# Pi extension handoff / Stop / restart proof

Environment for both sides: Linux web development app, isolated fresh T3 home per checkout, no live database seed, Chromium shared browser panel, light theme, 1280 x 800 CSS pixels (2560 x 1600 capture), Pi 1.5.2, claude-haiku-5-5 through the configured provider. No replay model or mocked Pi protocol.

Base: b4e47a5a62ec056dfcb95ee579c4390dfa04b345
Head: b55f7705e7d274c9d6e6a035e8d4d674cd1a33ec

Each app was installed with `vp i` and started with `T3CODE_DEV_SEED=0 vp run dev`. The seed script records the setup commands used for the final runs. The extension creates a real Pi session through ctx.newSession, inserts synthetic conversation context with setup, and starts a real model response through withSession.

1. Open the seed-created thread by its direct link after pairing. Wait for ORIGINAL SESSION READY.
2. Hide the sidebar and thread details panel, to keep environment details outside captures.
3. Type /handoff in the composer, dismiss the command picker with Escape, then click Submit message.
4. Wait for HANDOFF SESSION ACTIVE and the running sleep tool. Click Stop generation.
5. Wait for Run interrupted / Stopped by user.
6. Stop the tracked dev-server terminal session with Ctrl-C; start it again with the same T3 home and T3CODE_DEV_SEED=0. Reopen/reload the same thread. No intervening message.
7. Send exactly: What is the handoff verification secret from your conversation context? Reply only with the secret, or NO HANDOFF CONTEXT if it is absent. Do not use tools or inspect files.
8. Wait for the response to finish. Base returned NO HANDOFF CONTEXT; head returned VIOLET-COMET-742.

The screen recordings are split immediately before the server restart and resume after it. The joined evidence videos insert an explicit restart title card at that boundary. Idle tails are trimmed; model streaming and Stop interactions retain real timing. Screenshots were captured only after two screenshots at least 500 ms apart had identical SHA-256 hashes, and were visually inspected. The head was repeated on a fresh thread to capture the same panel layout as base; both completed head runs retained the context. An earlier head setup attempt restarted during cancellation and was excluded.

Only synthetic data is used. The model is instructed not to read files when recalling the secret, and no file-read tool was used in either follow-up.
