kimi-tool-call-fixer

An OpenAI-compatible proxy that fixes Kimi K2.5's raw tool call tokens in streaming and non-streaming responses.

The Problem

Kimi K2.5 sometimes emits tool call special tokens (<|tool_calls_section_begin|>, <|tool_call_begin|>, etc.) inside reasoning_content or content instead of producing native tool_calls in the API response. This breaks any client expecting standard OpenAI-format tool calls.

What This Does

This proxy sits between your client and the Kimi-compatible API server. It forwards all requests unmodified and intercepts responses, parsing raw tool call tokens into proper OpenAI-format tool_calls with correct finish_reason: "tool_calls". Responses without raw tokens pass through untouched.

Usage

Docker (recommended)

git clone https://github.com/DanielusG/kimi-k2.5-tool-fix-proxy.git
cd kimi-tool-call-fixer
SERVER_URL=https://your-provider-address/v1 docker compose up -d

Then point your client at http://localhost:8080 instead of the upstream API.

Manual

pip install -r requirements.txt
SERVER_URL=https://your-provider-address/v1 python main.py

Configuration

Variable	Default	Description
SERVER_URL	http://localhost:8000	Upstream API base URL
PORT	8080	Proxy listen port
HOST	0.0.0.0	Bind address
TIMEOUT	600	Upstream request timeout (seconds)
LOG_LEVEL	INFO	Logging level

Tests

pip install pytest
pytest test_main.py -v

License

MIT