Bug: reasoningSummary parameter sent to non-Claude models (GPT-5.2)

[vvgool]

Description

When using non-Claude models (e.g., GPT-5.2) in oh-my-opencode.json agent configuration, the framework still sends the reasoningSummary parameter which is Claude-specific. This causes API errors from OpenAI/LiteLLM.

Error Message

litellm.BadRequestError: OpenAIException - Unknown parameter: 'reasoningSummary'.
Received Model Group=openai.gpt-5.2
Available Model Group Fallbacks=None

and

litellm.BadRequestError: litellm.ContentPolicyViolationError: Unknown parameter: 'reasoningSummary'.
model=gpt-5.2. content_policy_fallback=None. fallbacks=None.

Environment

• oh-my-opencode version: 3.5.2
• @opencode-ai/plugin version: 1.1.59
• LiteLLM proxy backend

Configuration

opencode.json (model definitions):

{
  "provider": {
    "litellm": {
      "models": {
        "anthropic.novita.claude-opus-4-5-20251101": {
          "reasoning": true
        },
        "gpt-5.2": {
          "reasoning": false
        }
      }
    }
  }
}

oh-my-opencode.json (agent assignments):

{
  "agents": {
    "sisyphus": {
      "model": "litellm/anthropic.novita.claude-opus-4-5-20251101"
    },
    "sisyphus-junior": {
      "model": "litellm/gpt-5.2"
    },
    "oracle": {
      "model": "litellm/gpt-5.2"
    }
  }
}

Expected Behavior

When calling a model that doesn't support reasoning (like GPT-5.2), the framework should NOT include Claude-specific parameters like reasoningSummary in the API request.

Actual Behavior

The reasoningSummary parameter is sent to GPT-5.2 models, causing the API to reject the request with "Unknown parameter" error.

Suggested Fix

Filter out Claude-specific parameters (reasoningSummary, and potentially other extended thinking parameters) when the target model is not a Claude model, or when reasoning: false is set in the model configuration.

[solssak]

I've submitted a fix for this: #1794

The root cause was isGptModel() only matching openai/ and github-copilot/gpt- prefixes — models behind proxy providers like litellm/gpt-5.2 fell through to the Claude code path, getting thinking config which the runtime translated to reasoningSummary.