# Word wrap visual reproduction

Run each side from its own checkout and fresh T3 home with `T3CODE_DEV_SEED=0 vp run dev`. Both use Chrome, a light theme, a 1280 × 1100 viewport, and the same synthetic replay text. `seed.mjs <checkout> <data-directory>` creates the git sample and replay thread through the T3 CLI. Its only temporary source edit supplies synthetic replay content; it restores the recording afterwards. `drive.mjs <agent-browser-session> <capture-directory>` drives the chat and uncommitted diff flow. Open and authenticate the isolated app in that browser session first, with the wrap default on and workspace panels closed. The recording begins before the second block’s wrap toggle is clicked.

For PR Code, `forge-fixture.mjs <socket-path> <ready-file>` serves synthetic GitHub API responses on a local Unix socket. Give the dev server an isolated `GH_CONFIG_DIR` containing `config.yml` with `http_unix_socket: <socket-path>` and `GH_TOKEN=synthetic-fixture`. All PR data is synthetic; no pull request is created on GitHub. Add the sample repository's local origin as `https://github.com/wrap-demo/sample.git` and open its PR in the app using the synthetic repository, number 7, and the sample's project ID.

`drive-pr-tabs.mjs <agent-browser-session> <capture-directory>` starts on that PR page. It keeps Code mounted in the original tab and edits Appearance in a second tab. This matters: navigating the PR page itself to Settings remounts Code and masks the stale-state defect. The driver covers the wide toolbar, narrow dropdown, and overrides set with both controls. Panel maximization and viewport changes are identical on both sides.

Every final still is copied only after two byte-identical screenshots at least 500 ms apart. The chat recordings include Settings navigation and the initial local-toggle click. PR recordings follow the original Code tab while Appearance changes in the second tab; paired Settings stills show the actual off/on defaults. Both commits receive the same UI actions. JSON observations are supplementary runtime measurements of wrap state, node identity, and scroll position; the images and recordings are the evidence.