Why I built it
Codex can make a plausible improvement quickly. The harder problem is proving that the change helped, noticing what it traded away, and letting the next session continue without guessing which attempts already failed.
Autoresearch gives that work a measurement loop and a durable memory. It is useful when several small experiments can teach you more than one large rewrite.
How the loop works
A session records the goal, benchmark, metric, checks, budget, and allowed files. It measures a baseline, validates the benchmark, runs one packet at a time, and logs each result as a keep, discard, failure, or measurement.
Accepted work reaches a read-only finalization preview before any review branches are created. The optional dashboard reads the same ledger and never becomes a second control plane.
Where it stands
Version 2.7.1 is the current stable release. It is available through the Codex plugin marketplace, runs in local Git repositories, and supports both numeric benchmarks and accepted quality-gap checklists.