Autopilot — agents that pick themselves back up
Autopilot: agents that pick themselves back up. Turn it on in Agent Orchestration → AI Diagnosis and an agent that crashes or drops its connection mid-task is restarted in the same terminal with its session resumed and told to continue — or its seat's backup takes over. It never types into a question, picker or half-written prompt, gives up after three spaced tries with an Activity card instead of a restart loop, and logs every decision. Alongside it: projects can now live on a paired computer with no SSH, Reader mode is rebuilt as a real reading view with canvases, a prompts rail and its own composer, each diagnosis checker gets its own model, and the Git client opens files and diff lines on double-click.
Download v1.75.0
The download page always has the newest build; installers for this version are on GitHub.
What changed
New features
7- Autopilot — agents that pick themselves back up — turn it on in Agent Orchestration → AI Diagnosis (off by default). When a running task's process exits, the task goes quiet or a message to it stalls, Autopilot restarts the agent in the same terminal with its own resume command and asks it to continue where it left off; with no session to resume, a chart seat with a backup hands the task over. It waits 5 s, 20 s and 60 s between tries, restarts a terminal no more than 6 times an hour, then stops and pages you with an Activity card.
- Autopilot waits instead of typing over you — a question, a picker or a half-typed draft on screen leaves the terminal alone; idle tabs you never prompted are never checked; Pause stops every action without turning it off; and Recent decisions lists what it resumed, switched or left waiting.
- Pick the model each diagnosis checker runs on — the checker agents AI Diagnosis uses to read a stuck terminal's screen now each take their own model, or stay on the CLI's default.
- Projects that live on a paired computer — Add Project lists paired computers as "<name> · paired": attach one of its projects, pick a folder it already has or create a new one there, then add only the agents that computer has installed and prompt them from a mirrored terminal — no SSH required. First-run trust and update questions get one safe button, and Start again relaunches an agent that exited, with its session resumed.
- Windows Firewall stops being the silent blocker — on Windows the pairing panel says when inbound connections are blocked and offers Allow through Windows Firewall, a single administrator approval scoped to 1DevTool.
- Reader mode, redesigned as a real reading view — choose a Dark, Black, Sepia, Paper or Forest canvas or custom colors; Serif, Sans or Mono; Snug, Comfort or Wide measure; Document or Chat layout; and 1, 3, 5, 10 or all turns. A Prompts rail with ⌥↑/⌥↓ hopping, Find in transcript, per-turn Copy, Copy all for the shown turns and a reply composer make it a place to work, and the phone reader uses the same canvases and layouts.
- Double-click to open from the Git client — double-click a changed file to open it in the editor, or a diff line to land the cursor on that exact line.
Improvements
5- Pairing holds up on real networks — a session stuck on an old address reroutes to the computer's current one, and the machine roster shows each paired computer's 1DevTool version.
- Phone pairing advertises the address that works — the QR code never picks VPN or tunnel adapters and ranks host-only virtual adapters (Hyper-V, WSL, Docker) last, so a Windows PC puts a reachable address in the code.
- Joining a computer is faster and explains itself — all of a pairing code's addresses are probed at once, and a failure names each address with its reason and the next step to take.
- A missing SSH key says so in plain words — a machine whose key file isn't on this computer names the file and where to choose it again instead of surfacing a remote "Permission denied".
- Changing a saved SSH key carries onto the machine — editing a key in SSH settings now updates the machine roster's copy automatically.
Bug fixes
7- Fixed Codex account usage "resetting" after an app restart — the last known usage now survives a restart and falls back to Codex's own session log when the token lapses.
- Fixed Add Account for Claude failing with "Claude Code is not installed" in the packaged app when it was launched from Finder.
- Fixed clicking a file in File Explorer appearing to do nothing when the Editor lived in a collapsed sidebar — the sidebar now opens with the file.
- Fixed opening a file at a specific line landing on line 1 when no editor was open yet.
- Fixed prompts reporting "did not confirm acceptance" while the agent did the work — deliveries from links, the phone and paired computers now recognize wrapped and boxed composers (OpenCode, Grok, Cursor) and turns still in progress.
- Fixed OpenCode terminals printing their help and quitting on older installs — the launch flag now matches the installed version.
- Fixed Pi on Windows never accepting a prompt, and Grok on Windows stuck "not ready" after a long turn.
Screenshots


