Retornatus

Governance

Governance in Retornatus is local and explicit: Authority, Boundaries, Rules, Policy, and Human Decisions.

Authority

Actions carry Authority (e.g. DELEGATED for routine work, HUMAN when judgment is required). Policy uses Authority when evaluating effects.

Boundaries

Describe operational scope (filesystem root, isolation hints). Many are advisory — the host realizes worktrees/sandboxes. Retornatus records and projects; it does not replace the host sandbox.

Rules

Learning → Rule Candidate (inactive)
        → Human Decision (APPROVE_RULE_ACTIVATION)
        → Active Rule
retornatus rule propose --statement "Do not commit secrets" --applicability "commit secrets"
retornatus decision record --kind APPROVE_RULE_ACTIVATION \
  --subject R-0001 --summary "Approve" --confirm R-0001
retornatus rule activate R-0001 --decision D-0001

Setting authority=HUMAN on a Rule object is not enough — activation requires a Decision artifact whose subject and confirmation match the Rule id.

Active Rules are projected into Environment bridges (Cursor / CLAUDE.md / AGENTS.md). Canonical truth remains under .retornatus/.

Policy

retornatus policy check --effect "commit secrets to repo"
retornatus policy check --action C-0001/A-001
retornatus gate policy C-0001/A-001
retornatus run C-0001/A-001 --strict-policy

Verdicts: ALLOW · DENY · REQUIRE_HUMAN.

Deny-style Rules typically start with “Do not” / “Must not” and match via applicability.

Bypass

Skipping a gate without a durable Bypass + authority is rejected. Example: Skill research bypass with --force --reason.

Git blast radius

Structural gates are not enough — see Git governance for tiers:

TierMeaning
0 — LocalCommits on the machine
1 — SharePush / PR — only when the human asks
2 — ExternalMerge, deploy, PyPI publish — owner-only in this project

Cursor bridge: .cursor/rules/git-governance.mdc.

Design stance

> Learning informs; Rules constrain. Recurrence never creates authority automatically.

Details: PRD §§31–35, §48.

← Docs hub Product site →