AI SDK response model ID is discarded from assistant turns

[KarmCraft]

Description

OpenCode's AI SDK adapter receives the provider's normalized response.modelId on every finish-step, but drops it before the assistant turn is persisted. Clients therefore see only the requested alias (provider/auto), not the response model.

Preserve response.modelId as optional step-finish metadata so clients can report the model used for that turn. This does not persist response headers.

Related to #26091, which asks to expose raw response headers. This issue is the narrower structured-model path and blocks pingdotgg/t3code#6536.

Plugins

None

OpenCode version

1.15.13

Steps to reproduce

1. Configure an OpenAI-compatible provider with an alias such as auto.
2. Return a different model in the completion stream.
3. Inspect the assistant message events.
4. The assistant reports provider/auto; the response model ID is absent.

Screenshot and/or share link

pingdotgg/t3code#6536

Operating System

Windows 11

Terminal

T3 Code / OpenCode SDK

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• LLM response headers are discarded, preventing plugins from accessing proxy routing metadata #26091: LLM response headers are discarded — the same root cause (proxy routing metadata, actual model ID lost when using a proxy/router, response headers dropped in processor.ts). That issue also discusses preserving response headers so plugins and clients can see which model actually served the request.

[KarmCraft]

Thanks — #26091 is closely related. The distinction here is that this change preserves the AI SDK's structured response.modelId; it does not expose or persist arbitrary response headers. I'm keeping this issue focused on that safer client contract.