wtf is iffable?

Iffable is a guardrail tool for Claude Code, the AI coding assistant, aimed at people on a subscription plan where the most capable model has a limited weekly usage cap. That top model, referred to in the README as Claude Fable 5, drains its quota quickly if it is left to handle every small task, so iffable tries to push simple work toward cheaper models automatically. The tool only turns itself on when a session was started using the Fable model. If a session starts on any other model, it stays inactive except for one rule that always applies: it blocks a certain cheap model from being spawned as a helper at all, no matter which main model is running. When active, iffable adds three checks. It blocks spawning that lowest tier helper model entirely. It also blocks a large, unplanned piece of delegated work from starting unless a written requirements list, called a ledger, already exists in the project. And it blocks the AI from ending its turn while that ledger still has unfinished items checked off as open. If every check fails to parse correctly or hits an internal error, the tool is designed to step aside quietly rather than block anything, so a bug in iffable itself can never lock up a session. When armed, iffable also injects a written policy describing which model tier should handle which kind of work: the top model for judgment calls and debugging, a strong second tier model for heavy reasoning tasks, a simple tier for routine work like formatting or renaming, and the cheapest tier banned outright. Installation is done by running a shell script that copies the tool's files into a folder in the user's home directory, after which the user manually wires it into their Claude Code configuration. A handful of environment variables let you override its automatic behavior. It is released under the Apache 2.0 license.