Changelog
Every release across the Lite Suite ecosystem, newest first.
LiteSuiteLiteImageLiteSoundLiteModelerLiteBenchLiteDock
- One voice engine, three faces — conversation, navigation, and dictation share a single engine: one capture arbiter, one intent router, one speaking-state owner.
- All 50 voice-nav commands listed and bindable, with cross-package keybinding collision detection.
- Generative UI asks route to the asking seat's thread with modal/card surfaces and reload-safe replay.
- Frontier chat: kuro-native threads, streamed tool rows with real durations, provider routing.
- PTY replay preserves terminal modes — TUI reattach keeps cursor state; message streams dedupe-keyed.
- Bundles liteharness 0.4.0 (wheel + plugin 1.0.16); ends the 0.0.x era.
- Dead agents no longer pile up in the fleet roster, and a running agent's name can no longer be taken by the next one to start. An agent that registered through the LiteHarness command line rather than through a session hook never recorded which process owned it. Two separate mechanisms read that field, and both treat its absence as "already dead": the pass that reaps stale entries, and the guard that stops one agent claiming a name another is still using. So those agents were simultaneously unreapable and unprotected. Measured on a live roster: six entries for a single seat, four of them dead, and a running agent whose name was taken out from under it.
- The bundled LiteHarness was a version behind its own source. 0.0.56 shipped a 0.3.7 build made before that day's fixes landed, so the installer carried code the repository had already moved past — under an identical version number, which is exactly what made it invisible. 0.0.57 bundles 0.3.8, checked by reading the contents of the wheel inside the installer rather than by comparing version strings.
- The hourly cleanup now runs whether or not an agent has mail waiting. It sat behind an early return that only executed when there were unread messages, so on a quiet fleet it could go half a day without firing against a one-hour schedule.
- Listing online agents no longer counts one agent twice. Resuming a session registers once under a temporary identity and again under the real one, seconds apart from the same process. Both are genuinely alive, so no liveness check could separate them; they are now collapsed by the process that owns them.
- Running LiteHarness setup a second time no longer stacks duplicate hooks. The wizard tagged the entries it wrote so it could recognise them on the next run, but Claude Code strips unknown keys when it rewrites settings.json — and the wizard itself triggers that rewrite one step later, erasing its own marker every time. Two runs took a config from 32 hook groups to 45, with ten commands registered twice, including session registration firing twice per session. Idempotency now keys on the hook command itself, which nothing else can normalise away.
- The setup health check no longer reports a worse install as a better one. It counted hook groups, so a config with eleven duplicates read as "43 hooks registered" — a number that went up precisely because the install had gone wrong. It now counts distinct hooks per event and reports duplication as the fault it is.
- Setup no longer freezes for 30 to 60 seconds on the Claude and Codex panel steps. Installing those panels copied roughly 1.2 GB across 7,400 files without yielding, so the window stopped repainting and Windows marked it Not Responding partway through a successful install.
- The taskbar button now carries the LiteSuite icon rather than a generic one. The application identified itself to Windows with a different id than the one its own installer writes into the Start Menu shortcut, and that id is what Windows uses to tie a running window to its shortcut for grouping, pinning, jump lists and notifications.
- Setup says plainly when it edits the global instructions file. Adding its section to ~/.claude/CLAUDE.md affects every project opened on the machine, and it was reported as an unremarkable "wrote file" line among forty others.
- Terminal panes and tabs can be dragged out onto the canvas and back again. An individual tab can be pulled out into a pane of its own, or dropped onto another group to merge into it, and dragging a pane over another swaps their contents while the thin edges split and the tab strip merges. A HUD pill and a ghost of the resulting layout follow the cursor, replacing the region tinting that made it hard to tell what a drop would actually do.
- Dragging felt broken because the drop zones were painted underneath the terminal. They had been firing correctly for months; nothing that reported on them could see the problem, because hit-testing works on rectangles rather than on what is drawn on top. The reason it presented as "you can only drag to new tabs" is that the tab strip was the one target a person could actually see.
- Renaming a thread does something. The rename input was rendered, focusable and typable, and wired to nothing at all — the new name was discarded on submit, which is indistinguishable from a save that silently failed.
- One terminal is no longer shared between every project. Pane identifiers were minted from a per-context counter, so two different projects could independently produce the same id, and a merge keyed on that id handed both of them the same terminal session.
- Classic dictation survives a Voice Conversation. Leaving a conversation open left it holding the audio pipeline, so ordinary dictation went quiet afterwards with no error and no indication of what had taken it.
- The voice cancel hotkey no longer steals Escape from the whole machine. It was registered globally rather than scoped to a live voice session, so pressing Escape anywhere — in any other application — was being intercepted whenever LiteSuite was running.
- Wake-word toggles now write the setting the wake engine reads. The toggle wrote one key and the engine read another, so the control appeared to work, persisted its state correctly, and changed nothing.
- Compaction is backed up automatically. PreCompact and PostCompact are wired to the harness backup and log verbs, so the full conversation is written to disk before it is summarised rather than after it is gone.
- Agent inboxes refuse a stale identity. A resumed session could inherit an agent id belonging to a sibling that had already exited and quietly consume that agent's mail; the watcher now refuses an id it cannot verify, owns the heartbeat it used to merely carry, and records which identity it chose to disk so the decision can be audited afterwards.
- Personal architecture skills no longer ship. Three skills that index a codebase the end user does not have were being packaged twice over, in the plugin tree and inside the bundled harness wheel. The wheel is rebuilt carrying none of them, and the superseded copy is deleted rather than left beside the new one, because the loader takes the first filename that matches and would have gone on resolving the old one.
- The claim that they no longer ship was checked against the installer itself. Every previous privacy control read the source tree, which can only ever make claims about the source; the built artifact is a different object. It was scanned directly, 1,556 files, and the scanner was first made to fail on a planted token so that a clean result meant something.
- Agent bridges no longer hang on a child process that never closes its input. Six places spawned a helper and left its standard input open, so the child sat waiting for input that would never arrive. The first fix attempted here was a no-op, because the option it passed is silently ignored by the function it was passed to, so each site now records whether it was proven or merely defensive.
- The text-to-speech skill ships the scripts it cannot run without. Both were present on the machine that wrote them and tracked in no repository, so the skill worked there and failed everywhere else. They were the only two untracked scripts out of twenty-six in that tree.
- Spawned agents keep their own transcripts. A spawned agent inherited a flag that suppressed its session recording while its parent kept writing normally, so an agent could work for hours and leave nothing behind. The asymmetry is what hid it: the parent is always fine, so this is invisible to anything observing itself.
- Self-compaction leaves the agent pointing at its own state. The wake-up command is now queued into the terminal while it is still compacting, which means it arrives as the first thing the agent sees on the other side instead of being swallowed by a busy prompt.
- The auto-updater loads in every build that consumes it. The shared updater package is imported by three different builds and no single import form worked in all three; rather than picking one and breaking the other two, the mismatch is now resolved at each consumer.
- Generated CI stopped asserting things it had not measured. The "no unit test suite" step had been hand-typed into two repositories rather than derived, so it kept claiming there were no tests after tests existed. Python setup was inferred from two checks that were wrong in opposite directions, and an empty step left a stray blank line. The test step is now derived, with each repository still declaring its own dependencies.
- Agents work instead of narrating. Every tier now routes agent-to-agent traffic through the message inbox and reports through a single voice, rather than each agent repeating the same result outward at every hop.
- A message can no longer be sent wearing someone else’s identity. Omitting the sender did not mark it unknown — it guessed one from the most recent session transcript, so the message arrived attributed to a different live agent. The sender is now required, and the refusal explains what to pass.
- Self-compaction leaves an audit trail. The helper was silent unless asked to be verbose, which meant that afterwards nothing could distinguish it having run from it never having fired.
- The assistant terminal receives its command. Two independent defects, either one enough on its own: the terminal store is keyed by one id and was being looked up by another, and the watcher treated "no terminal yet" as "terminal closed" and shut itself down milliseconds in.
- `liteharness` 0.3.3 is bundled. 0.0.51 shipped a version with no prompts module at all, so the orchestrator creator died on its first call. All five call sites resolve.
- Tier preambles are delivered again. The shipped hooks had zero references to the code that sends them, so every agent ran without the constraints its tier defines.
- Your generated assistant survives upgrades. It used to be written into a version-pinned directory the installer replaces, so a completed setup was one update away from silently reverting to the generic default.
- **1.0.5 could not start. This fixes that.**
- Signed-in identity in the title bar. LiteBench was the only gated Lite app with no account avatar, so a licensed user saw no confirmation the license gate had let them through. Click it for your account and sign-out.
- About reports the real version. It was a hardcoded constant frozen at 1.0.0 and had read 1.0.0 since the 1.0.0 release; it now comes from the packaged app.
- Licensing corrected. LiteBench is proprietary and part of the Lite AI Suite. The MIT declaration was a leftover from the earliest builds, before it moved onto Lite Suite tooling. The MIT grant still covers the releases it shipped with.
- Sign-in completes on its own. The activation poll no longer overlaps itself, so its connection pool cannot collapse partway through. Measured on a clean Windows machine: 62 polls, 62 responses, 62 bodies read and zero exceptions, where the previous build managed 9 bodies out of 187 polls with 120 exceptions. The activation code now comes back in 3 milliseconds instead of 30 seconds.
- You finally get a status line. The code for one shipped months ago and was never once installed, because the setup wizard writes ~/.claude/settings.json by its own path and never called the function that adds it. It now installs on both setup routes and will never overwrite a status line you already have. It also shows which agent you are looking at by name and id — on a machine running several agents, four identical status lines are wallpaper.
- Setup no longer stalls for three minutes. The Codex Desktop step used to hold Node, Python, Git, both CLIs and the plugin behind a wait for an installer you had not finished clicking through. That wait now runs alongside the rest of setup instead of in front of it, and it says what it is waiting for rather than showing a motionless row.
- Honest labels on machines without the Microsoft Store. Setup used to finish by telling you to install Codex later from a Store that is not present — naming an action you cannot take. It now opens the download page in a browser found on disk, which also works on machines where Windows has no default browser registered.
- The updater is packaged correctly, and the build now fails rather than silently shipping without it. Editor extension downloads pick the build matching your platform instead of falling back to a mismatched one.
- Fixes the blank window in 0.0.36. A module-loading change made the production bundle fail to render at all while every test tier stayed green, because no gate ever loaded the packaged renderer. The release process now loads it and fails the build if the app mounts nothing.
- 0.0.36 is marked broken on GitHub so it cannot be installed by mistake.
- Withdrawn: this build shows a blank window on launch and should not be installed. Use 0.0.37 or later.
- MuseTalk is reachable. The talking-head engine was installed on disk with no route and no wiring connecting it to the app — present, and impossible to select. Engine, route and wiring all landed.
- An installed LTX no longer reads as offline. The readiness check itself crashed, so the endpoint meant to report LTX as available reported nothing at all, which looked identical to not having it.
- A retired engine is no longer the default. The talking-head path defaulted to an engine that had been deliberately shelved, and the interface still offered it.
- Three setup steps where a green result stood for something that could not actually run, and a weights test that was skipped but exited successfully — an absence is now declared rather than inferred from a zero.
- H3 Ref2VA: multiple reference images, video and audio. Wiring it up exposed and fixed a silent downgrade that had been quietly substituting a different video model.
- Engine-swap regression matrix, including a deliberately planted fault to prove the harness can report a failure, a partial run that can no longer overwrite a complete result, and a correction where the harness had reported a working model family as broken.
- Fixes a dangling import that broke packaging. The shared licensing package removed its protocol-handler API upstream, and LiteImage still imported it. The now-unreachable deep-link activation path has been removed along with it — activation by key and by browser sign-in are unaffected.
- Auto-updates. LiteSound now uses the shared auto-updater, so future versions install themselves instead of needing a manual download.
- LiteBench can be built again. The shared licensing package removed its protocol-handler API and LiteBench still imported it, so the bundler failed before producing an installer — 1.0.4 was the last version that could be packaged at all. The dangling imports and the unreachable deep-link activation path are gone; activation by key and by browser sign-in are unaffected.
- LiteModeler can be built again. Since the shared licensing package removed its protocol-handler API, the bundler failed before producing an installer — 1.0.4 was the last version that could be packaged at all. The dangling imports and the unreachable deep-link activation path are gone; activation by key and by browser sign-in are unaffected.
- Auto-updates. LiteModeler now uses the shared auto-updater.
- Speech generation, with three backends each in its own sidecar process: Chatterbox for zero-shot voice cloning from a 5-20 second reference, VibeVoice Realtime 0.5B for long-form expressive speech with eleven English presets, and Stable Audio Small — which had never actually run, having been loaded with the wrong pipeline.
- A narration-slate command turns a script file into named wavs plus a JSON manifest, for batch game-audio work.
- Stable Audio, Stable Audio Small and AudioLDM2 reported 'Not installed' after a successful download, so every click re-downloaded 4-15GB. The check assumed a root config.json that those repos do not publish; detection now uses the files each backend really downloads.
- Every download size was understated by 2.1x to 4.1x — Bark advertised 5GB and transfers 20.68GB. Sizes are now measured against the live repos, gated models say so before you click instead of failing with a raw 403, and Chatterbox installs the 5 files it needs (2.97GB) rather than the full 22GB repo.
- ACE-Step's default port moved off one that collides with LiteSuite's terminal service, and the orphan-port cleanup no longer terminates processes that are not LiteSound's.
- Audio generation runs from a repo checkout — the installer does not bundle the Python engine, and never has. pnpm python:setup-speech builds and verifies the three speech environments.
- Mesh cleanup is its own operation now. Floaters, remesh, simplify, decimate, hole fill and unwrap run as a selectable chain on any mesh you already have — previously they only happened as a side-effect of a texture job, which meant a prompt, model weights and GPU time just to decimate something.
- Edge-flow retopology: field-aligned quad retopology reaching 100% quads on real generated meshes in roughly 12-14 seconds.
- New Pipeline panel showing the same job graph as either a Kanban board or a node graph, and a Part Split panel with a backend dropdown.
- ARDY motion support for animation.
- Retopology used to reject meshes for 'not being manifold' when the real cause was mesh scale — identical geometry could fail at 1x and succeed at 100x. Meshes are now normalised instead. Separately, a sharp-edge preservation option that turned an 11-second solve into over eight minutes is no longer on by default.
- Not yet built, so you know what to expect: there is no before/after wireframe preview (the panel reports counts, quad percentage and timings as numbers), and per-region generation still goes through the Generate panel by hand.
- Your terminals now survive a restart. Closing LiteSuite closes windows, not shells — reopen the app and your terminals come back as the same running processes, with their scrollback replayed and still accepting input. Work left running in a terminal is no longer something a restart costs you.
- Terminal tabs can be dragged and docked where you want them, and the drag grip now sits in the tab bar's own slot instead of competing with the buttons beside it.
- A panel layout pass: panes arrange themselves into a grid automatically, the surrounding chrome is more compact so content gets the room back, terminals render full colour, and browser panes take keyboard focus properly.
- Long spoken replies are no longer cut off. Speech was ending at a fixed thirty seconds regardless of how much was left to say; there is now a Speech Timeout slider in Quick Voice, set to sixty seconds by default.
- The Codex panel finds Codex Desktop again on current builds, which now run under a different process name.
- The Codex panel works again. Codex Desktop updated itself to 0.147 and began asking LiteSuite for capabilities that did not exist yet, so every attempt to start a new chat failed — while the panel itself looked perfectly healthy, with nothing visibly erroring. LiteSuite now provides the six services and eleven message types the newer Codex expects.
- Starting a chat no longer drops the panel to a blank error screen. One piece of repository information LiteSuite reported back was missing two fields that Codex reads directly, which crashed the view outright instead of degrading. Failures of this kind now identify themselves rather than surfacing as an unnamed error.
- Otherwise identical to 0.0.33 — if your Codex panel is working, this update is optional.
- Terminals survive a restart. Closing or restarting LiteSuite now closes windows, not shells — your terminals keep running and reattach with their scrollback intact when the app comes back. Agents living in a terminal no longer die every time the app reloads. If the background service cannot start, terminals fall back to the old behaviour rather than failing to open.
- Paste an image straight into the orchestrator conversation with Ctrl+V, and the agent can read it.
- Images and video play in the canvas, with seeking. The video player's fullscreen button previously did nothing at all — fullscreen is a permission in Electron and it was being silently denied.
- Terminals open in the project you are actually working in, rather than defaulting to the workspace root.
- The orchestrator seat is now its own panel, separate from the agent list, and follows whichever agent currently holds the role — so handing the role over re-points the chat, voice and badge with no configuration.
- Nine entries in the Add Panel menu had never had an icon made for them and fell back to a generic glyph: Git, Orchestrator Chat, Agents View, Media, Design System, Screens, Claude Teams, Video Editor and LiteSound. They now have proper artwork. Plugin panels can show their own icon at all now — the menu had been substituting a plug for every one of them.
- Closing the app asks the window to shut down before forcing it. A forced kill mid-shutdown could truncate your settings and sign you out.
- LiteImage can now update itself. Previously it had no update mechanism at all, so every install stayed on the version it was first given until you noticed a new release and downloaded it by hand. It now checks on startup, fetches the update in the background, and offers to restart when it is ready — you can dismiss that and keep working.
- The 1.0.11 release was published without its installer. Only the zip was uploaded, so the download link on this site pointed at a file that was not there. The installer was built at the time; the upload simply never finished. 1.0.12 ships complete.
- Because 1.0.12 is the first version that can update itself, this is the last LiteImage update you have to install manually.
- The Runtime Engine installer works again. Two internal handlers had claimed the same name, and registering the duplicate threw — which silently abandoned every handler defined after it. The installer was the visible casualty, failing with an error that pointed nowhere near the actual cause.
- Repairing misplaced files reaches the right code again. The same name collision meant the repair button was calling into the wrong handler, which returned a result the screen could not read.
- A new Real-Repo Eval track measures whether a model can solve pinned software-engineering tasks inside isolated Docker containers, scored by each task's own test suite rather than by another model's opinion. Includes the 113-task public DeepSWE suite for comparable results, kept separate from your private regression tasks. The model never sees the verifier — it is injected only after the agent stops — and containers run with networking disabled and are destroyed after every run.
- The automated aesthetic judge is much harder to fool. It now runs three independent passes and takes the median, reports how much those passes agreed, and records whether a score came from consensus or a single pass. Previously an unreadable judge response silently became a score of 50, indistinguishable from a real verdict; it is now recorded as unavailable and dropped.
- The judge can score a rendered screenshot of the result, not just its source, falling back automatically for models without vision support.
- MiniMax-H3 video generation, with a real soundtrack. It is the only model here that produces a native audio track alongside the video — Wan, HunyuanVideo and LTX are all silent. Generated in one pass so a single soundtrack spans the whole clip, and a missing audio component now stops the run instead of quietly handing back a mute video.
- Face swap works again. It had been pointed at a folder that never held its models, so the engine could not load them and the settings panel offered a 1.1 GB re-download of files that were already on your disk.
- LTX 2.3 can be installed for the first time. Its text encoder pointed at a repository that does not exist, so every first-time install failed partway through with a download error and the feature could never finish setting up.
- Every feature now checks its own health when you open it, not just during first-run setup. It tells apart what previously all looked the same: genuinely missing, already on disk in the wrong place, waiting on a licence you need to accept, or removed after you installed it. Anything fixable without a download can be repaired in place — ControlNet models sitting unused on disk are wired up for you.
- Video models and text encoders are recognised by what they actually contain, so supporting files are no longer offered as image checkpoints and then fail to load.
- A failed generation clears when the next one starts, instead of leaving an error on the canvas that looks like every attempt is failing.
- War Room runs at a fraction of the GPU cost. A single background asset — a 2.7-million-triangle patrol drone, drawn four times — accounted for 91% of every triangle in the room. It has been replaced with lighter models picked at random per drone, and the scene now renders on demand at 30 Hz instead of a free-running 100 FPS loop. Triangles per frame dropped from 5,274,899 to 210,759, a 96% reduction, and the room's own GPU cost fell from +15.8 to +1.4 points.
- The browser is now a proper tabbed shell, with downloads and per-site permissions. Agent and user browser sessions no longer detach from one another mid-session.
- Downloads no longer silently overwrite an existing file. A page that chose a filename already on disk would clobber it with no warning and no way to tell it had happened.
- Patrol drones face the direction they are actually travelling, and patrol routes now genuinely patrol instead of drifting.
- The scene editor is reachable from a packaged build, not only from a development run.
- The status line no longer shows a frozen credit balance — the figure had been stuck since July 15.
- This is the first release where a fresh install can load the image engine. If you installed 1.0.8 on a clean machine, the setup wizard reported success while the engine could not start — please reinstall.
- The engine wizard now installs the complete runtime. It had been copying a hand-listed handful of files out of the engine archive and discarding the rest, including two libraries the engine links against directly. Fresh installs could not load it at all, and the readiness check still reported that everything was fine.
- CUDA acceleration is no longer skipped in silence. A missing backend library made the engine fall back to the CPU with no error and no log line, so generations simply took far longer than the hardware should allow.
- Engine runtime packs. Newer engine builds are opt-in downloads, selectable from Engine settings and fetched in-app with your signed-in identity — no browser round-trip. This is what unlocks MiniMax-H3 video with native audio.
- The virtual camera filter is back after going missing from the previous build.
- Feature Setup covers HunyuanVideo, downloads survive navigating away from the card that started them, and every feature that needs files now points at Feature Setup instead of failing quietly.
- Generation defaults are now per-checkpoint — CLIP skip, distilled CFG, and Flux defaulting to 1024 — so distilled and Lightning models stop coming out blown-out or over-sharpened.
- "Enable Video Mode" tells you why it is disabled instead of sitting greyed out with no explanation.
- Model weights and the CUDA backend are no longer bundled into the installer. Everything a feature needs installs on demand, the same way models do, so the download stays small and you only fetch what you actually use.
- Updates now have a real home. Command Center → Settings → Updates shows the version you are running, the current update status, when it last checked, and buttons to check, download, retry, or install and restart. All of that existed internally but had never been reachable from anywhere in the app.
- You can decide how updates behave. Two new settings: turn automatic checking on or off, and choose whether updates download by themselves or only after you say so — useful when a ~670 MB download would land on a slow or metered connection. Turning automatic updates off still leaves manual checking available.
- You can now see which version you are running. The status bar showed a hardcoded "v0.1" that had never been connected to anything; it now reports the real running build.
- Updates announce themselves. A badge appears next to your avatar when one is available, and a matching pill sits in the status bar — both say exactly what clicking will do (Update, Updating, or Restart to update) and both start the download or the restart directly. Previously the app could detect, download and install an update with no visible sign anywhere except a single button in the sidebar.
- The Codex panel works again. Projects and Recents populate, the "you don't have access to Codex yet" message is gone, and the composer is live. The cause was a field-name mismatch in the desktop bridge: the account was published as `planType` while the Codex client reads `plan` and `accountId`, so it saw no account at all and denied access — which, in Work mode, silently hid every local project and chat instead of showing an error.
- Projects are grouped by repository. The panel derives its workspace roots from your whole thread history instead of only the currently focused pane's directory, so Projects lists each repo with its own chats — and the workspace switcher finally has somewhere to switch to.
- Recents shows your local chats again. An earlier change had removed the local history feed and pointed it at a cloud endpoint that returns zero items for local-first users, so the list was permanently empty.
- The sidebar is seeded correctly at startup. LiteSuite now implements the bootstrap contract the Codex client expects (thread catalog, workspace roots and 16 global-state keys) rather than handing it an empty payload and relying on later feeds to repair a component that had already given up.
- Credit-balance scraping is switched off. claude.ai added bot protection, so the poll could never succeed and was retrying on every status-line tick.
- Automatic updates now work. They had never functioned in any previous release: there was no update feed to read, the Check for Updates menu item reached an orphaned code path that returned immediately without checking anything, and opening a window wiped any pending update notice. All three are fixed.
- The update feed is served from litesuite.dev rather than directly from the private source repository, so no credential is embedded in the installer.
- Codex config repair now covers every unsupported service_tier value. The Codex CLI accepts only two, and anything else made it refuse to load its config entirely while still reporting success — so every subsequent call failed silently.
- Codex panel: the panel is served from a secure origin, so its internal requests resolve instead of failing silently against a local file path. Correction: this entry originally said Projects and Recents populated again — they did not. Both stayed empty until 0.0.29, which fixed the actual cause.
- Codex panel: realtime voice and other staged-rollout features are enabled again. A request interceptor was answering the feature-gate endpoint with an empty success response, which read as "gates resolved, none found" and quietly switched every gated feature off.
- Account, onboarding, and feature-gate requests now forward to the backend with full authorization and integrity headers attached.
- Codex service tier handling was reworked during this release; see 0.0.28 for the corrected behaviour.
- New bridge instrumentation makes a sidebar fetch failure visible in the logs instead of silent.
- Claude Opus 5 is now the default model, with Fast Mode, Max Effort, Adaptive Reasoning and the Ultrathink keyword all resolving correctly for it.
- 1M-context sessions get their capabilities back. The context-window suffix in the model name was defeating every capability lookup, so a 1M session silently lost Fast Mode and the effort controls it should have had.
- Sonnet 5 is recognised by the model registry instead of falling through to defaults meant for an older model.
- Sprite sheets — pack a whole generated run into a single sheet with an atlas sidecar, instead of a folder of loose frames to assemble by hand.
- Frames that actually match. A locked seed, a shared LoRA stack and img2img chaining keep frame two looking like frame one.
- A Sprite Frames panel in the Generate sidebar, so a consistent run is something you drive rather than assemble.
- Pose authoring for ControlNet — draw an OpenPose skeleton in the app and hand it straight to generation, mirroring included.
- The TalkingHead and PersonaLive installers work again. They crashed on launch in every packaged build until now.
- A complete 3D-printing pipeline in a new Print panel — seven tools that carry a generated model all the way to a slicer-ready 3MF, entirely on your own machine.
- Orient, hollow and split — orientations scored on support volume and stability, distance-field hollowing (~40% less filament at a 2 mm wall), and bed-splitting with keyed alignment pins.
- Colour (AMS) — reduces a textured model to your filament slots per triangle, so it stays one closed solid instead of a pile of colour-split shells.
- De-light — strips baked shadows and highlights out of generated textures. Measured 68% tighter colour clustering, so AMS spends slots on real colours rather than lit-versus-shadowed variants of the same one.
- Relief & lithophane — turn any image into a printable carving or a backlit panel where dark areas print thicker.
- Models now arrive standing up: glTF is Y-up, slicer beds are Z-up, and that rotation is applied for you.
- Mesh quality control — a `qc` verb with subject-class profiles, `preflight` image validation before the GPU runs, and `clean_floaters` debris removal.
- Fixed panels bleeding through the War Room — native panel views ignore CSS entirely, so hiding a pane behind the 3D scene never worked; they are now hidden at the bounds controller while the scene is mounted.
- Sentinel chat — the pipeline stage display is a responsive 2×2 grid instead of overflowing on narrow panes.
- LiteDesign — token inputs grow to fit their contents, so long values no longer shear off the edge of the field.
- Harness — cross-vendor reviewer and OAuth CLI resolution tidied; bundled agent skill definitions synced to their durable sources.
- Camera-motion presets — an 18-preset catalogue of pushes, pulls, orbits, cranes and handheld moves, with a visual picker and up to three stacked at once.
- Hybrid local + cloud video — generate on your own GPU, or hand the shot to a frontier model through OpenRouter for trajectories local inference can't render. Async submit, poll, download and cancel.
- Cloud controls — seed, negative prompt, model discovery, and camera presets folded into the cloud prompt so the same preset works on either path.
- Fixed: the video path never passed `--lora-model-dir`, so motion LoRAs silently did nothing on video generation.
- Fixed: loading a profile wiped camera settings — `setVideoSettings` did a full replace instead of a merge.
- Glass Box token inspector — per-test inspection plus a standalone panel exposing local-model logprobs and entropy.
- Per-token logprob capture (additive, retry-safe) feeding the inspector.
- Restored the multimodal + OCR test suite and a multimodal-capable engine.
- Per-model harness profiles so each model is benchmarked fairly; benchmark artifacts hardened as hostile input.
- Cross-vendor code review — an OAuth-authenticated Codex CLI reviews Claude's diffs as a quality gate (no API keys).
- Model Cookbook — a ModelHub tab that scans your GPU and fit-scores local GGUF models against your VRAM, with one-click download and serve.
- Typed memory consolidation — deterministic, lock-guarded memory merges that never silently destroy a memory.
- Memory nudges and agent skill self-authoring, gated by a fail-closed security scan.
- Completed the Content-Security-Policy lockdown — locked style-src and isolated community widgets behind their own policy.
- Consolidated the built-in MCP tool surface and added native model and sound generation tools.
- Renderer performance and application-lifecycle hardening.
- Fixed the standalone MCP server (it was crashing on every tool call) and rebuilt on the current engine.
- Security hardening from an adversarial audit: HTTP auth parity with the WebSocket path, stricter command validation, a wider filesystem and credential guard, and tamper-proof license grace windows.
- Settings clarity redesign — every page leads with a plain-language intro, with power-user knobs behind an Advanced disclosure.
- The three voice subsystems (Dictation, Voice Navigation, Sentinel Voice) are now clearly named and separated.
- Claude pane parity (Opus 4.8, Fable 5, Sonnet 5) and a full slash-command palette; CI fully green; canvas, zoom, and quit stability fixes.
- Hardened, machine-bound license verification — signed ES256 entitlements re-checked every launch.
- Clearer trial-expiry and upgrade experience; stability and reliability improvements.
- Command Center simplification (18 views down to 14).
- Plugin runtime activation — all 20 panels are first-party plugins; community panels run as sandboxed WebViews.
- Glass Box token inspector (local-model logprob/entropy), footer role-presence, and git-as-memory dual-source.
- GGUF model-format support, a standalone Stable Diffusion server manager, a model-picker UI, and scanner/settings improvements.
- CLI stop verb with a clean two-process GPU clear, service-down exit hardening, and CLI contract probes.
- TripoSplat generation-runner rework and runtime wiring.
- ES256 asymmetric license verification, replacing the symmetric shared-secret baked into prior builds.
- Unified license-state UI across all Lite apps — shared login wall, trial status, and expiry.
- License-server security hardening — machine-bound activation and build-time secret injection.
- ES256 asymmetric license verification and a slimmed installer — runtime-provisioned CUDA/sd.cpp/ffmpeg (1.5 GB down to 256 MB).
- ES256 asymmetric license verification.
- Unified license-state UI and license-server hardening; ACE-Step GGUF backend and Radio Station Mode.
- ES256 asymmetric license verification.
- License storage fallthrough — no more re-signin on every launch.
- Live LiteHarness presence in the AGENTS panel; JWT secret inlined at build; repeatable release wrapper.
- Progress Theme system — 10 generation-time overlays (Bauhaus, Chronograph, Neural, DNA, HUD, and more).
- Avatar-cam refinements; Talking Head is now an opt-in post-install download; installer slimmed to 978 MB.
- First public release — local AI music and audio generation with prompt-to-WAV, continuation seeding, and granular quality controls.
- Windows NSIS installer with refreshed evolution-bridge token handling.
- First Windows NSIS installer — bootstrap prerequisite checks and the in-app setup wizards.
- Maintenance and stability release.
- First public release — local prompt-to-3D and image-to-3D generation with GLB export.
- First public release — a free, native Rust Stage Manager for Windows 11 with hotkey-driven window grouping.
- First public release — local GPU image generation with Flux, SDXL, and Stable Diffusion, plus face swap and avatar tools.
- Initial release — an LLM benchmark studio with live streaming results.