What's new.

Small UX fixes and a provider correctness bug.

• Sticky auto-scroll while a reply is streaming; pauses if you scroll up, resumes at the bottom.
• Ctrl+L focuses the composer from anywhere in the app.
• Ollama runtime context length now picks up Modelfile num_ctx and per-architecture context_length keys (llama.*, qwen2.*, etc.) instead of a fixed field.
• Deleting a chat no longer crashes the webview — switched off window.confirm (unsupported in Tauri) to the dialog plugin's ask().

System tray, configurable embedding chunk size, and quick-window polish.

• App lives in the system tray and keeps running when the main window is closed. Tray menu: Show/Hide Main Window, Show/Hide Quick Chat, Quit. The window's X hides to tray; Quit is the sole exit path.
• Configurable embedding chunk size (embedding_chunk_chars, default 1500, range 200–8000) for models with tight context windows. Exposed in the Embedding tab. Reindex required after changing.
• Quick window centers on the cursor's monitor whether you trigger it via Ctrl+Shift+Space or the tray menu, and won't get flipped by a focus-stealing tray click.

Light theme, side-chrome polish, and signed Linux builds.

• Light theme with a Sun/Moon toggle in the NavRail; both windows stay in sync via the shared pkma-theme key.
• Recessed side chrome (rail/history/titlebar) on --bg-side; titlebar simplified to just the vault name.
• Signed Linux Tauri build (AppImage + .deb / .rpm) and multi-platform versions.json/latest.json for the updater.
• Release notes now sourced from changelogs/v<version>.md.

First public build. Everything is provisional and the index format is not yet stable.

• Local-first chat against an Obsidian vault.
• Wikilink-aware retrieval with inline citations.
• OpenAI-compatible provider support (OpenAI, Anthropic, Google, Ollama, llama.cpp).
• Three agent modes: Quick, Deep, Research.
• Windows binary. macOS and Linux builds tracked in the roadmap.