Changelog

• Editor plugins 0.2.15: when neither Claude Code nor Codex is installed, the Forge panel in Unreal now shows how to get one (Install Claude Code / Install Codex) and a Check again button, instead of an empty picker.
• Forge for Unreal is coming to Fab; the website now links there.
• Update the plugin in each Unreal project from Forge (Update Forge).

• Editor plugins 0.2.14: Unreal Engine 5.7 support. Forge now installs into UE 5.6, 5.7 and 5.8 projects. Everything works on 5.7 except the Control Rig commands and merge_proxy, which answer clearly instead (5.7 changed the Control Rig Python API, and the engine's proxy builder crashes there); they still run on 5.6 and 5.8.
• Sturdier on every engine: reading a dialog's text can no longer crash the editor, and placing a Paper2D tile map no longer crashes UE 5.7.
• Update the plugin in each Unreal project from Forge (Update Forge).

• Editor plugins 0.2.13: about 112 new Unreal commands across five areas (UE 5.8 and 5.6).
• Replication: see what replicates and why, configure actors, wire authority logic, run listen/dedicated-server Play with lag and packet loss, prove RPCs and sync, find relevancy and ownership bugs, Iris support, and C++ scaffolding that only writes code when you confirm (with full rollback).
• Physics: physics assets, bodies, constraints and ragdolls, collision channels/profiles (project settings only when you ask), traces and collision tests, simulate in Play and prove the result, Chaos fields and destruction.
• Rendering: Lumen, Nanite, shadows and other renderer features (with every side effect listed first), post-process, lights, sky and time of day, look-dev presets, shot comparison and finding why a scene looks wrong, GPU profiling.
• AI: StateTree depth, Smart Objects, Mass crowds, ZoneGraph, one-call AI agent setup, perception and AI debugging in Play.
• Niagara: measure an effect while it runs (are particles really spawning?), find invisible or expensive effects and fix them, events, parameter collections, scalability, simulation stages, data channels and flipbook baking.
• Update the plugin in each Unreal project from Forge (Update Forge).

• Automatic checks can no longer change files your task didn't touch. Anything a check changes outside the task is put back, and that check never counts as passed or gets sent to your team. Files the open editor holds can be put back with one click once it's closed.
• Without restore points for a project, Forge doesn't run automatic checks there and tells you why, instead of sending unchecked work.
• Fewer false alarms: normal editor closes and restarts no longer show as crashes, busy editors show "loading" instead of "not answering", and Stop shows grey instead of red "Failed".
• Checks and Schedules say what really happened ("couldn't run", "you stopped it", "finished, still for you to do: …") instead of "did not pass" or "blocked", and Push anyway is offered whenever it's safe.
• Calmer chats: no repeated warnings for your own MCP servers, no raw connection errors, readable crash reasons, and playtests only count new errors as bugs.
• Desktop-only update; editor plugins remain 0.2.12.

• Editor plugins 0.2.12: Control Rig and IK in depth (Unreal, UE 5.8 and 5.6). 50 new commands and deeper versions of the existing ones:
• Control Rig: build and edit rig hierarchies, controls and graphs, FK and IK setups, constraints, physics, auto-rig and modular rigs, compile and evaluate poses, add a rig to an Anim Blueprint, key and bake Control Rig tracks in Sequencer, and check it in Play.
• IK: IK Rig solvers, goals, chains and bone settings in one go, retargeters with chain maps, ops and poses, IK nodes in Anim Blueprints, skeleton analysis, and checks that the pose is actually right (bones on the ground, distances) in the editor and in Play.
• Crash guards for bad indexes and abstract solvers/ops; a failed step inside a batch now cleans up after itself.
• Update the plugin in each Unreal project from Forge (Install / Update Forge).

• Automatic checks never start new work. Schedule checks and Send to team only check the work the chat itself did: the agent may fix only what its own changes broke, never touches anything else, and never changes project settings or config to make errors go away. Older or unrelated errors are listed and left alone.
• Tasks that only read or answer questions (no files changed) are no longer checked or sent. Tasks that only changed notes or docs are sent without a compile check.
• Fix Schedule steps being skipped as "blocked" when they needed an Unreal Project Launcher profile: agents now know they may read the engine install, including the saved launcher profiles.
• Desktop-only update; editor plugins remain 0.2.11.

• Fix the compile check failing on errors your project already had: Schedule checks and Send to team now fail only on errors the agent's own work caused (unrelated, older console errors are listed, not counted), so finished work is sent again.
• Changing the model, effort or Fast mid-chat asks first when it would make the agent re-read the whole conversation (more plan usage), with "Don't warn me again".
• Token saver: agents keep small tasks in the chat instead of fanning out to subagents (same results, several times fewer tokens).
• Desktop-only update; editor plugins remain 0.2.11.

• Fast / Ultrafast button in the chat header: on Codex models where your ChatGPT plan offers Ultrafast, the Fast button cycles Off → Fast → Ultrafast (purple, with sparks) per chat. Without Ultrafast it stays Off / Fast. Applies from the next message; Settings > Agents > Codex Fast speed is the default for new chats and Schedule steps.
• Desktop-only update; editor plugins remain 0.2.11.

• Codex Ultrafast: Settings > Agents > Codex Fast speed (Fast / Ultrafast). With Ultrafast picked, Fast on a Codex model that offers Ultrafast on your ChatGPT plan (e.g. Pro) uses it, with a glowing purple Ultrafast button; models without it use normal Fast. Changing it applies from the next message.
• Dev and test runs of Forge no longer take over the installed Forge's taskbar icon.
• Desktop-only update; editor plugins remain 0.2.11.

• Version Control replaces Agents in the sidebar (the agents overview is now on the Dashboard). It works with Forge Version Control, an optional companion app for teams on Azure DevOps, made for game projects (Git LFS with Unity and Unreal rules). Not installed: Get Forge Version Control downloads, checks and installs it for you. Installed: its status, updates (Forge asks it to close first and never ends it in the middle of a send; version 1.1.0 updates once you close it) and removing the old Harbor app once everything moved over.
• Push to team: a button at the bottom of every chat in a project Forge Version Control manages. The agent sums up what the chat changed and checks that the project still compiles; then Forge Version Control commits it with the agent's message and pushes it to your team. The button shows Summing up… and Pushing…, and Stop cancels it. When it does not compile, Push anyway sends it with that message (after asking), until the chat has a newer turn.
• Auto-send per project (Off, Schedules only, Every task): finished work is sent to your team once it still compiles, with the agent's own commit message. A project's next Schedule step waits for its send. Also on the project's Settings tab.
• Every Forge project is listed on the Version Control page; the ones not in Forge Version Control yet have Set up, which opens it on the project's folder; the project moves up by itself.
• Conflicts and sign-ins stay in Forge Version Control: Forge never resolves a conflict or force-pushes.
• Forge's own check and push messages in a chat are labelled Forge, not You. While you were away counts what was sent to your team.
• Fix Work until done stopping overnight on C++ work ("blocked: three-build limit"): Forge allowed only 3 rebuilds per message. Now what ends it is 3 failed builds in a row (a build that succeeds starts the count over), and with Work until done 6 in a row, up to 30 builds per task; the same errors coming back no longer stop a Work until done run. Rebuilds no longer use up the automatic-continue budget.
• Desktop-only update; editor plugins remain 0.2.11.

• Usage dashboard (Settings > Usage): what your Claude Code and Codex sessions would cost on the API, with a daily cost chart, Cost / Tokens / Limits tabs, 24 h to 90 days, totals (cached, uncached, output, cache savings) and a breakdown by model, day or project. Switch between all agents on this PC and Forge chats only. Read from the logs on your PC; nothing leaves it.
• Desktop-only update; editor plugins remain 0.2.11.

• Schedule steps can carry images: an Attach image button under each step, plus paste and drag-and-drop, like the chat box. The images go to the agent with that step's message.
• Desktop-only update; editor plugins remain 0.2.11.

• Scheduled sequences no longer stop overnight on one step. A step whose agent is blocked on something only you can give (an account, credentials) gets one more try and is then skipped; any step whose retries are used up is skipped too. You get a notice, and the next steps run.
• Desktop-only update; editor plugins remain 0.2.11.

• Fix Forge getting stuck after a failed C++ rebuild (a second rebuild waited on the first forever, holding the agent slot so scheduled steps waited for one).
• C++ rebuilds that only need out-of-date engine plugins rebuilt (e.g. MetaHuman after an engine update) are built again the way your IDE does. When the engine itself needs repair, the agent is told exactly that (Epic Launcher > Verify).
• Stability: 20+ fixes, including schedule steps that could be skipped wrongly, one bad chat row blanking the window, Fast lost on some starts, Restart to update closing Forge, safer .uproject writes on a full disk, and fewer startup-setting rewrites.
• Editor plugins 0.2.11: fixes a bp_add_node crash on Blueprints edited or compiled during a task; script safety check for agent-written Python/C#; Unity bridge comes back after script reloads in a background editor; Unity multiplayer testing; fewer failed editor calls. Use Update Forge on your projects.

• Token saver (Settings > Agents). Safe (default) keeps results the same and uses less plan usage: shell output is condensed by rtk (installed once by Forge, telemetry off, your own Claude/Codex configs untouched) and agents keep their reads small. Max also compacts long chats earlier and caps big tool outputs, for the most savings. Off works exactly like the CLIs on their own.
• Forge no longer gets stuck on Unreal's "modules are missing or built with a different engine version, rebuild them now?" prompt: editors Forge starts skip it, Forge rebuilds the C++ modules itself (the agent sees any compile errors) and opens the editor again.
• Editor plugins 0.2.10: Unreal screenshots default to 1280 px and editor_status is leaner, both to save tokens.

• Continue in a new chat: a button in the chat's bottom bar (highlighted once the chat passes 60% context) and in the chat menu opens a fresh chat with the same agent, model, effort, Fast, bridge and editor picks. Its message box is pre-filled with a short handoff of the old chat (requests, plan, last messages, files changed) to edit and send.
• Editor plugins 0.2.9: Unity gets the Forge chat panel in Window > Forge, like Unreal: sessions, Claude Code or Codex, model, effort (Ultracode / Ultra), Fast, Stop and example prompts. A running chat survives script reloads. Use Update Forge on your Unity projects.

• Fast mode for scheduled work: new-chat steps in Schedule and the project page's start box get the Fast button next to Effort, and a continue step can turn Fast on from that step.
• Editor plugins 0.2.8 (Unreal): pie_start can run every other player as its own game process (separate_processes), which online subsystems like EOS and Steam need for multiplayer testing. Agents now ask once for missing online credentials and keep working on the rest. Use Update Forge on your Unreal projects.

• Agent-requested editor restarts no longer give up on a busy Unity: when it is compiling or reloading scripts, Forge asks it again to save and quit for up to 90 seconds.
• An editor that saved everything and accepted the quit but is slow to close is now ended and reopened for the restart, instead of leaving the agent stuck with "Forge couldn't save and close the editor". Nothing unsaved is ever ended.
• Codex now waits up to 120 seconds for Forge's editor connection when a chat starts (it gave up after 30 seconds, often during a Unity script reload).
• Desktop-only update; editor plugins remain 0.2.7.

• New top effort levels in the chat header: Ultra for Codex models that offer it (Max is now shown for them too), and Ultracode for Claude models with Extra high (Extra high plus multi-agent workflows). Both glow in a flowing rainbow when switched on.
• Fix the whole window jumping up and leaving empty space at the bottom when ticking a step option in Schedule (Commit when done and the others). The same fix covers the restore dialog and project lists.
• When the editor is closed, Codex no longer shows a raw "MCP server forge-unreal failed to start" error: slash commands show nothing, other messages get one plain line.
• Desktop-only update; editor plugins remain 0.2.7.

• Add a Fast button to the chat header, next to Effort. It turns fast mode on or off per chat from the next message: Claude Opus uses its fast output mode, and Codex uses the model's fast tier when it has one. Fast mode answers quicker and uses more plan usage. The button only appears for models that support it.
• Editor plugins 0.2.7: actor_spawn accepts class: "static_mesh" together with a mesh asset instead of refusing it, so agents no longer need a retry. Use Update Forge on your projects.

• Scheduled messages and sequence tasks enable Work until done before their first turn. Sequence checks and optional follow-ups remain separate; a task that stops blocked is not counted as complete.
• Sequence steps with no progress for ten minutes stop and resume their agent once, using the guarded editor launch path when the editor is closed or crashed. A second stall skips the step and records why, allowing later prompts to run.
• Provider activity and build progress reset the watchdog. Paused runs, user-stopped chats, queued user messages and known usage-limit waits are protected. Pending input can be skipped after ten minutes, without granting approvals or dismissing recovery dialogs.
• Recovery attempts persist across restarts. Skipped steps remain visible in the run summary. Normal failed-step Retry/Pause/Skip preferences still apply.
• Desktop-only update; editor plugins remain 0.2.6.

• New Unity and Unreal chats default to opening a regular editor window. Missing or invalid launch preferences also fall back to Window.
• Hidden remains available as an explicit selection. Saved app/project preferences and existing chats keep their chosen mode.
• Desktop-only update; editor plugins remain 0.2.6.

• Attach MP4, WebM and MOV videos to Desktop chats using the video button, drag and drop, or paste. Clips can be up to 60 MB and five minutes long; codec support depends on the built-in browser.
• Play and seek original clips inside chat, including saved chats and restored drafts. Videos copied to multiple chats keep independent attachments.
• Claude and Codex receive six timestamped frames from each clip. Forge clearly labels this sampling: audio and motion between frames are not analyzed.
• This is a Desktop-only update. Editor plugins remain 0.2.6 and do not need reinstalling.

• Agent-requested editor restarts and C++ rebuilds work regardless of who launched the editor. Work until done is not required, and explicit restart requests are no longer capped at two. Saves, duplicate-editor checks and intentional user-close protection remain in place.
• Fix a stale running state when restarting an external editor before its first log scan completes.
• Tool groups say failed attempts, and Forge command rows show the actual dispatched command, including in older chats.
• Strengthen UI guidance for aligned margins, consistent gaps, localization, DPI, clipping and runtime screenshot verification.
• Editor packages 0.2.6 fix Blueprint insertion pin lifetime during macro reconstruction, reject incompatible insertions before disconnecting existing links, and accept additional unambiguous property/widget argument aliases.
• Download complete Desktop updates directly, avoiding the hosting service’s repeated differential-download failures.

• Add Update all Forge plugins beside the dashboard's Projects heading when updates are available. It checks all known projects, including hidden ones, and updates installed plugins one project at a time.
• Bulk updates skip open editors and leave missing plugins alone. Progress survives navigation, project cards share installation state, and a final summary identifies skipped or failed projects. Versions are checked again before updating.
• This is a Desktop-only update. Editor packages remain 0.2.5; current plugins do not need reinstalling.

• Failed messages and scheduled sends retain their text and images when you switch chats, including anything typed while the request was pending. Duplicate sends are blocked across navigation.
• Failed chat-history loads show an inline Retry action. Streaming revisions are coalesced while history loads, and retries cannot replace complete replies with stale partial text.
• Resumed chats reconnect to the currently installed editor bridge. When provider history is missing, Forge supplies a compact, redacted handoff from the saved chat instead of losing the task context.
• Editor packages 0.2.5 preserve long Unreal panel transcripts, load earlier messages on demand and handle chat-switch/reconnect races. Unity logging settings remain editable when C# execution is disabled. Both engines include guidance for resuming work safely.

• Remove the retired Codex Windows sandbox setting that caused an unrecognized-configuration warning in chats. Forge's hidden desktop runner and genuine configuration warnings remain available.
• Install Forge from project cards, Overview or New chat. Current editor packages are 0.2.4 for Unreal 5.6/5.8 and Unity.

• Install Forge appears only when missing; Update Forge appears only when an update or legacy migration is available. A successful install becomes a quiet “Forge up to date” status.
• Plugin status and installation progress persist across navigation. Project cards keep stable spacing while checks and installs finish.
• Installation failures remain visible, stale checks cannot overwrite newer results, and editor-close safeguards still apply.
• The website showcases Forge inside Unreal and includes setup documentation. Plugin installation remains in the desktop app.

• Install Forge and Update Forge stay visible on project cards, the project overview and New chat, including when the installed plugin is current. Editor-close and package verification safeguards still apply.
• The test launcher refuses to silently open a different running Forge build. Close that build, including its tray icon, then launch test3.

• Forge is the default editor bridge, with separate verified downloads for Unreal 5.6/5.8 and Unity. Install or update it from the project card, overview or New chat. Both engines must close before installation; Forge can save, close and reopen only an editor it controls.
• Choose another installed bridge, including NeoStack, in the new chat's Editor bridge options. Disabled plugins show instructions to enable them.
• Forge bridge crash recovery, graceful restart and agent reconnect work with both Claude Code and Codex. Ultra terse replies are the default; Normal remains available in Settings.
• The Unreal Forge tab uses the desktop fonts and colors, consistent picker spacing, example prompts and a compact Desktop benefit card. Unity's branded Forge window is under Window > Forge or Tools > Forge.

• Live build line: when an agent packages, cooks or builds your game (Unreal BuildCookRun or cook, Unity player build), the chat shows one clean line above the message box instead of a wall of log output: the stages (Compile, Cook, Stage, Package, Archive), cook progress (2,641 of 2,820 packages), overall %, time, a Low memory warning when the cook runs short, and how it ended, with buttons to open the build folder or the log. You also get a notification when it finishes or fails.

• Stop means stopped: Claude agents can no longer set their own timers or watchers (the reason a stopped chat kept posting "Still waiting on you"). If an agent still works on its own after you pressed Stop, Forge ends it; your next message picks the chat up again.

• Progress panel works with Claude too: Claude keeps its plan in its task list (TaskCreate / TaskUpdate), which Forge now shows as the plan with steps done, % and time left, like Codex's. Those calls no longer show as tool rows in the chat.
• Work until done asks the agent to track its steps in that list so the panel fills in.

• Crashed editor comes back while an agent works in it: if the editor crashes while an agent is using it (or a chat in that project has Work until done on), Forge restarts it and the agent continues, even when you opened the editor yourself. Only a real crash counts: an editor you close yourself is never reopened.

• Stuck on an editor dialog: when an agent's editor call hangs, Forge checks whether the editor is waiting on a dialog (like Unreal's "Change RepNotify Variable Name"). It tells you in the chat and on your phone. If nobody answers within your "If you don't answer" time (Settings › Agents, with Keep going on), Forge closes the dialog like No / Cancel. The agent then hears that the action didn't happen and does it another way.
• A call stuck on a dialog no longer makes Forge restart the agent as "not responding".
• Unreal agents are told never to trigger editor dialogs (RepNotify renames and the like).

• Fresh editor on request: an agent that needs a new editor session (a project rule like "no edits after PIE", a stuck editor) ends its reply with FORGE: RESTART EDITOR. Forge saves everything, closes the editor, opens it again and continues the agent. It never force-closes: an editor that can't save everything keeps running.
• Work until done: when the agent says it's blocked until someone reopens the editor, Forge now restarts the editor and keeps going instead of stopping. An editor you opened yourself is only restarted in a chat with Work until done on.

• Progress panel: shows only the plan made for your latest message. An old, finished plan no longer reads 100% while the agent works on something new.
• Codex: "Reconnecting… Reconnecting… 4/5" now reads "Reconnecting… 4/5".

• Work until done (for overnight runs): turn it on in a chat's Progress panel and Forge keeps the agent going on its own until it says the whole task is finished. A reply that stops halfway ("remaining work…", "should I continue?") gets an automatic "keep going"; a reported blocker gets one push to re-check it and work around it or do the rest first. It stops when the agent says it's done, when it is truly blocked twice, after 40 continues, or when turns stop doing anything.
• Progress panel beside each chat: the agent's plan with steps done, a percentage, how long it has been working, a rough time left and how many turns it took. Opens on wide windows; the panel button in the chat header shows or hides it. Claude's todo list now shows as a live plan too, like Codex's.
• Unreal: a save that fails because a file is locked is no longer treated as "the project is open twice" right away: the agent ends PIE, waits and retries first.

• Chats name themselves: after a chat's third message, Forge gives it a short title that says what it is about ("Hand Blur and FPS Pass"), using one tiny call to your provider's cheapest model (Claude Haiku or GPT-6 Luna). A chat you renamed yourself keeps your name.
• A chat that starts with a slash command (like /caveman ultra) no longer gets that as its name: the first real message is used.
• Older chats get their summary title the next time they finish a turn.

• Settings › Usage: turns counted before 1.2.2 now get an API price estimate too (it showed "—").

• ≈ API cost: each chat shows what it would have cost on the pay-per-token API ("≈ $4.12 API" next to Context), and Settings › Usage shows the total per period ("≈ $183 of API usage in the last 7 days, covered by your plan") and per project. Official list prices for GPT-6 Astra / Sol / Luna, GPT-5.x, Codex, Claude Opus / Sonnet / Haiku / Fable.
• Codex uses less of your plan: Codex now has its own default effort, Medium (Settings › Agents › Codex effort). High is still one click away per chat.
• Leaner agents: Forge's instructions no longer push extra checks on every edit: agents batch their editor edits, compile once, and only playtest when the task is about runtime behaviour.

• A message that is only a slash command (like /fast) no longer shows a "Kitwright isn't answering" warning: Forge never starts the editor for those, on purpose.
• The chat header no longer shows a red "kitwright failed" badge while the editor is simply closed.

• Rebuild C++ (Unreal C++ projects): one click on the project card, the Editor card or the chat's editor chip saves and closes Forge's own editor, builds the editor target with Unreal's Build.bat, shows the errors (and the log), then reopens the editor the way it was. Agents can ask for it too: they end a reply with FORGE: REBUILD and continue with the result. Forge never closes an editor you opened yourself — use Live Coding there.
• Review changes: every turn with a restore point gets a Review button: the files it changed with a readable diff, Unity scenes and prefabs grouped by object ("Player › Transform: position 0,1,0 → 0,2,0"), binary files by size.
• Project journal: each project keeps a short journal of what its chats did (pin, edit, delete, add notes on the project Overview). New chats read it, so they don't start from zero. Settings › Agents › Project journal.
• Usage (Settings › Usage): your Claude and ChatGPT limit windows with a forecast ("at this pace the 5-hour limit runs out around 02:10; it resets 03:00") and what used them, per project and day. Check before I leave uses the forecast.
• Playtest proof: new built-in recipe "Playtest & bug report", and Schedule steps can "Playtest when done": the agent plays the game in the editor, takes screenshots, reads the console and reports a bug list.
• Editor open: when your editor is already open, the start box and chat header say "Editor · Open" instead of offering a Hidden/Window pick that would change nothing.
• Boot animation: Forge opens on its logo glowing in over the night sky while it loads.
• Hidden (renders) · experimental for Unreal: a hidden editor that still renders off-screen, so agents can take screenshots with no window (about 6 GB RAM instead of 2.5). Off by default (Settings › Performance › Hidden Unreal, or the chat's Editor picker); Schedule steps with a screenshot or playtest use it. If it can't start twice, Forge goes back to headless.

Forge can now be left alone for a day.

• Schedule (new page in the sidebar): build a sequence of prompts — "do this; when it's done, do this; then this" — and click Start (now, at a time, or every day). Each step starts a new chat or continues the previous one, with its own model and effort. Optional per step: check it compiles afterwards, commit when done, screenshot when done; per run: work on a new git branch. Failed steps retry, pause or skip (your choice). Pause, Resume, Stop, Retry step and Skip step while it runs.
• Keep going (Settings › Agents, on by default): connection and provider errors retry with backoff for about an hour; a crashed or stuck agent (silent for 20 min) is restarted and continues; a failed start retries; usage limits wait for the reset and continue on their own; an expired sign-in continues once you sign in again; a full conversation continues in a fresh one with a handoff. Chats cut off by a Forge crash or a Windows restart resume on their own. A deliberate quit is never undone.
• Nobody answering: after the "If you don't answer" time, agent questions decide themselves, plans are approved, MCP forms are declined, the editor card continues without the editor, and permission requests are denied (never approved). Every card shows the countdown and "Wait for me".
• Forge comes back: after a crash Forge restarts itself (capped), and after a Windows restart it opens again at sign-in while work is pending (Settings › Agents › Reopen Forge).
• While you were away: the dashboard shows what happened since you left — done, failed, recovered, waiting for you, schedules.
• Phone notifications: paste an ntfy topic or a Discord webhook (Settings › Agents); Forge sends short status lines while you're away — never agent text or file contents.
• Check before I leave (Schedule page): accounts, usage headroom, NeoStack sign-in, disk space, Unity scene backups, phone and keep-going settings, with fix buttons.
• Unity scene backups: Forge's save + quit of a hidden Unity now really quits (it never did on the hidden desktop), and after a confirmed save leftover backups are set aside (Forge keeps a copy), so Unity stops asking "Recovering Scene Backups" after Forge closed it. Real crashes still keep the backups for you.
• Editors that keep crashing now get another automatic start after a cooldown (30 min, 1 h, 2 h) instead of pausing for good (Keep going).
• Held chats say why: "Usage limit · continues 03:02", "Retrying at 14:05", "Sign in needed".

• Send later: the clock next to Send (and Start chat) sends a message in 30 min to 4 h, at a time of day, or when a chat finishes (this one or another). Scheduled messages show above the composer, on the project page and under Agents, with Send now and Cancel, and survive a restart.
• If you don't answer: an agent's question or plan review decides itself after 3 minutes (Settings › Agents: 3, 10, 30 min or Always wait). The card counts down; Wait for me keeps it waiting. Permission requests always wait for you.
• Windows stays awake while an agent works or a message is scheduled (the screen can still turn off).

• Fix: automatic updates. Every check failed ("Could not reach the update server") because the updater never loaded in the installed app. Versions 1.0.0–1.0.12 cannot update themselves: download 1.0.13 once from the website; updates work from then on.

• Fix: Codex chats on Unreal that said the editor connection dropped and asked you to reconnect. Codex keeps the tool list it got when the chat opened, so a chat that opened while the editor was still loading never got NeoStack's editor tools. Forge now notices, reopens the chat with the editor connected and continues on its own.
• Agents wait for the editor (wait_for_editor) instead of asking you to reconnect, and check a timed-out editor step before redoing it.

• Cleaner chat bar: Autonomy and Recipes are gone from it. What agents may do without asking is now Permissions in Settings › Agents (and per project); recipes live on the Recipes page.
• Website: proper link preview (summit, logo, what Forge does) when the link is shared on Discord, X and friends.

• New logo: an F whose top is a snowy summit. App, taskbar, tray, installer and website.

• New look to match the website: night palette, snow-white buttons, ice-blue accents, Archivo headings, the summit behind the dashboard.
• The dashboard greets you by name (from Windows, or Settings › Appearance › Your name).

• Agents run out of sight: Claude Code and Codex run on Forge's hidden desktop, so programs they start (a test build of your game, tools) open there instead of on your screen. Toggle: Settings › Agents › Advanced › Run agents out of sight.
• Git never pops up a login window for hidden agents (saved logins still work).

• When an editor needs you first (Unity scene backups, Unreal package restore), the chat pauses with a card: Open Unity/Unreal once, or continue without the editor. After you answer, Forge closes it, starts it hidden again and continues.
• Agents are reminded every turn not to start or kill editors themselves, and to test gameplay in the editor Play mode instead of launching built games.
• Less noise when Kitwright/NeoStack is not ready.

• Unreal projects without NeoStack now say so: a "No NeoStack" line on the card and a note in the start box (agents can only edit C++ and config; Blueprint-only projects almost nothing). In a Blueprint-only project without NeoStack the agent says this first instead of trying.

• Cleaner project cards: name, one status line with a single Show/Open action, chats with a dot only when something is happening. Engine version, bridge badge (unless it needs attention), git line, CHATS label, provider icons and Done pills are gone from the card.
• No Done pills in project chat lists either.

• Refined look: Geist typeface, calmer status labels (Done is no longer a green pill), editor load text no longer cut off on cards.
• New website.

• A Unity project with scene backups from a crash is never started hidden (Unity would wait on an invisible question); Forge asks you to open it once.

• Low free memory only delays a chat for up to 45 s, then it starts anyway (it could stay queued forever with big editors open).
• Agents may no longer start or kill Unity/Unreal editors themselves; Forge owns the editors.

• Settings → Updates: version, update check when opened, download progress and Update now.
• A notification when an update is ready.

• Hidden / Window launch mode per chat, with an app and a project default. Show editor turns a hidden editor into a window.
• Background Unity runs on an invisible desktop (no window, taskbar button or pop-ups). Background Unreal runs headless.
• Crash saver: one editor per project, adoption of leftover hidden editors, handover when you open a project yourself, restore points with Undo, autosave recovery.
• Crash recovery for editors Forge started for agents (never for editors you closed), and resume of agents whose process died.
• NeoStack: agents get the installed skill guides, your NeoStack agent prompt and project notes. One Sign in action. No editor kills during sign-in.
• Model, effort and editor pickers. Honest context size. Claude and ChatGPT usage dropdowns.
• Notification icon and finish chime. Clearer image-generation errors.
• Codex installed with npm, pnpm or bun is found automatically.
• Isolated test mode. Cleaner UI with less clutter.