Added
- General settings now include a persisted app-language selector at the top of the tab. It currently exposes English only and is ready for future translation PRs to extend.
- Added a new option to display character/persona avatars in the Roleplay mode (as a side panel, bigger size). Access it in the Appearance Settings.
- NanoGPT support and improved image connection handling.
- Added a macOS Apple Silicon-only MLX backend for the local sidecar.
- Support for running different local models.
Fixed
- Installed Windows desktop and Start Menu shortcuts now launch Marinara Engine with the correct working directory, so packaged installs no longer open and close immediately.
- Windows installers and launchers now force the repo-pinned pnpm version through Corepack when available, so older global pnpm installs no longer break setup, and the batch installer restores the Marinara icon on the desktop shortcut.
- Conversation mode no longer forces OpenAI-compatible backends like NovelAI onto the non-streaming transport path, preventing immediate cancellations while keeping complete-message rendering in the UI.
- Character maker, persona maker, lorebook maker, prompt review, retry-agents, game setup, and other system tasks now obey the global Streaming Responses toggle instead of silently forcing streamed transport.
- Image Generation connections can now keep ComfyUI selected on non-default hosts and ports, so remote ComfyUI servers still expose checkpoint fetching and custom workflow JSON.
- Connection max-context limits now trim oversized prompts before generation and prompt inspection shows the fitted prompt that was actually sent upstream.
- OpenRouter connection provider preferences now carry through agent runs, game setup, GM/tool generations, and other helper flows instead of falling back to Auto router outside the main chat path.
- Inline reasoning blocks wrapped in
<thought>...</thought>or<|think|>...<|/think|>are now extracted into stored message thoughts, and game-mode JSON helpers strip those blocks before parsing model output. - Glued Side Panel roleplay avatars now fade and blur out more aggressively at the bottom so they merge into the message bubble instead of ending abruptly.
- Clean installs no longer warn that pnpm ignored build scripts for
onnxruntime-nodeandprotobufjs, so Windows users do not need to runpnpm approve-buildsor patchpackage.jsonby hand. - Added the no split mode flag to prevent the looping crash of Gemma-4 on multiple GPUs systems.
- Tracker agents can now use the built-in local sidecar through the normal Connection Override dropdown, and the Local Model card now provides a bulk action to point every built-in tracker at the local model.