yshmr/codex-goal-draft-policy

This project is an installable Skill for OpenAI's Codex, meant to be used before turning on Codex's long running /goal feature. Its job is to draft a completion contract, a written description of what counts as done, what evidence proves it, and when to stop, so a human can review it before any long running task actually starts. The Skill itself never activates a goal. It only produces a draft for a person to approve, which then gets turned into a real goal through a separate workflow. The problem the author is trying to solve is that turning a long task into a goal just by saying keep going leaves the finish line, the stopping conditions, and fair evaluation unclear, and the Skill might also fire when it should not, for example on a normal one off request instead of a real goal drafting request. So the project separates several questions: should the Skill even be used for this request, did it check the actual repository or files for evidence, did the draft honestly capture the outcome and stopping rules, and can a human then run the approved goal separately. Alongside the Skill, the repository includes a large evaluation setup: 16 test cases for whether the Skill triggers at the right times, 12 cases for draft quality, several end to end test cases, and a script that runs everything and grades it automatically where possible, with human review for cases judgment calls are needed. The README is unusually transparent about its own results. Basic checks that do not depend on an outside AI provider pass. But the comparison meant to prove the Skill actually helps, run against a baseline without it, is marked inconclusive due to test setup contamination across all attempts, and three separate remediation phases trying to fix that setup are each marked as rejected. Only one specific end to end test, run against an exact predefined goal, is reported as passing after a revision. The author is explicit that this does not prove the Skill generally works, and that further planned tests were never executed. The project is released under the MIT license and is aimed at people building or evaluating Codex Skills and agent workflows, not general end users.