Changelog

0.0.6-alpha.2 current 14 November 2026

Adopts a stricter version scheme, adds a release-channel choice for updates, and fixes the remaining plan-button state bugs.

• Release channels. Updates now follow the channel the build came from — alpha, beta or stable. Settings → About has a selector to override it, so you can stay on alpha or move to stable releases only.
• Versioning. Version numbers are now strict SemVer, and every rebuild bumps a counter. A same-version rebuild was previously invisible to the updater, so an updated build could go unnoticed.
• Fixed: a plan button could be disabled by an unrelated turn. An unexecuted plan was retired after any later chat turn or error, so its button stopped working before it had been used.
• Fixed: a plan button stuck at "Executing plan…". A successful run now finishes the button instead of leaving it mid-flight.
• Documentation link. Help → Documentation opens the user guide in your browser.

0.0.6-alpha superseded 13 November 2026

Consolidates the work from 0.0.5-alpha.1 into a tested build and fixes two defects found by the expanded regression suite.

• Fixed: "Execute plan" did nothing. The button could silently no-op when the backend connection was closed, when no folder was open, or when a plan card had been rebuilt from restored history. It now states why it cannot run and re-arms after four seconds. A finished turn also retires leftover plan buttons, so a plan cannot be run twice.
• Fixed: a crash when a turn ended through an unhandled error. The renderer assumed a cost value that was absent when the backend's exception wrapper ended a turn, which left the interface stuck showing a spinner after the turn had actually finished.
• Version number synchronised across the desktop app, the AI backend and the documentation.

0.0.5-alpha.1 superseded

Fixes to the agent turn lifecycle and to profile handling. All of these ship unchanged in 0.0.6-alpha.2.

• Context injection. Pressing Enter while a turn is running now injects your message into that running turn as context, rather than requiring separate Inject or Queue buttons.
• Stop button. Cancelling a turn now cancels the turn specifically, instead of closing the underlying connection.
• Long-path fallback. Agent turns no longer crash on Windows when a path exceeds the 260-character limit.
• Marketplace install guard. An install is only treated as successful when the install stream actually reports success.
• Work profile routing. Coding-plan entries carry per-profile metadata, so a Work profile routes to the Work sandbox rather than the default one.
• Claude Code CLI account management. Inline sign-in, with a tombstone written when a profile is locked.

Earlier alphas 0.0.1 – 0.0.5

The earlier alpha builds established the feature surface described on the features section: the Monaco editor with Python language-server support, the PTY-backed terminal, the git panel, agent mode with its four permission levels, local model support through Ollama, the hardware-fit Marketplace, model routing per task, and the cost meter and session cap.

• In-app sign-in replaced a sign-in flow that opened a separate console window and did not work reliably.
• Fixed a repeating freeze caused by profile cleanup running on the main process loop.
• Fixed a profile deletion failure that left a locked profile undeleted.

These builds are superseded and are not published for download.