Claude Code plugin · MIT · v0.1.0

falsify

A Claude Code plugin that takes a ticket to a pull request and refuses to skip the proof at any step.

What it refuses to skip

  1. 1 · evidence

    No branch until evidence shows the defect exists.

    A root cause written in the ticket is a hypothesis to disprove. A verdict of configuration, as designed, or not built ends the run with an explanation.

  2. 2 · probe

    No fix until one test fails with the value the reporter saw.

    The probe lives in the spec that owns the code. Its actual value has to equal the reported one. That equality is the proof.

  3. 3 · score

    No pull request until the computed score reads high.

    The score is printed from the gates' JSON results; the model cannot edit it. Medium opens a draft and waits for CI.

How a falsify run goes: ticket, evidence check, probe, worklist and build, gate battery, QA by a fresh agent, computed score, pull request

The gates

Each gate detects one class of defect, has a source showing the class is real, and cannot pass while the property is false.

Gate Catches
size Review blindness on large diffs
types Type errors, including in files the diff did not touch
fan-out The file you forgot
behaviors The fix does not do what the ticket says
tamper Tests weakened to pass
regression Broken existing behavior
complexity Code the next reader cannot hold
mutation Tests that run the code and assert nothing
real-world It works in the test and fails in the app
Sources and thresholds for every gate →

Install

Add the plugin, then run setup once per repository. Setup asks where tickets live, what the test command is and which branch is the base, and runs each answer before accepting it.

Claude Code
/plugin marketplace add maxymlyskov/falsify
/plugin install falsify@falsify
/falsify:setup

Every contribution went through it

Every open-source contribution on this site was found, proven and sent with falsify.

See the contributions →