What is fable_os?

This project is a ready made template that turns Claude Code, an AI coding assistant, into a small automated team that quietly maintains one of your software projects for you. Once a day, or whenever you run a single command, the team looks at what has changed recently, such as new code, open issues, or failing tests, and decides whether there is one small safe task worth doing, like cleaning up a warning or fixing a flaky test. Rather than editing your real code directly, the system does the work in a private copy of your project first. It checks its own work twice, once by having another AI review it and once by actually running your project's real tests, before showing you anything. Only after a particular type of task has proven itself successful many times in a row is it allowed to be applied automatically. Until then, everything is left for you to review and approve by hand. The work is split across four roles, each handled by a different AI model chosen to balance cost against capability. One role glances at what is new and decides if anything is worth doing, another picks the single best task and writes clear instructions for it, a third performs the work, and a fourth checks the finished result against those instructions. By default the thinking roles use the Opus model, while the hands on work uses the Sonnet model, with an optional setting to use a more capable model called Fable for the thinking roles if you have access to it. To use the system, you copy a folder along with two configuration files into the top level of any project that has a test command. Inside that folder are the instructions each AI role follows, a notebook style set of files that track what happened in each run and how trustworthy each type of task has proven to be, a place to define ongoing goals that get rechecked every day so they stay fixed, and a small library of reusable task recipes the team already knows, such as fixing lint warnings or updating dependencies. The project is built with shell scripts and designed for macOS. It was adapted from an online guide describing how to build this kind of automated agent system, with the original commands rewritten to work with Claude Code and Apple's built in command line tools.