Codex Autoresearch 3.0 adds investigations for work where the goal is clear and the method may change. Accept the objective, allowed changes, evidence requirements, and budget once. Preparation, failed attempts, repairs, confirmation, and delivery all use that same allowance.
- Define success with numeric targets or explicit conditions and counterexamples.
- Resume recorded actions without starting them twice. A packaged worker keeps process actions running when the observing CLI closes.
- Keep valid observations separate from improved metrics, reached targets, and accepted code. Changed inputs can make earlier evidence inapplicable without erasing its history.
- Save useful changes from a rejected candidate as a patch for later assessment. Saving them does not accept or apply them.
- Confirm candidates through GitHub Actions and deliver only when the current evidence and requested endpoint are verified. The CLI and read-only dashboard show the same decision.
Existing short benchmark sessions and direct Codex work remain available. Adoption preserves original 2.x session records and requires an explicit remaining allowance when historical costs are unknown. The docs also correct stale automatic two-failure pause guidance; retries use the accepted limit.
Read Bounded investigations for setup, adoption, execution, and delivery.
This release is based on engineering verification. No model-driven comparison was run, and no advantage over ordinary Codex or 2.9.0 is claimed. The optional comparison tooling stays disabled without a separately accepted study budget. Automated browser checks do not establish manual screen-reader validation or accessibility compliance.