Cerevisor 3.1.1: What's New
· v3.1.1
Cerevisor 3.1.1 puts the agents' conversation on the canvas beside the workflow, lets a new run start from the board its last run ended on, uses your own newer Claude Code install when you have one, and fixes a setting Codex CLI 0.153 no longer accepts.
3.1.1 is a small release on top of the board. It puts the agents’ conversation on the canvas next to the work, lets a run start from the board its last run ended on, and fixes two things that changed in outside tools: Claude Code models that the built-in runtime is too old for, and a Codex setting that Codex no longer accepts.
1. The team’s conversation, on the canvas
The notes agents leave for each other (updates, questions, answers, decisions, review exchanges) now show in a conversation panel anchored below the workflow on the canvas. It follows the canvas as you pan and zoom, shows who wrote to whom, the kind of note, the wave it came from and the reply it answers, and lets you filter by agent (notes they wrote and notes sent to them). A strip beside it shows each agent’s live activity or finished output and opens the full activity view. Live output is never shown as a message between agents. The panel follows new notes while you are at the end; scroll up and it holds still, with a catch-up button to return to the latest. It starts with the most recent sixty notes and shows earlier ones on request; a long note is trimmed with a “Read full note” control.
2. Start from where the board left off
When you run a workflow whose last run’s board is on screen, the new run’s board starts with those notes as saved context: the most recent ones up to the board’s usual seed limit, keeping their original authors and the run they came from, shown the way notes from an earlier run are shown. Only the exact run you were looking at is used, never “the newest one”. Nothing an agent wrote is ever turned into an instruction to the next run; it arrives as information, the way every note does.
3. Claude Code agents use your newer Claude Code
Until now an agent on the Claude Code provider always ran the runtime built into Cerevisor. When Anthropic shipped a model that runtime was too old for, every user was stuck until the next Cerevisor release, and the error advised running claude update for a program Cerevisor never used. Now, if the Claude Code installed on your computer is newer than the built-in runtime, Cerevisor uses yours; otherwise the built-in one runs exactly as before. When a model still needs a newer runtime, the message says which runtime spoke and what fixes it: update Claude Code, or pick another model. The built-in Anthropic SDKs were updated alongside.
4. Codex CLI 0.153
Codex removed its untrusted approval setting and refuses to start with it. Cerevisor’s read-only Codex profile now sets approval to never and relies on the read-only sandbox, which already blocks writes and shell. Other profiles are unchanged.
5. Housekeeping
The 68 dependency maps under docs/architecture/dependency-maps/ were regenerated for the 3.1 tree. One new read-only desktop channel lists a workflow’s saved run views (it is what “start from where the board left off” uses); nothing new leaves your machine.
6. What this was verified on
- The full automated test suite passed on this Windows development machine (the files that only time out under full-suite load re-run green in isolation). In CI, the macOS leg passed on the commit before the last test-only change; the Windows and Linux legs could not finish because GitHub Actions stopped starting jobs on the account’s billing. This line is updated when a full three-OS run completes.
- The canvas conversation panel has an interaction test suite and an end-to-end test that drives the built app with real board events over the orchestrator channel.
- Not exercised by hand in the running app for this version: the conversation panel, starting from a saved board, a Claude Code model that needs a newer runtime, and Codex 0.153. Nothing here claims otherwise.
Known limits
- The conversation panel is keyed to the workflow, like the Board tab, so two concurrent runs of one workflow share one view.
- Your own Claude Code is detected once when Cerevisor starts; updating it mid-session is seen after a restart.
- Everything listed under 3.1.0’s known limits still applies.
Updating
Existing installations receive 3.1.1 through Cerevisor’s built-in updater. New installers are available from cerevisor.com/download and the private GitHub release.