Before submitting
Area
apps/server
Problem or use case
OpenCode can expose an alias such as auto while a router chooses a different backend model for each turn. T3 only retains the alias, so I cannot see which model actually answered.
Related to #2481, but the selected model and the routed model are not the same. #5699 also touches provenance, for externally appended messages rather than provider responses.
Proposed solution
Let provider adapters attach optional, safe per-message metadata for the actual model, then show it in a small assistant-message footer, for example: OpenCode · auto → GPT-5.6 Luna · high.
If the provider does not expose this metadata, keep the current UI unchanged.
Why this matters
With dynamic routing, model attribution is needed to understand quality, latency, and cost without asking the assistant after every turn.
Smallest useful scope
Support optional actual-model metadata for OpenCode assistant messages and render it on web.
Contribution
Before submitting
Area
apps/server
Problem or use case
OpenCode can expose an alias such as
autowhile a router chooses a different backend model for each turn. T3 only retains the alias, so I cannot see which model actually answered.Related to #2481, but the selected model and the routed model are not the same. #5699 also touches provenance, for externally appended messages rather than provider responses.
Proposed solution
Let provider adapters attach optional, safe per-message metadata for the actual model, then show it in a small assistant-message footer, for example:
OpenCode · auto → GPT-5.6 Luna · high.If the provider does not expose this metadata, keep the current UI unchanged.
Why this matters
With dynamic routing, model attribution is needed to understand quality, latency, and cost without asking the assistant after every turn.
Smallest useful scope
Support optional actual-model metadata for OpenCode assistant messages and render it on web.
Contribution