ControlBird AI: The response limit (Max Tokens) can now be set to 0, leaving it to the model server to decide. Models that reason at length before answering vary enormously in how much they use, so any fixed limit truncates some replies and not others. On one model answering the same kind of request, that figure ranged from under a thousand to over seventeen thousand.
ControlBird AI: The model list now carries each model's context size and the longest reply it will serve, so the response limit can be set from what the model actually supports instead of guessed at.
Community Edition: The in-memory store can be enabled as an opt-in, for deployments that prefer it.
Bug Fixes
ControlBird AI: A long reply is no longer ended while it is still arriving. Any turn running longer than five minutes was cut off and reported as a timeout even though nothing had stalled, which most often hit exactly the requests that need the time: reasoning over an attached image.
ControlBird AI: Changing a provider setting now takes effect in a conversation that is already open. Adjusting the response limit, temperature or model previously reached only new chats, so following the on-screen advice to raise the limit and try again in the same chat could not work.
ControlBird AI: When a floor plan is approximated from the assistant's reading of your drawing rather than traced from the drawing itself, this is now reported. Both previously produced the same message, so an approximate plan looked identical to an accurate one.
NFC: A scan that is refused now reports that it was refused, rather than reporting the tag as unknown.
NFC: Recording a tag scan completes in a single request instead of waiting indefinitely.
Storage: File content is retained when running with the in-memory store.