AI-assisted; verified with the tests and tools listed above.

Use separate checkouts and empty T3 homes for the base and head. Run prepare-proof.py HOME to initialize each fresh home with projectSettingsFolded=true (the normal post-fold state). No database is copied or modified.

Place proof-driver.ts in apps/server/src in each checkout. Install dependencies with vp i. Start each app with T3CODE_DEV_SEED=0 T3CODE_AUTO_BOOTSTRAP_PROJECT_FROM_CWD=0 vp run dev --home-dir HOME --no-seed. Pair a Chromium browser using that app's startup URL, and wait for sign-in to finish before navigating again.

Run node apps/server/src/proof-driver.ts HOME NEW_WORKSPACE. It creates a small Git repository, calls the real t3_project_create/update/read MCP tools, launches a replay turn in a new worktree from main, waits for the turn, and checks SETUP_RAN.txt. Its stdout is the public runtime transcript; HOME/private-result.json is private navigation metadata and must not be uploaded.

Open project settings at /settings/projects?project=<URL-encoded environmentId:workspace>#project-actions. Viewport: 1280x800 CSS pixels, device scale 2, light theme, Linux Chromium web client. Captures crop the same 880x260 CSS-pixel settings region on both sides, excluding machine names, model controls and filesystem paths.

For the recordings, reset both projects using the same driver with a fourth argument reset, then start recording with the settings page already visible. Run the driver with a fourth argument update: it repeats the same project update and worktree launch, using recorded-setup as the branch. The recordings are the first 12 seconds at original speed. The public *-recorded-transcript.json files contain those launches' results. Screenshots show the final state; two captures at least 500 ms apart were pixel-identical in the published region before each still.
