Retornatus
Keep AI coding agents honest — with goals, memory, and proof.
Install once in your project. Your agent works toward a clear outcome, remembers across chats, and cannot call work “done” without evidence you can see.
Why people use it
- You stay in control. The agent proposes and builds; you keep the bar for “done.”
- Chat stops being the only memory. Intent and progress live in the repo, so the next session (or teammate) can continue.
- Fewer false finishes. Work closes when there is proof — not when the model sounds confident.
- Right amount of ceremony. Tiny fixes stay light; risky work earns more structure.
- Skills when you need them. Specialize for the problem at hand with current research — not a stale mega-pack.
How you use it
- Install the CLI and initialize it in your project.
- Open the project in Cursor, Claude Code, Codex, or your preferred agent.
- Describe what you want — Retornatus turns that into a clear change with an agreed finish line.
- Let the agent work under that agreement; check status anytime.
- Close when evidence matches the goal — then the next return starts wiser.
Your coding agent still writes the code. Retornatus governs the loop and keeps the record. Deep mechanics live in the docs.
The idea in one breath
Govern the work. Bound the agent. Verify the outcome. Demand what you need, agree on what “done” means, execute, then verify with proof that stays in git — so every return to the project is informed, not amnesiac.
Without
- Jumps to code and declares victory
- Each chat starts from zero
- You rediscover the same mistakes
With Retornatus
- Clear finish line before the build sprint
- Project memory survives the tab
- Proof required before “done”
Install
uv tool install retornatus
cd /path/to/your-app
retornatus init
retornatus integrate
retornatus doctor
Needs Python 3.11+ · uv recommended · also on PyPI. Then open the project in your AI coding agent and start a change.
What’s new
- One-screen progress — see goals, proof, and what’s next without digging through chat.
- Health at a glance — quick checks that separate “process needs care” from hard stops.
- Lessons that stick — a failed check can become guidance for the next return.
- Seedcore — brand mascot for the site and README (original illustration).
Release notes and technical depth: documentation · releases.