Pasta-Devs/Marinara-Engine v2.4.2 on GitHub

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 the v2.4.2 in-app What's New message with bundled images and looping videos so its update overview remains available without external media requests.
Let Game setup enable Illustrator's Dynamic LLM Prompt Generation before the first image request (#4964).
Added always-visible activation and vectorization status to lorebook entries in Professor Mari's Easy Viewer review cards: each entry now shows whether it is Constant, Selective, or Normal (using the lorebook editor's dot colors) and whether it is Vectorized, Not vectorized, or Vector excluded, with the before and after shown when Mari's edit changes them, so an entry's kind is clear at a glance instead of only when a setting happens to change (#4931).
Added a per-entry Reject control to Professor Mari's Easy Viewer: revert a single lorebook entry of a multi-entry change while keeping the rest applied, instead of only being able to Restore the whole change. Rejecting is refused for anything that could leave a dangling reference (an entry removed as part of deleting its lorebook, or a non-entry row), and the review card shrinks to the entries that remain (#4931).
Added a "View as prompt" preview to Professor Mari's Easy Viewer for character and preset edits: a before/after view of how the edit appears in the assembled prompt, with additions in green and removals in red. It is a synthetic preview assembled on its own (default preset, no persona or chat history), not a live chat prompt (#4931).
Added literal message search to Conversation and Roleplay chat headers, with compact results that jump to matching messages anywhere in the current chat (#4922).
Added a Roleplay Agents-menu stop control that appears only while the current chat has an active generation and can cancel server-side work that survived navigation or a reload (#4942).
Added a persisted Send on Enter toggle for Professor Mari and a reusable setup download on New Game's final step before the game starts (#4928, #4930).
Added Markdown previews to Conversation, Game, and section content in the preset editor, matching character and lorebook authoring (#4916).
Added Discord-style __underline__ and -# subtext rendering throughout chat history and other shared Markdown surfaces (#4914).
Added an Easy Viewer to Professor Mari's Keep/Restore review cards: instead of raw app_data JSON, her edits now show as a readable before/after diff with removed text in red and added text in green, lorebook entries laid out like the entry editor (name, keys, content, toggles), and a per-change Dismiss control to hide reviewed items and reduce clutter. Deletions and disables are flagged prominently since they are the most consequential, and you can toggle each card between Easy and Raw independently (#4919).
Let Professor Mari set every user-editable lorebook entry setting when she creates or edits entries, not just the keyword-matching controls: activation chance (probability, clamped 0-100), timing (sticky, cooldown, delay, ephemeral), recursion (preventRecursion, excludeRecursion, delayUntilRecursion), inclusion-group weight, per-entry scan depth, lock, folder placement, character/tag/trigger matching filters, and per-entry vectorization (which she only enables when an embedding model is configured) (#4791).
Added shared and per-character New Start context markers to group Roleplay, with an avatar target picker and character-colored dividers that let newly introduced characters begin from a later message without truncating the rest of the cast's history (#4905).
Added character-targeted Roleplay message hiding through /hide <character> <number/range>, including quoted names and the existing single, range, and comma-separated message selectors (#4906).
Added a search box, collapsible entries, and sorting to the Professor Mari Skills and Memories panels. Each entry is now a drawer that expands in place to edit; its description shows on wider screens and collapses to just the name on small screens; both panels can be sorted by name (A to Z or Z to A), newest, or oldest; persistent memories are pinned to the top and marked with a star that stays visible while collapsed; and the open editor stays in view so saving or toggling a row no longer scrolls it away (#4868).
Added one-click support diagnostics that copy the current version, build, platform, graphics adapter, and active text model for issue reports (#4878).
Added SwarmUI as a video-generation backend, including authenticated distributed ComfyUI workflow submission, base64 reference images, model discovery, and MP4 retrieval (#4885).
Added responsive batch selection to chat Galleries, with filtered Select All, confirmed multi-image downloads and deletes, and selection cleanup between chats (#4859).
Added optional image prompting instructions to image connections, applying them inside existing selfie and Illustrator prompt-writing calls before provider review or generation.
Added a Gallery image-agent picker that can run any active custom image-producing agent alongside the base Illustrator (#4846).
Added an editable Storyboard Agent shot-planner stage that inspects each generated keyframe before video generation, persists its suitability classification, and falls back to the planned motion when image-aware refinement is unavailable or invalid. The Storyboard Agent page now explains the four-stage prompt workflow and orders its shared prompt editors from illustration through image-aware grounding to video generation (#4839, Pasta-Devs/Marinara-Agents#296).
Added batch selection to Character and Persona image galleries so selected images can be downloaded or deleted together after confirmation (#4832).
Added a persisted Auto, Low, Medium, or High output-quality choice to GPT Image generation connections (#4831).
Added optional character and persona reference sheets under Sprites, with upload, size-bounded AI creation, safe selection cleanup, and an explicit generation-reference toggle that falls back to existing likeness art; both Galleries include the same AI creation entrypoint (#4786).
Gave Professor Mari granular preset editing: she can now read a preset's individual sections, groups, and choice-blocks with content previews and add, update, or delete any one of them in place through the Keep/Restore review, instead of only creating or replacing a whole preset — also exposed through a new mari presets CLI (#4812).
Gave Professor Mari a persistent, retrievable memory: standing preferences and behavior directives you save (or ask her to remember) that she retrieves index-and-fetch: only a lightweight title/description index of your saved memories stays in her prompt every turn (capped by both entry count and total size, and paged so a large store never truncates ids out of reach) and she pulls a memory's full text when it is relevant, so the store stays token-cheap as it grows. Memories marked Persistent inject their full body every turn (charged in full against one budget), so those are meant to be few and small. Saved memories take precedence over her default behavior where they conflict (for example, opting back into edit-without-asking). Every memory she writes goes through the Keep/Restore review and lands disabled until you turn it on (with the review card's Keep & Enable action or the new Memories panel), so nothing she reads can quietly activate a standing instruction, and a dedicated Memories manager lets you create, edit, enable, and pin them directly (#4851).
Let Professor Mari propose safe data-only custom Home widgets for explicit confirmation, with persistent responsive cards that users can reorder, hide, restore, and delete from the Widgets manager (#4801).
Added the Please Handle With Care achievement for dragging the Home navigation Professor Mari around the screen (#4805).
Rebuilt Home as a full-width, Firefox-inspired Marinara browser with theme-matched chrome, generated bookmark and tab artwork, a responsive draggable and hideable widget desk with equal spatial slots, space-aware large-widget packing, fine-pointer hover lift and illumination, and a focused five-widget first-run composition, Community shortcuts, compact mobile chrome with fitted destination tabs and a Bookmarks menu, a live timezone-aware cyan-accented clock and calendar, visit-ordered mode-colored recent chats with band-free scene backgrounds, expression sprites, and a fitted three-card mobile view, latest-unlock and nearest-goal achievement previews with distinct shelf and launcher copy, a generated trophy bookmark that follows the master Achievements setting, separate FAQ and widget-manager windows, a highlighted Home onboarding tour, mobile-fitted Guide and Character of the Day cards, and an optional Professor Mari navigator that finds Engine destinations and named resource editors locally without AI, remembers its bounded desktop position, and dangles by her hoodie in a generated pixel animation while dragged (#4763).
Moved Noodle into an optional downloadable package that keeps its familiar interface in a Home tab with its paired Noodle/NoodleR logo, blue/pink mobile navigation, and a persistent new-refresh badge, automatically preserves availability for existing profiles, follows the staging Agent catalog during feature review, and leaves timeline data ready for reinstall after removal (#4763, Pasta-Devs/Marinara-Agents#278).
Added Arli AI as a built-in text provider for chat connections, using OpenAI-compatible chat completions with models loaded from the Arli AI /models endpoint.
Added a custom 1–40 place target to the New Game wizard's AI world-map draft options (#4783).
Added a compact context-budget indicator to Professor Mari Workspace when message token usage is enabled (#4752).
Added per-chat image style overrides for custom image agents, alongside their existing connection overrides (#4718).
Added attributed marinara.fetch traffic totals and sustained-rate warnings for full-page extensions (#4720).
Added SwarmUI image generation, optional account-token authentication, model discovery, and ComfyUI workflow submission through SwarmUI's distributed generation queue (#4702).
Added an accent-colored divider above the current Roleplay new-start boundary (#4709).
Guided package onboarding can open the active Roleplay chat's Summary popover and assigned prompt preset Sections editor directly.
Added explicit step-by-step and immediate World Maps travel modes, with one committed movement per accepted Roleplay or Game turn and recoverable queued routes (#4618).
Added a single setting that reduces ambient animations and effects throughout the interface, including automatic support for the operating system's reduced-motion preference (#4631).
Added per-chat image-connection overrides and an on-demand snapshot button for image-generating custom agents (#4686).
Added a downgrade guard that blocks launches and updates onto an older build once chat data has been sharded, plus an offline unshard escape hatch to recover access to chat history (#4738).
Added a one-time notice explaining the storage migration after upgrading, with a plain-language summary, a technical-details toggle, and a link to the storage troubleshooting guide (#4762).
Added tiered exact, substring, and semantic resolution to Professor Mari's fetch command, letting her find characters, personas, chats, and lorebooks by partial name or description instead of an exact match only (#4778).
Added lorebook-authoring guidance for workspace Professor Mari, exposing selective, whole-word, case-sensitive, and regex matching fields and walking her through a user-gated fidelity review so generated entries are no longer skeletal (#4796).
Added mari CLI flags for lorebook entry secondary keys, selective matching, selective logic, whole-word matching, case sensitivity, and regex on add-entry and update-entry, matching the app-data entry editor (#4811).
Added lorebook-authoring strategy guidance, a worked multi-control example, and macro/recursion documentation to the lorebook entries guide (#4814).
Made Professor Mari's Keep/Restore undo durable across restarts, persisting pending reviews to disk with a 14-day retention window and a 50-item cap (#4828).
Added optional, quota-limited synthetic audience activity for public NoodleR posts, with separate likes, replies, reposts, and a manual refresh action.
Added a persisted local-day NoodleR fan activity plan with four deterministic platform runs and resumable activity application state.
Preserved creator-level fan audience inheritance when editing individual archetype weights, and surfaced source-action failures in NoodleR profile controls.
Added a configurable automatic audience-run count, kept manual audience runs available outside that daily budget, and prevented malformed generated activities from failing an entire run.
Exposed Max Parallel Agent Jobs in local-model runtime settings and used it for agent scheduling and llama-server parallel slots while retaining the configured context budget per request (#4609).

Changed

Expanded profile-import previews and quarantine to cover executable tools, personal extensions, persistent Professor Mari memories, and active themes, preserving their contents for deliberate review and re-enabling.
Bound downloaded Agent installs and updates to the version and checksum the user reviewed, quarantined imported executable tools until review, and moved custom scripts plus Professor Mari transforms into terminable or OS-sandboxed workers while retaining an explicit reviewed-script fallback on unsupported systems.
Repeated local checks now reuse native TypeScript and ESLint caches for unchanged client files (#4938).
Manual Agent retries now resolve the same current settings, custom tools, and permitted built-in tools as normal Agent generation while retaining retry-specific historical context and Spotify playback verification (#4894).
Localized achievement definitions and official capability-package Home metadata through stable locale-aware contracts with English fallback, and refreshed Home's welcome copy (#4803, #4806).
Positioned Professor Mari's navigation helper at the bottom-left on her first appearance while preserving every remembered user placement.
Removed the standalone Browser tab and its tutorial step, moved card downloads into split Download/Open Library controls in Characters and Personas, and replaced browser-native card destination and embedded-lorebook prompts with a themed Marinara import flow.
Sharded stored messages and message swipes into per-chat files instead of rewriting a chat's entire history on every saved message, migrating existing installs automatically on first boot (#4735).
Sharded fourteen more chat-keyed tables (memory, agent runs, game state, calls, gallery, influences, and notes) into per-chat files, cutting write amplification on every turn (#4754).
Moved Home Professor Mari's available-name lists out of the system prompt into a per-turn message, stabilizing prompt caching and capping each category at 100 entries (#4771).
Ranked Home Professor Mari's available-names list by relevance to the current conversation instead of an alphabetical slice, falling back to alphabetical when relevance data isn't available (#4779).
Advanced the stable release identity to v2.4.2 across the Engine, PWA manifest, Windows installer, Android bootstrap APK, update checks, Home page, and release references. Android uses versionName 2.4.2 with versionCode 43 so it updates over every previously published APK (#4610).
Required AI-agent contributions to update the Unreleased changelog for every bug fix, behavior change, and new feature, keeping release notes aligned with the implementation that introduced each change (#4613).

Fixed

Included Noodle in Professor Mari's official 32-package catalog knowledge and corrected the current downloadable-package documentation for the v2.4.2 Agent catalog (#4992).
Kept version-tagged Docker release images on the stable Marinara Agents catalog instead of following staging after publication.
Kept dependency security updates pinned to patched resolutions and made the existing pnpm launcher handoff checks a required pull-request gate, protecting upgrades without forcing users onto a new package-manager major version (#4988).
Kept inline Game dialogue and group character macros responsive on malformed or very large generated text without limiting valid authored content (#4984).
Removed the 2 GiB application ceiling from automatic and downloadable full backups, using streamed ZIP64 archives so larger local libraries remain restorable without weakening safety limits for ordinary profile imports (#4982).
Hardened imported and model-authored content against pathological parsing inputs and reserved metadata keys without removing supported import formats or local tools.
Kept automatic and downloadable backups working when local media exceeds the profile importer's ordinary per-file limit, while preserving bounded streaming imports and omitting only an individual unreadable or unarchivable asset instead of failing the whole backup; any omissions are listed in RESTORE.txt and automatic-backup status.
Kept visible Android browsers connected to a same-device Termux server responsive with a lightweight loopback-only heartbeat, without restoring hidden-tab or remote-client polling (#4968).
Kept long-running SwarmUI image requests alive across idle container networking while retaining the configured ComfyUI generation deadline (#4969).
Restored Professor Mari preset creation while keeping Engine-owned preset identifiers protected from user input.
Updated Roleplay message-edit shortcut hints to mention Ctrl+Enter on Windows as well as Cmd+Enter on macOS (#4965).
Kept the Message Avatar Scale and related art-size sliders wide enough to remain visible and repeatedly adjustable in Chrome, including at larger interface display sizes (#4974).
Kept an already-installed Noodle Agent visible and usable after downloading an update while its replacement waits for the required Marinara Engine restart (#4976).
Prevented crafted profile archives from retargeting saved credentials, serving active files as trusted app content, escaping asset directories through imported metadata, reconstructing blocked theme CSS, or exhausting the importer with oversized metadata.
Kept Professor Mari suggestion chips in a single horizontally scrollable row on mobile instead of stacking them above the composer.
Hardened local security without removing extension or remote-access capabilities: made Professor Mari filters data-only, encrypted webhook credentials, bounded ZIP extraction and rich-chat network/CSS behavior, restricted automatic network trust to detected runtimes, privatized local data and backups, verified downloads and release commits, and patched audited dependencies.
Authenticated APK-managed Android localhost sessions without restricting manual Termux or LAN use, verified the pinned F-Droid Termux APK before install, bound native bridge calls to the exact Engine origin, and made release signing and bootstrap source commits fail closed.
Imported preset regex entries even when they contain unsupported SillyTavern placements, warning about ignored placement values instead of discarding the entire regex, and let Music DJ try the next candidate when Spotify accepts but cannot verify the first selected track (#4959, #4960).
Refreshed stale same-version browser clients when the server commit changes, so staging fixes such as NovelAI generation-setting saves and connection imports no longer remain hidden behind an older cached interface (#4957).
Kept the Termux-hosted server awake while the launcher is running and released the Android wake lock when it stops, preventing backgrounded Termux sessions from freezing until the terminal is reopened (#4956).
Kept Professor Mari's generated continuation suggestions available after opening and closing a lorebook (#4953).
Made SwarmUI image requests honor the configured ComfyUI timeout instead of inheriting a shorter transport timeout (#4954).
Kept canceled Professor Mari prompts out of the composer, made the first message edit persist, and retained retry for the latest response (#4947).
Preserved the complete provider-ready prompt produced by Game Mode's dynamic image prompt generator through prompt review and image submission (#4948).
Stopped Professor Mari's workspace verification guard from treating ordinary task-completion wording as an unverified data mutation (#4949).
Enforced each connection's Max Parallel Agent Jobs limit across separate post-processing Agent batches (#4950).
Removed deleted lorebooks from the active chat context immediately instead of leaving a stale cached attachment label (#4951).
Made turn-game bot move prompts visible when UI debug mode or DEBUG_AGENTS is enabled (follow-up to #4945).
Included custom world fields and persona inventory in Agent Suite tracker editing without replacing adjacent tracker data (#4937).
Exposed active custom image agents in Conversation Gallery's Illustrate picker, matching the existing Roleplay and Game action (#4940).
Resumed pending turn-game bot seats after a regular Conversation reply so UNO cannot remain frozen when a chat send wins the per-chat generation lock (#4941).
Made the Game minimap's zoom icons follow the selected interface accent instead of retaining a fixed color (#4944).
Matched embedded Long-Term Memory chat settings to the compact spacing, control sizing, typography, and surfaces used by other Agent menus (Pasta-Devs/Marinara-Agents#309).
Repaired common malformed JSON in Roleplay and Conversation summaries through the shared JSON-repair path instead of falling back to raw model output (#4925).
Kept NovelAI generation defaults from reverting when Save follows a newly committed setup field (#4929).
Cleared embedded character-book data and pointers when Professor Mari deletes a whole lorebook, then restored both to the actual embedding character on Restore, including empty and multi-linked books (#4932).
Preserved an enabled Spotify repeat setting when Music DJ replaces playback with a multi-song selection, looping the selected set as one context (#4934).
Kept a character's embedded lorebook in sync when Professor Mari edits it: adding, updating, or deleting an entry (or restoring one of those changes) in a lorebook embedded in a character card now rebuilds that character's copy, instead of leaving it stale as only her edits did before (#4927).
Gave Professor Mari a safe way to delete a single lorebook entry (lorebook.deleteEntry) and told her to use it instead of a raw mari db delete, which previously let her remove far more entries than intended when asked to delete just one (#4924).
Preserved unsent Home Professor Mari messages across editor navigation and app restarts until they are sent or explicitly cleared (#4915).
Taught workspace Professor Mari to revise an existing saved memory when asked, instead of declining that a matching memory "already exists": her guidance now includes a read-then-rewrite recipe and worked example for editing a memory's content in place (the same pattern she uses for preset sections), which also works on an enabled or persistent memory without turning it off (#4917).
Made Force To Call Tool serialize native required-tool controls for Google Gemini, Vertex AI, and compatible Anthropic requests, while safely falling back to automatic choice for manual Claude extended thinking and Mythos (#4907).
Fixed the open-issues regression failing on Windows checkouts: its chat-summary reorder check matched the entry container and the touch-reorder row within a fixed character distance that CRLF line endings pushed just over the limit, so it now verifies the shared desktop and touch reorder surface on the row element directly, independent of line endings and where the row is defined (#4911).
Restored Music DJ Spotify playback by clearing the previous repeat mode before replacing a multi-song context, reapplying repeat only after the selected first track is verified, and surfacing persistent URI mismatches as failures instead of false pending successes (#4903).
Made persona saves reliable: the editor now sends only the fields you actually changed, keeps edits made while a save is still running, prevents saves and avatar replacements from overlapping, blocks its own Back and Discard controls mid-write, and explains exactly which field a rejected save objected to — including an empty persona name.
Made Reduce Ambient Animations & Effects flatten costly backdrop blur on desktop as well as mobile, including shared Conversation overlays (#4897).
Applied embedded character-card fields when replacing an existing character's avatar, so updated card PNGs refresh the matching editor sections (#4896).
Kept automatic summaries chronologically ordered and hidden when requested, while allowing users to rearrange summary entries by drag and drop on desktop and mobile (#4891).
Clarified Professor Mari's character-card field guidance so requested backstory, appearance, and other sections are edited independently without replacing unrelated content (#4874).
Rewrote the animated glow in the Card CSS Theming guide example to animate opacity on an overlay layer instead of the message bubble's box-shadow. Animating box-shadow forced a full repaint every frame and could pin a weak GPU; the example now demonstrates the GPU-composited pattern instead (#4897).
Matched Feature agent pages to the shared Agent editor chrome and Chroma accent, and correctly recognized their owning package as installed. (#4892)
Kept the Your Guide Professor Mari action icon synchronized with its label color.
Kept Sidecar tracker counts and bulk local assignment current when capability Agent packages hydrate or refresh, and centralized Agent result-type admission behind one shared compatibility-tested vocabulary.
Kept explicitly selected persona-linked lorebooks usable outside their owner persona while preserving automatic owner matching and chat exclusions (#4887).
Extended the Roleplay New Start divider across the full message body for user messages as well as assistant messages (#4886).
Applied Game image prompt overrides and the selected Illustrator prompt template to Game prompt-director requests, including manual portraits that use the Illustrator agent's image connection and explicitly customized prompt settings when the general dynamic-prompt toggle is off (#4880).
Made Game combat use character-sheet levels, resource pools, and typed abilities instead of deriving levels from HP and treating every ability as an attack (#4881).
Recorded the fallback provider and model on messages it actually generated instead of displaying the failed primary model (#4879).
Remembered the Echo Chamber corner independently for each chat and flushes position changes immediately so they survive tab changes and app restarts (#4882).
Made chat generation-parameter send toggles authoritative, so disabling Verbosity omits it from provider requests even when a preset still has a selected verbosity value (#4883).
Fixed the Windows installer falsely rejecting the required pnpm version when its version command emits LF-only output, while continuing to reject failed commands and mismatched versions (#4875).
Kept Random Model chats on a stable configured embedding source during Memory Recall refreshes instead of falling back to the local embedder (#4864).
Restored installed feature-only packages such as Noodle to the Agents management pane while keeping them out of ordinary chat-agent pickers (#4865).
Preserved bold formatting nested inside italic user messages (#4866).
Made server-present UI settings authoritative when an older browser cache has no trustworthy timestamp, while retaining local values only for settings absent from the server (#4867).
Protected Marinara's stock Universal Preset from deletion and direct edits, restored it when missing or changed, and preserved requested edits in a separate editable copy (#4871).
Re-enabled Character card lorebook embedding immediately after an embedded lorebook is removed (#4872).
Preserved Game Send-on-Enter and tutorial-dismissal preferences across their sync and local-storage boundaries, and made Session Logs honor persisted token counts and current-session markers after message metadata hydration (#4869).
Kept Recent Chats previews separate on narrow desktop windows by limiting the constrained two-column layout to the four newest chats while preserving the existing mobile and wide-screen counts (#4858).
Scoped image-connection prompt instructions to image-producing agents, bounded stored instructions to 20,000 characters, and kept each retry agent on its own configured image connection.
Kept chat settings profiles within their saved chat mode, migrated legacy Visual Novel profiles to Roleplay on import, and prevented reusable profiles from overwriting branch identity (#4849).
Served the Home shell directly for unknown routes so Termux mobile launches cannot recurse through the server's 404 handler (#4850).
Corrected legacy Google Gemini /v1 connection URLs to /v1beta for connection checks, model discovery, chat generation, and embeddings (#4854).
Removed the decorative address bar from mobile Home and the obsolete close action from Professor Mari's Home tab while retaining the full desktop browser frame (#4855).
Restored intuitive mobile swipe navigation for Conversation transcripts while preserving Roleplay swipes and interactive controls (#4841).
Made Memory Recall re-vectorization exclusive with background chunking so embedding-model changes cannot leave mixed vector dimensions (#4843).
Saved pending Lorebook vector settings before vectorization and surfaced provider or eligibility failures instead of reporting a misleading zero-vector success (#4844).
Recovered expected sharded chat history from its preserved pre-shard backup when a restored profile contains no message shards (#4845).
Kept Professor Mari's Home navigator enabled by default and visible with reduced effects, reset her to the default position when re-enabled, centered her drag handle, contained compact Field Notes and Community actions, and presented iPad sidebars as full-width overlays (#4826, #4827, #4829, #4830).
Sharpened Professor Mari's read-only guardrail so a "how do I…" question is answered or offered rather than performed, even when it names the change as its goal (for example "how do I make X have Y") — while a plainly-worded request to make that change, including a polite question form like "can you set X to Y", is still carried out — so she keys on intent rather than grammar and no longer edits without a clear instruction (#4838).
Reduced background autonomous-message polling to a lightweight candidate-id lookup instead of re-fetching the full chat list every 30 seconds (#4715).
Stopped the server's autonomous-messaging scheduler from sweeping every 60 seconds when no chats had changed, cutting idle load on self-hosted installs (#4716).
Stopped chat message history from being fully re-downloaded on reconnects, background generations, and Professor Mari workspace changes (#4719).
Replaced the .env config watcher's 2-second stat-poll with an event-driven directory watch (falling back to a slower 30-second poll only if needed), cutting a constant idle-wakeup cost on phone and Termux installs (#4723).
Moved the Chat Settings secret-plot reader off the shared chat transcript query, so Load More no longer disappears while older messages still exist and the secret-plot panel stays current after new messages arrive (#4725).
Replaced fixed 25ms Personal Extension sandbox polling with an adaptive cadence that idles down once traffic quiets, cutting background wakeups and flash writes for every running server extension on phone and Termux installs (#4727).
Bounded Professor Mari's workspace data reads so oversized character/lorebook fields are elided with a readable note and recoverable via drill-down, instead of being silently truncated mid-JSON (#4776).
Stopped Professor Mari from making unrequested changes when only asked to inspect, explain, or get how-to guidance, answering informational requests without modifying anything (#4816).
Routed every Professor Mari workspace create through the same Keep/Restore review as edits and deletes, so an unwanted creation can be undone in chat, and blocked creates from overwriting existing rows with colliding IDs (#4825).
Corrected Home and Appearance presentation regressions affecting semantic chat-mode colors, the Square avatar preview, Noodle notification placement, compact mobile Achievements, Discovery Desk spacing, mobile Bookmarks motion, and consistent widget-manager labels (#4809, #4817, #4818, #4819, #4820, #4821, #4823).
Restored Game party portrait file selection for Character, Persona, and NPC sheets whether or not the sheet editor is active (#4793).
Kept legacy reaction payloads out of rendered message content, accepted /title <name> in Professor Mari chats, and let her navigator open an exact saved chat title (#4782, #4794, #4800).
Standardized gradient-stop editing on Marinara's owned color controls and tightened Home responsiveness, including Guide overflow, deleted Recent Chats, semantic mode colors, decorative browser refresh behavior, larger Mari drag targeting, and pausing Home animations while the app is unfocused (#4797, #4799, #4802, #4804, #4807, #4808, #4809).
Kept blank messages and hidden command anchors out of Home's Recent Chats previews, counted paginated chat histories without materializing every message ID, stabilized history cursors across deletions, and rejected malformed cursors.
Applied the selected tag-import mode when a downloaded character card is imported as a persona.
Matched the Character and Persona Download/Open Library launchers to the Chats segmented controls with an exact centered split, fixed icon sizing, and single-line labels.
Kept Professor Mari's held navigation animation mounted, frame-snapped, paint-batched, and slightly wider so repeated grabs no longer flash the wrong sprite frame, stutter under pointer movement, or squash her proportions.
Mirrored Professor Mari's navigation speech-bubble tail from one canonical shape when her dialog switches sides and seated it flush against either dialog edge.
Centered the desktop Ask Me Where Things Are tutorial card and spotlighted Professor Mari's actual navigation bubble instead of her full-screen drag surface.
Standardized chat-mode icons across Home, Chats, creation flows, shortcuts, and linked chats: Conversation keeps its message bubble, Roleplay uses theater masks, and Game uses a gamepad without colliding with Lorebooks.
Kept mobile Chats, Settings, and tracker panels above the Home browser chrome so their controls remain tappable.
Stopped hidden Home and Professor surfaces from retaining navigation timers, pending focus frames, message-history requests, Discovery rotation, and workspace-status polling after their lifecycle ended.
Kept Professor Mari's navigation dialog controls clickable when her draggable sprite overlaps the dialog.
Kept Professor Mari suggestion chips visible after entering her Home tab instead of briefly showing them before loaded chat history made the composer jump.
Rejected malformed Persona create/edit fields and normalized downloaded Persona copies with fallback names plus unsafe-paint cleanup without changing saved Personas (#4857).
Kept NovelAI V4.5 style-plate file selection inside the Connection editor across Chromium platforms (#4777).
Deactivated and revealed lorebooks that lose their final owner, replaced Lorebook Keeper entry bodies without stacking duplicates, and added sandbox-safe structured file copy, move, and removal tools for Android/Termux (#4775).
Let readers scroll through Professor Mari history while her response is still streaming (#4774).
Replaced the browser-native single-chat deletion prompt in Professor Mari with Marinara's confirmation dialog (#4773).
Preserved Tracker panel controls at their configured width, kept the panel aligned beside the right sidebar, and made closing its detached window close the panel (#4772).
Displayed the actual error output beneath failed Professor Mari workspace tool events (#4770).
Quarantined an automatic-generation connection for six hours after three consecutive provider failures while keeping foreground chats available (#4769).
Moved smart group responder selection to the default Agent connection with Agent fallback and the connection's saved generation parameters instead of a fixed 512-token chat-connection call (#4765).
Shared Professor Mari's robust JSON repair waterfall with structured agents so repairable malformed or truncated output no longer wastes retries (#4753).
Revalidated replaceable user backgrounds so re-importing a different image under a deleted filename no longer displays stale browser-cached pixels (#4757).
Kept NovelAI V4.5 style-plate uploads inside the Connection editor by decoding oversized images directly to a bounded preview and avoiding redundant copies of their encoded data (#4748).
Replaced the Windows Chromium native Connection and Preset popups in new Conversation and Roleplay setup with readable themed selectors (#4750).
Kept explicitly disabled Boolean preset choices empty instead of injecting their first option into prompts (#4740).
Resolved current DeepSeek, MiMo, GLM, and Kimi output limits through NanoGPT, OpenRouter, and custom OAI-compatible connections, and stopped Professor Mari before executing tool calls from truncated output (#4741).
Sent explicit OpenRouter prompt-caching markers for every compatible model, including Gemini, instead of limiting the option to Claude (#4742).
Gave native Connection and Preset dropdown options readable theme colors in the new Conversation and Roleplay setup wizards on Chromium (#4743).
Prepared and bounded NovelAI V4.5 style-plate uploads before previewing them so large images no longer destabilize the Connection editor (#4744).
Let capability packages use the remote embedding source configured on their agent connection, retaining local MiniLM only as a fallback (#4745).
Reused matching avatars from the full Character library in both Game portrait-generation paths instead of repeatedly generating replacements (#4746).
Isolated the tracker panel's newest-message lookup from transcript pagination settings (#4724).
Preserved speaker names and timestamps in individual Conversation group history so characters can distinguish prior speakers (#4726).
Raised the default Termux Node.js heap limit for large profiles while preserving user-supplied limits (#4730).
Reconciled failed-generation prompts by submission ID so API errors do not show duplicate user messages (#4731).
Saved Echo Chamber resizing when pointer capture ends outside the resize handle, preserving its per-chat size across navigation and restarts (#4733).
Kept the selected Narrative Director push-story instruction in every individual group responder prompt, independent of queue and Secret Plot settings (#4734).
Allowed server extensions to finish storage writes registered through onCleanup during a polite stop.
Made edits to Conversation replies that contained character commands reach the model instead of the original wording (#4728).
Kept the full desktop Roleplay edit checkmark above overlapping message layers so its entire visible hit target remains clickable (#4700).
Made the Windows launcher and installer detect Node.js by running it directly instead of requiring where.exe (#4701).
Added the standard expanded editor and macro guide controls to Roleplay summary editors (#4710).
Matched Chats hover loading text and visible error details, including “Internal Server Error,” to the active accent/chroma text color (#4712).
Isolated the memoized Chat area from unrelated topbar state and stabilized message mutation callbacks, preventing unchanged Roleplay transcripts from rerendering when tabs open (#4713).
Applied the committed-message Markdown, action, and dialogue-color formatter to Roleplay output as it streams, while keeping updates coalesced to animation frames (#4714).
Show the localized validation error before restoring an invalid all-zero NoodleR fan activity weight.
Asked for confirmation before permanently deleting a prompt block from preset chat settings (#4698).
Let Professor Mari read trusted Wikipedia links through her structured wiki tools on Android while keeping raw shell networking sandboxed (#4691).
Stopped Illustrator from appending its generic anti-text list to NovelAI negative prompts, preserving the prompt supplied by the agent (#4692).
Converted NovelAI style-reference fidelity to the provider's inverse secondary-strength scale before sending the request (#4693).
Bounded delayed Roleplay typewriter catch-up and disabled competing Echo Chamber smooth scrolling during the main stream, keeping parallel-agent output fluid (#4694).
Kept Character and Persona gallery controls inside mobile cards, added deletion to their expanded image views, and matched Chat gallery delete actions to the configured accent (#4695).
Removed app-wide color-transition storms and synchronous cursor recoloring from animated accent ticks, and deferred offscreen gallery rendering to keep long sessions responsive (#4696).
Scoped cross-chat awareness in individual group conversations to the character currently replying (#4688).
Reconciled just-sent Roleplay messages with their saved IDs before editing so the first edit after stopping generation persists (#4678).
Retried one transient Roleplay message-edit save failure so remote sessions persist the edit without requiring another manual Save action (#4678).
Kept Persona tags, stats, saved statuses, avatar crops, tracker colors, and conversation settings intact across loading, editing, duplicating, restoring, and switching Personas by returning one consistent Persona API shape (#4646).
Included the active Conversation cast's avatar references and appearance descriptions in guided group selfies (#4676).
Added native Arli.ai image generation with authenticated text-to-image and image-to-image requests (#4672).
Let Professor Mari read complete lorebook entry bodies through her structured app-data tools instead of receiving only truncated entry previews (#4673).
Removed conflicting color-environment and unused React Compiler/Babel warnings from test and lint runs, refreshed smoke coverage for the current release, schema, agent, and Noodle settings surfaces, and preserved a reader's exact mobile chat position when the keyboard opens (#4670).
Kept source Chat Summary entries as inactive history after combining summaries in Roleplay mode instead of deleting them, and revealed that history immediately after each combine.
Removed local Character and Persona avatar files when their cards are deleted individually or through Danger Zone, including avatars retained only by deleted card-version history (#4668).
Installed the pinned pnpm automatically when the Windows installer cannot use Corepack, an existing pnpm, or its temporary runner (#4662).
Exposed saved prompt choices for active Roleplay tracker agents in Chat Settings, matching the existing Game-mode selector (#4663).
Centered the avatar-upload camera and inset the AI-generation action so both controls remain fully visible over the mini preview in Character and Persona editors (#4665).
Simplified proactive Conversation intent hints so check-ins follow the selected moment without extra tone instructions.
Distributed mobile topbar icons evenly across the available screen width without changing the desktop layout (#4666).
Kept the caret at the chosen insertion point while typing in expanded Character and Preset editors instead of repeatedly focusing the field and jumping to the end (#4656).
Restored a full, unobstructed hit target for Roleplay message edit controls so Save no longer reacts only near one corner (#4658).
Coalesced Roleplay streaming presentation to animation frames to limit Firefox accessibility-tree churn (#4659).
Added an independent Noodle timeline image-size setting, defaulting to the GPT-Image-compatible 1024x1536 portrait canvas instead of reusing the Illustrator dimensions (#4660).
Made NoodleR stage-profile drafts tolerate single-item array responses, extra model fields, and decorated handles from local models while ignoring model-provided disclosure modes, discarding invalid values, and applying the requested valid mode (#4626).
Removed deprecated generation parameters from single and bulk prompt-preset exports and ignored them when importing older Marinara or SillyTavern preset files (#4650).
Kept the selected prompt checkmark above the Presets avatar frame instead of clipping it into the rounded image border (#4651).
Kept Roleplay message editors open until the save is confirmed, so a delayed mobile save after stopping generation cannot briefly restore stale text or discard the first edit (#4649).
Kept the reasoning action visible when a provider reports hidden reasoning-token usage but omits the displayable summary, and explained the missing summary in the Model Thoughts panel.
Coalesced Professor Mari's streaming transcript work to animation frames and released her input lock before best-effort refreshes, preventing long replies and stalled cleanup requests from leaving the assistant unresponsive (#4628, #4637).
Stopped personal extensions and their policy from polling on every screen and in background tabs; existing query invalidation now refreshes them after changes (#4629).
Added each active Game tracker agent's available prompt templates to Chat Settings for quick per-chat selection (#4640).
Rebuilt incomplete Android/Termux output when a dist directory exists without its required entry file, preventing the server from starting into a browser and app 404 loop (#4639).
Raised profile ZIP import capacity from the former 1 GiB ceiling to the ZIP32 format limit while retaining per-entry and expanded-size safety limits (#4641).
Moved full Docker images to Debian Trixie so ARM64 sidecars can load the required glibc and libstdc++ symbols (#4638).
Matched the Add character to active chat button to the neighboring Character row actions in folders and standalone rows on desktop and mobile.
Stopped the NoodleR reserve poll from scanning the prepared-post and Noodle post tables every minute when automatic posting is off and no reserve posts exist, and backed the automatic timeline-refresh poll off to 15 minutes while refreshes are disabled, cutting idle CPU wake-ups on phone and Termux installs (#4630).
Removed duplicate safe-area padding from Noodle's mobile setup footer so the shared shell remains the single owner of spacing above Android navigation controls (#4586).
Kept the first edit to a sent user message after canceling generation instead of discarding it when transient swipe cleanup changes the active index (#4608).
Matched the Add character to active chat action to neighboring controls in folders and standalone Character rows, including a visible consistently sized icon on desktop and mobile (#4611).
Made Clear Trackers use the configured accent hover treatment instead of a destructive red state (#4612).
Made agent calls inherit the selected connection's temperature and parameter-send policy, with 0.7 as the default when no temperature is configured, and kept incompatible per-agent settings isolated during batching (#4614).
Reduced Spotify Music DJ execution to one full-context planning request followed by deterministic playback, and clarified per-round versus cumulative timing in agent debug output (#4615).
Added preset picture upload and replacement controls to the Presets panel and preset Overview editor, matching the existing lorebook picture workflow while cleaning up replaced and deleted artwork (#4624).
Prevented completed NoodleR profile-draft generations from restoring an editor after the user canceled, changed sources, or moved to another draft flow.