Agentic KDD v3.6
Autonomous development pipeline with persistent knowledge memory. Forces TDD, blocks spec violations before build, and eliminates repeated errors across sessions.
What it is
Agentic KDD is a framework that runs inside Cursor and Claude Code. It turns your AI assistant into a disciplined development pipeline with:
- KDD Memory — errors, patterns and decisions persist across sessions. The agent reads them before every task.
- TDD Gate — deterministic self-healing loop. Forces tests to pass before continuing. Not a suggestion — it's code.
- Spec Gate — blocks changes that contradict documented business rules before writing a single line.
- Security Gate — detects JWT bypass, cross-tenant access and debug flags before build.
- Regression Guard — protects healthy behaviors across cascading file changes.
- Contract Guard — tracks passing test patterns and escalates them to protected contracts over time.
Benchmark results (19 phases, Node.js + Python stacks)
| Metric | Without | With | Change |
|---|---|---|---|
| Errors per phase | 2.6 | 0.0 | 100% |
| Repeated errors | 3 | 0 | 100% |
| Tests passing first try | 79% | 100% | +27pp |
| Spec drift detections | 3 | 6 | +100% |
| Cascade files correct | 4 | 11 | +175% |
| Security issues detected | 1 | 2 | +100% |
| Autonomous fixes | 4 | 7 | +75% |
N=1 project per benchmark. Treat as directional, not definitive. Reproducible benchmark repo in progress.
Installation
npm install -g agentic-kddRequires Node.js 18+.
Setup
cd your-project
akdd initThat's it. Agentic KDD detects your stack (Node.js, Python, PHP), installs .agentic/, injects the dev:kdd script, and configures the pipeline.
For brownfield projects (existing codebase):
akdd onboardScans your project, maps the stack, suggests a first small task, and pre-populates memory with what it finds.
Usage
Every task goes through aa::
aa: implement the clients module with CRUD and tenant isolationThe pipeline runs automatically:
① Analyst — reads memory, recalls patterns and past errors
② Spec Gate — blocks if prompt contradicts HIGH-confidence rules
③ Security Gate — checks CRITICAL/SENSITIVE files before build
④ Regression Check — verifies changeset won't break protected behaviors
⑤ Build — implementation
⑥ TDD Gate — runs tests, self-heals up to 3 iterations
⑦ QA — validates against spec and memory
⑧ Preservation Gate — verifies contracts
⑨ Regression Register — snapshots healthy behaviors
⑩ Memory — writes errors, patterns and decisions
⑪ Creative — detects improvement opportunitiesCLI commands
akdd init # Install in a new project
akdd onboard # Analyze an existing project (brownfield)
akdd update # Pull latest from GitHub
akdd contracts # Contract Guard status
akdd contracts list # List all contracts
akdd graph # Knowledge graph state
akdd metrics # Project metrics
akdd health # System health check
akdd analyze # Cross-artifact consistency checkWhat's in the repo
.agentic/
├── agentes/ # Role instructions (Orchestrator, Analyst, Back, Front, QA, Memory)
├── grafo/ # Gate modules (tdd-gate, spec-gate, security-gate, regression-guard, contract-guard)
├── memoria/ # KDD memory files (errors, patterns, decisions, work)
├── config.md # Project configuration
└── PLAN.md # Active task
CLAUDE.md # Pipeline rules for Claude Code
.cursor/rules/ # Pipeline rules for CursorHow it compares
| OpenSpec | Spec-Kit | Agentic KDD | |
|---|---|---|---|
| Spec violation gate | blocks before build | ||
| Security gate | blocks before build | ||
| Error memory (KDD) | persists across sessions | ||
| TDD self-healing loop | Partial | deterministic, 3 iterations | |
| Regression protection | cascade-tested | ||
| Contract escalation | candidate → verified → protected | ||
| Python support | pytest + FastAPI | ||
| Brownfield onboarding | Partial | akdd onboard | |
| npm install | |||
| MCP tools | 25+ | — | 59 |
Stack support
- Node.js — Next.js, Express, Fastify, NestJS + Jest/Vitest
- Python — FastAPI, Django, Flask + pytest
- PHP — Laravel + PHPUnit
License
MIT — Adrián López (@Adrianlpz211)