hermes - 💡(How to fix) Fix Bug: Model switch mid-session not persisted to database (dashboard shows stale model)

Description

When a user switches models mid-session (via /model command in Telegram/Discord), the gateway updates the agent in memory and registers the session override, but does not update the database. This causes the dashboard "Models" page to show the stale original model instead of the actually-used model.

Root Cause

In hermes_state.py, the update_token_counts() method uses:

model = COALESCE(model, ?)

This only sets the model if the current value is NULL. Since the session was created with the original model, switching to a new model mid-session doesn't update the database record.

Additionally, the gateway's /model command handler (gateway/run.py, _handle_model_command()) updates the cached agent and stores the override in _session_model_overrides, but never calls any database update method to persist the new model.

Steps to Reproduce

1. Start a session with model A (e.g., xiaomi/mimo-v2.5-pro)
2. Switch to model B mid-session via /model xiaomi/mimo-v2.5
3. Continue using the session with model B
4. Check the dashboard "Models" page — it still shows model A

Expected Behavior

The dashboard should show the model that was actually used, not the original model from session creation.

Impact

• Dashboard analytics are inaccurate for sessions with model switches
• Users can't track which models they actually used
• The "Models Used" count may be incorrect

Suggested Fix

When a model switch occurs via /model, the gateway should update the database:

# After storing the session override
db.update_session_model(session_key, result.new_model)

Or modify update_token_counts() to always update the model when a new value is provided:

model = ?  -- instead of COALESCE(model, ?)

Environment

• Hermes Agent version: latest (dev)
• Platform: Telegram
• Provider: OpenRouter
• Models tested: xiaomi/mimo-v2.5-pro → xiaomi/mimo-v2.5