Repo-oriented Codex port of Andrej Karpathy's autoresearch with real git-backed improvement loops

Automate Codex CLI long tasks by starting one session, resuming it in a loop, and running until a strict completion protocol is met

A Human & Machine readable audit trail. Engineers need it because it prevents the two common failure modes of agent driven development: Agents make changes but nobody can quickly tell what changed, why, and whether it passed checks Work cannot be reliably resumed, reviewed, or automated because state and logs are scattered

Automate goal-driven Codex experiments to test, iterate, and find better results fast