# Context handoff steering reproduction

Use two isolated checkouts, one at each revision:
- Base: 2df750e0c94559c9e917d68826f642499e9c9a1c
- Head: ff5928eef07c45d03873d64640a7b21663111082

Run `vp i` in each. Start with fresh T3 state; never seed from live data.
Create two empty Codex homes per side. Put the authorized proxy credentials and a minimal proxy configuration in each home locally; do not publish either file.
Use `seed.py <checkout> <home-A> <home-B>` before starting the server.
Run `T3CODE_DEV_SEED=0 vp run dev` in each checkout, and pair the browser using that server's startup link.

Create an empty git repository for synthetic work. In each checkout run:

```sh
node apps/server/src/bin.ts project add "$WORKSPACE" --title 'Context handoff proof' --base-dir .t3
node apps/server/src/bin.ts thread start "$WORKSPACE" 'Run the shell command `sleep 90` and then reply with the single word DONE' --title 'Context handoff proof' --provider codex --model gpt-6.1-sol --base-dir .t3 --json
```

The start command dispatches through the real server and Codex adapter. Open the returned thread using its direct environment/thread URL.

Both captures use the web client in Chromium at a 1280x800 CSS viewport (device scale 2), light theme, with thread details closed before recording.

Recorded browser actions on both sides:
1. While Codex A shows `Running sleep`, start recording.
2. Open the GPT-6.1-Sol model picker.
3. Click the `Codex B` provider button.
4. Click `GPT-6.1-Sol , Codex B` in the model list. The timeline displays `Context handoff pending`, CO 6.1 Sol to CB 6.1 Sol.
5. Fill `[data-testid="composer-editor"]` with `Change of plan: stop waiting and reply with the single word STEERED.`
6. Press Escape to dismiss the picker, click the composer, and press Enter to send the steer.
7. Wait for `STEERED` and completion. Stop recording and capture the final timeline.

Observed: base completed with STEERED and no Context handoff row; head completed with STEERED and a persistent Context handoff CO 6.1 Sol to CB 6.1 Sol row above the steer.
Final screenshots were taken twice at least 500 ms apart and were byte-identical on each side.
No test assertion is used as evidence. The replay scenario runner was not applicable because its generic ACP agent cannot exercise separate Codex homes.
