🤖 fix: attach tool screenshots as image parts

[ibetitsmike]

What

Rewrite provider requests so tool-result media (e.g. Chrome DevTools take_screenshot) is not sent as large base64 blobs inside tool-result JSON.

Concretely:

• Replace tool-result { type: "media", data: "<base64>", mediaType } entries with small text placeholders.
• Emit a synthetic user message right after the assistant/tool message that attaches the extracted image(s) as multimodal parts.

This is applied both:

• When building the initial request from persisted history (AIService), and
• For internal streamText() steps via prepareStep, so tool results produced mid-stream are also rewritten before the next LLM call.

Why

Some providers treat tool-result payloads as text/JSON, so embedding screenshot base64 in tool results can easily exceed context limits ("prompt too long"). Attaching screenshots as proper image/file parts avoids that failure mode.

Notes

This is request-only: persisted history/tool output remains unchanged for UI rendering.

Validation

• Unit tests for the request transforms
• make static-check

---

Generated with mux • Model: openai:gpt-5.2 • Thinking: xhigh • Cost: 3119344{cost}