Second release — the autonomous director is now genuinely usable, plus safety and onboarding hardening. Driven by a multi-agent usability audit against real user journeys.
New
- Director self-stop — give it a goal and walk away: "get it under 500µs then stop" (target score) or "don't spend over $5" (cost ceiling); it halts itself.
- Sandboxed eval — run an untrusted candidate's correctness+benchmark inside a jail (
eval_sandbox: docker), so a hostile candidate can't reach the host. - Offline no-key demo —
chi run examples/offline.yamlimproves a champion with zero API keys (scripted fleet). - /director and /resume replay — see the rounds a director run completed, from the store.
- NoiseGuard for local noisy evals — median-of-N verification beyond leaderboards.
Fixed
- Director is stoppable —
/stop(and a bare "stop") now halt a running director; quitting warns instead of silently killing it. - Export ships the verified champion, not whatever the coder last wrote (data-integrity fix; every scored candidate is archived by hash).
- Domain-generic brain — the director's research/steering prompts are templated from your problem, not hardcoded to a CUDA kernel.
- A perma-plateau now escalates to research; zero-eval timeouts steer the coder to a smaller edit;
query_ledgersearches the right tables.
Install: uv tool install getchi then chi. Docs: https://getchi.dev