Cerevisor 3.0.0: What's New

Cerevisor 3.0 is about seeing what a run will do before it starts, and being able to prove what it did afterwards. Every run now shows you a plain summary first — which agent runs on which model, what it should cost, what it can reach — and asks again whenever a workflow has gained new powers since you last approved it. While a run works, the window lists the files it creates, changes and reads. When it is over, its record survives: the run window still opens tomorrow, run history keeps runs the old limit used to drop, and you can save a redacted copy of any run to a file you choose. Runs Cerevisor starts by itself — a schedule, a loop, a pipeline, a Power, the Operator — now check their folder and their models exactly as a button press does, and say plainly why they cannot start. Nothing leaves your machine; all of this is written to your own disk.

1. A summary before a run starts

• When you press Run Workflow, send from the workflow’s input box, or re-run a single agent, Cerevisor now shows what is about to happen first: which agent runs on which model, which connected servers it can reach, what it should cost, and anything it already knows will limit the run.
• Cancel or Run. Nothing is started, spent or reserved while you read it.
• A launch that stopped to ask you about a workflow’s capabilities shows the summary too, once you have answered.
• For an agent that runs on Antigravity, the summary now lists only the permissions you actually switched off for the workflow as ones Antigravity will not honour. It used to list all four even when every switch was on.

2. Cerevisor asks again when a workflow gains new powers

• If a workflow can do more than when you last approved it — new tools, new connected servers, broader file or command access — the summary grows an “Approve and run” step that lists what changed, in plain sentences.
• An automatic run in the same situation — scheduled, started by the Operator, or launched from your phone — is skipped instead, with a plain notice, and waits for you to look at it yourself.
• You can turn this off in Settings → Permissions → “Run history & approvals”.

3. “Changes during this run”

• The run window now lists the files a run created, changed or read, as it happens.
• It calls out anything changed that the workflow never declared.
• And it says, in one sentence per agent, what Cerevisor could actually hold that agent to — which differs by provider, and is now stated rather than assumed.

4. Runs you can still open tomorrow

• A run’s status, cost, files and timeline are read from the run’s own saved record, so a run window opens correctly long after the run finished — including after a restart, and including runs whose live view was never saved.
• Run history keeps runs that the old history limit used to drop. Older runs saved before this release still list exactly as they did.
• A run now says it started once, at the time it really started. It used to be written down as starting twice, the second time with the clock reading the moment it finished. Runs saved by earlier versions keep what they already had.
• A run that was stopped or that failed is never shown as completed — not in the run window, not in history, and not on the phone. That had been possible before, on both the desktop and the Companion.

5. Save a copy of a run

• A run’s record can be saved to a JSON file you choose. Secrets are removed before it is written, and it goes only where you point it — nothing is sent anywhere.
• It does contain file paths from this computer, so treat it like any other local export before sharing it.

6. The Operator and its runs are linked both ways

• An Operator activity row that started a run offers “View run” and opens that exact run.
• A run the Operator started is labelled “Started by the Operator” in the run window.
• The Operator’s own history is now written from what actually happened during each action, rather than kept as a separate note alongside it. Entries written by earlier versions are still shown, unchanged.

7. “What this run did” on your phone

• The Companion app can now show a run’s status, why it ran, what it cost, the files it touched, and a plain sentence per agent.
• If Cerevisor cannot tell which recorded run it is looking at, the phone says so rather than showing another run’s details as this run’s.
• This needs the updated Companion app. An older phone simply does not show the section, and nothing else about it changes.

8. Runs Cerevisor starts on its own prepare the same way you do

• A loop’s next round, a pipeline’s next stage, a Progressive increment, a Power a workflow reached for, a scheduled run and a run the Operator starts now all check that the workflow’s folder is still there and that every agent’s model is still available before they begin, exactly as when you press Run. If something is missing they say so plainly, in the same words, instead of quietly starting on a different model or in a folder that no longer exists.
• Those runs now also follow the approval setting you chose for the runs you start yourself, instead of always using the automatic one — so if you work in Oversight you will now be asked to approve steps inside a loop round, a pipeline stage or a Power that used to run without asking. A setting saved on the workflow itself still wins.
• Long runs Cerevisor starts on its own now trim their working memory at the same point your own runs do, using the setting you picked.
• One exception: a run the Operator starts while you are away has nobody to answer a question, so it runs hands-free and nothing waits for your approval. If you want it to work differently, give that workflow its own approval setting — the Operator’s run follows it. Which of the two applied is spelled out in one plain sentence in the copy you save from the run (“Save a copy of this run”).

9. When a run’s history cannot be saved

• If Cerevisor cannot write down what a run is doing, the run stops before its next action instead of carrying on unrecorded. The run window explains that it could not save the run’s history, so it stopped before doing anything else. It appears in history as “Stopped — history could not be saved”, and whatever it had already spent is kept.
• You can choose the other behaviour in Settings → Run history & approvals, with the setting “Keep my workflows running even if Cerevisor cannot record their history”. Runs then continue and show a notice instead.
• For the four providers that run their own agent loop — Codex, Cursor, Claude Agent SDK and Antigravity — the stop lands one agent later: that provider finishes the agent it is already running, and the next agent never starts. Cerevisor says so in that provider’s own disclosure sentence rather than pretending the two are the same.

10. Better background for Progressive Mode and the Operator

• Progressive Mode now plans with what Cerevisor already knows about you and your work — a short summary of what you are working toward, how you like work shaped, where you have told us we were wrong, and where the run stands. It follows that workflow’s memory setting: turn memory off for the workflow and nothing is added.
• The Operator now works from the same short summary before it picks its next move. It is background for better choices, not a new permission — the objectives, budget and limits you set still decide what it may do.
• The situation card can finally say you are on familiar ground. It never could before: the check for how much of your profile you had filled in was never switched on. Fill in a few sections and Cerevisor will say so.

11. Re-running one agent waits its turn

• When you re-run a single agent, Cerevisor now waits for a free slot instead of overloading your machine or your provider.
• The shell tool now says which interpreter really runs it, so a Windows agent is told it is talking to the Windows command prompt rather than guessing.
• Since 2.4.0, a web request that would change something on a site — posting, updating or deleting — asks you to confirm first in Auto mode; an unattended scheduled run cannot answer that question, so it cannot make that kind of request.

12. We tested the bad days

• Cerevisor was deliberately put through the failures that matter — the disk filling up, the app being killed mid-run, a provider stalling or rate-limiting, a connected server dying mid-call, a half-written history file, an approval nobody ever answers, a run you stop, an outside tool claiming success for a file it never wrote — and each one now has a written expectation for what you should see and what the record should say afterwards. Fifteen of the nineteen situations behave as they should; the four that do not are listed under Known limits below, rather than quietly left out.

13. One system, not two

Cerevisor used to keep two accounts of every run: the older files each screen read from, and a newer recording kept beside them for comparison. There is now one. Every run is recorded once, and that recording is what the run window, run history, the Operator’s history and your phone all answer from — which is why a run can still be opened, exported and trusted long after it finished. Files written by earlier versions still open exactly as before, and nothing about your saved workflows, worlds or transcripts changed shape. A large amount of the app’s internals were reorganised along the way; nothing you do should look or behave any differently.

14. 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. 3.0.0 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.

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

• Four situations from the failure testing behave worse than they should, and are not fixed in this release: a model call that never answers at all leaves that agent waiting indefinitely on Cerevisor’s own loop (calls that time out are handled); a tool result that was shortened to fit is not marked as shortened in the run’s history; a connected server that cannot be reached at all is not announced to you before or during the run; and money spent by an Operator action that was interrupted by a crash or a quit does not reach that day’s total.
• 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.
• Re-running a single agent does not get the automatic second try that agents inside a full run get when their output does not match what they declared.
• 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 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 3.0.0 through Cerevisor’s built-in updater. New installers are available from cerevisor.com/download and the private GitHub release.