Staged, Not Sent
Written later and filed to this date, as part of filling in every day of the year.
On the glasses I can open one of the fleet’s running coding sessions, watch its transcript scroll by on six lines, and answer it. There are two ways to answer. The session often ends its turn with a suggested reply, and a click on the ring accepts it. Or I hold the ring, speak, and a transcriber on the laptop turns the audio into text that goes into the session as my reply.
The accept path had been given a gate in early June: before injecting, the display shows the suggestion with “click:” in front of it, so the click is informed. The voice path had no gate at all. The instant the transcriber returned, the client sent the text. The handler was three steps long: result arrives, echo it, send it. A misheard sentence would drive real tools in a real repository with nobody having read it, which is precisely the risk the whole design exists to hold: every mutation triggered by a human, on the ring, on purpose.
Nobody had filed this. The agent that fixed it was auditing what it could build without hardware, found the top items all gated on glasses or a simulator, and read the handler instead.
The transcript is now staged. It appears on the display as “send:” followed by the text, in the row the suggestion would otherwise use, and only one such row ever shows, so a click means exactly one thing. The ring click sends it. Speaking again replaces it. Backing out to the fleet list, or the session opening a different job, discards it, because a reply belongs to the job it was spoken in. The echo of what was sent is written to the transcript only on send; a cancelled utterance leaves no line claiming I said it.
Nine tests, all on the display: the staged row shows, it reads as send, it wins over a suggestion, it clears, it truncates, it is never written into the scrollback.
On a six-line display the confirmation costs one of them. The row reads “send:” where the other reads “click:”, and both wait for the same finger. The difference is what comes after the colon: one is what the session proposed, the other is whatever the microphone thought I said.