codex-autoresearch

Overview of codex-autoresearch

What it is

An autonomous, iterative experimentation system for Codex that repeatedly modifies code, verifies the result, and retains improvements or reverts failures until a user-specified measurable target is reached. Inspired by Karpathy's autoresearch concept.

Key features

• Cycles through modify, verify, retain or discard, and repeat indefinitely
• Works with any measurable outcome a command can produce (test failures, coverage, latency, binary size, etc.)
• Runs in foreground (current Codex task with official Goal) or background (detached controller for long or overnight runs)
• Every trial is a Git commit; non-improving changes are reverted with git revert
• Full audit trail stored in uncommitted autoresearch-results/ directory (events.jsonl, logs, etc.)
• Safety model: out-of-scope edits, command failures, timeouts, and HEAD drift stop the run with an exact error

Use cases

• Automatically reducing error counts or test failures in a repository
• Improving code coverage, eliminating type errors or warnings
• Optimizing latency, binary size, or reproducible security findings
• Any scenario where the desired outcome can be measured by a single numeric metric

Getting started

text$skill-installer install https://github.com/leo-lilinxiao/codex-autoresearchOpen a clean Git repository with Full Access:bashcodex --dangerously-bypass-approvals-and-sandboxThen invoke the skill with a prompt like:textYou: $codex-autoresearch Reduce `python3 scripts/score.py` error_count to 0.

License and pricing

MIT license, as stated in the repository's LICENSE file.

Based on the project's public README and repository metadata.