Retornatus

Keep AI coding agents honest — with goals, memory, and proof.
Govern the work. Bound the agent. Verify the outcome.

Website → · Docs · Quick start · PyPI

Repo-native governance harness for AI coding agents (Cursor, Claude Code, Codex, and similar).

Agents are fast — and optimistic. They ship code, summarize what they think they did, and move on. Retornatus installs a repeatable contract into your repo: agree on the finish line in writing, execute under that agreement, and close only when evidence supports the goal. Intent, progress, and lessons live under .retornatus/ in git, not in a chat scrollback.

Your coding agent still writes the code. Retornatus governs the loop and keeps the record.

Package 1.4.0 is on PyPI. Maintainer release steps: CONTRIBUTING.

What it is · Install · Verify · First change · Checklist · How it works · What you get · Commands · Cloud agents · Docs · Credits

What it is

A governance harness for AI-assisted software work — not an IDE, not an LLM runtime, and not an agent marketplace. The host agent still writes code; Retornatus structures obligations, proof, and memory under .retornatus/.

After install, work moves through a durable loop you can inspect in files:

Demand → Situation (requirements) → Contract (WHAT + DONE) → Action (+ Tasks when needed) → Evidence → Assurance → Learning

Along that loop the harness also:

• Sizes ceremony to risk (QUICK / STANDARD / COMPLEX)
• Surfaces focused questions when the ask is fuzzy (change elicit)
• Scores readiness (Process vs Brakes) via doctor
• Stages freeform chat into Skill proposals only with human confirmation (intake analyze)
• Loads at most one specialization Skill when needed (skill need / skill create)

You approve product intent and consequential Rules. The agent implements. Push, merge, and publish stay on your terms (git governance).

1. Install

You need Python 3.11+. We recommend uv (installs Python CLIs quickly). If you do not have uv yet:

# Windows (PowerShell)
powershell -ExecutionPolicy ByPass -c "irm https://astral.sh/uv/install.ps1 | iex"

# macOS / Linux
curl -LsSf https://astral.sh/uv/install.sh | sh

Then run these in your application repository (the project the agent should change — not necessarily this harness repo):

# Install the Retornatus CLI once on your machine
uv tool install retornatus

cd /path/to/your-app

# Create .retornatus/ (config + durable memory layout)
retornatus init

# Install the hub skill into your agent environment (e.g. .cursor/skills/)
retornatus integrate

# Readiness check — Process vs Brakes scores
retornatus doctor

One-shot without a global install: uvx retornatus --help. Package page: PyPI.

Day to day you work in agent chat; the agent (following the hub skill) calls the CLI when a gate or record is needed.

Go deeper: Quick start · Environments · Cloud agents

2. Verify readiness

Re-run doctor after upgrades or machine changes. You are ready when:

• .retornatus/ exists and init has been run in this project
• The hub skill is visible to your agent (after integrate)
• doctor is not hard-stopping on “not initialized”
• Your AI coding agent can open the project and follow the hub skill

For brownfield repos, also consider retornatus project-init and retornatus wake --bridges so existing context is captured before the first Change.

3. Run your first change

Open your AI coding agent in the project and ask for a concrete goal, for example:

Create a Retornatus Change for GET /health returning 200. Contract and gates before code.

Ask it to follow the installed Retornatus hub skill. Prefer chat for product intent; use the CLI when you want mechanical gates yourself.

If the agent jumps straight to code: Stop. Finish Situation (requirements) + Contract (and pass gate contract) before the build sprint.

Go deeper: Quick start · Tutorials · How it works

Getting started checklist

• Python 3.11+ available (python --version)
• Ran uv tool install retornatus (or use uvx)
• In your app repo: retornatus init → integrate → doctor
• Opened the project in your AI coding agent and confirmed the hub skill is visible
• Asked for a written Change / Contract before implementation
• Know where docs live: Website · Docs hub

Stuck? Quick start · FAQ

How it works

A focused software-construction cycle — durable files under .retornatus/, gates as brakes:

Understand  →  Agree     →  Build              →  Prove           →  Learn
Situation      Contract     Action (+ Tasks)      Evidence            Learning
(requirements) (WHAT+DONE)  (agent writes code)   + verify            (+ Rules?)

When the ask is fuzzy, change elicit exits 1 and lists focused questions (with options). The agent should ask them in chat; you answer; record with --answer TOPIC=…. Clear asks can skip straight to a Contract.

Optional: change classify picks QUICK / STANDARD / COMPLEX so ceremony matches risk.

For a freeform chat prompt that might need specialization, prefer intake analyze: it stages the ask against .retornatus/, proposes a Skill only when you answer CREATE=yes, and never auto-creates from chat alone. Early signal without creating: skill need --prompt "…".

Status / overview are projections. If they disagree with files, the files win.

What you get — and why it helps

Requirements analysis (Situation)

Without it: “Add login” becomes three different products in three chats.

With Situation: The harness surfaces material questions (actors, scope, out of scope, how you’ll know it worked), reads kickoff files when present, and refuses to pretend the Contract is ready until those answers exist.

Memory — the repo remembers

Changes, Contracts, Evidence, and Learnings live under .retornatus/ in git. wake rebuilds continuity. Chat is a window; git is the source of truth.

Proof before “done”

A Contract states WHAT and DONE. Evidence binds to those claims. verify returns SATISFIED / NOT_SATISFIED / INCONCLUSIVE. Gates return non-zero = STOP.

Test, security-test, build, and lint results only count when Retornatus ran the command (evidence run … -- <command>). The record stores argv, exit code, timing, a SHA-256 of combined stdout/stderr, and the git HEAD (when the directory is a repository). evidence add is self-reported: for those types verify marks the evidence UNVERIFIED and does not exit 0. Narrative notes (review_result, repository_observation) can still be recorded with evidence add and are labeled self-reported. While migrating, verify --allow-self-reported or [assurance] allow_self_reported = true in .retornatus/config.toml restores the old acceptance. A recorded commit that no longer matches HEAD is a warning, not a failure — unless the owner named the commands that count.

Required checks. Put the real commands in [assurance] required_checks (name + run argv, optional types). verify then accepts execution evidence only when the argv matches one of those checks exactly, the exit code is 0, the recorded commit is the current HEAD, and the worktree has no uncommitted source changes. evidence run -- true does not pass. verify --run-checks and checks run execute the declared commands through the same capture path. allow_self_reported does not bypass a configured check.

Freshness. Uncommitted or untracked edits to a claim subject path make that evidence stale. Execution types fail by default; narrative types warn. Set [assurance] uncommitted_changes to "fail" or "warn" to override.

Diff gates and hooks. gate suppressions stops newly added skip and ignore markers. gate scope compares the git diff with the change’s Task resources (denied paths always fail; sensitive paths need a satisfied review_result or security_test claim). hooks install wires both into pre-commit, and a commit-msg hook that reads the message from $1.

Ceremony matches risk

QUICK for a typo; STANDARD for a normal feature; COMPLEX when security, payments, or high novelty need more depth.

Hub + Skills (human-controlled)

The hub skill is the map every turn. At most one specialization Skill while executing — research current sources when needed, not a mega-pack every message.

Two Skill worlds:

skill need --prompt can flag need early (no Action required). Agents must not invent Skills from a vague prompt without your confirmation.

Go deeper: How it works · Gates · Memory · Skills

Commands cheat sheet

Full map: CLI · hub skill after integrate.

Verify receipts

verify --receipt and receipt sign write an Ed25519 receipt under .retornatus/assurance/receipts/. The signature covers the verify verdict. receipt verify checks it.

What a receipt proves. Someone who held the private key signed that result. It does not prove the agent was sandboxed, and it is not a receipt network.

Where the keys live.

receipt keygen writes the public key into the repo and the private key under the user config dir ($XDG_CONFIG_HOME/retornatus or %APPDATA%\retornatus). receipt keygen --print prints the private key instead, for a CI secret.

Threat model. An agent that can read the private key can produce a valid signature. Keep RETORNATUS_SIGNING_KEY and the config-dir key out of the agent's environment. Verification needs only the committed public key, so a fresh clone can check a receipt without the secret.

CI. Store the private key as a GitHub Actions secret and sign in CI, where the agent that edits the repo does not see it:

- name: Sign verify receipt
  env:
    RETORNATUS_SIGNING_KEY: ${{ secrets.RETORNATUS_SIGNING_KEY }}
  run: retornatus verify C-0001 --receipt

Anyone who clones the repo (the public key is already in .retornatus/keys/) can run retornatus receipt verify path/to/receipt.json.

Legacy HMAC. Older HMAC-SHA256 receipts still verify only on a machine that has the old local key. receipt verify reports them as legacy_hmac with portable: false and prints a deprecation warning. That key is not written into the repo.

CLI mistakes print error: … instead of a traceback. An invalid id such as verify ../../../tmp exits 2. Search text with quotes or hyphens is matched as literal words (it does not crash). Gates and verify still use 0 for pass / SATISFIED and 1 for STOP / not SATISFIED.

What’s new (1.4.0)

• Required checks — when [assurance] required_checks is set, verify accepts execution evidence only for those commands (verify --run-checks, checks run)
• Cloud and remote agents — install the CLI on the clean VM, run hooks install on every fresh clone, and leave receipts unsigned
• PR gates — the consumer workflow pins retornatus==1.4.0 and runs blocking verify, gate suppressions, and gate scope
• Trusted Publishing — tag v* on main publishes after tests, a distribution contents check, and a matching changelog section
• Supply chain — GitHub Actions are pinned to commit SHAs; SECURITY.md and CodeQL cover private reports and Python analysis

What’s new (1.3.0)

• Verify receipts — Ed25519 signature checked with the committed public key; private key stays outside the repo (verify --receipt, receipt verify). Legacy HMAC receipts are legacy_hmac and not portable
• Action attempt budget — action budget --max N + gate budget stop runaway retries
• AGENTS.md map for host agents + CI check that docs HTML stays in sync with markdown
• From Spec Guardrails migration page + Landscape comparison with adjacent harnesses

What’s new (1.2.1)

• Prompt intake — intake analyze stages a freeform request against .retornatus/, proposes Skill only with human CREATE=yes
• Two Skill worlds — analyzed intake (controlled) or manual skill create
• Early skill need — skill need --prompt without requiring an Action

What’s new (1.2.0)

• Situation as requirements analysis — focused questions with options, --answer / --write, kickoff discovery
• Focused software cycle in hub/README: Understand → Agree → Build → Prove → Learn

What’s new (1.1.x)

• Change overview — claims, evidence, tasks, and next work in one view
• Doctor scores — process health vs hard brakes
• Lessons from gates — failed checks can guide the next return
• Ops loops — optional hygiene scans
• Public site — product landing + full HTML docs on GitHub Pages

Cloud and remote agents

A cloud or remote agent (Cursor cloud agent, Codex, Claude Code on a VM, a CI sandbox) starts clean. Install the CLI on that machine. Git does not version hooks, so every fresh clone runs retornatus hooks install. Leave receipts unsigned: the private signing key stays off the agent VM. The enforcement is the GitHub pull-request workflow (verify, gate suppressions --base, gate scope --base), wherever the agent ran.

uv tool install --force git+https://github.com/luizssantiago92/retornatus.git
export PATH="$HOME/.local/bin:$PATH"
retornatus hooks install
retornatus hooks status
retornatus doctor

templates/ci/retornatus-pr.yml pins uv tool install "retornatus==1.4.0". That release includes hooks install and the diff gates. Full notes: Cloud agents.

Documentation

Markdown sources for editors: docs/guide/. GitHub Pages generates the HTML from those files. To preview locally, run python scripts/build_docs_html.py (the output is gitignored).

Credits

Ideas are credited by influence, not by superficial similarity. Retornatus does not claim novelty for established software-engineering patterns; its contribution is how those guarantees are separated, constrained, and composed.

Direct predecessor — Spec Guardrails

Spec Guardrails (MIT) is the direct predecessor of Retornatus.

Retornatus is a separate successor architecture informed by building and dogfooding Spec Guardrails. It is not a fork, rename, or line-by-line rewrite.

Original work in Retornatus: Python domain model and CLI, .retornatus/ layout, Demand / Situation / Contract / Action (plus Finding / Question / Resolution), Evidence separated from Assurance, complexity lanes (QUICK / STANDARD / COMPLEX), on-demand specialization Skills with research gates, human-controlled prompt intake (intake analyze), doctor Process vs Brakes, overview / ops / lessons loops, and the public docs site.

Transitive lineage: Spec Guardrails itself credits upstream open-source work (spec-driven phases, task graphs, loop engineering, harness vocabulary, and related tools). Those influences arrive through Spec Guardrails unless Retornatus independently revisited them — see the full provenance write-up.

Full credits & lineage: credits on the website · credits-and-lineage.md · Spec Guardrails’ own credits

License

MIT — see LICENSE.