Reasoning-Content-Proxy: A Local Proxy Solution to Fill the Gap in Tool Calling Thought Chain for DeepSeek and Kimi

This article introduces Reasoning-Content-Proxy—a local proxy tool designed to fill the missing reasoning_content field in tool calls of domestic large models like DeepSeek-R1 and Kimi, resolving the invocation error issue when integrating with tools such as Claude Code. This solution intercepts and enhances model responses via a local proxy to achieve API compatibility, helping developers enjoy the cost advantages of domestic models while maintaining tool call stability.

In the tool calling process of modern large models, the thought chain (reasoning content) records the complete reasoning path of the model from problem analysis to tool selection, which is key to ensuring interpretability and debuggability. However, in the API responses of models like DeepSeek-R1 and Kimi, the reasoning_content field is often empty or missing, contrasting with the detailed thought feedback from models like Claude. This causes tools dependent on this field (e.g., Claude Code) to throw errors when parsing responses, interrupting the workflow.

The core process of Reasoning-Content-Proxy includes: 1. Request Forwarding: Receive OpenAI-format requests from clients and forward them to the target model API; 2. Response Parsing and Enhancement: Check the reasoning_content field—if missing, reversely generate a reasonable thought chain based on tool call information in the content; 3. Format Standardization: Convert the enhanced response into JSON that complies with the OpenAI Chat Completions API standard to ensure normal parsing by downstream tools.

This proxy has the following features: 1. Lightweight Local Deployment: Uses a lightweight architecture (e.g., FastAPI or Express) with low resource consumption and low latency; 2. Intelligent Thought Chain Generation: Parses the function name and parameters of tool calls, infers intent, and generates natural language thought descriptions (e.g., generating "To query Beijing's weather, the weather tool needs to be called" when get_weather is invoked); 3. Multi-Model Compatibility: Handles differences in authentication methods, response formats, and streaming responses of APIs from different vendors like DeepSeek/Kimi.

The main application scenarios of this tool include: 1. Claude Code Integration: Allows developers to use DeepSeek/Kimi in Claude Code, balancing cost and stability; 2. Unified Multi-Model Interface: Masks underlying model differences and simplifies client-side code; 3. Development and Debugging Assistance: The generated thought chain helps understand the model's decision-making process and optimize prompts.

The proxy has the following limitations: 1. Limited Interpretability: The generated thought chain is reversely constructed, only reflecting "what was done" rather than the real reasoning process; 2. Limited Debugging Value: It is difficult to provide sufficient information for complex errors; 3. Slight Latency: The proxy layer introduces a small amount of network latency. It is recommended to use it in development testing or cost-sensitive scenarios; for critical applications, prioritize models that natively support thought chains.

Reasoning-Content-Proxy is a practical solution to resolve tool calling compatibility issues for domestic models. Although it cannot restore the real reasoning process, it can meet most practical needs. We look forward to domestic model vendors providing complete thought chain support at the API level in the future, at which point the demand for such compatibility tools will gradually decrease.