v4.100.260

SDK

• An SDK session now runs the way the plugin's does when a host leaves a
  setting out. The extension always sends its own values. Every other host,
  including the CLI's unset fields and any program built on
  @cerebriline/sdk, got what the SDK did on its own. In several places that
  was not what the plugin runs:

  • Auto-compaction, checkpoints and the task checklist are on unless a host
    turns them off. Before this they were off unless a host turned them on.
  • On Ollama and xOllama, the SDK asks the model about itself before it sizes
    the session, as the plugin does. The window the model declares now sets the
    output cap, the thinking allowance and the compaction trigger. Until now
    these were sized without it, while the first request already used it.
  • Whether a local model reads images and calls tools now comes from what
    Ollama reports, so a text-only model is no longer sent screenshots.
  • An xOllama council compacts its own conversation, so client-side
    compaction stands down for one on every host, not only in the plugin.
  • opencoti gets the same system turn from every host, so its shared prompt
    cache is reused across sessions.
  • resolveSessionPromptTemplate asks Ollama or xOllama which family a local
    model belongs to, and picks that family's template, as the plugin does.

• "Prior reasoning: Automatic" now works on opencoti and llama-server. The
  check that decides how much earlier reasoning to send back posted to
  /v1/apply-template, where the server has no such route. So Automatic never
  measured anything, and every earlier reasoning block was dropped. This
  affected the plugin as well as the SDK. The check now goes to the server's
  root, and it runs before the session starts.

  • A llama-server behind the "OpenAI compatible" provider is checked too.
  • An endpoint that is not llama-server keeps what it did before.
  • The result is reported to the host as reasoningHistory in
    onConfigResolved: what was sent, by which route, and whether it was
    measured, set by hand, or left at the provider's default.

CLI

• On xOllama the CLI now picks a local model's family template, as it already
  did on Ollama.