Roles and authority

One person or agent can perform more than one role, but the authority and evidence for each role remain distinct. A reviewer must not approve its own unreviewed risk decision.

Role	Owns	Must not do by inference
Owner	Scope, priorities, exceptions, and remote authorization	Assume an agent has understood an unstated preference
Planner	Constraints, plan, risks, acceptance criteria, and validation design	Change product code while assigned plan-only work
Implementer	A scoped branch, batches, local commits, and evidence	Push, open a PR, merge, release, deploy, or delete without authorization
Reviewer	Independent assessment of correctness, regressions, and missing evidence	Treat unrun checks as passing or approve its own exception
Release operator	The authorized release or deployment procedure and operator evidence	Treat a successful push as a release or deployment

Modes and authorization

Agent modes defines the permitted local actions and required outcome for Plan, Inspect, Diagnose, Review, Implement, Verify, Deliver, and Monitor. The owner remains the only source of approval for remote or privileged actions.

Words such as “finish”, “urgent”, or “use Yeet” do not waive the authorization boundary. If a request says “use Yeet and push/open a PR”, those two remote actions are authorized; merge, release, and deployment still are not.

Reporting contract

At each meaningful handoff, report:

• the selected flow and completed batch number;
• changed areas and the local commit created for the batch;
• commands run and their result;
• unrun or blocked checks, including why;
• remote actions completed only when they actually succeeded.

Never include credentials, private endpoints, access tokens, or machine-specific paths in a public report or commit.