# Image recovery proof driver

On each commit, start the web app with T3CODE_DEV_SEED=0 and a fresh worktree-local T3 home. Set the default model to Codex gpt-6-astra (Medium), with binaryPath pointing to this wrapper:

```sh
#!/bin/sh
if [ "$1" = "app-server" ]; then sleep 15; fi
exec codex "$@"
```

Create a synthetic git project containing only a README. Open a new thread in the preview browser at 1280x800, light theme, details panel closed. Start recording before submitting.

Paste a 480x240 PNG produced by a canvas: white background, red circle centered at (120,120), radius 65; blue rectangle at (270,65), width 150 and height 110. The attachment is named shapes.png.

Send: Describe the attached image: what shapes and colours are in it?

The timing driver runs this JavaScript in the preview page:

```js
const send = [...document.querySelectorAll('button')].find(b => b.getAttribute('aria-label') === 'Submit message');
if (!send || send.disabled) throw Error('Send not ready');
send.click();
await new Promise(r => setTimeout(r, 6000));
const stop = [...document.querySelectorAll('button')].find(b => b.getAttribute('aria-label') === 'Stop generation');
if (!stop) throw Error('Stop missing');
stop.click();
```

Confirm the timeline says Run interrupted before provider start. Send this text-only follow-up: Clarification: answer in exactly two sentences. If you cannot see an image, say NO IMAGE.

Wait for completion and capture the transcript, original attachment, interruption, and reply. Stop recording.
