Console OpenAI variants send legacy SDK reasoning fields as raw JSON

[vimtor]

Description

Regression after #46134: V2 reads /api/v2/config, but existing Console OpenAI variants can store SDK-style reasoning fields in body. They now reach the upstream API unchanged and fail with 400 Unknown parameter: reasoningEffort. Native reasoning bodies work.

Plugins

Built-in OpenCode Console provider.

OpenCode version

0.0.0-dev-18806; earlier 0.0.0-beta-18743 interprets the existing configuration through /api/config.

Steps to reproduce

1. Use a working Console model with package aisdk:@ai-sdk/openai.
2. Configure a high variant with body: {"reasoningEffort":"high","reasoningSummary":"auto"}.
3. Select that variant and send a prompt.
4. Observe HTTP 400; the same model without the variant succeeds.

Expected: preserve existing variant behavior while accepting native body.reasoning and keeping Console configuration generic.

Operating System

macOS arm64.

[argszero]

I'd like to work on this.

I traced the regression to the variant body handling in the Console provider. In packages/core/src/plugin/provider/opencode.ts, variant options flow through ConfigProviderOptionsV1.model(withoutCredentials(options)) into existing.settings (line ~157-160). After #46134 switched V2 to /api/v2/config, SDK-style reasoning fields (reasoningEffort, reasoningSummary) stored in variant body are forwarded to the upstream API unchanged, producing 400 Unknown parameter: reasoningEffort — while the old /api/config path normalized them.

My plan:

1. Confirm exactly how variant body is serialized into the upstream request (settings → request body path).
2. Normalize SDK-style reasoning fields (reasoningEffort/reasoningSummary) in the variant body to the native body.reasoning shape when talking to upstream, so existing Console variants keep working.
3. Add a regression test covering a variant with body: {"reasoningEffort":"high","reasoningSummary":"auto"}.

I'll implement against v2 and open a PR linked to this issue.