hermes - 💡(How to fix) Fix Allow video analysis to use a separate auxiliary backend from image vision

Feature description

Hermes currently has auxiliary.vision for image/screenshot/video-like analysis routing. In practice, image and native video workloads often need different backends:

• Images/screenshots may work best on the active main model's native vision path.
• Native video analysis may require a provider/model that explicitly supports video input, such as Gemini video-capable models.

It would be useful for video_analyze to support a separate auxiliary.video config block while leaving auxiliary.vision and main-model native image routing untouched.

Motivation

Without a separate video backend, users must either route all vision tasks to a video-capable auxiliary model or risk image-capable-but-not-video-capable models receiving video inputs. This is especially awkward for profiles whose main model supports images but not native video.

Proposed solution

Support config such as:

agent:
  image_input_mode: native
auxiliary:
  vision:
    provider: auto
    model: ''
  video:
    provider: gemini
    model: gemini-3.1-flash-lite
    timeout: 180

Then video_analyze should prefer auxiliary.video when present, while image analysis continues to use auxiliary.vision / native image routing.

Local workaround tested

A local patch in tools/vision_tools.py reads auxiliary.video.provider/model/base_url/api_key/timeout for video_analyze_tool() and passes those values to the auxiliary client. This allowed images to remain on native model vision while videos used Gemini.

Environment

• Hermes checkout commit: 9f182bd7b
• OS: Ubuntu 24.04 / Linux