Summary

On the latest main, inbound Matrix m.image events still treat content.body (usually just the uploaded filename like 30.png) as the user text payload.

This becomes problematic when Hermes later prepends auto-vision analysis to the same message text: the model receives something semantically equivalent to:

[vision description of the image]

30.png

For image-only Matrix messages, that trailing raw filename can mislead the model into treating the turn as a file/path lookup instead of an image understanding request.

Why this matters

The media download / cache path itself is working correctly. The image is cached locally and passed into the vision enrichment pipeline. The issue is the semantic packaging of the inbound message.

In practice this can produce behavior like:

• user sends an image in Matrix DM
• Hermes caches it successfully
• Hermes then tries to search_files / reason about 30.png as a literal filename
• instead of responding to the image content

Current code path on latest main

1. gateway/platforms/matrix.py:

   • _handle_media_message() reads body = source_content.get("body", "") or ""
   • for m.image, that body is typically just the upload filename
   • after context resolution it constructs:

   msg_event = MessageEvent(
       text=body,
       message_type=msg_type,
       source=source,
       raw_message=source_content,
       message_id=event_id,
       media_urls=media_urls,
       media_types=media_types,
   )

2. gateway/run.py:

   • if event.media_urls contains images, Hermes calls _enrich_message_with_vision(user_text, image_paths)
   • that function prepends the vision description to the existing user_text

So for image-only Matrix messages, the filename survives as the message text and is mixed into the final LLM input.

Expected behavior

For Matrix image-only messages:

• if there is no real user caption, MessageEvent.text should probably be empty
• or be wrapped in an explicit semantic marker instead of a bare filename
• caption vs filename should be distinguished

This would align better with how image-only messages are typically represented in other channels, and avoids confusing the model while preserving the cached local image path for downstream vision tools.

Repro

1. Run Hermes with Matrix enabled and vision enabled.
2. Send an image-only message (no caption) in a Matrix DM.
3. Observe that the inbound body is the filename (for example 30.png).
4. Hermes may interpret the turn as being about the literal filename rather than the image contents.

Suggested fix directions

Possible fixes upstream:

1. In Matrix _handle_media_message(), treat m.image filename/body separately from user caption text.
2. For image-only events, set MessageEvent.text to empty string instead of raw filename.
3. If keeping some textual hint is desirable, wrap it semantically, e.g. [User sent an image: 30.png], rather than passing a naked filename.
4. More generally, preserve the cached local media path in media_urls, but avoid leaking transport-level filename metadata into the user-intent text unless it is actually a caption.

Notes

I confirmed this is still present on latest origin/main in local inspection, not just in an older downstream integration.