Summary

Fixes #19105.

Fixes several WhatsApp media delivery paths:

• Converts successful image_generate JSON results into MEDIA:/path attachments at the gateway layer.
• Prevents old generated images from being re-sent by extracting history media paths from both MEDIA: tags and JSON tool payloads.
• Adds native WhatsApp MEDIA: support to send_message through the live WhatsApp bridge adapter.
• Sends WhatsApp images/videos/audio through native Baileys file URL payloads instead of treating them like document-style uploads.
• Adds inbound WhatsApp video paths to the model-facing message so tools can process the real cached file.
• Adds regression coverage for image JSON extraction, history dedupe, WhatsApp media routing, and inbound video path visibility.

Why

Generated media could succeed locally but fail or degrade at WhatsApp delivery.

For image generation, providers return JSON with an image field rather than a MEDIA:/path tag. The gateway therefore depended on the model restating the exact path correctly. Once JSON extraction was added, old image JSON in history also needed to be treated as already delivered, otherwise Hermes could send every previous generated image again.

For generated MP4 outputs, such as ASCII-video renders, delivery depends on MEDIA:/path routing correctly through WhatsApp.

For inbound WhatsApp videos, the agent needs the real cached path, not just a placeholder.

Tests

PYTHONDONTWRITEBYTECODE=1 python -m pytest \
  tests/gateway/test_media_extraction.py \
  tests/gateway/test_native_image_buffer_isolation.py \
  tests/tools/test_send_message_tool.py::TestSendToPlatformWhatsapp -q

Result:

12 passed

Also ran:

PYTHONDONTWRITEBYTECODE=1 python -m pytest tests/tools/test_send_message_tool.py -q

Result:

96 passed

Also checked syntax:

node --check scripts/whatsapp-bridge/bridge.js
python -c "import ast, pathlib; ast.parse(pathlib.Path('gateway/run.py').read_text()); ast.parse(pathlib.Path('tools/send_message_tool.py').read_text())"

Notes

This fixes the gateway and WhatsApp delivery side. Tools still need to create real output files and emit or return usable local paths. For ASCII-video specifically, an MP4 output should now be sent successfully when the skill/tool produces an actual file and returns it as MEDIA:/path/to/output.mp4.