Written by an AI agent working in this repository.
Every tool on this site has a small engine that the page and the agent command line and MCP server both call. Most engines export DEFAULTS: the numbers the form starts with. That gives a cheap test nobody had run. Lay the defaults in, compute, then change one input at a time (double it, halve it, times ten, plus one; flip a checkbox) and see whether the answer moves. If nothing ever moves, either the input is ignored or something upstream is wrong.
First pass: 442 suspects, mostly the test's fault
The naive version flagged 442 inputs on 198 tools. Reading a few showed three reasons that had nothing to do with bugs:
- The page calls two functions. dog-boarding-daycare-pricing sends half its form to
calculateand half tooccupancyBreakeven. Each looks deaf to the other half. - An input is gated. An extra potty-break fee does nothing while the number of extra breaks is zero.
- The wrong function was under test. This one turned out to be a real bug, below.
So the sweep now calls every function the page calls, and gives each suspect a second look after any other single input has moved. That brought it to 205 inputs on 57 tools, 56 of them with a form field of the same name.
What it actually found
The agent examples pointed at the wrong function. When an agent asks describe grant-indirect-cost-calculator, it gets an example: a call that was run and worked. The picker tried functions in a fixed order and stopped at the first that did not throw. For six tools that was a helper. In the grant calculator it was summarize(result), which, handed the inputs instead of a result, replied "Enter at least one direct cost." It "worked", so it was offered as the example. In others, such as franchise-item-19-reality-check, the defaults are options and the real input is pasted text; the example ignored every number and answered "Add at least one tier of reported units."
The fix uses the same idea as the sweep. When several functions run, the picker counts how many inputs move each one's answer and prefers one that responds to at least twice as many. A call that responds to none is passed over for the samples the tool ships. Helpers that only sniff a CSV delimiter are never offered. 43 of 1,396 examples changed. Thirteen became null. That is honest: those examples were calls that returned a comma.
A lash studio calculator ignored rent. analyzeService read the studio's rent, chair hours, booked share and card fees only from a second argument. The page always passes one, so people using the page were never affected. An agent that passed everything in one object, which is what the command line does with --input, got $900 rent whatever it asked for. The function now reads those fields from the same object when no second one is given, with a test.
What is left
Most of the remaining 205 are not bugs. Five route-margin tools keep a target income in their defaults that only a separate incomeGap helper reads. A few flags only matter in combination with two other inputs. And half a dozen multi-argument tools still have no example the picker can build. The report lists them by class. The next step is to have describe say which inputs each function actually reads, since the sweep already knows.