aria-agent
Canonical clean-break agent harness: approval gates, execution tracing, and progressive-disclosure Agent Skills.
ARIA is an offline-first Python harness with bounded multi-hop execution, deterministic safety checks, retries, rate limits, ordered SSE events, dry-run replay, local memory search, and a checksummed portfolio evidence fixture. Hosted/team workflows, external notifications, OTLP protobuf/gRPC, mandatory infrastructure, and provider credentials remain deferred.
Why it matters
ARIA turns approval, safety, trace, cost, replay, and reproducibility boundaries into inspectable local code instead of relying on a hosted control plane.
Verified proof
200 Python tests, 12 focused Agent Skills tests, Ruff and Pyright checks, 31 frontend tests, checksum-verified offline evidence, and six Chromium smoke tests passed on 2026-08-15.