Thread titles

Generates thread titles for real thread shapes with the server's Codex title generator, then sets each new title against the old one for a blind review.

Run

cd evals
pnpm thread-titles --model <codex model> --out runs/titles
pnpm thread-titles --model <codex model> --out runs/titles-new --baseline runs/titles/results.json

• It uses your local Codex login, like the server does.
• --initial titles from the first message only, as a new thread does.
• A run writes results.json (titles and latency), review.json (titles A and B in random order, with a rubric), and answer-key.json. Score review.json before you open answer-key.json.

Add a case

Add an entry to cases.ts with an id, a source, the thread messages, the previousTitle, and a rubric that says what a good title names.