Real app reproduction, both commits

Surface: Linux web client, Chromium, 1280 x 800 CSS viewport, light theme.
Fresh independent T3 homes, T3CODE_DEV_SEED=0. Pi 1.5.2, isolated agent configuration.
Model: claude-haiku-4-5-20251001 via OpenAI-compatible chat completions.
Gateway: nginx 1.24.0 (Ubuntu), client_max_body_size 25m, real upstream provider.
No stub replies or injected orchestration failures.

1. Generate PNG files using seed.py, which creates deterministic synthetic noise images, 1600x1200, approximately 5.5 MiB each.
2. Create a demo git repository and add it with the T3 project add CLI.
3. Start a real Pi thread with Reply with just: ready. Wait for ready.
4. Open that thread by its environment/thread URL in the web client.
5. Set input[type=file] to screenshot-1.png through screenshot-4.png.
6. Type Reply with just: images received into [data-testid=composer-editor] and click button[aria-label="Submit message"]. Wait for images received.
7. Upload screenshot-5.png through screenshot-8.png through the same input. Send the identical prompt. Wait for Provider error.
8. Type Can you try again? in the composer and submit. Wait for the second Provider error.
9. Read only the error turn items from each isolated dev database. Save status and failure fields in failure-transcript.json.
10. Close thread details for still captures, blur the composer, verify two screenshots at least 500 ms apart are pixel-identical.

Recordings include steps 5-8. Both videos are cropped to the conversation column to exclude the environment's machine label, then downscaled to 732x800. No timing edits.
Screenshots retain the full 1280x800 CSS viewport (2560x1600 physical pixels).

Base: 789d10fb54adfce468dbf5c7ad8f118ceeed6c53
Head: caddb9b61a28f5189584e4beff4fe6be21e454c6

The default gateway accepted the oversized history during preliminary setup. A real local nginx proxy with the reported 25 MiB cap was therefore placed before the upstream for the final paired runs.
