FABULA
A local coding agent that will not call a task done until the project's own tests pass, ending every run either verified with a replayable receipt or explicitly marked as not done.
FABULA is a coding agent built for one small model on one machine, served by LM Studio or MLX, or any OpenAI-compatible endpoint the developer already has. It ships as a desktop app for macOS and as a local server with a web interface for Linux and Windows, and its design bet is that reliability comes from the machinery around the model rather than from the model itself: the rule that an edit is not finished until the project's tests run green after it lives in the agent's run loop, not in a prompt the model can argue past.
When source is edited and never verified, when the last test run was red, or when tools were used but tests never ran, the engine sends the turn back to work, bounded by a limited number of re-entries; past that limit the run ends explicitly as not done. A successful run produces a receipt recording the diff, the base commit, the exact verification command and its result, the model used, and a fingerprint of the working context, which a separate command can replay against a fresh checkout of the same base to confirm the result.