n8n-nodes-deepseek-reasoner

DeepSeek R1/Reasoner Chat Model for n8n with proper easoning_content handling for tool calls.

What This Fixes

When using DeepSeek R1 (deepseek-reasoner) with n8n's AI Agent node and tools, you get this error:

Missing reasoning_content field in the assistant message at message index 2

This happens because the built-in DeepSeek Chat Model node uses LangChain's ChatOpenAI, which doesn't preserve the easoning_content field from DeepSeek's API responses. During multi-step tool-calling loops, LangChain converts messages back to API format but drops easoning_content, causing the DeepSeek API to reject the request.

This node fixes that by implementing a custom LangChain BaseChatModel that:

1. Captures easoning_content from DeepSeek API responses
2. Stores it in the LangChain message's dditional_kwargs
3. Re-includes it when converting messages back to API format on subsequent tool-call turns

How to Use

1. Install this community node in your n8n instance
2. In your workflow, use the AI Agent node as usual
3. Instead of the built-in "DeepSeek Chat Model", connect DeepSeek Reasoner Chat Model as the Chat Model sub-node
4. Configure your DeepSeek API credentials
5. Connect your tools the reasoning_content error is now fixed!

Workflow Setup

[AI Agent] <-- Chat Model: [DeepSeek Reasoner Chat Model]
           <-- Tools: [Your Tool 1], [Your Tool 2], ...
           <-- Memory: [Optional Memory Node]

Credentials

Create a "DeepSeek Reasoner API" credential with:

• API Key: Your DeepSeek API key from https://platform.deepseek.com/
• Base URL: https://api.deepseek.com (default, change if using a proxy)

Models

| Model | Description | |-------|-------------| | deepseek-reasoner | DeepSeek R1 always uses thinking/reasoning mode | | deepseek-chat | DeepSeek V3 standard chat model | | deepseek-v4-pro | DeepSeek V4 Pro |

Options

| Option | Default | Description | |--------|---------|-------------| | Max Tokens | 8192 | Maximum tokens to generate (up to 65536) | | Temperature | 0.7 | Controls randomness (ignored by deepseek-reasoner) | | Top P | 1 | Nucleus sampling (ignored by deepseek-reasoner) | | Frequency Penalty | 0 | Penalize repeated tokens | | Presence Penalty | 0 | Penalize tokens already in text | | Max Retries | 2 | Retry attempts on API failure |

Technical Details

The core fix works by:

• On API response: When DeepSeek returns easoning_content in an assistant message (during thinking/reasoning), the node stores it in dditional_kwargs.reasoning_content on the LangChain AIMessage
• On next API call: When the AI Agent sends the conversation back for the next tool-call turn, the node's convertMessages method reads dditional_kwargs.reasoning_content and includes it as a top-level field in the API request body
• Result: The DeepSeek API receives the required easoning_content field and processes the tool-call loop correctly

Licenseit

MIT