[Bug] DeepSeek V4 fails with 'reasoning_content must be passed back' error in multi-turn conversations

[LieGeee]

Bug Description

When using DeepSeek V4 with OpenCode, multi-turn conversations fail with the error:

The reasoning_content in the thinking mode must be passed back to the API

Steps to Reproduce

1. Configure DeepSeek V4 in opencode.json:

"deepseek": {
  "npm": "@ai-sdk/openai-compatible",
  "name": "DeepSeek",
  "options": {
    "baseURL": "https://api.deepseek.com",
    "apiKey": "sk-xxx"
  },
  "models": {
    "deepseek-reasoner": {
      "name": "DeepSeek-V4",
      "limit": {
        "context": 1048576,
        "output": 262144
      }
    }
  }
}

2. Start a conversation with DeepSeek V4
3. Send a second message in the same conversation
4. Error occurs

Expected Behavior

Multi-turn conversations should work seamlessly, with reasoning_content properly preserved between turns.

Actual Behavior

The conversation fails because @ai-sdk/openai-compatible doesn't preserve reasoning_content from DeepSeek's thinking mode responses.

Environment

• OpenCode version: Latest
• DeepSeek model: deepseek-reasoner (V4)
• SDK: @ai-sdk/openai-compatible

Root Cause

DeepSeek V4 enables thinking mode by default and returns reasoning_content in responses. The OpenAI-compatible SDK doesn't handle this field, causing it to be lost between conversation turns.

Workaround

Switch to Anthropic API format, which natively supports reasoning modes:

"deepseekv4": {
  "npm": "@ai-sdk/anthropic",
  "name": "DeepSeek",
  "options": {
    "baseURL": "https://api.deepseek.com/anthropic",
    "apiKey": "sk-xxx"
  },
  "models": {
    "deepseek-v4-pro": {
      "name": "DeepSeek-V4-Pro",
      "limit": {
        "context": 1048576,
        "output": 262144
      },
      "options": {
        "thinking": {
          "type": "enabled",
          "budgetTokens": 8192
        }
      }
    }
  }
}

This configuration:

• Uses @ai-sdk/anthropic instead of @ai-sdk/openai-compatible
• Points to DeepSeek's Anthropic-compatible endpoint
• Properly preserves reasoning_content across conversation turns
• Allows explicit thinking mode configuration

Alternative Solutions

Community members have also suggested using transformRequestBody hooks to manually preserve reasoning_content, but the Anthropic API approach is cleaner and officially supported.

References

• DeepSeek API docs: https://api-docs.deepseek.com/
• Community discussion: feat (docs): add link to storage section vercel/ai#4768

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• DeepSeek thinking mode: reasoning_content must be passed back to API on conversation continuation #24104: DeepSeek thinking mode: reasoning_content must be passed back to API on conversation continuation (same error, same root cause)
• Support DeepSeek V4 Thinking Mode with reasoning_content handling #24114: Support DeepSeek V4 Thinking Mode with reasoning_content handling (same model, same issue)
• The reasoning_content in the thinking mode must be passed back to the API. #24083: The reasoning_content in the thinking mode must be passed back to the API (closed, exact same error message)
• Missing reasoning_content field in the assistant message for DeepSeek API implementation #17523: Missing reasoning_content field in the assistant message for DeepSeek API implementation (same underlying problem)

[lfaoro]

  "provider": {
    "deepseek": {
      "models": {
        "deepseek-v4-flash": {
          "interleaved": {
            "field": "reasoning_content"
          }
        },
        "deepseek-v4-pro": {
          "interleaved": {
            "field": "reasoning_content"
          }
        }
      }
    }
  }

adding this to opencode.json fixes it