Release Notes

v1.2.0 — October 14, 2026

A major feature release. TatsuCode can now find and change things across your whole project in an instant, it works with Autodesk Maya, it keeps a live checklist of its plan whichever model you use, and you can watch Claude Code's helpers at work while you keep typing. The Unity, Unreal, Godot and Blender plugins all got stronger, and the built-in browser now types, hovers and long-presses like a real person.

New Plugin

• Autodesk Maya — TatsuCode now works with Maya 2022 and newer. Ask it to connect and it sets everything up, then it can model, rig, animate, simulate, bake textures and render in your open Maya, or in one it opens for you. Many steps can run at once, and they undo together
• A Maya toolbox that grows — Tools the agent writes for you are saved in a TatsuScripts library you can put on shelf buttons, menus or hotkeys. They keep working even when TatsuCode isn't running

New Features

• Search everything, fast — Find text, files or folders in your project, in any folder, or across every drive, usually in a fraction of a second. Long searches keep going in the background while the agent works on, and a live card shows what's been found, with a Stop button
• Find & Replace across many files, safely — The agent shows you every change before anything is written, leaves out what shouldn't change, and can undo the whole batch in one step. It can also add a license or copyright header to a whole folder of files
• A live task checklist for every model — When the agent plans its work, one checklist card shows what's done, what's in progress and what's next, with a progress bar. It works with every model, and if you switch models halfway through, the new one carries on with the steps that are left
• Watch Claude Code's helpers at work — When Claude Code hands work to its own helper agents, a panel above the message box shows each one live. The chat no longer looks frozen while they run, and you can keep typing
• Jump between your own messages — Press Ctrl+↑ and Ctrl+↓ to move through the messages you sent
• A browser that behaves like a person — DevBrowser now types with real keystrokes, hovers to open menus, presses and holds (long-press menus, game controls), and right-clicks or double-clicks anywhere on the page, including inside embedded pages and canvas apps it couldn't reach before. Dropdowns can be picked by the label you see
• Unreal Engine 5.4 to 5.8 — The Unreal plugin now works with every engine version from 5.4 through 5.8, and always opens a project in its own engine version
• Close an app when you're done — Ask TatsuCode to close Unity, Unreal, Godot, Blender or Maya after a session to free up memory. Your work is saved first, and if the app asks "save changes?", the choice stays yours

New Models

• GPT-6.1 Sol (OpenAI Plus/Pro) — OpenAI's new default coding model: close to GPT-6 Astra on hard work, at a lower cost, with a very large context
• Claude Code: Sonnet 5.5 and Haiku 5.5 — Both can now be picked in the model menu and for delegated Claude Code runs

Improvements

• Your files stay exactly as they were — Edits change only what you asked for: a file's text format, line endings and permissions are kept, so your diffs stay clean. Files are saved in one step, never left half-written, and read-only files are left alone with a clear message
• More dependable edits — An edit changes exactly the text it targets and won't guess when that text appears more than once. "Replace all" now really replaces them all, and diffs show changes anywhere in a long file
• Faster and safer than shell scripts — When the agent reaches for a slow shell search or a risky find-and-replace script, TatsuCode points it to the faster built-in tools instead
• Messages you type while the agent works — The agent now sees what you wrote and reads it at its next safe moment. It also knows when you sent it, so it can tell what may already be done
• Several windows, one set of settings — Changing a setting in one TatsuCode window is no longer undone by another open window
• Unity: more versions, clearer answers — Works with Unity 2021.3 and 2023.1 as well as Unity 6. When Unity isn't responding, TatsuCode tells you why (a dialog waiting for you, or compile errors), and the agent can read the latest compile errors itself. Big asset-package imports no longer hold the agent up, and imported models can go straight into a scene
• Blender: exports and files — Exports to FBX or glTF with the options you'd choose by hand (textures, scale, selected objects only, animation), opens and creates files without ever throwing away unsaved work, and stays responsive during long jobs
• Sharper plugins all round — Cleaner FBX exports and captures from Maya, Unreal property changes that actually stick (such as assigning a mesh), and renders saved where you asked are kept
• No more install prompts for plugins — The agent tells you in one line what it's installing and where
• Reopened sessions look the way they did — Tool cards, edits, queued messages and stopped turns all come back as you saw them
• Smoother on OpenRouter — Claude, Gemini and other models reached through OpenRouter make fewer tool-call mistakes

Fixes

• Claude Code's final answer could go missing — After Claude Code asked you a question partway through a task, its final answer sometimes never appeared
• A message sent just as Claude Code started a task of its own — It failed with an error. Now it waits its turn and goes through
• Claude Code's background helpers — They no longer appear stuck in the agents panel, and they can't use TatsuCode's tools on their own
• Task checklists — An old task no longer keeps a checklist open forever, and a finished list no longer shows a second "All done" card
• Short answers — Replies like "42" or "No." were treated as empty and retried
• The @ file list — Files saved by some editors (Visual Studio, Office) dropped out of it
• Unity — Console messages are labeled correctly, Play mode no longer fills with errors in projects that use only the new Input System, and captures work while Unity is in the background
• Unreal — A project could open in an already-running editor of a different version
• Godot — Multi-angle captures came back blank
• Stopped Claude Code turns — Their work went missing when the session was reopened
• Accented characters — Large files saved in older formats showed them as broken symbols
• An empty model picker — Older settings could leave it empty after OpenAI retired its older models

Removed

• GitHub Copilot — The Copilot connection, switched off since June 4, is gone. Old Copilot models are removed from your list, and sessions that used one reopen on another connected model. OpenRouter reaches the same models
• Retired OpenAI models — OpenAI retired GPT-5.5, GPT-5.4 (with Mini and Nano) and GPT-5.3 Codex. A saved choice moves to GPT-6.1 Sol, or to GPT-6 Luna for the smaller ones

Notes

• Maya — After the first install, restart Maya once (or let TatsuCode open one for you). Final Arnold renders come from a Maya with a window: renders from a Maya running in the background carry Arnold's watermark
• Sonnet 5.5 and Haiku 5.5 need a recent Claude Code — If they don't start, run claude update
• Undo history is new in 1.2 — Undo history saved by 1.2 can't be read by older versions, so finish any undos before going back to one
• Unity 2019.4 isn't supported — The plugin needs Unity 2021.3 or later
• Compact Now with Claude Code — Claude Code compacts its own conversation when it fills up, so Compact Now tells you that instead of sending a command Claude Code would ignore

v1.1.3 — September 22, 2026

A second same-day model-support release. OpenAI's brand-new GPT-6 Sol and GPT-6 Luna, released today, are now selectable under OpenAI Plus/Pro, joining GPT-6 Astra.

New Models

• GPT-6 Sol (OpenAI Plus/Pro) — OpenAI's new GPT-6 workhorse for coding and everyday work, available as GPT-6 Sol in the model picker. It brings much of GPT-6 Astra's strength into a faster, cheaper model, and OpenAI reports it makes about half as many factual mistakes as GPT-5.6 Sol. It has an 872K-token context, reasoning effort up to Max, native web search, and computer use, and starts at High reasoning effort like the other Sol-class coding models
• GPT-6 Luna (OpenAI Plus/Pro) — OpenAI's fast, affordable GPT-6 tier for easier tasks, available as GPT-6 Luna. It's built for quick iteration where speed matters more than peak reasoning, keeps tool use and native web search, and also has an 872K-token context. OpenAI makes Luna available on most plans, including Free and Go

v1.1.2 — September 22, 2026

A same-day model-support release. Anthropic's brand-new Claude Opus 5.5, released today, is now selectable in Claude Code, both as your main model and for delegated agent runs.

New Models

• Claude Opus 5.5 (Claude Code) — Anthropic's newest Opus, released September 22, is available as Claude Code: Opus 5.5 in the model picker. Anthropic reports it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than Opus 5, with faster output. Its full 1-million-token context window is recognized from the first message, so the context bar and the /usage panel stay accurate. The existing Claude Code: Opus entry keeps following whatever your subscription treats as the current Opus model; the new entry pins Opus 5.5 specifically, so you always know exactly which model you're on
• Opus 5.5 for delegated agents — The Claude Code Manage panel's Model dropdown gains the same Opus 5.5 option, so tasks the main agent hands off to Claude Code can run on it too

Notes

• Switching keeps your conversation — Moving an existing Claude Code conversation to Opus 5.5 resumes Claude's own transcript, so no context is lost

v1.1.1 — September 18, 2026

A small follow-up to 1.1.0. OpenRouter now carries TypeSafe's Jev decision model, so TypeSafe AI can run on the OpenRouter key you already connected, with no TypeSafe account required.

New Features

• TypeSafe AI through OpenRouter — The TypeSafe AI setup in Provider Connections → Decision Models has a new Access choice. Pick TypeSafe API key to use your own key as before, or pick OpenRouter to reuse the key already under API Access; decisions then run as ~typesafe/jev-latest and appear on your OpenRouter bill. Both reach the same model, and the agent uses the TypeSafe tool exactly the same way on either, so nothing about your prompts or the typesafe-ai skill changes. If you have an OpenRouter key and no TypeSafe key, TatsuCode uses OpenRouter automatically and the skill works with no extra setup. The OpenRouter choice stays unavailable until an OpenRouter key is connected

Improvements

• Decision cost on the tool card — When a decision runs through OpenRouter, the TypeSafe tool card shows via OpenRouter and what the call cost, next to the model, timing, and token count
• Save & Verify proves the route works — For OpenRouter there is no key to type, so Save & Verify makes one small yes/no decision with your stored key to confirm it can reach the model. If the check fails, your previous Access choice is kept, so a working setup is never swapped for one that cannot answer
• A clear fix when OpenRouter blocks the provider — If your OpenRouter account limits Allowed Providers and TypeSafe is not on the list, the dialog says so in one sentence and links straight to your OpenRouter privacy settings. The message stays in the dialog until you retry instead of disappearing with a notification
• The Provider Connections row shows the route — TypeSafe AI reads Configured · via OpenRouter when that is how it is set up

Fixes

• A key or setting changed in one window now reaches your other open windows — Each TatsuCode window kept its own copy of your credentials for as long as it stayed open, so a key added or a provider setting changed in one window was not seen by the others until they were restarted, and saving something in one of those older windows could undo the change. Windows now notice within half a second when another one has saved, and a save always builds on the latest version. Saving is also done in one step, so a window can never read a half-written file
• A Claude Code window could show a name other windows could not message — The name in a window's title bar and footer badge is the address you give another window's agent ("ask tetris-7434b516 to run the tests"). Occasionally a window picked up a temporary name that Claude Code uses for a split second while starting, so the title showed something like tetris-81 and messages to it failed with "no agent by that name". The title and badge now always show the reachable name. For a name that is easier to say, rename the session with /session-rename and the address follows it
• The Send Message card says what happened — It now reads to the window it was sent to, and then delivered or not delivered with the reason, instead of leaving a failed delivery to a raw receipt
• TatsuCode started from a terminal inside Claude Code could not chat — The first message failed with "Claude Code never created its session transcript". TatsuCode now starts its own Claude Code session cleanly no matter where it was launched from
• Key fields in provider setup dialogs were narrower than the dropdown above them — A field's width followed the length of the help text under it, so a short hint produced a short field. Fields now fill the dialog the same way in every provider setup dialog

Notes

• On the OpenRouter route, requests pass through OpenRouter first — They are sent with your OpenRouter key, served by TypeSafe, and covered by both services' terms. See Third-Party Services and Decision Models for details

v1.1.0 — September 17, 2026

A feature release, and the reason for the version bump. TypeSafe AI joins Provider Connections as a decision-model provider, so the agent can classify, route, rank, and verify with typed answers and calibrated confidence instead of generated text. And Cross-Talk extends window-to-window messaging to every provider: a Gemini window, an Azure window, a local-model window, and a Claude Code window can all reach each other on the same machine.

New Features

• TypeSafe AI decision models — TypeSafe's System One model (Jev) answers typed questions about a piece of state: pick one of up to 255 options, place it on an ordered scale, or say yes or no. Every answer comes with per-option probabilities and a confidence figure, arrives in roughly 100 to 500 ms, and can never be schema-invalid because the answer space is the one you supplied. Load the new typesafe-ai skill and the agent gains a TypeSafe tool it uses inside a task (triage a bug list, rank search results, screen a document set, gate a risky action) and a reference for building the same calls into your own code: NPC and simulation decisions in Unity, Godot, or Unreal; intent routing and guardrails in an app; research screening; CI failure triage. Four bundled guides cover game loops, app patterns, agent control, and workflows by role. Bring your own key: Provider Connections → Decision Models → TypeSafe AI. The tool validates questions locally before spending a request and tells the agent, in one sentence, what to ask you for when the key is missing
• Cross-Talk between TatsuCode windows, for every provider — Models could already talk to other models inside one window, through task agents, delegated Claude Agent runs, and Cross-Agent Relay. Cross-Talk adds the second kind: messaging between windows. Every window now has an address, a name built from the project and the session, and a Peers tool with list, send, status, and whoami. A message delivered to an idle window starts a turn there under a distinct Message from <window> block; a message to a busy window is queued and handed to its model at the next safe point, and the sender's card says which happened. Sessions restore the blocks in place. Claude Code windows keep their native session messaging for other Claude sessions and use the same Cross-Talk bus, as tatsu_peers, to reach everything else

Improvements

• TypeSafe's mark in the UI — TypeSafe AI appears with its own icon in Provider Connections, in the tool cards, and in the working banner, so a decision call is easy to tell apart from a chat model call at a glance
• Decision Models row in Provider Connections — A new row under /connect holds the TypeSafe key with Save & Verify, which checks the key against TypeSafe's API before storing it locally alongside your other credentials

Notes

• Cross-Talk is loopback only — Windows find each other through per-window files under %UserProfile%\.tatsu\peers\, each carrying a random token that is required on every message, and the sender must present its own token too. Nothing listens on the network, and no message leaves the machine. Chains are capped at eight hops and senders are rate-limited, so two windows cannot talk each other into a loop
• TypeSafe is a separate service — Decision calls go from your machine straight to TypeSafe AI under your own key and their terms. TypeSafe processes requests in the United States and does not train on customer inputs; see Third-Party Services for the details and links