Summary

WhatsApp media delivery can fail or behave inconsistently even when Hermes successfully receives or generates the media file.

Observed workflows:

1. Sending a video/photo to Hermes over WhatsApp.
2. Asking Hermes to generate an image.
3. Asking Hermes to generate media output such as an ASCII-video MP4.

The failures appear across several related paths:

• Inbound WhatsApp videos can be surfaced to the model as placeholders instead of real cached file paths.
• image_generate returns JSON with an image path, but the gateway previously only auto-delivered tool results that already contained MEDIA:/path.
• Prior image-generation JSON results in session history can be rediscovered and re-sent, causing WhatsApp to send all previously generated images instead of only the latest.
• send_message supports MEDIA:/path native delivery for several platforms, but not WhatsApp.
• The WhatsApp bridge can send generated images in a way that appears as a file/document instead of a native image bubble.

Observed Behavior

• A WhatsApp video/photo can arrive as [video received] or similar without a usable local path.
• image_generate saves a PNG successfully, but WhatsApp delivery fails or only returns the path.
• Subsequent image generations may send multiple old generated images from conversation history.
• A tool-created MP4, such as an ASCII-video output, may be described in prose instead of attached as native media.
• Generated images may arrive as file/document attachments instead of inline image messages.

Expected Behavior

• Inbound WhatsApp video/photo attachments should expose their real local cached paths to the agent.
• Successful generated media outputs should be delivered as native WhatsApp attachments when Hermes has a real local path.
• Only media generated in the current turn should be delivered, not older media paths from history.
• send_message should support MEDIA:/path for WhatsApp using the live WhatsApp bridge adapter.
• PNG/JPEG/WebP/GIF outputs should appear as native WhatsApp images, not documents.
• MP4 outputs, including ASCII-video outputs, should route through WhatsApp as native video attachments when the file exists.

Fix Direction

• Normalize successful image-generation tool JSON results into MEDIA:/path at the gateway layer.
• Extract historical media paths from both literal MEDIA:/path tags and image-generation JSON payloads to prevent re-sending old images.
• Preserve existing MEDIA:/path behavior for tools that already emit it, including generated MP4 outputs.
• Route WhatsApp send_message media through the live WhatsApp adapter.
• Use native WhatsApp image/video/audio payloads in the bridge.
• Include inbound video local paths in model-facing message text.