Cerevisor 2.1.1: workflows checked before they spend, approval modes for Codex and Claude Code, and Operator goals you can manage

Two themes. Runs got harder to waste: Cerevisor now checks that a workflow can do what it asks for before it starts spending, and Codex and Claude Code let you decide up front who approves what. And Operator’s goals became something you can manage (edit, park, complete, delete, restore), belonging to the world you made them in, with a stop in chat that actually stops the work.

This note covers 2.1.0 and 2.1.1 together. 2.1.0 went out earlier the same day; 2.1.1 followed within the hour with the keyboard fixes described under Also fixed. If you are updating from 2.0.x, everything below is new to you.

---

Cerevisor checks a workflow before it spends anything

A workflow could ask for something its provider was never going to allow, such as writing outside the working folder or starting a program, and you would only find out partway through a paid run, from an agent that stalled or quietly produced something it had no way to verify.

Every run now starts with a quick check. Cerevisor reads the workflow for what its agents will need, asks the provider about the things it can confirm for free, and either gets out of the way or shows you what will not work. From there you can fix the workflow, or run it anyway with your eyes open.

If you do run anyway, the run window keeps the failed checks on screen for the whole run, rather than treating one click on one dialog as the last word. And if you change the workflow, the folder, or the provider after the check, the check is simply redone, because permission to run is never carried over to a run it was not granted for.

Decide who approves what on Codex and Claude Code

Codex and Claude Code run the agent themselves, which means Cerevisor sets the ground rules once, at the start, instead of approving each step as it happens. Until now those rules came only from the workflow’s permission switches, so if you were running on a plan you already pay for, there was no way to say “I trust this one, stop asking.”

Both now carry their own setting, chosen per provider in the Provider Hub and kept with that provider rather than with the workflow. Claude Code offers the same modes the tool itself has, and Codex offers its sandbox and approval combinations, plus an optional stricter sandbox on Windows.

Existing setups are untouched: a provider you have not given a mode to behaves exactly as it did before. Whatever rules a run ends up under are still written to the System Log the moment it starts, now including which mode produced them and who gets asked.

Answer an agent’s question by picking, not by typing

When an agent asked you which way to go, the options arrived as a paragraph and you had to type one of them back. The question is now its own card: tap an option, or tick several and submit, or write your own answer in Other. Once answered it collapses to a short summary of what you chose.

Single-choice questions commit on the first tap and lock, so a double tap can’t answer twice. Long or numerous options stay readable instead of overflowing, and the card respects your system’s reduce-motion setting.

Tell us something without leaving the app

There was no way to send feedback without going and finding an address. The contact avatar in the app now opens a small composer, and Cerevisor hands what you wrote to your own mail program as a prefilled draft.

Cerevisor sends nothing itself. The draft opens in whatever mail client your machine already uses, you decide whether to send it, and the message is never logged or stored on the way there.

---

Operator’s goals are now something you can manage, not just watch

Before this release, an Operator goal was mostly a name on a card. There was no way to correct it once adopted, no way to say “this is actually done,” and Park quietly kept the underlying work running behind the scenes regardless of what the button said.

Every goal card now offers Edit, Park/Resume, Complete, and Delete. Editing changes the title or brief without losing the goal’s history or spend. Park genuinely pauses the goal, and Operator stops advancing it until you resume. Complete marks it done and takes it off the active list. Delete asks you to confirm, states plainly that it never touches your workflows, files, or past runs, and leaves a Deleted recently view where you can restore it if you change your mind, for goals that already belong to a world. A goal carried over from before this update, one that hasn’t been assigned to a world yet, stays in its own review area and doesn’t get a Restore option if you delete it (see below).

Goals belong to the world you made them in

A goal adopted in one world could previously reappear, or quietly keep working, in a different world entirely, because nothing recorded which world it came from. Goals now carry that information: what you’re working on in one world stays there, and switching worlds shows you that world’s goals, not another one’s. Goals from before this update appear in their own review area and are never acted on automatically, so you decide what to do with them.

Telling Operator to stop in chat now actually stops it

Saying “stop this” or “do this instead” in chat used to have no effect on what Operator was actually doing, and the goal kept running underneath the conversation. A clear stop-or-replace message now pauses work on that one goal until you decide whether to keep it, replace it, or let it go, and the panel tells you plainly, in the moment, what’s happening and why.

---

Also fixed

• Space works on buttons again. Holding Space on the canvas switches to pan mode, and that shortcut was claiming every Space press in the whole app, so moving to a button with the keyboard and pressing Space did nothing. It had been written to stand aside for text fields, but not for buttons. It now stands aside for anything Space is meant to activate: buttons, links, checkboxes, radios, switches, tabs and menu items. Hold-to-pan on the canvas is unchanged.
• Closing the message composer puts you back where you were, instead of dropping your keyboard position to the top of the app.
• Settings no longer go missing when you switch workflows. Moving between workflows in a world rebuilt each one’s settings by hand, and anything that step forgot was dropped, so a setting could quietly reset just by switching away and back. Settings now carry across intact.
• Progressive stays put. A restored Progressive session no longer disappears from a workflow whose saved toggle was missing, and a session only ever shows on the workflow it belongs to.
• Claude Code runs in packaged builds again, after an update to the underlying tool moved its program into a separate per-platform package.

---

The short version

Feature	In one line
Pre-run checks	Every run starts with a quick check that the workflow can do what it asks; failed checks stay on screen for the whole run, and changing the workflow, folder or provider redoes the check.
Approval modes for Codex and Claude Code	Choose per provider, in the Provider Hub, who approves what, kept with the provider rather than the workflow; untouched setups behave exactly as before.
Manageable Operator goals	Edit, park, complete or delete any goal; Park genuinely pauses, and Delete confirms and can be undone from Deleted recently.
Goals stay in their world	A goal belongs to the world you made it in; switching worlds shows that world’s goals, not another one’s.
Stop in chat that stops	Telling Operator to stop or change course in chat now pauses the goal until you decide what to do with it.
Pick-an-option questions	An agent’s question arrives as a card where you tap an answer, tick several, or write your own, instead of a paragraph to type back.
Feedback without leaving	The contact avatar opens a composer that hands your message to your own mail program as a draft; Cerevisor sends and stores nothing.
Keyboard fixes	Space activates buttons again app-wide, and closing the composer returns your keyboard position instead of dropping it to the top.

---

Upgrading

Cerevisor updates itself. If you would rather not wait, Settings → Updates → Check for updates.

Nothing in this release changes your data, your settings, or anything Operator has already built. Goals you are already tracking carry forward automatically the first time you open the app after updating.