I got tired of babysitting Codex, so I built a Project Director skill with fail-closed task ownership
I like Codex a lot, but this nearly made me lose my mind.
I gave a pair of agents one clear job make a button work. I came back about 12 hours later and the button still did not work. They had written tests, investigated unrelated code, and made “progress” everywhere except the thing I actually asked for.
The first thing that genuinely helped was embarrassingly simple: a live checklist that forced the agent to keep naming the exact outcome, what was still broken, and what proof would count.
So I kept expanding that idea into **ProjectStart**, an open-source Codex skill for running a project more like a small software team:
- one Project Director and one master checklist
- planning and prior-art research before implementation
- separate top-level Codex tasks for real product domains
- exact task IDs/deeplinks so agents can actually find each other
- deadlines and drift alarms
- mockup skeleton approval before features
- durable correction records so the same mistake does not return 10 hours later
- evidence rules that stop a build or unit test from being called a working product
The rule I care about most: a developer, QA lane, visual auditor, integrator, or other lasting role **must be a real top-level task created with `create_thread`**. It cannot be hidden inside `spawn_agent`. The validator deliberately fails if a subagent is given durable project ownership.
It is MIT licensed, has an install button, and includes the full Markdown control kit plus PowerShell initialization/validation scripts:
**GitHub: https://github.com/LiquidGlek/ProjectStart\*\*
This came directly out of wasting a ridiculous amount of time babysitting agents and correcting the same failures. I have tested the package, fresh initialization, strict controls, clean extraction, and the negative case where a durable lane is incorrectly assigned to a subagent.
If Codex has ever spent hours being impressively busy without finishing the one thing you asked for, this is for that problem.
I would genuinely like people to try to break the workflow and tell me which failure mode I missed.
