🚀 New Features
- Audios studio on desktop — the Audios generation studio is now available in desktop releases and runs against the core and the unified catalog so you can generate audio clips without a dev build.
- Audio model tiles and catalog integration — audio models now show a logo, short description and license in the picker, and the app reads audio models from the unified catalog so the studio and downloads use the canonical catalog entries.
- Start local server with last-served model at launch — when "Start automatically" is enabled the Local API Server now loads the model it last served at launch so agents and tools see the expected model immediately.
🔧 Improvements & Fixes
- Composer Send behavior — the Send button now clearly names the model it will use and avoids auto-sending when no model resolves, preventing accidental sends; composer icon buttons now have consistent hit targets and aligned layout.
- Turn error card — chat turn errors are now readable with a collapsible Details region and preserve the same cause-and-action pattern, and settings-triggered restarts are recorded so you can see why a restart happened.
- Recover dropped local turns — a local engine that drops mid-stream now triggers one automatic "restarting and retrying" attempt per turn (preserving finished tool results) so failed local turns are retried when safe.
- Keep first turn after model load alive — fixes to model start/hold logic prevent requests from being orphaned when a model is loading and keep queued sends and composer text when selection changes.
- Agents: more reliable launch, config and disconnect — agent detection, installers and Connect/Disconnect are hardened (better installer stdout/stderr handling, Windows execution policy for PowerShell installers, inline environment for launches), writers now record what they overwrite so Disconnect can restore previous keys, and many agents (Copilot, Codex, Kilo, Goose, Zed, OpenClaw, Hermes, etc.) receive correct served-window and disconnect behavior.
- Agents: only count a config as ours on the server port — agent detection now associates environment URLs with the Local API Server port to avoid adopting unrelated local/LAN endpoints.
- Copilot and Codex launch fixes — the Launch page starts Copilot on the model recorded in providers.json and Codex now shows when an API key is required.
- Downloads: clearer heading and smarter disk handling — the Downloads panel always shows a neutral "Downloads" heading; downloads now measure free space at the destination (including APFS purgeable space on macOS), retry locked renames on Windows, refuse FAT32-overflow files up-front and tag oversized-file errors as disk_fs_limit.
- Downloads: resume and planning fixes — model downloads now plan a fresh download when no recorded plan matches and will only resume a saved plan when it matches the model revision and file identity, preventing cross-revision resume corruption.
- Images model downloads — the image model picker can plan and start a first download on a fresh or returning profile instead of rejecting it.
- Webview storage per profile — isolated profiles now keep WebView/WebView2 storage inside ATOMIC_CHAT_PROFILE_DIR so projects, onboarding and frontend settings no longer leak between profiles.
- Windows: registry PATH editing and close-to-tray behavior — the CLI now edits HKCU\Environment directly (preserving non-ASCII and %VAR% bytes) and closing the window while a local stream or download runs hides to tray with a one-time notice instead of quitting.
- Telemetry: clearer failure classification — telemetry schema updated (3 -> 4) so local engine failures, network errors and retries are classified correctly, retries are tagged as auto_retry, and attempts and timing fields (phase, ttft_ms, total_duration_ms, chunk counts) are reported for better analytics.
- Loopback HTTP transport errors — loopback transport failures are typed/prefixed ([transport:]) so the UI and telemetry classify them reliably across OS-localized error messages.
- API inspector context window — the Activity/API inspector now shows the actual context window each request ran with (ctx_len and cached tokens) instead of looking up by model id alone.
- Onboarding copy & handoff downloads — onboarding no longer shows a decorative step counter and the model step heading reads "Choose a model", and the handoff panel no longer duplicates download cards after setup.
- UX polish and robustness — the audio player avoids spinner flashes when switching clips, Activity durations round correctly before splitting minutes/seconds, Downloads waits for initial disk scan before showing empty states, a chat opened in a new window preserves its model/provider, and many cases where UI state could be lost (queued sends, stopped models, backend-mismatch prompts) are handled safely.
- 20+ stability, installer and UI fixes — a broad set of fixes across installers, integration detection, agent writers, downloads, engines, and the desktop shell improve reliability on Windows, macOS and Linux.
⚠️ Breaking Changes
- None in the app surface for end users.
🙏 Contributors
Thanks to @Vect0rM, @Ooooze for their contributions to this release!