Continue an interrupted run

What happens to a run when Cerevisor closes in the middle of it, and how Continue this run picks the work back up without paying for what already finished.

A run can be cut off part-way through: you close Cerevisor, the machine restarts, the app is killed by the operating system. Before 3.2 such a run was simply left as it was — still marked running, days later, with nothing that would ever move it. From 3.2 Cerevisor closes those runs out on its next start and offers to pick the work back up.

What happens on the next start

When Cerevisor starts, it looks over the runs it recorded in the last seven days (at most fifty of them, newest first) and finds any that begin but never end. Each one is closed with a single line saying the app closed before the run finished — nothing already recorded is rewritten or removed, and if the record cannot be read at all it is left exactly as it is rather than given an ending it does not have.

You see the result in two places:

The run stops claiming to be running. Its Overview carries the note Cerevisor closed before this run finished, so a run the app died in the middle of does not read as a failure the agents caused.
The Recovered interrupted work dialog lists it, with a Continue this run button beside it.

If Cerevisor is busy with a run of its own when it starts (a scheduled run that was already going, for instance), the tidy-up waits for the next start rather than filing the live run as interrupted.

Continue this run

Continue this run starts a new run seeded from the interrupted one. It is an ordinary launch: the pre-run summary opens first, exactly as it does for the Run button, and nothing spends anything until you confirm it.

The summary tells you, in three lists:

Keeps: the agents whose work is carried over. They do not run again and cost nothing. Their output is handed to whatever comes after them, just as it was on the first run.
Runs fresh: the agents that will actually run this time — the one that was interrupted, everything after it, and any agent Cerevisor could not carry over.
Continues its own session: agents running on an outside tool that can pick their own conversation back up (see below).

If an agent's earlier output cannot be read back, Cerevisor says so by name rather than quietly handing on an empty result: Cerevisor could not read back what X produced in the interrupted run, so it will look empty to the agents after it.

Everything on that summary is worked out by Cerevisor from the interrupted run's own record. There is no list to fill in and no options to get wrong — you press Continue, read what it is about to do, and confirm.

Changed while interrupted

Time passed while the app was closed, and files may have moved under the run. Before you confirm, the summary shows Changed while interrupted: the files in the workflow's folder that are different from what the run last saw. The list is capped (the rest are counted rather than named), and the agents that run are told about it too, as information.

Three answers are possible and they are deliberately different:

a list of files — these changed;
nothing — the folder was checked and nothing had changed;
Cerevisor could not check what changed while it was closed — the check itself could not finish. Cerevisor never prints an empty list when it means this.

Agents on outside tools

Whether an agent can resume its own conversation depends on the tool it runs on, and Cerevisor is straightforward about the difference:

Provider	On Continue
Codex CLI	Resumes its own session and carries on from where it was.
Claude Code	Resumes its own session.
Cursor Agent	Resumes its own session — cloud runs only. A workflow switched back to local execution runs the agent fresh.
Antigravity	Runs from scratch. Antigravity keeps no conversation Cerevisor can rejoin.
Grok Build	Runs from scratch. Grok Build has no session to resume.

Those last two are not a failure — the agent simply starts its part again — and the pre-run summary says which agents fall into which list, so you can see it before you confirm rather than discover it mid-run.

If a tool has forgotten the session it was told to resume, Cerevisor runs that agent fresh once instead of failing the run.

One thing a resumed cloud agent keeps

A Cursor cloud agent is created once, with the settings and any secrets it was given at the time, and those live on Cursor's side until that agent is deleted. Cerevisor cannot take them back. So a secret you have revoked from your vault since the interrupted run is still live inside a resumed Cursor cloud agent. This is said on the pre-run summary before you confirm: A cloud agent that continues its earlier session keeps the secrets it was started with until that agent is deleted. If that matters to you, let the agent run fresh instead of continuing its session.

What Continue does not do

It never runs by itself. Continuing an interrupted run always goes through the pre-run summary and always needs your confirmation. Nothing unattended — not the Operator, not a schedule, not a headless run — can decide on its own to pick up work the app died in the middle of.
It never widens anything. A Continue can only make the run smaller — a kept agent is one that does not run. It reads no permission, grants no tool, opens no path. The run it starts is prepared exactly as a Run-button run is, on the same settings you already approved.
It never continues someone else's run. A session belonging to a different workflow is refused in plain words.

Known limits in 3.2.0

The resumed-session paths for Codex, Cursor and Claude Code are covered by automated tests but have not been exercised against those real tools for this release.
Continuing a run that had already finished is accepted and simply does nothing.
A run interrupted more than seven days ago (or beyond the fifty most recent) is never closed out, and its steps can never be checked by hand — see Run and monitor.
The Overview does not yet show a note saying a run was continued from another one.