# Grouped subagent account badge reproduction

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

Use two isolated checkouts, one at each recorded commit, and fresh T3 homes. Disable real providers and set the default provider to the development replay agent. Start each with `T3CODE_DEV_SEED=0 vp run dev --no-seed`; expose the backend on all interfaces for mobile.

Create a synthetic git repository containing only a README. Add it with `node apps/server/src/bin.ts project add <workspace> --title "Badge demonstration" --base-dir .t3`.

Start a parent using `node apps/server/src/bin.ts thread start <workspace> "Inspect grouped account badges" --title "Account badge comparison" --provider devReplay --model question --base-dir .t3 --json`. The replay question holds the parent active while the driver delegates.

Pair the browser with the isolated server. Evaluate the seed-browser.js function with `{appDir: <checkout>, threadId: <returned id>}` in that browser. It uses the app's own authenticated RPC commands to configure four replay provider instances with AA, BW, GR and VS initials and accent colours, then delegates three real replay children. Wait for `3 done`.

Capture collapsed and expanded groups in dark and light themes. Add the fourth child using `window.addBadgeAgent(3)`, wait for `4 done`, and repeat. The web capture driver awaits finite animations, takes two screenshots at least 500 ms apart, and checks that the cropped group pixels match. All crop coordinates come from the real group's DOM bounds; the rendering is unchanged.

For native iOS, install the checkout's dev client with `vp run mobile:install ios --session <leased-session>`, start its Metro, and pair it to that checkout's backend. Open the same synthetic parent. A separate three-agent parent, seeded identically, retains the three-agent state after the four-agent web capture. Switch system appearance for each theme and expand/collapse the group. For the four-agent parent, collapse the replay question input and scroll upward so the group header remains on screen in both states. Leave the three-agent parent input open. Wait for all native layout changes before capturing. Native crops use 3 pixels per point, and paired captures 500 ms apart must match. Inspect each crop and full screenshot before upload.

The attempted automated proof did not reach the visual assertion: its initial seed used a completed parent and delegation correctly rejected it. The published evidence comes from the manual reproduction above, not those failed runner reports.