Cerevisor 3.0.1: What's New

3.0.1 is a housekeeping release on top of 3.0. It is about the moments a run does not go the way you expected. When a provider’s own reviewer stops one of your agents, you now see why and can approve that action in one click and keep the approval. The four bad days 3.0.0 listed as still broken — a model that never answers, a shortened tool result, a connected server that cannot be reached, and money spent by an interrupted Operator action — now behave. And the run investigation from 2 September landed in full: re-running one agent is checked like a full run, an agent that ran out of room to answer is labelled honestly instead of being called finished, and a handful of small things agents kept tripping over are gone. Nothing about your saved workflows, worlds or transcripts changed shape, and nothing new leaves your machine.

1. When a provider’s own reviewer stops an agent

• When a provider’s own reviewer stops an agent, Cerevisor now shows why and offers a one-click approval you can keep in the workflow — nothing is approved without you.
• The run window shows a card naming the exact action that was refused, with two choices: Approve only, which saves the approval on that agent for next time, and Approve and run again, which saves it and starts the agent again.
• Approved actions are listed on the agent’s own panel, each with a Remove button, so you can see and undo what you have allowed.
• An approval counts as new authority the workflow did not have before, so the pre-run summary asks you about it again the next time you launch — and “Approve and run again” itself goes through that summary rather than around it.

2. Re-running one agent gets the same checks as a full run

• Re-running a single agent now gets the same output checks that agent gets inside a full run, including the automatic second try when what it produced does not match what it declared. That closes the limit 3.0.0 shipped with.

3. Agents that ran out of room to answer

• Some models spend their room on thinking before they write anything. When that happens, Cerevisor now gives the agent one more try with a larger allowance instead of handing you an empty answer.
• If the second try also comes back empty, the agent is labelled ended early rather than shown as completed — a run that produced nothing no longer looks like a success.
• Models that let you choose how hard they think now get sensible limits: GLM 5 and newer and DeepSeek V3.1 and newer are capped so thinking cannot eat the whole answer, and models reached through OpenRouter no longer think without a bound.

4. The Rule of Two, explained in the run window

• An unattended run that has read something from the outside world and could then send something out is still refused — that is the Rule of Two, and it is deliberate.
• What is new is that the refusal explains itself. The run window now shows a card naming which agent was stopped, what it was about to reach, and why nobody could be asked, with a one-click Add a publishing agent fix that restructures the workflow so the reading and the sending are done by different agents. The card appears once per agent per run, not on every attempt.

5. Small things agents kept tripping over

• A version check like “curl –version” or “wget -h”, which names no website at all, is no longer treated as sending anything out.
• Read-only file tools may now read the skills you have installed, so an agent can look at a skill’s own instructions without being refused.
• Asking for a skill that lives inside another skill now works instead of being rejected as unknown.
• The general-purpose tool set now includes web search and web fetch, so a general-purpose agent can look things up without being given a custom tool list.
• An agent running on Antigravity is no longer stopped after five minutes of quiet work; it is given up to six hours, which is what a long piece of real work actually takes.
• A run you launch grants its own folder to the File Viewer, so you can open what the run is working on without adding the folder by hand.

6. Bad days that now behave

3.0.0 listed four situations from its failure testing that behaved worse than they should. All four are fixed:

• A model call that never answers at all no longer leaves that agent waiting indefinitely. It is stopped once the quiet period runs out, and the agent fails with a plain reason.
• A tool result that was shortened to fit is now marked as shortened in the run’s history, so you can tell a trimmed result from a complete one.
• A connected server that cannot be reached at all is now announced in the run window — naming the server and why — and written into the run’s history, instead of its tools silently going missing.
• Money spent by an Operator action that was interrupted by a crash or a quit now reaches that day’s total when the app comes back, instead of disappearing.

7. The Operator’s runs

• The run window now tells you plainly why a run the Operator started asks you for approvals, or why it never will. If the workflow saved its own setting, the note says that; only otherwise does it name the Operator’s own hands-free default.
• A Power reached for by a run the Operator started is now recorded as part of that Operator action, so the record joins up — but it still needs your hands-free acknowledgment like any other. Being started by the Operator does not waive it.
• The Operator’s history no longer loses an entry written in the last moments before midnight.

8. Housekeeping

• An agent pinned to a model provider you have since removed is now refused before the run starts, with a plain sentence naming the missing credential, instead of being quietly moved onto a different provider you never chose. This applies to the Operator’s own unattended runs too.
• A leftover setting from before 3.0 is removed from your settings file the first time 3.0.1 reads it.
• OpenAI’s GPT-6 Astra (released 3 September 2026) is now in the model picker with its cost, its 1.05M context, reasoning effort up to the highest setting, and image input. It is also offered for Codex CLI credentials, so it appears in the per-agent and chat-builder model pickers for Codex.
• Cerevisor’s own automated checks now run the test suite once per operating system rather than twice on one of them.

9. What this was verified on

• The automated test suite — unit and integration tests, the failure-day tests described above, and the tests that load workflow files saved by versions before 3.0 — passes on Windows.
• A built copy of the app was started on Windows and one small real run inside it was checked automatically: the run finished, what it did was written to this computer’s disk, and it still showed up in run history afterwards. That is one small run against a stand-in model, not a check of the installers, a real provider or a phone.
• The installers (Windows, macOS and Linux) and the Companion app on Android have not yet been checked by hand for this version. 3.0.1 goes out on the automated checks above; the by-hand checks follow, and this note is updated when they are done.
• The Companion has not been checked by hand on an iPhone yet either, and no App Store availability is claimed.
• Grok Build support keeps the statement it shipped with in 2.3.0: it is verified against recorded test data and unit tests only, not yet against a live end-to-end run.
• 3.0.1 adds automated tests of its own for everything above: the four failure days that used to be listed as broken now have passing tests where they had failing ones, and there are new tests for the refused-action card and its approvals, for re-running one agent, for the “ended early” label, for the Rule-of-Two card, for the unreachable-server warning, for the Operator’s midnight rollover and its interrupted spend, and for refusing an agent pinned to a removed provider.

The by-hand checks above are still open at the time this version goes out. This note is updated as each one is done.

Known limits

• Chatting with an agent, consulting a colleague inside an organization, and organization chat still make their tool decisions through the older, separate check rather than the single ten-step one every run uses. Their own protections are unchanged; this is about consistency, not safety, and it is planned next.
• A very long run keeps only its most recent steps in the summary the run window builds, and says so when that happens.
• A run started from inside another run shares the slot of the job that started it, so every agent in that group that runs at the same time rides on that one slot. Two paths take no slot at all: an organization chat that invokes a Power, and a helper agent whose caller holds none.
• Several things are exercised only by the test suite, not yet by hand in the running app: the sequences around the pre-run summary and the “Approve and run” step, the save-a-copy file dialog, the “View run” row and the “Started by the Operator” label, the run window restoring a run whose saved view is gone, the new background summary appearing in a live Progressive or Operator turn, the situation card’s familiar-ground label, and a single-agent re-run visibly waiting for a slot on a busy machine.
• The new “Approve and run again” step is in that same position: it is exercised by the test suite, not yet by hand in the running app.
• The cost estimate for GPT-6 Astra does not include the surcharge OpenAI applies above 272,000 input tokens, so a very long conversation on that model will cost more than the estimate shows.
• The by-hand checklists that came with 2.4.0 — for how many agents run at the same time, how providers that push back are handled, and where tools are allowed to reach — are still unticked.
• There is no single switch that turns off the background summary the Operator works from. Memory consent covers what leaves your machine and what is processed locally; the per-workflow memory setting covers Progressive Mode and your workflows, and the Operator does not have one.

Updating

Existing installations receive 3.0.1 through Cerevisor’s built-in updater. New installers are available from cerevisor.com/download and the private GitHub release.