Coming back to the app no longer briefly claims an agent has stopped while Cowboy reconnects. It now shows that it's reconnecting, and the agent's output comes back once it's connected.

- Write messages over several lines. Return adds a line break, and the send button sends. - Your message appears in the chat as soon as you send it. It's marked as pending until the agent picks it up. - If the agent doesn't react, Cowboy tells you, so you don't send the same thing twice. - Clearer connection messages that say why Cowboy can't reach your machine, instead of showing raw error text.

When Claude Code asks you a question with several answers, the answers now show up as buttons you can tap, the same way permission prompts already did. That includes the questions that have a short description under each answer.

• See your agents by workspace. A new Priority / Workspace switch on the home screen groups agents under the workspace they run in, with the most urgent one first. Cowboy remembers which view you picked. • Less clutter once you're set up. After your own machine connects, the demo host removes itself from your host list.

Adding a host no longer means pasting a private key into your phone. Pick Password, type the login you already use for that machine, and tap "Sign in and install a key". Cowboy signs in once, leaves a key of its own on the machine the way ssh-copy-id does, and forgets the password — it is never stored, on the phone or anywhere else. Undo it whenever you like by deleting that one line from ~/.ssh/authorized_keys. The first time you connect to a machine, Cowboy now shows you its SSH fingerprint and waits. Nothing is sent until you say it matches, so a password only ever reaches the machine you meant. Hosts behind a second factor work too: you answer it once, when you set the host up, rather than every time Cowboy reconnects. Pasting a private key still works for anyone who prefers it, and Tailscale SSH still needs nothing at all. Also fixed: on the host screen, "Forget host key" and "Delete host" ran off the edge of narrow phones.

Cowboy has been redrawn. • Agents sit on cards that name their state in a word as well as a colour — Blocked, Done, Working, Idle — so the one waiting for you stands out at a glance, and still reads if colour is hard to tell apart. • Write mode is unmistakable: the top bar turns amber while input is on, names the pane you are typing into, and counts down to switching itself off. • Terminal output keeps its own monospaced type, so what the agent printed still reads like a terminal. • New type throughout, with buttons, fields and dialogs that look like the rest of your phone instead of a terminal pastiche. Nothing about how Cowboy works has changed: read-only until you say otherwise, straight to your own machine over SSH, no account and no cloud.

Cowboy can now tell you when an agent stops and waits for an answer. Switch it on under hosts. You get one notification per agent, and it goes away by itself when the agent stops waiting or when you open Cowboy. Only agents waiting for an answer are announced, so nothing else interrupts you. iPhone decides how often it wakes Cowboy to look, so a notification can arrive a while after the agent asked. Also fixed: a host that wants an extra browser check, as Tailscale SSH does, used to leave Cowboy trying to connect for a long time without saying why. It now shows you the link, keeps waiting while you approve, and connects.

Cowboy is free to try for three days, then a one-time $9.99 unlock. No subscription. If you already have Cowboy, your three days start when you open this version — nothing is taken away before then. Adding hosts and checking connections stay free either way, and the built-in demo session stays free to poke at. Nothing else changed in this version.

Agent output now keeps itself up to date, and fits the screen. Live output. While an agent works, its output refreshes on its own every couple of seconds, and again when you come back to the app. The refresh button is gone: tap the status band to read again, and the time beside the folder tells you how fresh what you're reading is. Wrapped to the screen. Long lines fold at the edge instead of running off to the right, and a wrapped line lines up under the text it belongs to. Tables, file trees and box drawings keep their columns and scroll sideways on their own. More output in view. The space the toolbar took went back to the output. Long-press the output for the exact terminal text, and again to go back. Stays where you are. New output only pulls you to the bottom if you were already there, so you can read further up while an agent works. Fixed: after an agent finished, the output you could scroll back through shrank to a few lines until you refreshed.