One model by default, more room, and reviews that keep moving

Added

• One model by default: Send runs your request on one eligible model. Choose Auto, an agent system (Claude Code or Codex), or an exact supported model, and Winskel uses exactly that.
• Orchestrate when you choose: press Orchestrate to split suitable work into missions and run independent parts at the same time. Pressing it in Chat or Plan moves your unsent draft into Build, says so, and asks you to confirm before anything runs. Each mission's model follows the usual routing rules and your model choice, so several missions may use the same model.
• Collapsible sidebar: hide the sidebar with the button at the start of the title bar, or Command and backslash, and the conversation takes the full width. Bring it back the same way. Your conversation, your draft and any running work are unaffected.
• Preview supported apps: a result with its own start script now runs as an app, so its pages, styles, scripts, API and saved data work, including an app inside a subfolder of the result. It runs from a disposable copy of the result on your Mac, in a sandbox, reachable only on this Mac, and asks the first time. Data you save in the preview stays in that copy, never in your project folder. Results without a supported start script open as a static page, as before.
• Reviews of preserved work: when Winskel keeps work it could not verify, the independent review you asked for still reads exactly that kept code and is told which checks were blocked. If it finds blocking problems, choose Repair the findings: the repair starts from the same kept code and an independent review checks the repair.

Improved

• Shared memory: relevant project knowledge reaches both Claude Code and Codex, a corrected fact replaces the old one, forgotten facts stay gone, and turning access off stops anything from being shared.
• Clearer verification: checks run from the app's own folder and report what really happened, with the exit code and the test counts when the test runner prints them. A test followed by another command, such as an echo, is never read as a pass. Checks that could not run are shown as blocked, with the reason, never as passed or failed.
• Stop for Preview is per project: it stops only that project's preview and leaves nothing running.

Fixed

• Preview could open a result at the wrong location, its public folder, instead of the app itself.
• Stop could disappear after you left the conversation and came back while a preview was still running.
• Resume verification could run checks from the wrong application folder; it now uses the app's own folder.
• A brief connection blip could close open previews and terminals.

Known limitations

• Tests that start a local server remain blocked by Winskel's verification sandbox on Mac. Winskel preserves the work and explains which checks could not run. Independent review can still proceed, and its verdict is shown separately from verification.
• Send does not split work by itself; press Orchestrate when you want parts to run separately.
• Previews on your Mac share cookies with each other, and fully reloading the main window closes open previews.
• Claude Code and Codex run on your own accounts, and their usage limits still apply.
• Winskel runs on Apple Silicon Macs with macOS 12 or later.