Cerevisor 2.3.0: runs that survive restarts, a searchable World Library, and run results you can trust

Cerevisor 2.3 makes long-running work easier to trust, resume, and manage. Runs survive restarts, large worlds stay responsive, and the new World Library turns saved files into recognizable ongoing missions.

Worlds are easier to find and manage

• The home screen is now a searchable, sortable World Library with card and compact views.
• Worlds show their intent, team size, workflow count, working folder, and last-active time.
• Chat Builder can give a new world a concise mission name once it understands the job.
• You can rename a world from the library or canvas. A manual name is never overwritten automatically.
• Missing or malformed recent files degrade safely instead of breaking the library.

Large canvases stay smooth

• Dense worlds automatically reduce expensive decorative effects while preserving all controls and execution state.
• Agent cards switch to a clearer compact presentation at low zoom.
• Pan and zoom gestures pause non-essential animation, and edges no longer force unnecessary GPU layers.
• Rapid streaming updates are grouped into animation frames, reducing renderer churn during busy runs.
• Effective-input labels are indexed once instead of repeatedly recomputed across the canvas.

Runs survive app restarts

• Active and completed run state is persisted and restored with the world.
• World-level run badges and staleness hints make it clear what is still active and what was recovered from an earlier session.
• Concurrent runs are tracked by session so one workflow cannot overwrite another workflow’s status.

Run results now tell the truth

• A workflow no longer shows a clean green completion when agents failed, required checks were blocked, or a reviewer explicitly returned FAIL.
• Deliberately stopped runs are shown as stopped, not as errors.
• Failed agents are counted correctly in history and learning data.
• Undeclared file changes are visible in the execution log.
• Read-only Claude Code agents are prevented from writing files or running shell commands.

Provider failures recover more gracefully

• Missing API keys are caught before launch with a direct Settings path.
• An expired Claude Code sign-in is remembered for the session, preventing every remaining agent from repeating the same failed startup.
• Usage-limit failures can make one bounded attempt on a user-selected fallback provider, with spend attributed to the provider that actually ran.
• Optional auto-resume can continue a quota-limited run when the provider’s reset window arrives.
• Grok Build is now a first-class subscription provider with device sign-in, live permission approval, cost accounting, and post-run resume chat.

Better building and output controls

• Chat Builder clearly labels a response that only produced a plan and understands more direct build verbs.
• Mid-run steering is delivered at safe points instead of being stranded or dropped.
• Progressive building supports per-message step limits, a live steps-remaining count, pace changes, prior conversation context, and image delivery.
• A workflow-wide model control applies model and thinking settings to every agent as one undoable action.
• Workflows can ask their final agent to produce a phone-friendly HTML run report.
• Handover files now support Terse through Exhaustive length levels and an optional consistent, scannable structure.
• Slow activity messages rotate gently while making it clear the app is still working.

Reliability foundation

This release also completes Cerevisor 3.0 foundation Phases 0–3 under the hood: stricter type and IPC boundaries, a durable append-only run ledger, and a deterministic execution kernel with replay and recovery coverage. These changes are intentionally compatibility-first; incomplete Phase 4 gateway work is not included in 2.3.

The real project type gates now run in CI, the fatal-error baseline cannot grow silently, and the full test/build matrix covers Windows, macOS, and Linux before release artifacts are published.

Updating

Existing installations receive 2.3.0 through Cerevisor’s built-in updater. New installers are available from cerevisor.com/download and the private GitHub release.