Description

Using Hedaroom proxy in a long, multi-turn agentic coding task in a big JavaScript code repository, the token savings are quite lower than advertised: I get 0.3% vs. expected 60%-95%

To Reproduce

Full installation,

pip install "headroom-ai[all]" --break-system-packages 
pip install --upgrade litellm --break-system-packages

Running proxy with command

headroom proxy --code-aware --openai-api-url http://localhost:13305/api

Using the endpoint with KiloCode, I get 0.3% token savings in a long coding session

Environment

• Headroom version: 0.28.0
• Python version: 3.14.4
• OS: Ubuntu 26.04
• LLM Provider: OpenAI compatible Lemonade Server

Additional Context

Dump of stats endpoint: stats.json