v2.3.2

Important

Android APK notice: The APK is a Termux bootstrap + WebView shell, not a native Android server build. It opens an already-running local Marinara server, and on first launch it can download Termux from F-Droid, hand it to Android's installer, and start Marinara through Termux after Android permission prompts. Follow the Android wrapper guide if Android blocks the bootstrap handoff.

Added

• Added capability API 1.3 host services for downloadable packages, including safe model routing, persistence, history/checkpoints, resources, logging, transactions, client contribution loading, and visible readiness/retry states (#3690).
• Added reusable Game setup exports and imports. Initial Game Setup downloads now produce versioned .marinara-game-setup.json bundles that can refill the New Game wizard, remap available local resources, warn about missing ones, and require a replacement GM connection when necessary (#3701).
• Added Venice.ai as an Image Generation service, including authenticated live image-model discovery and native /image/generate support with model-aware sizing and validated base64 responses (#3682).
• Added responsive Background library folders, desktop and touch drag-and-drop organization, A-Z/Z-A/Newest/Oldest sorting, and collapsible tag filters without limiting the Background agent's available choices (#3678).
• Added Conversation, Roleplay, and Game compatibility badges to Download Agents, including catalog search by supported mode (#3676).

Changed

• Synchronized the stable release identity as v2.3.2 across the Engine, PWA manifest, Windows installer, Android bootstrap APK, update checks, Home release link, and What's New announcement. Android uses versionName 2.3.2 with versionCode 37 so it can update over every previously published APK.
• Moved Conversation About Me drafting from per-editor AI Write controls to Professor Mari. Character and Persona Convo editors no longer expose a separate model connection or source settings; Professor Mari can inspect a saved character or persona, write the bio in their voice, and save it directly to the real About Me field.
• Changed Noodle refreshes to choose active participants before first-time profile generation, skip characters that already have generated profiles, and send only the selected character cards to the timeline model. World/lore context and chat carryover now each have an 8,192-token budget so large invited rosters cannot inflate unrelated generations.
• Renamed current user-facing Conversation Calls references to Calls while preserving package IDs and legacy compatibility symbols (#3676).
• Moved Hierarchical Maps controls inside its active entry in Chat Settings → Agents instead of displaying a separate top-level settings section (#3679).

Fixed

• Kept new Roleplay chats Persona-less when the user selects no Persona, including the first optimistic message snapshot, provider prompt, scene generation, and combat context; only Conversation mode may fall back to the globally active Persona.
• Removed Roleplay Summary's 2,000-character truncation for each source message and the 64 KiB compiled-summary ceiling in chats.json, so chapter-length messages and accumulated summaries remain complete unless the selected model rejects the request.
• Preserved greetings, example dialogue, creator notes, system prompts, post-history instructions, character versions, and alternate greetings when Professor Mari or another app-data caller updates an unrelated Character Card field (#3708).
• Restored native undo and redo in shared text editors, including lorebook Content and Description fields, and made Tab / Shift+Tab indent or unindent every selected line without replacing the selection.
• Resolved {{user}}, {{char}}, and other prompt macros in opening greetings and /guided instructions at the final provider boundary, including lorebook routing and embedding scans, so late prompt injections cannot send raw identity placeholders to models (#3704).
• Raised Conversation routine-summary generations from a 512-token ceiling to an 8,192-token default, honored larger Connection overrides, and requested low reasoning effort so reasoning models still return visible summaries (#3696).
• Allowed selected custom agents to run in Conversation, Roleplay, and Game chats whenever Enable Agents is on, while preserving per-mode availability rules for official packages (#3692).
• Applied saved Connection Custom Parameters to every API-backed text generation that uses that connection, including Noodle and custom endpoints hosted locally, while preserving per-chat and per-call overrides.
• Routed bare Cohere API roots and versioned API URLs through Cohere's OpenAI-compatible endpoint.
• Kept the Noodle Carryover mode buttons equal-width while scaling their labels to remain fully visible with consistent spacing before each checkbox.
• Made installed Conversation games appear immediately in slash-command autocomplete and activate or deactivate their command and runtime contributions without an Engine restart, while route-bearing packages retain their safe restart path (#3699).
• Improved tracker-toolbar keyboard navigation with stable focus order, arrow/Home/End movement, and focus restoration when Escape closes the toolbar.
• Reported Agent catalog HTTP failures with their actual status and error instead of diagnosing every failure as an internet outage, while preserving installed agents offline (#3706, #3707).
• Made @handle mentions in Noodle replies open the referenced profile, and resolved active Persona, Character, and linked lorebook macros before Noodle refresh prompts reach the model (#3687).
• Let Connection Custom Parameters preserve arbitrary JSON values and convenient bare string values instead of discarding invalid drafts, and restored unified reasoning-effort requests for dynamically discovered OpenRouter models (#3688).
• Fixed the downloadable Conversation Calls package persisting a hardcoded assistant reply when typed-message generation failed. Package v1.0.4 no longer requires provider-native JSON mode and reports genuine provider failures instead of inventing character dialogue (#3685).
• Made native profile imports atomic across file-storage rows and assets. Present archive assets are now fully decompressed, CRC-checked, and staged before live mutation; table upserts run in a serialized transaction; and promoted files roll back if a later write or durable database flush fails, while missing assets remain warning-only (#3683).
• Fixed generated and manually replaced Game Journal NPC portraits failing to update when a reputation label was appended to the NPC name. Portrait matching now treats the display label and the tracked NPC as the same character throughout the live session (#3681).
• Increased the final Game setup generation watchdog from 300 to 500 seconds so large GM blueprints have enough time to finish without extending unrelated in-session generation limits (#3684).
• Fixed merged Roleplay group prompts stripping every historical speaker-tag example. The latest assistant message now keeps its <speaker> wrappers while older tags are still trimmed, and the instruction remains inside <output_format> when available or otherwise appends to the last user message (#3673).
• Preserved the Enable Agents master switch during the v2.3 capability migration, so upgrading cannot silently reactivate selected agents or their model calls. Hierarchical Maps now remains independently available when selected (#3669).
• Made the v2.3 capability migration restart-safe by writing its completion marker only after per-chat selections are migrated and flushed to durable storage; interrupted migrations retry idempotently on the next startup (#3670).
• Fixed Conversation Calls failing to recognize downloaded Local Whisper models when DATA_DIR was not explicitly configured. Downloaded capability runtimes now inherit Engine's resolved host data directory instead of deriving a private nested model path (#3671).
• Fixed Professor Mari returning blank character or lorebook generation turns, normalized common character-card field names, and made lorebook creation save generated entries atomically (#3674).
• Fixed Tic-Tac-Toe failing to render when an installed legacy game client expects React on the global scope (#3675).
• Kept cropped Character and Persona avatars contained inside the Colors message preview, including cards with additional sprites, instead of letting the preview image cover the editor (#3678).
• Kept the Roleplay-default star anchored in one place when its selected state changes (#3678).
• Vertically centered Chibi Professor Mari in her surprise-visit toast (#3678).