# Pi structured question evidence

Base: 6f5861fa27c9d1c21e5b48677ac0da1188bfd233
Head: 0b34c87e31384d64626a3c187b4fcdbb9a6d0a0e

Both are separate worktrees running `T3CODE_DEV_SEED=0 vp run dev`, each with a fresh T3 home. Real Pi 1.5.2, real @juicesharp/rpiv-ask-user-question 2.15.0, openai/gpt-6-luna. Browser viewport 1280x800 CSS pixels, device scale 2, light theme, details panel closed. Full screenshots are 2560x1600. Static composer crops use the identical rectangle on each side, excluding the unrelated timer and animated live-tool shimmer; two crops 500ms apart were compared pixel-for-pixel before retaining each still. Tool-result screenshots are uncropped and similarly checked. No application source was changed.

Prepare an empty synthetic Git project with one empty initial commit, then add it using `node apps/server/src/bin.ts project add PROJECT --title 'Question cards' --base-dir T3_HOME`.

Extract `npm pack @juicesharp/rpiv-ask-user-question@2.15.0` as `package/` next to the interleaving driver; install its runtime dependencies with `npm install --prefix package --ignore-scripts --legacy-peer-deps`. Enable Pi in the isolated T3 settings with launchArgs `--no-extensions --no-skills --no-mcp --no-context-files --extension ABSOLUTE_PATH_TO_PACKAGE/index.ts`.

Seed each thread through the real CLI using seed.py (basic or preview); prompt.txt and preview-prompt.txt are the exact prompts used. Open each thread by its environment/thread URL. The model calls the actual tool and says only Done afterwards; all answer evidence is the expanded Ask_user_question tool output, never a model recap.

1. Basic option selection: inspect the card, select Docker/Compose. Clicking the option submits it. Expand Worked for, Used 2 tools, Ask_user_question. Both sides return Docker/Compose; only head has the supplied Area header and separate labels/descriptions without Type something.
2. Basic custom answer (Literal custom answer proof): type 2 in the custom composer, click the exact Submit button. Wait until Worked for appears, then expand Worked for, Used 2 tools, Ask_user_question. Base actual tool output is Home Assistant; head actual tool output is literal 2. No second visible prompt appears on head. The paired custom-proof videos show this complete interaction.
3. Preview and full title: seed preview mode, inspect the first card, select Docker/Compose, inspect the second card, type 2 and submit. Questions are `Which area should we work on?` and `Which area should we work on? Only for staging?`, both with header Area and identical options. Docker/Compose preview is `services:\n  web: {}`. Expand Worked for, Used 3 tools, Ask_user_question. Preview text is retained in both; head renders each full question correctly and returns literal 2 for the longer question. Base returns Home Assistant.
4. Controlled interleaving: change only Pi's extension path to interleaved.ts, seed interleaved mode. This driver registers the published extension, executes its unchanged questionnaire implementation twice concurrently, and holds both select completions at a barrier. Type 2 then 1 into the two visible select cards, submitting each. Both answers are received before the driver asks an unrelated blank-placeholder input sharing the question's prefix. Type Ready there and submit. Then the barrier releases both real extension walkers. On head the two custom follow-up inputs are answered invisibly from the adapter's FIFO queue; there are no extra visible cards. Expand Worked for, Used 4 tools, Ask_user_question; inspect the output at its top, then scroll the output to its bottom. Base returns option Home Assistant then Docker/Compose. Head returns custom 2 then custom 1. Both preserve unrelated=Ready. interleaved-tool-output.json contains the exact DOM-read tool output from both sides.

The interleaving driver controls scheduling only; it does not replace the Pi process, native UI methods, T3 adapter, or published questionnaire code. This proves the controlled ordering, not that a production model will naturally issue the same concurrent calls.
