fix(provider): preserve GPT-5.6 max reasoning effort in local provider config

[LYY]

Description

OpenCode supports GPT-5.6 models with reasoning.effort: "max", but the current fix appears to work only for model metadata loaded from models.dev.

When the same model is defined manually in opencode.json, OpenCode exposes a max variant but does not bind it to reasoningEffort: "max". Selecting max then sends medium instead.

Environment

• OpenCode: local checkout containing PR fix(provider): derive variants from reasoning options #36543
• Provider SDK: @ai-sdk/openai
• Custom provider endpoint: OpenAI-compatible /v1 endpoint
• Model: gpt-5.6-luna
• Upstream proxy: Sub2API

Configuration

{
  "provider": {
    "openai-b1": {
      "id": "openai-b1",
      "npm": "@ai-sdk/openai",
      "options": {
        "baseURL": "https://example.invalid/v1",
        "apiKey": "REDACTED"
      },
      "models": {
        "gpt-5.6-luna": {
          "id": "gpt-5.6-luna",
          "reasoning": true,
          "reasoning_options": [
            {
              "type": "effort",
              "values": ["none", "low", "medium", "high", "xhigh", "max"]
            }
          ],
          "variants": {
            "low": {},
            "medium": {},
            "high": {},
            "xhigh": {},
            "max": {}
          }
        }
      }
    }
  }
}

Steps To Reproduce

1. Configure a GPT-5.6 model manually in opencode.json.
2. Add max to reasoning_options.
3. Add an empty max entry under variants.
4. Select the max reasoning variant.
5. Inspect the outgoing request or the upstream provider request log.

Expected Behavior

The outgoing OpenAI Responses request should contain:

{
  "reasoning": {
    "effort": "max"
  }
}

Actual Behavior

The max variant appears in the local configuration, but an empty variant does not generate the corresponding reasoning option. The request is sent with medium instead. high and xhigh work correctly.

An explicit variant configuration does work as a workaround:

"max": {
  "reasoningEffort": "max"
}

Suspected Cause

PR #36543 adds reasoning_options handling for models loaded from models.dev: fromModelsDevModel() copies the field and reasoningVariants() derives variants from it.

However, the local opencode.json provider model path rebuilds parsedModel without copying model.reasoning_options. It then falls back to ProviderTransform.variants(parsedModel), whose GPT-5.6 fallback does not include max.

The relevant path appears to be packages/opencode/src/provider/provider.ts around lines 1417-1505. The local variants merge preserves max: {}, but does not attach { reasoningEffort: "max" }.

Related Issues And PR

• GPT-5.6 models missing max reasoning effort variant #36141: GPT-5.6 models missing max reasoning effort variant
• [FEATURE]: expose reasoning effort "max" for GPT-5.6 (Codex Ultra parity) #36391: expose reasoning effort max for GPT-5.6
• fix(provider): derive variants from reasoning options #36543: fix(provider): derive variants from reasoning options

[LYY]

After further investigation, this is expected behavior rather than an OpenCode bug.

A local variant named max does not implicitly set reasoningEffort to max. The correct configuration is:

"max": {
  "reasoningEffort": "max"
}

With this explicit configuration, GPT-5.6 max reasoning works correctly in the reported setup. The reasoning_options field is model metadata used by the models.dev path; it is not the local opencode.json mechanism for configuring a variant.

Closing this issue as invalid.