v1.5.7

Added

• Guide for Game mode.
• Professor Mari can now create Lorebooks for you.
• Days tracker in Game mode that you can edit.
• Lorebook entry trigger mode can now be changed directly from the entry status dot.
• Game mode interrupt button that allows you to interrupt the GM (with or without consequences to your game).
• Various improvements to the Game mode's combat and inventory systems, more cinematic battles, better UI handling, and more overall mechanics.
• Game mode map scaling.
• New permanent tag that persists in Roleplay mode if a character passed you important information in Conversation mode.
• Improvements to the Knowledge Router agent.
• Storing the Conversation Theme background gradient separately for dark and light color schemes, so switching OS/browser theme automatically loads the correct gradient.
• Custom agents now have a chat memory.
• Prompt overrides the registry for image generation.
• Active filter tab in Lorebooks.
• Compressed Lorebooks.
• Customizable generation settings for local image generation.
• When generating schedules, they now receive context from the conversation chats you had with a character.
• Hide/unhide messages in Roleplay mode.
• Alternative display of logs for Game mode.
• Custom agents can now choose a result type, including Text Rewrite for post-processing agents that edit the generated reply.
• Setting to enable showing and editing image prompts before they're sent.
• Setting to change the image dimensions for generation.
• Various small QOL changes.
• Custom agents' outputs can now be edited in the Agents button in the Roleplay mode.
• Custom parameters field.
• Sliders to control the sprite's size and opacity in Roleplay mode.
• Custom activity statuses for the user.
• Vectorized Lorebook entries are now visibly marked.
• Character card version history with compare and restore controls.
• Prefills.
• File-backed storage is now the default: legacy SQLite data is imported into JSON files under DATA_DIR/storage, backups include those files, and STORAGE_BACKEND=sqlite remains as an advanced compatibility escape hatch.
• Allowed token size outputs in agents.
• Lorebook folders.
• Game mode setup remembers custom genre, tone, setting, and goal options from previous games.
• Optional trimming for incomplete model endings before generated messages are saved.
• Draft translation button option in chat Translation settings for Conversation, Roleplay, and Game modes.
• Native vs compatible export choices for profile, character, persona, and lorebook exports.
• PocketTTS is now available as a local TTS provider.
• Optional speech-to-text microphone buttons can be enabled for Conversation, Roleplay, and Game input fields.
• Character imports now ask before extracting embedded character-card lorebooks into standalone Marinara lorebooks.
• Home Assistant HACS integration that syncs Marinara custom tools and a Home Assistant agent for smart-home control.
• Updated the supported toolchain to Node.js 24 LTS and pnpm 10.33.2 across launchers, installers, Docker images, docs, and CI, plus refreshed dependencies within their compatible ranges.
• Lorebook entries can now be scoped by active characters, character tags, and generation triggers, and can scan selected character/persona fields as extra keyword-matching sources.
• Game mode now has an optional Lorebook Keeper that updates a game-scoped lorebook after session conclusion and automatically attaches it to that game.

Security

• Hardened default network access so loopback remains convenient while non-loopback private-network traffic fails closed unless Basic Auth, an allowlist, or an explicit unsafe opt-in is configured.
• Added global unsafe-method CSRF/origin protections, security headers, route throttling, and shared privileged-route gates for admin, update, backup/import, sidecar, haptics, and custom-tool operations.
• Added SSRF, path containment, upload validation, bulk-import capability tokens, and response-size guards around high-risk URL, file, and archive flows.
• Disabled or gated risky execution paths by default, including API-driven update apply, custom script tools, sidecar runtime installs, and remote haptic control.
• Removed the seeded default provider key, encrypted Spotify token storage, and redacted obvious secrets from profile export.
• Hardened chat HTML sanitization and SVG/image handling, then upgraded vulnerable production and build dependencies.
• Hardened Docker, Android WebView/backup, GitHub Actions action references, and Windows installer dependency verification.
• Breaking/default changes: privileged routes now require ADMIN_SECRET, Docker binds to localhost by default, and update apply, custom script tools, and sidecar runtime installs are disabled until operators opt in with the documented environment switches.
• Operators who intentionally need the old exposure model must set ADMIN_SECRET, choose a remote bind address for Docker/launchers, and explicitly enable only the required flows such as UPDATES_APPLY_ENABLED, CUSTOM_TOOL_SCRIPT_ENABLED, or SIDECAR_RUNTIME_INSTALL_ENABLED.

Fixed

• Addressed various security concerns.
• Game mode dark screen error addressed.
• Removed the persistent SQLite database as the default live storage path, reducing release-to-release migration failures.
• File-backed migration now merges every known legacy database location and performs a one-time repair for snapshots that missed chats during early v1.5.7 testing.
• On mobile Roleplay, the branch quick-switcher now lives inside the three-dot toolbar menu, so it no longer overlaps the Agents' controls.
• Settings Debug Mode now prints prompt, scene-analysis, party-turn, and game asset debug logs even when LOG_LEVEL is not set to debug.
• Switching chats doesn't stop the generation of the previously triggered one.
• Cross-conversations confusions addressed.
• {{user}} and {{char}} macros now work in all modes.
• Injections at a specific depth now work correctly.
• Added Spotify OAuth redirect URI handling and manual paste-back.
• [Start the game] is being sent twice upon starting the game.
• Expression Engine now retrieves all the available sprites correctly upon retry.
• Fixed unstable message pagination cursor.
• Various errors were addressed.
• Advanced parameters are now respected by local endpoints.
• Improved the quality of some prompts.
• Ensured the daily/weekly summaries trigger consistently.
• We now handle assets in Game mode better.
• Conversation mode characters no longer reply to themselves; instead, they reply to you.
• Drag-and-drop on mobiles now works.
• Custom agents can now rewrite your messages.
• Full-body sprites in game mode now get updated properly.
• Deleted characters from group chat no longer appear as Unknown.
• Roleplay setup and connection setup dialogs now fit short screens with internal scrolling, and Custom Parameters starts empty with an example placeholder.
• File-backed storage now supports Lorebook folders during generation and migration.
• Deleting one saved character card version now leaves the rest of the version history intact.
• Removed the legacy database setup step from the installer flow.
• Fresh installs no longer install the old better-sqlite3 or sql.js SQLite fallback packages.
• Browser-tab character imports now preserve embedded Chub lorebooks as linked Marinara lorebooks.
• OpenRouter Claude reasoning is requested with OpenRouter's unified reasoning payload again, restoring thinking capture for Sonnet/Opus reasoning models.
• Sprite sheet prompts now more explicitly require complete slicable grids for expression and full-body pose generation.
• Loopback LLM provider URLs are allowed by default again, so local model servers on 127.0.0.1, ::1, or localhost do not require the broad private-network URL opt-in.
• Restored the animated Marinara logo on the home screen while keeping the static logo as the inactive-page fallback.
• Tightened the home screen spacing so the logo, FAQ, credits, and special thanks fit more comfortably on desktop and mobile.
• Windows installer updates now force-refresh the release tag and verify the resolved tag commit instead of aborting on legitimate v1.5.7 hotfix retags.
• The v1.5.7 Android wrapper APK now uses a bumped versionCode for hotfix updates and the release workflow uploads an installable sideload APK.
• Game Lorebook Keeper now continues in the background after a session is concluded instead of holding the End Session response open.
• Launchers, installers, and in-app updates now fall back to installed or temporary pnpm when Corepack cannot resolve the exact pinned pnpm patch version.
• Explicit ComfyUI and AUTOMATIC1111 image-generation connections can use LAN/private-network hosts without the broad image URL opt-in.
• Restored scoped HTML/CSS rendering inside Roleplay messages and narrator bubbles.
• Backup and profile export failures now surface the specific server/admin-secret error instead of a generic failure toast.
• Haptic agent position commands now normalize PositionWithDuration-style outputs and continue executing later commands if one device command fails.
• Lorebook entry drawers now autosave edits, so the manual Save Entry button is no longer needed.
• Docker/LAN browser origins now pass CSRF checks when Marinara is reached through a mapped host port, and CSRF_TRUSTED_ORIGINS=* is honored as an explicit unsafe wildcard.
• Loopback backup/profile export requests no longer require ADMIN_SECRET by default; remote privileged requests still do.
• Turning off Conversation schedules now clears saved schedule metadata and resets affected character availability state.
• Removed the Workbox index.html navigation fallback that caused non-precached-url console noise.
• Various minor UI bugs.