fix(core)!: include multimodal blocks in get_buffer_string prefix format

[Mason Daugherty (mdrxy)]

Breaking change - v2 candidate

get_buffer_string now includes image/audio/video block references (e.g. [image: <url>]) in the default prefix format instead of dropping them.

---

The default (non-XML) get_buffer_string only kept text content and silently dropped image, audio, and video blocks. That is lossy: a user who asks "what does this screenshot show?" with an attached image URL ends up with a summary that no longer references the image at all. Only the format="xml" path preserved that information, so callers such as SummarizationMiddleware had to opt into XML to avoid the loss.

This updates the shared utility so every caller of the default format benefits: non-base64 image, image_url (OpenAI-style), audio, and video blocks are appended as a concise human-readable reference. Plain string content and the XML format are unchanged, and base64/data: media is still omitted to avoid dumping payloads.

Before / After

Given a multimodal HumanMessage:

from langchain_core.messages import HumanMessage, get_buffer_string

msg = HumanMessage(content=[
    {"type": "text", "text": "What does this screenshot show?"},
    {"type": "image_url", "image_url": {"url": "https://example.com/screenshot.png"}},
])
get_buffer_string([msg])

Before (image URL silently dropped):

Human: What does this screenshot show?

After (image reference preserved):

Human: What does this screenshot show? [image: https://example.com/screenshot.png]

Made by Open SWE

[open-swe]

✅ Open SWE Review: No issues found

Open SWE reviewed this PR and found no potential bugs to report.

Open in Web • View Open SWE trace