Add option to always default to currently loaded model when switching conversations

[Extraltodeus]

Which version of LM Studio?
LM Studio 0.4.6 (Build 1)

Which operating system?
Windows 10

What is the bug / feature request?
When switching to a conversation that was created with a different model, LM Studio requires manually re-selecting the model — even when there is already one loaded and ready.

The expected (and most common) intent is simply "use whatever is currently loaded." There's no option to make this the default behavior, so every conversation switch that involves a model mismatch results in a forced, unnecessary click.

To Reproduce

1. Load any model
2. Have two or more conversations created with different models (or no model assigned)
3. Switch between conversations
4. Observe that the model picker requires manual re-selection each time

Expected behavior / Feature request
Add a setting such as "Always use currently loaded model" that bypasses the per-conversation model assignment when a model is already loaded. This would eliminate the forced re-selection click on every conversation switch.

Workaround
None. The click is unavoidable.

[Staphylea]

From my experience with LM Studio 0.4.6 on Windows the current behavior is actually already reasonable in many cases:

• When switching between conversations that were created with the same model that is currently loaded, LM Studio reuses that model without any extra clicks – which is exactly what you’d expect.
• The extra click only appears when switching to a conversation that was created with a different model (e.g. vision vs. non‑vision, or a different base model).
• In those cases it’s actually desirable to require an explicit switch:
  • A conversation that relies on a vision model should not silently run on a non‑vision text model.
  • Likewise, conversations that are tailored to a specific model or tool setup should not be auto‑rerouted to some arbitrary currently loaded model.

Because of this, “always use whatever is currently loaded” would be convenient in some workflows, but could easily produce incorrect or broken behavior in mixed model setups (especially with vision models).