tseno26/probefish

probefish is a skill for Claude Code that catches a specific kind of AI agent mistake: silently losing data when it changes the shape of code. This often happens when an agent merges several near duplicate functions into one, and the resulting diff looks clean and the tests still pass, but a field or edge case that only one of the original functions handled has quietly disappeared. Nothing errors, and the loss might not surface until much later, or never. The tool works by having the agent plant a small, verified test called a marker before making a change that moves data between forms, such as from a form to app state to an API to a database. Each marker pins down one specific thing that must survive the change. To make sure a marker actually works, it has to call the real production code rather than a copy of itself, and the process includes deliberately breaking the protected behavior to confirm the marker turns red, proving it is not a test that would pass no matter what. Every planted marker is also recorded by hand in a separate file called a census. A reconciler test compares the markers actually present in the test suite against what the census says should exist, catching two problems: a marker that was quietly deleted later, called lost, and a marker present but never recorded, called an orphan. Keeping the census hand written and separate from the code is what makes a later deletion show up as a mismatch instead of just shrinking the test count unnoticed. The project is inspired by a lab technique called Fluorescence In Situ Hybridization, which uses a glowing probe to reveal one specific kind of genetic damage that would otherwise be invisible. It is meant to complement another skill called ponytail, which focuses on preventing unnecessary code bloat, while probefish focuses on making sure data survives when that code does get consolidated. Installation is a matter of copying a SKILL.md file into Claude Code's skills folder, either globally or per project.