YHH publishes open-source developer tools as esengine, including Reasonix, an AI coding assistant. Their writing examines how an agent’s conversation and prompt structure affect repeated model requests.[1][2][3]
Start with A byte-stable prompt prefix →
Contributions
Reasonix runs on a user’s computer and connects to a configured model provider. It reads project files, edits code and runs commands or tests while showing its actions. Desktop, terminal, browser and editor interfaces share the underlying program. The application distinguishes the local tool environment from the external model service, which receives the context sent to it.[2]
Ideas
In a first-person account of building Reasonix, YHH explains a design centered on keeping the beginning of successive prompts identical. Model services with prefix caching can reuse that repeated input instead of processing it from scratch. The engineering problem is preserving stable ordering and content as a conversation grows. The article concerns that particular caching workflow, rather than a universal cost guarantee for every provider.[3]
Tools, writing & talks
Writing, research & talks
Start here
An engineering account of preserving repeated prompt structure to use provider prefix caching.
Start with the article’s prompt-prefix explanation to understand the design decision. Familiarity with model API requests helps; the repository is the next step for trying the assistant.
Author[3]
Tools
- DeepSeek-Reasonix
github.comAn open-source coding assistant with local project tools and configurable external model services.
Creator[2][3]Original source