AgentOS Public — Documentation skill for Claude Code

AgentOS: project memory for coding agents

[](https://github.com/fasterquieter/AgentOS-Public/actions/workflows/ci.yml)

AgentOS helps a coding assistant remember what a software project is, what is happening now, what must not be broken, and what the evidence actually proves. It keeps that memory beside the code and routes only the useful slice into each task.

This repository has two supported roles:

1. the master AgentOS framework-development project; and
2. a complete, self-contained seed that can be cloned or duplicated to start a different project.

AgentOS in plain English

Imagine a very clever builder who arrives each morning with no memory of yesterday. Source code shows the builder what exists, but not every promise, reason, unfinished job, or past mistake. One enormous handbook would be just as unhelpful: the warning that matters could disappear in the pile.

AgentOS gives the project a small front desk. It tells the builder what matters, what is happening now, what must not be broken, what has actually been proved, and where to find deeper notes for this particular job. The guiding rule is: write down what is expensive to forget, label it honestly, and route it only when useful. Keep the memory system smaller than the confusion it prevents.

How it works

Three layers do the work, and all of them are plain files in the repository:

1. AGENTS.md is a small, always-visible router. Claude Code, Gemini CLI, and Cursor reach it through thin adapters (CLAUDE.md, GEMINI.md, .cursor/rules/); Codex reads it directly. Project truth is never copied into an adapter.
2. .agentos/index.json is an explicit, human-readable catalog. Each entry says what a document is, which task words and target paths make it relevant, what must be read with it, what kind of authority it carries (contract, decision, evidence, working, `historic