Devotd Github
NewAI agents

ranxianglei/billion-context — a plugin that compresses coding agent context hierarchically

ranxianglei/billion-context is a TypeScript plugin that compresses long coding agent conversations hierarchically so a session can continue in a small context window.

Repositoryranxianglei/billion-context ↗

ranxianglei/billion-context is a TypeScript plugin that compresses long coding agent conversations hierarchically so a session can continue in a small context window.

GitHub description: A context-compression plugin for month-long single sessions (billions of tokens) ,small context windows (a 100K context is enough) and token savings (5x fewer tokens) .上下文压缩插件,月级长会话会话, 兼顾小窗口(100k上下文足矣)省token(省5倍token) . billion-context is all you need

stars
★ 3K
Forks
246
Language
TypeScript
License
NOASSERTION
Latest release
v0.1.208
Read original source ↗View on GitHub ↗

ranxianglei/billion-context is a TypeScript plugin that hierarchically compresses the long conversation history of a coding agent so it can be carried on in a small context window. The latest release identifiable from the provided material is v0.1.208. What specifically changed in that version is not stated in the provided material.

The plugin is described as a structure that summarizes the conversation accumulating during long tasks step by step and holds it in the context window, and that passes the compressed content between multiple coding tools and models. Hierarchical compression is a way of keeping the needed context by stacking summaries stage by stage instead of shrinking the whole conversation in one pass. The goal is to keep working on a single session with a small context window even as the record grows long. The README cites long-running sessions on the scale of hundreds of millions of tokens and a 100K context window as its goals, and it explains integration with Claude Code, Codex, OpenCode and pi, along with how to check installation and cache state. The repository also contains a paper draft and English and Chinese documentation.

The README also claims a 5x reduction in tokens, but that is a figure written in the read material, and no independent evidence of reproduction has been confirmed. The design aimed at long-running sessions and a small context window is therefore confirmed, but expectations about the size of the reduction or the actual scale are hard to judge from the published description alone.