shuttle v0.1.0

Windows: this build needs the Microsoft Visual C++ runtime (VCRUNTIME140.dll); use v0.1.1.

The first usable release: shuttle drives the desktop on macOS and Windows.

• Daemon. cocoon-shuttle mcp starts cocoon-shuttle serve on the first tool call other than status, which the MCP server answers itself (on macOS from Cocoon Shuttle.app, so the permissions go to the app). The daemon runs the bundled cua-driver on a socket of its own, restarts it when it exits and stops it with itself; desk actions run one at a time.
• Tools. desk (windows, wait, snapshot, find, read, press, act, at, key, type, click, focus, close, launch, shot, screen), screenshot, gui_plan and script_run now run on macOS (AX) and Windows (UI Automation). gui_plan runs whole GUI plans, saves passing ones as flows and replays them; script_run runs a module on the bundled Node with the desk API. On Linux the desk and screenshot tools and the desk calls of a plan or script answer that there is no desk backend yet (P3); script_run and model calls run.
• Models, bring your own key. One OpenAI-compatible endpoint and key, OpenRouter by default: chat completions for ask, ground and decide over chat, and the decisions endpoint for decide and verify. The key comes from config.toml or an environment variable; scripts run without it in their environment.
• Safety controls on macOS and Windows.
  • cocoon-shuttle stop / resume, a global kill hotkey (⌃⌥⌘⎋ on macOS, Ctrl+Alt+Shift+Esc on Windows) and the indicator's Stop and Resume. Stop kills the running plan or script and ends the calls in flight; no MCP tool or plan step can resume.
  • An indicator: a menu-bar item on macOS, a tray icon on Windows.
  • Confirm mode: confirm = "coordinate" asks before a click at a point, "every" before every desk action. While a dialog is open, every other desk call fails.
  • An app allowlist ([safety] apps) that checks every desk call by window, element and point, and blacks out other apps in screenshots.
  • [tools] script_run = false; a config.toml that fails to load fails every tool call except status.
  • These are guard rails for tool use, not a sandbox; they hold only with script_run = false. docs/safety.md lists the limits.
• Speed. Batched macOS window lists, process-scoped window reads, the menu bar only in snapshots that need it, a plan step that reuses the plan's snapshot, a window wait in the daemon (desk wait) instead of polling, a 1 ms cursor glide, and a grounded click that takes at most 3 screen captures.
  • A plan whose last step is a shift key, an arrow in a plain window, or typing without a newline ends without the end-of-plan window wait.
  • The macOS window list skips the bundled cua-driver's own windows instead of reading them.
• Fixes. On Windows a dialog inside its window's tree, such as Notepad's unsaved-changes prompt, shows in desk snapshot and find and can be pressed; before, the walk skipped it.
• Bundled: cua-driver v0.36.0-loomhand.4 and Node.js 24.22.0, both checked by sha256 at build time.
• Docs: a first-time setup in docs/install.md (install, macOS permissions, your key, the safety settings to start with), and the status of every platform.