edwardyap90

edwardyap90 / counterfactual-engineering-skill

Public

Let Codex test multiple fixes in isolated worktrees, compare evidence, and apply only the safest proven solution.让 Codex 不再盲改代码，而是并行测试多个方案，用证据选出最稳的修复。

ai-agents codex codex-skill engineering-tools
16
0
89% credibility
Found May 31, 2026 at 16 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Shell
AI Summary

This project is a skill for AI coding assistants that helps you make risky code changes more safely. When you have a problem where multiple solutions are possible—like a login bug, payment fix, or complex refactor—it creates isolated copies of your project, tries different approaches in each one, runs the same tests on all of them, and generates a comparison report so you can apply only the best proven solution. It's designed for high-stakes changes where you want evidence before committing.

How It Works

1
💡 You have a tricky problem to solve

You realize your code change is risky—maybe it's about login, payments, or something where the root cause isn't clear. You want to try different approaches before committing to one.

2
🔧 You ask your AI assistant for help

You tell your AI coding assistant to use the counterfactual engineering skill, and it automatically reads your project structure, checks the current state, and understands what needs fixing.

3
🌿 Your assistant grows three parallel solutions

Instead of picking the first idea, your AI creates three separate workspaces and tries a different fix in each one—minimal fix, root cause fix, and a compatibility layer.

4
✅ Each solution gets tested the same way

Your AI runs the same tests, builds, and checks against each solution so you can compare them fairly based on real evidence, not just guesswork.

5
📊 You get a clear comparison report

A report is generated showing which solution passed all tests, which had risks, and how many files each changed—so you can make a confident decision.

🏆 You apply the winning fix with confidence

You pick the solution with the strongest evidence and apply it to your real project, knowing it was thoroughly tested before merging.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 16 to 16 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is counterfactual-engineering-skill?

This is a Codex skill written in Shell and Python that forces AI to stop guessing and start proving. When you have a tricky bug or risky change, it creates multiple isolated git worktrees, implements 2-3 different candidate fixes in each one, runs the same tests against every candidate, and generates a comparison report backed by evidence. Only the fix that actually passes verification gets applied back to your real branch. Think of it as A/B testing for code changes, but with git isolation and automated scoring.

Why is it gaining traction?

The hook is simple: it removes the "trust me, this should work" from AI-assisted coding. Instead of Codex picking the first reasonable-sounding fix, this skill makes it explore multiple paths and prove which one actually holds up. The comparison report scores candidates on verification results, diff size, and risk markers for sensitive areas like auth or payments. It auto-detects your package manager and runs appropriate verification commands like npm test or pytest. The cleanup script prevents worktree clutter, and dirty worktrees are protected by default.

Who should use this?

Backend engineers working on auth, payments, or database migrations where a wrong fix means real damage. Teams who want AI to explore multiple approaches before committing, not just generate one solution. Anyone tired of AI suggesting fixes that pass a linter but break production. It works best when you have existing test suites and a clear verification matrix.

Verdict

A clever concept that addresses a real problem: blind trust in AI-generated code. The credibility score of 0.8999999761581421% reflects a very early-stage project with 16 stars and minimal community validation. The demo exists and the scripts are functional, but test coverage and documentation depth are thin. Worth watching, but wait for more adoption before betting production code on it.

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.