codex-autoresearch
A Codex skill that autonomously and repeatedly modifies a codebase, measures a target outcome with a command you specify, and keeps or reverts each change until the target is reached.
Codex Autoresearch gives Codex a self-directed experimentation loop for any outcome a command can measure, such as test failures, code coverage, type errors, warnings, latency, binary size, or reproducible security findings. Given a goal and a numeric target, Codex inspects the repository, confirms the experiment, including which paths it may change, the metric command, an optional regression guard, and foreground or background mode, then repeatedly changes one focused thing, commits, measures the result, keeps the change if it improved and the guard passed, or reverts it, logging an audit event each time.
It can run in the foreground under a Codex Goal for watching and steering live, or in the background as a detached controller for long or overnight runs, with one worker process per iteration; a control script owns commits, verification, rollback, and state, while Codex proposes the changes. It requires a clean, dedicated Git branch and installs as a Codex skill.