wtf is codex-bug-reproducer?

Codex Bug Reproducer is a skill for the Codex AI coding assistant that hunts for real bugs in a codebase and proves whether they are actually bugs before touching any code. You do not need to already know what is broken. It scans contracts, boundaries, state transitions, error paths, and how functions are actually called elsewhere in the code, then ranks the strongest candidate bugs by how much evidence supports them, how reachable they are, and how much impact they would have. A candidate is never just declared a bug from reading the code alone. The tool only counts something as proven once it builds a focused, deterministic test, gets your approval to create it, and watches that test actually fail for the predicted reason. It clearly labels the outcome with an honest status, such as reproduced, not reproduced, inconclusive, or fix proven, rather than guessing. The workflow has two separate approval gates: one before it creates any test or reproduction file, and a second, separate approval before it changes any production code to fix the bug. If its plan changes along the way, it has to stop and ask again rather than assuming earlier permission still applies. It also supports reproduce only mode, which finds and proves bugs without ever touching production code. Every run produces a JSON evidence file and a Markdown report describing the original bug, expected versus actual behavior, the minimal reproduction, the root cause, and any remaining risks. It installs into Codex with a single npx command and is released under the MIT license.