bug(provider): GPT-5.6 *-pro models do not send reasoning.mode "pro" on Responses API

[tlerbao]

Summary

For GPT-5.6, OpenAI documents reasoning.mode: "pro" as a Responses API option that is orthogonal to reasoning.effort. In OpenCode/models.dev, derived *-pro model entries appear intended to represent the same base model plus reasoning.mode: "pro".

However, OpenCode's OpenAI Responses wire path appears to only forward reasoning.effort and reasoning.summary, dropping reasoning.mode. If so, selecting a GPT-5.6 *-pro model behaves like standard reasoning at the selected effort instead of Pro mode.

Environment

• Provider: built-in openai / @ai-sdk/openai
• Models: openai/gpt-5.6-pro, openai/gpt-5.6-sol-pro, openai/gpt-5.6-terra-pro, openai/gpt-5.6-luna-pro
• API path: OpenAI Responses API

Steps to reproduce

1. Select openai/gpt-5.6-sol-pro or another GPT-5.6 *-pro entry.
2. Set a reasoning variant, for example medium, via Ctrl+T.
3. Send any prompt and inspect the outbound /v1/responses JSON.

Expected behavior

The derived *-pro entry should keep the base model slug and send reasoning.mode: "pro" alongside the selected effort. For example:

{
  "model": "gpt-5.6-sol",
  "reasoning": {
    "mode": "pro",
    "effort": "medium",
    "summary": "auto"
  }
}

reasoning.mode and reasoning.effort are separate dimensions in the OpenAI GPT-5.6 Responses API; pro is not the same as xhigh or max.

Actual behavior

The request appears to include only effort/summary, for example:

{
  "model": "gpt-5.6-sol",
  "reasoning": {
    "effort": "medium",
    "summary": "auto"
  }
}

reasoning.mode is absent, so the request likely runs as standard reasoning.

Suspected implementation details

The high-level model/options merge appears able to preserve model options from derived model modes, but the OpenAI Responses lowering layer seems to only construct reasoning from reasoningEffort / reasoningSummary.

Relevant areas:

• packages/opencode/src/provider/provider.ts (fromModelsDevProvider, derived experimental modes)
• packages/opencode/src/session/llm/request.ts (options merge order)
• packages/llm/src/protocols/openai-responses.ts (Responses body/lowering)
• packages/llm/src/protocols/utils/openai-options.ts

Related issues

This is related to older Pro model discussions but the GPT-5.6 semantics are different: GPT-5.6 Pro is not a separate gpt-5.6-pro API model slug; it is the base GPT-5.6 model plus reasoning.mode: "pro".

Related, but not a duplicate:

• Codex (ChatGPT account) lists *-pro models it can't serve — gpt-5.5-pro shown then rejected by backend #32435

Notes

This may require either:

• adding first-class support for reasoning.mode in OpenCode's OpenAI Responses options/lowering, and/or
• coordinating with the upstream AI SDK path if that layer also drops reasoning.mode.

It would also be useful for the model picker/help text to make clear that *-pro means reasoning.mode: "pro", while Ctrl+T controls reasoning.effort independently.