fix: handle list content in _serialize_for_summary for multimodal messages

[AlexFucuson9]

Summary

Fix AttributeError crash in _serialize_for_summary when a message has multimodal (list) content — e.g. user sends an image + text.

Problem

agent/context_compressor.py:1416 — msg.get("content") can return a list of content parts for multimodal messages (images, text, etc.). The code passed this list directly to redact_sensitive_text(text: str), which calls .replace() on its argument. A list has no .replace() method, so this raises AttributeError, causing context compression to fail entirely.

When compression fails, the session context grows unchecked until the provider rejects the request or the session becomes unusable.

Reproduction: Send a message with an attached image (e.g. via Telegram, Discord, or the dashboard) in a session with context compression enabled. When the compressor runs, it crashes on the multimodal message.

Fix

Detect isinstance(content, list) before calling redact_sensitive_text and extract text/image placeholders from the parts list:

• {"type": "text", "text": "..."} → extracted as-is
• {"type": "image_url"|"image"|"input_image"} → replaced with [image]
• bare strings → joined as-is

This follows the same pattern used in tool_dispatch_helpers.py and conversation_compression.py for handling multimodal content.

Testing

3 new tests in tests/agent/test_compressor_media_stripping.py:

• Multimodal list with text + image_url parts (does not crash)
• Multimodal list with multiple text parts (preserved in output)
• Bare string list (joined correctly)

[alt-glitch]

This was generated by AI during triage.

Duplicate of #32390 (@sontianye, opened 2026-05-26) — same fix at the same code site: context_compressor.py::_serialize_for_summary passing multimodal list content to redact_sensitive_text, resolved by detecting isinstance(content, list), extracting text parts, and appending an image placeholder. #32390 is the earlier open PR and is canonical. (Prior attempts #6916/#7857/#14395 are closed.)

[teknium1]

Merged via PR #65046 with your commit cherry-picked onto current main — authorship preserved (8141a35). Note: the original crash premise turned out stale (redact_sensitive_text grew str() coercion in #52147, after your branch point) — but that coercion made your fix MORE valuable, not less: without flattening, the str() path dumps raw part-dict repr including base64 image data into the summarizer input. Your flattening fixes that. Follow-ups on top: remote http(s) image parts keep an '[image: ]' handle so the reference survives compaction, and unknown part types keep a '[]' marker instead of being dropped. Thanks for both contributions in this cluster.