OPENAI BUILD WEEK · DEVELOPER TOOLS

Let the agent write the fix. Never let it grade itself.

Codex authors a new regression test and an unapplied repair in your real repository. FORK freezes the red claim, binds human approval to one exact patch hash, and requires the same unchanged test to turn green. The result is fork-repair-evidence@2.

CODEX PLUGINLOCAL-FIRSTEXACT APPROVAL
30-SECOND JUDGE SANDBOX

Try the integrity boundary yourself

Open a sanitized artifact generated by the real verifier. Then alter one byte and watch the same browser-side validator fail closed.

This sanitized fixture comes from the verified Ghost Deploy session—not a simulated green screen. The browser revalidates the same evidence contract locally.

  1. 01Base source identity
  2. 02Post-session test
  3. 03Observed red run
  4. 04Exact human approval
  5. 05Applied production diff
  6. 06Same-test green
  7. 07Ordinary suite proof
01 · GPT-5.6 SOL · ADVISORY, NEVER AUTHORITY

Frames the review risk

Turns bounded red evidence and an exact diff into structured questions without gaining approval or verdict authority.

02 · CODEX

Discovers the missing claim

Inspects real code and authors a focused regression test after the FORK session begins.

03 · FORK

Separates authorship from authority

Requires observed red evidence and will not apply production bytes until a human approves their exact hash.

04 · EVIDENCE

Keeps one continuous story

Binds red, approval, applied diff, unchanged-test green, and the ordinary suite into one portable record.