Retornatus

Overview

Retornatus is a repo-native governance harness for AI-assisted software development.

It keeps agents inside a repeatable contract: understand the demand, activate obligations, execute under authority, and close with attributable proof — all durable under .retornatus/, not trapped in chat.

It is not:

Your coding environment executes. Retornatus governs and records.


The problem it solves

Failure modeWhat Retornatus does
Agent jumps to codeSituation elicitation + Contract before build
“Done” is a chat claimAssurance from Claim-bound Evidence
Session amnesiaFiles + wake continuity
One-size ceremonyTasks and Skills only when complexity earns them
Silent rule inventionRule Candidates need a Human Decision

The core loop

Demand
  → Situation
  → Contract          (WHAT + constraints + DONE)
  → Action            (+ Tasks when needed)
  → Skill?            (research current sources)
  → Execution         (host agent)
  → Evidence
  → Assurance         (SATISFIED | NOT_SATISFIED | INCONCLUSIVE)
  → Learning / Skill evolution

If work is blocked by discovery:

Finding → Question → Action → Evidence → Resolution

Status commands (status, loop next) are projections. If they disagree with files, the files win.


How much ceremony?

WorkTypical path
Trivial fixShort Contract, skip Skill (skill need), light Evidence
Normal featureFull Change loop + gates
High-risk / specializedSkill research gate + optional independent Assurance + Policy

Complexity must be earned — see Concepts and the PRD philosophy (§2).


How you interact

  1. Chat — ask the agent to follow the Retornatus hub skill (installed by integrate).
  2. CLI — intentions such as change create, gate contract, verify (source of truth for gates).
  3. Files — inspect .retornatus/changes/… when you need the durable record.

You approve product intent and Human Decisions (Rules). The agent implements. Git push / merge stay yours.


Next steps

← Docs hub Product site →