Image generation doesn't carry over context from earlier messages #185

Where: franklin.run/chat, composer in Image mode (GPT Image 2.5 Flare, 3:2)

Steps

1. In chat mode, send: "Search the web for the latest x402 usage data and report the 30-day transaction volume, transaction count, and unique paying agents." Franklin answers with figures from three sources (e.g. 75.41M transactions, $24.24M volume, 94.06K buyers).
2. Switch the composer to Image mode and send: "Turn these numbers into a clean, modern infographic showing x402's activity over the last 30 days. Highlight transaction volume, transaction count, and unique paying agents."

Expected

An infographic built from the figures in the previous answer.

Actual

A polished dashboard whose figures don't match the previous answer: $1,248,732 volume, 287,540 transactions, 18,732 paying agents, +42% / +56% / +48% vs. the previous 30 days, and a date range of Apr 12 – May 11, 2025.

Nothing in the UI says these numbers are generated, and the UI doesn't show the prompt that was sent to the image model, so there's no way to tell from the chat what went wrong.

Likely cause

From reading the code at 245e93b; I haven't traced the actual ImageGen call.

• Image-mode input is wrapped as Generate an image with the ${mediaModel} model, size ${imageSize} from this prompt:\n\n${text} (use-franklin-chat.ts#L421).
• The agent keeps the conversation history for this turn (same convId, so server.ts doesn't /clear), so the chat model could resolve "these numbers".
• But "from this prompt:" reads as "pass this text through", and ImageGen sends exactly the prompt it receives (imagegen.ts#L258). Its prompt parameter description doesn't say the prompt must be self-contained.
• So the image model most likely received "Turn these numbers into…" with no numbers, and invented its own.