Sort what you say while an agent works
Get the voice models from Models, dictate with ⇧⌘D, and your Mac sorts what you say mid-run: a change now, work for later, or not for the agent.
Said during a run, sorted on your Mac
When you talk while an agent works, the decision model on your Mac reads each sentence and decides whether it changes the work now, is new work for after it, or was not meant for the agent at all. Work for later waits under Queued in the order you said it, and a sentence that was not for the agent shows faint with “not sent”.
Only the model on this Mac is asked, and only while it is on in Settings. When it is unsure, or takes longer than a second and a half, the sentence steers the run unless it opens by putting the work off.
The voice models in one download
The speech models now come from the Models library like any other model: 529 MB, checked file by file. If they are missing, the voice window offers to download them and starts listening once they are in.
Keys that do one thing
Dictate into the focused prompt bar with ⇧⌘D, because ⌥⌘D belongs to macOS. Every key the menu shows now does what the menu says, and clicking the Thinking line in the voice window opens that run's reasoning in its chat.