Two Hard Gates
No code ships without an approved plan (Gate 1) and a passing review (Gate 2). The agent cannot self-approve.
MeowKit
Enforced discipline for AI coding agents
Hard gates, TDD, security scanning, and human approval — so your AI agent ships production-quality code, not untested prototypes.
workflow2 hard gates
- plan
- GATE 1human approval
- build · simplify · verify
- GATE 2review verdict
- ship
Enforcement
Hard stops the agent cannot talk its way past.
4-Layer Security
Prompt injection defense across input boundary, instruction anchoring, context isolation, and output validation.
Hook-Based Enforcement
Shell hooks block sensitive file reads and source writes before they happen — not after the agent has rationalized past the rule. Rules define why; hooks enforce what.
Adversarial Review
Three parallel reviewers — Blind Hunter, Edge Case Hunter, Criteria Auditor — catch 2-3x more bugs than single-pass review.
Intelligence
The right specialist, skill, and context for every task.
Specialist Agents
Each agent owns a specific concern — planning, testing, reviewing, shipping. No two agents modify the same files.
Domain Skills
From docs retrieval to multimodal analysis, code review to QA testing. Step-file architecture loads skills on demand, keeping context tight.
Cross-Session Memory
Lessons, patterns, and costs persist across sessions. After 10 sessions, the analyst proposes CLAUDE.md improvements.
Scale-Adaptive Intelligence
Auto-classifies task complexity by domain at Phase 0. Fintech and healthcare route to COMPLEX automatically — no manual guessing.
Workflow
Discipline that survives interruptions and long sessions.
TDD Opt-In
Opt-in via --tdd or MEOWKIT_TDD=1. When enabled, failing tests must exist before implementation. Default mode keeps spike work fast; production builds opt in for strict discipline.
Structured Task System
Template-driven task files help agents resume work without losing context. Five template types with acceptance criteria, constraints, and live agent state tracking.
Party Mode
/mk:party for multi-agent deliberation. 2-4 agents debate architecture decisions with forced synthesis before any code is written.
Navigation Help
/mk:help scans plans, reviews, tests, and git to determine where you are in the pipeline and prints the single next action. Re-orient instantly after any interruption.