oh-my-fable
A model-agnostic agent harness that plans a long task first, self-corrects at every step, and checkpoints its whole run, so a crash is just a pause instead of a restart from zero.
The entire run lives in one serializable RunContext, checkpointed after every step, so resume() continues exactly where a crashed process left off. A step that keeps failing is marked blocked and forces a replan instead of looping forever, and a final check verifies the stated success criteria before the run can declare the goal met.
It can drive Claude Code itself as the durable, tool-using agent, acting as the planner and reflector wrapped around an existing coding-agent subscription rather than requiring a separate, per-token API key.
It includes a scripted provider for deterministic, offline testing of the agent loop with no network calls, and is written in TypeScript with zero dependencies.