No more long silences during a task
Progress narration moves from "the model volunteers a sentence" to a framework guarantee: past a threshold of silent rounds or seconds, it generates a short spoken progress line. It skips when there is genuinely nothing new, and immersive skills can ask for quiet.
Give the robot something that takes a while — research, a multi-step procedure, a run of tool calls — and it used to say nothing at all until it finished. Whoever was waiting could not tell working from hung.
It speaks up now.
The mechanism changed, not the prompt
Narration used to rest on "the model volunteers a sentence and the system speaks it". Two problems showed up in practice. It could not be relied on — through long runs of tool calls the model often wrote nothing. And when it did write something, it was the wrong thing: internal reasoning, not a progress report for someone who is waiting.
The framework now guarantees it. Once enough rounds or seconds pass with no user-facing output, it generates a progress line and speaks it.
| What it says | What it is doing, what it found, what it plans next |
| How long | Under 40 characters, conversational |
| When it stays quiet | When there is genuinely no new progress — saying "still looking" three times is worse than silence |
It stays quiet when it should
Immersive skills — games, roleplay, performance — and sensor-driven autonomous behaviour nobody is waiting on do not want narration. A skill's own instructions take precedence: ask for quiet and you get quiet.
One known limit
The check happens between rounds. A tool call that runs for 90 seconds will not produce a report while it runs — only after it returns. What this does solve is the slow round: two rounds of 40 seconds each are well past the time threshold long before the round-count threshold would fire.