Cadence & Permissions
How often the agent checks in with you, and how it asks before doing anything real.
Cadence
Set in config.toml, adjustable live with /cadence. Tightest to most independent:
tight
Check in after every single step - a file worth noting, a proposal, a command about to run, a result just in. Short bursts, pause for a reaction before the next thing.
checked-in
Check in at each natural milestone - before running something non-trivial, right after a result lands. A couple of related actions can happen back-to-back, but nothing longer without surfacing where things stand.
loose
Work through a small cluster of related steps (propose, run, log the result) before summarizing - check in once that whole arc is done, unless something in the middle needs a decision only you can make.
autonomous
Work through a full hypothesis or task with minimal check-ins - propose, run, log, and only surface on a real result, a blocker, or a decision only you can make. Permission asks are the checkpoints.
Permissions
Every real action - running a command, writing a file, launching an experiment - is proposed before it runs. When it asks, you choose how long that answer holds:
| Reply | Scope |
|---|
| once | Just this one call. |
| session | Anything matching, for the rest of this session. |
| project | Saved to this project - every future session here. |
| always | Saved globally - every project. |
| reject | No, and don't ask again this turn. |