Cerevisor 3.6.0: What's New

3.6.0 keeps several worlds open at once and can show two different worlds side by side in one window. The app now starts on Home, with picking a folder as the first step. Agents on external tools get stuck less often and say why when they stop. You choose whether workflow agents use your personal Claude Code settings. Decision cards read what you gave them, and dialogs now truly block what is behind them.

1. Several worlds open at once

The pills in the top-left dock are now your open worlds, not the workflows of one world. Open another from the dock’s + → Open world…, or double-click a world file, and it joins the others as its own pill. Clicking a pill switches at once, with no save prompt. The world you leave keeps its unsaved edits, its camera and its runs, which keep going. Ctrl+Tab and Ctrl+Shift+Tab move between worlds.

Each pill shows unsaved changes and whether a run in it is going, waiting for you, done or failed. Closing a world, or the app, asks about every world with unsaved changes. On restart, the worlds you had open come back, waiting in the dock and on Home, one click away.

2. Two worlds side by side

Choose Open beside from a pill’s menu, or drag the pill onto the band at the window’s right edge. That world opens in the right half, fully editable, with its own header, Save and Close. You can edit, run and chat in both worlds at once. Each save writes only its own file, and a run’s questions appear on the side that started it. Drag the divider to resize. Swap sides trades the two, and Close split brings the world back as a pill with its unsaved changes intact.

A Stop on one side’s run pill now stops only that side’s run. Before, it stopped every run in the app.

This replaces 3.5.0’s split view, which showed two views of the same world and could also split to the bottom.

3. A simpler first run

• Cerevisor always starts on Home, including after a restart with worlds open. Provider setup and the first-project guide no longer pop up over it. The main button now reads Pick the folder to work in. A new Model providers button in the title bar opens the Provider Hub when you want it.
• Entering a world opens the Workflow Builder chat, whether you picked a folder, opened a saved world or imported one. In a new world the builder sits beside the workflow rail, with the empty workflow to its right.
• The guide is optional. Guide me through my first project starts or resumes it. It explains a workflow in everyday language. On a guided first run the canvas follows the working agent, with a status and a pause control, and at the end points you to the finished file. It never clicks or opens anything for you, and it respects reduced motion.

4. Fewer stuck agents

• Claude Code agents no longer lose the ability to ask for approval mid-run. A quiet Chat with agent reply stops after 10 minutes with a plain sentence instead of hanging.
• Codex agents allowed to change files can use the internet whenever web access is on, and the note before a run says exactly what Codex can reach. A Codex question waits 30 minutes, not counting paused time. A long command is no longer mistaken for a stuck agent.
• Cursor: Stop works while an agent is still starting. Antigravity: a sign-in done in a terminal is picked up within 2 minutes. Grok Build: the Workflow Builder’s Approve and Deny buttons now work on it.
• Questions on the board wait up to 30 minutes, then the agent carries on and the board says nobody answered.
• Agents on Cerevisor’s own engine get four times the token allowance per run, and the run says so if one uses it all.
• A brief Windows file lock from antivirus or a backup no longer stops a run. Cerevisor retries for about a second without freezing.
• Scheduled and Hands-free runs with web tools say up front what agents cannot do with nobody there to approve. After a web refusal, the error card offers Open web-access settings.
• If your plan no longer includes connected tools, the run names each one it left out instead of dropping it silently.

5. Your Claude Code settings, your choice

Each Claude Code connection in the Provider Hub has a new checkbox, Use my personal Claude Code settings in workflows. It is on by default, so workflow agents get your skills, plugins, hooks, allowed commands and extra folders, as in 3.5.0, and the note before a run says so. Turn it off to give agents only the project folder’s settings.

Secrets you assign to a Claude Code agent from the vault now reach it, in runs and in Chat with agent.

6. Decision cards that read what you gave them

• A decision card with no agent before it now judges what you typed in the launcher’s message box, as well as its notes. An empty input file on the card is treated as a placeholder for that text, not a failure.
• The Workflow Builder no longer puts input files on a decision card, and says so.
• Decisions reach the next agent as one table: question, answer, how sure, who answered and a note. Every decision arrives.
• With TypeSafe Jev connected, many decisions go out together in a few requests. If you turn the decision model off or clear its key mid-run, nothing more is sent to TypeSafe.
• The canvas warns when a decision’s input reads like an instruction, or its instructions go unused.

7. Dialogs that block, and a calmer canvas

Settings, Agent Config, the Provider Hub, the pre-run summary and every other blocking dialog now block what is behind them, including the chat builder and the workflow rail. Focus moves into the dialog and returns when it closes, and the dialog you can click is always the one on top.

While you work in the chat builder or the run window, the rest of the app’s controls fade until you point at them. Selecting an agent keeps it and the cards wired to it bright. An agent asking you something never fades.

Also fixed

• A launch refused before it starts ends the run and says why, instead of spinning on “pending”.
• Stopping a run reads as a stop, with its cost, not an error.
• Sharing works for a folder reached through a link, such as a symlink or a Windows junction.
• The Workflow Builder no longer refuses an agent the model named with a short name it never declared.
• The secret picker on an agent card opens beside the agent’s chat, and stays inside the window when the window is made narrower.
• After Cerevisor closed during a run, View recovered work opens the run window. Before, from Home it showed nothing.

Known limits

• The world beside: one per window, on the right only. A world with a run in progress cannot be opened beside, swapped or brought back until the run finishes. Undo history does not cross the split. At zoom other than 100 % it can sit in the wrong place, and a few menus and floating panels can still sit under it.
• Ctrl+R closes the split. A saved world beside loses edits since its last automatic save, and a run going there loses its window.
• A run of a workflow added inside the world beside is not seen by the main side, which may let the split or the app close without asking.
• On a narrow window, or with a long world name, the right end of the app, including Run, can sit past the window’s edge.
• The phone follows the main side only.
• A Codex agent with web access ignores your approved-sites list and the Off setting; the note before the run says so.
• Canvas shortcuts such as Delete and Ctrl+Z still act on the canvas behind an open dialog.
• Everything listed under 3.1.0’s through 3.5.0’s known limits still applies.

What this was verified on

The whole automated suite, the type checks, the desktop-channel contract check, a production build and the full end-to-end suite were run on one Windows development machine on 2026-09-24, on code identical to this release’s.

The type checks passed with no type errors anywhere in the app’s code, and the desktop-channel contract check passed. In the full run, 1,036 of 1,042 test files and 12,656 of 12,670 tests passed. Five files ran out of time while the whole suite was running at once; all five then passed on their own (35 tests). The production build succeeded. In the full end-to-end suite, 56 checks passed and 3 were skipped. Four did not pass: one clipboard check that then passed on its own, and three that have failed the same way on every build for several releases (the contact-avatar composer and two phone-companion checks, which need a real phone setup).

Putting the pieces together found and fixed three problems that no single piece of work showed on its own: restarting with worlds open skipped Home, View recovered work showed nothing from Home, and the secret picker could leave a narrowed window.

Nothing in this release has been exercised by hand in an installed app. Open worlds and the split, the new first run, and the agent fixes against real Claude Code, Codex, Cursor, Antigravity and Grok Build accounts were checked only by automated tests. The phone-app checks and the coverage check were not run for this release.

Platforms

Available for Windows, macOS (Intel and Apple silicon) and Linux (AppImage and Debian package), all built from the same tagged tree by the release workflow.

Updating

Existing installations receive the update through Cerevisor’s built-in updater, and the installers are available from cerevisor.com/download and the private GitHub release. Your workflows, run history and settings carry forward untouched; no file format changed.