fix: handle multimodal content in context compression summarization

[teknium1]

Summary

Fixes a bug where multimodal user messages (containing images, audio, etc.) would produce garbled output during context compression summarization.

The bug

In _generate_summary(), message content was accessed as:

content = msg.get("content") or ""

When content is a multimodal list like:

[
    {"type": "text", "text": "What is in this image?"},
    {"type": "image_url", "image_url": {"url": "data:image/png;base64,..."}},
]

The `or """ fallback doesn't trigger (non-empty lists are truthy), and then:

• len(content) returns 2 (list length) instead of character count
• content[:1000] slices list items instead of characters
• The resulting summary prompt contains raw dict representations instead of readable text

The fix

New _content_to_text() static method on ContextCompressor

Safely converts any content format to plain text:

Input type	Output
str	Returned as-is
None	Empty string
list (multimodal)	Text parts joined with newlines; image_url → [image]; other types → [type_name]
dict / other	JSON serialization, str() fallback

Changed line in _generate_summary()

# Before:
content = msg.get("content") or ""

# After:
content = self._content_to_text(msg.get("content"))

Tests added

• TestContentToText class with 6 test cases:
  • String passthrough
  • None → empty string
  • Multimodal text + image parts
  • Mixed content types (text, audio, etc.)
  • Dict content JSON serialization
  • End-to-end: multimodal messages through _generate_summary() — verifies the summarization prompt contains "What is in this image?" and "[image]" instead of raw dicts

What does NOT change

• The compression algorithm, role alternation, or trigger logic
• How string content is handled (identical behavior for the common case)
• The content=None handling that already worked via the existing TestGenerateSummaryNoneContent tests

Inspired by PR #776 by @kshitijk4poor.

[teknium1]

Thanks for the PR. I looked into this, but it doesn’t correspond to a real Hermes message-flow path today. We don’t keep raw multimodal image blocks in the conversation history that context compression summarizes; image handling goes through the vision tool, and the resulting text description is what gets stored in chat history. Because of that, this edge case isn’t currently actionable in the live architecture, so I’m closing this for now. If we later add a path that stores raw multimodal content directly in chat history, we can revisit it then.