Real app / Pi 1.5.2 / local synthetic HTTP fixture
Base: df15bd4f7bfcb5c41282e46d9de1092c01043b00
Head: 0a4f5b184c1a8dfc5362c7a2c0895b95e5f633dd

Gemini: POST /v1beta/models/gemini-2.5-flash:streamGenerateContent?alt=sse -> 429
Mistral: POST /v1/chat/completions -> 429
Both requests observed on both app revisions.
Base toast, both models: Text generation failed in generateCommitMessage: Pi could not generate text: the model provider reported an error.
Head toast, both models: Text generation failed in generateCommitMessage: Pi could not generate text: the model provider returned HTTP 429.
Neither raw error body nor either synthetic credential marker appeared in the toast.

Browser: Chromium preview, 1280x800 CSS viewport, light theme, Linux.
Each side uses its own fresh server database, synthetic git repository and replay-seeded thread.
Pi itself performs text generation; replay only seeds the thread history.
Still pairs were pixel-identical before saving (captures separated by at least 500 ms).
Videos redact only the environment hostname in a narrow black bar; error text remains visible.