# Manual real-app reproduction

Environment on both sides: Linux web client, Chrome preview browser, 1280 x 800 CSS pixels, light theme, Codex CLI 0.169.0, API-key authentication. Separate worktrees and fresh databases; no production seed. A synthetic empty Git project titled "Speed proof" is registered on each server. All times in the protocol excerpts are UTC.

Base: e9b48b7809974913515b10e4f2f7a0fd256a985c.
Head: ab0a37a23e1bd06274aee0d117065439152d57db.

1. Install with `vp i`; start each app with `T3CODE_DEV_SEED=0 vp run dev`.
2. Register the synthetic project with `node apps/server/src/bin.ts project add --base-dir .t3 --title 'Speed proof' "$PROJECT"`.
3. Start a native Codex goal through the server CLI, which calls the real thread launch service:

```sh
node apps/server/src/bin.ts thread start --base-dir .t3 \
  --provider codex --model gpt-6-luna --option effort=medium \
  --option serviceTier=fast --title 'Codex speed proof' --json "$PROJECT" \
  '/goal Do exactly 6 steps, strictly one step per turn. In each step run the shell command "sleep 40" once, then reply "step N done" and end the turn.'
```

The composer displays Medium (the model default). The initial native collaboration mode uses the model default reasoning effort; the picker persists explicit `reasoningEffort=medium` when Standard is selected.

4. Pair the browser with each fresh app, navigate directly to the returned thread using `/<environmentId>/<encoded threadId>`, close Thread details (to omit environment identity), and record the tab.
5. While step 1 is running `sleep 40`, click the composer button named `GPT-6-Luna, Medium Fast`, click radio `Standard`, and press Escape. Send no message. The exact same picker interaction is used on both sides.
6. Observe subsequent native goal turns. Let all six steps complete. Open the picker again to show Standard remains selected, and capture the final app state.
7. Extract a minimal, public-safe protocol transcript with `extract-evidence.py CHECKOUT THREAD OUTPUT.json`. The extractor reads only that run's provider log, its own native rollout, and its own read-only database. It emits timestamps, method names, speed settings, and synthetic `sleep 40` commands. Full logs are deliberately not uploaded.

The before video has only the environment-name rectangle blanked for privacy and is scaled to 1280 x 800; no speed control or conversation content is masked. The after video uses the same scale. Captures with a protocol panel append an explicitly labelled excerpt below the untouched app screenshot; the JSON transcripts are the underlying source.

Observed base: picker persisted default, no outgoing settings update after it, last applied native tier priority, six sleep commands and seven native turns (including a final summary).
Observed head: picker persisted default, outgoing thread/settings/update default, native thread_settings_applied default and incoming thread/settings/updated default before goal turn 2.

Recorded with GPT-6.1-Sol through Codex.
