Cerevisor 3.5.0: What's New

3.5.0 brings new models and agents you can talk to while they work. Claude Opus 5.5 and OpenAI’s GPT-6 Sol and GPT-6 Luna are available. You can now send instructions to a Claude Code agent while it runs, and a Codex agent can stop and ask you a question instead of guessing. Split view is easier to use from the world dock, each pane now keeps its own world, and signing in to Claude uses Claude’s own browser login.

1. New models

• Claude Opus 5.5 ($4 / $20 per million tokens) is available wherever Claude models are offered. A model setting that simply says “Opus” now means Opus 5.5. A workflow saved with a specific model keeps that exact model. Opus 5.5 always thinks before it answers: the thinking picker still sets how hard it works, but it cannot be switched off.
• GPT-6 Sol ($2 / $10) and GPT-6 Luna ($0.10 / $0.50) are available for Codex and OpenAI-compatible connections, with reasoning effort and image input. Codex’s default model is unchanged, and no saved choice is moved.

2. Talking to an agent while it works

Claude Code agents take new instructions mid-run. A message you send to a running Claude Code agent is delivered at its next tool step, or just before it would finish, so it can change course. Sending a message never grants a permission and never interrupts a tool that is already running. For a tool that cannot take instructions mid-run, the message is not sent and Cerevisor tells you so, instead of losing it silently.

Codex agents can ask you a question. When a Codex agent needs a decision from you, it now shows the question (with options, when it has them) in Cerevisor’s question panel and continues with your answer. Before, it stopped with a report saying it was blocked. Your answer never loosens the agent’s sandbox or tool limits.

Agents that run in their own tool are told their real file scope. Codex, Cursor, Antigravity and Grok Build agents are now told what they may actually touch: their working folder plus the files the workflow declares. Before, they were given the declared files alone, which described a narrower limit than the one really in place.

3. Split view from the world dock

Drag a world’s chip from the dock to the right or bottom edge of the canvas to open a split. This now works for every world, including the one you are in. Open in Right Split and Open in Bottom Split from the chip’s menu work the same way. If a split cannot open, for example because the file has only one world, Cerevisor says why instead of doing nothing.

Each pane now keeps its own world. Switching worlds from the dock or with Alt/Ctrl+1–9 moves only the pane that shows the chosen world, and the other pane’s world no longer slides away. Keyboard shortcuts and canvas controls act only on the pane you are working in. Dropping a dragged chip also works reliably with a real mouse now. Before, drops were often lost.

4. Signing in to Claude

Sign in with Claude now uses Claude Code’s own browser login, and Claude keeps and refreshes that sign-in itself. A successful sign-in is then confirmed by a real request, not taken on trust. When a check fails, the message now says Claude rejected the credentials, rather than claiming you are not signed in.

Also fixed

• Progressive mode uses the builder model you chose rather than borrowing the background model, and pressing Stop while it is planning stops it.

Known limits

• Mid-run instructions reach Claude Code agents and Cerevisor’s own agents only. Codex, Cursor, Antigravity and Grok Build agents refuse them with a sentence. Stop the agent and send your instructions to restart it.
• A Codex agent asks only when it decides to. It asks by ending its turn with a question. A Codex request for desktop (Computer Use) access still cannot be approved from inside Cerevisor. That work is planned in docs/plans/2026-09-22-computer-use-approval-recovery.md and is not in this release.
• Everything listed under 3.1.0’s through 3.4.0’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 six end-to-end specs were run on one Windows development machine on 2026-09-23, on code identical to this release’s.

The type checks passed with no type errors left anywhere in the app’s code, for the first time. The desktop-channel contract check passed. In the full run, 973 of 979 test files and 11,914 tests passed. Five files did not: four of the slowest ran out of time while the whole suite was running at once, and one was a generated index that had not been refreshed. The index was regenerated, and all five then passed on their own (103 tests). The production build succeeded, and all 20 end-to-end checks passed across split view and edge drop, the new models, the agent question panel, first launch and a real recorded run.

Nothing in this release has been exercised by hand in an installed app. Sending instructions to a real Claude Code agent mid-run, a real Codex agent asking a question, a real Claude browser sign-in and the new models against a real account were checked only by automated tests. The phone-app checks, the coverage check and the full end-to-end suite 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.