χchi

Changelog

What shipped in each chi release.

Changelog

The docs on this site track the latest release. Older versions link to their GitHub release notes from the version selector in the header.

v0.2.0

The autonomous director became 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” sets a target score, “don’t spend over $5” sets a cost ceiling. The director halts itself when the goal or cap is met.
  • 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.yaml improves a champion with zero API keys (a scripted fleet). See Getting started.
  • /director and /resume replay — see the rounds a director run completed, read back from the run store.
  • NoiseGuard for local noisy evals — median-of-N verification of an apparent improvement, not just for leaderboard problems.

Fixed

  • The 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 — every scored candidate is archived by hash and the export is hash-checked.
  • Domain-generic brain — the director’s research and steering prompts are templated from your problem, no longer hardcoded to a specific CUDA kernel.
  • A perma-plateau now escalates to a research round; a zero-eval timeout steers the coder to a smaller edit; query_ledger searches the right tables.

v0.1.0

First public release: the autoresearch harness, the conversational operator, the multi-agent fleet, the SQLite blackboard with a negative-results ledger, the deterministic watchdog, two-tier eval with submission gating, and the Textual UI.