Contributing to CAIO

Thanks for your interest. CAIO is a reference prototype synthesizing ideas validated separately across several sibling projects (see the README's "Relationship to other projects" section) — contributions that strengthen the actual security properties (lineage integrity, capability conservation, divergence detection, audit tamper-evidence) are especially valuable.

Getting set up

git clone https://github.com/kaymyg/caio.git
cd caio
pip install -r requirements-dev.txt
pip install -e .

python -m caio.demo   # see it work end to end
pytest -q --cov=caio --cov-report=term-missing   # 10 tests, 100% on core.py

Before opening a PR

CI runs the test suite with coverage and the demo script, across Python 3.10–3.12, on every push and PR. Please run that locally first.

If you add a new branch or status code to Gateway.transact(), add a test that exercises it through gw.transact() itself, not just the underlying helper it calls — test_transact_level_status_mapping exists specifically because the original test suite covered EvidenceGraph.evaluate() directly but never asserted what transact() actually mapped that result to.

What especially matters here

This is a reference prototype, not a deployable network protocol. The agent key-binding check in Gateway.transact() is a local sign/verify-with-a-fresh-nonce consistency check, not a real challenge issued over an untrusted wire — be precise about this distinction in any docs or code you touch, don't let it get oversold.
Lease.expires_at is checked; Lease.issued_at is not. A lease "valid" from a future timestamp still passes Gateway.valid(). This is a known, documented gap (see README "Known Limitations"), not an oversight to silently patch — if you fix it, make sure you understand why it was left out (clock-skew tolerance) before tightening it.
The host-divergence detector checks two specific action strings (CRITICAL_TAMPER, UNSANCTIONED_EXEC), not general anomaly detection. A generic claim/host mismatch only reaches CAPABILITY_REDUCTION, not QUARANTINE — that's intentional graduated response, not a bug, but don't describe it as catching "any divergence" in docs.
Nonce replay protection burns the nonce only after key-binding succeeds — preserve that ordering if you touch transact(); marking a nonce used before authentication would let an attacker burn a legitimate agent's nonce with a failed forged request.

Good places to start

A real network-level challenge-response protocol to replace the local key-binding consistency check
issued_at validation (with configurable clock-skew tolerance)
Persisting Ledger/Audit/EvidenceGraph state instead of in-memory only
Expanding the host-action taxonomy beyond the current two hardcoded "bad" actions, ideally via a configurable policy rather than more hardcoded strings
A real transport binding (HTTP middleware, gRPC interceptor) around Gateway.transact()

For anything larger, please open an issue first to discuss the approach.

Commit style

Small, focused commits with a clear one-line summary. No strict format enforced beyond that.