Two things that made the session's own address a lie. The code on the rock now points at the piece's published URL, and nothing published it until the first edit — so a fresh session advertised `@jeffrey/dafu` and `@jeffrey/dafu` 404ed, every time. Auto-publish was on and waiting for a save that had not happened yet. The blank publishes now, as soon as the handle resolves, which is the earliest moment there is a route to publish to. That reverses a deliberate rule: an untouched session used to leave nothing behind, out there or in the workspace. It still leaves nothing in the workspace — the local blank is discarded on exit if it was never edited — but the published copy stays. A published blank is a smaller problem than a code that resolves to nothing, now that the code is the piece's real address rather than a throwaway channel. The second is that a prompt typed during a running turn was thrown away with "A turn is already running." Thinking ahead of the machine is the normal way to use this — you read the first half of an answer and already know the next instruction — and the sentence was discarded, so you waited and retyped it. Those lines queue now and start when the turn reports completed, one at a time, routed back through the same path a typed line takes so a queued `/command` is still a command. A turn that fails before it starts drains the queue too; it emits no turn/completed and the queue would otherwise wait for a turn that never comes. Ctrl-c drops the queue. An interrupt is a decision about everything you were about to say, not only the turn that happened to be running, and having three stale instructions fire after you stopped the machine is the wrong reading. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01A1xZ1BonXHyds6EzF5H9Hb