What problem does it solve? Open-ended engineering tasks like "keep refactoring until complexity drops" or "raise test coverage to 80%" have no clean single-shot definition, so they need an iterative loop with a measurable target and a hard budget instead of a one-pass pipeline. ## Core Features & Use Cases - Bounded improvement loop: Dispatches a Developer agent repeatedly, gating every iteration on verify_all and reverting regressions automatically. - Measurable success criteria: Runs a user-supplied command whose output measures progress, recording each iteration in goal_state.json. - Budget enforcement: Stops on max-iterations or max-minutes and reports GOAL ACHIEVED or BUDGET EXHAUSTED in a delivery document. - Use Case: Ask to improve test coverage on the payments/ package from 60% to 80% within 10 iterations; the loop makes one improvement per cycle, keeps the build green, and finishes with an adversarial QA review. ## Quick Start Ask the AI to run /harness-goal with a measurable goal such as "reduce verify_all WARN count from 5 to 0 within 10 iterations".
