[FEATURE]: expose reasoning effort "max" for GPT-5.6 (Codex Ultra parity) #36391

Summary

OpenAI GPT-5.6 supports reasoning.effort: "max", which corresponds to Codex UI "Ultra". OpenCode Ctrl+T variants for native OpenAI GPT-5.6 models currently stop at xhigh, so users cannot request max effort from the TUI.

Environment

• Provider: built-in openai / @ai-sdk/openai
• Models: openai/gpt-5.6, openai/gpt-5.6-sol, openai/gpt-5.6-terra, openai/gpt-5.6-luna
• UI: model reasoning variant cycling via Ctrl+T

Steps to reproduce

1. Select openai/gpt-5.6-sol or another GPT-5.6 OpenAI model.
2. Cycle reasoning variants with Ctrl+T.
3. Observe the available variants.

Expected behavior

GPT-5.6 models that support OpenAI reasoning.effort: "max" should expose a max variant, matching Codex UI "Ultra" semantics.

The outgoing Responses request should be able to include:

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

Actual behavior

The variant list appears to stop at:

none, low, medium, high, xhigh

There is no max / Ultra variant exposed for GPT-5.6.

Suspected implementation details

ProviderTransform.variants() appears to use openaiReasoningEfforts(). For GPT-5.6-style versions, this currently maps to a list that does not include max. Separately, the OpenAI Responses option handling appears to filter max out of OpenAI reasoning efforts.

Relevant areas:

• packages/opencode/src/provider/transform.ts
• packages/llm/src/protocols/utils/openai-options.ts

Related issues

This is related to previous reasoning effort UI/API discussions, but it is not the same as missing xhigh or API-side rejection of unsupported values:

• Reasoning Effort Levels Not Reflected for OpenAI Models (GPT-5.2-Codex Extra Level Missing) #10068
• [BUG] Model options (reasoning/thinking) not forwarded to API for @ai-sdk/openai-compatible providers in headless mode #27361

Notes

Codex exposes "Ultra" for newer models and maps it to request-level max effort. It would be useful for OpenCode to expose the same capability for GPT-5.6 where the provider/model supports it.