Structural gates that block AI agents from breaking architecture. Not warnings. Not suggestions. Enforcement.
Speed without structure is organized chaos. Here's what happens when AI agents run free.
The AI calls functions that don't exist in your codebase, creating phantom dependencies.
React components importing directly from database adapters. Architecture? Ignored.
API keys hardcoded in utility files. One commit away from exposure.
You asked for a login page. You got social auth, password strength meter, and SSO — unrequested.
A simple CRUD endpoint becomes a 4-layer abstraction with a factory pattern.
By message 50, your system prompt is a distant memory. The AI drifts further each turn.
CLAUDE.md and Cursor Rules are suggestions. Harness gates are enforcement.
PostToolUse hooks fire on every file write. The AI literally cannot save a file that violates the rules.
| Gate | What It Catches | Status |
|---|---|---|
check-secrets | 35+ secret patterns (AWS, GitHub, Stripe, OpenAI, …) | Default |
check-boundaries | Unauthorized imports and dependencies | Default |
check-structure | Files in wrong directories | Default |
check-spec | Incomplete specs (no TODO/TBD allowed) | Default |
check-layers | 3-tier violations (P→D skip, reverse deps) | Default |
check-security | SAST scan (Semgrep / Bandit / built-in) | Default |
check-deps | Known vulnerability audit | Default |
check-complexity | Cyclomatic complexity threshold | Opt-in |
check-mutation | Mutation testing coverage | Opt-in |
check-performance | Performance regression | Opt-in |
check-ai-antipatterns | Common AI code generation mistakes | Opt-in |
A Socratic interview measures ambiguity across 4 dimensions before a single line of code is written.
Socratic interview measuring 4D ambiguity — goal, constraints, success criteria, context.
Crystallize into an immutable seed spec. Changes mean new versions, never mutations.
Double Diamond execution: Discover → Define → Design → Deliver. D→L→P order enforced.
3-stage verification: Mechanical gates → Semantic check → Human judgment.
Wonder/Reflect/Re-seed loop. Iterates until ontology similarity ≥ 0.95.
Navigator-Driver pair programming. 3 plans generated per task. Independent test design.
Auto-detects 30+ frameworks and maps layer rules to your stack. No layer skipping. No reverse deps.
# Simplest way to get started
/plugin marketplace add studioKjm/ai-harness-template
/plugin install harness@studioKjm-harness
# Clone the template git clone https://github.com/studioKjm/ai-harness-template.git # Lite — bash only, zero dependencies ./ai-harness-template/init.sh /path/to/your-project # Pro — Python 3.11+, MCP server + observability ./ai-harness-template/pro/install.sh /path/to/your-project
Start with Lite. It delivers 80% of the value with zero dependencies. MIT Licensed.
Open source. Framework-agnostic. Drop it into any project.