Saying I'm leaving now turns off every light, and reports what didn'tSaying I'm leaving now turns off every light, and reports what didn't

September 4, 2026 · A voice command now turns off every light and names any that failed, and a same-day switch to a second AI dispatch tool got reverted.Saying I am leaving now turns off every light and reports which ones did not respond, and a same-day switch to a second AI tool got reverted.

Saying "I'm leaving" now turns off every light, and reports what didn't

The Alexa skill got a new voice intent today, and it's the mirror image of one that already existed. "I'm here" turns every Home Assistant light on; "I'm leaving" (or "I'm heading out") now turns them all off, over the same listLights()/setLightPower() path, so it reaches lights that were never individually set up as Smart Home endpoints, which is most of them. There's no confirmation step, on purpose. Asking "are you sure?" of someone already walking out the door costs an exchange every single time, to guard against an outcome that one more spoken phrase undoes anyway.

Building it surfaced a real bug in the intent it mirrors. Both intents fan a command out to every light at once, and that fan-out used to wait on all of them together. So one dropped socket among thirteen calls answered the whole request with a flat "reached server, but had error," even though eleven lights had already switched. It now waits on each call independently and answers with a count of the ones that didn't respond. Two of the thirteen real lights were unavailable on the day this shipped, so a partial result turns out to be the normal case here, not an edge one. A real round trip against the live house confirmed it: the office light went off, then came back on at its old brightness.

A second AI CLI joined the dispatch queue, then got pulled back within hours

Separately, the queue that hands Linear tickets to an AI session gained the ability to run that session through Codex's CLI instead of Claude Code: a second engine behind the same four dispatch roles (triage, auditor, release, the job that writes this blog). That took a real fix: the prompt every dispatched session reads described itself as "running as claude -p" with a Bash tool and a turn ceiling, in plain prose, regardless of which CLI it actually ran in. The first role dispatched onto Codex would have been told it held tools it doesn't have and warned about a ceiling that doesn't apply to it, and its attempt to satisfy an instruction to call a nonexistent tool while also writing a prose fallback would likely have produced neither. That prompt now describes whichever CLI actually resolved for the role. Codex also needed its own project brief (it only auto-reads a file called AGENTS.md, confirmed by testing it against a token no model could guess), which didn't exist here before today.

With that in place, all four roles were switched onto Codex in one push. Then, a few hours later, the switch itself was reverted. The plan behind this move required one real, completed Linear dispatch per role before trusting it, and that never happened: the first live session hit its 160-turn limit immediately after asking to restart, landing seven commits with no finished proof that a role could actually complete work this way. Triage is the role that routes every other ticket, so a bad bet there would have been a fleet-wide failure, not a contained one. Nothing was mid-flight, so reverting cost no work in progress. It also cost no ground: the plumbing (the adapter, the schema translation, the project brief, the registry) is still live; only the "which CLI runs this role" dial went back to Claude, to be flipped one role at a time, each only after a real dispatch proves it.

Saying "I'm leaving" now turns off every light, and reports what didn't

The voice assistant that controls the smart home learned a new phrase today, and it's the flip side of one it already knew. Saying "I'm here" already turned every light in the house on; saying "I'm leaving" (or "I'm heading out") now turns them all off, using the same underlying method, so it reaches lights that were never individually set up to be controlled by name, which is most of them. There's no "are you sure?" step, on purpose. Asking that of someone already walking out the door costs a back-and-forth every single time, just to guard against a mistake that saying one more sentence would fix anyway.

Building this turned up a real bug in the existing "I'm here" command. Both commands send an "off" or "on" request to every light at the same time, and the code used to wait for every single one to succeed before saying anything back. So if even one light failed to respond, out of thirteen, the whole command reported a flat "something went wrong," even though eleven lights had already switched. Now each light is checked separately, and the response says exactly how many didn't respond. Two of the house's thirteen real lights happened to be offline on the day this shipped, so getting a partial result turns out to be the normal case, not a rare one. It was tested against the real house before shipping: the office light was switched off, then switched back on to its previous brightness.

A second AI assistant joined the automated work queue, then got pulled back the same day

Separately, the system that hands off coding tickets to an AI assistant to work on autonomously gained the ability to run those sessions through a second AI tool, Codex, instead of the one it always used, Claude. That took a real fix first: the instructions given to every automated session described themselves as running inside the first tool, with a specific set of capabilities and a specific time limit, no matter which tool was actually running the session. The first ticket handed to Codex would have been told it had abilities it doesn't have and warned about a time limit that doesn't apply to it, and trying to follow both a broken instruction and a working one at the same time would likely have produced neither. That description now matches whichever tool is actually running. Codex also needed its own copy of the project's written brief, in a file name and location it actually reads automatically (confirmed by testing it with a made-up secret word no AI could have guessed), which hadn't existed until today.

With that fixed, four different kinds of automated tickets were switched over to run on Codex all at once. Then, a few hours later, that switch was undone. The plan for this change required proving each kind of ticket could actually be finished successfully on Codex before trusting it with real work, and that proof never came in: the first real attempt ran out of its turn limit right after asking to be restarted, leaving behind seven small changes but no finished, successful example. One of the four ticket types decides where every other ticket in the system gets routed, so a bad decision there would have broken the whole system, not just one part of it. Nothing important was left half-done, so undoing the switch didn't lose any work, and it didn't lose any progress either. The underlying groundwork that makes Codex usable at all is still in place. Only the decision of which tool runs which kind of ticket went back to the original one, to be switched over again later, one ticket type at a time, each only once a real success proves it works.