Profiles

A profile is a per-repo declaration of how strict the merge gate is. It lives in registry.yaml and gates which verbs can act without explicit human approval.

The three profiles

Profile	Use for	Effect
`strict`	Production / control-plane / shared infrastructure	All write verbs require human merge. `request_merge_when_green` is rejected unconditionally.
`standard`	Most production repos	Human merge by default. `request_merge_when_green` permitted only for fleets explicitly listed in `fleets_allowed`.
`permissive`	Sandboxes, scratch repos, dogfood instances	Auto-merge available to any allowed fleet (or human). Use sparingly.

What `request_merge_when_green` actually does

It opens a PR (same surface as request_review) and sets GitHub’s auto-merge label. Branch protection + required CI checks still gate the merge — auto-merge fires after CI goes green, not on PR open. So even on a permissive-profile repo, a red CI run blocks the merge.

The profile gate is therefore about whether the verb is permitted at all, not about whether merges happen instantly.

Choosing a profile when registering a repo

A simple decision tree:

Does this repo have CODEOWNERS or required review approvers on its default branch? → strict.
Does this repo have CI but no required reviewers? → standard.
Is this repo a sandbox you’d be fine with bots merging into freely? → permissive.

When in doubt, strict is the safe default. Profile is editable later via a request_review PR against registry.yaml; tightening is cheap, loosening should be deliberate.

What strict-profile rejection looks like

$ petrova request_merge_when_green petrova-hq --input ...
request_merge_when_green → failed
  ✗ PROFILE_PERMITS_AUTOMERGE  repo 'petrova-hq' has profile 'strict';
    auto-merge is forbidden — use request_review instead

The recovery is mechanical: switch to request_review (drop merge_method from the params) and request a human approver.

petrova-hq is strict

The control plane itself is strict-profile and fleets_allowed: []. Changes to PETROVA-HQ go through human PR review only — no agent automation. This is intentional: the control plane is the substrate that audits everything else, and an agent that could rewrite the substrate would invalidate every audit it then performed.