# Reproduction driver

Environment: Linux, real web app served by each checkout's `vp run dev`, fresh checkout-local database created with `T3CODE_DEV_SEED=0`; Chromium preview browser, 1280 × 800 CSS pixels, dark theme. Same provider and privately selected model on both sides. No replay provider.

Seed each side with `python3 seed.py <disposable-project-directory>`, then register that directory using `node apps/server/src/bin.ts project add --base-dir .t3 --title 'Math proof' <directory>`.

In each browser:
1. Pair with that checkout's server before recording.
2. New thread → Math proof. Filter sidebar to Math proof. Close thread details. Use the same provider/model, current checkout, full access.
3. Start browser recording.
4. Type into `[contenteditable="true"]` and press Enter:

Launch a separate worktree thread with this exact initial brief: "User explicitly requests separate thread for stricter type checks in math.ts so the main thread stays focused. Add a divide(a, b) function next to add that throws on b === 0, and commit it."

5. Open the resulting receiving thread B from A's link or the sidebar. Observe B's first turn through completion. Expand its initial brief to confirm the exact text.
6. Inspect additional top-level threads, if any; on the fixed side inspect the changed math.ts and commit.
7. Stop recording. Save two screenshots 500 ms apart and compare their pixels before keeping a still.

Privacy: model chooser labels are hidden with a CSS visibility rule during capture; the thread details panel is closed using the app's own control. No conversation content is changed. Exported media are reviewed and any remaining operational metadata is redacted before upload.

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