Skip to content
bucker

Review mode

Review mode decides how the changes that agents and integrations propose are reviewed before they take effect. It lives in Settings, Review mode, and a project can override it from Settings, Projects. Only an organization owner can change it; every member can read it.

A person's own changes are never held by this setting. It governs what an agent, a service account or an integration does on your behalf.

The two modes

Mode What it does
Automatic (the default) Your approval policy decides, exactly as it did before review mode existed. Set a risk ceiling and the decision model can approve more, within the goals you set.
Human review required Every change an agent or integration proposes waits for a person, including changes your approval policy would have allowed.

An organization with nothing configured is Automatic with no risk ceiling, which lifts nothing: no existing organization changes behaviour until an owner picks Human review or sets a ceiling.

What always stays with a person

In either mode, these are never approved automatically:

  • merging code;
  • anything your approval policies deny;
  • actions the product reserves for a person, such as billing changes and key minting;
  • any change that needs more than one approver, or that your policy marks as always asking.

The goals

In Automatic mode a change that your policy would ask about is approved by the decision model only when a risk ceiling is set, the change's risk is at or below it, and every other goal you set also holds. A goal that cannot be measured counts as not met.

Goal Meaning
Risk ceiling The highest risk class that can be approved automatically: Elevated or Sensitive. Critical is never a ceiling.
Protected environments Changes that reach these are never approved automatically. Production is always protected, even if you remove it from the list.
Change windows Days, a start and end time, and a time zone. When you add any, approval happens only inside one. A window that ends before it starts runs past midnight.
Audience and traffic caps The largest share of users or traffic a change may reach, as a percentage. Blank is no limit.
Error budget floor The least error budget that must remain, as a percentage. Blank is no limit.
Require a passing verification The change needs a named check that passed.
Require a way to undo it The change needs a declared inverse. A deploy that only someone outside Bucker can undo has none, so with this on it still waits for a person.
Context Up to 500 characters on what your team is doing right now, shown to the decision model. It is scrubbed for secrets and is never treated as an instruction.

A project can override the risk ceiling, the three share limits and the two require switches. Protected environments, windows and the deploy switches stay the organization's.

Automatic deploys

A deploy is a critical change, so a risk ceiling never covers it. It is approved automatically only when both of these hold, whatever the ceiling says:

  1. the environment it reaches is switched on under Automatic deploys, and
  2. the deploy has a named verification signal that passed.

Any other goal you have configured still applies.

Production has its own switch, and it needs you present. Turning it on asks for your current password or a security key in the same request; the session you signed in with last week is not proof. The organization records who switched it on and when, and the other owners are notified. Turning it off needs no proof and takes effect immediately. An environment that cannot be identified is treated as production.

The audit trail

  • Every change to the mode, the goals, a project override and the production switch is written to the audit log with what it was before and after.
  • Every change the decision model approves is written as review_mode.auto_allow, with the goals that held and the facts it read. Recent automatic decisions on the Review mode page lists the newest, read from the audit log (owners only; reading the log is itself recorded).