hermes - 💡(How to fix) Fix Gateway image auto-routing ignores session model overrides [1 pull requests]

Bug Description

In gateway sessions, agent.image_input_mode: auto decides native-vs-text image handling from the persisted config/default model instead of the effective per-session model selected with /model.

This is especially visible in Telegram:

• config/default model is text-only (for example mimo-v2.5-pro)
• the current Telegram session is switched with /model to a vision-capable model
• screenshots are still pre-analyzed through vision_analyze, so the main model receives a text summary instead of native pixels

The inverse can also happen:

• config/default model is vision-capable
• the current session is switched to a text-only model
• auto may route images natively even though the effective session model cannot consume pixels

Expected Behavior

agent.image_input_mode: auto should resolve the same effective model/provider that the upcoming gateway turn will use, including session-scoped /model overrides, before choosing:

• native for vision-capable effective models
• text / vision_analyze fallback for text-only effective models

Actual Behavior

Image preprocessing runs before the agent runtime globals are set for the turn. The gateway's image-routing helper reads the main model/provider from config/runtime defaults, so it can ignore the session-scoped /model override.

Reproduction Shape

1. Configure a text-only default model.
2. In Telegram, switch the current session/topic to a vision-capable model with /model.
3. Send a screenshot.
4. Observe that auto uses text pre-analysis instead of attaching pixels natively.

A mirror regression exists with a vision default + text-only session override.

Notes

I have a small fix branch that makes the gateway image-routing decision resolve the same session runtime as the upcoming agent turn and adds regression tests for both directions.