-
-
Judge path: one local Python command, zero credentials, standalone HTML and JSON.
-
Ancient Paths proof: 360 shipped levels, 69 tests, and 180 future runway levels.
-
The exact $50 boundary defect is reproduced, repaired, and replayed.
-
PatchCase proves itself: 5/5 release claims verified, including path-escape protection.
-
PatchCase turns every agent-made change into a replayable case file.
-
From claim to counterexample to deterministic check to preserved evidence.
Inspiration
I use Codex on all my projects. I'm not great at writing code, so I rely heavily on AI to help bring my ideas into reality. One thing I never trusted was when GPT would give a confident answer about a change in code/etc, and whether it was better, or more accurate than the previous version.
I wanted something that I would ACTUALLY use. It's a case file that is made to ensure the agent states what it believes, what could disprove it and what was/how it was checked.
What it does
It's a plugin and local Python program where I ask Codex to challenge the current change made and generate a report. It reads the report and scrutinizes it heavily. It argues against itself (with help from Python, so it's not "biased") and determines truth/validity of what it initially claimed. In other words, Python calculates final verdicts deterministically - and this happens outside the model.
How I built it
Entirely with Codex. It did all the research and coding logic for me.
Challenges I ran into
Finding a suitable name for this project - to begin with. Making it useful for others beyond my personal use of it.
What I'm proud of
The fact that it found a packaging defect in itself in its release. And it helped me with my own game I am developing. It's a Codex skill that anyone can use. I made it accessible to the public on GitHub. It has real project evidence from what I've tested it on, and is self-testing. It runs on every project wothout me primpting Codex to do so, whixh makes my life easier.
What I learned
In short, the evidence has to be allowed to disagree with the agent and/or developer. AI can't assume the developer is right, let alone "think" it is right either.
What's next for PatchCase
That remains to be seen. But, getting a team onboard to make this an even more useful tool would be the next logical step.
Log in or sign up for Devpost to join the conversation.