning. A message that only answers still does.
• "Save model requests" now saves the answer too. Beside each saved request
  there is now a .response.json with the response exactly as the model
  streamed it back, its reasoning included. A request alone was not the whole
  turn. The model's answer only showed up in the next request, after the
  history had been reshaped: older reasoning dropped, and the latest sometimes
  moved into the reply text. A response that was cut off is saved with what
  arrived and complete: false. A request that got no answer is saved with
  the error.
• On plain Ollama, saved requests are now filed under their conversation
  instead of no-session.

SDK

• saveModelRequests on ClineCore.create does what the plugin's "Save
  model requests" setting does, with the same files in the same layout:
  <dataDir>/logs/requests/<session>/<time>-<n>.json and
  <time>-<n>.response.json. Pass true, or { dir, responses }. It is off
  by default, as in the plugin. applyModelRequestCapture switches it while
  running. createModelRequestWriter is the writer itself, for a host that
  wants the same layout somewhere else.
• A writer installed with Llms.setModelRequestCapture can ask for each
  request's response by returning a function, which receives
  CapturedModelResponse (status, the raw body, complete, error).
  Headers are never saved.