Skip to content
v0.1 — Now Available

teststop
Break it before your users do.

Trigger AI to test any software system the way a real adversarial user would — someone who never read the docs, retries when things are slow, and does what no spec ever imagined.

Get Started → How It Works GitHub ↗ 
go install github.com/shaifulshabuj/teststop/cmd/teststop@latest
teststop run

The Problem With Test Coverage

Every test ever written was written by someone who knew how the system works.

Real users don't know. That's why production still surprises us.

We call it test coverage. What we actually have is assumption coverage.

🎯
### Zero Configuration `teststop run` works on any project with no setup. Point it at any directory — Go, Python, TypeScript, Ruby, anything — and it learns the system by reading it.
🤖
### AI-Native, No API Keys Uses `claude` or `copilot` CLI already on your PATH. No SDK, no secrets to manage, no lock-in. The AI thinks adversarially because the mandate tells it to.
📉
### Tests Reduce Over Time Confidence scores persist per area. Proven stable areas get tested less. New or changed areas get tested more. After ~15 clean passes, an area retires.
🔗
### Agent-Native Output JSON output is designed for AI coding agents. Structured exit codes signal deploy safety. teststop fits cleanly into any automated workflow.
🛡️
### Sandbox Isolation On macOS with Apple Container, teststop runs the AI inside an isolated VM. The AI cannot touch your host filesystem beyond the mounted project path.
🌍
### Universal Works on any language, any age, any system type — web apps, APIs, CLIs, data pipelines, libraries. If it has code, teststop can test it.

How It Works in 30 Seconds

1

Scan

teststop walks your project tree, detects the language and system type, and extracts routes, flows, and dependencies — all statically, no code execution.

2

Compose Mandate

Injects project context and accumulated memory into the base mandate — the adversarial instruction that tells the AI how a real user would break this specific system.

3

Generate Scenarios

Calls claude -p or copilot -p with the mandate. The AI returns structured JSON: scenario IDs, steps, chaos factors, failure modes, and priorities.

4

Update Memory

Confidence scores update per area. High confidence → less testing next run. Retirement at 0.95+ with ≥15 passes. The system gets smarter with every run.

5

Report & Exit

Outputs JSON, text, or markdown. Exits with a machine-readable code: 0 = safe to deploy, 1 = review needed, 2 = critical failures found.

Quick Install

go install github.com/shaifulshabuj/teststop/cmd/teststop@latest

Download the latest binary for your platform from GitHub Releases.

# macOS arm64 example
curl -L https://github.com/shaifulshabuj/teststop/releases/latest/download/teststop_Darwin_arm64.tar.gz \
  | tar xz
sudo mv teststop /usr/local/bin/

git clone https://github.com/shaifulshabuj/teststop
cd teststop
go build -o teststop ./cmd/teststop

Prerequisite: claude or copilot CLI must be on your PATH.

Full installation guide →

Part of a Trilogy

DocuFlow  → gives AI the context to act with purpose
Waymark   → gives humans the reason to trust and step back
teststop  → gives systems the confidence to prove themselves
Tool Role Link
DocuFlow MCP server — LLM wiki for AI context GitHub
Waymark MCP middleware — AI agent governance GitHub
teststop CLI — adversarial testing trigger This repo