GPT-OSS Adapter

A middleware proxy that manages chain-of-thought reasoning for GPT-OSS models.

Overview

This adapter sits between client applications and inference servers to manage chain-of-thought reasoning for GPT-OSS models. GPT-OSS models work better when they receive reasoning context from prior tool calls, but most clients don't include this context in subsequent requests. It handles this by:

• Mapping: Automatically maps fields based on the target provider (LM Studio, llama.cpp)

• Caching: Stores reasoning content from tool call responses and automatically injects it into subsequent requests

• Proxying: Sits between clients and inference servers to manage reasoning content

Installation

Build from source:

go build -o gpt-oss-adapter

Usage

gpt-oss-adapter --target http://localhost:8080 --listen :8005

Command Line Options

• --target, -t: Target server URL (required)
• --listen, -l: Server listen address (default: :8005)
• --verbose, -v: Enable debug logging
• --provider, -p: Target provider type (lmstudio, llama-cpp)

Provider Support

The adapter automatically handles field mapping based on the target provider:

LM Studio (lmstudio)

• Reasoning field: reasoning
• Reasoning effort: reasoning_effort

llama.cpp (llamacpp)

• Reasoning field: reasoning_content
• Reasoning effort: chat_template_kwargs.reasoning_effort

Reasoning Effort Support

The adapter automatically extracts reasoning.effort from client requests and maps it to the appropriate provider field:

• Input: {"reasoning": {"effort": "high"}}
• LM Studio: Maps to reasoning_effort
• llama.cpp: Maps to chat_template_kwargs.reasoning_effort

Examples

# Proxy requests to LM Studio
gpt-oss-adapter \
  --target http://localhost:1234 \
  --listen :8005 \
  --provider lmstudio \
  --verbose

# Proxy requests to llama.cpp server
gpt-oss-adapter \
  --target http://localhost:8000 \
  --listen :8005 \
  --provider llama-cpp \
  --verbose

Supported Endpoints

The adapter handles these OpenAI-compatible endpoints:

• /v1/chat/completions
• /chat/completions

Other endpoints pass through unchanged.

License

MIT