I built a small OSS tool for agent workflows: GitHub - Olie0512/goal-run: Tiny CLI that runs a GOAL.md until a verifier says done · GitHub (MIT).
Problem: coding agents love checklists and mark them done without proof.
Idea: GOAL.md + shell checks. goal-run done refuses self-grade — every check must exit 0.
pip install -e . # venv recommended
goal-run init "Ship MVP"
goal-run check # RED / GREEN
goal-run check --json # for CI
goal-run done # archives only if green
Also ships a GitHub Action wrapper. No LLM inside the CLI.
Would love feedback from folks wiring agents into CI — does this fit your loop, or do you already solve it differently?