crew 5.2.0

The big change here is how Voice OS handles talking to sessions that aren't on screen. Say "checkout, is the build green?" and you stay where you are. Voice OS says "Sent to checkout", the answer plays in full with its name on it, and follow-ups keep going there for a minute. Switches are said out loud now, "go back" walks your last five screens, and "where am I?" gets an answer. Updates from other sessions no longer talk over you. They wait for a quiet moment and come together as one "Meanwhile, …" line, and "what did I miss?" plays them right away.

Voice OS also works in any language now. Every place that read what you meant with an English pattern (a yes, a stop, "mute that", "for checkout?") asks a small Haiku judge instead, and when it can't tell or times out it takes the safe side: nothing gets approved and no words get dropped. There's a "Languages you speak" picker in the listening menu that feeds the hints to Soniox. The speech layer itself (stop words, "end of turn", the wake word) stays English on purpose.

Two smaller things worth knowing. Permission asks no longer read the whole command at you, paths and all. You hear "run grep" or "edit App.tsx", and the page still shows the full command. And if a remote machine is on an older release, the main now runs crew update there over SSH and reconnects, once per version, instead of refusing it. A remote that's newer than the main is never downgraded; the card tells you to update this machine instead. I also fixed the live transcript showing a local session with another machine's prefix, and a few judge regressions the evals caught around mute and listening.

Pull requests

• Voice OS: "speak main" no longer shows as another machine's session (#20)
• Voice OS: permission asks say what a call does, never the command (#22)
• Voice OS: context routing and speech turn-taking (#23)
• Voice OS: works in any language — a Haiku judge replaces the English guards; languages are a setting (#24)
• Voice OS: the main updates a remote that runs an older release (#25)
• Voice OS: fix the eval regressions from the judge (mute and listening) (#26)