v1.16

## Release Notes

### Speech stack refactored into independent plugins

The legacy core speech system has been split into two built-in plugins — **Kokoro TTS** and **Whisper STT** — that can be enabled or toggled independently. Browser-native TTS remains available as a fallback when Kokoro is off, and the Whisper mic UI only appears when its plugin is enabled. Legacy speech settings and APIs have been removed in favor of engine-agnostic service brokers.

### Timezone and time display preferences

A new user-configurable timezone and 12/24-hour clock preference flows through the scheduler, notifications, backups, memory timestamps, and all frontend time displays. UTC remains the internal boundary for absolute instants; user-facing dates render in the selected timezone.

### Screenshot and capture privacy improvements

- Browser, Desktop, and computer-use screenshots are now **ephemeral and context-scoped** by default — routed through in-process image refs consumed by vision payloads instead of being written to disk.
- Explicit user-initiated screenshot paths remain durable.

### Model preset handling

Deep-merging model preset slots with the active configuration now preserves custom context windows, rate limits, and nested kwargs across preset switches. Omitted utility and embedding slots inherit from the currently configured models.

### Plugin system improvements

- **Scanner false-positive reduction** — scanner prompts are calibrated around demonstrated risk rather than the mere presence of common capabilities like network calls or filesystem access.
- **Simplified activation toggle** — the plugin list now uses a one-click ON/OFF switch instead of the previous dropdown.
- **Configurable error retry count** — the `_error_retry` plugin reads its retry limit from settings instead of using a hardcoded single retry.
- **Skill visibility controls** — users can hide skills from the model-facing catalog via the chat Skills selector; hidden skills are filtered from listing, search, recall, and prompt injection.
- **Connector skill activation** — a new `skills_activate` endpoint lets connector clients activate skills in live chats.

### Editor and document artifact changes

- Text Editor now owns Markdown saves, canvas-open intent, refresh, and stale-save protection.
- Office artifacts are renamed from `document_artifact` and remain Desktop-only with Office formats.
- File Browser routes Markdown files to Editor, txt/Office documents to Desktop, and browser-renderable files to Browser.
- Root-level Markdown files can be opened in Editor from the File Browser.

### Surface and UI improvements

- Unified header action rail across Browser, Desktop, and Editor floating modals with consistent ordering and separators.
- Icon markers now render as Material Symbols in composer progress overlays.
- Browser screenshots open in the image viewer; live previews still open the Browser surface.
- Codex OAuth setup UI gains Main/Utility model selectors and account quota usage bars on the dashboard.

### Browser runtime recovery

Clearer settings and recovery guidance when the host-browser runtime fails, directing users back to the internal Docker browser.

### Fixes

- Fixed message queue persistence after auto-send.
- WhatsApp connector now includes the sender number in user messages.
- Updated Computer Use rearm guidance to reference the new `/computer-use on` CLI access model.

### Breaking changes

- **Legacy speech settings and APIs removed.** TTS and STT are now managed through the `_kokoro_tts` and `_whisper_stt` built-in plugins. Existing speech configuration will not carry over automatically.
- **`document_artifact` renamed to `office_artifact`.** Retired shims and facades have been removed; skills and integrations referencing the old name will need updating.