[rekram1-node]

Problem

assistant in to-llm-message.ts wraps message.model and model fields in String(...) even though both are typed as Model.Ref (unlike modelSwitched in the same file, which compares providerID and id directly).

Change

Compare Model.Ref fields directly in assistant.

 assistant(message, model)
-  sameProvider = String(message.model.providerID) === String(model.providerID)
-  sameModel = sameProvider && String(message.model.id) === String(model.id)
+  sameProvider = message.model.providerID === model.providerID
+  sameModel = sameProvider && message.model.id === model.id

Out of scope

• Changing provider metadata reuse rules.

[opencode-agent]

The change looks sound and leaves behaviour unchanged. With fewer than two eligible inline mentioned images, the dedupe loop can't find a duplicate, so it returns the same parts as the removed early-return path. Model.Ref.id and providerID are branded strings, so === matches the old String(...) comparison. In packages/core, session-runner-message.test.ts passes, including the dedupe and provider-metadata reuse cases, and typecheck passes.

The public API is unchanged. I have no concerns.

[GiuseppeBellamacina]

With fewer than two eligible inline mentioned images, the dedupe loop can't find a duplicate, so dropping the pre-filter returns the same parts as the old early return. The direct providerID / id comparisons match what modelSwitched already does in this file. Looks good to me.